content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Throughout, $\kk$ is a field and all algebras are $\kk$-algebras.
All isomorphisms should be read as `isomorphisms as $\kk$-algebras'.
Our primary source for all definitions is \cite{BG}.
\begin{hypothesis}
\label{hyp.cga}
$A$ is a connected graded algebra finitely
generated over $\kk$ in degree $1$.
\end{hypothesis}
Let $R$ and $S$ be algebras satisfying Hypothesis \ref{hyp.cga}
with bases $\{x_i\}$ and $\{y_i\}$, respectively.
Then $R$ and $S$ are isomorphic as graded algebras
if there exists an algebra isomorphism $\Phi:R \rightarrow S$
such that $\Phi(x_i) = \sum \alpha_{ij} y_j$, $\alpha_{ij} \in \kk$,
for each $x_i$.
If $a_{ij}\neq 0$, then we say $y_j$ is a {\sf summand} of $\Phi(x_i)$.
Because of the following result,
we will often assume without comment that isomorphisms
between graded algebras are graded isomorphisms.
\begin{thm}[Bell, Zhang {\cite[Theorem 0.1]{BZ}}]
Let $A$ and $B$ be two algebras satisfying Hypothesis \ref{hyp.cga}.
If $A \iso B$ as ungraded algebras, then $A \iso B$ as graded algebras.
\end{thm}
A square matrix $\bp = (p_{ij}) \in\cM_n(\kk^\times)$ is
{\sf multiplicatively antisymmetric} if
$p_{ii} = 1$ and $p_{ij}=p_{ji}\inv$ for all $i \neq j$.
Let $\cA_n \subset \cM_n(\kk^\times)$ be the subset of
multiplicatively antisymmetric matrices.
A matrix $\bq \in \cA_n$ is a {\sf permutation} of $\bp$
if there exists a permutation $\sigma \in \cS_n$
such that $q_{ij}=p_{\sigma(i)\sigma(j)}$ for all $i,j$.
For $\bp \in \cA_n$, the {\sf (multiparameter) quantum affine $n$-space}
$\cO_{\bp}(\kk^n)$ is defined as the algebra with basis $\{x_i\}$,
$1 \leq i \leq n$, subject to the relations $x_ix_j=p_{ij}x_jx_i$
for all $1 \leq i,j \leq n$.
\begin{thm}[Theorem \ref{thm.qas}]
$\cO_{\bp}(\kk^n) \iso \cO_{\bq}(\kk^m)$ if and only if $m=n$ and $\bp$ is a permutation of $\bq$.
\end{thm}
Mori proved Theorem \ref{thm.qas} when $n=3$ \cite[Example 4.10]{M}
and this was extended to all $n$ by Vitoria \cite[Lemma 2.3]{V}.
We present a simple, self-contained proof that does not
rely on the noncommutative projective algebraic geometry associated to $\cO_{\bp}(\kk^n)$.
{\sf (Multiparameter) quantized Weyl algebras} may be regarded
as $\gamma$-difference operators on $\cO_{\bp}(\kk^n)$.
Let $\bp \in \cA_n$ and
$\gamma=(\gamma_1,\hdots,\gamma_n) \in (\kk^\times)^n$.
Then $A_n^{\bp,\gamma}(\kk)$ is the algebra with basis $\{x_i,y_i\}$,
$1 \leq i \leq n$, subject to the relations
\begin{align*}
y_iy_j &= p_{ij} y_jy_i & & (\text{all } i,j) \\
x_ix_j &= \gamma_i p_{ij} x_jx_i & & (i < j) \\
x_iy_j &= p_{ji} y_jx_i & & (i < j) \\
x_iy_j &= \gamma_j p_{ji} y_jx_i & & (i > j) \\
x_jy_j &= 1 + \gamma_j y_jx_j + \sum_{l<j}(\gamma_l-1)y_lx_l & &
(\text{all } j).
\end{align*}
In the case of $n=1$, the parameter $\bp$ plays no role
and so we refer to the single parameter simply as $p$ and write $A_1^p(\kk)$.
By \cite[Section 6]{G}, $A_1^p(\kk) \iso A_1^q(\kk)$ if and only if $p=q^{\pm 1}$.
Goodearl and Hartwig solved the isomorphism problem for
quantized Weyl algebras when no $\gamma_i$
is a root of unity \cite[Theorem 5.1]{GH}.
They proved that if $A_n^{\bp,\gamma}(\kk) \iso A_m^{\bq,\mu}(\kk)$,
then $n=m$ and there exists a sign vector $\ep \in \{\pm 1\}^n$ such that
\begin{align}
\label{eq.hwa1}
\mu_i = \gamma_i^{\ep_i}
\end{align}
and $\bp,\bq$ satisfy
\begin{align}
\label{eq.hwa2}
q_{ij} = \begin{cases}
p_{ij} & \text{if } (\ep_i,\ep_j) = (1,1), \\
p_{ji} & \text{if }(\ep_i,\ep_j) = (-1,1), \\
\gamma_i\inv p_{ji} & \text{if } (\ep_i,\ep_j) = (1,-1), \\
\gamma_i p_{ij} & \text{if } (\ep_i,\ep_j) = (-1,-1).
\end{cases}
\end{align}
The isomorphisms they give hold regardless of
the root-of-unity condition.
Levitt and Yakimov have extended this result to
the case where all parameters are roots of unity
and $A_n^{\bp,\gamma}(\kk), A_m^{\bq,\mu}(\kk)$ are free over their centers
by utilizing noncommutative discriminants \cite[Corollary 6.4]{LY}.
Several intermediate cases are still open.
The quantized Weyl algebras are not graded and so we consider their homogenizations.
The {\sf homogenized (multiparamenter) quantized Weyl algebra}
$H_n^{\bp,\gamma}$ has algebra basis $\{z,x_i,y_i\}$, $1 \leq i \leq n$,
where $z$ commutes with the $x_i$ and $y_i$.
The relations in $H_n^{\bp,\gamma}$ are the same as those for $A_n^{\bp,\gamma}(\kk)$
except the final relation type is replaced by its homogenization,
\[ x_jy_j = z^2 + \gamma_j y_jx_j + \sum_{l<j}(\gamma_l-1)y_lx_l.\]
The isomorphisms defined by Goodearl and Hartwig extend to isomorphisms in
the homogenized case by fixing the homogenizing variable.
The next theorem suggests that the Goodearl-Hartwig
result should hold in general.
\begin{thm}[Theorem \ref{thm.hwaiso}]
Let $\gamma,\mu \in (\kk^\times)^n$ and $\bp,\bq \in \cA_n$
Then $\Phi:H_n^{\bp,\gamma} \rightarrow H_m^{\bq,\mu}$ is an isomorphism
if and only if $n=m$ and there exists $\ep \in \{\pm 1\}^n$ such that
$\gamma$ and $\mu$ satisfy \eqref{eq.hwa1}
while $\bp$ and $\bq$ satisfy \eqref{eq.hwa2}.
\end{thm}
Fix parameters $\lambda \in \kk^\times$ and $\bp \in \cA_n$.
The {\sf (multiparameter) quantum ($n\times n$) matrix algebra}, $\cO_{\lambda,\bp}(\cM_n(\kk))$,
is the algebra with basis $\{X_{ij}\}$, $1 \leq i,j \leq n$,
subject to the relations
\begin{align*}
X_{lm}X_{ij} = \begin{cases}
p_{li}p_{jm}X_{ij}X_{lm} + (\lambda-1)p_{li}X_{im}X_{lj} & l > i, m > j\\
\lambda p_{li}p_{jm} X_{ij}X_{lm} & l > i, m \leq j \\
p_{jm}X_{ij}X_{lm} & l = i, m > j.
\end{cases}
\end{align*}
By \cite[Theorem 2]{AST}, $\gk(\cO_{\lambda,\bp}(\cM_n(\kk))) = n^2$ if and only if $\lambda \neq -1$.
We show in Theorem \ref{thm.pma} that the ideal of $\cO_{\lambda,\bq}(\cM_n(\kk))$
generated by degree one normal elements is $\langle X_{1n},X_{n1}\rangle$
when $\lambda \neq 1$.
On the other hand, every degree one generator of $\cO_{\bq}(\kk^m)$ is normal.
This, combined with the fact that $\gk(\cO_{\bq}(\kk^m))=m$
(\cite[Lemma II.9.7]{BG}) proves that
$\cO_{\lambda,\bp}(\cM_n(\kk)) \iso \cO_{\bq}(\kk^m)$ if and only if $m=n^2$ and $\lambda = 1$.
Hence, we ignore the cases $\lambda = \pm 1$ henceforth.
The isomorphism problem in the single parameter case of $\cO_{\lambda,\bp}(\cM_n(\kk))$ for all $n$
was solved in \cite[Proposition 3.1]{G},
as was the multiparameter case for $n=2$ \cite[Proposition 4.2]{G}.
We extend these results in the following theorem.
\begin{thm}[Theorem \ref{thm.mpqma}]
$\cO_{\lambda,\bp}(\cM_n(\kk)) \iso \cO_{\mu,\bq}(\cM_m(\kk))$ if and only if
$n=m$ and one of the following cases holds:
\begin{enumerate}
\item $\lambda=\mu$ and $\bp = \bq$;
\item $\lambda=\mu$ and $p_{ij}=\lambda\inv q_{ji}$
for all $i,j$;
\item $\lambda=\mu\inv$ and $p_{ij}=q_{n+1-i,n+1-j}$ for all $i,j$;
\item $\lambda=\mu\inv$ and $p_{ij}=\lambda\inv q_{n+1-j,n+1-i}$ for all $i,j$.
\end{enumerate}
\end{thm}
\section{Quantum affine spaces}
\label{sec.qas}
\begin{lemma}
\label{lem.taudef}
Let $\Phi:R \rightarrow S$ be a graded isomorphism
between algebras satisfying Hypothesis \ref{hyp.cga}
with homogeneous generators
$\{x_i\}$ and $\{y_i\}$, respectively, $1 \leq i \leq n$.
Then there exists a permutation $\tau \in \cS_n$ such that
$y_{\tau(i)}$ is a summand of $\Psi(x_i)$ for each $i$.
\end{lemma}
\begin{proof}
Write $\Phi(x_i)=\sum \gamma_{ij} y_j$ and let $M = (\gamma_{ij})_{ij}$.
Because $\Phi$ is an isomorphism, $\det(M) \neq 0$.
For $k$, $1 \leq k \leq n$, let $I=\{1,\hdots,k\}$.
Choose $J \subset \{1,\hdots,n\}$ such that
$|J|=k$ and such that the minor $M_{IJ} \neq 0$.
We claim there exists an injective map of sets $\tau:I \rightarrow J$.
If $k=1$ and $M_{1j}$ denotes the $(1,j)$-minor of $M$, then we have
\[ 0 \neq \det(M) = \sum_{j=1}^{n} (-1)^{j+1} \gamma_{1j} M_{1j}.\]
Hence, there exists $j$ such that $\gamma_{1j} M_{1j} \neq 0$. Set $\tau(1)=j$.
Suppose inductively that the claim holds for some $k<n$.
Since $\det(M) \neq 0$, there exists a nonzero minor
$N$ of size $(k+1) \times (k+1)$.
Let $I',J' \subset \{1,\hdots,n\}$ such that $M_{I'J'}=N$, so $|I'|=|J'|=k+1$.
As above, we can choose $\ell$ such that $\gamma_{i_1 j_\ell} N_{1 j_\ell} \neq 0$
and set $\tau(i_1)=j_\ell$.
Now apply the inductive hypothesis to
$I=\{i_2,\hdots,i_{k+1}\}$ and $J=\{j_1,\hdots,\wh{j_\ell},\hdots,j_{k+1}\}$.
The result follows by setting $N=M$.
\end{proof}
For the remainder of this section, let $\{x_i\}$, $\{y_i\}$
be bases for $\cO_{\bp}(\kk^n)$ and $\cO_{\bq}(\kk^n)$, respectively.
and suppose $\Phi:\cO_{\bp}(\kk^n) \rightarrow \cO_{\bq}(\kk^n)$ is a isomorphism.
We claim $\bp = \tau.\bq$ where $\tau \in \cS_n$
is determined by Lemma \ref{lem.taudef}.
\begin{lemma}
\label{lem.qparam1}
If $r,s \in \{1,\hdots,n\}$ such that $p_{rs} \neq 1$,
then $p_{rs}=q_{\tau(r)\tau(s)}$.
\end{lemma}
\begin{proof}
Write $\Phi(x_r)=\sum \alpha_i y_i$ and $\Phi(x_s)=\sum \beta_i y_i$. Then,
\[ 0 = \Phi(x_rx_s - p_{rs} x_sx_r)
= (1-p_{rs})\left(\sum_{d=1}^n \alpha_d \beta_d y_d^2\right)
+ \sum_{1 \leq i \neq j \leq n}
\left(\alpha_i \beta_j -p_{rs} \alpha_j\beta_i \right) y_iy_j.
\]
Since $p_{rs} \neq 1$, then $\alpha_d=0$ or $\beta_d = 0$ for each $d$. Thus,
\begin{align}
0 &= \Phi(x_rx_s - p_{rs} x_sx_r) \\
&= \sum_{1 \leq i < j \leq n}
\left[ (\alpha_i \beta_j -p_{rs} \alpha_j\beta_i)
+ q_{ji}(\alpha_j\beta_i - p_{rs}\alpha_i\beta_j)\right] y_iy_j \notag \\
&= \sum_{1 \leq i < j \leq n} \left[ (\alpha_j\beta_i(q_{ji}-p_{rs})
+ \alpha_i\beta_j (1 - q_{ji}p_{rs})\right]y_iy_j. \label{not1}
\end{align}
By Lemma \ref{lem.taudef}, $\alpha_{\tau(r)}$, $\beta_{\tau(s)} \neq 0$.
Thus, $\alpha_{\tau(s)}=0$ and $\beta_{\tau(r)}=0$.
If $\tau(r) > \tau(s)$, then by \eqref{not1}
the coefficient of $y_{\tau(s)}y_{\tau(r)}$ is
$\alpha_{\tau(r)}\beta_{\tau(s)}(q_{\tau(r)\tau(s)} - p_{rs})$.
Therefore, $p_{rs} = q_{\tau(r)\tau(s)}$.
One the other hand, if $\tau(r) < \tau(s)$,
then the coefficient of $y_{\tau(r)}y_{\tau(s)}$ is $\alpha_{\tau(r)}\beta_{\tau(s)}(1-q_{\tau(s)\tau(r)}p_{rs})$.
Therefore, $p_{rs} = q_{\tau(s)\tau(r)}\inv = q_{\tau(r)\tau(s)}$.
Because $p_{rs} \neq 1$, then $r \neq s$ and so,
because $\tau$ is a permutation, $\tau(r) \neq \tau(s)$.
\end{proof}
\begin{lemma}
\label{lem.qparam2}
If $r,s \in \{1,\hdots,n\}$ such that $p_{rs} = 1$,
then $p_{rs}=q_{\tau(r)\tau(s)}$.
\end{lemma}
\begin{proof}
Define $\bp^{\#}=\{p_{ij} \in \bp \mid p_{ij} \neq 1\}$
and $\bq^{\#}$ similarly.
By Lemma \ref{lem.qparam1}, $\bp^{\#} \leq \bq^{\#}$.
Because $\Phi$ is an isomorphism,
then we can apply Lemma \ref{lem.qparam1} to $\Phi\inv$
to get that $\bq^{\#} \leq \bp^{\#}$.
Thus, $\bp^{\#} = \bq^{\#}$.
It follows that $p_{rs}=1$ implies $q_{\tau(r)\tau(s)}=1$
so $p_{rs} = q_{\tau(r)\tau(s)}$ for all $r,s \in \{1,\hdots,n\}$.
\end{proof}
\begin{theorem}
\label{thm.qas}
$\cO_{\bp}(\kk^n) \iso \cO_{\bq}(\kk^m)$ if and only if $m=n$ and $\bp$ is a permutation of $\bq$.
\end{theorem}
\begin{proof}
Suppose $n=m$ and there exists $\sigma \in \cS_n$ such that $\bp=\sigma.\bq$.
Define a map $\Psi: \cO_{\bp}(\kk^n) \rightarrow \cO_{\bq}(\kk^n)$ by $x_i \mapsto y_{\sigma(i)}$.
For all $i,j$, $1 \leq i,j \leq n$,
\[ \Psi(x_i)\Psi(x_j) - p_{ij}\Psi(x_j)\Psi(x_i)
= y_{\sigma(i)}y_{\sigma(j)} - q_{\sigma(i)\sigma(j)}y_{\sigma(j)}y_{\sigma(i)}=0. \]
Hence, $\Psi$ extends to a bijective homomorphism.
Thus, $\cO_{\bp}(\kk^n) \iso \cO_{\bq}(\kk^n)$.
If $\cO_{\bp}(\kk^n) \iso \cO_{\bq}(\kk^m)$, then $n=\gk(\cO_{\bp}(\kk^n))=\gk(\cO_{\bq}(\kk^m))=m$, so $n=m$.
By Lemmas \ref{lem.taudef}, \ref{lem.qparam1}, and \ref{lem.qparam2}
there exists a permutation $\tau \in \cS_n$ such that $\bp = \tau.\bq$.
\end{proof}
\section{Degree one normal elements}
\label{sec.deg1}
In order to consider the isomorphism problem for
homogenized quantized Weyl algebras and quantum matrix algebras,
we identify homogeneous degree one normal elements.
Let $A$ be an algebra satisfying Hypothesis \ref{hyp.cga}.
We define the ideals the ideals
$I_0 \subset I_1 \subset \cdots \subset I_n$ of $A$
where $I_k/I_{k-1}$ is generated by all
(homogeneous) degree one normal elements in $A/I_{k-1}$.
By convention we set $I_k=0$ for $k<0$.
We frequently identify elements in $A$ with their images in $A/I_{k-1}$.
An algebra $B = A/I_k$ for some $k$ is an
{\sf iterative quotient of $A$ (by degree one normal elements)}.
It is clear that $B$ also satisfies Hypothesis \ref{hyp.cga}.
Moreover, if $\Phi:A \rightarrow A'$ is an
isomorphism of algebras satisfying Hypothesis \ref{hyp.cga},
then $A/I_k \iso A'/\Phi(I_k)$ for all $k > 0$.
Our goal is to identify degree one normal
elements in each such quotient.
\begin{theorem}
\label{thm.wanorm}
In the case of $H_n^{\bp,\gamma}$,
$I_0 = \langle z \rangle$ and for $0 < k \leq n$,
$I_k/I_{k-1} = \langle x_k,y_k \rangle$.
\end{theorem}
\begin{proof}
Suppose $u \in H_n^{\bp,\gamma}$
is a degree one normal element and write
\[u = a z + \sum_{i=1}^m (\alpha_{1i} x_i + \alpha_{2i} y_i),\]
with $a,\alpha_i,\beta_i \in \kk$ and $m \leq n$.
We may assume $\alpha_{1m}$ or $\alpha_{2m}$ is nonzero.
Both cases are similar so assume the former. Then
\[
u y_m = ay_m z
+ \alpha_{1m} \left(z^2 + \gamma_m y_m x_m
+ \sum_{l < m} (\gamma_l - 1) y_l x_l \right)
+ \sum_{i=1}^{m-1} \alpha_{1i} (p_{mi} y_m x_i)
+ \sum_{i=1}^m \alpha_{2i} y_iy_m.
\]
By normality, there exists $r \in H_n^{\bp,\gamma}$ such that $uy_m = ru$. Write
\[ r = b z + \sum_{i=1}^m (\beta_{1i} x_i + \beta_{2i} y_i).\]
Then
\begin{align*}
ru
&= \sum_{i=1}^m \left(
(b\alpha_{1i} + a\beta_{1i} )x_iz +
(b\alpha_{2i} + a\beta_{2i} )y_iz
\right) +
\sum_{i=1}^m (\alpha_{1i}\beta_{1i} x_i^2 + \alpha_{2i}\beta_{2i} y_i^2) \\
&\quad + \sum_{i=1}^{m-1} \left(
(\alpha_{1i}\beta_{1m} + \gamma_i p_{im} \alpha_{1m}\beta_{1i}) x_ix_m +
(\alpha_{2i}\beta_{2m} + p_{im} \alpha_{2m}\beta_{2i}) y_iy_m \right) \\
&\quad + \sum_{i=1}^{m-1} \left(
(\alpha_{1i}\beta_{2m} + p_{mi}\alpha_{2m}{\beta_1i}) y_mx_i +
(\alpha_{1m}\beta_{2i} + \gamma_i p_{im} \alpha_{2i}\beta_{1m}) y_ix_m
\right) \\
&\quad + \text{(additional terms in $y_ix_i$ and $z^2$)}.
\end{align*}
We now evaluate several coefficients in $0=ru-uy_m$.
The coefficient of $x_m^2$ is $\alpha_{1m}\beta_{1m}$, so $\beta_{1m}=0$
by our hypothesis on $\alpha_{1m}$.
It follows that the coefficient of $x_ix_m$ is
$\gamma_i p_{im}\alpha_{1m}\beta_{1i}$, so $\beta_{1i}=0$ for all $i=1,\hdots,m$.
The coefficient of $x_m z$ is $b\alpha_{1m}$ so $b=0$.
Finally, it follows that the coefficient of $y_ix_m$ for $i<m$
is $\alpha_{1m}\beta_{2i}$ so $\beta_{2i}=0$ for $i<m$. Thus
\[
ru = \beta_{2m} y_m \left( a z + \sum_{i=1}^m (\alpha_{1i} x_i + \alpha_{2i} y_i) \right)
= \beta_{2m} \left( ay_mz + \sum_{i=1}^m (\alpha_{1i} y_mx_i + \alpha_{2i} p_{mi} y_iy_m ) \right).
\]
But then the coefficient of $z^2$ in $ru-uy_m$ is $-\alpha_{1m}$, a contradiction.
Since $z \in I_0$ it follows that $I_0=(z)$.
The general statement is similar.
Let $B=H_n^{\bp,\gamma}/I_{k-1}$ be an iterative quotient. Write
\[ v = \sum_{i=k}^m (\alpha_{1i} x_i + \alpha_{2i} y_i) + I_{k-1}.\]
It is clear that $x_k,y_k \in I_k$ and an analysis as
above shows that $v = \alpha_{1k} x_k + \alpha_{2k}y_k$.
\end{proof}
\begin{corollary}
\label{cor.hwprops}
$H_n^{\bp,\gamma}$ is Artin-Schelter regular of global and GK dimension $2n+1$.
\end{corollary}
\begin{proof}
Observe that $H_n^{\bp,\gamma}/I_{n-1} \iso \cO_q(\kk^2)$, the quantum plane with $q=\gamma_1$,
and $\cO_q(\kk^2)$ is Artin-Schelter regular of global and GK dimension $2$ \cite{AS}.
By Theorem \ref{thm.wanorm}, $(z,x_1,y_1,\hdots,x_{n-1},y_{n-1})$
is a regular normalizing sequence.
Hence, $H_n^{\bp,\gamma}$ has the required properties by \cite[Lemma 5.7 and Corollary 5.10]{L}.
\end{proof}
\begin{remark}
\label{rmk.pma}
Observe from the defining relations of $\cO_{\lambda,\bp}(\cM_n(\kk))$
that for every pair of generators $X_{im},X_{lj}$
with $l > i$ and $m > j$ there exists a unique pair
$X_{ij},X_{lm} \notin \{X_{im},X_{lj}\}$
such that $X_{im}X_{lj}$ is a linear
combination of $X_{ij}X_{lm}$ and $X_{lm}X_{ij}$.
Moreover, if $l=i$ or $m=j$, then there is no such pair.
\end{remark}
\begin{theorem}
\label{thm.pma}
In the case of $\cO_{\lambda,\bp}(\cM_n(\kk))$,
\[I_k/I_{k-1} = \langle
X_{1(n-k)},X_{(n-k)1},X_{2(n-k+1)},X_{(n-k+1)2},\hdots,X_{(k+1)n},X_{n(k+1)}
\rangle,
\quad
1 \leq k < n.\]
\end{theorem}
\begin{proof}
Let $B=\cO_{\lambda,\bp}(\cM_n(\kk))/I_{k-1}$ and recall that in case $k=0$ we have $I_{k-1}=0$.
It is clear from the defining relations that the given generators
are degree one normal elements in $B$.
Let $u$ be a degree one normal element of $B$.
Suppose by way of contradiction that $u$
has a summand $X_{ij}$ such that
\[ X_{ij} \notin \{ X_{1(n-k)},X_{(n-k)1},X_{2(n-k+1)},X_{(n-k+1)2},\hdots,X_{(k+1)n},X_{n(k+1)} \}.\]
Hence, there exists $X_{lm}$ such that
$l>i$ and $m>j$ or $l<i$ and $m<j$.
We consider the first case though the second is similar.
For simplicity, write
$x_1=X_{ij}$, $x_2=X_{im}$,
$x_3=X_{lj}$, $x_4=X_{lm}$, and
\[u = ax_1 + bx_2 + cx_3 + dx_4 + \bx\]
where $a,b,c,d \in \kk$ and $\bx$ is a degree one element
such that $x_1,\hdots,x_4$ are not summands.
By hypothesis, $a \neq 0$. Then
\[ x_4 u
= a(p_{li}p_{jm} x_1x_4 + (\lambda-1)p_{li}x_2x_3)
+ b \lambda p_{li} x_2x_4 + cp_{li}x_3x_4 + dx_4^2 + x_4 \bx.\]
Clearly $x_1x_4,x_2x_4,x_3x_4,x_4^2$ are not summands of $x_4\bx$.
Since $x_1,x_4$ is the unique pair such that $x_2x_3$ is a linear
combination of $x_1x_4$ and $x_4x_1$, and $x_1$ is not a summand of $\bx$,
then by Remark \ref{rmk.pma}, $x_2x_3$ is also not a summand of $x_4\bx$.
Because $u$ is normal, there exists $r \in B$ such that $x_4u=ur$. Write
\[ r = a'x_1 + b'x_2 + c'x_3 + d'x_4 + \by\]
where $a',b',c',d' \in \kk$ and $\by$ is a degree one element
such that $x_1,\hdots,x_4$ are not summands.
On the other hand,
\begin{align*}
ur &= (ax_1 + bx_2 + cx_3 + dx_4 + \bx)(a'x_1 + b'x_2 + c'x_3 + d'x_4 + \by) \\
&= (aa' x_1^2 + bb' x_2^2 + cc' x_3^2 + dd' x_4^2)
+ (ab' + p_{jm}a'b)x_1x_2
+ (ac'+\lambda p_{li} a'c) x_1x_3
+ (ad' + p_{li}p_{jm}a'd) x_1x_4 \\
&\qquad+ (bd' + \lambda p_{li}b'd) x_2x_4
+ (cd' + p_{jm}c'd)x_3x_4
+ (bc' + \lambda p_{li}p_{mj} b'c + (a'd)(\lambda-1))x_2x_3
+ \bx r + u\by.
\end{align*}
It is clear that $x_1^2$ is not a summand of $x_4\bx$, $\bx r$, or $u\by$.
Hence, because $x_4u=ur$, we must have $0 = aa'$.
But $a\neq 0$ by hypothesis so $a'=0$.
The expression for $ur$ now reduces to
\begin{align*}
ur &= (bb' x_2^2 + cc' x_3^2 + dd' x_4^2)
+ ab'x_1x_2
+ ac'x_1x_3
+ ad'x_1x_4 \\
&\qquad+ (bd' + \lambda p_{li}b'd) x_2x_4
+ (cd' + p_{jm}c'd)x_3x_4
+ (bc' + \lambda p_{li}p_{mj} b'c)x_2x_3
+ \bx r + u\by.
\end{align*}
Similarly, because $x_2=X_{im}$ and $x_1=X_{ij}$,
then by Remark \ref{rmk.pma},
$x_1x_2$ is not a summand of $x_4\bx$, $\bx r$, or $u\by$.
Thus, $ab'=0$ so $b'=0$.
The same logic applies to $x_3$ and $x_1$ so $c'=0$. Hence,
\begin{align*}
ur &= dd' x_4^2 + ad'x_1x_4 + bd' x_2x_4 + cd' x_3x_4
+ \bx r + u\by.
\end{align*}
Finally, we can apply similar reasoning to
conclude that $x_2x_3$ is not a summand of $ur$.
Hence, the coefficient of $x_2x_3$ in $x_4u$ must be zero.
It follows that $a(\lambda-1)p_{li}=0$, so $\lambda = 1$,
a contradiction.
\end{proof}
\section{Homogenized quantized Weyl algebras}
The following was proved in \cite[Proposition 5.4.5]{gad-thesis}.
However, in light of Theorem \ref{thm.wanorm},
we give a much simpler proof that also
outlines the strategy for the general case.
\begin{proposition}
$H(\wap) \iso H(\wa)$ if and only if $p=q^{\pm 1}$.
\end{proposition}
\begin{proof}
Let $\{X,Y,Z\}$ (resp. $\{x,y,z\}$) be the standard basis of $H(\wap)$ (resp. $H(\wa)$).
If $p=q$, then there is nothing to prove.
If $p=q\inv$, then the map given by
$X \mapsto -q y$, $Y \mapsto x$, and $Z \mapsto z$
clearly extends to a bijective homomorphism.
Conversely, suppose $\Phi:H(\wap) \rightarrow H(\wa)$ is a graded isomorphism.
Since $\Phi((Z)) = (z)$ by Theorem \ref{thm.wanorm},
then there is an induced isomorphism
\[\cO_p(\kk^2) \iso H(\wap)/(Z) \iso H(\wa)/(z) \iso \cO_q(\kk^2).\]
By Theorem \ref{thm.qas}, $p=q^{\pm 1}$.
\end{proof}
We now move to the general case.
\begin{lemma}
\label{lem.hwa1}
Suppose $\Phi:H_n^{\bp,\gamma} \rightarrow H_n^{\bq,\mu}$ is an isomorphism.
For all $k$, $1 \leq k \leq n$,
$\mu_k = \gamma_k^{\pm 1}$.
That is, $\gamma$ and $\mu$ satisfy \eqref{eq.hwa1}.
\end{lemma}
\begin{proof}
Fix $k$, $1 \leq k \leq n$.
Suppose first that $\gamma_k \neq 1$.
By Theorem \ref{thm.wanorm},
$I_k = \langle X_k,Y_k \rangle$ is mapped bijectively to
$J_k = \Phi(I_k) = \langle x_k,y_k \rangle$.
Thus, there exists $a,b,c,d \in \kk$ with $ab-cd \neq 0$ such that
\[
\Phi(X_k) = ax_k + by_k + J_{k-1} \text{ and }
\Phi(Y_k) = cx_k + dy_k + J_{k-1}.
\]
Then
\begin{align*}
0
&= \Phi(X_kY_k - \gamma_k Y_kX_k + I_{k-1}) \\
&= (ax_k + by_k)(cx_k + dy_k)
- \gamma_k (cx_k + dy_k)(ax_k + by_k) + J_{k-1} \\
&= (1-\gamma_k)(ac x_k^2 + bd y_k^2)
+ (ad - \gamma_k bc)(x_ky_k) + (bc-\gamma_k ad) y_kx_k + J_{k-1} \\
&= (1-\gamma_k)(ac x_k^2 + bd y_k^2)
+ (ad - \gamma_k bc)(\mu_k y_kx_k) + (bc-\gamma_k ad) y_kx_k + J_{k-1} \\
&= (1-\gamma_k)(ac x_k^2 + bd y_k^2)
+ \left[ ad(\mu_k - \gamma_k) + bc(1-\mu_k\gamma_k) \right] y_kx_k + J_{k-1}.
\end{align*}
Because $\gamma_k \neq 1$, then the coefficients of $x_k^2$ and $y_k^2$
must be zero on the above. Hence, $ac=0$ and $bd=0$.
But $ad-bc \neq 0$ and so we have two cases.
In the first case, $b=c=0$ and $a,d \neq 0$ so $\mu_k = \gamma_k$.
In the second case, $a=d=0$ and $b,c \neq 0$ so $\mu_k = \gamma_k\inv$.
If $\mu_k \neq 1$, then an argument as above
shows that $\gamma_k = \mu_k^{\pm 1}$.
Thus, $\gamma_k=1$ if and only if $\mu_k = 1$.
\end{proof}
\begin{lemma}
\label{lem.hwa2}
Suppose $\Phi:H_n^{\bp,\gamma} \rightarrow H_n^{\bq,\mu}$ is an isomorphism.
Then $\bp$ and $\bq$ satisfy \eqref{eq.hwa2}.
\end{lemma}
\begin{proof}
Let $V_k = \Span_\kk \{z,x_\ell,y_\ell \mid \ell < k\}$.
Because $\Phi$ is assumed to be graded,
then it follows from Theorem \ref{thm.wanorm} that for $i < j$,
\[
\Phi(Y_i) = ax_i + by_i + K
\quad\text{and}\quad
\Phi(Y_j) = cx_j + dy_j + L\]
for some $K \in V_i$ and $L \in V_j$.
We have,
\begin{align*}
0 &= \Phi(Y_iY_j-p_{ij}Y_jY_i) \\
&= (ax_i + by_i + K)(cx_j + dy_j + L) - p_{ij}(cx_j + dy_j + L)(ax_i + by_i + K) \\
&= ac(x_ix_j-p_{ij}x_jx_i) + ad(x_iy_j-p_{ij}y_jx_i)
+ bc(y_ix_j-p_{ij}x_jy_i) + bd(y_iy_j-p_{ij}y_jy_i) \\
&\hspace{5em} + (ax_i + by_i + K)L-p_{ij}L(ax_i + by_i + K) \\
&= ac(\mu_iq_{ij}-p_{ij})x_jx_i + ad(q_{ji}-p_{ij})y_jx_i
+ bc(1-p_{ij}\mu_iq_{ij})x_jy_i + bd(q_{ij}-p_{ij})y_jy_i \\
&\hspace{5em} + (ax_i + by_i + K)L-p_{ij}L(ax_i + by_i + K).
\end{align*}
The defining relations clearly imply that the monomials $x_jx_i$,
$y_jy_i$, $x_jy_i$, and $y_jy_i$ do not appear as summands of
$(ax_i + by_i + K)L-p_{ij}L(ax_i + by_i + K)$.
Hence, the coefficients of those monomials in the above must
each be zero.
Suppose $\gamma_i,\gamma_j \neq 1$.
By the proof of Lemma \ref{lem.hwa1},
exactly one of $ac,ad,bc,bd$ is nonzero,
corresponding to the signs of the exponents in
$\gamma_i = \mu_i^{\pm 1}$ and $\gamma_j = \mu_j^{\pm 1}$.
If $\gamma_i=\mu_i$ and $\gamma_j=\mu_j$,
then $a,d \neq 0$ and so $q_{ij}=p_{ij}$.
The other cases are similar.
If $\gamma_i=1$ or $\gamma_j=1$,
then the argument is similar.
However, we get additional restrictions on the parameters.
For example, if $\gamma_i=1$ and $\gamma_j \neq 1$,
then we may have $ad,bd \neq 0$ implying $q_{ij}=p_{ij}=q_{ji}$,
so $q_{ij}=\pm 1$.
\end{proof}
\begin{theorem}
\label{thm.hwaiso}
Let $\gamma,\mu \in (\kk^\times)^n$ and $\bp,\bq \in \cA_n$
Then $\Phi:H_n^{\bp,\gamma} \rightarrow H_m^{\bq,\mu}$ is an isomorphism
if and only if $n=m$ and there exists $\ep \in \{\pm 1\}^n$ such that
$\gamma$ and $\mu$ satisfy \eqref{eq.hwa1}
while $\bp$ and $\bq$ satisfy \eqref{eq.hwa2}.
\end{theorem}
\begin{proof}
Suppose $H_n^{\bp,\gamma} \iso H_m^{\bq,\mu}$.
By Corollary \ref{cor.hwprops},
$2n+1=\gk(H_n^{\bp,\gamma}) = \gk(H_m^{\bq,\mu})=2m+1$, so $n=m$.
By Lemma \ref{lem.hwa1}, $\gamma$ and $\mu$ satisfy \eqref{eq.hwa1}
and by Lemma \ref{lem.hwa2}, $\bp$ and $\bq$ satisfy \eqref{eq.hwa2}.
Conversely, suppose there exists $\ep \in \{\pm 1\}^n$ such that
$\lambda$ and $\mu$ satisfy \eqref{eq.hwa1}
while $\bp$ and $\bq$ satisfy \eqref{eq.hwa2}.
The isomorphisms provided in \cite{GH}
are affine and hence extend to bijective
homomorphisms $H_n^{\bp,\gamma} \rightarrow H_n^{\bq,\mu}$.
\end{proof}
\section{Quantum matrix algebras}
Throughout, let $\{X_{ij}\}$ and $\{Y_{ij}\}$ be
the standard generators for
$\cO_{\lambda,\bp}(\cM_n(\kk))$ and $\cO_{\mu,\bq}(\cM_n(\kk))$, respectively.
As in Section \ref{sec.deg1},
we let $I_k/I_{k-1}$ be the ideal
in $\cO_{\lambda,\bp}(\cM_n(\kk))/I_{k-1}$ generated by
the degree one normal elements.
\begin{proposition}
\label{prop.trans}
For $i>j$, let $p_{ij}=\lambda\inv q_{ji}$.
The map $\Phi:\cO_{\lambda,\bp}(\cM_n(\kk)) \rightarrow \cO_{\lambda,\bq}(\cM_n(\kk))$ defined by
$\Phi(X_{ij})=Y_{ji}$ extends to an isomorphism.
\end{proposition}
\begin{proof}
We claim the map $\Phi$ satisfies the defining relations for $\cO_{\lambda,\bp}(\cM_n(\kk))$,
whence $\Phi$ extends to a homomorphism.
As $\Phi$ maps onto the generators of $\cO_{\lambda,\bq}(\cM_n(\kk))$,
it is clearly surjective.
Injectivity now follows because $\cO_{\lambda,\bp}(\cM_n(\kk))$ and $\cO_{\mu,\bq}(\cM_n(\kk))$
are domains and $\gk(\cO_{\lambda,\bp}(\cM_n(\kk))) = n^2 = \gk(\cO_{\mu,\bq}(\cM_n(\kk)))$.
Case 1 ($l=i$, $m>j$)
\[
\Phi(X_{im})\Phi(X_{ij}) - p_{jm}\Phi(X_{ij})\Phi(X_{im})
= Y_{mi}Y_{ji} - \lambda q_{mj} Y_{ji}Y_{mi}
= (\lambda q_{mj}q_{ii} Y_{ji}Y_{mi}) - \lambda q_{mj} Y_{ji}Y_{mi}
= 0.
\]
Case 2 ($l>i$, $m < j$)
\begin{align*}
\Phi(X_{lm})\Phi(X_{ij}) - \lambda p_{li} p_{jm}\Phi(X_{ij})\Phi(X_{lm})
&= Y_{ml}Y_{ji} - \lambda (\lambda\inv q_{il})(\lambda\inv q_{mj}) Y_{ji}Y_{ml} \\
&= Y_{ml}Y_{ji} - \lambda\inv q_{il} q_{mj} (\lambda q_{jm} q_{li}
Y_{ml}Y_{ji}) = 0.
\end{align*}
Case 3 ($l>i$, $m=j$)
\begin{align*}
\Phi(X_{lj})\Phi(X_{ij}) - \lambda p_{li} \Phi(X_{ij})\Phi(X_{lm})
&= Y_{jl}Y_{ji} - \lambda (\lambda\inv q_{il}) Y_{ji}Y_{jl}
= (q_{il} Y_{ji}Y_{jl}) - q_{il} Y_{ji}Y_{jl}
= 0.
\end{align*}
Case 4 ($l>i$, $m>j$)
\begin{align*}
\Phi(X_{lm})\Phi(X_{ij}) &- p_{li} p_{jm}\Phi(X_{ij})\Phi(X_{lm})
+ (\lambda-1)p_{li}X_{im}X_{lj}) \\
&= Y_{ml}Y_{ji} - (\lambda\inv q_{il})(\lambda q_{mj}) Y_{ji}Y_{ml}
+ (\lambda-1)(\lambda\inv q_{il}) Y_{mi}Y_{jl} \\
&= (q_{mj}q_{il} Y_{ji}Y_{ml} + (\lambda-1)q_{mj}Y_{jl}Y_{mi})
- q_{il}q_{mj} Y_{ji}Y_{ml}
+ (\lambda-1)(\lambda\inv q_{il}) (\lambda q_{mj}q_{li}) Y_{jl}Y_{mi} = 0.
\qedhere
\end{align*}
\end{proof}
\begin{proposition}
\label{prop.flip}
Let $p_{ij}=q_{n+1-i,n+1-j}$ and $\lambda=\mu\inv$.
The map $\Phi:\cO_{\lambda,\bp}(\cM_n(\kk)) \rightarrow \cO_{\mu,\bq}(\cM_n(\kk))$ defined by
$\Phi(X_{ij})=Y_{n+1-i,n+1-j}$ extends to an isomorphism.
\end{proposition}
\begin{proof}
We claim the map $\Phi$ satisfies the defining relations for $\cO_{\lambda,\bp}(\cM_n(\kk))$,
whence $\Phi$ extends to a bijective homomorphism by an analogous
argument as in Proposition \ref{prop.trans}.
Set $r=n+1-l$, $s=n+1-m$, $u=n+1-i$, and $v=n+1-j$.
Case 1 ($l=i, m>j$) We have $r=u$ and $s>v$. Thus,
\[ \Phi(X_{lm})\Phi(X_{ij}) - p_{jm}\Phi(X_{ij})\Phi(X_{lm})
= Y_{rs}Y_{uv} - p_{jm}Y_{uv}Y_{rs}
= (q_{vs}-p_{jm})Y_{uv}Y_{rs} = 0.
\]
Case 2 ($l>i$, $m \leq j$) We have $r<u$ and $v\leq s$. Thus,
\[ \Phi(X_{lm})\Phi(X_{ij}) - \lambda p_{li}p_{jm} \Phi(X_{ij})\Phi(X_{lm})
= Y_{rs}Y_{uv} - \lambda p_{li}p_{jm} Y_{uv}Y_{rs}
= (1-\lambda p_{li}p_{jm}\mu q_{ur}q_{sv})Y_{rs}Y_{uv} = 0.
\]
Case 3 ($l>i$, $m>j$) We have $r<u$ and $s < v$. Thus,
\begin{align*}
\Phi(X_{lm})\Phi(X_{ij}) &- p_{li}p_{jm} \Phi(X_{ij})\Phi(X_{lm}) - (\lambda-1)p_{li}\Phi(X_{im})\Phi(X_{lj}) \\
&= Y_{rs}Y_{uv} - p_{li}p_{jm} Y_{uv}Y_{rs} - (\lambda-1)p_{li}Y_{us}Y_{rv} \\
&= Y_{rs}Y_{uv} - p_{li}p_{jm} (q_{ur}q_{sv}Y_{rs}Y_{uv} +(\mu-1)q_{ur}Y_{rv}Y_{us}) - (\lambda-1)p_{li}(\mu q_{ur}q_{vs})Y_{rv}Y_{us} \\
&= -p_{li}q_{ur} ((\mu-1)p_{jm} + (\lambda-1)\mu q_{vs}) =0. \qedhere
\end{align*}
\end{proof}
The key question is whether there are any more types of isomorphisms.
\begin{lemma}
\label{lem.sum}
Suppose $\Phi:\cO_{\lambda,\bp}(\cM_n(\kk)) \rightarrow \cO_{\mu,\bq}(\cM_n(\kk))$ is a graded isomorphism.
(1) If $p_{jm} \neq 1$,
then $\Phi(X_{ij})$ and $\Phi(X_{im})$
do not share any summands.
(2) If $\lambda p_{li} \neq 1$,
then $\Phi(X_{ij})$ and $\Phi(X_{lj})$
do not share any summands.
(3) For all $i,j$, $\Phi(X_{ij})$ and $\Phi(X_{ji})$ do not share
any summands.
\end{lemma}
\begin{proof}
(1) WLOG, assume $m > j$.
Suppose $Y_{uv}$ is a summand of
both $\Phi(X_{ij})$ and $\Phi(X_{im})$
with coefficient $a$ and $b$, respectively.
The coefficient of $Y_{uv}^2$ in
$0=\Phi(X_{im}X_{ij}-p_{jm}X_{ij}X_{lm})$ is $ab(1-p_{jm})$.
Hence, $a=0$ or $b=0$, a contradiction.
(2) is similar.
(3) WLOG, assume $i>j$.
Suppose $Y_{uv}$ is a summand of
both $\Phi(X_{ij})$ and $\Phi(X_{ji})$
with coefficient $a$ and $b$, respectively.
The coefficient of $Y_{uv}^2$ in
$0=\Phi(X_{ii}X_{jj}-X_{jj}X_{ii}-(\lambda-1)p_{ij}X_{ji}X_{ij})$
is $-(\lambda-1)p_{ij}ab$.
Hence, $a=0$ or $b=0$, a contradiction.
\end{proof}
\begin{lemma}
\label{lem.lammu}
$\cO_{\lambda,\bp}(\cM_n(\kk)) \iso \cO_{\mu,\bq}(\cM_n(\kk))$ implies $\lambda=\mu^{\pm 1}$.
\end{lemma}
\begin{proof}
By Theorem \ref{thm.pma}, the degree one normal elements of $\cO_{\lambda,\bp}(\cM_n(\kk))$
(resp. $\cO_{\mu,\bq}(\cM_n(\kk))$) are $X_{1n}$ and $X_{n1}$ (resp. $Y_{1n}$ and $Y_{n1}$).
Consequently, $\Phi(X_{1n})$ and $\Phi(X_{n1})$
are linear combinations of $Y_{1n},Y_{n1}$. Write
\begin{align*}
\Phi(X_{11}) &= a_1 Y_{11} + b_1 Y_{nn} + c_1 Y_{1n} + d_1 Y_{n1} + \bx,
& & \Phi(X_{1n}) = c_3 Y_{1n} + d_3 Y_{n1}, \\
\Phi(X_{nn}) &= a_2 Y_{11} + b_2 Y_{nn} + c_2 Y_{1n} + d_2 Y_{n1} + \by ,
& & \Phi(X_{n1}) = c_4 Y_{1n} + d_4 Y_{n1}.
\end{align*}
where $a_i,b_i,c_i,d_i \in \kk$
and $\bx,\by$ are linear terms not containing
$Y_{11},Y_{nn},Y_{1n},Y_{n1}$ as summands.
By Proposition \ref{prop.flip} and Lemma \ref{lem.sum} (3),
we can reduce to the case $d_3=c_4=0$. Then
\begin{align*}
0 &= \Phi(X_{1n}X_{11} - p_{1n}X_{11}X_{1n}) \\
&= c_3 \left[ a_1(q_{1n}-p_{1n}) Y_{11}Y_{1n}
+ b_1(1- \mu p_{1n}q_{n1}) Y_{1n}Y_{nn} \right. \\
&\quad + \left. c_1(1-p_{1n}) Y_{1n}^2
+ d_1(\mu q_{1n}^2 - p_{1n}) Y_{1n}Y_{n1} \right] \\
&\quad + \text{(additional terms not in $Y_{11}Y_{1n}, Y_{1n} Y_{nn},
Y_{1n}^2, Y_{1n}Y_{n1}$)}, \\
0 &= \Phi(X_{n1}X_{11} - \lambda p_{n1}X_{11}X_{n1}) \\
&= d_4 \left[ a_1 (\mu q_{n1}-\lambda p_{n1}) Y_{11}Y_{n1}
+ b_1(1-\lambda p_{n1}q_{1n}) Y_{n1}Y_{nn} \right. \\
&\quad + \left. c_1 (1 - \lambda \mu p_{n1}q_{n1}^2) Y_{1n}Y_{n1}
+ d_1(1-\lambda p_{n1})Y_{n1}^2 \right] \\
&\quad + \text{(additional terms not in $Y_{11}Y_{n1}, Y_{11} Y_{1n},
Y_{1n}Y_{n1}, Y_{n1}^2$)}.
\end{align*}
If $a_1\neq 0$, then $q_{1n}=p_{1n}$ and $\mu q_{n1}=\lambda p_{n1}$,
so $\mu=\lambda$.
If $b_1 \neq 0$, then $1=\mu p_{1n}q_{n1}$ and $1=\lambda p_{n1}q_{1n}$,
so $\mu p_{1n}q_{n1} = \lambda p_{n1}q_{1n}$ implies $\lambda=\mu\inv$.
A similar argument using $X_{nn}$ in place of $X_{11}$ shows that if
$a_2\neq 0$ or $b_2\neq 0$, then $\lambda = \mu^{\pm 1}$.
Moreover, one can show that if $X_{11}$ or $X_{nn}$
are summands of either $\Phi\inv(Y_{11})$ or $\Phi\inv(Y_{nn})$
then $\lambda = \mu^{\pm 1}$.
We now reduce to the case that $a_1,a_2,b_1,b_2 = 0$. In
$0=\Phi(X_{11}X_{nn}-X_{nn}X_{11} + (\lambda-1)p_{n1} X_{1n}X_{n1})$
the coefficient of $Y_{1n}Y_{n1}$ is
\[ (1-\mu q_{n1}^2)(c_1d_2-c_2d_1) + (\lambda-1)p_{n1} c_3d_4. \]
If $p_{n1}, \lambda p_{n1} \neq 1$, then
$c_1,d_1,c_2,d_2=0$ by Lemma \ref{lem.sum} (1,2)
and this coefficient reduces to $(\lambda-1)p_{n1} c_3d_4 \neq 0$,
a contradiction. Thus, $p_{n1} = 1$ or $\lambda p_{n1}=1$.
We may further assume that $q_{n1}=1$ or $\lambda q_{n1}=1$.
Thus, we have four cases to consider.
Note that
\begin{align*}
0 = \Phi(X_{n1}X_{1n} - \lambda p_{n1}^2 X_{1n}X_{n1})
= c_3d_4 (\mu q_{n1}^2 - \lambda p_{n1}^2) Y_{1n}Y_{n1}.
\end{align*}
Thus, $\lambda p_{n1}^2=\mu q_{n1}^2$.
If $p_{n1}=q_{n1}=1$, then $\lambda=\mu$.
If $\lambda p_{n1}=q_{n1}=1$, then $p_{n1}=\mu$ so $\lambda = \mu\inv$.
The other two cases follow similarly.
\end{proof}
Suppose $\cO_{\lambda,\bp}(\cM_n(\kk)) \iso \cO_{\mu,\bq}(\cM_n(\kk))$.
By Lemma \ref{lem.lammu}, $\lambda=\mu^{\pm 1}$.
If $\lambda=\mu\inv$, then by Proposition \ref{prop.flip}
we can replace $\cO_{\mu,\bq}(\cM_n(\kk))$ by $\cO_{\mu\inv,\bq'}(\cM_n(\kk))$.
Thus, it suffices to consider the isomorphism problem
for the case where $\lambda=\mu$.
\subsection{The $n=2$ case}
\label{sec.n2}
As in the homogenized quantized Weyl algebra case,
we consider the $n=2$ case to illustrate methods
in the general case.
We will prove an extension of \cite[Proposition 4.2]{G},
removing any restriction on the parameters.
Let $\bp,\bq \in \cA_2$ and set $p=p_{12}$ and $q=q_{12}$. We denote
$\cO_{\lambda,\bp}(\cM_2(\kk))$ by $M_{\lambda,p}$ and
$\cO_{\mu,\bq}(\cM_2(\kk))$ by $M_{\mu,q}$.
Recall that, by assumption, $\lambda^2,\mu^2 \neq 1$.
For simplicity, we write the basis $\{X_{ij}\}$ as
$x_1 = X_{11}$, $x_2 = X_{12}$, $x_3 = X_{21}$, and $x_4 = X_{22}$.
The defining relations for $M_{\lambda,p}$ are then
\begin{align*}
&x_2x_1 = p x_1x_2
& &x_4x_3 = p x_3x_4
& &x_3x_2 = \lambda p^2 x_2 x_3 \\
&x_3x_1 = \lambda p x_1x_3
& & x_4x_2 = \lambda p x_2x_4
& & x_4x_1 = x_1x_4 + (\lambda p - p) x_2x_3.
\end{align*}
We rewrite the basis and defining relations for $M_{\mu,q}$ similarly.
\begin{lemma}
\label{lem.mlp}
If $M_{\lambda,p} \iso M_{\lambda,q}$,
then $q=p$ or $q=\lambda\inv p\inv$.
\end{lemma}
\begin{proof}
Let $\Phi: M_{\lambda,p} \rightarrow M_{\lambda,q}$
be an isomorphism.
Write $\Phi(x_i) = \sum_{k=1}^4 a_{ik} y_k$.
By Theorem \ref{thm.pma},
any normal element in $M_{\lambda,p}$ has the form
$bx_2 + cx_3$ and similarly for $M_{\lambda,q}$. Hence,
\begin{align*}
\Phi(x_2) = a_{22} y_2 + a_{23} y_3 \text{ and }
\Phi(x_3) = a_{32} y_2 + a_{33} y_3,
\end{align*}
with $a_{22}a_{33}-a_{23}a_{32} \neq 0$.
By Lemma \ref{lem.sum}, $\Phi(x_2)$ and $\Phi(x_3)$
may not share any summands.
Thus, we conclude that there are two cases,
$a_{23}=a_{32}=0$, or $a_{22}=a_{33}=0$.
By Theorem \ref{thm.pma},
$\Phi$ restricts to a graded isomorphism
\[ \kk[x_1,x_4] \iso M_{\lambda,p}/\langle x_2,x_3 \rangle
\iso M_{\lambda,q}/\langle y_2,y_3 \rangle \iso \kk[y_1,y_4].\]
Hence, $a_{11}a_{44}-a_{14}a_{41} \neq 0$. Write,
$\Phi(x_1) = a_{11} y_1 + a_{14} y_4 + K$ with
$K \in \Span_{\kk}\{y_2,y_3\}$. Then
\begin{align*}
\Phi(x_2x_1-px_1x_2)
&= (a_{22} y_2 + a_{23} y_3)(a_{11} y_1 + a_{14} y_4 + K)
- p(a_{11} y_1 + a_{14} y_4 + K)(a_{22} y_2 + a_{23} y_3) \\
&= a_{11}a_{22}(q-p) y_1y_2 + a_{14}a_{22} (1-\lambda pq) y_2y_4
+ a_{11}a_{23}(\lambda q-p) y_1y_3 + a_{14}a_{23}(1-pq) y_3y_4 \\
&\hspace{5em} + (a_{22} y_2 + a_{23} y_3)K - K(a_{22} y_2 + a_{23} y_3).
\end{align*}
It is clear that the monomials $y_1y_2,y_1y_3,y_2y_4,y_3y_4$
do not appear as summands in $(a_{22} y_2 + a_{23} y_3)K - K(a_{22} y_2 + a_{23} y_3)$.
Hence, the coefficients of those monomials must be zero.
In the first case, $a_{22} \neq 0$. Then
$a_{11} \neq 0$ implies $q=p$ and
$a_{14} \neq 0$ implies $q=\lambda\inv p\inv$.
In the second case, $a_{23} \neq 0$. Then
$a_{11} \neq 0$ implies $q=\lambda\inv p$ and
$a_{14} \neq 0$ implies $q=p\inv$.
\end{proof}
\begin{theorem}
\label{thm.mlp}
If $M_{\lambda,p} \iso M_{\mu,q}$, then
$(\mu,q)$ is one of $(\lambda,p)$, $(\lambda\inv,p\inv)$,
$(\lambda,\lambda\inv p\inv)$, or $(\lambda\inv,\lambda p)$.
\end{theorem}
\begin{proof}
If $(\mu,q)=(\lambda,p)$, then there is nothing to prove.
If $(\mu,q)=(\lambda,\lambda\inv p\inv)$, then there is an
isomorphism by Proposition \ref{prop.trans}.
If $(\mu,q)=(\lambda\inv,p\inv)$, then there is an isomorphism
by Proposition \ref{prop.flip}.
Composing these isomorphisms gives the final case.
Conversely, suppose $\Phi:M_{\lambda,p} \rightarrow M_{\mu,q}$
is a graded isomorphism.
By Lemma \ref{lem.lammu}, $\mu = \lambda^{\pm 1}$.
By Proposition \ref{prop.flip}, we may assume $\mu=\lambda$.
Lemma \ref{lem.mlp} now completes the proof.
\end{proof}
\subsection{The general case}
Assume $n>2$. We retain our original
notation and let $\{X_{ij}\}$ and $\{Y_{ij}\}$ be
the standard homogeneous generators for
$\cO_{\lambda,\bp}(\cM_n(\kk))$ and $\cO_{\mu,\bq}(\cM_n(\kk))$, respectively.
\begin{lemma}
\label{lem.g1}
Let $\Phi:\cO_{\lambda,\bp}(\cM_n(\kk)) \rightarrow \cO_{\lambda,\bq}(\cM_n(\kk))$ be an isomorphism
that maps the ideals $(X_{1n})$ and $(X_{n1})$ to
$(Y_{1n})$ and $(Y_{n1})$, respectively.
For all $\ell < n$,
if $X_{ij} \in I_\ell/I_{\ell-1}$,
then the ideal $(X_{ij}) + I_{\ell-1}$ is mapped
to $(Y_{ij}) + J_{\ell-1}$.
\end{lemma}
\begin{proof}
Assume throughout that $j>i$.
The case with $j<i$ is similar.
Throughout we use $\Phi$ to denote the given isomorphism
as well as any induced isomorphism
$\cO_{\lambda,\bp}(\cM_n(\kk))/I_k \rightarrow \cO_{\lambda,\bq}(\cM_n(\kk))/I_k$.
The lemma is true when $\ell=0$ by assumption.
Assume it holds for all ideals $I_k$ with $k<\ell$.
Let $S$ be the set of
$X_{ij}$ in $I_\ell/I_{\ell-1}$,
$S^+= \{ X_{ij} \in S \mid i<j\}$, and
$S^-= \{ X_{ij} \in S \mid i>j\}$.
Define $T, T^+,T^-$ similarly for the $Y_{ij}$.
{\bf Claim 1:}
If $X_{lm} \in S^+$ with $X_{lm} \neq X_{ij}$,
then $Y_{ij}$ is a summand of $\Phi(X_{ij})$ and
not a summand of $\Phi(X_{lm})$.
Either $l>i$ and $m>j$ or $l<i$ and $m<j$.
Assume the former.
Note that $Y_{ij}$ and $Y_{lm}$ are the only
two generators in $T^+$ such that a linear combination
of $Y_{ij}Y_{lm}$ and $Y_{lm}Y_{ij}$ is nonzero in $(Y_{im})$.
Let $r$ be minimal such that $X_{im} \in I_r$.
Because $m > j$ then it follows that $r<\ell$.
Hence, by the inductive hypothesis,
$\Phi( (X_{im}) + I_{r-1}) = (Y_{im}) + J_{r-1}$.
Let $U=\{Y_{uv} \in J_\ell \mid (u,v) \neq (i,j),(l,m)\}$.
Then
\begin{align*}
\Phi(X_{ij}) = a_1Y_{ij} + b_1Y_{lm} + K \text{ and }
\Phi(X_{lm}) = a_2Y_{ij} + b_2Y_{lm} + L
\end{align*}
for some $K,L \in \Span_{\kk}\{U\}$
and $a_i,b_j \in \kk$ with $a_1b_2-b_1a_2 \neq 0$.
We will show that $b_1=a_2=0$.
Suppose $b_1$ and $a_2$ are nonzero.
Write $\Phi(X_{im}) = \alpha Y_{im}+ J_{r-1}$, $\alpha \in \kk^\times$. Then
\begin{align}
\label{eq.pmarel}
\notag 0 &= \Phi(X_{im}X_{ij} - p_{jm} X_{ij}X_{im} + I_{r-1}) \\
\notag &= \alpha\left( Y_{im}(a_1Y_{ij} + b_1Y_{lm} + K)
\notag - p_{jm}(a_1Y_{ij} + b_1Y_{lm} + K)Y_{im} \right) + J_{r-1} \\
\tag{$\star$} &= \alpha \left[
a_1 (q_{jm} - p_{jm}) Y_{ij} Y_{im}
+ b_1 (1-\lambda q_{li}p_{jm}) Y_{im}Y_{lm}\right]
+ \alpha(Y_{im}K-p_{jm}KY_{im}) + J_{r-1}.
\end{align}
Since $Y_{ij} Y_{im}, Y_{im}Y_{lm}$ are not summands
in $(Y_{im}K-p_{jm}KY_{im})$, then our hypothesis that
$b_1 \neq 0$ implies $p_{jm}=\lambda\inv q_{il}$.
Similarly, write $\Phi(X_{mi}) = \beta Y_{mi} + J_{r-1}$ with $\beta \in \kk^\times$. Then
\begin{align*}
0 &= \Phi(X_{lm}X_{im} - \lambda p_{li} X_{im}X_{lm}) \\
&= \alpha\left( (a_2Y_{ij} + b_2Y_{lm} + L) Y_{im}
- \lambda p_{li} Y_{im} (a_2Y_{ij} + b_2Y_{lm} + L) \right) + J_{r-1} \\
&= \alpha\left( a_2(1-\lambda p_{li}q_{jm}) Y_{ij}Y_{im}
+ b_2(\lambda q_{li}-\lambda p_{li}) Y_{im}Y_{lm} \right)
+ \beta(LY_{im} - \lambda p_{li}Y_{im}L) + J_{r-1}.
\end{align*}
Hence, $a_2 \neq 0$ implies $p_{li} = \lambda\inv q_{mj}$. Finally,
\begin{align*}
0 &= \Phi(X_{lm}X_{ij} - p_{li}p_{jm} X_{ij}X_{lm} + (X_{im}) ) \\
&= (a_2Y_{ij} + b_2Y_{lm} + L)(a_1Y_{ij} + b_1Y_{lm}+K)
- p_{li}p_{jm}(a_1Y_{ij} + b_1Y_{lm}+K)(a_2Y_{ij} + b_2Y_{lm} + L) \\
&\hspace{5em} + (Y_{im}) + J_{r-1} \\
&= (1-p_{li}p_{jm})(a_1a_2 Y_{ij}^2 + b_1b_2 Y_{lm}^2)
+ \left[ a_1b_2(q_{li}q_{jm}-p_{li}p_{jm})
+ a_2b_1(1-p_{li}p_{jm}q_{li}q_{jm}) \right] Y_{ij}Y_{lm} \\
&\hspace{5em} + (LK-p_{li}p_{jm}KL) + (Y_{im}) + J_{r-1}.
\end{align*}
By hypothesis, $a_2b_1 \neq 0$.
Suppose further that $a_1b_2 \neq 0$.
Then $a_1a_2, b_1b_2 \neq 0$ and so $p_{li}p_{jm}=1$.
But then the coefficient of $Y_{ij}Y_{lm}$
reduces to $(q_{li}q_{jm}-1)(a_1b_2-a_2b_1)$,
so $q_{li}q_{lm}=1$ and
\[ 1 = p_{li}p_{jm} = (\lambda\inv q_{mj})(\lambda\inv q_{il}) = \lambda^{-2}.\]
Thus, $\lambda^2=1$, violating our hypothesis.
Hence, $a_1b_2 = 0$ Since $a_2b_1 \neq 0$, then
\[ 1
= p_{li}p_{jm}q_{li}q_{jm}
= (\lambda\inv q_{mj})(\lambda\inv q_{il})q_{li}q_{lm} = \lambda^{-2}.\]
Thus, $\lambda^2=1$, again violating our hypothesis.
{\bf Claim 2:}
If $X_{lm} \in S^-$, then $Y_{lm}$ is not a summand of $\Phi(X_{ij})$.
Suppose $Y_{lm} \in T^-$ is a summand of $\Phi(X_{ij})$.
If $(l,m)=(j,i)$, then the claim follows by Lemma \ref{lem.sum}. By Claim 1,
$\Phi(X_{ij}) = a_1Y_{ij} + b_1Y_{lm} + K$ and
$\Phi(X_{ji}) = a_2Y_{ji} + L$
for some $K \in \Span_{\kk}\{U\bs \{S^+ \cup \{Y_{lm}\}\} \}$,
$L \in \Span_{\kk}\{U\bs S^-\}$ and $a_i,b_j \in \kk$
with $a_1,a_2 \neq 0$.
We will show that $b_1=0$.
If $l>j$ and $m>i$, then we have
\begin{align*}
0 &= \Phi(X_{ji}X_{ij} - \lambda p_{ji}^2 X_{ij}X_{ji}) \\
&= (a_2Y_{ji} + L)(a_1Y_{ij} + b_1Y_{lm} + K)
- \lambda p_{ji}^2 (a_1Y_{ij} + b_1Y_{lm} + K)(a_2Y_{ji} + L) \\
&= a_1a_2\lambda(q_{ji}^2 - \lambda p_{ji}^2) Y_{ij}Y_{ji}
+ a_2b_1\left[ (1- \lambda p_{ji}^2 q_{lj}q_{im})Y_{ji}Y_{lm}
- (\lambda-1)\lambda p_{ji}^2 q_{lj} Y_{jm}Y_{li}\right] \\
&\hspace{5em} + a_2(Y_{ji}K-\lambda p_{ji}^2 KY_{ji})
+ a_1( L(a_1Y_{ij} + b_1Y_{lm})-\lambda p_{ji}^2 (a_1Y_{ij} + b_1Y_{lm})L)
+ (KL - \lambda p_{ji}^2 LK).
\end{align*}
Observe that $Y_{ji}, Y_{lm}$ is the unique pair
such that $Y_{jm}Y_{li}$ is a linear combination
of $Y_{ji}Y_{lm}$ and $Y_{lm}Y_{ji}$.
Moreover, $Y_{ji}$ and $Y_{lm}$ are not summands of $L$ or $K$.
Hence, we have $a_2b_1=0$, so $b_1=0$ as claimed.
\end{proof}
We our now ready to prove our main theorem.
\begin{theorem}
\label{thm.mpqma}
$\cO_{\lambda,\bp}(\cM_n(\kk)) \iso \cO_{\mu,\bq}(\cM_m(\kk))$ if and only if
$n=m$ and one of the following cases holds:
\begin{enumerate}
\item $\lambda=\mu$ and $\bp = \bq$;
\item $\lambda=\mu$ and $p_{ij}=\lambda\inv q_{ji}$
for all $i,j$;
\item $\lambda=\mu\inv$ and $p_{ij}=q_{n+1-i,n+1-j}$ for all $i,j$;
\item $\lambda=\mu\inv$ and $p_{ij}=\lambda\inv q_{n+1-j,n+1-i}$ for all $i,j$.
\end{enumerate}
\end{theorem}
\begin{proof}
First, we will establish the indicated isomorphisms.
(1) is obtained as from the map $X_{ij} \mapsto Y_{ij}$.
(2) and (3) are obtained from Propositions \ref{prop.trans}
and \ref{prop.flip}, respectively, while
(4) is the composition of those.
Conversely, let $\Phi:\cO_{\lambda,\bp}(\cM_n(\kk)) \rightarrow \cO_{\mu,\bq}(\cM_m(\kk))$ be an isomorphism.
Then $n=m$ since
\[ n^2 = \gk(\cO_{\lambda,\bp}(\cM_n(\kk))) = \gk(\cO_{\mu,\bq}(\cM_m(\kk)))=m^2.\]
By Lemma \ref{lem.lammu}, we may assume $\lambda=\mu$.
By Theorem \ref{thm.pma} and Lemma \ref{lem.sum},
we may assume that $\Phi(X_{1n})$ and $\Phi(X_{n1})$ either fix
their respective positions or interchange them.
We consider the case where they are fixed and claim that $\bp=\bq$.
Choose $j,m \in \{1,\hdots,n\}$, $j < m$.
By Lemma \ref{lem.g1}, the coefficient of $Y_{1j}Y_{1m}$ in
$\Phi(X_{1m}X_{1j}-p_{jm}X_{1j}X_{1m})$ is $p_{jm}-q_{jm}$
(see \eqref{eq.pmarel}).
Hence, $p_{jm}=q_{jm}$ for all $j,m$ and so $\bp=\bq$.
\end{proof}
\noindent {\bf Acknowledgements.}
The author thanks Ken Goodearl for helpful discussions
and the referee for offering several corrections and
points of clarification.
\bibliographystyle{amsplain}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Appendix}
We now prove all theoretical results of the paper.
We need to introduce some technical denotation.
From now on $f$ denotes one from the following functions: $\sin$, $\cos$, $sign$ or a linear rectifier. We call the set of these functions $\mathcal{F}$.
For two vectors $v,w$ we denote by $v \cdot w$ their dot product.
We denote by $G_{struct}^{i}$ for $i=1,...,\frac{k}{m}$ the building blocks of the structured matrix constructed according to the $\mathcal{P}$-model that are vertically stacked to
produce the final structured matrix.
Let $v^{1},v^{2} \in \mathbb{R}^{n}$ be two datapoints from the preprocessed input-dataset $D_{1}HD_{0}\mathcal{X}$.
Let $d$ be a fixed integer constant.
Let $R=\{i_{1},...,i_{r}\}$ be some $r$-element subset of the set $\{1,...,m\}$, where $m$ stands for the number of rows used in the construction of matrices $G_{struct}^{i}$ (key building blocks of our structured mechanism). Finally, let $\alpha_{1},...,\alpha_{r}$ be positive integers such that
$\alpha_{1}+...+\alpha_{r} = d$.
\begin{definition}
For three vectors: $v,w,z \in \mathbb{R}^{n}$ and a given nonlinear function $f \in \mathcal{F}$ we denote:
$$\phi(v,w,z) = f(z \cdot v) f(z \cdot w).$$
\end{definition}
We will show that for a variety of functions $\Psi : \mathbb{R}^{r} \rightarrow \mathbb{R}$ the expected value of the expression $T^{G,d}_{v^{1},v^{2}}(\mathcal{R}, \alpha_{1},...,\alpha_{r})$ given by the formula:
\begin{equation}
\label{eq-unstructured}
\Psi(\phi_{1}(v^{1},v^{2},g^{i_{1}})^{\alpha_{1}},...,\phi_{r}(v^{1},v^{2},g^{i_{r}})^{\alpha_{r}}),
\end{equation}
where $g^{1},...,g^{m}$ is the set of $m$ gaussian vectors forming gaussian matrix $G$, each obtained by sampling independently $n$ values from the distribution $\mathcal{N}(0,1)$ and $\phi_{i}$s differ by the choice of nonlinear mapping $f_{i} \in \mathcal{F}$, can be accurately approximated by its structured version $T^{A,d}_{v^{1},v^{2}}((\mathcal{R}, \alpha_{1},...,\alpha_{r})$
which is of the form:
\begin{equation}
\label{eq-structured}
\Psi(\phi_{1}(v^{1},v^{2},a^{i_{1}})^{\alpha_{1}},...,\phi_{r}(v^{1},v^{2},a^{i_{r}})^{\alpha_{r}}),
\end{equation}
where $a^{1},...,a^{m}$ are rows of the structured matrix $A = G_{struct}^{i}$.
The importance of $T^{G,d}_{v^{1},v^{2}}(\mathcal{R}, \alpha_{1},...,\alpha_{r})$ and $T^{A,d}_{v^{1},v^{2}}(\mathcal{R}, \alpha_{1},...,\alpha_{r})$ lies in the fact that $d^{th}$ moments of the random variables approximating considered kernels in the unstructured and structured mechanism can be expressed as weighted sums of the expressions of the form $T^{G,d}_{v^{1},v^{2}}(\alpha_{1},...,\alpha_{r})$
and $T^{A,d}_{v^{1},v^{2}}(\alpha_{1},...,\alpha_{r})$ respectively if $\Psi(x_{1},...,x_{r}) = x_{1} \cdot ... \cdot x_{r}$.
Thus if $T^{A,d}_{v^{1},v^{2}}(\alpha_{1},...,\alpha_{r})$ closely approximates $T^{G,d}_{v^{1},v^{2}}(\alpha_{1},...,\alpha_{r})$ then the corresponding moments are similar. That, as we will see soon, implies several theoretical guarantees for the structured method. In particular, this means that the variances are similar. Since in the unstructured setting the variance is of the order $O(\frac{1}{m})$, that will be also the case for the structured setting. This in turn will imply concentration results providing theoretical explanation for the observations from the experimental section that show the quality of the proposed structured setting.
We need to introduce a few definitions.
\begin{definition}
We denote by $\Delta^{\xi}_{s}$ the supremum of the expression $\|\xi(y_{1},...,y_{m}) - \xi(y^{\prime}_{1},...,y^{\prime}_{m})\|$ over all pairs of vectors $(y_{1},...,y_{m}), (y^{\prime}_{1},...,y^{\prime}_{m})$ from the domain $\mathcal{D}$ that differ on at most one dimension and by at most $s$.
We say that a function $\xi : \mathbb{R}^{m} \rightarrow \mathbb{R}$ is $M$-bounded in the domain $\mathcal{D}$ if $\Delta^{\xi}_{\infty} = M$.
\end{definition}
Note that the value of the function $\phi_{i}(v^{1},v^{2},g^{i})^{\alpha_{i}}$
depends only on the projection $g^{i}_{proj}$ of $g^{i}$ on the $2$-dimensional space spanned by $v^{1}$ and $v^{2}$. Thus for a given pair $v^{1},v^{2}$ function $\phi$ is in fact a function $B^{v^{1},v^{2}}_{i}$ of this projection.
\begin{definition}
Define:
\begin{align}
\begin{split}
p_{\lambda,\epsilon} = \sup_{i, v^{1},v^{2}, \|\zeta|_{\infty} \leq \epsilon}
\mathbb{P}[|B^{v^{1},v^{2}}_{i}(g^{i}_{proj}+\zeta)-\\
B^{v^{1},v^{2}}_{i}(g^{i}_{proj})| > \lambda],
\end{split}
\end{align}
where the supremum is taken over all indices $i=1,...,m$, all pairs of linearly independent vectors from the domain, all coordinate systems in $span(v^{1},v^{2})$ and vectors $\zeta$ of $L_{1}$-norm at most $\epsilon$ in some of these coordinate systems.
\end{definition}
We will use the following notation: $\sigma_{i,j}(n_{1},n_{2}) = \vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}$.
To compress the statements of our theoretical results, we will use also the following notation:
$$\xi(i_{i},i_{2}) = 2\chi(i_{1},i_{2}) \sqrt{\sum_{1 \leq n_{1} < n_{2} \leq n} (\sigma_{i_{1},i_{2}}(n_{1},n_{2}))^{2}},$$
We will also denote:
$\lambda(i_{1},i_{2}) = \sum_{j=1}^{n}|\sigma_{i_{1},i_{2}}(j,j)|$ and
$\tilde{\lambda}(i_{1},i_{2}) = |\sum_{j=1}^{n}\sigma_{i_{1},i_{2}}(j,j)|$
for $1 \leq i_{1} \leq i_{2} \leq m$ (see: \ref{sec:p_model}).
Note first that the preprocessing step preserves kernels' values since transformation $HD_{0}$
is an isometry and considered kernels are spherically-invariant.
We start with Lemma \ref{unbiasedness_lemma}.
\begin{proof}
Note that it suffices to show that for any two given vectors $x,y \in \mathbb{R}^{n}$ the following holds:
\begin{equation}
\label{mean_eq_1}
\mathbb{E}[f(G^{i}_{struct}x) \cdot f(G^{i}_{struct}y)]=\mathbb{E}[f(Gx) \cdot f(Gy)],
\end{equation}
where $G$ is the unstructured gaussian matrix. Let $g^{i,j}_{struct}$ be the $j^{th}$ row of $G^{i}_{struct}$
and let $g^{j}$ be the $j^{th}$ row of $G$. Note that we have:
\begin{equation}
\label{mean_eq_2}
\mathbb{E}[f(g^{i,j}_{struct} \cdot x) f(g^{i,j}_{struct} \cdot y)]=\mathbb{E}[f(g^{j} \cdot x)f(g^{j} \cdot y)].
\end{equation}
The latter follows from the fact that $g^{i,j}_{struct}$ has the same distribution as $g$. To see this note that
$g^{i,j}_{struct} = g \cdot P_{i}$. Thus dimensions of $g^{i,j}_{struct}$ are projections of $g$ onto columns of $P_{i}$.
Each projection is trivially gaussian from $\mathcal{N}(0,1)$ (that is implied by the fact that each column is normalized).
The independence of different dimensions of $g^{i,j}_{struct}$ comes from the observation that different columns are orthogonal.
Thus we can use a simple property of gaussian vectors stating that the projections of a gaussian vector on mutually orthogonal directions are
independent. The equation \ref{mean_eq_1} implies equation \ref{mean_eq_2} by the linearity of expectation and that completes the proof.
\end{proof}
Now we prove Theorem \ref{main_theorem_1}.
This one is easily implied by a more general result that we state below.
We will assume that function $\Psi$ from equations: \ref{eq-unstructured}, \ref{eq-structured} is $M$-bounded for some given $M>0$.
We will assume that expected values defining $T^{A,d}$ are not with respect to the random choices determining $P_{i}s$.
\begin{theorem}
\label{main_theorem}
Let $v^{1},v^{2} \in \mathbb{R}^{n}$ be two vectors from a dataset $\mathcal{X}$.
Let $\mathcal{R}=\{i_{1},...,i_{r}\} \in \{1,...,m\}$ and let $\alpha_{1},...,\alpha_{r}$ be the set
of positive integers such that $\alpha_{1} + ... + \alpha_{r} = d$.
Assume that each structured matrix $G_{struct}^{i}$ consists of $m$ rows
and either $\sup_{1 \leq i_{1} < i_{2} \leq m} \lambda(i_{1},i_{2}) = o(\frac{n}{\log^{2}(n)})$
if $P_{i}s$ were constructed deterministically or $sup_{1 \leq i_{1} < i_{2} \leq m} \mathbb{E}[\tilde{\lambda}(i_{1},i_{2})] =
o(\frac{n}{\log^{2}(n)})$ if $P_{i}s$ were constructed randomly.
In the latter case assume also that for any $1 \leq i_{1} < i_{2} \leq m$ and $1 \leq n_{1} < n_{2} \leq n$ the $n_{1}^{th}$
column of $P_{i_{1}}$ is chosen independently from the $n_{2}^{th}$ column of $P_{i_{2}}$.
Denote by $\Psi_{max}$ the maximal value of the function $\Psi$ for the datapoints from $\mathcal{X}$.
Let $q_{v^{1},v^{2}}^{d} = |T^{A,d}_{v^{1},v^{2}}(\mathcal{R},\alpha_{1},...,\alpha_{r})-T^{G,d}_{v^{1},v^{2}}(\mathcal{R},\alpha_{1},...,\alpha_{r})|$ denote the absolute value of the difference of the two fixed terms on the weighted sum for the $d$-moments of the kernel's approximation in the structured $\mathcal{P}$-model setting and the fully unstructured setting.
Then for any $\lambda, \epsilon > 0$, $T>0$, $n$ large enough and $P_{i}s$ chosen deterministically we have:
$$q_{v^{1},v^{2}}^{d} \leq
(p_{gen} + p_{struct})\Psi_{max} + \sum_{i=0}^{d} p_{f}^{i}(iM + (d-i)\Delta^{\Psi}_{\lambda}),$$ where:
\begin{equation}
p_{gen} = \frac{4r}{\sqrt{2 \pi T}} e^{-\frac{T}{2}} + 4ne^{-\frac{\log^{2}(n)}{8}},
\end{equation}
\begin{equation}
p_{f}^{i} = {d \choose i} (p_{\lambda, \epsilon})^{i}
\end{equation}
and
\begin{align}
\begin{split}
p_{struct} = 4\sum_{i=1}^{m}\chi(i,i)e^{-\frac{1}{2\xi^{2}(i,i)}\frac{n^{2}}{\log^{6}(n)}}\\
+2\sum_{1 \leq i_{1} \leq i_{2} \leq m}\chi(i_{1},i_{2})e^{-\frac{\epsilon^{2}n^{\frac{3}{2}}}{2\xi^{2}(i_{1},i_{2})T\log^{4}(n)}}\\
\end{split}
\end{align}
If $P_{i}s$ are chosen from the probabilistic model then the above holds with probability at least
$1-p_{wrong}$, where $p_{wrong} = 2\sum_{i \leq i_{1} < i_{2} \leq m}e^{-\frac{n^{2}}{8\log^{6}(n)\sum_{j=1}^{n}(\sigma_{i_{1},i_{2}}(j,j))^{2}}}$.
\end{theorem}
\begin{proof}
Consider the expression $$q^{d}_{v_{1},v_{2}} = |T^{A,d}_{v^{1},v^{2}}(\mathcal{R}, \alpha_{1},...,\alpha_{r}) -
T^{G,d}_{v^{1},v^{2}}(\mathcal{R}, \alpha_{1},...,\alpha_{r})|.$$
We will use formulas for $T^{G,d}$ and $T^{A,d}$ given by equations: \ref{eq-unstructured} and \ref{eq-structured}.
Without loss of generality we will assume that $A=G^{i}_{struct}D_{1}$ i.e. in our theoretical analysis we will make $D_{1}$
a part of the structured mechanism and move it away from the preprocessing phase (obviously both ways are equivalent because of the
associative property of matrix mutliplication).
We have already noted that each argument of the function $\Psi$ from equations: \ref{eq-unstructured} and \ref{eq-structured}
depends only on the projections of $a^{i_{1}},...,a^{i_{r}}$ on the $2$-dimensional space spanned by $v^{1}$ and $v^{2}$.
Denote these projections as: $a^{i_{1}}_{proj}$,...,$a^{i_{r}}_{proj}$ respectively and fix some orthonormal basis $\mathcal{B}$ of this $2$-dimensional space.
As we will see soon, in the $\mathcal{P}$-model setting the coordinates of $a^{i}_{proj}s$ in $\mathcal{B}$ can be expressed as $g \cdot s^{i,j}$ for $j=1,2$,
where $g$ is a vector representing a budget of randomness of the corresponding $\mathcal{P}$-model and $s^{i,j}$s are some vectors from $\mathbb{R}^{t}$
(parameter $t$ stands for the length of $g$).
We will show that $s^{i,j}$s, even though not necessarily pairwise orthogonal, are close to be pairwise orthogonal with high probability.
Let us assume now that vectors $s^{i,j}$ can be chosen in such a way that each $s^{i,j}$ satisfies: $s^{i,j} = w^{i,j} + \rho(i,j)$,
where vectors $w^{i,j}$ are mutually orthogonal, we have $\|s^{i,j}\|_{2} = \|w^{i,j}\|_{2}$ and furthermore
$\|\rho(i,j)\|_{2} \leq \rho$ for some given $\rho > 0$. We call this property the $\rho$-orthogonality property.
We will later show that the $\rho$-orthogonality property depends on the random diagonal matrix $D_{1}$.
Assume now that the $\rho$-orthogonality property is satisfied. Denote by $g^{\mathcal{H}}$ the projection of the ``budget-of-randomness''
vector $g$ onto $2r$-dimensional linear space $\mathcal{H}$ spanned by vectors from $\{s^{i,j}\}$.
Note that then the coordinates of $a^{i}_{proj}s$ in $\mathcal{B}$ can be rewritten as $g \cdot w^{i,j} + \epsilon(i,j)$,
where $|\epsilon(i,j)| \leq \epsilon$ and $\epsilon = \|g^{\mathcal{H}}\|_{2} \rho$.
Thus each $\psi_{i}$ in the formula from equation \ref{eq-structured} can be then
expressed as $B_{i}^{v^{1},v^{2}}(g^{i}_{proj} + \epsilon(i))$, where $g^{i}_{proj}s$ stand for the
projections onto $2$-dimensional linear space spanned by $v^{1}$ and $v^{2}$ of independent copies of gaussian vectors $g^{i}$. Each $g^{i}$ is of the same distribution as the corresponding
structured vector $a^{i}$ and $\epsilon(i)s$ are vectors with the $L_{1}$-norm satisfying
$\|\epsilon(i)\| \leq \epsilon$. The independence comes from the fact that variables of the form $g \cdot w^{i,j}$ are independent. That, as in the
proof of Lemma \ref{unbiasedness_lemma} is implied by the well known fact that dot products of a given gaussian vector with orthogonal vectors
are independent.
Note that if not the term $\epsilon(i)$ then the formula for $T^{A,d}$ would collapse to its unstructured counterpart $T^{G,d}$.
We will argue that both expressions are still close to each other if $\epsilon(i)$ have small $L_{1}$-norm.
Let us fix $\lambda > 0$. Our goal is to count these indices $i$ that satisfy the following:
$|\psi_{i}(v^{1},v^{2},g^{i})^{\alpha^{i}} - \psi_{i}(v^{1},v^{2},g^{i})^{\alpha^{i}}| > \lambda$,
where $g^{i}s$ corresponds to the aforementioned independent counterparts of $a^{i}s$. We call them
\textit{bad indices}.
Based on what we have said so far, we can conclude that the latter inequality can be expressed as
$|B_{i}^{v^{1},v^{2}}(g^{i}_{proj} + \epsilon(i)) - B_{i}^{v^{1},v^{2}}(g^{i}_{proj})| > \lambda$.
Let us first find the upper bound on the probability of the event that the number of bad indices is
$j$ for some fixed $1 \leq j \leq d$. Note that since $g^{i}s$ are independent, we can use Bernoulli scheme
to find that upped bound. Using the definition of $p_{\lambda, \epsilon}$ we obtain an upper bound of the form
$p_{upper} \leq {d \choose j} (p_{\lambda, \epsilon})^{j}$. If the number of bad indices is $j$ then
by the definition of $M$ and $\Delta^{\Psi}_{\lambda}$ we see that $T^{A,d}$ differs from $T^{G,d}$ by at most
$iM + (d-i)\Delta^{\Psi}_{\lambda}$. Summing up over all indices $j$ we get the second term of the upper bound on
$q^{d}_{v^{1},v^{2}}$ from the statement of the theorem.
However the $\rho$-orthogonality does not have to hold. Note that (by the definition of $\Psi_{max}$) to finish the proof
of the theorem it suffices to show that the probability of $\rho$-orthogonality not to hold is at most $p_{gen} + p_{struct}$.
\begin{lemma}
\label{technical_lemma}
The $\rho$-orthogonality property holds with probability at least $1 - (p_{gen} + p_{struct})$.
\end{lemma}
\begin{proof}
We need the following definition.
\begin{definition}
Let $x=(x_{1},...,x_{n})$ be a vector with $\|x\|_{2}=1$. We say that $x$ is $\theta$-balanced if $|x_{i}| \leq \frac{\theta}{\sqrt{n}}$
for $i=1,...,n$.
\end{definition}
For a fixed pair of vectors $v^{1},v^{2} \in \mathcal{X}$ choose some orthonormal basis $\mathcal{B} = \{x^{1},x^{2}\}$
of the $2$-dimensional space spanned by $v^{1}$ and $v^{2}$.
Let $\tilde{x}^{1}$ and $\tilde{x}^{2}$ be the images of $x^{1}$ and $x^{2}$ under transformation $HD_{0}$, where $H$
is a Hadamard matrix and $D_{0}$ is a random diagonal matrix.
We will show now that with high probability $\tilde{x}^{1}$ and $\tilde{x}^{2}$ are $\log(n)$-balanced.
Indeed, the $i^{th}$ dimension of $\tilde{x}^{1}$ is of the form:
$\tilde{x}^{1}_{i} = h_{i,1}x^{1}_{1} + ... + h_{i,n}x^{1}_{n}$, where
$h_{i,j}$ stands for the entry in the $i^{th}$ row and $j^{th}$ column of a matrix $HD_{0}$.
We need to find a sharp upper bound on $\mathbb{P}[|h_{i,1}x^{1}_{1} + ... + h_{i,n}x^{1}_{n}| \geq a]$ for $a = \frac{\log(n)}{\sqrt{n}}$.
We will use the following concentration inequality, calles \textit{Azuma's inequality}
\begin{lemma}
Let $X_{1},...,X_{n}$ be a martingale and assume that $-\alpha_{i} \leq X_{i} \leq \beta_{i}$ for some positive constants $\alpha_{1},...,\alpha_{n}, \beta_{1},...,\beta_{n}$.
Denote $X = \sum_{i=1}^{n} X_{i}$.
Then the following is true:
$$\mathbb{P}[|X - \mathbb{E}[X]| > a] \leq 2e^{-\frac{a^{2}}{2\sum_{i=1}^{n}(\alpha_{i} + \beta_{i})^{2}}}$$
\end{lemma}
In our case $X_{j} = h_{i,j}x^{1}_{j}$ and $\alpha_{i} = \beta_{i} = \frac{1}{\sqrt{n}}$.
Applying Azuma's inequality, we obtain the following bound:
$\mathbb{P}[|h_{i,1}x^{1}_{1} + ... + h_{i,n}x^{1}_{n}| \geq \frac{\log(n)}{\sqrt{n}}] \leq 2e^{-\frac{\log^{2}(n)}{8}}$.
The probability that all $n$ dimensions of $\tilde{x}^{1}$ and $\tilde{x}^{2}$ have absolute value at most $\frac{\log(n)}{\sqrt{n}}$ is, by the union bound, at least $p_{balanced} = 1 - 2n \cdot 2e^{-\frac{\log^{2}(n)}{8}} = 1-4n e^{-\frac{\log^{2}(n)}{8}}$.
Thus this a lower bound on the probability that $\tilde{x}^{1}$ and $\tilde{x}^{2}$ are $\log(n)$-balanced.
We will use this lower bound later. Now note that it does not depend on the particular form of the structured matrix since it is only related to the preprocessing phase, where linear mappings $D_{0}$ and $H$ are applied.
For simplicity we will now denote $\hat{x}^{1}$ and $\hat{x}^{2}$ simply as $x^{1}$ and $x^{2}$, knowing these are the original vectors after applying linear transformation $HD_{0}$.
Let us get back to the projections of $a^{i}s$ onto $2$-dimensional linear space spanned by $v^{1}$ and $v^{2}$.
Note that we have already noticed that $a^{i} \cdot x^{j}$ ($j=1,2$) is of the form $g \cdot s^{i,j}$ for some vector $s^{i,j} \in \mathbb{R}^{t}$, where $t$ is the size of the ``budget of randomness'' used in the given $\mathcal{P}$-model.
From the definition of the $\mathcal{P}$-model we obtain:
\begin{equation}
s^{i,j}_{l} = d_{1}p^{i}_{l,1}x^{j}_{1} + ... + d_{n}p^{i}_{l,n}x^{j}_{n}
\end{equation}
for $l=1,...,t$, where $s^{i,j}_{l}$ stands for the $l^{th}$ dimension of $s^{i,j}$,
$p^{i}_{l,k}$ is the entry in the $l^{th}$ row and $k^{th}$ column of $P_{i}$
and $d_{r}s$ are the values on the diagonal of the matrix $D_{0}$.
As we noted earlier, we want to show that $s^{i,j}s$ are close to be mutually orthogonal.
To do it, we will compute dot products $s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}$.
We will first do it for $i_{1}=i_{2}$.
We have:
\begin{align}
\begin{split}
s^{i_{1},j_{1}} \cdot s^{i_{1},j_{2}} = x^{j_{1}}_{1}x^{j_{2}}_{1}\sum_{l=1}^{t}(p^{i_{1}}_{l,1})^{2}+...+x^{j_{1}}_{n}x^{j_{2}}_{n}\sum_{l=1}^{t}(p^{i_{1}}_{l,n})^{2}\\
+2 \sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}(\sum_{i=1}^{t} p^{i_{1}}_{l,n_{1}}p^{i_{2}}_{l,n_{2}})
\end{split}
\end{align}
Now we take advantage of the normalization property of the matrices $P_{i}$ and the fact that $x^{1}$ is orthogonal to $x^{2}$ and conclude that the first term on the RHS of the equation above is equal to $0$.
Thus we have:
\begin{equation}
s^{i_{1},j_{1}} \cdot s^{i_{1},j_{2}} = 2 \sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}\sigma_{i_{1},i_{1}}(n_{1},n_{2}).
\end{equation}
Note that if for any fixed $P_{i}$ any two different columns of $P_{i}$ are orthogonal then
$\sigma_{i_{1},i_{1}}(n_{1},n_{2}) = 0$ and thus $s^{i_{1},j_{1}} \cdot s^{i_{1},j_{2}} = 0$.
This is the case for many structured matrices constructed according to the $\mathcal{P}$-model, for instance circulant, Toeplitz or Hankel matrices.
Let us consider now $s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}$ for $i_{1} \neq i_{2}$.
By the previous analysis, we obtain:
\begin{align}
\begin{split}
s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}} = \sigma_{i_{1},i_{2}}(1,1)x^{j_{1}}_{1}x^{j_{2}}_{1}+...+\sigma_{i_{1},i_{2}}(n,n)x^{j_{1}}_{n}x^{j_{2}}_{n}\\
+ 2 \sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}\sigma_{i_{1},i_{2}}(n_{1},n_{2}).
\end{split}
\end{align}
This time in general we cannot get rid of the first term in the RHS expression. This can be done if columns of the same indices in different $P_{i}s$ are orthogonal. This is in fact again the case for circulant, Toeplitz or Hankel matrices.
Let us now fix some $1 \leq i_{1} \leq m$ and $\kappa > 0$.
Our goal is to find an upper bound on the following probability: $\mathbb{P}[|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \kappa]$.
We have:
\begin{align}
\begin{split}
\mathbb{P}[|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \kappa] = \\
\mathbb{P}[|\sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}2\sigma_{i_{1},i_{2}}(n_{1},n_{2})| > \kappa].
\end{split}
\end{align}
For $\{n_{1},n_{2}\}$ such that $n_{1} \neq n_{2}$ and $\sigma_{i_{1},i_{1}}(n_{1},n_{2}) \neq 0$
let us now consider random variables $Y_{n_{1},n_{2}}$ that are defined as follows
\begin{equation}
Y_{n_{1},n_{2}} = 2d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}\sigma_{i_{1},i_{1}}(n_{1},n_{2}).
\end{equation}
From the definition of the chromatic number $\chi(i_{1},i_{1})$ we can deduce that the set of all this random
variables can be partitioned into at most $\chi(i_{1},i_{1})$ subsets such that random variables in each subset
are independent.
Let us denote these subsets as: $\mathcal{L}_{1},...,\mathcal{L}_{r}$, where $r \leq \chi(i_{1},i_{1})$.
Note that an event $\{|\sum_{1 \leq n_{1} < n_{2} \leq n}d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}2\sigma_{i_{1},i_{1}}(n_{1},n_{2})| > \kappa\}$
is contained in the sum of the events: $\mathcal{E} = \mathcal{E}_{1} \cup ... \cup \mathcal{E}_{r}$, where each $\mathcal{E}_{j}$ is defined as follows:
\begin{equation}
\mathcal{E}_{j} = \{|\sum_{Y \in \mathcal{L}_{j}} Y| \geq \frac{\kappa}{\chi(i_{1},i_{1})}\}.
\end{equation}
Thus, from the union bound we get:
\begin{equation}
\mathbb{P}[\mathcal{E}] \leq \sum_{i=1}^{\chi(i_{1},i_{1})} \mathbb{P}[\mathcal{E}_{i}].
\end{equation}
Now we can use Azuma's inequality to find an upper bound on $\mathcal{P}[\mathcal{E}_{i}]$
and we obtain:
\begin{equation}
\mathbb{P}[\mathcal{E}_{i}] \leq 2e^{-\frac{\frac{\kappa^{2}}{\chi^2(i_{1},i_{1})}}{2 \sum_{1 \leq n_{1} <
n_{2} \leq n} (2\sigma_{i_{1},i_{1}}(n_{1},n_{2}))^{2}(x^{j_{1}}_{n_{1}})^{2}(x^{j_{2}}_{n_{2}})^{2}}}.
\end{equation}
Now, if we assume that the vectors of the orthonormal basis $\mathcal{B}$ are $\log(n)$-balanced, then
by the union bound we obtain the following upper bound on the probability $\mathbb{P}[\mathcal{E}]$:
\begin{equation}
\mathbb{P}[\mathcal{E}] \leq 2\chi(i_{1},i_{1})e^{-\frac{\kappa^{2} n^{2}}{2\log^{4}(n)\chi^{2}(i_{1},i_{1})\sum_{1 \leq n_{1} < n_{2} \leq n}
(2\sigma_{i_{1},i_{1}}(n_{1},n_{2}))^{2}}}.
\end{equation}
We can conclude, using the union bound again, that for a $\log(n)$-balanced basis $\mathcal{B}$ the probability that there exist $i_{1},j_{1},j_{2}$ such that:
$|s^{i_{1},j_{1}} \cdot s^{i_{1},j_{2}}| > \kappa$ is at most
\begin{equation}
p_{1,bad}(\kappa) \leq 2\sum_{i=1}^{m} \chi(i,i)e^{-\frac{\kappa^{2}}{2\xi^{2}(i,i)}\frac{n^{2}}{\log^{4}(n)}}.
\end{equation}
Now let us find an upper bound on the expression $p_{2,bad}(\kappa) = \mathbb{P}[\exists_{i_{1},i_{2},j_{1},j_{2}, i_{1} \neq i_{2}} : |s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \kappa]$,
where $i_{1} \neq i_{2}$.
We will assume that vectors of the basis $\mathcal{B}$
are $\log(n)$-balanced.
Using the formula on $s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}$ for $i_{1} \neq i_{2}$, we get:
\begin{align}
\begin{split}
\mathbb{P}[|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \kappa ] = \\ \mathbb{P}[|\sigma_{i_{1},i_{2}}(1,1)x^{j_{1}}_{1}x^{j_{2}}_{1}+...+\sigma_{i_{1},i_{2}}(n,n)x^{j_{1}}_{n}x^{j_{2}}_{n}\\
+ 2 \sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}\sigma_{i_{1},i_{2}}(n_{1},n_{2})|> \kappa].
\end{split}
\end{align}
Assume first that $P_{i}s$ are chosen deterministically.
Note that by $\log(n)$-balanceness, we have:
\begin{equation}
|\sum_{n_{1}=1}^{n} \sigma_{i_{1},i_{2}}(n_{1},n_{1})x^{j_{1}}_{1}x^{j_{2}}_{1}| \leq \frac{\log^{2}(n)}{n}\lambda(i_{1},i_{2}).
\end{equation}
Thus, by the triangle inequality, we have:
\begin{align}
\begin{split}
\mathbb{P}[|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \kappa ] \\ \leq \mathbb{P}[|2 \sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}\sigma_{i_{1},i_{2}}(n_{1},n_{2})|
\geq \\
\kappa - \frac{\log^{2}(n)}{n}\lambda(i_{1},i_{2})].
\end{split}
\end{align}
Using the same analysis as before, we then obtain the following bound on $p_{bad}(\kappa, \theta)$:
\begin{equation}
p_{2,bad}(\kappa) \leq 2\sum_{1 \leq i_{1} < i_{2} \leq m} \chi(i_{1},i_{2})e^{-\frac{(\kappa-\frac{\log^{2}(n)}{n}\lambda(i_{1},i_{2}))^{2}}{2\xi^{2}(i,i)}\frac{n^{2}}{\log^{4}(n)}}.
\end{equation}
We can conclude that in the setting where $P_{i}s$ are chosen deterministically, under our assumptions on $\lambda(i_{1},i_{2})$, for $\kappa > 0$
that does not depend on $n$ and
$n$ large enough the following is true. The probability that there exist two different vector $s^{i_{1},j_{1}}$, $s^{i_{2},j_{2}}$
such that $|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \kappa$ satisfies:
\begin{equation}
p_{bad}(\kappa) \leq 2\sum_{1 \leq i_{1} \leq i_{2} \leq m} \chi(i_{1},i_{2})e^{-\frac{(\kappa-\frac{\log^{2}(n)}{n}\lambda(i_{1},i_{2}))^{2}}{2\xi^{2}(i,i)}\frac{n^{2}}{\log^{4}(n)}}.
\end{equation}
Now let us assume that $P_{i}s$ are chosen probabilistically. In that setting we also assume that columns of different indices are chosen
independently (this is the case for instance for the FastFood Transform).
Let us now denote:
\begin{equation}
Y_{j} = \sigma_{i_{1},i_{2}}(j,j)x^{j_{1}}_{j}x^{j^{2}}_{j}
\end{equation}
for $j=1,...,n$.
Denote $Y = \sum_{i=1}^{n} Y_{1} + ... + Y_{n}$. Note that the condition on $\tilde{\lambda}(i_{1},i_{2})$
from the statement of the theorem implies that $\mathbb{E}[Y] = o_{n}(1)$.
From the condition regarding independence of columns of different indices we deduce that $Y_{i}s$ are independent.
Therefore we can apply Azuma's inequality and obtain the following bound on the expression: $\mathbb{P}[|Y - \mathbb{E}[Y]| > a]$:
\begin{align}
\begin{split}
\mathbb{P}[|Y-\mathbb{E}[Y]| > a] \leq 2e^{-\frac{a^{2}}{ 8\frac{\log^{4}(n)}{n^{2}} \sum_{j=1}^{n}(\sigma_{i_{1},i_{2}}^{max}(j,j))^{2}}}.
\end{split}
\end{align}
If we now take $a=\frac{1}{\log(n)}$ and under $\log(n)$-balanceness assumption, we obtain:
\begin{equation}
\mathbb{P}[|Y-\mathbb{E}[Y]| > a] \leq 2e^{-\frac{n^{2}}{ 8\log^{6}(n) \sum_{j=1}^{n}(\sigma_{i_{1},i_{2}}^{max}(j,j))^{2}}}.
\end{equation}
Assume now that $|Y-\mathbb{E}[Y]| \leq \frac{1}{\log(n)}$. This happens with probability
at least $1-p_{wrong}$ with respect to the random choices of $P_{i}s$,
where $p_{wrong} = 2e^{-\frac{n^{2}}{ 8\log^{6}(n) \sum_{j=1}^{n}(\sigma_{i_{1},i_{2}}^{max}(j,j))^{2}}}$.
But then random variable $|Y|$ is of the order $o_{n}(1)$.
Note that we have:
\begin{align}
\begin{split}
\mathbb{P}[|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \kappa ] = \\ \mathbb{P}[|Y
+ 2 \sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}\sigma_{i_{1},i_{2}}(n_{1},n_{2})|> \kappa].
\end{split}
\end{align}
Thus, using our bound on $Y$ for a fixed $\kappa$ and $n$ large enough we can repeat previous analysis and conclude that in the probabilistic setting of $P_{i}s$
the following is true:
\begin{equation}
p_{bad}(\kappa) \leq 2\sum_{1 \leq i_{1} \leq i_{2} \leq m} \chi(i_{1},i_{2})e^{-\frac{(\frac{\kappa}{2})^{2}}{2\xi^{2}(i,i)}\frac{n^{2}}{\log^{4}(n)}}.
\end{equation}
Thus we can conclude that in both the deterministic and probabilistic setting for $P_{i}s$ we get:
\begin{equation}
p_{bad}(\kappa) \leq 2\sum_{1 \leq i_{1} \leq i_{2} \leq m} \chi(i_{1},i_{2})e^{-\frac{\kappa^{2}}{8\xi^{2}(i,i)}\frac{n^{2}}{\log^{4}(n)}}.
\end{equation}
Now we will show that the squared lengths of vectors $s^{i,j}$ are well concentrated around their means and that these means are equal to $1$.
Let us remind that we have:
\begin{equation}
s^{i,j}_{l} = d_{1}p^{i}_{l,1}x^{j}_{1}+...+d_{n}p^{i}_{l,n}x^{j}_{n}.
\end{equation}
Thus we get:
\begin{align}
\begin{split}
\|s^{i,j}\|^{2}_{2} = \sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}2\sigma_{i,i}(n_{1},n_{2}) + \\
\sum_{n_{1}=1}^{n} (\sigma_{i,i}(n_{1},n_{1}))^{2}(x^{j}_{n_{1}})^{2}=\\
\sum_{1 \leq n_{1} < n_{2} \leq n} d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}2\sigma_{i,i}(n_{1},n_{2}) + 1,
\end{split}
\end{align}
where the last inequality comes from the fact that each column of each $P_{i}$ has $l_{2}$-norm equal to $1$.
Since obviously $\mathbb{E}[d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}2\sigma_{i,i}(n_{1},n_{2}) ] = 0$,
then indeed $\mathbb{E}[\|s^{i,j}\|^{2}_{2}] = 1$.
Let us find the upper bound on the following probability:
$\mathbb{P}[|\|s^{i,j}\|^{2}_{2}-1| > \frac{1}{\log(n)}]$.
We have:
\begin{align}
\begin{split}
\mathbb{P}[|\|s^{i,j}\|^{2}_{2}-1| > \frac{1}{\log(n)}] = \\
\mathbb{P}[|d_{n_{1}}d_{n_{2}}x^{j_{1}}_{n_{1}}x^{j_{2}}_{n_{2}}2\sigma_{i,i}(n_{1},n_{2})| > \frac{1}{\log(n)}].
\end{split}
\end{align}
We can again apply Azuma's inequality and the union bound as we did before and obtain:
\begin{align}
\begin{split}
\mathbb{P}[ \exists_{i,j} : |\|s^{i,j}\|^{2}_{2}-1| > \frac{1}{\log(n)}]
\leq p_{s},
\end{split}
\end{align}
where $p_{s} = 4\sum_{i=1}^{m}\chi(i,i)e^{-\frac{1}{2\xi^{2}(i,i)\log^{2}(n)}\frac{n^{2}}{\log^{4}(n)}}$.
We will assume now that all $s^{i,j}$ satisfy:
$|\|s^{i,j}\|^{2}_{2}-1| \leq \frac{1}{\log(n)}$, in particular:
\begin{equation}
\sqrt{1-\frac{1}{\log(n)}} \leq \|s^{i,j}\|_{2} \leq \sqrt{1 + \frac{1}{\log(n)}}.
\end{equation}
Let us assume right now that the above inequality holds. Let $\{w^{i,j}\}$ be a set of vectors
obtained from $\{s^{i,j}\}$ by the Gram-Schmidt process. Without loss of generality we can assume that $\|w^{i,j}\|_{2} = \|s^{i,j}\|_{2}$. Note that the size of the set $\{s^{i,j}\}$
is in fact not $2m$, but $2r$ and in all practical application $r \ll m$.
Assume now that $|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| \leq \kappa$ for any two different vectors $s^{i_{1},_{j_{1}}}, s^{i_{2},j_{2}}$ and some fixed $\kappa >0$.
Now, one can easily note that directly from the description of the Gram-Schmidt process that it leads to the set of vectors $\{w^{i,j}\}$ such that $\|s^{i,j}-w^{i,j}\|_{2} \leq \kappa \Gamma(2r)$, where $\Gamma$ is some constant that depends just on the size of the set $\{s^{i,j}\}$.
Thus if we want $\rho$-orthogonality with $\rho = \frac{\epsilon}{\|g^{\mathcal{H}}\|_{2}}$,
where $g^{\mathcal{H}}$ stands for the random projection of a vector $g$ onto $2r$-dimensional linear space spanned by vectors from $\{s^{i,j}\}$,
then we want to have:
\begin{equation}
\frac{\epsilon}{\|g^{\mathcal{H}}\|_{2}} = \kappa \Gamma(2r).
\end{equation}
Thus we need to take:
\begin{equation}
\kappa = \frac{\epsilon}{\Gamma(2r)\|g^{\mathcal{H}}\|_{2}}.
\end{equation}
Note that $g^{\mathcal{H}}$ is a $2r$-dimensional gaussian vector.
Now let us take some $T>0$. By the union bound the probability that $g^{\mathcal{H}}$
has $l_{2}$ norm greater than $\sqrt{2r} \cdot \sqrt{T}$ is at most: $2r \mathbb{P}[|\hat{g}|^{2} > T]$, where $\hat{g}$ stands for a gaussian random variable taken from $\mathcal{N}(0,1)$.
Now we use the following inequality for a tail of the gaussian random variable:
\begin{equation}
\mathbb{P}[|\hat{g}| > x] \leq 2\frac{e^{-\frac{x^{2}}{2}}}{x\sqrt{2\pi}}.
\end{equation}
Thus we can conclude that the probability that $g^{\mathcal{H}}$ has $l_{2}$ norm larger than
$\sqrt{2r} \cdot \sqrt{T}$ is at most $p_{gauss}(T) \leq \frac{4r}{\sqrt{2\pi T}}$.
In such a case we need to take $\kappa$ of the form:
\begin{equation}
\kappa = \frac{\epsilon}{\Gamma(2r)\sqrt{2r} \sqrt{T}}.
\end{equation}
We are ready to finish the proof of Lemma \ref{technical_lemma}.
Take $\kappa = \frac{\epsilon}{\Gamma(2r)\sqrt{2r} \sqrt{T}}$.
Let us first take the setting where $P_{i}s$ are chosen deterministically.
Take an event $\mathcal{E}_{bad}$ which is the sum of the events which probabilisites are upper-bounded by $p_{gauss}(T)$, $1-p_{balanced}$, $p_{bad}(\kappa)$ and $p_{s}$.
By the union bound, the probability of that event is at most $p_{gauss} + (1-p_{balanced}) + p_{bad}(\kappa) + p_{s}$ which is upper-bounded by $p_{gen} + p_{struct}$ for $n$ large enough.
Note that if $\mathcal{E}_{bad}$ does not hold then $\rho$-orthogonality is satisfied.
Now let us take the probabilistic setting for choosing $P_{i}s$. We proceed similarly.
The only difference is that right now we need to assume that the event upper-bounded by $p_{wrong}$ does not hold (this one depends only on the random choices for setting up $P_{i}s$). Thus again we get the statement of the lemma.
That completes the proof of Lemma \ref{technical_lemma}.
\end{proof}
As mentioned above, the proof of Lemma \ref{technical_lemma} completes the proof of the theorem.
\end{proof}
Now we prove Theorem \ref{main_variance_theorem}.
\begin{proof}
Fix some $\vv{x}, \vv{z} \in \mathbb{R}^{n}$. Assume that a matrix $\vv{A}$ is used to compute the approximation of the kernel $k(\vv{x},\vv{z})$.
Matrix $\vv{A}$ is either a truly random Gaussian matrix as it is the case in the unstructured computation or a structured matrix produced according to the $\mathcal{P}$-model.
We assume that $\vv{A}$ has $k$ rows and consists of $\frac{k}{m}$ blocks stacked vertically.
If $\vv{A}$ is produced via the $\mathcal{P}$-model then each block is a structured matrix $G^{i}_{struct}$.
The approximation of the kernel $\tilde{k}_{\mathcal{P}}(\vv{x},\vv{z})$ is of the form:
$\tilde{k}_{\vv{A}}(\vv{x},\vv{z}) = \frac{1}{k} \sum_{i=1}^{\frac{k}{m}} \sum_{j=1}^{m} [\phi(a^{i,j} \cdot \vv{x}, a^{i,j} \cdot \vv{y})]$,
where $a^{i,j}$ stands for the $j^{th}$ row of the $i^{th}$ block and
$\phi : \mathbb{R}^{2} \rightarrow \mathbb{R}$ is either of the form $\phi(a,b) = f(a)f(b)$, where $f$ is a ReLU/sign function or $\phi(a,b) = \cos(a)\cos(b) + \sin(a)\sin(b)$.
The latter formula for $\phi$ is valid if a kernel under consideration is Gaussian.
Let use denote the random variable: $\phi(a^{i,j} \cdot \vv{x}, a^{i,j} \cdot \vv{y})$ as $X_{i,j}$.
Then we have:
\begin{equation}
\tilde{k}_{\vv{A}}(\vv{x},\vv{z}) = \frac{1}{k}\sum_{i=1}^{\frac{k}{m}}\sum_{j=1}^{m}X_{i,j}.
\end{equation}
Thus we have:
\begin{align}
\begin{split}
\label{var_equation}
Var(\tilde{k}_{\vv{A}}(\vv{x},\vv{z})) = Var(\frac{1}{k}\sum_{i=1}^{\frac{k}{m}}\sum_{j=1}^{m}X_{i,j}) = \\
\frac{1}{k^{2}}Var(\sum_{i=1}^{\frac{k}{m}}\sum_{j=1}^{m}X_{i,j})=\frac{1}{k^{2}}[\sum_{i=1}^{\frac{k}{m}}\sum_{j=1}^{m} Var(X_{i,j}) + \\
\sum_{i,j_{1} \neq j_{2}}Cov(X_{i,j_{1}},X_{i,j_{2}})].
\end{split}
\end{align}
The last inequality in Eqn.\ref{var_equation} is implied by the fact that different blocks of the structured matrix are computed
independently and thus covariance related to rows from different blocks is $0$.
Therefore we obtain:
\begin{align}
\begin{split}
Var(\tilde{k}_{\vv{A}}(\vv{x},\vv{z})) = \frac{1}{k^{2}}\sum_{i=1}^{\frac{k}{m}}\sum_{j=1}^{m}Var(X_{i,j}) + \\
\frac{1}{k^{2}}\sum_{i, j_{1} \neq j_{2}}
(\mathbb{E}[X_{i,j_{1}},X_{i,j_{2}}]-\mathbb{E}[X_{i,j_{1}}]\mathbb{E}[X_{i,j_{2}}]).
\end{split}
\end{align}
Now note that the first expression on the RHS above is the same for both the structured and unstructured setting.
This is the case since one can note that $X_{i,j}$ has the same distribution in the unstructured and structured setting.
For the same reason the expression $\mathbb{E}[X_{i,j_{1}}]\mathbb{E}[X_{i,j_{2}}]$ is the same for the structured and
unstructured setting. Thus if $\vv{G}$ stands for the fully unstructured model and we denote
$\tilde{k}_{\vv{A}}(\vv{x},\vv{z}) = \tilde{k}_{\mathcal{P}}(\vv{x},\vv{z})$ if $A$ is constructed according to the $\mathcal{P}$-model,
then we get:
\begin{align}
\begin{split}
|Var(\tilde{k}_{\vv{G}}(\vv{x},\vv{z}))-Var(\tilde{k}_{\mathcal{P}}(\vv{x},\vv{z}))| \leq \\
\frac{1}{k^{2}} \sum_{i,j_{1} \neq j_{2}}|\mathbb{E}[X^{\mathcal{P}}_{i,j_{1}}X^{\mathcal{P}}_{i,j_{2}}]-
\mathbb{E}[X^{\vv{G}}_{i,j_{1}}X^{\vv{G}}_{i,j_{2}}]|,
\end{split}
\end{align}
where $X^{\mathcal{P}}_{i,j}$ stands for the version of $X_{i,j}$ if $\vv{A}$ was costructed via the $\mathcal{P}$-model
and $X^{\vv{G}}_{i,j}$ stands for the fully unstructured one.
Therfore we have:
\begin{align}
\begin{split}
|Var(\tilde{k}_{\vv{G}}(\vv{x},\vv{z}))-Var(\tilde{k}_{\mathcal{P}}(\vv{x},\vv{z}))| \leq \\
\frac{1}{k^{2}}\cdot \frac{k}{m} \sum_{j_{1} \neq j_{2}}|\mathbb{E}[X^{\mathcal{P}}_{1,j_{1}}X^{\mathcal{P}}_{1,j_{2}}]-
\mathbb{E}[X^{\vv{G}}_{1,j_{1}}X^{\vv{G}}_{1,j_{2}}]|,
\end{split}
\end{align}
where the latter inequality is implied by the fact that different blocks are constructed independently.
Therefore we get:
\begin{equation}
|Var(\tilde{k}_{\vv{G}}(\vv{x},\vv{z}))-Var(\tilde{k}_{\mathcal{P}}(\vv{x},\vv{z}))| \leq \frac{1}{k^{2}} \cdot \frac{k}{m} {m \choose 2} \beta,
\end{equation}
where $\beta$ is an upper bound as in Theorem \ref{main_theorem} for $d=2$. Now we can proceed in the same way as in the proof of Theorem \ref{main_theorem_1}
and the proof is completed.
\end{proof}
Now we prove Theorem \ref{ldrm-intro}.
\begin{proof}
The fact that $\mu[\mathcal{P}] \leq \kappa$ comes directly from the definition of the coherence number and the sparse setting of semi-gaussian matrices.
To see that, note that any given column $col$ of any matrix $\vv{P}_{i}$ in the related $\mathcal{P}$-model has a nonzero dot-product with at most $\kappa^{2}$ other columns
of any matrix $\vv{P}_{j}$. This in turn is implied by the fact that different columns are obtained by applying skew-circulant shifts blockwise, thus the number of
columns from $\vv{P}_{j}$ that have nonzero dot product with $col$ is at most the product of the number of nonzero dimensions of $col$ and $\vv{P}_{j}$. This is clearly upper
bounded by $\kappa^{2}$. This leads to the upper bound on the coherence $\mu[\mathcal{P}]$.
The new formula for $p_{wrong}$ is derived by a similar analysis to the one used to obtain the formula on $p_{wrong}$ in the proof of Theorem \ref{main_theorem_1}.
This time random variables under analysis are not independent though, but using
the same trick as the one we used in the proof of Theorem \ref{main_theorem_1}
to decouple dependent random variables in the sum to be estimated and applying Azuma's inequality (we omit details since the analysis is exactly the same as
in the aforementioned proof),
we obtain the following: $\mathbb{P}[|\vv{P}_{i,n_{1}}^{T} \vv{P}_{j,n_{1}}| >c] \leq e^{-\Omega (rc^{2})}$ for $i \neq j$ and any constant $c>0$. Taking the union bound
over all the pairs of columns and fixing $c=\frac{1}{\log^{2}(n)}$ and $r=3\log^{5}(n)$, we can conclude that with probability at
least $1-o(\frac{1}{n})$ the absolute value of the expression $\lambda(i,j)$
from the proof of Theorem \ref{main_theorem_1} is of the order
$o(\frac{n}{\log^{2}(n)})$. That enables us to finish tha analysis in the same way as in the proof of Theorem \ref{main_theorem_1} and derive similar conclusions.
The bound regarding the chromatic number is implied by the observation that each coherence graph in the corresponding $\mathcal{P}$-model has degree at most $\kappa^{2}$.
That follows directly from the observation we used to prove the upper bound on $\mu[\mathcal{P}]$. But now we can use Lemma \ref{graph_lemma} and that completes the proof
of Theorem \ref{ldrm-intro}.
\end{proof}
Below we present the proof of Theorem \ref{ldrm-extra}.
\begin{proof}
Fix two columns $\vv{P}_{i,n_{1}}$ and $\vv{P}_{j,n_{2}}$ and consider the expression $\vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}$.
We have already mentioned in the previous proof the right approach to finding strong upper bound on $|\vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}|$.
We first note that $\vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}$ can be written as a sum $w_{1} + ... + w_{nr}$, where $w_{i}s$ are not necessarily independent but can be
partitioned into at most three sets such that wariables in each of these sets are independent. This is true since $G^{i}_{struct}$ is produced by skew-circulant shifts
and the corresponding coherence graphs has verrtices of degree at most $2$.
Note also that each $w_{k}$ satisfies: $|w_{k}| \leq \frac{1}{\alpha r}$.
In each of the sum we get rid of these $w_{i}s$ that are equal to $0$.
Then, by applying Azuma's inequality independently on each of these subsets
and taking union bound over these subsets, we conclude that for any $a>0$:
\begin{equation}
|\vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}} > a| \leq 3e^{-\frac{a^{2}\alpha r}{O(1)}}
\end{equation}
Now we can take the union bound over all pairs of columns and notice that for every columcn $col$ in $\vv{P}_{i}$ and any $\vv{P}_{j}$ there exists at most $\kappa$
columns in $\vv{P}_{j}$ that have nonzero dot product with $col$. We can then take $a = \frac{\tau}{\kappa}$ and the proof is completed.
\end{proof}
Let us now switch to dense semi-gaussian matrices.
The following is true.
\begin{theorem}
\label{random_theorem}
Consider the setting as in Theorem \ref{main_theorem_1}.
Assume that entries of any fixed column of $P_{i}$ are chosen independently at random.
Assume also that for any $1 \leq i \leq j \leq m$ and any fixed column $col$ of $P_{i}$
each column of $P_{j}$ is a downward shift of $col$ by $b$ entries (possibly with signs of dimensions swapped) and that $b=0$ for $O(1)$ columns in $P_{j}$.
Then for and $T >0$ and $n$ large enough the following holds:
\begin{align}
\begin{split}
|\mathbb{E}[\tilde{k}_{\mathcal{P}}^{d}(\textbf{x},\textbf{z})]-\mathbb{E}[\tilde{k}_{\vv{G}}^{d}(\textbf{x},\textbf{z})]| \leq
O(\Delta),
\end{split}
\end{align}
where
$\Delta = p_{gen}(T) + p_{struct}(T) + d\epsilon + e^{-n^{\frac{1}{3}}}$
and
$$\epsilon = \frac{\log^{3}(n)}{n}\left(n^{\frac{2}{3}} + \max_{1 \leq i \leq j \leq m} |\sum_{1 \leq n_{1} < n_{2} \leq n} \vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}|\right).$$
As a corollary:
\begin{align}
\begin{split}
|Var(\tilde{k}_{\mathcal{P}}(\textbf{x},\textbf{z})) -Var(\tilde{k}_{\vv{G}}(\textbf{x},\textbf{z}))|=
O(\frac{m-1}{2k}\Delta).
\end{split}
\end{align}
\end{theorem}
\begin{proof}
The proof of this result follows along the lines of the proof of Theorem \ref{main_theorem_1} and Theorem \ref{main_variance_theorem}.
Take the formulas for $s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}$ derived in the proof of Theorem \ref{main_theorem}. Note that we want to have:
$|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| \leq \frac{\epsilon}{\Gamma(2d)\|g^{\mathcal{H}}\|_{2}}$, where $\Gamma$ is a constant that depends only on
the degree $d$. Each $s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}$ is a sum of random variables that can be decoupled into $O(1)$ subsums such that variables
in each subsum are independent (here we use exactly the same trick as in the proof of Theorem \ref{main_theorem_1}). In each subsum we apply Azuma's inequality.
Straightforward computations lead to the conclusion that if one sets up $\epsilon$ as in the statement of Theorem \ref{random_theorem} then the probability that
there exist different $s^{i_{1},j_{1}}$, $s^{i_{2},j_{2}}$ such that $|s^{i_{1},j_{1}} \cdot s^{i_{2},j_{2}}| > \frac{\epsilon}{\Gamma(2d)\|g^{\mathcal{H}}\|_{2}}$
is of the order $e^{-n^{\frac{1}{3}}}$ for $n$ large enough. That is the extra term in the formula for $\Delta$ that was not present in the staement of Theorem \ref{main_theorem_1}.
The variance results follows immediately by exactly the same analysis as in the proof of Theorem \ref{main_variance_theorem}.
\end{proof}
Note that introduced dense semi-gaussian matrices trivially satisfy conditions of Theorem \ref{random_theorem}
(look for the description of matrices $\vv{P}_{i}$ from Subsection: \ref{subsec:semi}).
The role of rank is similar as in the sparse setting, i.e. larger values of $r$ lead to sharper concentration results.
Theorem \ref{random_theorem} can be applied to classes of matrices for which $|\sum_{1 \leq n_{1} < n_{2} \leq n} \vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}|$
is small and random dense semi-gaussian matrices satisfy this condition with high probability.
\section{Empirical Support}
In this section, we compare feature maps obtained with fully Gaussian,
Fastfood, Circulant, and Toeplitz-like matrices with increasing displacement rank.
Our goal is to lend support to the theoretical contributions of this paper by showing that high-quality
feature maps can be constructed from a broad class of structured matrices as instantiations of the proposed $\mathcal{P}$-model.
{\bf Kernel Approximation Quality}: In Figure~\ref{fig:g50c}, we report relative Frobenius error in reconstructing the Gram matrix, i.e. $\frac{\|\vv{K} - \vv{\tilde{K}}\|_{fro}}{\|\vv{K}\|_{fro}}$ where $\vv{K}, \vv{\tilde{K}}$ denote the exact and approximate Gram matrices, as a function of the number of random features. We use the g50c dataset which comprises of $550$ examples drawn from multivariate Gaussians in $50$-dimensional space with means separated such that the Bayes error is $5\%$. We see that Circulant matrices and Toeplitz-like matrices with very low displacement rank (1 or 2) perform as well as Fastfood feature maps. In all experiments, for Toeplitz-like matrices, we used skew-Circulant parameters (the $\vv{h}$ vectors in Eqn.~\ref{eq:circ_skewcirc}) with average sparsity of $5$. As the displacement rank is increased, the budget of randomness increases and the reconstruction error approaches that of Gaussian Random features, as expected based on our theoretical results.
\begin{figure}[h]
\begin{center}
\includegraphics[height=4cm,width=0.8\linewidth]{g50c-crop.pdf}\\\label{fig:g50c}
\end{center}
\caption{Lower blue curves (better reconstruction) correspond to Toeplitz-like matrices with increasing displacement rank.}
\end{figure}
\begin{table*}[t]
\begin{center}
\caption{Kernel approximation (first row) and classification error (second row) in percentage for Complex Exponential (Gaussian Kernel).}\label{tab:complexexp}
{\scriptsize
\begin{tabular}{cccc|cccccc}
& & Gaussian & QMC (Halton) & Fastfood & Circulant & ToeplitzLike(1) & ToeplitzLike(5) & ToeplitzLike(10) & ToeplitzLike(20)\\
\hline
& \multirow{2}{*}{USPS (k=256)} & 5.06 & 5.05 & 6.76 & 7.61 & 9.66 & 7.55 & 6.86 & {\bf 6.68}\\
& & 7.12 & 6.90 & 7.37 & 7.54 & 7.72 & 7.44 & 7.46 & {\bf 7.29}\\
& \multirow{2}{*}{USPS (k=1280)} & 2.32& 2.15 & 3.06 & 3.32 & 4.41& 3.35& 3.16& {\bf 3.00}\\
& & 4.52& 4.73& 4.62& 4.53& 4.62& 4.58 & {\bf 4.53} & 4.65\\
\hline
&\multirow{2}{*}{DNA (k=80)} & 3.6 & 3.51 & 5.01 & 4.62 & 6.26 & 4.65 & 4.40 & {\bf 4.10} \\
& &31.04 & 30.94 & 31.04 & 30.94 & 31.35 & 30.82 &{\bf 30.29} & 30.70 \\
&\multirow{2}{*}{DNA ($k=900$)} & 1.61 & 1.59 & 2.23 & 2.06 & 2.88& 2.09 & 1.93 & {\bf 1.83} \\
& &16.5 & 15.01 & 16.94 & 16.63 & 16.82 & {\bf 16.34} & 16.57 & 16.57\\
\hline
&\multirow{2}{*}{COIL ($k=1024$)} & 2.74 & 2.41 & {\bf 3.67} & 4.45 & 5.60 & 4.47 & 4.09 & 3.79\\
& & 0.52 & 1.11 & {\bf 0.49} & 0.62 & 0.62 & 0.48 & 0.57 & 0.52\\
& \multirow{2}{*}{COIL ($k=2048$)} & 1.92 & 1.87 & {\bf 2.64} & 3.14 & 4.18 & 3.04 & 2.87 & 2.76\\
& & 0.17 & 0.28 & {\bf 0.15} & 0.19 & 0.19 & 0.20 & 0.19 & 0.19\\
\hline
\end{tabular}
}\end{center}
\end{table*}
Results on publicly available real-world classification datasets, averaged over $100$ runs, are reported in Table~\ref{tab:complexexp} for complex exponential nonlinearity (Gaussian kernel). Results with ReLU (arc-cosine) are similar but not shown for lack of space. As observed in previous papers, better Gram matrix approximation is not often correlated with higher classification accuracy. Nonetheless, it is clear that the design of space of valid feature map constructions based on structured matrices is much larger than what has so far been explored in the literature: Circulant and Toeplitz-like matrices are very competitive with Fastfood, and sometimes give better results particularly with increasing displacement rank. The effectiveness of such feature maps for nonlinearities other than the complex exponential also validates our theoretical contributions. Among the unstructured baselines, we also include Quasi-Monte Carlo (QMC) feature maps of~\cite{qmc} using Halton low-discrepancy sequences. The use of structured matrices to accelerate QMC techniques building on~\cite{DickFastQMC} is of interest for future work.
\iffalse
\begin{table*}
\caption{Kernel approximation error (first row) and classification error (second row) in percentage for ReLU (Arc-cosine ($p=1$) kernel).}\label{tab:relu}
\begin{center}
{\scriptsize
\begin{tabular}{cccc|c|ccccc}
& & Gaussian & QMC (Halton) & Fastfood & Circulant & ToeplitzLike(1) & ToeplitzLike(5) & ToeplitzLike(10) & ToeplitzLike(20)\\
\hline
& \multirow{2}{*}{USPS ($k=1280$)} & 6.14 & 4.52 & {\bf 7.33} & 7.64 & 8.58 & 7.81 & 7.75 & 7.37\\
&& 6.2 & 7.30 & 6.41 & 6.52 & 6.27 & 6.27& 6.49 & {\bf 6.29}\\
\hline
&\multirow{2}{*}{DNA ($k=180$)} & 19.43 & 16.89 & 23.44 & 20.55 & 24.4 & 22.29 & 21.48 & {\bf 20.28}\\
& &16.11 & 14.84 & 16.96 & {\bf 14.65} & 16.68 & 15.78 & 15.61 & 15.68 \\
\hline
&\multirow{2}{*}{COIL ($k=1024$)} & 4.72 & 2.07 & 5.68 & 5.05 & 7.50 & 5.44 & 5.07& {\bf 4.83} \\
& & 0.11 & 0.0 & {\bf 0.07} & 0.11 & 0.1 & 0.13 & 0.09 & 0.08\\\hline
\end{tabular}
}\end{center}
\end{table*}
\fi
\begin{figure}[h]
\begin{center}
\caption{Lower blue curves (smaller speedup) correspond to Toeplitz-like matrices with increasing displacement rank.}
\includegraphics[height=4.3cm, width=0.8\linewidth]{icml16_speedup-crop.pdf}\label{fig:speedup}
\end{center}
\end{figure}
{\bf Speedups}: Figure~\ref{fig:speedup} shows the speedup obtained in featuremap construction time using structured matrices relative to
using unstructured Gaussian random matrices (on a 6-core 32-GB Intel(R) Xeon(R) machine running Matlab R2014a). The benefits of
sub-quadratic matrix-vector multiplication with FFT-variations tend to show up beyond $1024$ dimensions. Circulant-based feature
maps are the fastest to compute. Fastfood (with DCT instead of Hadamard matrices) is about as fast as Toeplitz-like matrices with
displacement rank 1 or 2. Higher displacement rank matrices show speedups at higher dimensions as expected. Fastfood with inbuilt
\texttt{fwht} routine in Matlab performed poorly in our experiments.
\section{Introduction}
\label{sec:intro}
\input{introduction.tex}
\input{theory.tex}
\input{experiments.tex}
\section{Conclusions}
\label{sec:conclusions}
We have theoretically justified and empirically validated the use of a broad family of structured matrices
for accelerating the construction of random embeddings for approximating various kernel functions.
In particular, the class of Toeplitz-like semi-Gaussian matrices allows our construction to span highly compact to fully random matrices.
\section{Background and Preliminaries}
\subsection{Random Embeddings, Nonlinearities and Kernels}
Random feature maps may be viewed as arising from Monte-Carlo approximations to integral representations of kernel functions. The original construction by ~\citet{RahimiRecht} was motivated by a classical result that characterizes the class of shift-invariant positive definite functions.
\begin{theorem}[Bochner's Theorem~\cite{bochner1933}] A continuous shift-invariant kernel
function $k(\vv{x},\vv{z})\equiv \phi(\vv{x}-\vv{z})$ on $\mathbb{R}^n$ is positive definite if and
only if it is the Fourier transform of a unique finite non-negative measure on
$\mathbb{R}^n$.
That is, for any $\vv{x},\vv{z} \in \mathbb{R}^d $,
\begin{equation}
k(\vv{x},\vv{z}) = \int_{\mathbb{R}^n} e^{-i (\vv{x}-\vv{z})^T\vv{w}} p(\vv{w}) d\vv{w} =
\mathbb{E}_{\vv{w}\sim p}[ e^{-i (\vv{x}-\vv{z})^T\vv{w}}]~.\nonumber\label{eq:bochner}
\end{equation}
\end{theorem} \label{thm:bochner_R_d}
Without loss of generality, the non-negative measure above can be assumed to be
a probability measure with the associated density $p$. Bochner's theorem
establishes one-to-one correspondence between shift-invariant kernel functions
and probability densities on $\mathbb{R}^n$, via the Fourier transform. A Monte Carlo approximation to the integral above leads to the following,
\begin{equation}
k(\vv{x}, \vv{z}) \approx \frac{1}{k}\sum_{j=1}^k e^{-i \vv{x}^T \vv{w}_j} e^{i \vv{z}^T \vv{w}_j} =
\Psi(\vv{x})^T \Psi(\vv{z})
\end{equation} where the feature map $\Psi(\vv{x})$ has the form given in Eqn.~\ref{eq:featuremap}, with rows of $\vv{M}$ drawn from the density $p$. For the Gaussian kernel with bandwidth $\sigma$, $k(\vv{x}, \vv{z}) = e^{\frac{-\|\vv{x} - \vv{z}\|}{2\sigma^2}}$, the associated density $p$ is also Gaussian, with $\sigma^{-2}$ times the identity as the covariance matrix.
Equivalently, entries of $\vv{M}$ may be drawn i.i.d from a standard Gaussian distribution, if the complex exponential $s(x) = e^{i \frac{x}{\sigma}}$ is used as the scalar nonlinearity.
While studying synergies between kernel methods and deep learning~\cite{cho} introduce $b^{th}$-order {\it arc-cosine} kernels via the following integral representation:
$$
k_{b}(\vv{x}, \vv{z}) = \int_{\mathbb{R}^d} i(\vv{w}^T \vv{x}) i(\vv{w}^T \vv{z}) (\vv{w}^T\vv{x})^b (\vv{w}^T \vv{z})^b \ p(\vv{w}) d\vv{w}
$$
where $i(\cdot)$ is the step function, i.e. $i(x) = 1$ if $x>0$ and $0$ otherwise; and the density $p$ is chosen to be standard Gaussian.
These kernels evaluate inner products in the representation induced by a infinitely wide single hidden layer neural network with random
Gaussian weights: for $b=0$, the activation function is hard thresholding, while $=1$ corresponds to using rectified linear units (ReLU).
These kernels admit closed form expressions, in terms of the angle $\theta = cos^{-1} (\frac{\vv{x}^T\vv{z}}{\|\vv{x}\|_2\|\vv{z}\|_2})$ between $\vv{x}$ and $\vv{z}$:
\begin{eqnarray}
k_0(\vv{x}, \vv{z}) &=& 1 - \frac{\theta}{\pi}\label{eq:stepkernel}\\
k_1(\vv{x}, \vv{z}) &=& \frac{\|\vv{x}\|_2 \|\vv{z}\|_2}{\pi} [sin(\theta) + (\pi - \theta) cos(\theta)] \label{eq:relukernel}
\end{eqnarray} where $\| \cdot\|_2$ denotes $l_2$ norm. Therefore, feature maps of the form in Eqn.~\ref{eq:featuremap} with hard-thresholding ($s(x) = i(x)$) correspond to the angular similarity kernel in Eqn.~\ref{eq:stepkernel}, while ReLU activation $s(x) = max(x, 0)$ leads to the trigonometric kernel in Eqn.~\ref{eq:relukernel}.
\subsection{Structured Matrices}
\begin{figure*}[t]
\begin{minipage}{.3\linewidth}
~~~~~~~~~~~~~~~~(i) \textit{f-Circulant}
\begin{scriptsize}
\begin{eqnarray}
\left[\begin{array}{cccc}
{\bf \color{red}{g_0}} & f {\bf g_{n-1}} & \ldots & f {\bf \color{blue} g_1}\\
{\bf \color{blue} g_1}& {\bf \color{red} g_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & f{\bf g_{n-1}} \\
{\bf g_{n-1}} & \ldots & {\bf \color{blue} g_1} & {\bf \color{red}g_0}
\end{array}
\right] \nonumber
\end{eqnarray}
\end{scriptsize}
\end{minipage}\hspace{-0.75cm}
\begin{minipage}{.3\linewidth}
~~~~~~~~~~~~~~~~(ii) \textit{Toeplitz}
\begin{scriptsize}
\begin{eqnarray}
\left[\begin{array}{cccc}
{\bf \color{red}{t_0}} & {\bf \color{green} t_{-1}} & \ldots & {\bf \color{cyan}{t_{-(n-1)}}}\\
{\bf \color{blue} t_1}& {\bf \color{red} t_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & {\bf \color{green} t_{-1}} \\
{\bf t_{n-1}} & \ldots & {\bf \color{blue} t_1} & {\bf \color{red}t_0}
\end{array}
\right] \nonumber
\end{eqnarray}
\end{scriptsize}
\end{minipage}\hspace{-0.75cm}
\begin{minipage}{.5\linewidth}
~~~~~~~~~~~~~~~~(iii) \textit{Displacement rank r}
\begin{scriptsize}
\begin{eqnarray}
\sum_{i=1}^r \left[\begin{array}{cccc}
{\bf \color{red}{g^i_0}} & {\bf g^i_{n-1}} & \ldots & {\bf \color{blue} g^i_1}\\
{\bf \color{blue} g^i_1}& {\bf \color{red} g^i_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & {\bf g^i_{n-1}} \\
{\bf g^i_{n-1}} & \ldots & {\bf \color{blue} g^i_1} & {\bf \color{red}g^i_0}
\end{array}
\right]
\left[\begin{array}{cccc}
{\bf \color{red}{h^i_0}} & -{\bf h^i_{n-1}} & \ldots & -{\bf \color{blue} h^i_1}\\
{\bf \color{blue} h^i_1}& {\bf \color{red} h^i_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & -{\bf h^i_{n-1}} \\
{\bf h^i_{n-1}} & \ldots & {\bf \color{blue} h^i_1} & {\bf \color{red}h^i_0}
\end{array}
\right]
\nonumber
\end{eqnarray}
\end{scriptsize}
\end{minipage}\\
\caption{Structured Matrices}\label{fig:structured}
\end{figure*}
A $m\times n$ matrix is called a structured matrix is it satisfies the following two properties: (1) it has much fewer degrees of freedom than
$mn$ independent entries, and hence can be implicitly stored more efficiently than general matrices, and (2) the structure in the matrix can
be exploited for fast linear algebra operations such as fast matrix-vector multiplication, factorizations and solution to linear systems.
Examples include the Discrete Fourier Transform matrix (DFT) and the Discrete Cosine Transform matrix (DCT). Here, we give other examples particularly relevant to this paper.
{\bf Hadamard Matrices:} These matrices are defined for $m = n = 2^k$ recursively as follows:
\begin{equation}
\vv{H}_0 = 1,~~~~~~~~~~~\vv{H}_k = \frac{1}{\sqrt{2}} \left(\begin{array}{cc} \vv{H}_{k-1} & \vv{H}_{k-1}\\\vv{H}_{k-1} & -\vv{H}_{k-1} \end{array}\right)
\end{equation} Note that $\frac{1}{\sqrt{n}}\vv{H}$ is an orthonormal matrix.
Hadamard matrices support $O(n~\log n)$ matrix-vector multiplication time via the Fast Hadamard transform, a variant of FFT.
\def\textrm{krylov}{\textrm{krylov}}
{\bf Circulant Matrices}: These matrices are intimately associated with circular
convolutions and have been used for fast compressed sensing in~\cite{CirculantCS}.
A $n\times n$ Circulant matrix is completely determined by its first column/row, i.e., $n$ parameters. Each column/row of a Circulant
matrix is generated by cyclically down/right-shifting the previous column/row. For righ-shifting one can define a Circulant matrix from $\mathbb{R}^{m \times n}$
for $m \leq n$ by taking first $m$ rows of the matrix described above.
An elementary matrix that implements a basic {\it downward shift-and-scale} transformation involving a scalar $f$ is called an $f$-unit-Circulant matrix,
and is given below.
\begin{definition}[$f$-unit-Circulant Matrix]
For a real-valued scalar $f$, the $(n\times n)$ f-unit-Circulant matrix, denoted by $\vv{Z}_f$, is defined as follows,
\begin{eqnarray}
\vv{Z}_f = \left[ \begin{array}{cc} 0_{1\times (n-1)} & {\color{red}f}\\ \vv{I}_{n-1} & 0_{(n-1)\times 1}\end{array}\right]\nonumber
\label{def:unitcirc}
\end{eqnarray}
\end{definition} where the notation $\vv{I}_n$ denotes a $n\times n$ identity matrix.
The matrix-vector product $\vv{Z}_f \vv{v}$ shifts the elements of the column vector $\vv{v}$ ``downwards", scales by $f$ and brings the last element $v_n$ to the ``top", resulting in $[{\color{red} f} v_n, v_1, \ldots v_{n-1}]^T$. Define the matrix
\textrm{krylov}$(\vv{A}, \vv{v})$ by:
\begin{equation}
\textrm{krylov}(\vv{A}, \vv{v}) = [\vv{v}~~\vv{A}\vv{v}~~\vv{A}^2 \vv{v} \ldots \vv{A}^{n-1} \vv{v}]\label{def:krylov}
\end{equation}
Then it is easy to see that a Circulant matrix whose first column is $\vv{g}$ is simply $\textrm{krylov}(\vv{Z}_1, \vv{g})$.
We use the notation $\vv{Z}_1[\vv{g}]$ for a Circulant matrix. For $f=-1$, the matrix $\textrm{krylov}(\vv{Z}_{-1}, \vv{g})$ is called skew-Circulant. For a general scalar $f$, the structure of $\vv{Z}_f[\vv{v}] = \textrm{krylov}(\vv{Z}_{f}, \vv{g})$ referred to as an $f$-Circulant matrix is shown in Figure~\ref{fig:structured} (i).
The class of $f$-circulant matrices admit $O(n~\log~n)$ matrix-vector multiplication as they are diagonalized by the DFT matrix~\cite{Pan}.
\iffalse
:
\begin{theorem}[Diagonalization of $f$-circulant matrices]
For any $f\neq 0$, let $${\bf f} = [1, f^\frac{1}{d}, f^\frac{2}{d}, \ldots f^{\frac{d-1}{d}}]^T \in \mathbb{C}^d,~~\textrm{and}~~\vv{D}_{\vv{f}} = diag(\vv{f})$$ and let $\Omega_d$ denote the matrix of the Fast Fourier Transform, i.e., $[\Omega_d]_{ij} = {\omega_d}^{ij}$ where $\omega_d = \exp{\left(\frac{i 2\pi}{d}\right)}$ denotes the $d$-th root of unity. Then,
\begin{equation}
\vv{Z}_f[\vv{v}] = \vv{D}_{\vv{f}}^{-1}~\Omega_d^{-1}~diag(\Omega_d (\vv{f} \circ\vv{v})) \Omega_d \vv{D}_{\vv{f}}
\end{equation}
\end{theorem}\label{thm:f_circulant_diagonalization}
\fi
{\bf Toeplitz and Hankel Matrices:} These matrices implement discrete linear convolution and arise naturally in dynamical systems and time series analysis. Toeplitz matrices are characterized by constant diagonals (Fig~\ref{fig:structured}(ii)), while closely related Hankel matrices have constant anti-diagonals. Matrix-vector multiplication for these matrices taken from $\mathbb{R}^{m \times n}$ can be reduced to $O(n~\log~m)$ Circulant matrix-vector multiplication. For detailed properties of Circulant and Toeplitz matrices, we point the reader to~\cite{GrayCirculantToeplitz}
{\bf Structured Matrices with Low-displacement Rank}: The notion of displacement operators and displacement rank~\cite{Golub, Pan, KKM79} can be used to broadly generalize various classes of structured matrices. For example, under the action of the {\it Sylvester displacement operator} defined as $$L[\vv{T}] = \vv{Z}_1 \vv{T} - \vv{T} \vv{Z}_{-1},$$ every Toeplitz matrix can be transformed into a matrix of rank at most $2$ using shift and scale operations. For a given displacement rank parameter $r$, the class of matrices for which the rank of $L[\vv{T}]$ is at most $r$ is called {\it Toeplitz-like}, and in fact admits a closed-form parameterization:
\begin{theorem}[Parameterization of Toeplitz-like matrices with displacement rank r] (Pan, 2003): If a $n\times n$ matrix $\vv{T}$ satisfies $rank(\vv{Z}_1 \vv{T} - \vv{T} \vv{Z}_{-1}) \leq r$ then $\vv{T}$ can be written as,
\begin{equation}
\vv{T} = \sum_{i=1}^r \vv{Z}_1[\vv{g}^i] \vv{Z}_{-1}[\vv{h}^i] \label{eq:circ_skewcirc}
\end{equation}
\end{theorem}
Note that the notation $\vv{Z}_f$ refers to $f$-unit-Circulant matrix (Definition~\ref{def:unitcirc}) while $\vv{Z}_f[\vv{g}] = \textrm{krylov}(\vv{Z}_f, \vv{g})$ (see Eqn.~\ref{def:krylov}).
Instead of $T \in \mathbb{R}^{n \times n}$ defined above one can also take its first $m$ rows ending up with a structured matrix from $\mathbb{R}^{m \times n}$.
This family of matrices expressible by Eqn.~\ref{eq:circ_skewcirc} for a suitable choice of vectors $\{\vv{g}^i, \vv{h}^i\}_{i=1}^r$, is very rich, i.e., it covers (i)
all Circulant and Skew-circulant matrices for $r=1$, (ii) all Toeplitz matrices and their inverses for $r=2$, (iii) Products, inverses, linear combinations of distinct
Toeplitz matrices with increasing $r$, and (iv) all $n\times n$ matrices for $r=n$.
Since Toeplitz-like matrices under the parameterization of Eqn.~\ref{eq:circ_skewcirc}
are a sum of products between Circulant and Skew-circulant matrices, their fast FFT based matrix-vector multiplication is of order $O(nr log~m)$,
where $r$ is the displacement rank. Hence, $r$ provides a knob with which to control the storage requirements,
computational constraints and statistical capacity of Toeplitz-like structured matrices. Recently such matrices were used in the context of
learning mobile-friendly neural networks in~\cite{StructuredTransforms}. We note in passing that the displacement rank framework generalizes
to other types of base structures (e.g. Vandermonde); see~\cite{Pan}.
\subsection{FastFood}
In the context of fast kernel approximations,~\cite{LSS13} introduce the Fastfood technique where the matrix $\vv{M}$ in Eqn.~\ref{eq:featuremap} is parameterized by a
product of diagonal and simple matrices as follows:
\begin{equation}
\vv{F} = \frac{1}{\sqrt{d}}\vv{S}\vv{H}\vv{G} \vv{P} \vv{H} \vv{B}.
\end{equation} Here, $\vv{S}, \vv{G}, \vv{B}$ are diagonal random matrices, $\vv{P}$ is a permutation matrix and $\vv{H}$ is the Walsh-Hadamard matrix.
The $k\times n$ matrix $\vv{M}$ is obtained by vertically stacking $k/n$ independent copies of $\vv{F}$. Multiplication
against such a matrix can be performed in time $O(k\log~n)$. The authors prove that (1) the Fastfood approximation is
unbiased, (2) its variance is at most the variance of standard Gaussian random features with an additional $O(\frac{1}{k})$
term, and (3) for a given error probability $\delta$, the pointwise approximation error of a $n\times n$ block of Fastfood is
at most $O(\sqrt{\log(n/\delta)})$ larger than that of standard Gaussian random features. However, note that the Fastfood analysis
is limited to the Gaussian kernel and their variance bound uses properties of the complex exponential. The authors also conjecture
that the Hadamard matrix $\vv{H}$ above, can be replaced by any matrix $\vv{T}$ such that $\vv{T}/\sqrt{n}$ is orthonormal, the
maximum entry in $\vv{T}$ is small, and matrix-vector product against $\vv{T}$ can be computed in $O(n\log~n)$ time.
\section{Background and Preliminaries}
We start by giving a brisk background on random feature maps and structured matrices.
\subsection{Random Embeddings, Nonlinearities and Kernels}
Random feature maps may be viewed as arising from Monte-Carlo approximations to integral representations of kernel functions. The original construction by ~\citet{RR07} was motivated by a classical result that characterizes the class of shift-invariant positive definite functions.
\begin{theorem}[Bochner's Theorem~\cite{bochner1933}] A continuous shift-invariant scaled kernel
function ${\cal K}(\vv{x},\vv{z})\equiv \phi(\vv{x}-\vv{z})$ on $\mathbb{R}^n$ is positive definite if and
only if it is the Fourier transform of a unique finite probability measure $p$ on
$\mathbb{R}^n$.
That is, for any $\vv{x},\vv{z} \in \mathbb{R}^d $,
\begin{equation}
{\cal K}(\vv{x},\vv{z}) = \int_{\mathbb{R}^n} e^{-i (\vv{x}-\vv{z})^T\vv{w}} p(\vv{w}) d\vv{w} =
\mathbb{E}_{\vv{w}\sim p}[ e^{-i (\vv{x}-\vv{z})^T\vv{w}}]~.\nonumber\label{eq:bochner}
\end{equation}
\end{theorem} \label{thm:bochner_R_d}
Bochner's theorem stablishes one-to-one correspondence between shift-invariant kernel functions and probability densities on $\mathbb{R}^n$, via the Fourier transform. In the case of the Gaussian kernel with bandwidth $\sigma$, the associated density is also Gaussian with covariance matrix $\sigma^{-2}$ times the identity.
While studying synergies between kernel methods and deep learning,~\cite{cho} introduce $b^{th}$-order {\it arc-cosine} kernels via the following integral representation:
$$
{{\cal K}}_{b}(\vv{x}, \vv{z}) = \int_{\mathbb{R}^d} i(\vv{w}^T \vv{x}) i(\vv{w}^T \vv{z}) (\vv{w}^T\vv{x})^b (\vv{w}^T \vv{z})^b \ p(\vv{w}) d\vv{w}
$$
where $i(\cdot)$ is the step function, i.e. $i(x) = 1$ if $x>0$ and $0$ otherwise; and the density $p$ is chosen to be standard Gaussian.
These kernels evaluate inner products in the representation induced by an infinitely wide single hidden layer neural network with random
Gaussian weights, and admit closed form expressions in terms of the angle $\theta = cos^{-1} (\frac{\vv{x}^T\vv{z}}{\|\vv{x}\|_2\|\vv{z}\|_2})$ between $\vv{x}$ and $\vv{z}$:
\begin{eqnarray}
{\cal K}_{0}(\vv{x}, \vv{z}) &=& 1 - \frac{\theta}{\pi}\label{eq:stepkernel}\\
{\cal K}_{1}(\vv{x}, \vv{z}) &=& \frac{\|\vv{x}\|_2 \|\vv{z}\|_2}{\pi} [sin(\theta) + (\pi - \theta) cos(\theta)] \label{eq:relukernel}
\end{eqnarray} where $\| \cdot\|_2$ denotes $l_2$ norm.
Monte Carlo approximations to the integral representations above lead to the following,
\begin{equation}
{\cal K}(\vv{x}, \vv{z}) \approx \frac{1}{k}\sum_{j=1}^k s(\vv{x}^T \vv{w}_j) s(\vv{z}^T \vv{w}_j) =
\Psi(\vv{x})^T \Psi(\vv{z})
\end{equation} where the feature map $\Psi(\vv{x})$ has the form given in Eqn.~\ref{eq:featuremap}, with rows of $\vv{M}$,
i.e. $\vv{w}_j$ vectors, drawn from the Gaussian density, and the nonlinearity $s$ set to the following: complex exponential, $s(x) = e^{i \frac{x}{\sigma}}$, for the Gaussian kernel with bandwidth $\sigma$; hard-thresholding, $s(x) = i(x)$, for the angular similarity kernel in Eqn.~\ref{eq:stepkernel}; and ReLU activation, $s(x) = \max(x, 0)$, for the first order arc-cosine kernel in Eqn.~\ref{eq:relukernel}.
\subsection{Structured Matrices}
\iffalse
\begin{figure*}[t]
\begin{minipage}{.3\linewidth}
~~~~~~~~~~~~~~~~(i) \textit{f-Circulant}
\begin{scriptsize}
\begin{eqnarray}
\left[\begin{array}{cccc}
{\bf \color{red}{g_0}} & f {\bf g_{n-1}} & \ldots & f {\bf \color{blue} g_1}\\
{\bf \color{blue} g_1}& {\bf \color{red} g_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & f{\bf g_{n-1}} \\
{\bf g_{n-1}} & \ldots & {\bf \color{blue} g_1} & {\bf \color{red}g_0}
\end{array}
\right] \nonumber
\end{eqnarray}
\end{scriptsize}
\end{minipage}\hspace{-0.75cm}
\begin{minipage}{.3\linewidth}
~~~~~~~~~~~~~~~~(ii) \textit{Toeplitz}
\begin{scriptsize}
\begin{eqnarray}
\left[\begin{array}{cccc}
{\bf \color{red}{t_0}} & {\bf \color{green} t_{-1}} & \ldots & {\bf \color{cyan}{t_{-(n-1)}}}\\
{\bf \color{blue} t_1}& {\bf \color{red} t_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & {\bf \color{green} t_{-1}} \\
{\bf t_{n-1}} & \ldots & {\bf \color{blue} t_1} & {\bf \color{red}t_0}
\end{array}
\right] \nonumber
\end{eqnarray}
\end{scriptsize}
\end{minipage}\hspace{-0.75cm}
\begin{minipage}{.5\linewidth}
~~~~~~~~~~~~~~~~(iii) \textit{Displacement rank r}
\begin{scriptsize}
\begin{eqnarray}
\sum_{i=1}^r \left[\begin{array}{cccc}
{\bf \color{red}{g^i_0}} & {\bf g^i_{n-1}} & \ldots & {\bf \color{blue} g^i_1}\\
{\bf \color{blue} g^i_1}& {\bf \color{red} g^i_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & {\bf g^i_{n-1}} \\
{\bf g^i_{n-1}} & \ldots & {\bf \color{blue} g^i_1} & {\bf \color{red}g^i_0}
\end{array}
\right]
\left[\begin{array}{cccc}
{\bf \color{red}{h^i_0}} & -{\bf h^i_{n-1}} & \ldots & -{\bf \color{blue} h^i_1}\\
{\bf \color{blue} h^i_1}& {\bf \color{red} h^i_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & -{\bf h^i_{n-1}} \\
{\bf h^i_{n-1}} & \ldots & {\bf \color{blue} h^i_1} & {\bf \color{red}h^i_0}
\end{array}
\right]
\nonumber
\end{eqnarray}
\end{scriptsize}
\end{minipage}\\
\vspace{-0.7cm} \caption{Structured Matrices: Parameter sharing is illustrated by color.}\label{fig:structured}
\end{figure*}
\fi
A $m\times n$ matrix is called a structured matrix if it satisfies the following two properties: (1) it has much fewer degrees of freedom than
$mn$ independent entries, and hence can be implicitly stored more efficiently than general matrices, and (2) the structure in the matrix can
be exploited for fast linear algebra operations such as fast matrix-vector multiplication. Examples include the Discrete Fourier Transform (DFT), the Discrete Cosine Transform (DCT) and the Walsh-Hadamard Transform (WHT) matrices. Here, we give other examples particularly relevant to this paper. The matrices described below are square. Rectangular matrices can be obtained by appropriately selecting rows or columns.
\iffalse
{\bf Hadamard Matrices:} These matrices are defined for $m = n = 2^k$ recursively as follows:
\begin{equation}
\vv{H}_0 = 1,~~~~~~~~~~~\vv{H}_k = \frac{1}{\sqrt{2}} \left(\begin{array}{cc} \vv{H}_{k-1} & \vv{H}_{k-1}\\\vv{H}_{k-1} & -\vv{H}_{k-1} \end{array}\right)
\end{equation} Note that $\frac{1}{\sqrt{n}}\vv{H}$ is an orthonormal matrix.
Hadamard matrices support $O(n~\log n)$ matrix-vector multiplication time via the Fast Hadamard transform, a variant of FFT.
\fi
\def\textrm{krylov}{\textrm{krylov}}
{\bf Circulant Matrices}: These matrices are intimately associated with circular
convolutions and have been used for fast compressed sensing in~\cite{CirculantCS}.
A $n\times n$ Circulant matrix is completely determined by its first column/row, i.e., $n$ parameters. Each column/row of a Circulant
matrix is generated by cyclically down/right-shifting the previous column/row. A {\it skew-Circulant} matrix has identical structure to Circulant, except that the upper triangular part of the matrix is negated. This general structure looks like,
{\small
\begin{eqnarray}
\left[\begin{array}{cccc}
{\bf \color{red}{g_0}} & f {\bf g_{n-1}} & \ldots & f {\bf \color{blue} g_1}\\
{\bf \color{blue} g_1}& {\bf \color{red} g_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & f{\bf g_{n-1}} \\
{\bf g_{n-1}} & \ldots & {\bf \color{blue} g_1} & {\bf \color{red}g_0}
\end{array}
\right] \nonumber
\end{eqnarray}}
with $f=1$ for Circulant and $f=-1$ for skew-Circulant matrix. Both these matrices admit $O(n~\log~n)$ matrix-vector multiplication as they are diagonalized by the DFT matrix~\cite{Pan}. We will use the notation $\texttt{circ}[\vv{g}]$ and $\texttt{scirc}[\vv{g}]$ for Circulant and skew-Circulant matrices respectively.
\iffalse
An elementary matrix that implements a basic {\it downward shift-and-scale} transformation involving a scalar $f$ is called an $f$-unit-Circulant matrix,
and is given below.
\begin{definition}[$f$-unit-Circulant Matrix]
For a real-valued scalar $f$, the $(n\times n)$ f-unit-Circulant matrix, denoted by $\vv{Z}_f$, is defined as follows,
\begin{eqnarray}
\vv{Z}_f = \left[ \begin{array}{cc} 0_{1\times (n-1)} & {\color{red}f}\\ \vv{I}_{n-1} & 0_{(n-1)\times 1}\end{array}\right]\nonumber
\label{def:unitcirc}
\end{eqnarray}
\end{definition} where the notation $\vv{I}_n$ denotes a $n\times n$ identity matrix.
The matrix-vector product $\vv{Z}_f \vv{v}$ shifts the elements of the column vector $\vv{v}$ ``downwards", scales by $f$ and brings the last element $v_n$ to the ``top", resulting in $[{\color{red} f} v_n, v_1, \ldots v_{n-1}]^T$. Define the matrix
\textrm{krylov}$(\vv{A}, \vv{v})$ by:
\begin{equation}
\textrm{krylov}(\vv{A}, \vv{v}) = [\vv{v}~~\vv{A}\vv{v}~~\vv{A}^2 \vv{v} \ldots \vv{A}^{n-1} \vv{v}]\label{def:krylov}
\end{equation}
Then it is easy to see that a Circulant matrix whose first column is $\vv{g}$ is simply $\textrm{krylov}(\vv{Z}_1, \vv{g})$.
We use the notation $\vv{Z}_1[\vv{g}]$ for a Circulant matrix. For $f=-1$, the matrix $\textrm{krylov}(\vv{Z}_{-1}, \vv{g})$ is called skew-Circulant. For a general scalar $f$, the structure of $\vv{Z}_f[\vv{v}] = \textrm{krylov}(\vv{Z}_{f}, \vv{g})$ referred to as an $f$-Circulant matrix is shown in Figure~\ref{fig:structured} (i).
\fi
\iffalse
:
\begin{theorem}[Diagonalization of $f$-circulant matrices]
For any $f\neq 0$, let $${\bf f} = [1, f^\frac{1}{d}, f^\frac{2}{d}, \ldots f^{\frac{d-1}{d}}]^T \in \mathbb{C}^d,~~\textrm{and}~~\vv{D}_{\vv{f}} = diag(\vv{f})$$ and let $\Omega_d$ denote the matrix of the Fast Fourier Transform, i.e., $[\Omega_d]_{ij} = {\omega_d}^{ij}$ where $\omega_d = \exp{\left(\frac{i 2\pi}{d}\right)}$ denotes the $d$-th root of unity. Then,
\begin{equation}
\vv{Z}_f[\vv{v}] = \vv{D}_{\vv{f}}^{-1}~\Omega_d^{-1}~diag(\Omega_d (\vv{f} \circ\vv{v})) \Omega_d \vv{D}_{\vv{f}}
\end{equation}
\end{theorem}\label{thm:f_circulant_diagonalization}
\fi
{\bf Toeplitz and Hankel Matrices:} These matrices implement discrete linear convolution and arise naturally in dynamical systems and time series analysis. Toeplitz matrices are characterized by constant diagonals as follows,
{\small \begin{eqnarray}
\left[\begin{array}{cccc}
{\bf \color{red}{t_0}} & {\bf \color{green} t_{-1}} & \ldots & {\bf \color{cyan}{t_{-(n-1)}}}\\
{\bf \color{blue} t_1}& {\bf \color{red} t_0} & \ldots & \vdots\\
\vdots & \vdots & \vdots & {\bf \color{green} t_{-1}} \\
{\bf t_{n-1}} & \ldots & {\bf \color{blue} t_1} & {\bf \color{red}t_0}
\end{array}
\right] \nonumber
\end{eqnarray}} Closely related Hankel matrices have constant anti-diagonals. Toeplitz-vector multiplication can be reduced to $O(n~\log~n)$ Circulant-vector multiplication. For detailed properties of Circulant and Toeplitz matrices, we point the reader to~\cite{GrayCirculantToeplitz}
{\bf Structured Matrices with Low-displacement Rank}: The notion of displacement operators and displacement rank~\cite{Golub, Pan, KKM79} can be used to broadly generalize various classes of structured matrices. For example, under the action of the {\it Sylvester displacement operator} defined as $L[\vv{T}] = \vv{Z}_1 \vv{T} - \vv{T} \vv{Z}_{-1}$, every Toeplitz matrix can be transformed into a matrix of rank at most $2$ using elementary shift and scale operations implemented by matrices of the form $\vv{Z}_f = [\vv{e}_2 \vv{e}_3\ldots \vv{e}_n~f \vv{e}_1]$ for $f=\pm 1$ where $\vv{e}_1\ldots \vv{e}_n$ are column vectors representing the standard basis of $\mathbb{R}^n$.
For a given displacement rank parameter $r$, the class of matrices for which the rank of $L[\vv{T}]$ is at most $r$ is called {\it Toeplitz-like}. Remarkably, this class of matrices admits a closed-form parameterization in terms of the low-rank factorization of $L[\vv{T}]$:
\begin{theorem}[Parameterization of Toeplitz-like matrices with displacement rank $r$~\cite{Pan}]: If an $n\times n$ matrix $\vv{T}$ satisfies $rank(\vv{Z}_1 \vv{T} - \vv{T} \vv{Z}_{-1}) \leq r$, then it can be written as,
\begin{equation}
\vv{T} = \sum_{i=1}^r \texttt{circ}[\vv{g}^i]~\texttt{scirc}[\vv{h}^i] \label{eq:circ_skewcirc}
\end{equation} for some choice of vectors $\{\vv{g}^i, \vv{h}^i\}_{i=1}^r \in \mathbb{R}^n$.
\end{theorem} The family of matrices expressible by Eqn.~\ref{eq:circ_skewcirc} is very rich~\cite{Pan}, i.e., it covers (i)
all Circulant and Skew-circulant matrices for $r=1$, (ii) all Toeplitz matrices and their inverses for $r=2$, (iii) Products, inverses, linear combinations of distinct
Toeplitz matrices with increasing $r$, and (iv) all $n\times n$ matrices for $r=n$.
Since Toeplitz-like matrices under the parameterization of Eqn.~\ref{eq:circ_skewcirc}
are a sum of products between Circulant and Skew-circulant matrices, they inherit fast FFT based matrix-vector multiplication with cost $O(nr log~n)$,
where $r$ is the displacement rank. Hence, $r$ provides a knob on the degree of structure imposed on the matrix with which storage requirements, computational constraints and statistical capacity can be explicitly controlled. Recently such matrices were used in the context of learning mobile-friendly neural networks in~\cite{StructuredTransforms}.
We note in passing that the displacement rank framework generalizes to other types of base structures (e.g. Vandermonde); see~\cite{Pan}.
\subsection{FastFood}
In the context of fast kernel approximations,~\cite{LSS13} introduce the Fastfood technique where the matrix $\vv{M}$ in Eqn.~\ref{eq:featuremap} is parameterized by a
product of diagonal and simple matrices as follows:
\begin{equation}
\vv{F} = \frac{1}{\sqrt{n}}\vv{S}\vv{H}\vv{G} \vv{P} \vv{H} \vv{B}.\label{eq:fastfood}
\end{equation} Here, $\vv{S}, \vv{G}, \vv{B}$ are diagonal random matrices, $\vv{P}$ is a permutation matrix and $\vv{H}$ is the Walsh-Hadamard matrix.
The $k\times n$ matrix $\vv{M}$ is obtained by vertically stacking $k/n$ independent copies of the $n\times n$ matrix $\vv{F}$. Multiplication against such a matrix can be performed in time $O(k\log~n)$. The authors prove that (1) the Fastfood approximation is
unbiased, (2) its variance is at most the variance of standard Gaussian random features with an additional $O(\frac{1}{k})$
term, and (3) for a given error probability $\delta$, the pointwise approximation error of a $n\times n$ block of Fastfood is
at most $O(\sqrt{\log(n/\delta)})$ larger than that of standard Gaussian random features. However, note that the Fastfood analysis
is limited to the Gaussian kernel and their variance bound uses properties of the complex exponential. The authors also conjecture
that the Hadamard matrix $\vv{H}$ above, can be replaced by any matrix $\vv{T}$ such that $\vv{T}/\sqrt{n}$ is orthonormal, the
maximum entry in $\vv{T}$ is small, and matrix-vector product against $\vv{T}$ can be computed in $O(n\log~n)$ time.
\section{Structured Matrices from Gaussian Vectors}
In this section, we present a general structured matrix model that allows a small Gaussian vector to be recycled in order to mimic the properties of a Gaussian random matrix suitable for generating random features. We first introduce some basic concepts in our construction. Note that we emphasize intuitions in our exposition - formal proofs are provided in our supplementary material.
\subsection{The $\mathcal{P}$-model}
\label{sec:p_model}
{\bf Budget of Randomness}: Let $t$ be some given parameter. Consider the column vector $\vv{g}=(g_{1},...,g_{t})^T$, where each entry is an independent Gaussian taken from $\mathcal{N}(0,1)$. This vector stands for the ``budget of randomness'' used in our structured matrix construction scheme.
Our goal is to recycle the Gaussian vector $\vv{g}$ to construct random matrices with desirable properties. This is accomplished using a sequence of matrices which we call the $\mathcal{P}$-model.
\begin{definition}[$\mathcal{P}$-model]
\label{def:p_model}
Given the budget of uncertainty parameter $t$, a sequence of
$m$ matrices with unit $l_{2}$ norm columns, denoted as $\mathcal{P} = \{\vv{P}_i\}_{i=1}^m$, where $\vv{P}_i \in \mathbb{R}^{t \times n}$, specifies a $\mathcal{P}$-model.
Such a sequence defines an $m\times n$ random matrix of the form:
\begin{equation}
\vv{S}[\mathcal{P}] = \left(\begin{array}{c}\vv{g}^T\vv{P}_1 \\\vv{g}^T\vv{P}_2\\ \vdots\\\vv{g}^T\vv{P}_m\end{array}\right)\label{eq:pmodel}
\end{equation} where $\vv{g}$ is a Gaussian random vector of length $t$.
\end{definition}
In the constructions of interest to us, the sequence $\mathcal{P}$ is designed to separate structure from Gaussian randomness;
though elements of $\mathcal{P}$ can be deterministic or itself random, Gaussianity is restricted to the vector $\vv{g}$.
The ability of $\mathcal{P}$ to recycle a Gaussian vector effectively depends on certain structural constants that we now define.
\begin{definition}[Coherence of a $\mathcal{P}$-model] For $\mathcal{P} = \{ \vv{P}_i \}_{i=1}^m$, let $\vv{P}_{ij}$ denote the $j^{th}$
column of the $i^{th}$ matrix. The coherence of a $\mathcal{P}$-model is defined as,
\begin{equation}
\mu[\mathcal{P}] = \max_{1 \leq i \leq j \leq m}\sqrt{\frac{\sum_{1 \leq n_{1} < n_{2} \leq n} (\vv{P}^{T}_{i,n_{1}} \vv{P}_{j,n_{2}})^{2}}{n}}
\end{equation}
\end{definition}
Note that $\mu[\mathcal{P}]$ is a maximum over all pairs of rows $1 \leq i \leq j \leq m$ of the rescaled sums of cross-correlations $\vv{P}^{T}_{i,n_{1}} \vv{P}_{j,n_{2}}$
for all pairs of different column indices $n_{1},n_{2}$.
Lower values of $\mu[\mathcal{P}]$ will lead to better quality models. In practice, as we will see in subsequent analysis, it suffices if $\mu[\mathcal{P}] = O(poly(\log(n)))$ which is the case for instance for Toeplitz and Circulant matrices.
The coherence of the $\mathcal{P}$-model is an extremal statistic of pairwise correlations. We couple it with another set of objects describing global structural properties of the model, namely the \textit{coherence graphs}.
\begin{definition}[Coherence Graphs for $\mathcal{P}$-model and their Chromatic Numbers]
Let $1 \leq i,j \leq m$. We define by $\mathcal{G}_{i,j}$ an undirected graph
with the set of vertices $V(\mathcal{G}_{i,j}) = \{\{n_{1},n_{2}\} : 1 \leq n_{1} \neq n_{2}
\leq n$ and $\vv{P}^{T}_{i,n_{1}}\vv{P}_{j,n_{2}} \neq 0\}$ and the set of edges
$E(\mathcal{G}_{i,j}) = \{\{\{n_{1},n_{2}\},\{n_{2},n_{3}\}\}:\{n_{1},n_{2}\},\{n_{2},n_{3}\}
\in V(\mathcal{G}_{i,j})\}$. In other words, edges are between these vertices such that their corresponding $2$-element subsets intersect.
The chromatic number $\chi(i, j)$ of a graph $\mathcal{G}_{i,j}$ is the smallest number of colors that can be used to color all vertices
of $\mathcal{G}_{i,j}$ in such a way that no two adjacent vertices share the same color.
\end{definition}
The chromatic number of a $\mathcal{P}$-model is defined as follows:
\begin{definition}[Chromatic number of a $\mathcal{P}$-model]
The chromatic number $\chi[\mathcal{P}]$ of a $\mathcal{P}$-model is given as:
$$\chi[\mathcal{P}] = \max_{1 \leq i \leq j \leq m} \chi(i, j),$$
where $\mathcal{G}_{i,j}$ are associated coherence graphs.
\end{definition}
As it was the case for the coherence $\mu[\mathcal{P}]$, smaller values of the chromatic number $\chi[\mathcal{P}]$ lead to better theoretical results regarding
the quality of the model.
Intuitively speaking, coherence graphs encode in a compact combinatorial way correlations between different rows of the structured matrix produced
by the $\mathcal{P}$-model.
The chromatic number $\chi[\mathcal{P}]$ is a single combinatorial parameter measuring quantitatively these dependencies. It can be easily computed
or at least upper-bounded (which is enough for us) for $\mathcal{P}$-models related to all structured matrices considered in this paper.
The following is a well-known fact from graph theory:
\begin{lemma}
\label{graph_lemma}
The chromatic number $\chi(G)$ of an undirected graph $G$ with maximum degree $d_{max}$ satisfies: $\chi(G) \leq d_{max} + 1$.
\end{lemma}
For all instantiations of $\mathcal{P}$-models considered in this paper leading to various structured matrices, the vertices of associated coherence graphs will turn out to have small degrees and hence, by Lemma \ref{graph_lemma}, small chromatic numbers.
We will introduce one more structural parameter of the $\mathcal{P}$-model, depending on whether it is specified deterministically or randomly.
\begin{definition}
The uni-coherence $\tilde{\mu}[\mathcal{P}]$ of the $\mathcal{P}$-model is defined as follows.
If matrices $\vv{P}_{i}$ are constructed deterministically then
$\tilde{\mu}[\mathcal{P}] = \max_{1 \leq i < j \leq m} \sum_{n_{1}=1}^{n}|\vv{P}^{T}_{i,n_{1}}\vv{P}_{j,n_{1}}|.$
If the matrices that specify $\mathcal{P}$ are constructed randomly, then we take
$\tilde{\mu}[\mathcal{P}] = \max_{1 \leq i < j \leq m} \mathbb{E}[|\sum_{n_{1}=1}^{n}\vv{P}^{T}_{i,n_{1}}\vv{P}_{j,n_{1}}|].$
\end{definition}
It turns out that the sublinearity in $n$ of uni-coherence $\tilde{\mu}[\mathcal{P}]$ helps to establish strong theoretical results
regarding the quality of the $\mathcal{P}$-model.
\iffalse
LDRM mechanism for gaussian and arc-cosine kernels approximation}
\label{sec:theorems}
In this section we will present structured mechanism for fast computation of the gaussian and arc-cosine kernels with low displacement rank matrices. Arc-cosine kernel is parametrized by one integer parameter $n=0,1,...$ (see:...). For $n=0$ its computation reduces to the computation of the angular distance.
We will focus here on $n=1$ even though our method can be extended to arbitrary values of the parameter $n$. As mentioned before, the advantage of using low displacement rank matrices lies in smooth transition from the unstructured setting to the fully structured one by manipulating the rank parameter.
\subsection{Basics}
We need to introduce few definitions. Let $A \in \mathbb{R}^{m \times m}$, $B \in \mathbb{R}^{n \times n}$
The \textit{Stein displacement operator} $L=\Delta_{A,B} : \mathbb{R}^{m \times n} \rightarrow \mathbb{R}^{m \times n}$ is defined as follows:
\begin{equation}
\Delta_{A,B}[M] = M - AMB.
\end{equation}
To express the structured matrix with low displacement rank in a straightforward way by its low displacement generators, we will use the following known result:
\begin{theorem_k}
\label{krylov_theorem}
If an $n \times n$ matrix $M$ is such that $\Delta_{A,B}[M] = GH^{T}$, where $G=[g^{1},...g^{r}] \in \mathbb{R}^{n \times r}$ and $H=[h^{1},...,h^{r}] \in \mathbb{R}^{n \times r}$ and the operator matrices satisfy $A^{n} = aI$, $B^{n} = bI$ for some scalars $a,b$ then $M$ can be expressed as:
\begin{equation}
\label{krylov_equation}
M = \frac{1}{1-ab} \sum_{j=1}^{r} krylov(A,g^{j}) krylov(B^{T}, h^{j})^{T},
\end{equation}
where: $krylov(A,v) =[v, Av, A^{2}v, ... ,A^{n-1}v]$.
\end{theorem_k}
We will focus on operator matrices $A,B$ for which $a,b \in \{-1,1\}$.
Parameter $r$ in the expression above is called the \textit{rank of the related displacement operator}.
Larger values of $r$ enable us to cover ``more unstructured'' matrices with the additional computational cost, while smaller values of $r$ correspond to more structured setting, where the computations regarding these matrices (such as matrix-vector product computations) can be significantly sped up.
\begin{definition}
We say that a matrix $M \in \mathbb{R}^{n \times n}$ is a \textit{general cyclic matrix} if it corresponds to the linear transformation of the form $(v_{0},...,v_{n-1}) \rightarrow (s_{0}v_{\sigma^{-1}(0)},...,s_{n-1}v_{\sigma^{-1}(n-1)})$,
where $\sigma : \{0,...,n-1\} \rightarrow \{0,...,n-1\}$ is a permutation with one cycle in its cyclic decomposition and $s_{i} \in \{-1,+1\}$
for $i=0,...,n-1$.
\end{definition}
\begin{figure}[H]
\vspace{-0.12in}
\vspace{-0.1in}
\centering
\includegraphics[width = 2.5in]{images/cycle.png}
\vspace{-0.1in}
\caption{On the left: Cycle representng permutation $\sigma$. On the right: corresponding general cyclic matrix $M$.\\}
\label{fig:cycle}
\vspace{-0.32in}
\end{figure}
In the theory of low displacement rank matrices of special interests are matrices: $Z_{1}$ and $Z_{-1}$ which stand for: circulant and skew-circulant matrices respectively (see:...) Their importance is reflected by the fact that if in Theorem \ref{krylov_theorem} one takes: $A = Z_{1}$ and $B=Z_{-1}$ then equation \ref{krylov_equation} covers rich class of structured matrices such as: all $n \times n$ corculant and skew-circulant matrices (for $r \geq 1$), all $n \times n$ Toeplitz matrices (for $r \geq 2$), inverses of Toeplitz matrices (for $r \geq 2$), all products of the form $A_{1} \cdot ... \cdot A_{t}$ (for $r \geq 2t$), all linear combinations of the form $\sum_{i=1}^{p}\beta_{i} A^{i}_{1} \cdot ... \cdot A^{i}_{t}$ (for $r \geq 2tp$), finally all $n \times n$ matrices for $r = n$, where $A^{i}_{j}$s and $A_{i}$s stand for Toeplitz matrices or inverses of Toeplitz matrices.
As one can easily see $Z_{1}$ and $Z_{-1}$ are special cases of general cyclic matrices. In the presented structured mechanism we will show how to use low displacement rank matrices that, when represented in the form given by equation \ref{krylov_equation}, use as $A$ and $B$ arbitrary general cyclic matrices.
That will give us much more flexibility in setting up the structured pipeline.
\fi
\subsection{Examples of $\mathcal{P}$-model structured matrices}
\label{sec:examples}
Below we observe that various structured random matrices can be constructed according to the $\mathcal{P}$-model, i.e. by specifying a sequence of matrices $\vv{P}_i$ in Eqn.~\ref{eq:pmodel}. We note that chromatic numbers and coherence values of these $\mathcal{P}$-models are low. In the next section, we show that this implies that we can get unbiased, low-variance kernel approximations from these matrices, for various choices of nonlinearities. Here we consider square structured matrices for which $m = n$, or rectangular matrices with $m < n$ obtained by selecting first $m$ rows of a structured matrix.
\subsubsection{Circulant matrices}
Circulant matrices can be constructed via the $\mathcal{P}$-model with budget of randomness $t=n$
and matrices $\{\vv{P}_{i}\}_{i=1}^m$ of entries in $\{0,1\}$. See Fig. \ref{fig:cycle} for an illustrative construction.
The coherence of the related $\mathcal{P}$-model trivially satisfies: $\mu[\mathcal{P}] = O(1)$ and $\tilde{\mu}[\mathcal{P}] = 0$.
The coherence graphs are vertex disjoint cycles. Since each cycle can be colored with at most $3$ colors, the chromatic number of the
$\mathcal{P}$-model satisfies: $\chi[\mathcal{P}] \leq 3$.
\begin{figure}[h]
\vspace{-0.05in}
\vspace{-0.1in}
\centering
\includegraphics[height=6cm, width = 3.35in]{images/graph.png}
\vspace{-0.1in}
\caption{Top left: Circulant gaussian matrix $\mathcal{C}$. Top right: matrices $\vv{P}_{1},\vv{P}_{2},\vv{P}_{3},\vv{P}_{4}$ from
the $\mathcal{P}$-model generating $\mathcal{C}$ from the ``budget of randomness'' $(g_{1},...,g_{5})$.
Bottom: Graph $\mathcal{G}_{i_{1},i_{2}}$ corresponding to two highlighted rows of $\mathcal{C}$.
Graphs obtained from circulant matrices are collections of cycles thus their chromatic number is at most $3$.\\}
\label{fig:cycle}
\vspace{-0.32in}
\end{figure}
\subsubsection{Toeplitz and Hankel matrices}
The associated $\mathcal{P}$-models are obtained in a similar way as for circulant matrices, in particular each column of each $\vv{P}_{i}$
is a binary vector. The corresponding coherence graphs have vertices of degrees at most $2$ and thus the chromatic number $\chi[\mathcal{P}]$
is at most $3$. As for the previous case, coherence $\mu[\mathcal{P}]$ is of the order $O(1)$ and $\tilde{\mu}[\mathcal{P}] = 0$.
\subsubsection{Fastfood matrices}
The Fastfood~\cite{LSS13} approach is a very special case of the $\mathcal{P}$-model.
Note that the core term in the Fastfood transform, Eqn.~\ref{eq:fastfood}, is the structured matrix $\textbf{H}\textbf{G}$, where $\textbf{H}=\{{h}_{i,j}\}$ is Hadamard and $\textbf{G}$
is a random diagonal gaussian matrix (the rightmost terms $\vv{H}\vv{B}$ in Eqn.~\ref{eq:fastfood} implement data preprocessing
to make all datapoints dense, and normalization is implemented by the leftmost scaling matrix $\vv{S}$). The matrix $\textbf{H}\textbf{G}$ can be constructed via the $\mathcal{P}$-model
with the fixed budget of randomness $\vv{g}=(g_{1},...,g_{n})$ and using the sequence of matrices
$\mathcal{P}=(\textbf{P}_{1},...,\textbf{P}_{n})$, where each $\textbf{P}_{i}$ is a random diagonal matrix with
entries on the diagonal of the form: $h_{i,1},...,h_{i,n}$.
The quality of the FastFood approach can be now explained in the general $\mathcal{P}$-model method framework. One can easily see that the graphs related to the model are empty (since $\vv{P}^{T}_{i,n_{1}} \vv{P}_{j,n_{2}} = 0$ for $n_{1} \neq n_{2}$).
The sublinearity of $\tilde{\mu}[\mathcal{P}]$ comes from the fact that with high probability any two rows of $\vv{H}\vv{G}$ are close to be orthogonal.
\subsubsection{Toeplitz-like semi-Gaussian matrices}
\label{subsec:semi}
Consider Toeplitz-like matrices expressible by Eqn.~\ref{eq:circ_skewcirc} with displacement rank $r$.
We will assume that $\vv{g}^{1},...,\vv{g}^{r} \in \mathbb{R}^{n}$ defining the Circulant-components in Eqn.~\ref{eq:circ_skewcirc} are independent Gaussian vectors. They will serve as a
``budget of randomness'' in the related $\mathcal{P}$-model that we are about to describe, with $r$ allowing a tunable tradeoff between structure and randomness. The vectors $\vv{h}^{1},...,\vv{h}^{r}$defining the skew-Circulant components in Eqn.~\ref{eq:circ_skewcirc} can be defined in different ways. Below we present two general schemes:
\textbf{Random discretized vectors $\vv{h}^{i}$:} Each dimension of each $\vv{h}^{i}$ is chosen independently at random from the
binary set $\{-\frac{1}{\sqrt{nr}},\frac{1}{\sqrt{nr}}\}$.
\textbf{Sparse setting:} Each $\vv{h}^{i}$ is sparse (but nonzero), i.e. has only few nonzero entries.
Furthermore, the sign of each $\vv{h}^{i}_{j}$ is chosen independently at random and
the following holds: $\|\vv{h}^{1}\|^{2}+...+\|\vv{h}^{r}\|^{2}=1$.
This setting is characterized by a parameter $\kappa$ defining the
size of the set of dimensions that are nonzero for at least one
$\vv{h}^{i}$.
We refer to such matrices as Toeplitz-like semi-Gaussian matrices. We now sketch how they can be obtained from the $\mathcal{P}$-model.
We take $t=nr$ and $\vv{g} = (g^{1}_{1},...,g^{1}_{n},...,g^{r}_{1},...,g^{r}_{n})^{T}$.
The matrix $\vv{P}_{1}$ is constructed by vertically stacking $r$ matrices $\vv{S}_{j}$ for $j=1,...,r$,
where each $\vv{S}_{j}$ is constructed as follows. The first column of $\vv{S}_{j}$ is $\vv{h}^{j}$ and the
subsequent columns are obtained from previous by skew-Circulant downward shifts.
Matrix $\vv{P}_{i}$ for $i>1$ is obtained from $\vv{P}_{i-1}$ by upward Circulant shifts, independently for each column at each block $\vv{S}_{j}$.
Matrices constructed according to this procedure satisfy conditions regarding certain structural parameters of the $\mathcal{P}$-model (see: Theorem \ref{ldrm-extra}).
In particular, in the sparse semi-Gaussian setting the corresponding coherence graphs have vertices of degrees bounded by a constant; thus,
by Lemma \ref{graph_lemma} the $\mathcal{P}$-models associated with them have low chromatic numbers.
\subsection{Construction of Random Feature Maps}
\label{sec:alg}
Given $S[\mathcal{P}]$, the $m\times n$ structured random matrix defined by a $\mathcal{P}$-model, in lieu of using the $k \times n$
Gaussian random matrix $\vv{M}$ in Eqn.~\ref{eq:featuremap}, the feature map for a data vector $\vv{x}$ is constructed as follows.
\begin{compactitem}
\item Preprocessing phase: Compute $\vv{x}' = D_{1}HD_{0}\vv{x}$, where $H \in \mathbb{R}^{n \times n}$ is a $l_{2}$-normalized Hadamard matrix and $D_{0},D_{1} \in \{-1,+1\}^{n \times n}$ are independent random diagonal matrices. Note that this transformation does not change the values of Gaussian or Arc-cosine kernels, since they are spherically-invariant. This preprocessing densifies the input data vector.
\item Compute $\vv{x}'' = S[\mathcal{P}] \vv{x} \in \mathbb{R}^m$.
\item Compute $\vv{\bar{x}} \in \mathbb{R}^k$ by concatenating random instantiations of the vector $\vv{x}''$ above obtained from $k/m$ independent constructions of $S[\mathcal{P}]$.
\item Return $ \Psi(\vv{x}) = \frac{1}{\sqrt{k}} s(\vv{\bar{x}})$
\end{compactitem}
\iffalse
\textbf{Step 1}
First we preprocess $\mathcal{X}$ by passing it over a linear transformation $D_{1}HD_{0}$, where $H \in \mathbb{R}^{n \times n}$ is a $l_{2}$-normalized Hadamard matrix and
$D_{0},D_{1} \in \{-1,+1\}^{n \times n}$ are independent random diagonal matrices
Note that this transformation does not change the values of our kernels, since they are spherically-invariant.
\textbf{Step 2}
Instead of computing $f(\textbf{Gx})f(\textbf{Gy})$ in the preprocessed dataset, one computes the ``structured version'' of that product,
which is: $f(\textbf{Ax})f(\textbf{Ay})$, where $\textbf{A}$ is the structured version of $\textbf{G}$.
Matrix $\textbf{A} \in \mathbb{R}^{k \times n}$ is build of $\frac{k}{m}$ blocks stacked vertically.
Each block (for $i=1,...,\frac{k}{m}$) is a structured matrix $G_{struct}^{i} \in \mathbb{R}^{m \times n}$
obtained with the use of the $\mathcal{P}$-model and defined as $G_{struct}^{i} = \vv{S}[\mathcal{P}]$ (see: Def. \ref{def:p_model}).
Different structured matrices $G_{struct}^{i}$ are constructed independently.
The preferable $\mathcal{P}$-models for the above algorithm are those for which $\chi[\mathcal{P}],\mu[\mathcal{P}] = O(1)$
and $\tilde{\mu}[\mathcal{P}] = o(\frac{n}{\log^{2}(n)})$, as it will become clear in section~\ref{sec:theory}. All the structured matrices discussed in this paper satisfy these conditions.
\fi
Note that the displacement rank $r$ for low displacement rank matrices and the number of rows $m$ of a single structured block can be used to control the ``budget of randomness"; $m=1$ reduces to a completely unstructured matrix.
\section{Theoretical results}
\label{sec:theory}
In this section we provide concentration results regarding $\mathcal{P}$-model for Gaussian and arc-cosine kernels,
showing in particular that the variance of the computed structured approximation of the kernel
is close to the unstructured one. We also present results targeting specifically low displacement rank structured matrices, and show how the displacement rank knob can be used to increase the budget of randomness and reduce the variance.
Let us denote by $\tilde{{\cal K}}_{\mathcal{P}}(\textbf{x},\textbf{z})$ the approximation of the kernel for two vectors
$\vv{x},\vv{z} \in \mathbb{R}^{n}$ if the $\mathcal{P}$-model is used.
By $\tilde{{\cal K}}_{\vv{G}}(\textbf{x},\textbf{z})$ we denote the approximation of the kernel for two vectors
$\vv{x},\vv{z} \in \mathbb{R}^{n}$ if the fully unstructured setting with truly random Gaussian matrix $\vv{G}$ is applied.
All the proofs are in the Appendix. We start with the following result.
\begin{lemma}[Unbiasedness of the $\mathcal{P}$-model]
\label{unbiasedness_lemma}
Presented $\mathcal{P}$-model mechanism gives an unbiased estimation of the Gaussian and $b^{th}$-order arc-cosine kernels for $b \in \{0,1\}$
if for every $\vv{P}_{i}$ any two different columns $\vv{P}_{i,j}$,$\vv{P}_{i,k}$ of $\vv{P}_{i}$ satisfy $\vv{P}^{T}_{i,j}\vv{P}_{i,k}=0$.
Thus, $\mathbb{E}[\tilde{{\cal K}}_{\mathcal{P}}(\textbf{x},\textbf{z})] = {\cal K}(\vv{x},\vv{z}).$
\end{lemma}
The orthogonality condition $\vv{P}^{T}_{i,j}\vv{P}_{i,k}=0$ is trivially satisfied by Hankel, circulant or Toeplitz structured matrices produced by the $\mathcal{P}$-model
as well as Toeplitz-like semi-Gaussian matrices, where each $\vv{h}^{i}$ has one nonzero entry.
It is also satisfied in expectation (which in practice suffices) for all presented Toeplitz-like semi-Gaussian matrices.
For a $\mathcal{P}$-model, where matrices $\vv{P}_{i}$ were chosen randomly we denote as $\eta[\mathcal{P}]$ the maximum possible value that
a random variable $(\vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{1}})^{2}$ can take for $1 \leq i < j \leq m, 1 \leq n_{1} \leq n$.
Without loss of generality we will assume that data vectors are drawn from the ball $\mathcal{B}(0,1)$ centered at $0$ of unit $l_{2}$ norm. Below we state results regarding $d^{th}$ moments of the obtained kernel's approximation via the $\mathcal{P}$-model
that lead to the concentration results.
\begin{theorem}
\label{main_theorem_1}
Let $\vv{x},\vv{z} \in \mathcal{B}(0,1)$ and let $d \in \mathbb{N}$.
Assume that each structured block of a matrix $\vv{A}$ (see: Section \ref{sec:alg}) produced according to the $\mathcal{P}$-model has $m$ rows
and $\tilde{\mu}[\mathcal{P}] = o(\frac{n}{\log^{2}(n)})$.
If matrices $\vv{P}_{i}$ of the $\mathcal{P}$-model are chosen randomly then assume furthermore that for
any $1 \leq i < j \leq m$ and $1 \leq n_{1} < n_{2} \leq n$ the $n_{1}^{th}$
column of $\vv{P}_{i}$ is chosen independently from the $n_{2}^{th}$ column of $\vv{P}_{j}$.
If matrices $\vv{P}_{i}$ are chosen deterministically then for any $T, \epsilon > 0$ the following is true for $n$ large enough:
$$|\mathbb{E}[\tilde{{\cal K}}_{\mathcal{P}}^{d}(\textbf{x},\textbf{z})]-\mathbb{E}[\tilde{{\cal K}}_{\vv{G}}^{d}(\textbf{x},\textbf{z})]| \leq O(p_{gen}(T) + p_{struct}(T) + d \epsilon),$$
where:
\begin{equation}
p_{gen}(T) = \frac{4d}{\sqrt{2 \pi T}} e^{-\frac{T}{2}} + 4ne^{-\frac{\log^{2}(n)}{8}},
\end{equation}
\begin{align}
\begin{split}
p_{struct}(T) = 4\sum_{i=1}^{m}\chi(i,i)e^{-\frac{1}{8\mu^{2}[\mathcal{P}]\chi^{2}[\mathcal{P}]}\frac{n}{\log^{6}(n)}}\\
+2\sum_{1 \leq i \leq j \leq m}\chi(i,j)e^{-\frac{\epsilon^{2}\sqrt{n}}{8\mu^{2}[\mathcal{P}]\chi^{2}[\mathcal{P}]T\log^{4}(n)}}\\
\end{split}
\end{align}
and expectations are taken in respect to random choice for a Gaussian vector $\vv{g}$.
If $\vv{P}_{i}s$ are chosen from the probabilistic model then the above holds with probability at least
$1-p_{wrong}$ in respect to random choices of $\vv{P}_{i}s$, where
$$p_{wrong} = 2\sum_{i \leq i < j \leq m}e^{-\frac{n}{8\log^{6}(n)\eta[\mathcal{P}]}}.$$
\end{theorem}
Let us comment on the result above. The upper bound is built from two main components: $p_{gen}$ and $p_{struct}$.
The first one depends on the general parameters of the setting: dimensionality of the data $n$ and order of the computed moment $d$.
The second one is crucial to understand how the structure of the matrix influences the quality of the model.
We can immediately see that low chromatic numbers $\chi(i,j)$ (see: Section \ref{sec:p_model}) improve quality since they decrease computed upper bound.
Furthermore, low values of the coherence $\mu[\mathcal{P}]$ and chromatic number $\chi[\mathcal{P}]$ also
lead to stronger concentration results. Both observations were noticed by us before, but now we see how they are implied by general theoretical results.
Finally, for all considered settings, where matrices $\vv{P}_{i}$ are constructed randomly
parameter $\eta[\mathcal{P}]$ is of order $O(1)$ thus $p_{wrong}$ in negligibly small.
In particular, if both the chromatic number $\chi[\mathcal{P}]$ and the coherence $\mu[\mathcal{P}]$ are of the order $O(poly(\log(n)))$ then $p_{struct}$
if inversely proportional to the superpolynomial function of $n$ thus is negligible in practice.
That, as we will see soon, will be the case for proposed Toeplitz-like semi-Gaussian matrices with sparse vectors $h^{i}$.
Let us also note that Theorem \ref{main_theorem_1} can be straightforwardly applied to the structured matrix from the Fastfood model since the condition
regarding $\tilde{\mu}[\mathcal{P}]$ is satisfied and so is the independence condition.
Since all the chromatic numbers are equal to zero (because corresponding graphs are empty), $p_{struct} = 0$ and thus the theorem holds.
Theorem \ref{main_theorem_1} implies also that variances of the kernel approximation for the structured $\mathcal{P}$-model case and
unstructured setting are very similar (we borrow denotation from Theorem \ref{main_theorem_1}).
\begin{theorem}
\label{main_variance_theorem}
Consider the setting as in Theorem \ref{main_theorem_1}. If matrices $\vv{P}_{i}$ are chosen deterministically then for any $T, \epsilon > 0$
the following is true for $n$ large enough:
\begin{equation}
|Var(\tilde{{\cal K}}_{\mathcal{P}}(\textbf{x},\textbf{z})) -Var(\tilde{{\cal K}}_{\vv{G}}(\textbf{x},\textbf{z}))|= O(\frac{m-1}{2k}\Delta),
\end{equation}
where $Var$ stands for the variance and $\Delta = p_{gen}(T) + p_{struct} + \epsilon$.
If $\vv{P}_{i}s$ are chosen from the probabilistic model then the above holds with probability at least
$1-p_{wrong}$, where $p_{wrong}$ is as in Theorem \ref{main_theorem_1}.
\end{theorem}
Note that in practice it means that the variance in the structured and unstructured setting is similar.
In particular, choosing $\epsilon=O(\frac{1}{m^{2}})$, $T > 7\log(m)$, one can deduce that the variance in the structured setting is of the order $O(\frac{1}{m})$ for $n$ large enough (the well known fact is that the unstructured variance is of the order $O(\frac{1}{m})$).
Note also that as expected, for $m=1$ the structured setting becomes an unstructured one, since each structured block consists of just one row and different blocks are constructed independently.
{\bf Toeplitz-like semi-Gaussian Low-displacement rank matrices}: Note that the structure of a matrix affects only the $p_{struct}$ factor in the statements above. Thus, we will focus on the structured parameters of the $\mathcal{P}$-model.
We will show that Toeplitz-like semi-Gaussian matrices can be set up so that the above parameters are of required order.
\begin{theorem}
\label{ldrm-intro}
Consider Toeplitz-like semi-Gaussian matrices with sparse skew-Circulant factors (as in Subsection \ref{subsec:semi}). Let $\kappa$
denote the number of dimensions that are nonzero for at least one $\vv{h}^{i}$. Then for $1 \leq i \leq j \leq m$ we have: $\chi(i,j) \leq \kappa^{2} + 1$.
Furthermore, $\mu[\mathcal{P}] \leq \kappa$ and the bound on
$|\mathbb{E}[\tilde{{\cal K}}_{\mathcal{P}}^{d}(\textbf{x},\textbf{z})]-\mathbb{E}[\tilde{{\cal K}}_{\vv{G}}^{d}(\textbf{x},\textbf{z})]|$ derived in Theorem \ref{main_theorem_1}
is valid also here if $r \geq 3log^{5}(n)$ and for $p_{wrong}$ of the order $o(\frac{1}{n})$.
\end{theorem}
The richness of the low displacement rank mechanism comes from the fact that the budget of randomness can be controlled by the rank parameter $r$ and increasing $r$ leads to better quality approximations. In particular, we have:
\begin{theorem}
\label{ldrm-extra}
Consider Toeplitz-like semi-Gaussian matrices with sparse skew-Circulant factors and parameter $\kappa$. Assume that each $\vv{h}^{i}$ has exactly $\alpha$ nonzero dimensions,
each nonzero dimensions taken independently at random from $\{-\frac{1}{\alpha r}, \frac{1}{\alpha r}\}$.
Then,
$\mathbb{P}[|\mu[\mathcal{P}]| > \tau] \leq 4n^{2}e^{-\frac{\tau^{2}\alpha r}{O(\kappa^{2})}}.$
\end{theorem}
Note that increasing rank $r$ leads to sharper upper bounds on the coherence $\mu[\mathcal{P}]$ (in practice $r$ polynomial in $log(n)$ suffices) and thus,
from what we have said so far, to better concentration results for the entire structured scheme.
Analogous variance bounds can also be derived for Toeplitz-like semi-Gaussian matrices where the $\vv{h}^i$ vectors are chosen to be dense. But due to lack of space, these results are included in our supplementary material.
\iffalse
Let us consider now Toeplitz-like semi-Gaussian matrices where the $\vv{h}^i$ vectors are chosen to be dense.
Theoretical results for this case is implied by another general result regarding certain class of $\mathcal{P}$-models.
Structured matrices constructed in a fully probabilistic way belong to this class with high probability.
Again, we borrow notation used in Theorem \ref{main_theorem_1}.
\begin{theorem}
\label{random_theorem}
Consider the setting as in Theorem \ref{main_theorem_1}.
Assume that entries of any fixed column of $P_{i}$ are chosen independently at random.
Assume also that for any $1 \leq i \leq j \leq m$ and any fixed column $col$ of $P_{i}$
each column of $P_{j}$ is a downward shift of $col$ by $b$ entries (possibly with signs of dimensions swapped) and that $b=0$ for $O(1)$ columns in $P_{j}$.
Then for and $T >0$ and $n$ large enough the following holds:
\begin{align}
\begin{split}
|\mathbb{E}[\tilde{{\cal K}}_{\mathcal{P}}^{d}(\textbf{x},\textbf{z})]-\mathbb{E}[\tilde{{\cal K}}_{\vv{G}}^{d}(\textbf{x},\textbf{z})]| \leq
O(\Delta),
\end{split}
\end{align}
where
$\Delta = p_{gen}(T) + p_{struct}(T) + d\epsilon + 2e^{-n^{\frac{1}{3}}}$
and
$$\epsilon = \frac{\log^{3}(n)}{n}\left(n^{\frac{2}{3}} + \max_{1 \leq i \leq j \leq m} |\sum_{1 \leq n_{1} < n_{2} \leq n} \vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}|\right).$$
As a corollary:
\begin{equation}
|Var(\tilde{{\cal K}}_{\mathcal{P}}(\textbf{x},\textbf{z})) -Var(\tilde{{\cal K}}_{\vv{G}}(\textbf{x},\textbf{z}))|=
O(\frac{m-1}{2k}\Delta).
\end{equation}
\end{theorem}
Note that introduced dense semi-gaussian matrices trivially satisfy conditions of Theorem \ref{random_theorem}
(look for the description of matrices $\vv{P}_{i}$ from Subsection: \ref{subsec:semi}).
The role of rank is similar as in the sparse setting, i.e. larger values of $r$ lead to sharper concentration results.
Theorem \ref{random_theorem} can be applied to classes of matrices for which $|\sum_{1 \leq n_{1} < n_{2} \leq n} \vv{P}_{i,n_{1}}^{T}\vv{P}_{j,n_{2}}|$
is small and random Toeplitz-like semi-Gaussian matrices with dense skew-Circulant factors satisfy this condition with high probability.
\fi |
\section{Introduction}
Recently, CFT techniques have been extensively used to extract the critical behavior of the various quantum field theories in dimensions larger than two. Among them, the numerical conformal bootstrap is one of the most successful tools \cite{ElShowk:2012ht}. In a paper by Rychkov-Tan \cite{Rychkov:2015naa}, the multiplet recombination was used to gain the (irrelevant) critical exponents of the Wilson-Fisher fixed point of the $\phi^4$-theory in $(4-\epsilon)$ dimensions. They assumed that the $\phi^3$ operator should be a descendant state of the elementary $\phi$ field and thus the quantum dimension of $\phi^3$ is related to the one of $\phi$ and then the field $\phi^3$ should appear in the OPE as the descendant. Although, at the Gaussian fixed point, these two operators are both primary but for the WF fixed point, $\phi^3$ is a member of the $\phi$ conformal multiplet. They called this situation ``multiplet recombination''.
They have studied the two-point functions and the three-point functions by using the multiplet recombination and the OPE and they have found the anomalous dimensions of the composite operators $\phi^n$ at the leading order in $\epsilon$ without any diagrammatic calculation.
The method developed in \cite{Rychkov:2015naa} was immediately applied to other types of the ``WF fixed point'', for example, the $\phi^6$-theory in $(3-\epsilon)$ dimensions \cite{Basu:2015gpa} and the Gross-Neveu model in $(2+\epsilon)$ dimensions \cite{Raju:2015fza,Ghosh:2015opa}, where the anomalous dimensions of various composite operators are derived at the leading order in $\epsilon$. In all these examples, it is important that the anomalous dimension of the elementary field is $O(\epsilon^2)$ and the ones of the composite operators are $O(\epsilon)$. Then the naive application of the method \cite{Rychkov:2015naa} to the $\phi^3$-theory in $(6-\epsilon)$ dimensions did not work due to the fact that in the six-dimensional $\phi^3$-theory the wave function renormalization of the field $\phi$ starts with the one-loop graph.
This method is quite generic and does not require the perturbative treatment, which implies that potentially one can predict the non-perturbative results beyond the leading order approximation. However the consideration of only the two- and three-point functions and the OPE give the leading expression for the critical exponents at $O(\epsilon)$ or $O(\epsilon^2)$.\footnote{The next leading order of the anomalous dimensions was studied in \cite{Sen:2015doa} by using the CFT techniques and the unitarity.}
In this paper, we will study and re-organize the Rychkov-Tan method from the more perturbative and more Lagrangian-based point of view in order to adapt the method also to the six-dimensional $\phi^3$-theory. The similar analysis was already done in \cite{Anselmi:1998bh, Belitsky:2007jp} and recently in \cite{Giombi:2016hkj}. The main point of the re-organized method is the use of the Schwinger-Dyson equation without contact terms which do not contribute to the discussion here. This is not the same as the multiplet recombination used in \cite{Rychkov:2015naa} since the multiplet recombination generally can not be derived from the Lagrangian-based approach. However, for the leading order calculation, these are effectively equivalent.
The Schwinger-Dyson equation for the scalar field theory is schematically written as
\begin{align}
\braket{\Box_x \phi(x) O_1(x_1) O_2 (x_2) \cdots } = \braket{\frac{g}{3!}\phi^3(x) O_1(x_1) O_2 (x_2) \cdots },
\end{align}
where we considered the $\phi^4$-theory and neglected the contact terms. This equation should be quantum-mechanically regarded as the renormalized one.
However, when we estimate the right-hand side at $O(g)$, the tree-level evaluation of the correlation function suffices for our purpose and then all the quantities can be reduced to the bare quantities. The Schwinger-Dyson equation reduces to the classical equation of motion.
If we restrict our attention to the conformal field theory and to the two- or three-point functions of the conformal primary operators, the left-hand side can be fixed up to the constant coefficient. Due to the derivative $\Box_x$, the left-hand side is proportional to the anomalous dimensions of the primary operators. Equating both sides, we will find the values of the anomalous dimensions as the function of the coupling.
For the Wilson-Fisher fixed point we need to know the critical coupling $g_*$. Our claim in this paper is that considering the two- and three-point functions and employing the Schwinger-Dyson equation (classical equation of motion), we can decide the value of $g_*$ without any input from the loop calculation. The use of the Schwinger-Dyson equation for the calculation of the anomalous dimensions is very reminiscent of the story that the renormalized equation of motion was used for relating the critical exponents and for reducing the number of the independent exponents \cite{Amit:1984ms,Brezin:1974zr}.
We will study the $\phi^3$-theory in $(6 - \epsilon)$ dimensions and find the leading critical exponent at $O(\epsilon)$ without any input from the Feynman diagrammatic calculation. The similar analysis was recently carried out in \cite{Giombi:2016hkj}, where however the critical coupling $g_*$ is decided from the $\beta$ function derived from the perturbative calculation. The calculation in this paper is very respecting the method by Rychkov-Tan \cite{Rychkov:2015naa} in a sense that all the calculations include no Feynman diagrammatic calculation and use the general forms of the two- and three-point functions of the conformal primary operators although our method is more closer to the perturbative method than the one by Rychkov-Tan \cite{Rychkov:2015naa}.
As the result, when we study the higher order behavior of the critical exponents, we will inevitably need the perturbative calculation of the Feynman graphs but the use of the equation of motion will reduce the complexity of the perturbative calculation.
The organization of the paper is as follows. In first two sections we will study the $\phi^6$-theory in $(3-\epsilon)$ dimensions and the $\phi^4$-theory in $(4-\epsilon)$ dimensions respectively. These two sections serve as an explanation of our method.
In Section 4, we will consider the $\phi^3$-theory in $(6-\epsilon)$ dimensions.
In Section 5, we generalize the result of Section 4 by introducing the additional $O(N)$ scalar fields and modifying the interaction.
In Section 6, we summarize the results and discuss the potential future directions.
\section{$\phi^6$-theory in $(3-\epsilon)$ dimensions}
We will first consider the $\phi^6$-theory in $d=3- \epsilon$ dimensions as an illustration of our strategy. This theory has an analog of the Wilson-Fisher fixed point in $(4-\epsilon)$ dimensions. The perturbative treatment can be found, for example, in \cite{Pisarski:1982vz,Hager:2002uq,Lewis:1978zz} and the conformal method using the multiplet recombination was performed in \cite{Basu:2015gpa}. In the following calculation, we assume that all the $\phi^n$ operators except for $n=5$ are the primary operators and that the conformal symmetry appears at the ``Wilson-Fisher fixed point''.
\subsubsection*{Set-up}
The action of the $\phi^6$-theory is
\begin{align}
S=\int d^d x \, \left( \frac{1}{2} \partial \phi^2 +\frac{g \mu ^{2 \epsilon}}{6!} \phi^6 \right),~~~d=3-\epsilon
\end{align}
and we will maintain the coupling constant in a dimensionless one. We are only interested in the leading order calculations and then the $\mu^{2\epsilon}$ factor does not play any role. Therefore we will omit this factor in the following discussion. For the purpose of the perturbative (Feynman diagrammatic) calculation, we have to include the other renormalizable terms, but in our calculation these are not required.
The scaling dimensions for the field $\phi$ and the composite operators $\phi^n$ are defined as
\begin{align}
\Delta_1 &:= \Delta_{\phi} =\frac{1-\epsilon}{2} +\gamma_1 \\
\Delta_n &:= \Delta_{\phi^n} =n\left( \frac{1-\epsilon}{2}\right) +\gamma_n.
\end{align}
For the calculation of the lowest order anomalous dimension, the multiplet recombination employed in \cite{Rychkov:2015naa} is nothing but the classical equation of motion
\begin{align}
\Box \phi =\frac{g \mu^{2\epsilon} }{5!} \phi^5.
\end{align}
in our approach.
In the paper \cite{Rychkov:2015naa}, the multiplet recombination relation is defined as $\Box [\phi]_R = \alpha(\epsilon) [\phi^5]_R$, where $[\cdots]_R$ means the renormalized operator at the Wilson-Fisher fixed point and $\alpha(\epsilon)$ is a certain function which becomes zero at the $\epsilon \rightarrow 0$ limit. Although this is a more generic operator identity and would be necessary for the higher order calculation, but in the lowest order calculation of the anomalous dimensions, the tree-level relation, namely the classical equation of motion is adequate.
\subsubsection*{Two-point function}
As a first step of determining the anomalous dimension from the conformal symmetry and the classical equation of motion, we will study the two-point function of the elementary field $\phi$. This step is completely the same as \cite{Rychkov:2015naa,Basu:2015gpa,Giombi:2016hkj}. The two-point function is given by
\begin{align}
\braket{\phi(x) \phi(y)} \overset{\mathrm{tree-level}}{=} \frac{1}{4 \pi} \frac{1}{|x-y|}\\
\braket{\phi(x) \phi(y)} =c \, |x-y|^{-2 \Delta_1},
\end{align}
where the tree-level value of $c$ is $1/4\pi$.
Correctly speaking, in the second equation, the operator should be a renormalized one $[\phi]_R$. In the following calculation we will multiply the above function by the differential operator $\Box:=\partial_\mu \partial^\mu$ twice and use the equation of motion $\Box \phi =\frac{g \mu^{2\epsilon} }{5!} \phi^5$. Since the equation of motion introduces the coupling dependence and all the correlation functions will be evaluated at the tree-level, we can replace all the renormalized operator with the tree-level (bare) ones. Therefore we can neglect the wave function renormalization in the following discussion. For the same reason, the renormalization factor of the coupling constant can be dropped.
As mentioned above, taking the derivatives of the above two-point function first by $\Box_x$, we have
\begin{align}
\braket{ \Box_x \phi(x) \phi(y)} &= c \Box_x |x-y|^{-2 \Delta_1} \nonumber \\
&= 2c \Delta_1(2\Delta_1 +2-d) |x-y|^{-2\Delta_1- 2} \nonumber \\
&\sim \frac{\gamma_1}{2 \pi} |x-y|^{-5},
\end{align}
where in the last line we have evaluated this at the lowest level in the coupling constant and the epsilon parameter. Since we eventually find the critical coupling $g_* =g(\epsilon)$, neglecting the $\epsilon$-dependence above will suffice for our purpose. We can alternatively calculate the above correlation functions by using the classical equation of motion as
\begin{align}
\left. \braket{ \Box \phi(x) \phi(y)}\right|_{\mathrm{lowest}} &= \left. \frac{g \mu^{2\epsilon}}{5!} \ \braket{\phi^5(x) \phi(y)} \right|_{\mathrm{lowest}}=0,
\end{align}
which means $\gamma_1=O(g^2)$. Next we will further apply the differential operator $\Box_y$ in addition to $\Box_x$,
\begin{align}
\braket{ \Box \phi(x) \Box \phi(y)} &= c \Box_x \Box_y |x-y|^{-2 \Delta_1} \nonumber \\
&= 2c \Delta_1(2\Delta_1 +2) (2\Delta_1 +2-d) (2\Delta_1+4-d) |x-y|^{-2\Delta_1 -4} \nonumber \\
&\sim \frac{3}{\pi} \gamma_1 |x-y|^{-5}.
\end{align}
We can again evaluate the two-point function of $\Box \phi$ by using the classical equation of motion as follows. At $O(g^2)$, all the calculations are carried out in a tree-level approximation.
\begin{align}
\left. \braket{ \Box \phi(x) \Box \phi(y)} \right|_{O(g^2)} &= \left( \frac{g \mu^{2\epsilon}}{5!} \right)^2 \braket{\phi^5(x) \phi^5(y)} \nonumber \\
&=\frac{g^2}{5!} \frac{1}{(4 \pi)^5} |x-y|^{-5}
\end{align}
In the calculation above we again retain the lowest order. Comparing these two results, we obtain
\begin{align}
\gamma_1 =\frac{1}{3 \cdot 4^5 \cdot 5! \, \pi^4} g^2.
\end{align}
As the calculation above shows, all the renormalized quantities $[\phi]_R$ and $g_R$ can be replaced by the bare quantity for the lowest order calculation.
Notice that here we are using the conformal properties of the primary operators so we should tune the coupling constant $g$ to a special value to realize the Wilson-Fisher fixed point. Usually we can calculate $g_* =g(\epsilon)$ by the Feynman diagrammatic approach, but we will alternatively find this critical value by considering the three-point functions without any diagrammatic input, which is based on the philosophy of \cite{Rychkov:2015naa} (and for the extended works see \cite{Basu:2015gpa, Raju:2015fza,Ghosh:2015opa}).
\subsubsection*{Three-point functions}
Next we will study the three-point functions. The conformal symmetry restricts the coordinate dependence of the three-point function of the (quasi-)primary operators as
\begin{align}
\braket{\phi(x) \phi^n (y) \phi^{n+1} (z) } &= C_{1,n,n+1} |x-y|^{\Delta_{n+1}-\Delta_{1}-\Delta_{n}} |y-z|^{\Delta_{1}-\Delta_{n}-\Delta_{n+1}} |x-z|^{\Delta_{n}-\Delta_{1}-\Delta_{n+1}},
\end{align}
where we assume $n \ge 2$ and the tree-level value of $C_{1,n,n+1}$ is given by
\begin{align}
C_{1,n,n+1}^{\mathrm{tree}} &=(n+1)! \frac{1}{(4 \pi)^{n+1}}.
\end{align}
These three-point functions actually include the non-primary operator for $n=4,5$ since $\phi^5$ is assumed to be a descendant at the conformal fixed point we want to analyze. Then for $n=4,5$ the right-hand side is not correct and includes the various terms like
\begin{align}
\sum_{a,b} C_{a,b}|x-y|^a|y-z|^b|z-x|^{-a-b-\Delta_{1}-\Delta_{n}-\Delta_{n+1}}.
\end{align}
However, since we are now interested in the lowest order values of the anomalous dimensions, the above restricted form is adequate even for $n=4$ and $5$. This will be proven in the last paragraph of this section.
In the same way as the two-point function, we apply the differential operator $\Box_x$ and evaluate it at the lowest-level,
\begin{align}
\braket{ \Box_x \phi(x) \phi^n (y) \phi^{n+1} (z) } &\sim C_{1,n,n+1} \gamma_1 |y-z|^{-n} |x-z|^{-3} \nonumber \\
& \qquad -C_{1,n,n+1} (\gamma_1+\gamma_{n}-\gamma_{n+1}) |x-y|^{-2} |y-z|^{-n+2} |x-z|^{-3}.
\end{align}
The terms proportional to $\gamma_1$ will start with $O(g^2)$, then these do not contribute to the following discussion with $O(g)$. Using the classical equation of motion, we find
\begin{align}
\braket{ \Box_x \phi(x) \phi^n (y) \phi^{n+1} (z) } &=\frac{g \mu^{2\epsilon}}{12 (4 \pi)^{n+3}} n(n-1)(n+1)! |x-y|^{-2} |y-z|^{-n+2} |x-z|^{-3}.
\end{align}
Equating these two results we obtain the recursion relation
\begin{align}
\gamma_{n+1} -\gamma_{n}=\frac{n(n-1)}{12 (4 \pi)^2}g +O(g^2) ~~~\mbox{for}~ n \ge 2,
\end{align}
where we have used the tree-level value of $C_{1,n,n+1}$.
Next we will consider the following three-point function
\begin{align}
\braket{\phi(x) \phi(y) \phi^2(z) } &=C_{1,1,2} |x-y|^{\Delta_2 -2\Delta_1} |y-z|^{-\Delta_2} |x-z|^{-\Delta_2},
\end{align}
where the tree-level value of the coefficient $C_{1,1,2}$ is given by
\begin{align}
C_{1,1,2}^{\mathrm{tree}}=\frac{1}{8 \pi^2}.
\end{align}
First, we take the derivative by $\Box_x$ and evaluate it at the lowest order.
\begin{align}
\braket{\Box_x \phi(x) \phi(y) \phi^2(z) } &=C_{1,1,2} (2 \Delta_1 -\Delta_2) (2\Delta_1 +2-d) |x-y|^{\Delta_2 -2\Delta_1-2}|y-z|^{-\Delta_2}|x-z|^{-\Delta_2} \nonumber \\
& \qquad +C_{1,1,2} \Delta_2 (2 \Delta_1+2-d) |x-y|^{\Delta_2 -2 \Delta_1} |y-z|^{-\Delta_2} |x-z|^{-\Delta_2-2} \nonumber \\
&\qquad \quad -C_{1,1,2} \Delta_2 (2\Delta_1 -\Delta_2) |x-y|^{\Delta_2 -2\Delta_1-2}|y-z|^{-\Delta_2+2}|x-z|^{-\Delta_2-2} \nonumber \\
&\sim 2C_{1,1,2} \gamma_1 |y-z|^{-1}|x-z|^{-3}-C_{1,1,2} (2\gamma_1 -\gamma_2) |x-y|^{-2}|y-z| |x-z|^{-3}
\end{align}
If we use the equation of motion and evaluate the left-hand side at the lowest order, the result is vanishing, which means that the non-vanishing terms will start with $O(g^2)$. This implies that the anomalous dimension $\gamma_2$ will start at $O(g^2)$. We further take a derivative by $\Box_y$ and we find
\begin{align}
\braket{\Box_x \phi(x) \Box_y \phi(y) \phi^2(z) } \sim -2 C_{1,1,2} (2\gamma_1-\gamma_2 )|x-y|^{-4} |y-z|^{-1} |x-z|^{-1} +O(g^3).
\end{align}
By using again the equation of motion, the left-hand side becomes
\begin{align}
\braket{\Box_x \phi(x) \Box_y \phi(y) \phi^2(z) } &= \left( \frac{g}{5!} \right)^2 \braket{\phi^5 (x) \phi^5 (y) \phi^2} \nonumber \\
&\sim \frac{g^2}{12 (4 \pi)^6} |x-y|^{-4} |y-z|^{-1} |x-z|^{-1}
\end{align}
Comparing these two results we find
\begin{align}
2\gamma_1- \gamma_2 =-\frac{g^2}{3 \cdot 4^6 \pi^4}
\end{align}
and the recursion relation above can be solved by noticing that $\gamma_2 =0\cdot g^1 +O(g^2)$,
\begin{align}
\gamma_n= \frac{g}{3^2 4^3 \pi^2}n(n-1)(n-2) +O(g^2),~~~\mbox{for}~~ n \ge 2
\end{align}
Note again that the above derivation uses the conformal properties so we need to find the critical coupling $g_*$. This will be given in the following.
\subsubsection*{Determination of $g_*$}
We should notice that $\Delta_5$ can be represented in two ways,
\begin{align}
\Delta_5 &=\Delta_1 +2 =\frac{5 -\epsilon}{2} +\gamma_1 \\
&=5\left( \frac{1-\epsilon}{2} \right) +\gamma_5
\end{align}
Then we get the relation
\begin{align}
\gamma_5-\gamma_1 =2\epsilon.
\end{align}
At $O(g)$, this means $\gamma_5 =2 \epsilon$, which fixes the value of $g_*$ as
\begin{align}
g_*=\frac{2^5 \cdot 3 \pi^2 }{5} \epsilon
\end{align}
Inserting this to the anomalous dimensions obtained above, we find
\begin{gather}
\gamma_1=\frac{1}{1000} \epsilon^2+O(\epsilon^3),~~\gamma_2=\frac{3}{100} \epsilon^2 +O(\epsilon^3)\\
\gamma_{n \ge 3} =\frac{1}{30} n(n-1)(n-2)\epsilon+O(\epsilon^2).
\end{gather}
These results are consistent with the perturbative calculation \cite{Hager:2002uq, Lewis:1978zz} and the conformal method \cite{Basu:2015gpa}. In the paper \cite{Basu:2015gpa}, the value of $\gamma_2$ is not derived.
\subsubsection*{OPE coefficients}
Next we will consider the OPE coefficients which are vanishing in the limit $g \rightarrow 0$.
Let us consider the following three-point function
\begin{align}
\braket{\phi(x) \phi(y) \phi^4 (z)} &=C_{1,1,4} |x-y|^{\Delta_4-2\Delta_1} |y-z|^{-\Delta_4} |x-z|^{-\Delta_4} \nonumber \\
\braket{ \Box_x \phi(x) \phi(y) \phi^4 (z)} & \sim 2 C_{1,1,4} |x-y|^{-1} |x-z|^{-4}.
\end{align}
Again, by the use of the equation of motion, the left-hand side becomes
\begin{align}
\braket{ \Box_x \phi(x) \phi(y) \phi^4 (z)} &=\frac{g}{5!} \braket{\phi^5(x) \phi(y) \phi^4(z)} \nonumber \\
&\sim \frac{g}{(4 \pi)^5} |x-y|^{-1} |x-z|^{-4}.
\end{align}
Then we obtain the lowest order coefficient
\begin{align}
\left. C_{1,1,4} \right|_{O(g)} =\frac{g}{2(4 \pi)^5}.
\end{align}
In a similar way, considering the three-point function $\braket{\phi (x) \phi(y) \phi^6(z)}$, we obtain
\begin{align}
\left. C_{1,1,6} \right|_{O(g)} =\frac{g}{(4 \pi)^6}.
\end{align}
\subsubsection*{Justification of the form of the 3-pt function for $n=3,4$}
In the above calculation, we assume the following form of the three-point function. This is correct for the (quasi-)primary operators but not for their descendants. For the descendant fields, we generically have the different terms. Here we will justify that in the lowest order calculation it is enough to restrict the form of the three-point function to the following form.
\begin{align}
\braket{\phi(x) \phi^n (y) \phi^{n+1} (z) } &= C_{1,n,n+1} |x-y|^{\Delta_{n+1}-\Delta_{1}-\Delta_{n}} |y-z|^{\Delta_{1}-\Delta_{n}-\Delta_{n+1}} |x-z|^{\Delta_{n}-\Delta_{1}-\Delta_{n+1}},
\end{align}
Let us first consider the three-point function,
\begin{align}
\braket{\phi(x) \phi^4(y) \phi^5(z)}.
\end{align}
Since $\phi^5(z)$ is a descendant of a single $\phi(z)$ field, this is proportional to
\begin{align}
\frac{1}{g} \braket{\phi (x) \phi^4(y) \Box_z \phi(z) }.
\end{align}
This correlation function is constrained by the conformal symmetry as follows.
\begin{align}
\braket{\phi (x) \phi^4(y) \Box_z \phi(z) } &=C_{1,4,1} \Box_z \left( |x-y|^{-\Delta_4} |y-z|^{-\Delta_4} |x-z|^{\Delta_4-2 \Delta_1}\right) \nonumber \\
&=C_{1,4,1} \Delta_4 (2\Delta_1 +2-d) |x-y|^{-\Delta_4} |y-z|^{-\Delta_4-2} |x-z|^{\Delta_4-2\Delta_1} \nonumber \\
&\quad +C_{1,4,1} (2\Delta_1 -\Delta_4) (2\Delta_1 +2-d) |x-y|^{-\Delta_4} |y-z|^{-\Delta_4} |x-z|^{\Delta_4 -2\Delta_1-2} \nonumber \\
&\qquad -C_{1,4,1} \Delta_4 (2 \Delta_1 -\Delta_4) |x-y|^{-\Delta_4 +2} |y-z|^{-\Delta_4 -2} |x-z|^{\Delta_4-2\Delta_1-2}
\end{align}
Since the first and second lines have the pre-factors $2\Delta_1 +2-d=2\gamma_1=O(g^2)$ and $C_{1,4,1}=O(g)$, then including the $1/g$ factor, these two lines are totally the $O(g^2)$ contributions. Since we are calculating the $O(g)$ contribution to the anomalous dimensions for $\gamma_{n \ge 3}$, these terms can be neglected. On the other hand, the third line is $\frac{1}{g}\, C_{1,4,1} \Delta_4 (2 \Delta_1-\Delta_4) = O(g^0)$ and remains at the lowest-order calculation. If we notice the relation
\begin{align}
\Delta_5 =\Delta_1 +2,
\end{align}
the last term can be recast in
\begin{align}
C_{1,4,5}|x-y|^{-\Delta_1 -\Delta_4 +\Delta_5} |y-z|^{-\Delta_4-\Delta_5 +\Delta_1} |x-z|^{\Delta_4-\Delta_1-\Delta_5},
\end{align}
where we combine all the $O(g^0)$ coefficients into $C_{1,4,5}$. This is the reason why the above form of the three-point function is adequate for our purpose. For the correlation function $\braket{\phi(x) \phi^5 (y) \phi^{6} (z) } $, the same argument can be applied. This situation is applied for the $\phi^4$-theory in $(4-\epsilon)$ dimensions. But for the $\phi^3$-theory in $(6-\epsilon)$ dimensions, the situation is quite different since the wave function renormalization start from the one-loop graph. Then for the $\phi^3$-theory, we should include the terms neglected here.
\section{$\phi^4$-theory in $(4-\epsilon)$ dimensions}
A next example is a $\phi^4$-theory in $(4-\epsilon)$ dimensions. This can be analyzed in the same manner as the previous section, where the forms of the three-point functions including the descendant field can be approximated as if they are the correlation functions only with the primaries.
\subsubsection*{Set-up}
The action and the anomalous dimensions are defined as
\begin{align}
S&=\int d^d x \, \left( \frac{1}{2} \partial \phi^2 +\frac{g \mu ^{ \epsilon}}{4!} \phi^4 \right),~~~d=4-\epsilon \\
\Delta_1 &:= \Delta_{\phi} =\left( 1-\frac{\epsilon}{2} \right) +\gamma_1,~~~
\Delta_n := \Delta_{\phi^n} =n\left( 1-\frac{\epsilon}{2} \right)+\gamma_n.
\end{align}
The classical equation of motion relates $\Box \phi$ to $\phi^3$,
\begin{align}
\Box \phi =\frac{g \mu^{\epsilon} }{3!} \phi^3.
\end{align}
\subsubsection*{Two-point function}
We start with the two-point function of $\phi$:
\begin{align}
\braket{\phi(x) \phi(y)} &\overset{\mathrm{tree-level}}{=} \frac{1}{4 \pi^2} \frac{1}{|x-y|^2}\\
\braket{\phi(x) \phi(y)} &=c \, |x-y|^{-2 \Delta_1}.
\end{align}
Taking the derivatives of the above function, we have
\begin{align}
\braket{ \Box \phi(x) \Box \phi(y)} &= c \Box_x \Box_y |x-y|^{-2 \Delta_1} \nonumber \\
&= 2c \Delta_1(2\Delta_1 +2) (2\Delta_1 +2-d) (2\Delta_1+4-d) |x-y|^{-2\Delta_1 -4} \nonumber \\
&\sim \frac{8}{\pi^2} \gamma_1 |x-y|^{-6}
\end{align}
We can evaluate the left-hand side at $O(g^2)$ by using the classical equation of motion.
\begin{align}
\braket{ \Box \phi(x) \Box \phi(y)} &= \left( \frac{g \mu^{\epsilon}}{3!} \right)^2 \braket{\phi^3(x) \phi^3(y)}\nonumber \\
&=\frac{g^2}{3!} \frac{1}{(4 \pi^2)^3} |x-y|^{-6}
\end{align}
Therefore we obtain
\begin{align}
\gamma_1 =\frac{1}{3 \cdot 4^3} \left( \frac{g}{4 \pi^2} \right)^2.
\end{align}
\subsubsection*{Three-point function}
Next, we will study the three-point functions in order to derive the anomalous dimensions of the composite operators and to find the critical value of the coupling $g_*$. Let us start with the correlator $\braket{\phi(x) \phi(y) \phi^2(z) }$, whose general form is constrained from the conformal symmetry as
\begin{align}
\braket{\phi(x) \phi(y) \phi^2(z) } &=C_{1,1,2} |x-y|^{\Delta_2 -2\Delta_1} |y-z|^{-\Delta_2} |x-z|^{-\Delta_2},
\end{align}
where the tree-level coefficient of $C_{1,1,2}$ is obtained from the simple calculation of the Wick contractions and it is
\begin{align}
C_{1,1,2}^{\mathrm{tree}}=\frac{1}{8 \pi^4}.
\end{align}
First, we will take the derivative by $\Box_x$ and evaluate the right-hand side at the lowest order.
\begin{align}
\braket{\Box_x \phi(x) \phi(y) \phi^2(z) } &= C_{1,1,2}(2\Delta_1-\Delta_2) (2 \Delta_1+2-d) |x-y|^{\Delta_2-2\Delta_1-2} |y-z|^{-\Delta_2} |x-z|^{-\Delta_2} \nonumber \\
&\qquad +C_{1,1,2}\Delta_2 (2 \Delta_1+2-d) |x-y|^{\Delta_2-2\Delta_1} |y-z|^{-\Delta_2} |x-z|^{-\Delta_2-2} \nonumber \\
&\qquad \quad -C_{1,1,2} \Delta_2 (2\Delta_1-\Delta_2) |x-y|^{\Delta_2-2\Delta_1-2} |y-z|^{-\Delta_2+2} |x-z|^{-\Delta_2-2} \nonumber \\
&=C_{1,1,2}(2\gamma_1 -\gamma_2)2\gamma_1 |x-y|^{\Delta_2-2\Delta_1-2} |y-z|^{-\Delta_2} |x-z|^{-\Delta_2} \nonumber \\
&\qquad +C_{1,1,2} \Delta_2 2\gamma_1 |x-y|^{\gamma_2-2\gamma_1} |y-z|^{-\Delta_2} |x-z|^{-\Delta_2-2} \nonumber \\
&\qquad \quad -C_{1,1,2} \Delta_2 (2\gamma_1-\gamma_2) |x-y|^{\Delta_2-2\Delta_1-2} |y-z|^{-\Delta_2+2} |x-z|^{-\Delta_2-2} \nonumber \\
& \sim 4C_{1,1,2} \gamma_1 |y-z|^{-\Delta_2} |x-z|^{-\Delta_2-2} -4C_{1,1,2} \gamma_1 |x-y|^{-2} |x-z|^{-4} \nonumber \\
&\qquad + C_{1,1,2} \Delta_2 \gamma_2 |x-y|^{\gamma_2-2} |y-z|^{-\Delta_2+2} |x-z|^{-\Delta_2-2} +O(g^3)\nonumber \\
&\sim 2C_{1,1,2} \gamma_2 |x-y|^{-2}|x-z|^{-4} +O(g^2),
\end{align}
where we used the fact that $\gamma_1$ starts with $O(g^2)$, which is the result from the study of the two-point function.
By using the classical equation of motion, the left-hand side becomes
\begin{align}
\braket{\Box_x \phi(x) \phi(y) \phi^2(z) } &= \frac{g}{3!} \braket{\phi^3 (x) \phi (y) \phi^2 (z)} \nonumber \\
&\sim \frac{g}{ (4 \pi^2)^3} |x-y|^{-2} |x-z|^{-4} +O(g^2)
\end{align}
Then, comparing these two, we find
\begin{align}
\gamma_2=\frac{g}{16 \pi^2}. \label{g2}
\end{align}
One may expect that if we apply further the derivative $\Box_y$ to the above three-point function, we get the $O(g^2)$ result of the anomalous dimension $\gamma_2$. However this is not the case since we will get the uninteresting relation
\begin{align}
\gamma_2^2 =O(g^2) .
\end{align}
For the composite operators $\phi^n$, we study the following three-point functions.
\begin{align}
\braket{\phi(x) \phi^n (y) \phi^{n+1}(z)} =C_{1,n,n+1}|x-y|^{\Delta_{n+1} - \Delta_{1}-\Delta_{n}} |y-z|^{\Delta_{1}-\Delta_{n}-\Delta_{n+1}} |x-z|^{\Delta_{n}-\Delta_{1}-\Delta_{n+1}}
\end{align}
For $n=2,3$, the correlator includes a descendant field $\phi^3$, then the right-hand side generally involves other terms. However the coefficients of such terms start with $O(g)$ or more higher power of $g$. In the following we will consider the $O(g)$ contribution, so these terms can be neglected as before. Notice that we take a derivative of this correlator and this manipulation gives rise to another $O(g)$ pre-factor, so in the above we have to only retain the terms with $O(g^0)$ coefficients. The tree-level coefficient $C_{1,n,n+1}$ is given by
\begin{align}
C_{1,n,n+1}^{\mathrm{tree}}=\frac{(n+1)!}{(4\pi^2)^{n+1}}.
\end{align}
As mentioned above, we will take a derivative by $\Box_x$ and compare it with the result from the classical equation of motion.
\begin{align}
\braket{\Box_x \phi(x) \phi^n (y) \phi^{n+1}(z)} &= 2C_{1,n,n+1}(\gamma_{n+1}-\gamma_n) |x-y|^{-2} |y-z|^{-2n+2} |x-z|^{-4} +O(g^2)\\
\braket{\Box_x \phi(x) \phi^n (y) \phi^{n+1}(z)} &=\frac{g}{3!}\braket{ \phi^3(x) \phi^n (y) \phi^{n+1}(z)} \nonumber \\
&\sim g\, n! \, _{n+1}C_2 \frac{1}{(4 \pi^2)^{n+2}} |x-y|^{-2} |y-z|^{-2n+2} |x-z|^{-4}
\end{align}
Then we find the following recurrence relation
\begin{align}
\gamma_{n+1} -\gamma_n =\frac{g}{4 \pi^2} \frac{n}{4}.
\end{align}
Remembering the result \eqref{g2}, $\gamma_2=\frac{g}{16 \pi^2}$, this is easily solved as
\begin{align}
\gamma_n=\frac{g}{32 \pi^2}n(n-1),~~~\mbox{for}~~ n \ge 2.
\end{align}
From the assumption $\Delta_3 = \Delta_1+2 $, we have
\begin{align}
\gamma_3-\gamma_1=\epsilon
\end{align}
and the critical coupling is found to be
\begin{align}
g_*=\frac{16 \pi^2}{3} \epsilon.
\end{align}
The anomalous dimensions of the composite operators are summarized as
\begin{align}
\gamma_1=\frac{\epsilon^2}{108} ,~~~\gamma_{n \ge 2} =\frac{1}{6}n(n-1)\epsilon +O(\epsilon^2 ),
\end{align}
which is consistent with the perturbative results \cite{Kleinert:2001ax, Kehrein:1992fn} and the conformal method developed in \cite{Rychkov:2015naa}.
\section{$\phi^3$-theory in $(6-\epsilon)$ dimensions}
In this section we study the $\phi^3$-theory in $(6-\epsilon)$ dimensions. As claimed earlier, in this case we are very careful about the three-point functions including the $\phi^2$ fields since $\phi^2$ is a descendant and the three-point functions become more involved.
The bare Lagrangian is
\begin{align}
S=\int d^d x \, \left( \frac{1}{2} \partial \phi^2 +\frac{g \mu ^{ \epsilon/2}}{3!} \phi^3 \right),~~~d=6-\epsilon
\end{align}
and the scaling dimensions are defined as
\begin{align}
\Delta_1 &:= \Delta_{\phi} =\left( 2-\frac{\epsilon}{2} \right) +\gamma_1,~~~\Delta_n := \Delta_{\phi^n} =n\left( 2-\frac{\epsilon}{2} \right)+\gamma_n.
\end{align}
\if0
For $\Delta_1$ and $\Delta_2$, there is a constraint coming from the multiplet recombination.
\begin{align}
\Delta_2 &=\Delta_1+2=4-\frac{\epsilon}{2}+\gamma_1 \\
&= 4- \epsilon + \gamma_2 \\
\gamma_2-\gamma_1&=\frac{\epsilon}{2}
\end{align}
\fi
The classical equation of motion is
\begin{align}
\Box \phi =\frac{g \mu^{\epsilon/2} }{2} \phi^2.
\end{align}
\subsubsection*{Two-point function}
For two-point functions, the manipulation is completely the same as the previous two sections.
\begin{align}
\braket{\phi(x) \phi(y)} &\overset{\mathrm{tree-level}}{=} \frac{1}{4 \pi^3} \frac{1}{|x-y|^4}\\
\braket{\phi(x) \phi(y)} &=c \, |x-y|^{-2 \Delta_1}
\end{align}
For the case where a single d'Alembertian $\Box_x$ acts on the above two-point function, the resulting correlation function $\braket{\phi^2 (x) \phi(y) }$ is vanishing at the tree-level approximation. Then we need to take the derivatives of the two-point function by $\Box_x$ and $\Box_y$,
\begin{align}
\braket{ \Box \phi(x) \Box \phi(y)} &= c \Box_x \Box_y |x-y|^{-2 \Delta_1} \nonumber \\
&= 2c \Delta_1(2\Delta_1 +2) (2\Delta_1 +2-d) (2\Delta_1+4-d) |x-y|^{-2\Delta_1 -4} \nonumber \\
&\sim 4^2 \cdot 6 c \gamma_1 |x-y|^{-2\Delta_1 -4} \nonumber \\
&\sim \frac{24}{\pi^3} \gamma_1 |x-y|^{-8}, \\
\braket{ \Box \phi(x) \Box \phi(y)} &= \left( \frac{g \mu^{\epsilon/2}}{2} \right)^2 \braket{\phi^2(x) \phi^2(y)} \nonumber \\
&=\frac{g^2}{2} \frac{1}{(4 \pi^3)^2} |x-y|^{-8}
\end{align}
Therefore we find that the $O(g)$ contribution to $\gamma_1$ is absent and
\begin{align}
\gamma_1 =\frac{1}{3 \cdot 4^3} \frac{g^2}{4 \pi^3}.
\end{align}
\subsubsection*{Three-point function}
First, let us assume that the field $\phi^n$ is a conformal primary with the scaling dimension $\Delta_n$ except for $n=2$. The three-point function of $\phi$ is fixed by the conformal symmetry as
\begin{align}
\braket{\phi(x) \phi(y) \phi(z)} &=C_{1,1,1} |x-y|^{-\Delta_1} |y-z|^{-\Delta_1} |x-z|^{-\Delta_1}.
\end{align}
Since the above three-point function is zero for the free theory limit, $C_{1,1,1}=O(g)$.
By taking a derivative by $\Box_x$, we obtain
\begin{align}
\braket{\Box_x \phi(x) \phi(y) \phi(z)} &=C_{1,1,1} \Delta_1 (2\Delta_1 +2-d) |x-y|^{-\Delta_1-2} |y-z|^{-\Delta_1} |x-z|^{-\Delta_1} \nonumber \\
&\qquad + C_{1,1,1} \Delta_1 (2\Delta_1 +2-d) |x-y|^{-\Delta_1} |y-z|^{-\Delta_1} |x-z|^{-\Delta_1-2} \nonumber \\
&\qquad \quad - C_{1,1,1} (\Delta_1)^2 |x-y|^{-\Delta_1-2} |y-z|^{-\Delta_1+2} |x-z|^{-\Delta_1-2}\nonumber \\
&\sim -4C_{1,1,1} |x-y|^{-4} |x-z|^{-4}+O(g^2).
\end{align}
Using the classical equation of motion, we find
\begin{align}
\braket{\Box_x \phi(x) \phi(y) \phi(z)} &=\frac{g}{2} \braket{\phi(x)^2 \phi(y) \phi(z)} \nonumber \\
&\sim \frac{g}{(4 \pi^3)^2} |x-y|^{-4} |x-z|^{-4}+O(g^2),
\end{align}
and then
\begin{align}
C_{1,1,1}=-\frac{g}{64 \pi^6}+O(g^2).
\end{align}
Next, we will consider the three-point function including the descendant $\phi^2$. This can be obtained from the three-point function $\braket{\phi(x) \phi(y) \phi(z)}$ since we are now assuming $\phi^2$ is a descendant. At the leading order we find
\begin{align}
\braket{\phi(x) \phi(y) \phi^2(z)} &= \frac{2}{g} \braket{\phi(x) \phi(y) \Box_z \phi(z)} \nonumber \\
&\sim -\frac{\gamma_1}{8 \pi^6} |x-y|^{-2} |y-z|^{-4} |x-z|^{-2} -\frac{\gamma_1}{8 \pi^6}|x-y|^{-2} |y-z|^{-2} |x-z|^{-4} \nonumber \\
&\qquad + \frac{1}{8 \pi^6} |x-y|^{-\Delta_1+2}|y-z|^{-\Delta_1-2} |x-z|^{- \Delta_1-2},
\end{align}
where we omitted the unnecessary terms and factors for our purpose of the $O(g^2)$ computation. For example, in the second line, the terms are proportional to $\gamma_1$, so the powers of $|x-y|, |y-z|$ and $|x-z|$ can be replaced with the classical values. Notice that the difference between this expression and the 3d or 4d ones. In the previous two sections, we can drop the first two terms since the anomalous dimension $\gamma_1$ started with $O(g^2)$ and the one of the descendant field starts with $O(g)$. However, in the six-dimensional $\phi^3$-theory, these two quantities would start with the same order, then we should retain all the three terms above.
Being multiplied by $\Box_x$ and $\Box_y$, the right-hand side becomes
\begin{align}
\braket{ \Box_x \phi(x) \Box_y \phi(y) \phi^2(z)} &= \left(\frac{4 \gamma_1}{\pi^6} +\frac{(2\gamma_1-\epsilon)}{\pi^6} \right)|x-y|^{-4} |y-z |^{-4}|x-z|^{-4}.
\end{align}
On the other hand, applying the classical equation of motion, we find
\begin{align}
\braket{ \Box_x \phi(x) \Box_y \phi(y) \phi^2(z)} &=\frac{g^2}{4} \braket{\phi^2(x) \phi^2(y) \phi^2(z)} \nonumber \\
&\sim \frac{g^2 2^3}{4(4 \pi^3)^3}|x-y|^{-4} |y-z |^{-4}|x-z|^{-4}.
\end{align}
Comparing these two results, we obtain
\begin{align}
\gamma_1=\frac{\epsilon}{6} +\frac{g^2}{3 \cdot 4^3 \pi^3}.
\end{align}
Combining this with
\begin{align}
\gamma_1 =\frac{1}{3 \cdot 4^3} \frac{g^2}{4 \pi^3},
\end{align}
we find the critical coupling and the critical exponent
\begin{align}
g_*^2 =-\frac{2 \cdot 4^3 \pi^3 }{3} \epsilon +O(\epsilon^2),~~~\gamma_1(g^2_*) =-\frac{1}{18} \epsilon+O(\epsilon^2).
\end{align}
This result is consistent with the perturbation results (see, for example, \cite{Fisher:1978pf,deAlcantaraBonfim:1980pe,deAlcantaraBonfim:1981sy}). The conformal method developed in \cite{Rychkov:2015naa} was not applicable to this theory since they crucially used the conditions that $\gamma_1=O(\epsilon^2)$ and $\gamma_{n>1} =O(\epsilon)$. However our method does not rely on these condition and can be widely applied.
\section{$\phi_i \phi_i \sigma $ interaction in $(6-\epsilon)$ dimensions}
This section generalizes the $\phi^3$-theory in $d=(6-\epsilon)$ dimensions by adding the $\phi_i \phi_i \chi $ interaction. We will also include the $\chi^3$ term.
\begin{align}
S=\int d^dx \, \frac{1}{2} (\partial \phi_i)^2 +\frac{1}{2} (\partial \chi)^2 +\frac{g_1}{2} \chi \phi_i \phi_i +\frac{g_2}{3!} \chi^3 ,~~~i=1,\cdots ,N
\end{align}
The theory is recently studied in \cite{Fei:2014yja,Giombi:2016hkj} (for an old reference, see for example, \cite{Mikhailov:1985cm}).
The $\beta$ function of $g_1$ is very similar to the one of the four-dimensional QCD. When we limit to the case with $g_2=0$ and $\epsilon =0$, the theory is asymptotically free for $N<8$ and we have the IR free phase for $N >8$, . For $N=8$, the $\beta$ function is vanishing at least from the one-loop analysis and it would show the conformal symmetry and we can use the conformal method.
The theory can also have the Lee-Yang zero for the imaginary values of the couplings. For the large $N$ limit, the Wilson-Fisher fixed point is realized at the real values of the couplings.
The classical equations of motion for $\phi_i$ and $\chi$ are
\begin{align}
\Box \phi_i &=g_1 \chi \phi_i,~~~~~\Box \chi =\frac{g_1}{2} \phi_k \phi_k +\frac{g_2}{2} \chi^2
\end{align}
The scaling dimensions are defined as
\begin{align}
\Delta_{\phi} &=(2 -\frac{\epsilon}{2}) +\gamma_\phi,~~~\Delta_{\chi} ~=(2 -\frac{\epsilon}{2}) +\gamma_\chi
\end{align}
\subsubsection*{Two-point function}
The theory includes the two types of the scalar fields, $\phi_i$ and $\chi$, then we have the two propagators.
\begin{align}
\braket{\phi_i(x) \phi_j (y)} &=c_{\phi} \delta_{ij} |x-y|^{-2 \Delta_{\phi}} \\
\braket{\chi (x) \chi(y) } &=c_{\chi} |x-y|^{-2 \Delta_{\chi}}
\end{align}
Applying the two differential operators, $\Box_x$ and $\Box_y$, we obtain the following result.
\begin{align}
\gamma_\phi &=\frac{1}{6} \frac{g_1^2 }{(4 \pi)^3} \label{gammaphi} \\
\gamma_\chi &= \frac{1}{12} \left( \frac{g_1^2 N}{(4 \pi)^3} +\frac{g_2^2}{(4 \pi)^3} \right), \label{gammachi}
\end{align}
where we used the classical equations of motion. This is precisely identical to the one-loop calculation.
\subsubsection*{Three-point function}
Next, let us study the three-point functions which effectively determine the $\beta$ functions at the leading order and the critical values of the couplings.
The analysis of the three-point function is more subtle since now the theory has the two types of the scalar fields and the equations of motion should involves the various terms.
We start with the correlation function, $\braket{ \phi_i (x) \phi_j (y) \chi (z)}$. By the assumption that the fields $\phi_i$ and $\chi$ are the primary fields, the coordinate dependence is completely fixed as
\begin{align}
\braket{ \phi_i (x) \phi_j (y) \chi (z)} =C_{ij \chi} |x-y|^{\Delta_{\chi} -2\Delta_{\phi}} |y-z|^{-\Delta_{\chi}} |x-z|^{-\Delta_{\chi}},
\end{align}
where the three-point function above is classically vanishing, then the coefficient is $O(g_1)$. We will first determine the leading behavior of $C_{ij \chi}$. We multiply the above function by the differential operator $\Box_z$, we find
\begin{align}
\braket{ \phi_i (x) \phi_j (y) \Box_z \chi (z)} &=C_{ij \chi} \Box_z |x-y|^{\Delta_{\chi} -2\Delta_{\phi}} |y-z|^{-\Delta_{\chi}} |x-z|^{-\Delta_{\chi}} \nonumber \\
&\sim -4 C_{ij \chi} |y-z|^{-4} |x-z|^{-4} \\
\braket{ \phi_i (x) \phi_j (y) \Box_z \chi (z)} &=\braket{\phi_i (x) \phi_j (y) \left( \frac{g_1}{2} \phi_k \phi_k +\frac{g_2}{2} \chi^2 \right)} \nonumber \\
&\sim \frac{g_1 \delta_{ij}}{(4 \pi^3)^2} |y-z|^{-4}|x-z|^{-4},
\end{align}
where in the third line above, we used the classical equation of motion for $\chi$. Comparing the second and fourth lines, the coefficient is determined to be
\begin{align}
\left. C_{ij \chi} \right|_{\mathrm{lowest}} =-\frac{g_1}{64 \pi^6} \delta_{ij}
\end{align}
at the leading order of the perturbation. The similar manipulation for the three-point function $\braket{ \chi (x) \chi (y) \chi (z)} $ gives
\begin{align}
\left. C_{\chi \chi \chi} \right|_{\mathrm{lowest}} =-\frac{g_2}{64 \pi^6}.
\end{align}
Next we will study the correlation functions which involve the descendant field. The form of the correlation function is fixed by the equation of motion and the three-point function of the primary fields.
\begin{align}
\braket{ \phi_i (x) \phi_j (y) (\phi_k \phi_k) (z)} &= \frac{2}{g_1} \braket{ \phi_i (x) \phi_j (y) (\Box \chi -\frac{g_2}{2} \chi^2 ) (z)} \nonumber \\
&= \frac{2}{g_1}\Box_z \braket{\phi_i (x) \phi_j (y) \chi(z)} -\frac{g_2}{g_1} \braket{ \phi_i (x) \phi_j (y) \chi^2 (z)}
\end{align}
We multiply the above expression by the derivatives $\Box_x$ and $\Box_y$,
\begin{align}
\braket{ \Box \phi_i (x) \Box \phi_j (y) (\phi_k \phi_k) (z)} &= \frac{2}{g_1}C_{ij\chi} \Box_x \Box_y \Box_z |x-y|^{\Delta_{\chi} -2\Delta_{\phi}}|y-z|^{-\Delta_{\chi}} |x-z|^{-\Delta_{\chi}} \nonumber \\
& \qquad \quad -\frac{g_2}{g_1} \braket{ \Box\phi_i (x)\Box \phi_j (y) \chi^2 (z)} \label{pppp}
\end{align}
The left-hand side at the leading order can be evaluated by the classical equation of motion as
\begin{align}
\braket{ \Box \phi_i (x) \Box \phi_j (y) (\phi_k \phi_k) (z)}
&=g_1^2\braket{ \chi \phi_i(x) \chi \phi_j (y) (\phi_k \phi_k) (z) } \nonumber \\
&\sim \frac{2g_1^2 \delta_{ij}}{64 \pi^9} |x-y|^{-4}|y-z|^{-4} |x-z|^{-4}
\end{align}
The right-hand side in \eqref{pppp} is also evaluated as foliows.
\if0
\begin{align}
&\Box_z |x-y|^{\Delta_{\chi} -2\Delta_{\phi}}|y-z|^{-\Delta_{\chi}} |x-z|^{-\Delta_{\chi}} \nonumber \\
&=\Delta_{\chi}(2 \Delta_{\chi} +2-d) |x-y|^{\Delta_{\chi} -2\Delta_{\phi}} |y-z|^{-\Delta_{\chi}-2} |x-z|^{-\Delta_\chi} \nonumber \\
&\qquad +\Delta_{\chi}(2 \Delta_{\chi} +2-d) |x-y|^{\Delta_{\chi} -2\Delta_{\phi}} |y-z|^{-\Delta_{\chi}}|x-z|^{-\Delta_{\chi}-2} \nonumber \\
&\qquad \qquad-\Delta_\chi^2 |x-y|^{\Delta_{\chi} -2\Delta_{\phi}+2} |y-z|^{-\Delta_{\chi}-2} |x-z|^{-\Delta_\chi-2} \nonumber \\
&\sim 4 \gamma_\chi |x-y|^{-2} |y-z|^{-4} |x-z|^{-2} +4 \gamma_\chi |x-y|^{-2} |y-z|^{-2} |x-z|^{-4} \nonumber \\
& \qquad -\Delta_\chi^2 |x-y|^{\Delta_{\chi} -2\Delta_{\phi}+2} |y-z|^{-\Delta_{\chi}-2} |x-z|^{-\Delta_\chi-2}
\end{align}
\fi
\begin{align}
& \frac{2}{g_1}C_{ij\chi} \Box_x \Box_y \Box_z |x-y|^{\Delta_{\chi} -2\Delta_{\phi}}|y-z|^{-\Delta_{\chi}} |x-z|^{-\Delta_{\chi}} \nonumber \\
& \qquad \qquad \sim \frac{2 \delta_{ij}}{\pi^6} (2 \gamma_\phi +\gamma_\chi -\frac{\epsilon}{2}) |x-y|^{-4}|y-z|^{-4} |x-z|^{-4}
\end{align}
\begin{align}
-\frac{g_2}{g_1} \braket{ \Box\phi_i (x)\Box \phi_j (y) \chi^2 (z)} &=- g_1g_2 \braket{ \chi \phi_i (x) \chi \phi_i (y) \chi^2 (z) } \nonumber \\
& \sim -\frac{2g_1g_2 \delta_{ij}}{64 \pi^9} |x-y|^{-4}|y-z|^{-4} |x-z|^{-4}
\end{align}
Comparing the both sides we find
\begin{align}
2\gamma_\phi +\gamma_\chi =\frac{\epsilon}{2} + \frac{g_1^2}{64 \pi^3} +\frac{g_1g_2}{64\pi^3}.
\end{align}
If we substitute the results \eqref{gammaphi} and \eqref{gammachi} in the above, we find
\begin{align}
-\frac{\epsilon}{2} +\frac{1}{(4\pi)^3} \frac{(N-8)g_1^2 -12g_1g_2 +g_2^2}{12} =0
\end{align}
This is precisely the same as the $\beta$ function for $g_1$,
\begin{align}
\beta_1 =-\frac{\epsilon}{2}g_1 +\frac{1}{(4\pi)^3} \frac{(N-8) g_1^3 -12g_1^2g_2 +g_1g_2^2 }{12}
\end{align}
and it is consistent with the fact that for $g_2=0$, $N=8$ and $\epsilon=0$, the theory is conformal at the leading order of the perturbation.
Next, we would like to find the $\beta$ function for $g_2$. Then it is plausible to study the correlation functions, $\braket{\chi(x) \chi(y) \chi^2 (z)} $ and $\braket{\chi(x) \chi(y) \chi (z)} $. Since $\chi^2$ is related to $\Box \chi$ and $\phi_k \phi_k$ by the equation of motion, we have the following relation.
\begin{align}
\braket{\chi(x) \chi(y) \chi^2 (z)} &=\frac{2}{g_2} \braket{\chi(x) \chi(y) \left( \Box_z \chi -\frac{g_1}{2} \phi_k \phi_k \right)(z)}\\
\braket{\Box_x\chi(x) \Box_y\chi(y) \chi^2 (z)} &=\frac{2}{g_2}\Box_x \Box_y \Box_z \braket{\chi(x) \chi(y) \chi (z)} -\frac{g_1}{g_2} \braket{\Box_x \chi(x) \Box_y \chi(y) \phi_k \phi_k (z) } \label{ccc}
\end{align}
The left-hand side in \eqref{ccc} can be evaluated as
\begin{align}
\braket{\Box_x\chi(x) \Box_y\chi(y) \chi^2 (z)} &= \braket{\left( \frac{g_1}{2} \phi_k \phi_k +\frac{g_2}{2} \chi^2 \right)(x) \left( \frac{g_1}{2} \phi_k \phi_k +\frac{g_2}{2} \chi^2 \right)(y) \chi^2 (z)} \nonumber \\
&\sim \frac{g_2^2}{4}\braket{\chi^2 (x) \chi^2(y) \chi^2(z)} \nonumber \\
&\sim \frac{2g_2^2}{ 64 \pi^9} |x-y|^{-4}|y-z|^{-4} |x-z|^{-4}.
\end{align}
The two terms in the right-hand side of \eqref{ccc} become
\begin{align}
\frac{2}{g_2}\Box_x \Box_y \Box_z \braket{\chi(x) \chi(y) \chi (z)} &= \frac{2}{\pi^6} (3 \gamma_\chi -\frac{\epsilon}{2}) |x-y|^{-4}|y-z|^{-4} |x-z|^{-4} \\
-\frac{g_1}{g_2} \braket{\Box \chi(x) \Box \chi(y) \phi_k \phi_k (z) } &=-\frac{2Ng_1^3}{g_2} |x-y|^{-4}|y-z|^{-4} |x-z|^{-4},
\end{align}
where we used the general form of the three-point function $\braket{\chi(x) \chi(y) \chi (z)}$, which can be constrained by the conformal symmetries by the assumption that $\chi$ is a primary field. Comparing the both sides we obtain
\begin{align}
\frac{2g_2^2}{64 \pi^3} =-\frac{2Ng_1^3/g_2 }{64 \pi^3} +\frac{2}{\pi^6} (3 \gamma_\chi -\frac{\epsilon}{2}),
\end{align}
which is equivalent to
\begin{align}
-\frac{\epsilon}{2} +\frac{1}{(4\pi)^3} \frac{-4N g_1^3 +Ng_1^2g_2 -3 g_2^3}{4g_2} =0.
\end{align}
Again, this is consistent with the $\beta$ function for $g_2$
\begin{align}
\beta_2 =-\frac{\epsilon}{2}g_2 +\frac{-4N g_1^3 +Ng_1^2g_2 -3 g_2^3}{4 (4 \pi)^3}.
\end{align}
In this way we correctly reproduce the anomalous dimensions for $\phi_i$ and $\chi$ and the conditions that the $\beta$ functions vanish. It seems difficult to study the composite operators $\phi_k \phi_k$ and $\chi^2$ since these two operators mix with each other and solving the mixing will inevitably require the one-loop computation. It would be interesting to calculate the mixing matrix from the tree-level calculation and this direction will be left as a future direction.
\section{Summary and Discussion}
Motivated by the recent paper \cite{Rychkov:2015naa}, we here derived the leading expression of the anomalous dimensions of the operators $\phi^n$ for the $\phi^6$-, $\phi^4$-, and $\phi^3$-theories in $(3-\epsilon), (4-\epsilon)$ and $(6-\epsilon)$ dimensions respectively. The method developed here does not rely on the Feynman diagrammatic technique but on the conformal symmetry and on the classical equations of motion. This is very parallel and similar to the method developed in \cite{Rychkov:2015naa}.
One of the differences is that in \cite{Rychkov:2015naa}, the operator relation, $\Box \phi =\alpha \phi^3$ is not the same as the equation of motion but just an operator identity coming from the assumption that $\phi^3$ is a descendant of the conformal primary field $\phi$ at the Wilson-Fisher fixed point. In this sense, the method in \cite{Rychkov:2015naa} does not rely on the perturbative approach. On the other hand, in this paper, we used the classical equation of motion, which is just the lowest order form of the Schwinger-Dyson equation without the contact term and clearly relying on the Lagrangian-based (perturbative) approach.
The other difference is that in \cite{Rychkov:2015naa} they studied the three-point functions, $\braket{\phi^n(x) \phi^{n+1}(y)\phi(z) } $ and $\braket{\phi^n(x) \phi^{n+1}(y)\phi^3(z) } $, by using the OPE between $\phi^n$ and $\phi^{n+1}$. Then this is effectively equivalent to studying the three-point functions, $\braket{\phi^n(x) \phi^{n+1}(y)\phi(z) } $ and $\braket{\phi^n(x) \phi^{n+1}(y) \Box \phi(z) } $, in our approach. However, as we have mentioned in Section 4, it is important to study the three-point functions, $\braket{\phi(x) \phi(y)\phi^2(z) }$ and $\braket{\Box \phi(x) \Box\phi(y)\phi^2(z) }$ in the $\phi^3$-theory in $(6-\epsilon)$ dimensions. Since the later one includes the two differential operators, the information on $\braket{\Box \phi(x) \Box\phi(y)\phi^2(z) }$ is not taken into account in the method \cite{Rychkov:2015naa}. This is the reason why the method \cite{Rychkov:2015naa} can not be directly applied to the six-dimensional $\phi^3$-theory.
In this paper, we found the critical coupling $g_*=g_*(\epsilon)$ at the leading order without any perturbative input. This can be carried out by considering the two- and three-point functions and their derivatives. The coordinate dependence of these functions is constrained from the conformal symmetry. By combining these results with the classical equations of motion we obtained the anomalous dimensions and the critical coupling. It would be interesting to go beyond the leading order approximation. For example, if we want to know the next-leading order, we will need to do the one-loop computation since the use of the equation of motion reduces the loop order required for calculating the physical quantities. Thus we must inevitably deal with the regularization and the renormalization at this order.
In Section 5, we study the theory with the $N+1$ scalar fields,
\begin{align}
S=\int d^6x \, \frac{1}{2} (\partial \phi_i)^2 +\frac{1}{2} (\partial \chi)^2 +\frac{g_1}{2} \chi \phi_i \phi_i +\frac{g_2}{3!} \chi^3,~~i=1,\cdots ,N
\end{align}
and we found the anomalous dimensions of the operators $\phi_i$ and $\chi$. For the composite operators, we have to be careful with the operator mixing, for example, between $\phi_i\phi_i$ and $\chi^2$. In our approach it seems difficult to solve the mixing only from the tree-level calculation. It would be interesting to find the techniques to avoid the one-loop computation.
It would be important to apply the method to more complicated theories including the various kinds of the scalar and fermion fields \cite{Raju:2015fza,Ghosh:2015opa} and also important to know if the conformal techniques developed here can be applied to the gauge theories.
\section*{Acknowledgments}
I would like to thank Slava Rychkov, Masashi Hayakawa, Dileep Jatkar and Yu Nakayama for valuable comments and helpful discussions.
\bibliographystyle{ieeetr}
|
\section{Introduction} \label{sec:intro}
We consider spectral properties of sparse random matrices. One of the most prominent examples in the class of sparse random matrices is the (centered) adjacency matrix of the Erd\H{o}s--R\'{e}nyi graph on~$N$ vertices, where an edge is independently included in the graph with a fixed probability $p\equiv p(N)$. Introduced in~\cite{ER,ERG2,Gi}, the Erd\H{o}s--R\'{e}nyi graph model $G(N,p)$ serves as a null model in the theory of random graphs and has numerous applications in many fields including network theory. Information about a random graph can be obtained by investigating its adjacency matrix, especially the properties of its eigenvalues and eigenvectors.
The sparsity of a real symmetric $N$ by $N$ random matrix may be measured by the sparsity parameter $q\equiv q(N)$, with $0\le q\le N^{1/2}$, such that the expected number of non-vanishing entries is $q^2$. For example for the adjacency matrices of the Erd\H{o}s--R\'enyi graph we have $q^2 \simeq Np$, while for standard Wigner matrices we have $q=N^{1/2}$. We call a random matrix sparse if $q$ is much smaller than $N^{1/2}$.
For Wigner matrices, one of the fundamental inputs in the proof of universality results is the local semicircle law~\cite{ESY1,ESY2,ESY3,EYY}, which provides an estimate of the local eigenvalue density down to the optimal scale. The framework built on the local law can also help understanding the spectral properties of sparse random matrices~\cite{EKYY1}. However, in contrast to Wigner matrices, the local eigenvalue density for a sparse random matrix depends on its sparsity. For this reason, the universality of the local eigenvalue statistics for sparse random matrices was proved at first only for $q\ge N^{1/3}$ in~\cite{EKYY1,EKYY2}. Recently, bulk universality was proved in~\cite{HLY} under the much weaker condition $q \geq N^{\epsilon}$, for any $\epsilon > 0$. The main obstacle in the proof of the edge universality is that the local law obtained in~\cite{EKYY1} deteriorates at the edge of the spectrum.
Our first main results is a local law for sparse random matrices up to the edge. More precisely, we show a local law for the eigenvalue density in the regime $q\ge N^\epsilon$, for arbitrarily small $\epsilon>0$. The main observation is that, although the empirical spectral measure of sparse random matrices converges in the large $N$ limit to the semicircle measure, there exists a deterministic correction term that is not negligible for large but finite $N$. As a result, we establish a local law that compares the empirical spectral measure not with the semicircle law but with its refinement. (See Theorem~\ref{theorem:local} and Corollary~\ref{corollary density of states} for more detail.)
The largest eigenvalue $\mu_1$ of a real symmetric $N$ by $N$ Wigner matrix (whose entries are centered and have variance $1/N$) converges almost surely to two under the finite fourth-moment condition, and $N^{2/3}(\mu_1 -2)$ converges in distribution to the GOE Tracy--Widom law. For sparse random matrices the refinement of the local semicircle law reveals that the eigenvalues at the upper edge of the spectrum fluctuate around a deterministic number larger than two, and the shift is far greater than $N^{-2/3}$, the typical size of the Tracy--Widom fluctuations.
Our second main result is the edge universality that states the limiting law for the fluctuations of the rescaled largest eigenvalues of a (centered) sparse random matrices is given by the Tracy--Widom law when the shift is taken into consideration and if $q\ge N^{1/6+\epsilon}$, where $\epsilon>0$ is arbitrarily small. We expect the exponent one-sixth to be critical. (See Theorem~\ref{thm tw} and the discussion below it for more detail.) For the adjacency matrices of the Erd\H{o}s--R\'{e}nyi graphs, the sparsity conditions corresponds to $p\ge N^{-2/3 + \epsilon}$, for any $\epsilon > 0$, and our result then assures that the rescaled second largest eigenvalue has GOE Tracy--Widom fluctuations; see Corollary~\ref{cor: TW for A}.
In the proof of the local law, we introduce a new method based on a recursive moment estimate for the normalized trace $m$ of the Green function, \ie we recursively control high moments of $|P(m)|$, for some polynomial $P$, by using lower moments of $|P(m)|$, instead of fully expanding all powers of~$m$; see Section~\ref{sec:outline} for detail. This recursive computation relies on cumulant expansions which were used in the random matrix theory literature many times, especially in the study of linear eigenvalue statistics~\cite{KKP,LP}.
Our proof of the Tracy--Widom limit of the extremal eigenvalues relies on the Green function comparison method~\cite{EYY2,EYY}. However, instead of applying the conventional Lindeberg replacement approach, we use a continuous flow that interpolates between the sparse random matrix and the Gaussian Orthogonal Ensemble (GOE). The main advantage of using a continuous interpolation is that we may estimate the rate of change of $m$ along the flow even if the moments of the entries in the sparse matrix are significantly different from those of the entries in the GOE matrix. The change of $m$ over time is offset by the shift of the edge. A similar idea was used in the proof of edge universality of other random matrix models in~\cite{LS14,LS14b}.
This paper is organized as follows: In Section~\ref{sec:results}, we define the model, present the main results and outline applications to adjacency matrices of the Erd\H{o}s--R\'{e}nyi graph ensemble. In Section~\ref{sec:outline}, we explain the main strategy of our proofs. In Section~\ref{sec:rho}, we prove several properties of the deterministic refinement of Wigner's semicircle law. In Section~\ref{sec:proof main}, we prove the local law using our technical result on the recursive moment estimate, Lemma~\ref{lem:claim}. In Section~\ref{sec:stein}, we prove Lemma~\ref{lem:claim} with technical detail. In Section~\ref{sec:tw}, we prove our second main result on the edge universality using the Green function comparison method.
{\it Notational conventions:}
We use the symbols $O(\,\cdot\,)$ and $o(\,\cdot\,)$ for the standard big-O and little-o notation. The notations $O$, $o$, $\ll$, $\gg$, refer to the limit $N\to \infty$ unless otherwise stated. Here $a\ll b$ means $a=o(b)$. We use~$c$ and~$C$ to denote positive constants that do not depend on $N$, usually with the convention $c\le C$. Their value may change from line to line. We write $a\sim b$, if there is $C\ge1$ such that $C^{-1}|b|\le |a|\le C |b|$. Throughout the paper we denote for $z\in\C^+$ the real part by $E=\re z$ and the imaginary part by $\eta=\im z$. For $a\in\R$, we let $(a)_+=\max (0,a)$, and $(a)_-= -\min(a,0)$. Finally, we use double brackets to denote index sets, \ie for $n_1,n_2\in\mathbb{R}$, $\llbracket n_1,n_2\rrbracket:=[n_1,n_2]\cap \mathbb{Z}$.
{\it Acknowledgement:} We thank L\'aszl\'o Erd\H{o}s for useful comments and suggestions. Ji Oon Lee is grateful to the department of mathematics, University of Michigan, Ann Arbor, for their kind hospitality during the academic year 2014--2015.
\section{Definitions and main results} \label{sec:results}
\subsection{Motivating examples}\label{subsection motivating examples}
\subsubsection{Adjacency matrix of Erd\H os--R\'enyi graph}\label{le subsection adjacency matrix ER graph}
One motivation for this work is the study of adjacency matrices of the {\it Erd\H os--R\'enyi random graph model} $G(N,p)$. The off-diagonal entries of the adjacency matrix associated with an Erd\H os--R\'enyi graph are independent, up to the symmetry constraint, Bernoulli random variables with parameter $p$, \ie the entries are equal to $1$ with probability~$p$ and~$0$ with probability $1-p$. The diagonal entries are set to zero, corresponding to the choice that the graph has no self-loops. Rescaling this matrix ensemble so that the bulk eigenvalues typically lie in an order one interval we are led to the following random matrix ensemble. Let $A$ be a real symmetric $N\times N$ matrix whose entries, $A_{ij}$, are independent random variables (up to the symmetry constraint $A_{ij}=A_{ji}$) with distributions
\begin{align}\label{le probability AER}
\P \Big( A_{ij}=\frac{1}{\sqrt{Np(1-p)}}\Big)=p\,, \quad\qquad \P( A_{ij}=0)= 1-p\,,\qquad\quad\P ( A_{ii}=0)=1\,,\qquad (i\not=j)\,.
\end{align}
Note that the matrix $ A$ typically has $N(N-1)p$ non-vanishing entries. For our analysis it is convenient to extract the mean of the entries of $A$ by considering the matrix $\wt A$ whose entries, $\wt A_{ij}$, have distribution
\begin{align*}
\p \Big( \wt A_{ij} = \frac{1-p}{\sqrt{Np(1-p)}} \Big) = p\,,\qquad \p \Big( \wt A_{ij} = -\frac{p}{\sqrt{Np(1-p)}} \Big) = 1-p\,, \qquad \p\big( \wt A_{ii}=0\big)=1\,,
\end{align*}
with $i\not=j$. A simple computation then reveals that
\begin{align}\label{assumption ER1}
\E \wt A_{ij} = 0\,,\qquad \qquad \E \wt A_{ij}^2 =\frac{1}{ N}\,,
\end{align}
and
\begin{align}\label{assumption ER2}
\E \wt A_{ij}^k = \frac{(-p)^k (1-p) + (1-p)^k p}{(Np(1-p))^{k/2}} =\frac{1}{ N d^{(k-2)/2}}(1 + O(p))\,,\qquad\qquad (k\ge 3)\,,
\end{align}
with $i\not=j$, where $d\deq pN$ denotes the~{\it expected degree} of a vertex, which we allow to depend on $N$. As already suggested by~\eqref{assumption ER2}, we will assume that $p\ll 1$.
\subsubsection{Diluted Wigner matrices}\label{sec:diluted w}
Another motivation for this work are {\it diluted Wigner matrices}. Consider the matrix ensemble of real symmetric $N\times N$ matrices of the form
\begin{align}
D_{ij}=B_{ij}V_{ij} \,,\qquad\qquad (1\le i\le j\le N)\,,
\end{align}
where $(B_{ij}:i\le j)$ and $(V_{ij}:i\le j)$ are two independent families of independent and identically distributed random variables. The random variables $(V_{ij})$ satisfies $\E V_{ij}^2=1$ and $\E V_{ij}^{2k}\le (Ck)^{ck}$, $k\ge 4$, for some constants $c$ and $C$, and their distribution is, for simplicity, often assumed to be symmetric. The random variables $B_{ij}$ are chosen to have a Bernoulli type distribution given by
\begin{align}
\P \Big( B_{ij} = \frac{1}{\sqrt{Np}} \Big) =p\,, \qquad \p ( B_{ij} = 0) =1- p\,,\qquad \p(B_{ii}=0)=1\,,
\end{align}
with $i\not=j$. We introduce the {\it sparsity parameter $q$} through
\begin{align}\label{le q0}
p=\frac{q^2}{N}\,,
\end{align}
with $0<q\le N^{1/2}$. We allow $q$ to depend on $N$. We refer to the random matrix $D=(D_{ij})$ as a {\it diluted Wigner matrix} whenever $q\ll N^{1/2}$. For $q=N^{1/2}$, we recover the usual Wigner ensemble.\newpage
\subsection{Notation}
In this subsection we introduce some of the notation and conventions used.
\subsubsection{Probability estimates}
We first introduce a suitable notion for high-probability estimates.
\begin{definition}[High probability event]
We say that an $N$-dependent event $\Xi \equiv \Xi^{(N)}$ holds with high probability if, for any (large) $D>0$,
\begin{align}
\p\big(\Xi^{(N)}\big) \geq 1 - N^{-D}\,,
\end{align}
for sufficiently large $N\ge N_0(D)$.
\end{definition}
\begin{definition}[Stochastic domination] \label{def:domination}
Let $X\equiv X^{(N)}$, $Y\equiv Y^{(N)}$ be $N$-dependent non-negative random variables. We say that $Y$ stochastically dominates $X$ if, for all (small) $\epsilon>0$ and (large)~$D>0$,
\begin{align}
\P\big(X^{(N)}>N^{\epsilon} Y^{(N)}\big)\le N^{-D}\,,
\end{align}
for sufficiently large $N\ge N_0(\epsilon,D)$, and we write $X \prec Y$.
When
$X^{(N)}$ and $Y^{(N)}$ depend on a parameter $u\in U$ (typically an index label or a spectral parameter), then $X(u) \prec Y (u)$, uniformly in $u\in U$, means that the threshold $N_0(\epsilon,D)$ can be chosen independently of $u$. A slightly modified version of stochastic domination appeared first in~\cite{EKY}.
\end{definition}
In Definition~\ref{def:domination} and hereinafter we implicitly choose $\epsilon>0$ strictly smaller than $\phi/10>0$, where $\phi>0$ is the fixed parameter appearing in~\eqref{le phi} below.
The relation $\prec$ is a partial ordering: it is transitive and it satisfies the arithmetic rules of an order relation, {\it e.g.}, if $X_1\prec Y_1$ and $X_2\prec Y_2$ then $X_1+X_2\prec Y_1+Y_2$ and $X_1 X_2\prec Y_1 Y_2$. Furthermore, the following property will be used on a few occasions: If $\Phi(u)\ge N^{-C}$ is deterministic, if $Y(u)$ is a nonnegative random variable satisfying $\E [Y(u)^2]\le N^{C'}$ for all~$u$, and if $Y(u) \prec \Phi(u)$ uniformly in~$u$, then, for any $\epsilon>0$, we have $\E [Y(u)] \le N^\epsilon \Phi(u)$ for $N\ge N_0(\epsilon)$, with a threshold independent of $u$. This can easily be checked since
$$
\E [Y(u) \lone(Y(u) > N^{\epsilon/2} \Phi)] \leq \left(\E[ Y(u)^2] \right)^{1/2} \big( \p[ Y(u) > N^{\epsilon/2} \Phi] \big)^{1/2} \leq N^{-D}\,,
$$
for any (large) $D > 0$, and $\E [Y(u) \lone(Y(u) \leq N^{\epsilon/2} \Phi(u))] \leq N^{\epsilon/2} \Phi(u)$, hence $\E [Y(u)] \leq N^{\epsilon} \Phi(u)$.
\subsubsection{Stieltjes transform} Given a probability measure $\nu$ on $\R$, we define its Stieltjes transform as the analytic function $m_\nu\,:\,\C^+\rightarrow \C^+$, with $\C^+\deq\{ z=E+\ii\eta\,:\, E\in\R, \eta>0\}$, defined by
\begin{align}
m_{\nu}(z)\deq\int_\R\frac{\dd\nu(x)}{x-z}\,,\qquad\qquad (z\in\C^+)\,.
\end{align}
Note that $\lim_{\eta\nearrow 0}\ii \eta \, m_{\nu}(\ii\eta)=-1$ since $\nu$ is a probability measure. Conversely, if an analytic function $m\,:\,\C^+\rightarrow \C^+$ satisfies $\lim_{\eta\nearrow 0}\ii \eta \, m(\ii\eta)=-1$, then it is the Stieltjes transform of a probability~measure.
Choosing $\nu$ to be the standard semicircle law with density $\frac{1}{2\pi}\sqrt{4-x^2}$ on $[-2,2]$, on easily shows that $m_{\nu}$, for simplicity hereinafter denoted by $\msc$, is explicitly given by
\begin{align}\label{le msc}
\msc(z) = \frac{-z + \sqrt{z^2 -4}}{2}\,, \qquad \qquad(z \in \C^+)\,,
\end{align}
where we choose the branch of the square root so that $\msc(z) \in \C^+$, $z\in\C^+$. It directly follows that
\begin{align}\label{le sce for msc}
1 + z \msc(z) + \msc(z)^2 = 0\,, \qquad \qquad(z \in \C^+)\,.
\end{align}
\subsection{Main results}
In this section we present our main results. We first generalize the matrix ensembles derived from the Erd\H os--R\'enyi graph model and the diluted Wigner matrices in Section~\ref{subsection motivating examples}.
\begin{assumption} \label{assumption H}
Fix any small $\phi>0$. We assume that $H = (H_{ij})$ is a real symmetric $N \times N$ matrix whose diagonal entries are almost surely zero and whose off-diagonal entries are independent, up to the symmetry constraint $H_{ij} = H_{ji}$, identically distributed random variables. We further assume that $(H_{ij})$ satisfy the moment conditions
\begin{align}\label{le moment condition}
\E H_{ij} = 0\,, \qquad \E (H_{ij})^2 = \frac{1-\delta_{ij}}{N}\,, \qquad \E |H_{ij}|^k \leq \frac{(Ck)^{ck}}{N q^{k-2}}\,, \quad\qquad (k\ge 3)\,,
\end{align}
with sparsity parameter $q$ satisfying
\begin{align}\label{le phi}
N^{\phi}\le q\le N^{1/2}\,.
\end{align}
\end{assumption}
We assume that the diagonal entries satisfy $H_{ii}=0$ a.s., yet this condition can easily be dropped. For the choice $\phi=1/2$ we recover the {\it real symmetric Wigner ensemble} (with vanishing diagonal). For the rescaled adjacency matrix of the Erd\H{o}s--R\'enyi graph, the sparsity parameter~$q$, the edge probability~$p$ and the expected degree of a vertex $d$ are linked by $q^2=pN=d$.
We denote by $\kappa^{(k)}$ the $k$-th {\it cumulant} of the i.i.d.\ random variables $(H_{ij}:i<j)$. Under Assumption~\ref{assumption H} we have $\kappa^{(1)}=0$, $\kappa^{(2)}=1/N$, and
\begin{align}\label{le cumulant bound}
|\kappa^{(k)}| \leq \frac{(2Ck)^{2(c+1)k}}{N q^{k-2}}\,, \quad\qquad (k\ge 3)\,.
\end{align}
We further introduce the {\it normalized cumulants}, $s^{(k)}$, by setting
\begin{align}\label{normalized cumulants}
s^{(1)}\deq 0\,,\qquad\quad s^{(2)}\deq 1\,,\qquad\quad s^{(k)}\deq Nq^{k-2}\kappa^{(k)}\,,\qquad(k\ge 3)\,.
\end{align}
In case $H$ is given by the centered adjacency matrix~$\wt A$ introduced in~Subsection~\ref{le subsection adjacency matrix ER graph}, we have that $\nm^{(k)}=1+O(d/N)$, $k\ge 3$, as follows from~\eqref{assumption ER2}.
We start with the local law for the Green function of this matrix ensemble.
\subsubsection{Local law up to the edges for sparse random matrices}
Given a real symmetric matrix $H$ we define its Green function, $G^H$, and the normalized trace of its Green function, $m^H$, by setting
\begin{align}\label{se green functions}
G^H(z)\deq \frac{1}{H-z\id}\,,\qquad\quad m^H(z)\deq\frac{1}{N}\mathrm{Tr} \,G^H(z)\,,\qquad\qquad (z\in\C^+)\,.
\end{align}
The matrix entries of $G^H(z)$ are denoted by $G^{H}_{ij}(z)$. In the following we often drop the explicit $z$-dependence from the notation for $G^H(z)$ and $m^H(z)$.
Denoting by $\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_N$ the ordered eigenvalues of $H$, we note that $m^H$ is the Stieltjes transform of the empirical eigenvalue distributions, $\mu^H$, of $H$ given by
\begin{align}
\mu^H\deq\frac{1}{N}\sum_{i=1}^N \delta_{\lambda_i}\,.
\end{align}
We further introduce the following domain of the upper-half plane
\begin{align}\label{second domain}
\caE\deq \{ z = E+ \ii \eta \in \C^+ : |E| < 3, \, 0< \eta \le 3 \}\,.
\end{align}
Our first main result is the local law for $m^H$ up to the spectral edges.
\begin{theorem} \label{theorem:local}
Let $H$ satisfy Assumption \ref{assumption H} with $\phi>0$. Then, there exists an algebraic function $\wt m : \C^+ \to \C^+$ and $2<L<3$ such that the following hold:
\begin{enumerate}
\item The function $\wt m$ is the Stieltjes transform of a deterministic symmetric probability measure $\wt \rho$, \ie $\wt m(z)=m_{\wt\rho}(z)$. Moreover, $\supp \wt\rho=[-L,L]$ and $\wt\rho$ is absolutely continuous with respect to Lebesgue measure with a strictly positive density on $(-L,L)$.
\item The function $\wt m\equiv \wt m(z)$ is a solution to the polynomial equation
\begin{align}\begin{split} \label{eq:poly la}
P_{ z}(\wt m)& \deq 1 + z \wt m + \wt m^2 + \frac{{ \nmf }}{q^2}\wt m^4=0\,,\qquad\qquad (z\in \C^+)\,.
\end{split}\end{align}
\item The normalized trace $m^H$ of the Green function of $H$
satisfies
\begin{align}\label{le local law}
|m^H (z) - \wt m (z)| \prec \frac{1}{q^2}+\frac{1}{N\eta}\,,
\end{align}
uniformly on the domain $\caE$, $z=E+\ii\eta$.
\end{enumerate}
\end{theorem}
Some properties of $\wt\rho$ and its Stieltjes transform $\wt m$ are collected in Lemma~\ref{lem:w} below.
The local law~\eqref{le local law} implies estimates on the {\it local density of states} of $H$. For $E_1< E_2$ define
\begin{align*}
\frak{n}(E_1,E_2)\deq\frac{1}{N}|\{ i\,:\, E_1<\lambda_i\le E_2\}|\,,\qquad\quad n_{\wt\rho}(E_1,E_2)\deq \int_{E_1}^{E_2}\wt\rho(x)\,\dd x\,.
\end{align*}
\begin{corollary}\label{corollary density of states}
Suppose that $H$ satisfies Assumption \ref{assumption H} with $\phi>0$. Let $E_1,E_2\in\R$, $E_1<E_2$. Then,
\begin{align}\label{le estimate on density of states}
|\frak{n}(E_1,E_2)-n_{\wt\rho}(E_1,E_2)|\prec \frac{E_2-E_1 }{q^2}+ \frac{1}{N}\,.
\end{align}
\end{corollary}
The proof of Corollary~\ref{corollary density of states} from Theorem~\ref{theorem:local} is a standard application of the Helffer-Sj\"ostrand calculus; see e.g., Section 7.1 of~\cite{EKYY13} for a similar argument.
An interesting effect of the sparsity of the entries of $H$ is that its eigenvalues follow, for large~$N$, the deterministic law $\wt\rho$ that depends on the sparsity parameter $q$. While this law approaches the standard semicircle law $\rho_{sc}$ in the limit $N\rightarrow\infty$, its deterministic refinement to the standard semicircular law for finite $N$ accounts for the non-optimality at the edge of results obtained in~\cite{EKYY1}, \ie when~\eqref{le local law} is compared with~\eqref{le EYYY1 1} below.
\begin{proposition}[Local semicircle law, Theorem 2.8 of \cite{EKYY1}] \label{local semiclrcle law}
Suppose that $H$ satisfies Assumption~\ref{assumption H} with $\phi>0$. Then, the following estimates hold uniformly for $z \in \caE$:
\begin{align}\label{le EYYY1 1}
|m^H(z) - \msc(z)| \prec \min \left\{ \frac{1}{q^2 \sqrt{\kappa+\eta}}, \frac{1}{q} \right\} + \frac{1}{N\eta}\,,
\end{align}
where $\msc$ denote the Stieltjes transform of the standard semicircle law, and
\begin{align}\label{le EYYY1 2}
\max_{1 \leq i, j \leq N} |G^H_{ij}(z) - \delta_{ij} \msc(z)| \prec \frac{1}{q} + \sqrt{\frac{\im \msc(z)}{N\eta}} + \frac{1}{N\eta}\,,
\end{align}
where $\kappa\equiv \kappa(z)\deq |E-2|$, $z=E+\ii\eta$.
\end{proposition}
We remark that the estimate~\eqref{le EYYY1 1} is essentially optimal as long as the spectral parameter $z$ stays away from the spectral edges, \eg for energies in the bulk $E\in[-2+\delta,2-\delta]$, $\delta>0$. For the individual Green function entries, $G_{ij}$, we believe that the estimate~\eqref{le EYYY1 2} is already essentially optimal ($\msc$ therein may be replaced by $\wt m$ without changing the error bound). A consequence of Proposition~\ref{local semiclrcle law} is that all eigenvectors of $H$ are completely delocalized.
\begin{proposition}[Theorem~2.16 and Remark~2.18 in~\cite{EKYY1}]\label{cor: delocalization}
Suppose that $H$ satisfies Assumption~\ref{assumption H} with $\phi>0$. Denote by $(\bsu_i^H)$ the $\ell^2$-normalized eigenvectors of $H$.~Then,
\begin{align}
\max_{1\le i\le N} { \|\bsu_i^H\|}_\infty\prec\frac{1}{\sqrt{N}}\,.
\end{align}
\end{proposition}
Using~\eqref{le EYYY1 1} as {\it a priori} input it was proved in~\cite{HLY} that the local eigenvalue statistics in the bulk agree with the local statistics of the GOE, for $\phi>0$; see also~\cite{EKYY2} for $\phi>1/3$.
When combined with a high moment estimates of~$H$ (see Lemma~4.3 in~\cite{EKYY1}), the estimate in~\eqref{le EYYY1 1} implies the following bound on the operator norm~of~$H$.
\begin{proposition}[Lemma 4.4 of \cite{EKYY1}] \label{a priori norm bound}
Suppose that $H$ satisfies Assumption \ref{assumption H} with $\phi>0$. Then,
\begin{align}
| \|H\| -2| \prec \frac{1}{q^2}+\frac{1}{N^{2/3}}\,.
\end{align}
\end{proposition}
The following estimates of the operator norm of $H$ sharpens the estimates of Proposition~\ref{a priori norm bound} by including the deterministic refinement to the semicircle law as expressed by Theorem~\ref{theorem:local}.
\begin{theorem} \label{prop:norm bound}
Suppose that $H$ satisfies Assumption \ref{assumption H} with $\phi>0$. Then,
\begin{align}\label{le estimate on operator norm}
\left|\|H\|-L\right|\prec \frac{1}{q^4} +\frac{1}{N^{2/3}}\,,
\end{align}
where $\pm L$ are the endpoints of the support of the measure $\wt \rho$ given by
\begin{align}\label{le L}
L=2+\frac{\nmf}{q^2}+O(q^{-4})\,.
\end{align}
\end{theorem}
Here and above, we restricted the choice of the sparsity parameter $q$ to the range $N^\phi\le q\le N^{1/2}$ for arbitrary small $\phi>0$. Yet, pushing our estimates and formalism we expect also to cover the range $(\log N)^{A_0\log\log N}\le q\le N^{1/2}$, $A_0\ge 30$, considered in~\cite{EKYY1}. In fact, Khorunzhiy showed for diluted Wigner matrices (\cf Subsection~\ref{sec:diluted w}) that $\|H\|$ converges almost surely to $2$ for $q\gg(\log N)^{1/2}$, while~$\|H\|$ diverges for $q\ll(\log N)^{1/2}$; see Theorem~2.1 and Theorem~2.2 of~\cite{Kho1} for precise statements.
As noted in Theorem~\ref{prop:norm bound}, the local law allows strong statements on the locations of the extremal eigenvalues of $H$. We next discuss implications for the fluctuations of the rescaled extremal eigenvalues.
\subsubsection{Tracy--Widom limit of the extremal eigenvalues}
Let $W$ be a real symmetric Wigner matrix and denote by $\lambda_1^{W}$ its largest eigenvalue. The {\it edge universality} for Wigner matrices asserts that
\begin{align}\label{le TW}
\lim_{N \to \infty} \p \Big( N^{2/3} (\lambda_1^{W} -2) \leq s \Big) = F_1 (s)\,,
\end{align}
where $F_1$ is the Tracy--Widom distribution function~\cite{TW1,TW2} for the GOE. Statement~\eqref{le TW} holds true for the smallest eigenvalue $\lambda_N^{W}$ as well. We henceforth focus on the largest eigenvalues, the smallest eigenvalues can be dealt with in exactly the same way.
The universality of the Tracy--Widom laws for Wigner matrices was first proved in~\cite{SiSo1,So1} for real symmetric and complex Hermitian ensembles with symmetric distributions. The symmetry assumption on the entries' distribution was partially removed in~\cite{PeSo1,PeSo2}. Edge universality without any symmetry assumption was proved in~\cite{TV2} under the condition that the distribution of the matrix elements has subexponential decay and its first three
moments match those of the Gaussian distribution, \ie the third moment of the entries vanish. The vanishing third moment condition was removed in~\cite{EYY}. A necessary and sufficient condition on the entries' distribution for the edge universality of Wigner matrices was given in~\cite{LY}.
Our second main result shows that the fluctuations of the rescaled largest eigenvalue of the sparse matrix ensemble are governed by the Tracy--Widom law, if the sparsity parameter~$q$ satisfies $q\gg N^{1/6}$.
\begin{theorem} \label{thm tw}
Suppose that $H$ satisfies Assumption \ref{assumption H} with $\phi>1/6$. Denote by $\lambda^H_1$ the largest eigenvalue of~$H$. Then,
\begin{align} \label{eq:main}
\lim_{N \to \infty} \p \left( N^{2/3} \big( \lambda^H_1 -L\big)\leq s \right) = F_1 (s)\,,
\end{align}
where $L$ denotes the upper-edge of the deterministic measure $\wt\rho$ given in~\eqref{le L}.
\end{theorem}
The convergence result~\eqref{eq:main} was obtained in Theorem~2.7 of~\cite{EKYY2} under the assumption that the sparsity parameter $q$ satisfies $q\gg N^{1/3}$, \ie $\phi > 1/3$ (and with $2$ replacing $L$).
In the regime $N^{1/6}\ll q\le N^{1/3}$, the deterministic shift of the upper edge by $L-2=O(q^{-2})$ is essential for~\eqref{eq:main} to hold since then $q^{-2} \ge N^{-2/3}$, the latter being the scale of the Tracy--Widom fluctuations. In other words, to observe the Tracy--Widom fluctuations in the regime $N^{1/6}\ll q\le N^{1/3}$ corrections from the fourth moment of the matrix entries' distribution have to be accounted for. This is in accordance with high order moment computations for diluted Wigner matrices in~\cite{Kho2}.
It is expected that the order of the fluctuations of the largest eigenvalue exceeds $N^{-2/3}$ if $q \ll N^{1/6}$. The heuristic reasoning is that, in this regime, the fluctuations of the eigenvalues in the bulk of the spectrum are much larger than $N^{-2/3}$ and hence affect the fluctuations of the eigenvalues at the edges. Indeed, the linear eigenvalue statistics of sparse random matrices were studied in~\cite{BGM,ShTi}. For sufficiently smooth functions $\varphi$, it was shown there that
\begin{align*}
\frac{q}{\sqrt N} \sum_{i=1}^N \varphi(\lambda_i) - \E \bigg[ \frac{q}{\sqrt N} \sum_{i=1}^N \varphi(\lambda_i) \bigg]
\end{align*}
converges to a centered Gaussian random variable with variance of order one. This suggests that the fluctuations of an individual eigenvalue in the bulk are of order $N^{-1/2} q^{-1}$, which is far greater than the Tracy--Widom scale~$N^{-2/3}$ if $q \ll N^{1/6}$.
\begin{remark} \label{k main}
Theorem \ref{thm tw} can be extended to correlation functions of extreme eigenvalues as follows: For any fixed~$k$, the joint distribution function of the first $k$ rescaled eigenvalues converges to that of the GOE, \ie if we denote by $\lambda_1^{\GOE} \geq \lambda_2^{\GOE} \geq \ldots \geq \lambda_N^{\GOE}$ the eigenvalues of a GOE matrix independent of $H$, then
\begin{align} \label{eq: k main}
&\lim_{N \to \infty} \p \left( N^{2/3} \big( \lambda^H_1 -L\big) \leq s_1 \,, N^{2/3} \big( \lambda_2^H - L\big) \leq s_2 \,, \ldots, N^{2/3} \big( \lambda_k^H -L\big) \leq s_k \right)\nonumber \\
&\quad= \lim_{N \to \infty} \p \left( N^{2/3} \big( \lambda_1^{\GOE} - 2 \big) \leq s_1 \,, N^{2/3} \big( \lambda_2^{\GOE} - 2 \big) \leq s_2 \,, \ldots, N^{2/3} \big( \lambda_k^{\GOE} - 2 \big) \leq s_k \right)\,.
\end{align}
\end{remark}
We further mention that all our results also hold for complex Hermitian sparse random matrices with the GUE Tracy--Widom law describing the limiting edge fluctuations.
\subsubsection{Applications to the adjacency matrix of the Erd\H os--R\'enyi graph}
We briefly return to the adjacency matrix~$A$ of the Erd\H os--R\'enyi graph ensemble introduced in Subsection~\ref{le subsection adjacency matrix ER graph}. Since the entries of~$A$ are not centered, the largest eigenvalue $\lambda_1^A$ is an outlier well-separated from the other eigenvalues. Recalling the definition of the matrix~$\wt A$ whose entries are centered, we notice that
\begin{align}\label{le triv shift}
A=\wt A+f\ket{\boldsymbol{e}}\bra{\boldsymbol{e}}-a \id\,,
\end{align}
with $f\deq q(1-q^2/N)^{-1/2}$, $a\deq f/N$ and $\boldsymbol{e}\deq N^{-1/2} (1,1,\ldots,1)^\mathrm{T}\in\R^N$. (Here, $\ket{\boldsymbol{e}}\bra{\boldsymbol{e}}$ denotes the orthogonal projection onto $\boldsymbol{e}$.) The expected degree $d$ and the sparsity parameter $q$ are linked by $$d=pN=q^2.$$ Applying a simple rank-one perturbation formula and shifting the spectrum by $a$ we get from Theorem~\ref{theorem:local} the following corollary whose proof we leave aside.
\begin{corollary}\label{cor:local law A}
Fix $\phi>0$. Let $A$ satisfies~\eqref{assumption ER1} and~\eqref{assumption ER2} with expected degree $N^{2\phi}\le d\le N^{1-2\phi}$. Then
the normalized trace $m^A$ of the Green function of $A$
satisfies
\begin{align}\label{le local law bis}
|m^A (z) - \wt m (z+a)| \prec \frac{1}{d}+\frac{1}{N\eta}\,,
\end{align}
uniformly on the domain $\caE$, where $z=E+\ii\eta$ and $a=\frac{q}{N}{(1-q^2/N)^{-1/2}}$.
\end{corollary}
Let $\lambda^A_1\ge \lambda^A_2\ge \ldots \ge \lambda^A_N$ denote the eigenvalues of $A$. The behavior of the largest eigenvalue $\lambda_1^A$ was fully determined in~\cite{EKYY1}, where it was shown that it has Gaussian fluctuations, \ie
\begin{align}
\sqrt{\frac{N}{2}}(\lambda_1^A-\E\lambda_1^A)\rightarrow\mathcal{N}(0,1)\, ,
\end{align}
in distribution as $N\to\infty$, with $\E\lambda_1^A=f-a+\frac{1}{f}+O({q^{-3}})$; see Theorem~6.2 in~\cite{EKYY1}.
Combining Theorem~\ref{thm tw} with the reasoning of Section~6 of~\cite{EKYY2}, we have the following corollary on the behavior of the second largest eigenvalue $\lambda_2^A$ of the adjacency matrix $A$.
\begin{corollary}\label{cor: TW for A}
Fix $\phi>1/6$ and $\phi'>0$. Let $A$ satisfies~\eqref{assumption ER1} and~\eqref{assumption ER2} with expected degree $N^{2\phi}\le d\le N^{1-2\phi'}$. Then,
\begin{align} \label{eq:main A}
\lim_{N \to \infty} \p \left( N^{2/3} \big( \lambda^A_2 - L-a\big)\leq s \right) = F_1 (s)\,,
\end{align}
where $L=2+d^{-1}+O(d^{-2})$ is the upper edge of the measure $\wt \rho$; see~\eqref{le L}.
\end{corollary}
We skip the proof of Corollary~\ref{cor: TW for A} from Theorem~\ref{thm tw}, since it is essentially the same as the proof of Theorem~2.7 in~\cite{EKYY2}, where the result was obtained for $\phi>1/3$, with $L$ replaced by $2$. In analogy with Remark~\ref{k main}, the convergence result in~\eqref{eq:main A} extends in an obvious way to the eigenvalues $\lambda_{k-1}^A,\ldots,\lambda_2^A$, for any fixed $k$. The analogous results apply to the $k$-smallest eigenvalues of~$A$. We leave the details to the interested reader.
\begin{remark} \label{community}
The largest eigenvalues of sparse random matrices, especially the (shifted and normalized) adjacency matrices of the Erd\H os--R\'enyi graphs, can be used to determine the number of clusters in automated community detection algorithms~\cite{BS, Lei} in stochastic block models. Corollary~\ref{cor: TW for A} suggests that the test statistics for such algorithms should reflect the shift of the largest eigenvalues if $N^{-2/3} \ll p \ll N^{-1/3}$, or equivalently, $N^{1/3} \ll d \ll N^{2/3}$. If $p \ll N^{-2/3}$, the test based on the edge universality of random matrices may fail as we have discussed after Theorem~\ref{thm tw}.
In applications, the sparsity should be taken into consideration due to small $N$ even if $p$ is reasonably large. For example, in the Erd\H os--R\'enyi graph with $N \simeq 10^3$, the deterministic shift is noticeable if $p \simeq 0.1$, which is in the vicinity of the parameters used in numerical experiments in~\cite{BS, Lei}.
\end{remark}
\section{Strategy and outline of proofs} \label{sec:outline}
In this section, we outline the strategy of our proofs. We begin with the local law of Theorem~\ref{theorem:local}.
\subsection{Wigner type matrices}
We start by recalling the approach initiated in~\cite{ESY1,ESY2,ESY3} for Wigner matrices. Using Schur's complement (or the Feshbach formula) and large deviation estimates for quadratic forms by Hanson and Wright~\cite{HW}, one shows that the normalized trace $m^W(z)$ approximately satisfies the equation $1+zm^W(z)+m^W(z)^2\simeq 0$, with high probability, for any~$z$ in some appropriate subdomain of $\mathcal{E}$. Using that $\msc$ satisfies~\eqref{le sce for msc}, a local stability analysis then yields $|m^W(z)-m_{sc}(z)|\prec (N\eta)^{-1/2}$, $z\in\mathcal{E}$. In fact, the same quadratic equation is approximately satisfied by each diagonal element of the resolvent, $G_{ii}^W$, and not only by their average $m^W$. This observation and an extension of the stability analysis to vectors instead of scalars then yields the {\it entry-wise local law}~\cite{EYY2,EYY3,EYY}, $|G_{ij}^W(z)-\delta_{ij} \msc(z)|\prec (N\eta)^{-1/2}$, $z\in\mathcal{E}$. (See Subsection~\ref{Example: Local law for the GOE} for some details of this argument.) Taking the normalized trace of the Green function, one expects further cancellations of fluctuations to improve the bound. Exploring the {\it fluctuation averaging mechanism} for $m^W$ and refining the local stability analysis, one obtains the {\it strong local law} up to the spectral edges~\cite{EYY}, $|m^W(z)-\msc(z)|\prec(N\eta)^{-1}$, $z\in\mathcal{E}$. It was first introduced in~\cite{EYY3} and substantially extended in~\cite{EKY,EKYY13} to generalized Wigner matrices. We refer to~\cite{E,EKYY13} for reviews of this general approach. Parallel results were obtained in~\cite{TV1,TV2}. For more recent developments see~\cite{AEK,BES15b,BKY,CMS,GNT,LS}.
The strategy outlined in the preceding paragraph was applied to sparse random matrices in~\cite{EKYY1}. The sparsity of the entries manifests itself in the large deviation estimate for quadratic forms, \eg letting $(H_{ij})$ satisfy~\eqref{le moment condition} and choosing $(B_{ij})$ to be any deterministic $N\times N$ matrix,~Lemma~3.8 of~\cite{EKYY1} assures that
\begin{align}\label{le LDE}
\Big|\sum_{k,l}H_{ik}B_{kl}H_{li}-\frac{1}{N}\sum_{k=1}^NB_{kk}\Big|\prec\frac{{\max_{k,l}} |B_{kl}|}{q}+\Big(\frac{1}{N^2}\sum_{k,l} |B_{kl}|^2\Big)^{1/2}\,,
\end{align}
for all $i\in\llbracket 1,N\rrbracket$. Using the above ideas the entry-wise local law in~\eqref{le EYYY1 2} was obtained in~\cite{EKYY1}. Exploiting the fluctuation averaging mechanism for the normalized trace of the Green function, an additional power of $q^{-1}$ can be gained, leading to~\eqref{le EYYY1 1} with the deteriorating factor $(\kappa+\eta)^{-1/2}$.
To establish the local law for the normalized trace of the Green function which does not deteriorate at the edges, we propose in this paper a novel recursive moment estimate for the Green function. When applied to the proof of the strong local law for a Wigner matrix, it is estimating $\E|1+zm^W(z)+m^W(z)^2|^D$ by the lower moments $\E|1+zm^W(z)+m^W(z)^2|^{D-l}$, $l\ge 1$. The use of recursive moment estimate has three main advantages over the previous fluctuation averaging arguments: (1) it is more convenient in conjunction with the cumulant expansion in Lemma~\ref{le stein lemma}, (2) it is easier to track the higher order terms involving the fourth and higher moments if needed, and (3) it does not require to fully expand the higher power terms and thus simplifies bookkeeping and combinatorics. The same strategy can also be applied to individual entries of the Green function by establishing a recursive moment estimate for $\E|1+zG^W_{ii}(z)+\msc(z) G^W_{ii}(z)|^D$ and $\E|G_{ij}|^D$ leading to the entry-wise local law.
We illustrate this approach for the simple case of the GOE next.
\subsection{Local law for the GOE}\label{Example: Local law for the GOE}
Choose $W$ to be a GOE matrix. Since $\msc\equiv \msc(z)$, the Stieltjes transform of the semicircle law, satisfies $1+z\msc+\msc^2=0$, we expect that moments of the polynomial $1+zm(z)+m(z)^2$, with $m\equiv m^W(z)$ the normalized trace of the Green function~$G\equiv G^W(z)$, are small. We introduce the subdomain $\caD$ of $\caE$ by setting
\begin{align} \label{domain}
\caD \deq \{ z = E+ \ii \eta \in \C^+ : |E| < 3, \, N^{-1} < \eta < 3 \}\,.
\end{align}
We are going to derive the following {\it recursive moment estimate} for $m$. For any $D\ge 2$,
\begin{align}
&\E [|1 + zm + m^2|^{2D}]\nonumber \\
&\qquad\leq \E \left[ \frac{\im m}{N\eta} |1 + zm + m^2|^{2D-1} \right] + (4D-2) \E \left[ \frac{\im m}{(N\eta)^2} |z+2m| \cdot |1 + zm + m^2|^{2D-2} \right]\,,\label{example GOE}
\end{align}
for $z\in\C^+$. Fix now $z\in \caD$. Using Young's inequality, the second order Taylor expansion of $m(z)$ around $\msc(z)$ and the {\it a priori} estimate $|m(z)-\msc(z)|\prec 1$, we conclude with Markov's inequality from~\eqref{example GOE} that
\begin{align}\label{le SCE for GOE}
\left|\alpha_{\mathrm{sc}}(z) (m(z)-\msc(z))+(m(z)-\msc(z))^2\right|\prec \frac{|m(z)-\msc(z)|}{N\eta}+\frac{\im \msc(z)}{N\eta}+\frac{1}{(N\eta)^2}\,,
\end{align}
where $\alpha_{\mathrm{sc}}(z)\deq z+2\msc(z)$; see Subsection~\ref{subsection proof of prposition prop:local} for a similar computation. An elementary computation reveals that $|\alpha_{\mathrm{sc}}(z)| \sim \im \msc(z)$. Equation~\eqref{le SCE for GOE} is a {\it self-consistent equation} for the quantity $m(z)-\msc(z)$. Its local stability properties up to the edges were examined in the works~\cite{EYY2,EYY3}. From these results and~\eqref{le SCE for GOE} it follows that, for fixed $z\in \caD$, $|m(z)-\msc(z)|\prec 1$, implies $|m(z)-\msc(z)|\prec \frac{1}{N\eta}$. To obtain the local law on $\caD$ one then applies a continuity or bootstrapping argument~\cite{ESY1,EYY2,EYY3} by decreasing the imaginary part of the spectral parameter from $\eta\sim 1$ to $\eta\ge N^{-1}$. Using the monotonicity of the Stieltjes transform, this conclusion is extended to all of $\caE$. This establishes the local law for the~GOE,
\begin{align}\label{le local law for GOE}
|m(z)-\msc(z)|\prec\frac{1}{N\eta}\,,
\end{align}
uniformly on the domain $\mathcal{E}$.
Hence, to obtain the strong local law for the GOE, it suffices to establish~\eqref{example GOE} for fixed $z\in\caD$. By the definition of the normalized trace, $m\equiv m(z)$, of the Green function we have
\begin{align} \label{expansion 0}
\E [ |1 + zm + m^2|^{2D} ] = \E \bigg[ \bigg( \frac{1}{N} \sum_{i=1}^N (1 + z G_{ii}) + m^2 \bigg) (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \bigg]\,.
\end{align}
We expand the diagonal Green function entry $G_{ii}\equiv G^W_{ii}$ using the following identity:
\begin{align} \label{green identity}
1 + z G_{ii} = \sum_{k=1}^N W_{ik} G_{ki}\,,
\end{align}
which follows directly from the defining relation $(W-z\id)G = \id$. To some extent~\eqref{green identity} replaces the conventional Schur complement formula. We then obtain
\begin{align} \label{expansion wigner}
\E [ |1 + zm + m^2|^{2D} ] = \E \bigg[ \bigg( \frac{1}{N} \sum_{i, k} W_{ik} G_{ki} + m^2 \bigg) (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \bigg]\,.
\end{align}
Using that the matrix entries $W_{ik}$ are Gaussians random variables, integration by parts shows that
\begin{align}\label{le SL}
\E_{ik} [W_{ik} F(W_{ik})] = \frac{1+\delta_{ik}}{N} \E_{ik} [ \partial_{ik} F(W_{ik})]\,,
\end{align}
for differentiable functions $F\,:\,\R\to\C$, where $\partial_{ik} \equiv \partial/(\partial W_{ik})$. Here we used that $\E W_{ij}=0$ and $\E W^2_{ij}=(1+\delta_{ij})/N$ for the GOE. Identity~\eqref{le SL} is often called {\it Stein's lemma} in the statistics literature. Combining~\eqref{expansion wigner} and~\eqref{le SL} we obtain
\begin{align} \begin{split} \label{expansion wigner 2}
\E [ |1 + zm + m^2|^{2D} ]&= \frac{1}{N^2} \E \bigg[ \sum_{i, k}(1+\delta_{ik}) \partial_{ik} \bigg( G_{ki} (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \bigg) \bigg] \\
&\qquad + \E \big[ m^2 (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \big]\,.
\end{split} \end{align}
We next expand and estimate the first term on the right side of \eqref{expansion wigner 2}. It is easy to see that
\begin{align} \begin{split}\label{expansion wigner 3}
&(1+\delta_{ik})\partial_{ik} \left( G_{ki} (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \right) \\
&= -G_{ii} G_{kk} (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D - G_{ki} G_{ki} (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \\
&\quad - \frac{2(D-1)}{N} G_{ki} (z+2m) \sum_{j=1}^N G_{jk} G_{ij} (1 + zm + m^2)^{D-2} (\ol {1 + zm + m^2})^D \\
&\quad - \frac{2D}{N} G_{ki} (\ol z + 2\ol m) \sum_{j=1}^N \ol{G_{jk}} \ol{G_{ij}} (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^{D-1}\,.
\end{split} \end{align}
After averaging over the indices $i$ and $k$, the first term on the right side of~\eqref{expansion wigner 3} becomes
\begin{align*}
-\frac{1}{N^2} \E \bigg[ \sum_{i, k} G_{ii} G_{kk} (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \bigg] = - \E[ m^2 (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D]\,,
\end{align*}
which exactly cancels with the second term on the right side of~\eqref{expansion wigner 2}. The second term on the right side of~\eqref{expansion wigner 3} can be estimated as
\begin{align*} \begin{split}
&\bigg|\E \bigg[ \frac{1}{N^2} \sum_{i, k} G_{ki} G_{ki} (1 + zm + m^2)^{D-1} (\ol {1 + zm + m^2})^D \bigg] \bigg|\\
&\qquad\qquad\leq \E \bigg[ \frac{1}{N^2} \sum_{i, k} |G_{ki}|^2 |1 + zm + m^2|^{2D-1} \bigg] = \E \bigg[ \frac{\im m}{N\eta} |1 + zm + m^2|^{2D-1} \bigg]\,,
\end{split} \end{align*}
where we used the identity
\begin{align}\label{ward identity}
\frac{1}{N}\sum_{i=1}^N|G_{ik}(z)|^2=\frac{\im G_{kk}(z)}{N\eta}\,,
\end{align}
which we refer to as the {\it Ward identity} below. It follows from the spectral decomposition of $W$.
For the third term on the right side of~\eqref{expansion wigner 3} we have that
\begin{align} \begin{split} \label{expansion wigner 3 third}
&\bigg|\E \bigg[ \frac{1}{N^3} \sum_{i, j, k} G_{ki} G_{jk} G_{ij} (z+2m) (1 + zm + m^2)^{D-2} (\ol {1 + zm + m^2})^D \bigg]\bigg| \\
&\leq \E \bigg[ \frac{|\Tr G^3|}{N^3} |z+2m| \cdot |1 + zm + m^2|^{2D-2} \bigg] \leq \E \bigg[ \frac{\im m}{(N\eta)^2} |z+2m| \cdot |1 + zm + m^2|^{2D-2} \bigg]\,,
\end{split} \end{align}
where we used that
\begin{align}\label{sum rule 1}
\frac{1}{N^3}|\Tr G^3| \leq \frac{1}{N^3}\sum_{\alpha=1}^N \frac{1}{|\lambda_{\alpha} - z|^3} \leq \frac{1}{N^2\eta^2} \frac{1}{N}\sum_{\alpha=1}^N \frac{\eta}{|\lambda_{\alpha} - z|^2} = \frac{\im m}{N^2\eta^2}.
\end{align}
The fourth term on the right side of~\eqref{expansion wigner 3} can be estimated in a similar manner since
\begin{align}\label{sum rule 2}
\frac{1}{N^3}\bigg| \sum_{i, j, k} G_{ki} \ol{G_{ij}}\ol{G_{jk}} \bigg| =\frac{1}{N^3} \left| \Tr G \ol G^2 \right| \leq \frac{1}{N^3}\sum_{\alpha=1}^N \frac{1}{|\lambda_{\alpha} - z|^3}\le\frac{\im m}{N^2\eta^2}\,.
\end{align}
Returning to~\eqref{expansion wigner 2}, we hence find, for $z\in\C^+$, that
\begin{align*} \begin{split}
&\E [|1 + zm + m^2|^{2D}] \\
&\qquad\leq \E \left[ \frac{\im m}{N\eta} |1 + zm + m^2|^{2D-1} \right] + (4D-2) \E \left[ \frac{\im m}{(N\eta)^2} |z+2m| \cdot |1 + zm + m^2|^{2D-2} \right]\,,
\end{split} \end{align*}
which is the recursive moment estimate for the GOE stated in~\eqref{example GOE}.
\begin{remark}\label{remark on entrywise local law of GOE}
The above presented method can also be used to obtain the entry-wise local law for the Green function of the GOE. Assuming the local law for $m^W$ has been obtained, one may establish a recursive moment estimate for $1+zG^W_{ii}(z)+\msc(z) G^W_{ii}(z)$ to derive
\begin{align}\label{diagonal entrywise law for GOE}
\big|G^W_{ii}(z)-\msc(z)\big|\prec\sqrt{\frac{\im \msc(z)}{N \eta}}+\frac{1}{N\eta}\,,
\end{align}
uniformly in $i\in\llbracket 1,N\rrbracket$ and $z\in\mathcal{E}$. (One may also consider high moments of $1+z G^W_{ii}(z)+m^W(z)W_{ii}(z)$ to arrive at the same conclusion.) We leave the details to the reader. Yet, for later illustrative purposes in Section~\ref{sec:stein} and Section~\ref{sec:tw}, we sketch the derivation of recursive moment estimate for the off-diagonal Green function entries $G_{ij}\equiv G^W_{ij}(z)$, $i\not=j$. Let $z\in\caD$. Using the relation $(W-z\id)G = \id$, we~get
\begin{align*}
\E\Big[ z|G_{ij}|^{2D}\Big] = z\E\Big[ G_{ij} G_{ij}^{D-1} \ol{G_{ij}^D}\Big] = \sum_{k=1}^N\E\Big[ W_{ik} G_{kj} G_{ij}^{D-1} \ol{G_{ij}^D}\Big]=\sum_{k=1}^N \frac{1+\delta_{ik}}{N} \E \left[ \partial_{ik} \left( G_{kj} G_{ij}^{D-1} \ol{G_{ij}^D} \right) \right]\,,
\end{align*}
where we used Stein's lemma~in~\eqref{le SL} in the last step. Upon computing the derivative we get, for $i\not=j$,
\begin{align} \begin{split} \label{GOE local off-diagonal}
\E \left[ z|G_{ij}|^{2D} \right] &=
-\frac{1}{N} \sum_{k=1}^N \E \left[ G_{kk} G_{ij} G_{ij}^{D-1} \ol{G_{ij}^D} \right] -\frac{1}{N} \sum_{k=1}^N \E \left[ G_{ki} G_{kj} G_{ij}^{D-1} \ol{G_{ij}^D} \right] \\
& \qquad -\frac{D-1}{N} \sum_{k=1}^N \E \left[ G_{kj}^2 G_{ii} G_{ij}^{D-2} \ol{G_{ij}^D} \right] -\frac{D-1}{N} \sum_{k=1}^N \E \left[ G_{kj} G_{ik} G_{ij}^{D-1} \ol{G_{ij}^D} \right] \\
& \qquad -\frac{D}{N} \sum_{k=1}^N \E \left[ |G_{kj}|^2 \ol{G_{ii}} G_{ij}^{D-1} \ol{G_{ij}^{D-1}} \right] -\frac{D}{N} \sum_{k=1}^N \E \left[ G_{kj} \ol{G_{ik}} G_{ij}^{D-1} \ol{G_{ij}^D} \right] \,.
\end{split} \end{align}
The first term in the right side of~\eqref{GOE local off-diagonal} equals $-\E \big[m|G_{ij}|^{2D} \big]$. Using the Ward identity~\eqref{ward identity} we~have
\begin{align*}
\frac{1}{N}\sum_{k=1}^N |G_{ki}G_{kj}|\le \frac{1}{2N} \sum_{k=1}^N |G_{ki}|^2+\frac{1}{2N} \sum_{k=1}^N |G_{kj}|^2=\frac{\im G_{ii}+\im G_{jj}}{2N\eta}\,.
\end{align*}
Thus using~\eqref{diagonal entrywise law for GOE}, we get the bound
\begin{align*}
\frac{1}{N}\sum_{k=1}^N |G_{ki}G_{kj}|\prec \frac{\im \msc}{N\eta}+\frac{1}{(N\eta)^2}\,,
\end{align*}
uniformly on $\mathcal{E}$. We can now easily bound the right side of~\eqref{GOE local off-diagonal}, for example, any given $\epsilon > 0$,
\begin{align*}
\bigg| \frac{D}{N} \sum_{k=1}^N \E \left[ |G_{kj}|^2 \ol{G_{ii}} G_{ij}^{D-1} \ol{G_{ij}^{D-1}} \right] \bigg| \leq N^\epsilon \E \left[ \bigg(\frac{\im \msc}{N\eta}+\frac{1}{(N\eta)^2} \bigg) \big|G_{ij}\big|^{2D-2} \right]\,,
\end{align*}
for $N$ sufficiently large. Thus we get from~\eqref{GOE local off-diagonal} that, for $i\not=j$,
\begin{align}\begin{split}\label{GOE local off-diagonal 2}
|z+\msc|\,\E \big[|G_{ij}|^{2D}\big]&\le \E \Big[ |m-\msc| \big|G_{ij}\big|^{2D} \Big] + N^{\epsilon} \E \bigg[\bigg(\frac{\im \msc}{N\eta}+\frac{1}{(N\eta)^2} \bigg)|G_{ij}|^{2D-2} \bigg]\\ &\qquad\qquad+ N^{\epsilon} \E \bigg[ \bigg(\frac{\im \msc}{N\eta}+\frac{1}{(N\eta)^{2}} \bigg)|G_{ij}|^{2D-1} \bigg]\,,
\end{split}\end{align}
uniformly on $\mathcal{E}$, for $N$ sufficiently large. Since $|z+\msc| > c$ on $\mathcal{E}$, for some $N$-independent constant $c>0$, we find from~\eqref{GOE local off-diagonal 2} and~\eqref{le local law for GOE} by Young's and Markov's inequality that
\begin{align}
\big|G^W_{ij}(z)\big| \prec \sqrt{\frac{\im \msc(z)}{N \eta}}+\frac{1}{N\eta}\,,\qquad\qquad(i\not=j)\,,
\end{align}
for fixed $z\in\mathcal{D}$. Using continuity and monotonicity of $G_{ij}(z)$ the bound can be made uniform on the domain $\mathcal{E}$. Together with~\eqref{diagonal entrywise law for GOE}, this shows the entry-wise local law for the GOE.
\end{remark}
\subsection{Local law for sparse matrices}
When applying the strategy of Subsection~\ref{Example: Local law for the GOE} to sparse matrices we face two difficulties. First, since the matrix entries are not Gaussian random variables, the simple integration by parts formula~\eqref{le SL} needs to replaced by a full-fletched cumulant expansion. Second, since the higher order cumulants are not small (in the sense that the $(\ell+2)$-nd cumulant is only $O(N^{-1} q^{-\ell})$) we need to retain higher orders in the cumulant expansion. The following result generalizes~\eqref{le SL}.
\begin{lemma}[Cumulant expansion, generalized Stein lemma]\label{le stein lemma}
Fix $\ell\in \N$ and let $F\in C^{\ell+1}(\R;\C^+)$. Let $Y$ be a centered random variable with finite moments to order $\ell+2$. Then,
\begin{align}\label{le stein}
\E[Y F(Y)] = \sum_{r=1}^\ell \frac{\kappa^{(r+1)}(Y)}{r!} \E \big[ F^{(r)}(Y) \big]+\E\big[\Omega_\ell(YF(Y))\big]\,,
\end{align}
where $\E$ denotes the expectation with respect to $Y$, $\kappa^{(r+1)}(Y)$ denotes the $(r+1)$-th cumulant of $Y$ and~$F^{(r)}$ denotes the $r$-th derivative of the function $F$. The error term~$\Omega_\ell(YF(Y))$ in~\eqref{le stein} satisfies
\begin{align}
\big|\E\big[\Omega_\ell(YF(Y))\big]\big|\le C_\ell \E[ |Y|^{\ell+2}]\sup_{|t|\le Q}|F^{(\ell+1)}(t)|+ C_\ell \E [|Y|^{\ell+2} \lone(|Y|>Q)]\sup_{t\in \R} |F^{(\ell+1)}(t)| \,,
\end{align}
where $Q\ge 0$ is an arbitrary fixed cutoff and $C_\ell$ satisfies $C_\ell\le \frac{(C\ell)^\ell}{\ell!}$ for some numerical constant~$C$.
\end{lemma}
For proofs we refer to Proposition~3.1 in~\cite{LP} and Section~II of~\cite{KKP}. In case $Y$ is a standard Gaussian we recover~\eqref{le stein} and thus we sometimes refer to Lemma~\ref{le stein lemma} as generalized Stein lemma.
Let $H$ be a sparse matrix satisfying Assumption~\ref{assumption H} with $\phi>0$. Recall the polynomial $P\equiv P_z$ and the function~$\wt m\equiv \wt m(z)$ of Theorem~\ref{theorem:local} that satisfy$P(\wt m)=0$. Let $m\equiv m^{H}(z)$ be given by~\eqref{se green functions}. Following the ideas of Subsection~\ref{Example: Local law for the GOE}, we derive in Section~\ref{sec:stein} a recursive estimate for $\E |P(m)|^{2D}$, for large~$D$ with $z\in\mathcal{D}$; see Lemma~\ref{lem:claim} for the precise statement and Section~\ref{sec:stein} for its proof. We start~with
\begin{align}\label{the source}
\E [ |P|^{2D} ] = \E \Big[\Big(1+zm+m^2+\frac{\nm^{(4)}}{q_t^{2}}m^4\Big )P^{{D-1}}\ol{P^D}\Big ]\,,
\end{align}
for $D\geq2 $, and expand $zm$ using the identity
\begin{align}\label{expansion indentity}
zG_{ij} = \sum_{k=1}^N H_{ik} G_{kj}-\delta_{ij}\,,
\end{align}
which follows from the definition of the Green function. We then obtain the identity
\begin{align}\label{expansion}
\E\big[(1+z m)P^{{D-1}}\ol{P^D}\big]= \E \bigg[ \bigg( \frac{1}{N} \sum_{i\not= k} H_{ik} G_{ki} \bigg) P^{D-1} \ol{P^D} \bigg]\,.
\end{align}
Using the generalized Stein lemma, Lemma~\ref{le stein lemma}, we get
\begin{align}\begin{split}\label{ziwui eins}
\E\big[(1+ zm)P^{{D-1}}\ol{P^D}\big]&= \frac{1}{N} \sum_{r=1}^\ell \frac{\kappa^{(r+1)}}{r!} \E \bigg[ \sum_{i \neq k} \partial_{ik}^r \Big( G_{ki} P^{D-1} \ol{P^D} \Big) \bigg]+\E\Omega_{\ell}\Big((1+zm) P^{D-1} \ol{P^D} \Big)\,,
\end{split}\end{align}
where $\partial_{ik} = \partial/(\partial H_{ik})$ and $\kappa^{(k)}$ are the cumulants of $H_{ij}$, $i\not=j$. The detailed form of the error $\E\Omega_\ell(\cdot)$ is discussed in Subsection~\ref{subsection truncation of the cumulant expansion}. Anticipating the outcome, we mention that we can truncate the expansion at order $\ell\ge 8D$ so that the error term becomes sufficiently small for our purposes.
From the discussion in Subsection~\ref{Example: Local law for the GOE}, we see that the leading term on the right side of~\eqref{ziwui eins} is $\E [m^2 P^{D-1}\ol{P^D}]$ coming from the $r=1$ term. For the other terms with $2\le r\le \ell$, we need to separate relevant from negligible contributions (see beginning of~Section~\ref{sec:stein} for quantitative statement of negligible contributions). Some of the negligible contributions can be identified by power counting, while others require further expansions using cumulant series and ideas inspired by the GOE computation in Remark~\ref{remark on entrywise local law of GOE} above. In Lemma~\ref{summary expansions}, we will show that the remaining relevant terms stem from the term $r=3$ and are, after further expansions, eventually identified to be $s^{(4)}q^{-2}\E[m^4 P^{D-1}\ol{P^D}]$. As {\it a priori} estimates for this analysis we rely on Proposition~\ref{cor: delocalization}, stating that the eigenvectors of $H$ are completely delocalized, as well as on the rough bounds $|G_{ij}(z)|\prec 1$, $z\in\mathcal{D}$. Returning to~\eqref{ziwui eins}, we then observe that the relevant terms in~\eqref{ziwui eins} cancels with the third and fourth term on the right side of~\eqref{the source}. This yields the recursive moment estimate for $P(m)$, respectively $m$.
As we will see in Section~\ref{sec:rho}, the inclusion of the fourth moment $\nmf/q^2$ in $P(m)$ enables us to compute the deterministic shift of edge which is of order $q^{-2}$. While it is possible to include a higher order correction term involving the sixth moment, $\nm^{(6)}/q^4$, this does not improve the local law in our proof since the largest among the negligible contributions originates from the $r=3$ term in~\eqref{ziwui eins}. (More precisely, it is $I_{3, 2}$ of~\eqref{e1}.)
In Section~\ref{sec:proof main}, we then prove Theorem~\ref{theorem:local} and Theorem~\ref{prop:norm bound} using the recursive moment estimate for~$m$ and a local stability analysis. The local stability analysis relies on some properties of the Stieltjes transform~$\wt m$ of the deterministic distribution~$\wt\rho$ obtained in Section~\ref{sec:rho}.
\subsection{Tracy--Widom limit and Green function comparison} \label{Tracy-Widom limit and Green function comparison}
To establish the edge universality (for $\phi>1/6$), we first show in Subsection~\ref{preliminaries of edge universality} that the distribution of the largest eigenvalue of $H$ may be obtained as the expectation (of smooth functions) of the imaginary part of $m(z)$, for appropriately chosen spectral parameters $z$. Such a relation was the basic structure for proving the edge universality in~\cite{EYY,EKYY2}, and the main ingredients in the argument are the local law, the square-root decay at the edge of the limiting density, and an upper bound on the largest eigenvalue, which are Theorem~\ref{theorem:local}, Lemma~\ref{lem:w}, and Theorem~\ref{prop:norm bound} for the case at hand.
For the sake of self-containment, we redo some parts of these estimates in Subsection~\ref{preliminaries of edge universality}.
In Subsection~\ref{subsection green function comparison}, we then use the {\it Green function comparison} method~\cite{EYY2,EYY} to compare the edge statistics of $H$ with the edge statistics of a GOE matrix. Together with the argument of Subsection~\ref{preliminaries of edge universality}, this will yield the Tracy--Widom limit of the largest eigenvalue. However, the conventional discrete Lindeberg type replacement approach to the Green function comparison does not work due to the slow decaying moments of the sparse matrix. We therefore use a continuous flow that interpolates between the sparse matrix ensemble and the GOE. Such an approach has shown to be effective in proving edge universality for deformed Wigner matrices~\cite{LS14} and for sample covariance matrices~\cite{LS14b}.
More concretely, we consider the {\it Dyson matrix flow} with initial condition $H_0$ defined by
\begin{align} \label{eq:A(t)}
H_t \deq \e{-t/2} H_0 + \sqrt{1-\e{-t}} W^{\GOE}\,, \qquad\qquad( t\ge 0)\,,
\end{align}
where $W^{\GOE}$ is a GOE matrix independent of $H_0$. In fact, since we will choose $H_0$ to be a sparse matrix~$H$ with vanishing diagonal entries, we assume with some abuse of terminology that $W^{\GOE}=(W^{\GOE})^*$ has vanishing diagonal, \ie we assume that $W^{\GOE}_{ii}=0$ and that $(W^{\GOE}_{ij},i<j)$ are independent centered Gaussian random variables of variance $1/N$. It was shown in Lemma~3.5 of~\cite{LY} that the local edge statistics of $W^{\GOE}$ is described by the GOE Tracy--Widom statistics.
Let $\kappa_t^{(k)}$ be the $k$-th cumulant of $(H_t)_{ij}$, $i\not=j$. Then, by the linearity of the cumulants under the addition of independent random variables, we have $\kappa_t^{(1)} =0$, $\kappa_t^{(2)}=1/N$ and $\kappa_t^{(k)}=\e{-kt/2}\kappa^{(k)}$, $k\ge 3$. In particular, we have the bound
\begin{align}\label{eq:cumulant}
|\kappa_t^{(k)}|\le \e{-t}\frac{(Ck)^{ck}}{N q_t^{k-2}}\,,\qquad\qquad (k\ge 3)\,,
\end{align}
where we introduced the {\it time-dependent sparsity parameter}
\begin{align}
q_t \deq q\,\e{t/2}\,.
\end{align}
Choosing $t=6\log N$, a straightforward perturbation argument shows that the local statistics, at the edges and in the bulk, of $H_t$ and $W^{\GOE}$ agree up to negligible error. It thus suffices to consider $t\in[0,6\log N]$.
We first establish the local law for the normalized trace of the Green function of~$H_t$.~Let
\begin{align}\label{le timedependent green functions}
G_t(z)\equiv G^{H_t}(z)=\frac{1}{H_t-z\id}\,,\qquad\quad m_t(z)\equiv m^{H_t}(z)=\frac{1}{N}\sum_{i=1}^N (G_t)_{ii}(z)\,,\qquad\qquad (z\in\C^+)\,.
\end{align}
\begin{proposition} \label{prop:local}
Let $H_0$ satisfy Assumption~\ref{assumption H} with $\phi>0$. Then, for any $t\ge0$, there exists a algebraic function $\wt m_t : \C^+ \to \C^+$ and $2\le L_t<3$ such that the following holds:
\begin{enumerate}
\item $\wt m_t$ is the Stieltjes transform of a deterministic symmetric probability measure $\wt \rho_t$, \ie $\wt m_t(z)=m_{\wt\rho_t}(z)$. Moreover, $\supp \wt\rho_t=[-L_t,L_t]$ and $\wt\rho_t$ is absolutely continuous with respect to Lebesgue measure with a strictly positive density on $(-L_t,L_t)$.
\item $\wt m_t \equiv \wt m_t (z)$ is a solution to the polynomial equation
\begin{align}\begin{split} \label{eq:poly}
P_{t, z}(\wt m_t)& \deq 1 + z \wt m_t + \wt m_t^2 + \e{-t} q_t^{-2} \nmf\wt m_t^4 \\
&=1 + z \wt m_t + \wt m_t^2 + \e{-2t} q^{-2}\nmf \wt m_t^4 = 0\,.
\end{split}\end{align}
\item
The normalized trace of the Green function
satisfies
\begin{align}
|m_t (z) - \wt m_t (z)| \prec \frac{1}{q_t^2}+\frac{1}{N\eta}\,,
\end{align}
uniformly on the domain $\caE$ and uniformly in $t\in[0,6\log N]$.
\end{enumerate}
\end{proposition}
Note that Theorem~\ref{theorem:local} is a special case of Proposition~\ref{prop:local}. Given Proposition~\ref{prop:local}, Corollary~\ref{corollary density of states} extends in the obvious way from $H$ to $H_t$. Proposition~\ref{prop:local} is proved in Subsection~\ref{subsection proof of prposition prop:local}.
The endpoints $\pm L_t$ of the support of $\wt \rho_t$ are given by $L_t = 2 + \e{-t} \nmf q_t^{-2} + O(\e{-2t}q_t^{-4})$ and satisfy
\begin{align}\label{le L dot}
\dot L_t = -2 \e{-t} \nmf q_t^{-2} + O(\e{-2t}q_t^{-4})\,,
\end{align}
where $\dot L_t$ denotes the derivative with respect to $t$ of $L_t$; \cf Remark~\ref{rem:L_+ estimate} below.
Choose now $q\ge N^{\phi}$ with $\phi>1/6$. In our proof of the Green function comparison theorem, Proposition~\ref{prop:green}, we estimate the rate of change of~$m_t$ along the Dyson matrix flow over the time interval $[0,6\log N]$, where it undergoes a change of $o(1)$. The continuous changes in $m_t$ can be compensated by letting evolve the spectral parameter $z\equiv z(t)$ according to~\eqref{le L dot}. This type of cancellation argument appeared first in~\cite{LS14} in the context of deformed Wigner matrices. However, one cannot prove the Green function comparison theorem for sparse random matrices by directly applying the cancellation argument since the error bound for the entry-wise local law in Proposition~\ref{local semiclrcle law} is not sufficiently small. Thus the proof of the Green function comparison theorem requires some non-trivial estimates on functions of Green functions as is explained in Subsection~\ref{subsection green function comparison}.
\section{The measure $\wt \rho$ and its Stieltjes transform} \label{sec:rho}
In this section, we prove important properties of $\wt m_t\equiv \wt m_t(z) $ in Proposition \ref{prop:local}. Recall that $\wt m_t$ is a solution to the polynomial equation $P_{t,z}(\wt m_t) = 0$ in \eqref{eq:poly} and that $q_t=\e{t/2}q$.
\begin{lemma} \label{lem:w}
For any fixed $z = E + \ii \eta \in \caE$ and any $t\ge0$, the polynomial equation $P_{t,z}(w_t) = 0$ has a unique solution $w_t \equiv w_t(z)$ satisfying $\im w_t > 0$ and $|w_t| \leq 5$. Moreover, $w_t$ has the following properties:
\begin{enumerate}
\item There exists a probability measure $\wt \rho_t$ such that the analytic continuation of $w_t$ coincides with the Stieltjes transform of $\wt \rho_t$.
\item The probability measure $\wt \rho_t$ is supported on $[-L_t, L_t]$, for some $L_t \ge 2$, has a strictly positive density inside its support and vanishes as a square-root at the edges, \ie letting
\begin{align}\label{le kappa}
\kappa_t\equiv\kappa_t(E) \deq \min \{ |E+L_t|, |E-L_t| \}\,,
\end{align}
we have
\begin{align}
\wt \rho_t(E) \sim \kappa_t^{1/2}(E)\,,\qquad\qquad (E\in(-L_t,L_t))\,.
\end{align}
Moreover, $L_t = 2 + \e{-t} q_t^{-2}\nmf + O(\e{-2t}q_t^{-4})$.
\item The solution $w_t$ satisfies that
\begin{align} \begin{split}
\im w_t(E+\ii\eta)& \sim \sqrt{\kappa_t + \eta} \qquad \text{ if }\; E \in [-L_t, L_t]\,, \\
\im w_t(E+\ii\eta) &\sim \frac{\eta}{\sqrt{\kappa_t + \eta}} \qquad\! \text{ if }\; E \notin [-L_t, L_t]\,.
\end{split} \end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
For simplicity, we abbreviate $P \equiv P_{t, z}$. Let
\begin{align}
Q(w)\equiv Q_{t,z}(w_t) \deq -\frac{1}{w_t} - w_t - \e{-t} q_t^{-2}\nmf w_t^3.
\end{align}
By definition, $P(w) = 0$ if and only if $z = Q(w)$. It is easily checked that the derivative
\begin{align}
Q'(w) = \frac{1}{w^2} - 1 - 3 \e{-t} q_t^{-2}\nmf w^2
\end{align}
is monotone increasing on $(-\infty, 0)$. Furthermore, we have
$$
Q'(-1) =-3\e{-t}q_t^{-2}\nmf<0\,, \quad Q'(-1 + 2 q_t^{-2}\nmf) = (4-3\e{-t}) q_t^{-2}\nmf + O(q_t^{-4}) > 0\,.
$$
Hence, $Q'(w) = 0$ has a unique solution on $(-1, -1+ 2 q_t^{-2}\nmf)$, which we will denote by $\tau_t$, and $Q(w) \equiv Q_t(w_t)$ attains its minimum on $(-\infty, 0)$ at $w_t = \tau_t$. We let $L_t\deq Q_t(\tau_t)$, or equivalently, $w_t=\tau_t$ if $z=L_t$. We remark that there is no $w_t \in (-\infty, 0)$ satisfying $Q_t(w_t) < L_t$.
A direct calculation shows that
\begin{align} \label{eq:L_+ estimate}
\tau_t = -1 + \frac{3 \e{-t}}{2} q_t^{-2}\nmf + O(\e{-2t}q_t^{-4})\,, \qquad L_t = 2 + \e{-t} q_t^{-2}\nmf + O(\e{-2t}q_t^{-4})\,.
\end{align}
For simplicity we let $L \equiv L_t$ and $\tau\equiv\tau_t$. Choosing now $w = Q^{-1}(z)$ we have the expansion
\begin{align} \begin{split}
z &= Q(\tau) + Q'(\tau)(w-\tau) + \frac{Q''(\tau)}{2} (w-\tau)^2 + O(|w-\tau|^3) \\
&= L + \frac{Q''(\tau)}{2} (w-\tau)^2 + O(|w-\tau|^3)\,,
\end{split} \end{align}
in a $q_t^{-1/2}$-neighborhood of $\tau_t$. We hence find that
\begin{align} \label{L+ square root}
w = \tau + \left( \frac{2}{Q''(\tau)} \right)^{1/2} \sqrt{z-L} + O(|z-L|)\,,
\end{align}
in that neighborhood. In particular, choosing the branch of the square root so that $\sqrt{z-L} \in \C^+$, we find that $\im w > 0$ since $Q''(\tau)>0$.
We can apply the same argument with a solution of the equation $Q'(w) = 0$ on $(1-2q_t^{-2}\nmf, 1)$, which will lead us to the relation
\begin{align} \label{L- square root}
w = -\tau + \left( \frac{2}{Q''(\tau)} \right)^{1/2} \sqrt{z+L} + O(|z+L|)\,,
\end{align}
in a $q_t^{1/2}$-neighborhood of $-\tau$. We note that there exists another solution with negative imaginary part, which corresponds to the different branch of the square root.
For uniqueness, we consider the disk $B_5 = \{ w \in \C : |w| < 5 \}$. On its boundary $\partial B_5$
\begin{align}
|w^2 + zw + 1| \geq |w|^2 - |z| \cdot |w| - 1 > 1 > \big| q_t^{-2} w^4\nmf \big|\,,
\end{align}
for $z \in \caE$. Hence, by Rouch\'e's theorem, the equation $P(w) = 0$ has the same number of roots as the quadratic equation $w^2 + zw + 1 = 0$ in $B_5$. Since $w^2 + zw + 1 = 0$ has two solutions on $B_5$, we find that $P(w) = 0$ has two solutions on it. For $z = L + \ii q_t^{-1}$, we can easily check that the one solution of $P(w) = 0$ has positive imaginary part (from choosing the branch of the square root as in~\eqref{L+ square root}) and the other solution has negative imaginary part. If both solutions of $P(w) = 0$ are in~$\C^+ \cup \R$ (or in $\C^- \cup \R$) for some $z = \wt z \in \C^+$, then by continuity, there exists $z'$ on the line segment joining $L + \ii q_t^{-1}$ and $\wt z$ such that $P_{t, z'}(w') = 0$ for some $w' \in \R$. By the definition of $P$ this cannot happen, hence one solution of $P(w)=0$ is in~$\C^+$ and the other in~$\C^-$, for any $z \in \caE$. This shows the uniqueness statement of the lemma.
Next, we extend $w \equiv w(z)$ to cover $z \notin \caE$. (With slight abuse of notation, the extension of $w$ will also be denoted by $w \equiv w(z)$.) Repeating the argument of the previous paragraph, we find that $P(w) = 0$ has two solutions for $z \in (-L, L)$. Furthermore, we can also check that exactly one of them is in~$\C^+$ by considering $z = \pm L \mp q_t^{-1}$ and using continuity. Thus, $w(z)$ forms a curve on $\C^+$, joining $-\tau$ and $\tau$, which we will denote by $\Gamma$. We remark that, by the inverse function theorem, $w(z)$ is analytic for $z \in (-L, L)$ since $Q'(w) \neq 0$ for such $z$.
By symmetry, $\Gamma$ intersects the imaginary axis at $w(0)$. On the imaginary axis, we find that
\begin{align}
Q'(w) = \frac{1}{w^2} - 1 + O(\e{-t}q_t^{-2}) < 0 \qquad \text{if } \, |w| < 5\,.
\end{align}
Thus, we get from $Q(w)=z$ that
\begin{align}
\frac{\dd w}{\dd \eta} = \frac{1}{Q'(w)} \frac{\dd z}{\dd \eta} = \frac{\ii}{Q'(w)}\,,
\end{align}
which shows in particular that $w(\ii)$ is pure imaginary and $\im w(\ii) < \im w(0)$. By continuity, this shows that the analytic continuation of $w(z)$ for $z \in \C^+$ is contained in the domain $D_{\Gamma}$ enclosed by $\Gamma$ and the interval $[-L, L]$. We also find that $|w(z)| < 5$, for all $z \in \C^+$.
To prove that $w(z)$ is analytic in $\C^+$, it suffices to show that $Q'(w) \neq 0$ for $w \in D_{\Gamma}$. If $Q'(w) = 0$ for $w \in D_{\Gamma}$, we have
\begin{align}
w^2 Q'(w) = 1 - w^2 - 3 \e{-t} q_t^{-2}\nmf w^4 = 0\,.
\end{align}
On the circle $\{ w \in \C : |w| = 5 \}$,
\begin{align}
|1 - w^2| \geq 24 > \big| 3 \e{-t} q_t^{-2} w^4\nmf \big|\,.
\end{align}
Hence, again by Rouch\'e's theorem, $w^2 Q'(w) = 0$ has two solutions in the disk $\{ w \in \C : |w| < 5 \}$. We already know that those two solutions are $\pm \tau$. Thus, $Q'(w) \neq 0$ for $w \in D_{\Gamma}$ and $w(z)$ is analytic.
Let $\wt \rho$ be the measure obtained by the Stieltjes inversion of $w \equiv w(z)$. To show that $\wt \rho$ is a probability measure, it suffices to show that $\lim_{y \to \infty} \ii y \, w(\ii y) = -1$. Since $w$ is bounded, one can easily check from the definition of $w$ that $|w| \to 0$ as $|z| \to \infty$. Thus,
\begin{align}
0 = \lim_{z \to \ii \infty} P(w(z)) = \lim_{z \to \ii \infty} (1 + zw)\,,
\end{align}
which implies that $\lim_{y \to \infty} \ii y \, w(\ii y) = -1$. This proves the first property of $\wt \rho$. Other properties can be easily proved from the first property and Equations~\eqref{L+ square root} and~\eqref{L- square root}.
\end{proof}
\begin{remark} \label{rem:L_+ estimate}
Recall that $q_t=\e{t/2}q$. As we have seen in \eqref{eq:L_+ estimate},
\begin{align}
L_t = 2 + \e{-t} q_t^{-2}\nmf + O(\e{-2t}q_t^{-4}) = 2 + \e{-2t} q^{-2}\nmf + O(\e{-4t}q^{-4})\,.\nonumber
\end{align}
Moreover, the time derivative of $L_t$ satisfies
\begin{align}
\dot L_t = \frac{\dd}{\dd t} Q(\tau) = \frac{\partial Q}{\partial t}(\tau) + Q'(\tau) \dot \tau = \frac{\partial Q}{\partial t}(\tau) = 2\e{-2t} q^{-2}\nmf \tau^3\,,\nonumber
\end{align}
hence, referring once more to~\eqref{eq:L_+ estimate},
\begin{align}
\dot L_t = -2 \e{-t} q_t^{-2}\nmf + O(\e{-2t}q_t^{-4}) =-2 \e{-2t} q^{-2}\nmf + O(\e{-4t}q^{-4})\,.\nonumber
\end{align}
\end{remark}
\begin{remark} \label{rem:stability of wt m}
It can be easily seen from the definition of $P_{t, z}$ that $w_t \to \msc$ as $N \to \infty$ or $t \to \infty$. For $z \in \caE$, we can also check the {\it stability condition }
$|z+w_t| > 1/6$ since
\begin{align}
|z+ w_t| = \frac{|1+ \e{-2t} q^{-2}\nmf |w_t|^4|}{|w_t|}
\end{align}
and $|w_t|<5$, as we have seen in the proof of Lemma \ref{lem:w}.
\end{remark}
\section{Proof of Proposition \ref{prop:local} and Theorem~\ref{prop:norm bound}} \label{sec:proof main}
\subsection{Proof of Proposition \ref{prop:local}}\label{subsection proof of prposition prop:local}
In this section, we prove Proposition \ref{prop:local}. The main ingredient of the proof is the recursive moment estimate for $P(m_t)$. Recall the subdomain $\caD$ of $\caE$ defined in~\eqref{domain} and the matrix $H_t$, $t\ge 0$, defined in~\eqref{eq:A(t)}. We have the following result.
\begin{lemma}[Recursive moment estimate]\label{lem:claim}
Fix $\phi>0$ and suppose that $H_0$ satisfies Assumption~\ref{assumption H}. Fix any $t\ge 0$. Recall the definition of the polynomial $P\equiv P_{t,z}$ in \eqref{eq:poly}. Then, for any $D > 10$ and (small) $\epsilon > 0$, the normalized trace of the Green function, $m_t \equiv m_t (z)$, of the matrix $H_t$ satisfies
\begin{align} \label{eq:claim}
\E \left|P(m_t) \right|^{2D}& \leq N^{\epsilon} \, \E \bigg[ \bigg( \frac{\im m_t}{N\eta} + q_t^{-4} \bigg) \big| P(m_t) \big|^{2D-1} \bigg] + N^{-\epsilon/8} q_t^{-1} \, \E \bigg[ |m_t - \wt m_t|^2 \big| P(m_t) \big|^{2D-1} \bigg]\nonumber \\
&\qquad+ N^{\epsilon} q_t^{-1} \, \sum_{s=2}^{2D} \sum_{s'=0}^{s-2} \E \bigg[ \bigg( \frac{\im m_t}{N\eta} \bigg)^{2s-s'-2} \big| P'(m_t) \big|^{s'} \big| P(m_t) \big|^{2D-s} \bigg] + N^\epsilon q_t^{-8D}\\
&\qquad+ N^{\epsilon} \, \sum_{s=2}^{2D} \E \bigg[ \bigg( { \frac{1}{N\eta} } + \frac{1}{q_t}\bigg({\frac{\im m_t}{N\eta}}\bigg)^{1/2} + \frac{1}{q_t^{2}} \bigg) \bigg( \frac{\im m_t}{N\eta} \bigg)^{s-1} \big| P'(m_t) \big|^{s-1} \big| P(m_t) \big|^{2D-s} \bigg]\,,\nonumber
\end{align}
uniformly on the domain $\caD$, for $N$ sufficiently large.
\end{lemma}
The proof of Lemma~\ref{lem:claim} is postponed to Section~\ref{sec:stein}. We are now ready to prove Proposition~\ref{prop:local}.
\begin{proof}[Proof of Proposition \ref{prop:local} and Theorem~\ref{theorem:local}]
Fix $t\in[0,6\log N]$. Let $\wt m_t$ be the solution $w_t$ in Lemma~\ref{lem:w}. The first two parts were already proved in Lemma~\ref{lem:w}, so it suffices to prove the third part of the proposition. For simplicity, we omit the $z$-dependence. Let
\begin{align}
\Lambda_t \deq |m_t - \wt m_t|\,.
\end{align}
We remark that from the local law in Lemma \ref{local semiclrcle law}, we have $\Lambda_t\prec 1$, $z\in\caD$. We also define the following $z$-dependent deterministic parameters
\begin{align}
\alpha_1 \deq \im \wt m_t\, ,\qquad \alpha_2 \deq P'(\wt m_t)\,, \qquad \beta \deq \frac{1}{N\eta} +\frac{1}{ q_t^2}\,,
\end{align}
with $z=E+\ii\eta$. We note that
\begin{align*}
|\alpha_2| \geq \im P'(\wt m_t) = \im z + 2 \im m_t + 4 \e{-t}\nmf q_t^{-2} \left( 3 (\re \wt m_t)^2 \im \wt m_t - (\im \wt m_t)^3 \right) \geq \im \wt m_t = \alpha_1\,,
\end{align*}
since $|\wt m_t| \leq 5$ as proved in Lemma \ref{lem:w}. Recall that $\wt m_t(L) = \tau$ in the proof of Lemma \ref{lem:w}. Recalling the definition of $\kappa_t\equiv \kappa_t(E)$ in~\eqref{le kappa} and using~\eqref{L+ square root} we have
\begin{align*}
|\wt m_t - \tau| \sim \sqrt{z - L} \sim \sqrt{\kappa_t + \eta}\,.
\end{align*}
By the definitions of $\tau$ and $L$ in the proof of Lemma \ref{lem:w}, we also have that
\begin{align*} \begin{split}
L + 2\tau + 4\e{-t} q_t^{-2}\nmf \tau^3 &= -\frac{1}{\tau} - \tau - \e{-t} q_t^{-2} \tau^3\nmf + 2\tau + 4\e{-t} q_t^{-2} \tau^3 \nmf\\ &= -\tau \big( \frac{1}{\tau^2} - 1 - 3\e{-t} q_t^{-2} \tau^2\nmf \big)=0\,,
\end{split} \end{align*}
hence
\begin{align} \label{eq:P' expansion}
P'(\wt m_t) = z + 2\wt m_t + 4\e{-t} q_t^{-2} {\wt m_t}^3\nmf = (z-L) + 2(\wt m_t -\tau) + 4 \e{-t} q_t^{-2}\nmf ({\wt m_t}^3 - \tau^3)\,,
\end{align}
and we find from~\eqref{L+ square root} that
$$
|\alpha_2|=|P'(\wt m_t)| \sim \sqrt{\kappa_t + \eta}\,.
$$
We remark that the parameter $\alpha_1$ is needed only for the proof of Theorem~\ref{prop:norm bound}; the proof of Proposition \ref{prop:local} can be done simply by substituting every $\alpha_1$ below with $|\alpha_2|$.
Recall that, for any $a, b \geq 0$ and ${\textsf p},{\textsf q} > 1$ with ${\textsf p}^{-1} + {\textsf q}^{-1} = 1$, Young's inequality states
\begin{align}\label{Young inequality}
ab \leq \frac{a^{\textsf p}}{{\textsf p}} + \frac{b^{\textsf q}}{{\textsf q}}\,.
\end{align}
Let $D\geq 10$ and choose any (small) $\epsilon>0$. All estimates below hold for $N$ sufficiently large (depending on $D$ and $\epsilon$). For brevity, $N$ is henceforth implicitly assumed to be sufficiently large.
Using first that $\im m_t \leq \im \wt m_t + |m_t -\wt m_t| = \alpha_1 + \Lambda_t $ and then applying~\eqref{Young inequality} with ${\textsf p}=2D$ and ${\textsf q} =2D/(2D-1)$, we get for the first term on the right side of \eqref{eq:claim} that
\begin{align} \begin{split} \label{young1}
&N^{\epsilon} \left( \frac{\im m_t}{N\eta} + q_t^{-4} \right) |P(m_t)|^{2D-1} \leq N^{\epsilon} \frac{\alpha_1 + \Lambda_t}{N\eta} |P(m_t)|^{2D-1} + N^{\epsilon} q_t^{-4} |P(m_t)|^{2D-1} \\
&\qquad\leq \frac{N^{(2D+1)\epsilon}}{2D} \left( \frac{\alpha_1 + \Lambda_t}{N\eta} \right)^{2D} + \frac{N^{(2D+1)\epsilon}}{2D} q_t^{-8D} + \frac{2(2D-1)}{2D} \cdot N^{-\frac{\epsilon}{2D-1}} |P(m_t)|^{2D}\,.
\end{split} \end{align}
Similarly, for the second term on the right side of \eqref{eq:claim}, we have
\begin{align} \label{young2}
N^{-\epsilon/8} q_t^{-1} \Lambda_t^2 \left| P(m_t) \right|^{2D-1} \leq \frac{N^{-(D/4 -1)\epsilon}}{2D} q_t^{-2D} \Lambda_t^{4D} + \frac{2D-1}{2D} N^{-\frac{\epsilon}{2D-1}} |P(m_t)|^{2D}\,.
\end{align}
From the Taylor expansion of $P'(m_t)$ around $\wt m_t$, we have
\begin{align}
|P'(m_t) - P'(\wt m_t) - P''(\wt m_t) (m_t - \wt m_t)| \le C q_t^{-2} \Lambda_t^2\,,
\end{align}
and $|P'(m_t)| \leq |\alpha_2| + 3\Lambda_t$, for all $z\in\mathcal{D}$, with high probability since $P''(\wt m_t) = 2 + O(q_t^{-2})$ and $\Lambda_t\prec 1$ by assumption. We note that, for any fixed $s\ge 2$,
\begin{align*}\begin{split}
(\alpha_1 + \Lambda_t)^{2s-s'-2} (|\alpha_2| + 3\Lambda_t)^{s'} &\leq N^{\epsilon/2} (\alpha_1 + \Lambda_t)^{s-1} (|\alpha_2| + 3\Lambda_t)^{s-1}\\
&\leq N^{\epsilon} (\alpha_1 + \Lambda_t)^{s/2} (|\alpha_2| + 3\Lambda_t)^{s/2}
\end{split}\end{align*}
with high probability, uniformly on $\caD$, since $\alpha_1 \leq |\alpha_2| \leq C$ and $\Lambda_t \prec 1$. In the third term of \eqref{eq:claim}, note that $2s-s'-2 \geq s$ since $s' \leq s-2$. Hence, for $2 \leq s \leq 2D$,
\begin{align} \label{young3} \begin{split}
& {N^{\epsilon}}q_t^{-1} \left( \frac{\im m_t}{N\eta} \right)^{2s-s'-2} \left| P'(m_t) \right|^{s'} \left| P(m_t) \right|^{2D-s}\\
&\qquad\qquad\leq {N^{\epsilon} }q_t^{-1} \beta^s (\alpha_1 + \Lambda_t)^{2s-s'-2} (|\alpha_2| + 3\Lambda_t)^{s'} \left| P(m_t) \right|^{2D-s} \\
&\qquad\qquad\leq N^{2\epsilon} q_t^{-1} \beta^s (\alpha_1 + \Lambda_t)^{s/2} (|\alpha_2| + 3\Lambda_t)^{s/2} \left| P(m_t) \right|^{2D-s} \\
&\qquad\qquad\leq N^{2\epsilon} q_t^{-1} \frac{s}{2D} \beta^{2D} (\alpha_1 + \Lambda_t)^D (|\alpha_2| + 3\Lambda_t)^{D} + N^{2\epsilon} q_t^{-1} \frac{2D-s}{2D} \left| P(m_t) \right|^{2D}
\end{split} \end{align}
uniformly on $\mathcal{D}$ with high probability. For the last term in~\eqref{eq:claim}, we note that
\begin{align}
{ \frac{1}{N\eta} } + q_t^{-1}\bigg({\frac{\im m_t}{N\eta}}\bigg)^{1/2} + q_t^{-2} \prec \beta\,,
\end{align}
uniformly on $\caD$. Thus, similar to \eqref{young3} we find that, for $2 \leq s \leq 2D$,
\begin{align} \label{young4}
& N^{\epsilon} \bigg( { \frac{1}{N\eta} } + q_t^{-1}\bigg({\frac{\im m_t}{N\eta}}\bigg)^{1/2} + q_t^{-2} \bigg) \bigg( \frac{\im m_t}{N\eta} \bigg)^{s-1} \left| P'(m_t) \right|^{s-1} \left| P(m_t) \right|^{2D-s}\nonumber \\
&\qquad\leq N^{2\epsilon} \beta \cdot \beta^{s-1} (\alpha_1 + \Lambda_t)^{s/2} (|\alpha_2| + 3\Lambda_t)^{s/2} \left| P(m_t) \right|^{2D-s} \nonumber\\
&\qquad\leq \frac{s}{2D} \left( N^{2\epsilon} N^{\frac{(2D-s)\epsilon}{4D^2}} \right)^{\frac{2D}{s}} \beta^{2D} (\alpha_1 + \Lambda_t)^D (|\alpha_2| + 3\Lambda_t)^{D} + \frac{2D-s}{2D} \left( N^{-\frac{(2D-s)\epsilon}{4D^2}} \right)^{\frac{2D}{2D-s}} \left| P(m_t) \right|^{2D} \nonumber\\
&\qquad\leq N^{(2D+1)\epsilon} \beta^{2D} (\alpha_1 + \Lambda_t)^D (|\alpha_2| + 3\Lambda_t)^{D} + N^{-\frac{\epsilon}{2D}} \left| P(m_t) \right|^{2D}\,,
\end{align}
for all $z\in\caD$, with high probability. We hence have from \eqref{eq:claim}, \eqref{young1}, \eqref{young2}, \eqref{young3} and \eqref{young4} that
\begin{align} \begin{split}
\E \left[|P(m_t)|^{2D}\right] &\leq N^{(2D+1)\epsilon} \E \left[ \beta^{2D} (\alpha_1 + \Lambda_t)^D (|\alpha_2| + 3\Lambda_t)^D \right] + \frac{N^{(2D+1)\epsilon}}{2D} q_t^{-8D} \\
&\qquad + \frac{N^{-(D/4 -1)\epsilon}}{2D} q_t^{-2D} \E \left[\Lambda_t^{4D}\right] + C N^{-\frac{\epsilon}{2D}} \E \left[ |P(m_t)|^{2D} \right]\,,
\end{split} \end{align}
for all $z\in\mathcal{D}$. Note that the last term on the right side can be absorbed into the left side. Thence
\begin{align} \begin{split} \label{eq:claim'}
&\E \left[|P(m_t)|^{2D}\right] \\
&\leq C N^{(2D+1)\epsilon} \E \left[ \beta^{2D} (\alpha_1 + \Lambda_t)^D (|\alpha_2| + 3\Lambda_t)^D \right] + \frac{N^{(2D+1)\epsilon}}{D} q_t^{-8D} + \frac{N^{-(D/4 -1)\epsilon}}{D} q_t^{-2D} \E \left[\Lambda_t^{4D}\right] \\
&\leq N^{3D\epsilon} \beta^{2D} |\alpha_2|^{2D} + N^{3D\epsilon} \beta^{2D} \E \left[ \Lambda_t^{2D} \right] + N^{3D \epsilon} q_t^{-8D} + N^{-D \epsilon/8} q_t^{-2D} \E \left[\Lambda_t^{4D}\right],
\end{split} \end{align}
uniformly on $\mathcal{D}$, where we used that $D> 10$ and the inequality
\begin{align}\label{convex inequality}
(a+b)^{\textsf p} \leq 2^{{\textsf p}-1}(a^{\textsf p} + b^{\textsf p})\,,
\end{align}
for any $a, b \geq 0$ and ${\textsf p} \geq 1$, to get the second line.
Next, from the third order Taylor expansion of $P(m_t)$ around $\wt m_t$, we have
\begin{align}\label{le taylor of P}
\left| P(m_t) - \alpha_2 (m_t - \wt m_t) - \frac{P''(\wt m_t)}{2} (m_t - \wt m_t)^2 \right| \le C q_t^{-2} \Lambda_t^3
\end{align}
since $P(\wt m_t) = 0$ and $P'''(\wt m_t)=4!\e{-t} q_t^{-2}\nmf\wt m_t$. Thus, using $\Lambda_t\prec 1$ and $P''(\wt m_t)=2+O(q_t^{-2}) $ we~get
\begin{align}
\Lambda_t^2\prec 2|\alpha_2|\Lambda_t+2|P(m_t)|\,,\qquad\qquad (z\in\mathcal{D})\,.
\end{align}
Taking the $2D$-power of the inequality and using once more~\eqref{convex inequality}, we get after taking the expectation
\begin{align}
\E [\Lambda_t^{4D}]\le 4^{2D}N^{\epsilon/2}|\alpha_2|^{2D}{\E[ \Lambda_t^{2D}]}+4^{2D}N^{\epsilon/2}\E[ |P(m_t)|^{2D}]\,,\qquad\qquad(z\in\mathcal{D})\,.
\end{align}
Replacing from~\eqref{eq:claim'} for $\E[ |P(m_t)|^{2D}]$ we obtain, using that $4^{2D}\le N^{\epsilon/2}$, for $N$ sufficiently large,
\begin{align}\begin{split}\label{annoying}
\E [\Lambda_t^{4D}]&\le N^{\epsilon}|\alpha_2|^{2D}{\E[ \Lambda_t^{2D}]} + N^{(3D+1)\epsilon} \beta^{2D} |\alpha_2|^{2D} +N^{(3D+1)\epsilon} \beta^{2D} \E \left[ \Lambda_t^{2D} \right]+N^{(3D+1) \epsilon} q_t^{-8D}\\ &\qquad+ N^{-D \epsilon/8+\epsilon} q_t^{-2D} \E \big[\Lambda_t^{4D}\big]\,,
\end{split}\end{align}
uniformly on $\mathcal{D}$. Applying Schwarz inequality to the first term and the third term on the right, absorbing the terms $o(1)\E[\Lambda_t^{4D}]$ into the left side and using $q_t^{-2}\le\beta$ in the fourth term, we get
\begin{align}\begin{split}\label{ESCL}
\E [\Lambda_t^{4D}]&\le N^{2\epsilon}|\alpha_2|^{4D} + N^{(3D+2)\epsilon} \beta^{2D} |\alpha_2|^{2D} +N^{(3D+2)\epsilon} \beta^{4D}\,,
\end{split}\end{align}
uniformly on $\mathcal{D}$. Feeding~\eqref{ESCL} back into~\eqref{eq:claim'} we get, for any $D\ge 10$ and (small) $\epsilon>0$,
\begin{align} \begin{split} \label{eq:claim''}
\E \left[|P(m_t)|^{2D}\right]
&\leq N^{3D\epsilon} \beta^{2D} |\alpha_2|^{2D} + N^{3D\epsilon} \beta^{2D} \E \left[ \Lambda_t^{2D} \right] + N^{(3D+1) \epsilon} \beta^{4D} + q_t^{-2D} |\alpha_2|^{4D} \\
&\le N^{5D\epsilon} \beta^{2D} |\alpha_2|^{2D} + N^{5D \epsilon} \beta^{4D} + q_t^{-2D} |\alpha_2|^{4D} \,,
\end{split} \end{align}
uniformly on $\mathcal{D}$, for $N$ sufficiently large, where we used Schwarz inequality and once more~\eqref{ESCL} to get the second line.
By Markov's inequality, we therefore obtain from~\eqref{eq:claim''} that for fixed $z\in\mathcal{D}$, $|P(m_t)|\prec |\alpha_2|\beta+\beta^2 + q_t^{-1} |\alpha_2|^2$. It then follows from the Taylor expansion of $P(m_t)$ around $\wt m_t$ in~\eqref{le taylor of P} that
\begin{align} \begin{split} \label{eq:claim'''}
|\alpha_2(m_t-\wt m_t)+(m_t-\wt m_t)^2|\prec \beta{\Lambda_t^2}+|\alpha_2|\beta+\beta^2 + q_t^{-1}|\alpha_2|^2,
\end{split}
\end{align}
for each fixed $z\in\mathcal{D}$, where we used that $q_t^{-2}\le \beta$. To get a uniform bound on $\mathcal{D}$, we choose $18 N^{8}$~lattice points $z_1, z_2,\ldots, z_{18 N^{8}}$ in $\caD$ such that, for any $\wt z \in \caD$, there exists $z_n$ satisfying $|\wt z-z_n| \leq N^{-4}$. Since
$$
|m_t(\wt z) - m_t(z_n)| \leq |\wt z-z_n| \sup_{z \in \caD} \left| \frac{\partial m_t(z)}{\partial z} \right| \leq |\wt z-z_n| \sup_{z \in \caD} \frac{1}{(\im z)^2} \leq N^{-2}
$$
and since a similar estimate holds for $|\wt m_t(\wt z) - \wt m_t(z)|$, a union bound yields that~\eqref{eq:claim'''} holds uniformly on $\mathcal{D}$ with high probability. In particular, for any (small) $\epsilon>0$ and (large) $D$ there is an event $\wt\Xi$ with $\P(\wt\Xi)\ge 1-N^D$ such that, for all $z\in\mathcal{D}$,
\begin{align}\label{SCL}
|\alpha_2(m_t-\wt m_t)+(m-m_t)^2|\le N^{\epsilon}\beta{\Lambda_t^2}+N^{\epsilon}|\alpha_2|\beta+N^{\epsilon}\beta^2 +N^{\epsilon} q_t^{-1}|\alpha_2|^2,
\end{align}
on $\wt\Xi$, for $N$ sufficiently large.
Recall next that there is a constant $C_0>1$ such that $C_0^{-1}\sqrt{\kappa_t(E)+\eta}\le |\alpha_2|\le C_0 \sqrt{\kappa_t(E)+\eta}$, where we can choose $C_0$ uniform in $z\in\mathcal{E}$. Note further that $\beta=\beta(E+\ii\eta)$ is for fixed $E$ a decreasing function of~$\eta$ while $\sqrt{\kappa_t(E)+\eta}$ is increasing. Thus, there exists $\wt \eta_0 \equiv \wt \eta_0(E)$ such that $\sqrt{\kappa(E)+ \wt\eta_0} = C_0 q_t \beta(E+\ii \wt\eta_0)$. We then consider the subdomain $\wt\caD \subset \caD$ defined by
\begin{align}
\wt\caD \deq \left\{ z=E+\ii\eta\in\caD\,:\, \eta> \wt\eta_0(E) \right\}.
\end{align}
On $\wt\caD$, $\beta \leq q_t^{-1} |\alpha_2|$, hence we obtain from the estimate~\eqref{SCL} that
\begin{align*}
|\alpha_2(m_t-\wt m_t)+(m-m_t)^2|\le N^{\epsilon}\beta{\Lambda_t^2}+ 3N^{\epsilon} q_t^{-1}|\alpha_2|^2
\end{align*}
and thus
\begin{align*}
|\alpha_2|\Lambda_t\le (1+N^\epsilon\beta)\Lambda_t^2+ 3N^{\epsilon} q_t^{-1}|\alpha_2|^2\,,
\end{align*}
uniformly on $\wt\caD$ on $\wt\Xi$. Hence, we get on $\wt\Xi$ that either
\begin{align}\label{le first dichotomy}
|\alpha_2|\le 4 \Lambda_t\qquad\qquad\textrm{ or }\qquad\qquad \Lambda_t\le 6N^{\epsilon} q_t^{-1}|\alpha_2|\,,\qquad\qquad (z\in\wt\caD)\,.
\end{align}
Note that any $z \in \caE$ with $\eta=\im z = 3$ is in $\wt\caD$. When $\eta =3$, we easily see that
$$
|\alpha_2| \geq |z + 2\wt m_t| - Cq_t^{-2} \geq \eta=3 \gg 6N^{\epsilon} q_t^{-1}|\alpha_2|\,,
$$
for sufficiently large $N$. In particular we have that either $3/4\le\Lambda_t$ or $\Lambda_t\le 6N^{\epsilon} q_t^{-1}|\alpha_2|$ on $\wt\Xi$ for $\eta=3$. Moreover, since $m_t$ and $\wt m_t$ are Stieltjes transforms, we have
$$
\Lambda_t \leq \frac{2}{\eta}=\frac{2}{3}\,.
$$
We conclude that, for $\eta=3$, the second possibility, $\Lambda_t\le 6N^{\epsilon} q_t^{-1}|\alpha_2|$ holds on $\wt\Xi$. Since $6N^{\epsilon} q_t^{-1} \ll 1$ on $\wt\caD$, in particular $6N^{\epsilon} q_t^{-1}|\alpha_2| < |\alpha_2|/8$, we find from~\eqref{le first dichotomy} by continuity that
\begin{align}\label{eq:claim' mod}
\Lambda_t \le 6N^{\epsilon} q_t^{-1}|\alpha_2|\,,\qquad\qquad (z\in\wt\caD)\,,
\end{align}
holds on the event $\wt\Xi$. Putting the estimate~\eqref{eq:claim' mod} back into~\eqref{eq:claim'}, we find that
\begin{align} \begin{split} \label{eq:claim4}
\E\left[|P(m_t)|^{2D}\right] &\leq N^{4D\epsilon} \beta^{2D} |\alpha_2|^{2D} + N^{3D \epsilon} q_t^{-8D} + q_t^{-6D} |\alpha_2|^{4D} \\
&\leq N^{6D\epsilon} \beta^{2D} |\alpha_2|^{2D} + N^{6D \epsilon} \beta^{4D} \,,
\end{split} \end{align}
for any (small) $\epsilon>0$ and (large) $D$, uniformly on $\wt\caD$. Note that, for $z \in \caD \backslash \wt\caD$, the estimate $\E[|P(m_t)|^{2D}] \leq N^{6D\epsilon} \beta^{2D} |\alpha_2|^{2D} + N^{6D \epsilon} \beta^{4D}$ can be directly checked from \eqref{eq:claim''}. Considering lattice points $\{ z_i \} \subset \caD$ again, a union bound yields for any (small) $\epsilon>0$ and (large) $D$ there is an event~$\Xi$ with $\P(\Xi)\ge 1-N^D$ such that
\begin{align}\label{SCL2}
|\alpha_2(m_t-\wt m_t)+(m-m_t)^2|\le N^{\epsilon}\beta{\Lambda_t^2}+N^{\epsilon}|\alpha_2|\beta+N^{\epsilon}\beta^2,
\end{align}
on $\Xi$, uniformly on $\caD$ for $N$ sufficiently large.
Next, recall that $\beta=\beta(E+\ii\eta)$ is for fixed $E$ a decreasing function of~$\eta$ while $\sqrt{\kappa_t(E)+\eta}$ is increasing. Thus there is $\eta_0\equiv\eta_0(E)$ such that $\sqrt{\kappa(E)+\eta_0}=10C_0N^\epsilon\beta(E+\ii\eta_0)$. Further notice that~$\eta_0(E)$ is a continuous function. We consider the three subdomains of $\caE$ defined by
\begin{align}\nonumber\begin{split}
\caE_1&\deq\left\{z=E+\ii\eta\in\caE\,:\, \eta\le \eta_0(E),10N^{\epsilon} \le N\eta\right\},\\
\caE_2&\deq\left\{z=E+\ii\eta\in\caE\,:\, \eta> \eta_0(E),10N^{\epsilon} \le N\eta\right\},\\
\caE_3&\deq\left\{z=E+\ii\eta\in\caE\,:10N^{\epsilon}> N\eta\right\}.
\end{split}\end{align}
Note that $\caE_1\cup\caE_2\subset\mathcal{D}$. We split the stability analysis of~\eqref{SCL2} according to whether $z\in\caE_1$, $\caE_2$ or $\caE_3$.
{\it Case 1:} If $z\in \caE_1$, we note that $|\alpha_2|\le C_0 \sqrt{\kappa(E)+\eta}\le 10C_0^2 N^\epsilon\beta(E+\ii\eta)$. We then obtain from~\eqref{SCL2} that
\begin{align}
\Lambda_t^2\le |\alpha_2|\Lambda_t+N^{\epsilon}\beta\Lambda_t^2+N^{\epsilon}|\alpha_2|\beta+ N^{\epsilon} \beta^2\le 10C_0^2N^{\epsilon}\beta\Lambda_t+N^{\epsilon}\beta\Lambda_t^2+(10C_0^2N^\epsilon+1)N^{\epsilon}\beta^2\,,\nonumber
\end{align}
on $\Xi$. Thus,
\begin{align}\label{jjj}
\Lambda_t\le CN^{\epsilon}\beta\,,\qquad\qquad (z\in\caE_1)\,,
\end{align}
on $\Xi$, for some finite constant $C$.
{\it Case 2:} If $z\in\caE_2$, we obtain from~\eqref{SCL2} that
\begin{align}
|\alpha_2|\Lambda_t\le (1+N^\epsilon\beta)\Lambda_t^2+|\alpha_2|N^{\epsilon}\beta+N^{\epsilon}\beta^2\,,
\end{align}
on $\Xi$. We then note that $C_0|\alpha_2|\ge \sqrt{\kappa_t(E)+\eta}\ge 10C_0 N^\epsilon\beta$, \ie $N^\epsilon\beta\le |\alpha_2|/10$, so that
\begin{align}
|\alpha_2|\Lambda_t\le 2 \Lambda_t^2+(1+ N^{-\epsilon})|\alpha_2|N^{\epsilon}\beta\,,
\end{align}
on $\Xi$, where we used that $N^\epsilon\beta\le 1$. Hence, we get on $\Xi$ that either
\begin{align}
|\alpha_2|\le 4 \Lambda_t\qquad\qquad\textrm{ or }\qquad\qquad \Lambda_t\le 3N^\epsilon\beta\,,\qquad\qquad (z\in\caE_2)\,.
\end{align}
We follow the dichotomy argument and the continuity argument that were used to obtain~\eqref{eq:claim' mod}. Since $3N^{\epsilon} \beta\le |\alpha_2|/8$ on $\caE_2$, we find by continuity that
\begin{align}\label{jjj2}
\Lambda_t \le 3N^{\epsilon} \beta\,,\qquad\qquad (z\in\caE_2)\,,
\end{align}
holds on the event $\Xi$.
{\it Case 3:} For $z\in\caE_3=\caE\backslash (\caE_1\cup\caE_2)$ we use that $|m'_t(z)|\le \im m_t(z)/\im z$, $z\in\C^+$, since $m_t$ is a Stieltjes transform. Set now $ \wt\eta\deq 10N^{-1+\epsilon}$. By the fundamental theorem of calculus we can estimate
\begin{align}\begin{split}\nonumber
|m_t(E+\ii\eta)|&\le \int_{\eta}^{\wt\eta}\frac{\im m_t(E+\ii s)}{s}\dd s+|m_t(E+\ii\wt\eta) |\\
&\le \int_{\eta}^{\wt\eta}\frac{s \im m_t(E+\ii s)}{s^2}\dd s+\Lambda_t(E+\ii\wt\eta)+|\wt m_t(E+\ii\wt\eta)|\,.
\end{split}\end{align}
Using that $s\rightarrow s \im m_t(E+\ii s)$ is a monotone increasing function as is easily checked from the definition of the Stieltjes transform, we find that
\begin{align}\begin{split}
|m_t(E+\ii\eta)|&\le \frac{2\wt\eta}{\eta} \im m_t(E+\ii\wt\eta)+\Lambda_t(E+\ii\wt\eta)+|\wt m_t(E+\ii\wt\eta)|\\
&\le C\frac{N^{\epsilon}}{N\eta}\big(\im \wt m_t(E+\ii\wt \eta)+\Lambda_t(E+\ii\wt \eta)\big) +|\wt m_t(E+\ii\wt\eta)|\,,
\end{split}\end{align}
for some $C$ where we used $\wt\eta\deq 10N^{-1+\epsilon}$ to get the second line. Thus noticing that $z=E+\ii\wt \eta\in \caE_1\cup\caE_2$, hence, on the event $\Xi$ introduced above, we have $\Lambda_t(E+\ii\wt\eta)\le CN^\epsilon\beta(E+\ii\wt \eta) \leq C$ by~\eqref{jjj} and~\eqref{jjj2}. Using moreover that $\wt m_t$ is uniformly bounded by a constant on $\caE$, we then get that, on the event $\Xi$,
\begin{align}\label{jjj3}
\Lambda_t\le CN^\epsilon\beta\,,\qquad \qquad( z\in\caE_3)\,.
\end{align}
Combining~\eqref{jjj},~\eqref{jjj2} and~\eqref{jjj3}, and recalling the definition of the event $\Xi$, we get $\Lambda_t\prec \beta$, uniformly on $\caE$ for fixed $t\in[0,6\log N]$. Choosing $t=0$, we have completed the proof of Theorem~\ref{theorem:local}. To extend this bound to all $t\in[0,6\log N]$, we use the continuity of the Dyson matrix flow. Choosing a lattice $\mathcal{L}\subset[0,6\log N]$ with spacings of order $N^{-3}$, we get $\Lambda_t\prec \beta$, uniformly on $\mathcal{E}$ and on $\mathcal{L}$, by a union bound. By continuity we extend the conclusion to all of $[0,6\log N]$. This proves~Proposition~\ref{prop:local}.
\end{proof}
\subsection{Proof of Theorem~\ref{prop:norm bound}}
We start with an upper bound on the largest eigenvalue $\lambda_1^{H_t}$ of $H_t$.
\begin{lemma} \label{le lemma norm bound} Let $H_0$ satisfy Assumption~\eqref{assumption H} with $\phi>0$.
Let $L_t$ be deterministic number defined in Lemma \ref{lem:w}. Then,
\begin{align} \label{eq:norm bound}
\lambda_1^{H_t} - L_t\prec \frac{1}{q_t^4} +\frac{1}{N^{2/3}}\,,
\end{align}
uniformly in $t\in[0,6\log N]$.
\end{lemma}
\begin{proof}
To prove Lemma~\ref{le lemma norm bound} we follow the strategy of the proof of Lemma~4.4 in~\cite{EKYY1}.
Fix $t\in[0,6\log N]$. Recall first the deterministic $z$-dependent parameters
\begin{align}
\alpha_1 \deq \im \wt m_t\,, \qquad \alpha_2 \deq P'(\wt m_t)\,, \qquad \beta \deq\frac{1}{q_t^2}+ \frac{1}{N\eta} \,.
\end{align}
We mostly drop the $z$-dependence for brevity. We further introduce the $z$-independent quantity
\begin{align}
\wt\beta \deq \left( \frac{1}{q_t^4}+\frac{1}{N^{2/3}} \right)^{1/2}.
\end{align}
Fix a small $\epsilon > 0$ and define the domain $\caD_{\epsilon}$ by
\begin{align}
\caD_{\epsilon} \deq \bigg \{ z = E + \ii \eta : N^{4\epsilon}\wt\beta^2 \leq \kappa_t \leq q_t^{-1/3}\,, \; \eta = \frac{N^{\epsilon}}{N \sqrt{\kappa_t}} \bigg\}\,,
\end{align}
where $\kappa_t\equiv \kappa_t(E)=E-L_t$. Note that on $\caD_\epsilon$,
\begin{align*}
N^{-1+\epsilon}\ll \eta \leq \frac{N^{ -{\epsilon}}}{N\wt\beta}\,, \qquad\qquad \kappa \geq N^{5\epsilon} \eta\,.
\end{align*}
In particular we have $N^\epsilon\wt\beta\le (N\eta)^{-1}$, hence $N^\epsilon q_t^{-2}\le C (N\eta)^{-1}$ so that $q_t^{-2}$ is negligible when compared to $(N\eta)^{-1}$ and $\beta$ on $\caD_\epsilon$. Note moreover that
\begin{align}\label{the alphas}\begin{split}
|\alpha_2| &\sim \sqrt{\kappa_t+\eta}\sim\sqrt{\kappa_t}= \frac{N^{\epsilon} }{N\eta}\sim N^\epsilon\beta\,,\\
\alpha_1 &= \im \wt m_t \sim \frac{\eta}{\sqrt{\kappa_t + \eta}}\sim\frac{\eta}{\sqrt{\kappa_t}} \leq N^{-5\epsilon}\sqrt{\kappa_t} \sim N^{-5\epsilon} |\alpha_2| \sim N^{-4\epsilon} \beta\,.
\end{split}\end{align}
In particular we have $\alpha_1\ll |\alpha_2|$ on $\caD_\epsilon$.
We next claim that
\begin{align*}
\Lambda_t\deq |m_t - \wt m_t| \ll \frac{1}{N\eta}
\end{align*}
with high probability on the domain $\caD_{\epsilon}$.
Since $\caD_{\epsilon} \subset \caE$, we find from Proposition~\ref{prop:local} that $\Lambda_t \leq N^{\epsilon'}\beta$ for any $\epsilon' > 0$ with high probability. Fix $0 < \epsilon' < \epsilon/7 $. From \eqref{eq:claim'}, we get
\begin{align}
\E \left[|P(m_t)|^{2D}\right]&\leq C N^{(4D-1)\epsilon'} \E \left[ \beta^{2D} (\alpha_1 + \Lambda_t)^D (|\alpha_2| + 3\Lambda_t)^D \right] \nonumber\\ &\qquad\quad+ \frac{N^{(2D+1)\epsilon'}}{D} q_t^{-8D} + \frac{N^{-(D/4 -1)\epsilon'}}{D} q_t^{-2D} \E \left[\Lambda_t^{4D}\right]\nonumber \\
&\leq C^{2D} N^{6D \epsilon'} \beta^{4D} + \frac{N^{(2D+1)\epsilon'}}{D} q_t^{-8D} + \frac{N^{4D\epsilon'}}{D} q_t^{-2D} \beta^{4D}\nonumber\\
&\leq C^{2D} N^{6D \epsilon'} \beta^{4D} \,,\nonumber
\end{align}
for $N$ sufficiently large, where we used that $\Lambda_t\le N^{\epsilon'}\beta\ll N^\epsilon\beta$ with high probability and, by~\eqref{the alphas}, $\alpha_1\ll|\alpha_2|$, $|\alpha_2|\le CN^\epsilon\beta$ on $\caD_\epsilon$.
Applying $(2D)$-th order Markov inequality and a simple lattice argument combined with a union bound, we get $|P(m_t)| \leq CN^{3\epsilon'} \beta^{2}$ uniformly on $\caD_\epsilon$ with high probability. From the Taylor expansion of $P(m_t)$ around $\wt m_t$ in~\eqref{le taylor of P}, we then get that
\begin{align}\label{choco}
|\alpha_2| \Lambda_t \leq 2 \Lambda_t^2 + CN^{3\epsilon'} \beta^{2}\,,
\end{align}
uniformly on $\caD_\epsilon$ with high probability, where we also used that $\Lambda_t\ll 1$ on $\caD_\epsilon$ with high probability.
Since $\Lambda_t\le N^{\epsilon'} \beta\le C N^{\epsilon'-\epsilon}|\alpha_2|$ with high probability on $\caD_\epsilon$, we have $|\alpha_2|\Lambda_t\ge CN^{\epsilon-\epsilon'}\Lambda_t^2\gg 2\Lambda_t^2$. Thus the first term on the right side of~\eqref{choco} can be absorbed into the left side and we conclude that
\begin{align}\nonumber
\Lambda_t \leq C N^{3\epsilon'}\frac{\beta}{|\alpha_2|}\beta\le CN^{3\epsilon'-\epsilon}\beta\,,
\end{align}
must hold with high probability on $\caD_\epsilon$. Hence, using that $0<\epsilon'<\epsilon/7$, we obtain that
\begin{align}\nonumber
\Lambda_t \leq N^{-\epsilon/2}\beta\le 2\frac{N^{-\epsilon/2}}{N\eta}\,,
\end{align}
with high probability on $\caD_\epsilon$. This proves the claim that $\Lambda_t \ll (N\eta)^{-1}$ on $\caD_{\epsilon}$ with high probability. Moreover, this also shows that
\begin{align}\label{the neck}
\im m_t \leq \im \wt m_t + \Lambda_t=\alpha_1+\Lambda_t \ll\frac{1}{N\eta}\,,
\end{align}
on $\caD_{\epsilon}$ with high probability, where we used~\eqref{the alphas}.
Now we prove the estimate \eqref{eq:norm bound}. If $\lambda_1^{H_t} \in [E-\eta, E+\eta]$ for some $E \in [L_t +N^\epsilon(q_t^{-4}+ N^{-2/3}), L_t +{q}^{-1/3}]$ with $z = E + \ii \eta \in \caD_{\epsilon}$,
\begin{align}
\im m_t (z) \geq \frac{1}{N} \im \frac{1}{\lambda_1^{H_t} -E - \ii \eta} = \frac{1}{N} \frac{\eta}{(\lambda_1^{H_t}-E)^2 + \eta^2} \geq \frac{1}{5N\eta}\,,
\end{align}
which contradicts the high probability bound $\im m_t \ll (N\eta)^{-1}$ in~\eqref{the neck}. The size of each interval $[E-\eta, E+\eta]$ is at least $N^{{ -1+\epsilon}} q_t^{1/6}$. Thus, considering $O({ N})$ such intervals, we can conclude that $\lambda_1 \notin [L + N^\epsilon(q_t^{-4}+N^{-2/3}), L +q_t^{-1/3}]$ with high probability. From Proposition~\ref{a priori norm bound}, we find that $\lambda_1^{H_t}-L_t\prec q_t^{-1/3}$ with high probability, hence we conclude that \eqref{eq:norm bound} holds, for fixed $t\in[0,6\log N]$. Using a lattice argument and the continuity of the Dyson matrix flow, we easily obtain~\eqref{eq:norm bound} uniformly in $t\in[0,6\log N]$.
\end{proof}
We are now well-prepared for the proof of Theorem~\ref{prop:norm bound}. It follows immediately from the next result.
\begin{lemma}\label{lemma friday evenig}
Let $H_0$ satisfy Assumption~\eqref{assumption H} with $\phi>0$. Then, uniformly in $t\in[0,6\log N]$,
\begin{align} \label{eq:norm bound t}
\left|\| H_t\| - L_t\right|\prec \frac{1}{q_t^4} +\frac{1}{N^{2/3}}\,.
\end{align}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lemma friday evenig} and Theorem~\ref{prop:norm bound}]
Fix $t\in[0,\log N]$. Consider the largest eigenvalue $\lambda_1^{H_t}$. In Proposition~\ref{le lemma norm bound}, we already showed that $(L_t-\lambda_1^{H_t})_-\prec q_t^{-4}+N^{-2/3}$. It thus suffices to consider $(L_t-\lambda_1^{H_t})_+$. By Lemma~\ref{lem:w} there is $c>0$ such that $ c(L_t-\lambda_1^{H_t})_+^{3/2}\le n_{\wt\rho_t}(\lambda_1^{H_t},L_t)$. Hence, by Corollary~\ref{corollary density of states} (and its obvious generalization to $H_t$), we have the estimate
\begin{align}
(L_t-\lambda_1^{H_t})_+^{3/2}\prec \frac{(L_t-\lambda_1^{H_t})_+}{q_t^2}+\frac{1}{N}\,,
\end{align}
so that $(L_t-\lambda_1^{H_t})_+\prec q_t^{-4}+N^{-2/3}$. Thus $|\lambda_1^{H_t}-L_t|\prec q_t^{-4}+N^{-2/3}$. Similarly, one shows the estimate $|\lambda_N^{H_t}+L_t|\prec q_t^{-4}+N^{-2/3}$ for the smallest eigenvalue $\lambda_N^{H_t}$. This proves~\eqref{eq:norm bound t} for fixed $t\in[0,6\log N]$. Uniformity follows easily from the continuity of the Dyson matrix flow.
\end{proof}
\section{Recursive moment estimate: Proof of Lemma \ref{lem:claim}} \label{sec:stein}
In this section, we prove Lemma \ref{lem:claim}. Recall the definitions of the Green functions $G_t$ and $m_t$ in~\eqref{le timedependent green functions}. We fix $t\in[0,6\log N]$ throughout this section, and we will omit $t$ from the notation in the matrix~$H_t$, its matrix elements and its Green functions. Given a (small) $\epsilon>0$, we introduce the $z$-dependent control parameter $\Phiepsi\equiv\Phiepsi(z)$ by setting
\begin{align} \begin{split} \label{eq:phi}
\Phiepsi(z) &\deq N^{\epsilon} \, \E \bigg[ \bigg(\frac{1}{q_t^4}+ \frac{\im m_t}{N\eta} \bigg) \big| P(m_t) \big|^{2D-1} \bigg] + N^{-\epsilon/4} q_t^{-1} \, \E \bigg[ |m_t - \wt m_t|^2 \big| P(m_t) \big|^{2D-1} \bigg] \\
& \qquad+ {N^{\epsilon}}{ q_t}^{-1}\, \sum_{s=2}^{2D} \sum_{s'=0}^{s-2} \E \bigg[ \bigg( \frac{\im m_t}{N\eta} \bigg)^{2s-s'-2} \big| P'(m_t) \big|^{s'} \big| P(m_t) \big|^{2D-s} \bigg] +N^{\epsilon} q_t^{-8D}\\
&\qquad+ N^{\epsilon} \, \sum_{s=2}^{2D} \E \bigg[ \bigg( { \frac{1}{N\eta} } + \frac{1}{q_t}\bigg({\frac{\im m_t}{N\eta}}\bigg)^{1/2} + \frac{1}{q_t^{2}} \bigg) \bigg( \frac{\im m_t}{N\eta} \bigg)^{s-1} \big| P'(m_t) \big|^{s-1} \big| P(m_t) \big|^{2D-s} \bigg] \,.
\end{split} \end{align}
Recall the domain $\mathcal{D}$ defined in~\eqref{domain}. Lemma~\ref{lem:claim} then states that, for any (small) $\epsilon>0$,
\begin{align}
\E [ |P|^{2D}(z) ] \le\Phiepsi(z)\,,\qquad\qquad( z\in\mathcal{D})\,,
\end{align}
for $N$ sufficiently large. We say that a {\it random variable $Z$ is negligible} if $|\E[Z]| \leq C \Phiepsi$ for some $N$-independent constant $C$.
To prove the recursive moment estimate of Lemma~\ref{lem:claim}, we return to~\eqref{ziwui eins} which reads
\begin{align}\begin{split}\label{ziwui}
\E\big[(1+ zm)P^{{D-1}}\ol{P^D}\big]&= \frac{1}{N} \sum_{r=1}^\ell \frac{\kappa_t^{(r+1)}}{r!} \E \bigg[ \sum_{i \neq k} \partial_{ik}^r \Big( G_{ki} P^{D-1} \ol{P^D} \Big) \bigg]+\E\Omega_{\ell}\Big((1+zm) P^{D-1} \ol{P^D} \Big)\,,
\end{split}\end{align}
where $\partial_{ik} = \partial/(\partial H_{ik})$ and $\kappa_t^{(k)}\equiv\kappa_t^{(k)}$ are the cumulants of $(H_t)_{ij}$, $i\not=j$. The detailed form of the error $\E\Omega_\ell(\cdot)$ is discussed in Subsection~\ref{subsection truncation of the cumulant expansion}.
It is convenient to condense the notation a bit. Abbreviate
\begin{align}\label{first time Ir}
\Ir\equiv \Ir(z,m,D)\deq (1+zm) P(m)^{D-1}\ol{P(m)^D}\,.
\end{align}
We rewrite the cumulant expansion~\eqref{ziwui} as
\begin{align}\label{the IR}
\E\Ir=\sum_{r=1}^\ell\sum_{s=0}^r w_{\Ir_{r,s}}\E\Ir_{r,s}+\E\Omega_\ell(I)\,,
\end{align}
where we set
\begin{align}\begin{split}\label{e1}
\Ir_{r,s}\deq {N\kappa_t^{(r+1)}} \frac{1}{N^2} \sum_{i \neq k} \big( \partial_{ik}^{r-s} G_{ki} \big) \big( \partial_{ik}^s \big( P^{D-1} \ol{P^D} \big) \big)\,.
\end{split}\end{align}
(By convention $\partial_{ik}^0G_{ki}=G_{ki} $.) The weights $ w_{\Ir_{r,s}}$ are combinatoric coefficient given by
\begin{align}
w_{\Ir_{r,s}}\deq\frac{1}{r!}\binom{r}{s}=\frac{1}{(r-s)!s! } \,.
\end{align}
Returning to~\eqref{the source}, we have in this condensed form the expansion
\begin{align}\begin{split}\label{the short guy}
\E [ |P|^{2D} ] &= \sum_{r=1}^\ell\sum_{s=0}^r w_{\Ir_{r,s}}\E\Ir_{r,s}+ \E \bigg[ \Big( m^2 + \frac{\nm^{(4)}}{q_t^{2}} m^4 \Big) P^{D-1} \ol{P^D} \bigg] + \E\Omega_{\ell}(I)\,.
\end{split}\end{align}
\subsection{Truncation of the cumulant expansion}\label{subsection truncation of the cumulant expansion}
In this subsection, we bound the error term $\E\Omega_\ell(I)$ in~\eqref{the short guy} for large $\ell$. We need some more notation. Let $E^{[ik]}$ denote the $N\times N$ matrix determined~by
\begin{align}
(E^{[ik]})_{ab}=\begin{cases}\delta_{ia}\delta_{kb}+\delta_{ib}\delta_{ka} \qquad &\textrm{if}\; i\not=k\,,\\
\delta_{ia}\delta_{ib} &\textrm{if}\; i=k\,,
\end{cases}
\qquad\qquad (i,k,a,b\in\llbracket1,N\rrbracket)\,.
\end{align}
For each pair of indices $(i,k)$, we define the matrix $H^{(ik)}$ from $H$ through the decomposition
\begin{align}\label{s notation}
H=H^{(ik)}+H_{ik} E^{[ik]}\,.
\end{align}
With this notation we have the following estimate.
\begin{lemma}\label{le lemma first bound for error terms}
Suppose that $H$ satisfies Assumption~\ref{assumption H} with $\phi>0$. Let $i,k\in\llbracket1,N\rrbracket$, $D\in\N$ and $z\in\caD$. Define the function $F_{ki}$ by
\begin{align}\label{le F function}
F_{ki}(H)\deq G_{ki} P^{D-1}\ol P^D\,,
\end{align}
where $G\equiv G^H(z)$ and $P\equiv P(m(z))$. Choose an arbitrary $\ell\in\N$. Then, for any (small) $\epsilon>0$,
\begin{align}\label{le first bound for error terms}
\E\bigg[\sup_{x\in\R, \,|x|\le q_t^{-1/2}}|\partial_{ik}^\ell F_{ki}(H^{(ik)}+xE^{[ik]})|\bigg]\le N^\epsilon\,,
\end{align}
uniformly $z\in\mathcal{D}$, for $N$ sufficiently large. Here $\partial_{ik}^\ell$ denotes the partial derivative $\frac{\partial^\ell}{\partial H_{ik}^\ell}$.
\end{lemma}
\begin{proof}
Fix two pairs of indices $(a,b)$ and $(i,k)$. From the definition of the Green function and~\eqref{s notation} we easily get
\begin{align*}
G^{H^{(ik)}}_{ab}=G^{H}_{ab}+H_{ik}\big(G^{H^{(ik)}}E^{[ik]}G^{H}\big)_{ab}=G^{H}_{ab}+H_{ik}G_{ai}^{H^{(ik)}}G_{kb}^{H}+H_{ik} G_{ak}^{H^{(ik)}}G_{ib}^{H}\,,
\end{align*}
where we omit the $z$-dependence. Letting $\Lambda_o^{H^{(ik)}}\deq\displaystyle{\max_{a,b}}| G^{H^{(ik)}}_{ab}|$ and $\Lambda_o^{H}\deq\displaystyle{\max_{a,b}}| G^{H}_{ab}|$, we get
\begin{align*}
\Lambda_o^{H^{(ik)}}\prec \Lambda_o^{H}+\frac{1}{q_t}\Lambda_o^{H}\Lambda_o^{H^{(ik)}}\,.
\end{align*}
By~\eqref{le moment condition} we have $|H_{ik}|\prec q_t^{-1}$ and by~\eqref{le EYYY1 2} we have $\Lambda_o^{H}\prec 1$, uniformly in $z\in\mathcal{D}$. It follows that $\Lambda_o^{H^{(ik)}}\prec \Lambda_o^{H}\prec 1$, uniformly in $z\in\mathcal{D}$, where we used~\eqref{le EYYY1 2}. Similarly, for $x\in\R$, we have
\begin{align*}
G^{H^{(ik)}+xE^{[ik]}}_{ab}=G^{H^{(ik)}}_{ab}-x\big(G^{H^{(ik)}}E^{[ik]}G^{H^{(ik)}+xE^{[ik]}}\big)_{ab}\,,
\end{align*}
and we get
\begin{align}\label{le nilpferd}
\sup_{|x|\le q_t^{-1/2}}\max_{a,b}|G^{H^{(ik)}+xE^{[ik]}}_{ab}|\prec \Lambda_o^{H^{(ik)}}\prec 1\,,
\end{align}
uniformly in $z\in\mathcal{D}$, where we used once more~\eqref{le EYYY1 2}.
Recall that $P$ is a polynomial of degree $4$ in $m$. Then $F_{ki}$ is a multivariate polynomial of degree $4(2D-1)+1$ in the Green function entries and the normalized trace $m$ whose number of member terms is bounded by $4^{2D-1}$. Hence $\partial_{{ik}}^{\ell}F_{ki}$ is a multivariate polynomial of degree $4(2D-1)+1+\ell$ whose number of member terms is roughly bounded by $4^{2D-1}\times(4(2D-1)+1+2l)^l$. Next, to control the individual monomials in~$\partial_{{ik}}^{\ell}F_{ki}$, we apply~\eqref{le nilpferd} to each factor of Green function entries (at most $4(2D-1)+1+\ell$ times). Thus, altogether we obtain
\begin{align}\label{le biranimi}
\E\bigg[\sup_{|x|\le q_t^{-1/2}}|(\partial_{{ik}}^{\ell}F_{ki})(H^{(ik)}+xE^{[ik]})|\bigg]\le 4^{2D}(8D+\ell) N^{(8D+\ell)\epsilon'}\,,
\end{align}
for any small $\epsilon'>0$ and sufficiently large $N$. Choosing $\epsilon'=\epsilon/(2(8D+\ell))$ with get~\eqref{le first bound for error terms}.
\end{proof}
Recall that we set $\Ir= {(1+zm)} P(m)^{D-1}\overline{P(m)^D}$ in~\eqref{first time Ir}. To control the error term $\E\Omega_{\ell}(I)$ in~\eqref{the short guy}, we use the following result.
\begin{corollary}\label{le corollary for error terms omega}
Let $\E\Omega_\ell(I)$ be as in~\eqref{the short guy}. With the assumptions and notation of Lemma~\ref{le lemma first bound for error terms}, we have, for any (small) $\epsilon>0$,
\begin{align}\label{final bound on the error term in the Stein}
|\E \Omega_{\ell}\big( I\big)\big|\le N^\epsilon \left(\frac{1}{q_t}\right)^\ell,
\end{align}
uniformly in $z\in\mathcal{D}$, for $N$ sufficiently large.
In particular, the error $\E\Omega_{\ell}(I)$ is negligible for $\ell\ge 8D$.
\end{corollary}
\begin{proof}
First, fix a pair of indices $(k,i)$, $k\not=i$. Recall the definition of $F_{ik}$ in~\eqref{le F function}. Denoting $\E_{ik}$ the partial expectation with respect to $H_{ik}$, we have from Lemma~\ref{le stein lemma}, with $Q=q_t^{-1/2}$,
\begin{align}\label{le error term in general stein}\begin{split}
|\E_{ik}\Omega_\ell(H_{ik}F_{ki})|&\le C_\ell \E_{ik}[ |H_{ik}|^{\ell+2}]\sup_{|x|\le q_t^{-1/2}}|\partial_{ik}^{\ell+1}F_{ki}(H^{(ik)}+xE^{[ik]})|\\ &\qquad+ C_\ell \E_{ik} [|H_{ik}|^{\ell+2} \lone(|H_{ik}|>q_t^{-1/2})]\sup_{x\in \R} |\partial_{ik}^{\ell+1}F_{ki}(H^{(ik)}+xE^{[ik]})| \,,
\end{split}\end{align}
with $C_\ell\le (C\ell)^\ell/\ell!$, for some numeral constant $C$. To control the full expectation of the first term on the right side, we use the moment assumption~\eqref{le moment condition} and Lemma~\ref{le lemma first bound for error terms} to conclude that, for any $\epsilon>0$,
\begin{align*}\begin{split}
C_\ell\E \bigg[\E_{ik}[ |H_{ki}|^{\ell+2}]\sup_{|x|\le q_t^{-1/2}}\big|\partial_{ik}^{\ell+1}F_{ki}\big(H^{(ik)}+xE^{[ik]}\big)\big|\bigg]&\le C_\ell\frac{(C(\ell+2))^{c(\ell+2)}}{Nq_t^{\ell}}N^\epsilon\le \frac{N^{2\epsilon}}{Nq_t^{\ell}}\,,
\end{split}\end{align*}
for $N$ sufficiently large. To control the second term on the right side of~\eqref{le error term in general stein}, we use the deterministic bound $\|G(z)\|\le \eta^{-1}$ to conclude that
\begin{align*}
\sup_{x\in\R}\big|\partial_{ik}^{\ell+1}F_{ki}\big(H^{(ik)}+xE^{[ik]}\big)\big|\le 4^{2D}(8D+\ell)\left
(\frac{C}{\eta}\right)^{(8D+\ell)}, \qquad\qquad( z\in\C^+)\,;
\end{align*}
\cf the paragraph above~\eqref{le biranimi}. On the other hand, we have from H\"older's inequality and the moment assumptions in~\eqref{le moment condition} that, for any $D'\in\N$,
\begin{align*}
\E_{ik} [|H_{ik}|^{\ell+2} \lone(|H_{ik}|>q_t^{-1/2})]\le \left(\frac{C}{q}\right)^{D'}\,,
\end{align*}
for $N$ sufficiently large. Using that $q\ge N^\phi$ by~\eqref{assumption H}, we hence obtain, for any $D'\in\N$,
\begin{align}\label{le biranimi 2}
C_\ell \E_{ik} [|H_{ik}|^{\ell+2} \lone(|H_{ik}|>q_t^{-1/2})]\sup_{x\in \R} |\partial_{ik}^{\ell+1}F_{ki}(H^{(ik)}+xE^{[ik]})| \le\left(\frac{C}{q}\right)^{D'}\,,
\end{align}
uniformly on $\C^+$, for $N$ sufficiently large.
Next, summing over $i$, $k$ and choosing $D'\ge \ell $ sufficiently large in~\eqref{le biranimi 2} we obtain, for any~$\epsilon>0$
\begin{align*}
\bigg|\E\bigg[\Omega_\ell\Big( (1+ zm)P^{{D-1}}\ol{P^D}\Big)\bigg]\bigg|=\bigg|\E\bigg[\Omega_\ell\Big( \frac{1}{N}\sum_{i\not=k}H_{ik}F_{ki}\Big)\bigg]\bigg|\le \frac{N^{\epsilon}}{q_t^{\ell}}\,,
\end{align*}
uniformly on $\mathcal{D}$, for $N$ sufficiently large. This proves~\eqref{final bound on the error term in the Stein}.
\end{proof}
\begin{remark}\label{remark on other expansions}
We will also consider slight generalizations of the cumulant expansion in~\eqref{ziwui}. Let $i,j,k\in\llbracket1,N\rrbracket$. Let $n\in \N_0$ and choose indices $a_1,\dots, a_{n},$ $ b_1,\ldots, b_n\in\llbracket1,N\rrbracket$. Let $D\in\N$ and choose $s_1,s_2,s_3,s_4\in\llbracket0,D\rrbracket$. Fix $z\in\mathcal{D}$. Define the function $F_{ki}$ by setting
\begin{align}\label{le F function bis}
F_{ki}(H)\deq G_{ki}\prod_{l=1}^n G_{a_lb_l}P^{D-s_1} \ol{P^{D-s_2}} \left( P' \right)^{s_3} \left( \ol{P'} \right)^{s_4} \,.
\end{align}
It is then straightforward to check that we have cumulant expansion
\begin{align}\label{le general general stein}
\E \bigg[\frac{1}{N} \sum_{i \neq k} H_{ik} F_{ki} \bigg] = \sum_{r=1}^\ell \frac{\kappa_t^{(r+1)}}{r!} \E \bigg[\frac{1}{N} \sum_{i \neq k} \partial_{ik}^r F_{ki} \bigg] +\E\Omega_{\ell}\bigg( \frac{1}{N} \sum_{i\not=k} H_{ik}F_{ki} \bigg)\,,
\end{align}
where the error $\E\Omega_\ell(\cdot)$ satisfies the same bound as in~\eqref{final bound on the error term in the Stein}. This follows easily by extending Lemma~\ref{le lemma first bound for error terms} and Corollary~\ref{le corollary for error terms omega}.
\end{remark}
\subsection{Truncated cumulant expansion}
Armed with the estimates on $\E\Omega_\ell(\cdot)$ of the previous subsection, we now turn to the main terms on the right side of~\eqref{the short guy}.
In the remainder of this section we derive the following result from which Lemma~\ref{lem:claim} follows directly. Recall the definition of $\Phiepsi$ in~\eqref{eq:phi}.
\begin{lemma}\label{summary expansions}
Fix $D\ge 2$ and $\ell\ge 8D$. Let $\Ir_{r,s}$ be given by~\eqref{e1}. Then we have, for any (small) $\epsilon>0$,
\begin{gather}\label{le summary equation 1} \begin{aligned}
w_{\Ir_{1,0}} \E[\Ir_{1,0}]&=-\E\big[m^2 P(m)^{D-1}\ol{P(m)^D}\big]+O(\Phiepsi)\,, & w_{\Ir_{2,0}}\E[\Ir_{2,0}]&=O(\Phiepsi)\,,\\
w_{\Ir_{3,0}}\E[\Ir_{3,0}]&=-\frac{s^{(4)}}{q_t^2}\E\big[ m^4 P(m)^{D-1}\ol{P(m)^D}\big]+O(\Phiepsi)\,, &w_{\Ir_{r,0}}\E[\Ir_{r,0}]&=O(\Phiepsi)\,,\quad\;\; (4\le r\le \ell)\,,
\end{aligned}\end{gather}
uniformly in $z\in\mathcal{D}$, for $N$ sufficiently large. Moreover, we have, for any (small) $\epsilon>0$,
\begin{align}\label{le summary equation 2}
w_{\Ir_{r,s}}|\E[\Ir_{r,s}]|\le \Phiepsi\,, \qquad (1\le s\le r\le \ell)\,,
\end{align}
uniformly in $z\in\mathcal{D}$, for $N$ sufficiently large.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lem:claim}]
By the definition of $\Phiepsi$ in~\eqref{eq:phi} (with a sufficiently large (small) $\epsilon>0$), it suffices to show that $\E [ |P|^{2D}(z) ] \le\Phiepsi(z)$, for all $z\in\mathcal{D}$, for $N$ sufficiently large. Choosing~$\ell\ge 8D$, Corollary~\ref{le corollary for error terms omega} asserts that $\E \Omega_\ell(I)$ in~\eqref{the short guy} is negligible. By Lemma~\ref{summary expansions} the only non-negligible terms in the expansion of the first term on the right side of~\eqref{the short guy} are $w_{\Ir_{1,0}}\E\Ir_{1,0}$ and $w_{\Ir_{3,0}}\E\Ir_{3,0}$, yet these two terms cancel with the middle term on the right side of~\eqref{the short guy}, up to negligible terms. Thus the whole right side of~\eqref{the short guy} is negligible. This proves Lemma~\ref{lem:claim}.
\end{proof}
We now choose an initial (small) $\epsilon>0$. Below we use the factor $N^\epsilon$ to absorb numerical constants in the estimates by allowing $\epsilon$ to increase by a tiny amount from line to line. We often drop $z$ from the notation; it is always understood that $z\in\mathcal{D}$ and all estimates are uniform on $\mathcal{D}$. The proof of Lemma~\ref{summary expansions} is done in the remaining Subsections~\ref{sub:e111}--\ref{sub:e113} where $\E\Ir_{r,s}$ are controlled.
\subsection{Estimate on $\Ir_{1, s}$} \label{sub:e111}
Starting from the definition of $\Ir_{1,0}$ in~\eqref{the IR}, a direct computation yields
\begin{align} \label{ea111}
\E\Ir_{1,0}&=\frac{\kappa_t^{(2)}}{N} \E \bigg[ \sum_{i_1 \neq i_2} \big( \partial_{i_1i_2} G_{i_2i_1} \big) P^{D-1} \ol{P^D} \bigg] = -\E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} (G_{i_2i_2} G_{i_1i_1}+G_{i_1i_2}G_{i_2i_1}) P^{D-1} \ol{P^D} \bigg] \nonumber \\
&= -\E \bigg[ m^2 P^{D-1} \ol{P^D} \bigg] + \E \bigg[\frac{1}{N^2} \sum_{i_1=1}^N (G_{i_1i_1})^2 P^{D-1} \ol{P^D} \bigg] - \E \bigg[ \frac{1}{N^2} \sum_{i_1 \neq i_2} (G_{i_2i_1})^2 P^{D-1} \ol{P^D} \bigg]\,.
\end{align}
The middle term on the last line is negligible since
\begin{align*}
\bigg| \frac{1}{N^2} \E \bigg[ \sum_{i_1=1}^N (G_{i_1i_1})^2 P^{D-1} \ol{P^D} \bigg] \bigg| \leq \frac{N^{\epsilon}}{N} \E \Big[ P^{D-1} \ol{P^D} \Big]\,,
\end{align*}
where we used $|G_{i_1i_1}|\prec 1$, and so is the third term since
\begin{align*}
\frac{1}{N^2} \bigg| \E \bigg[ \sum_{i_1 \neq i_2} (G_{i_2i_1})^2 P^{D-1} \ol{P^D} \bigg] \bigg| \leq N^{\epsilon} \E \bigg[ \frac{\im m}{N\eta} |P|^{2D-1} \bigg]\,,
\end{align*}
where we used Lemma~\ref{lem:2 off-diagonal}. We thus obtain from~\eqref{ea111} that
\begin{align} \begin{split} \label{e111}
\Big|\Ir_{1,0} + \E \big[ m^2 P^{D-1} \ol{P^D}\, \big] \Big| \leq \Phiepsi\,,
\end{split} \end{align}
for $N$ sufficiently large. This proves the first estimate in~\eqref{le summary equation 1}.
Consider next $\Ir_{1,1}$. Similar to~\eqref{expansion wigner 3}, we have
\begin{align*}\begin{split}
\E\Ir_{1,1}=\frac{1}{N^2}\sum_{i_1\not=i_2} \E \bigg[ G_{i_2i_1} \partial_{i_1i_2} ( P^{D-1} \ol{P^D} )\bigg] &= -\frac{2(D-1)}{N^3}\E\bigg[ \sum_{i_1\not=i_2}\sum_{i_3=1}^N G_{i_2i_1} G_{i_2i_3}G_{i_3i_2}P'(m) P^{D-2} \ol{P^D}\bigg] \\ &\qquad- \frac{2D}{N^3} \sum_{i_1\not=i_2}\sum_{i_3=1}^N \E\bigg[G_{i_2i_1} \ol{G_{i_3i_2}G_{i_2i_3}P'(m)}P^{D-1} \ol{P^{D-1}}\bigg]\,.
\end{split}\end{align*}
Here the {\it fresh summation index} $i_3$ originated from $\partial_{i_1i_2}P(m)=P'(m)\frac{1}{N}{\sum_{i_3=1}^N}\partial_{i_1i_2}G_{i_3i_3}$. Note that we can add the terms with $i_1=i_2$ at the expense of a negligible error, so that
\begin{align}\begin{split}\label{le soul}
\E\Ir_{1,1} &= -2(D-1)\E\bigg[\frac{1}{N^3}\Tr G^3 P'(m) P^{D-2} \ol{P^D}\bigg] \\ &\qquad\qquad - 2D \,\E\bigg[\frac{1}{N^3} G (G^*)^2\ol{P'(m)}P^{D-1} \ol{P^{D-1}}\bigg]+O(\Phiepsi)\,.
\end{split}\end{align}
In the remainder of this section, we will freely include or exclude negligible terms with coinciding indices. Using~\eqref{sum rule 1} and~\eqref{sum rule 2} we obtain from~\eqref{le soul} the estimate
\begin{align} \label{e1211}
|\E\Ir_{1,1}|=\bigg| \E \bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2} G_{i_2i_1} \partial_{i_1i_2} ( P^{D-1} \ol{P^D} ) \bigg] \bigg| \leq (4D-2) \E \bigg[ \frac{\im m}{(N\eta)^2} |P'| |P|^{2D-2} \bigg]+\Phiepsi\,,
\end{align}
for $N$ sufficiently large. This proves~\eqref{le summary equation 2} for $r=s=1$.
\subsection{Estimate on $\Ir_{2,0}$} \label{sub:e112}
We start with a lemma that is used in the power counting arguments~below.
\begin{lemma} \label{lem:2 off-diagonal}
For any $i,k\in\llbracket1,N\rrbracket$,
\begin{align}\label{generalized Ward}
\frac{1}{N} \sum_{j=1}^N |G_{ij}(z)G_{jk}(z)| \prec \frac{\im m(z)}{N\eta}\,,\qquad\frac{1}{N}\sum_{j=1}^N|G_{ij}(z)|\prec \left(\frac{\im m(z)}{N\eta} \right)^{1/2}\,,\qquad\qquad (z\in\C^+)\,.
\end{align}
Moreover, for fixed $n\in \N$,
\begin{align}\label{generalized Ward 2}
\frac{1}{N^n}\sum_{j_1,j_2,\ldots ,j_n=1}^N|G_{ij_1}(z)G_{j_1j_2}(z)G_{j_2j_3}(z)\cdots G_{j_nk}(z)|\prec\bigg(\frac{\im m(z)}{N\eta} \bigg)^{n/2}\,,\qquad\qquad (z\in\C^+)\,.
\end{align}
\end{lemma}
\begin{proof}
Let $\lambda_1^{H_t} \geq \lambda_2^{H_t} \geq \dots \geq \lambda_N^{H_t}$ be the eigenvalues of $H_t$, and let $\bsu_1,\ldots,\bsu_N$, $\bsu_{\alpha}\equiv \bsu_{\alpha}^{H_t}$, denoted the associated normalized eigenvectors. Then, by spectral decomposition, we get
\begin{align*} \begin{split}
\sum_{j=1}^N |G_{ij}|^2 = \sum_{j=1}^N \sum_{\alpha, \beta} \frac{\bsu_{\alpha}(i) \ol{\bsu_{\alpha}(j)}}{\lambda_{\alpha} -z} \frac{\ol{\bsu_{\beta}(i)} \bsu_{\beta}(j)}{\lambda_{\beta} - \ol z} = \sum_{\alpha, \beta} \frac{\bsu_{\alpha}(i) \langle \bsu_{\alpha}, \bsu_{\beta} \rangle \ol{\bsu_{\beta}(i)}}{(\lambda_{\alpha} -z)(\lambda_{\beta} - \ol z)} = \sum_{\alpha=1}^N \frac{|\bsu_{\alpha}(i)|^2}{|\lambda_{\alpha} -z|^2}\,.
\end{split} \end{align*}
Since the eigenvectors are delocalized by Proposition~\ref{cor: delocalization}, we find that
\begin{align*}
\frac{1}{N}\sum_{j=1}^N |G_{ij}|^2 \prec \frac{1}{N^2} \sum_{\alpha=1}^N \frac{1}{|\lambda_{\alpha} -z|^2} = \frac{\im m}{N\eta}\,.
\end{align*}
This proves the first inequality in~\eqref{generalized Ward} for $i=k$. The inequality for $i\not=k$, the second inequality in~\eqref{generalized Ward} and~\eqref{generalized Ward 2} then follow directly from Schwarz inequality.
\end{proof}
Recalling the definition of $\Ir_{r,s}$ in~\eqref{e1} we have
\begin{align*}
\Ir_{2,0}\deq N\kappa_t^{(3)}\frac{1}{N^2} \sum_{i_1 \neq i_2} \big( \partial_{i_1i_2}^{2} G_{i_2i_1} \big) P^{D-1} \ol{P^D} \,.
\end{align*}
We then notice that $\Ir_{2,0}$ contains terms with one or three off-diagonal Green function entries $G_{i_1i_2}$. We split accordingly
\begin{align}\label{le split Ir2}
w_{\Ir_{2,0}}\Ir_{2,0}=w_{\Ir_{2,0}^{(1)}}\Ir_{2,0}^{(1)}+w_{\Ir_{2,0}^{(3)}}\Ir_{2,0}^{(3)}\,,
\end{align}
where $\Ir_{2,0}^{(1)}$ contains all terms with one off-diagonal Green function entries (and, necessarily, two diagonal Green function entries) and where $\Ir_{2,0}^{(3)}$ contains all terms with three off-diagonal Green function entries (and zero diagonal Green function entries), and $w_{\Ir_{2,0}^{(1)}}$, $w_{\Ir_{2,0}^{(3)}} $ denote the respective weights. Explicitly,
\begin{align}\begin{split}\label{definition of the IR21}
\E\Ir_{2,0}^{(1)}&= {N\kappa_t^{(3)}}\E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} G_{i_2i_1} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D} \bigg]\,,\\
\E\Ir_{2,0}^{(3)}&= {N\kappa_t^{(3)}}\E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} (G_{i_2i_1})^3 P^{D-1} \ol{P^D} \bigg]\,,
\end{split}\end{align}
and $w_{\Ir_{2,0}}=1$, $w_{\Ir_{2,0}^{(1)}}=3$, $w_{\Ir_{2,0}^{(3)}}=1$.
We first note that $\Ir_{2,0}^{(3)}$ satisfies, for $N$ sufficiently large,
\begin{align} \label{e1123}
|\E\Ir_{2,0}^{(3)}|& \leq\frac{N^\epsilon s^{(3)}}{q_t } \E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} |G_{i_1i_2}|^2 |P|^{2D-1} \bigg] \leq \frac{N^{\epsilon}}{ q_t} \E \bigg[ \frac{\im m}{N\eta} |P|^{2D-1} \bigg] \leq \Phiepsi\,.
\end{align}
\begin{remark}\label{remark power counting}[Power counting I] Consider the terms $I_{r,0}$, $r\ge 1$. For $n\ge 1$, we then split
\begin{align}\label{smino}
w_{\Ir_{2n}}\Ir_{2n,0}=\sum_{l=0}^n w_{\Ir_{2n,0}^{(2l+1)}}\Ir_{2n,0}^{(2l+1)}\,,\qquad
w_{\Ir_{2n-1}}\Ir_{2n-1,0}=\sum_{l=0}^n w_{\Ir_{2n-1,0}^{(2l)}}\Ir_{2n-1,0}^{(2l)}\,,
\end{align}
according to the parity of $r$. For example, for $r=1$, $\Ir_{1,0}=\Ir_{1,0}^{(0)}+\Ir_{1,0}^{(2)}$, with
\begin{align*}
\E\Ir_{1,0}^{(0)}= -\E \bigg[\frac{1}{N^2} \sum_{i \neq k} G_{kk} G_{ii} P^{D-1} \ol{P^D} \bigg]\,,\qquad \E\Ir_{1,0}^{(2)}=-\E \bigg[\frac{1}{N^2} \sum_{i \neq k} G_{ik}G_{ki} P^{D-1} \ol{P^D} \bigg]\,;
\end{align*}
\cf~\eqref{ea111}. Now, using a simple power counting, we bound the summands in~\eqref{smino} as follows. First, we note that each term in $\Ir_{r,0}$ contains a factor of $q_t^{(r-2)_+}$. Second, for $\E\Ir_{2n,0}^{(2l+1)}$ and $\E\Ir_{2n-1,0}^{(2l)}$, with $n\ge 1$, $l\ge 1$, we can by Lemma~\ref{lem:2 off-diagonal} extract one factor of $\frac{\im m}{N\eta}$ (other Green function entries are bounded using $|G_{ik}|\prec 1$). Thus, for $n\ge 1$, $l\ge 1$,
\begin{align}\label{trivially estimated higher order terms}
|\E\Ir_{2n,0}^{(2l+1)}|\le\frac{ N^{\epsilon}}{q_t^{2n-2}}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)|P|^{2D-1}\bigg]\,,\qquad\quad |\E\Ir_{2n-1,0}^{(2l)}|\le \frac{N^{\epsilon}}{q_t^{(2n-3)_+}}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)|P|^{2D-1}\bigg]\,,
\end{align}
for $N$ sufficiently large, and we conclude that all these terms are negligible.
\end{remark}
We next consider $\E\Ir_{2,0}^{(1)}$ that is not covered by~\eqref{trivially estimated higher order terms}. Using $|G_{ii}|\prec 1$ and Lemma~\ref{lem:2 off-diagonal} we~get
\begin{align} \begin{split}\label{a factor of q to be gained}
|\E\Ir_{2,0}^{(1)}|\leq \frac{N^{\epsilon}s^{(3)}}{ q_t} \E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} |G_{i_1i_2}| |P|^{2D-1} \bigg]\le \frac{N^\epsilon}{ q_t}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2}|P|^{2D-1}\bigg]\,,
\end{split} \end{align}
for $N$ sufficiently large. Yet, this bound is not negligible. We need to gain an additional factor of $q_t^{-1}$ with which it will become negligible. We have the following result.
\begin{lemma}\label{le first round lemma q}
For any (small) $\epsilon>0$, we have, for all $z\in\mathcal{D}$,
\begin{align}\label{a factor q gained from a round}
|\E\Ir_{2,0}^{(1)}|\le \frac{N^\epsilon}{q_t^2}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2}|P|^{2D-1}\bigg]+\Phiepsi\le N^\epsilon\E \bigg[ \bigg(q_t^{-4}+ \frac{\im m}{N\eta} \bigg) \big| P(m) \big|^{2D-1} \bigg] +\Phiepsi\,,
\end{align}
for $N$ sufficiently large. In particular, the term $\E\Ir_{2,0}$ is negligible.
\end{lemma}
\begin{proof}
Fix a (small) $\epsilon>0$. Recalling~\eqref{definition of the IR21}, we have
\begin{align}\label{preumathched begins}
\E\Ir_{2,0}^{(1)}={N\kappa_t^{(3)}} \E \bigg[\frac{1}{N^2} \sum_{i_1\not=i_2} G_{i_2i_1}G_{i_1i_1}G_{i_2i_2} P^{D-1} \ol{P^D} \bigg] \,.
\end{align}
The key feature here is that the Green function entries are $ G_{i_2i_1}G_{i_1i_1}G_{i_2i_2}$, where at least one index, say $i_2$, appears an odd number of times. (This index $i_2$ can be considered as ``unmatched''.) Using the resolvent formula~\eqref{expansion indentity} we expand in the unmatched index $i_2$ to get
\begin{align} \label{unmatched begins}\begin{split}
z \E\Ir_{2,0}^{(1)}&={N\kappa_t^{(3)}}\E \bigg[\frac{1}{N^2}\sum_{i_1\not=i_2 \not=i_3} H_{i_2i_3} G_{i_3i_1} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D}\bigg]\,.
\end{split}\end{align}
We now proceed in a similar way as in Remark~\ref{remark on entrywise local law of GOE} where we estimated $|G^W_{i_1i_2}|$, $i_1\not=i_2$, for the GOE. Applying the cumulant expansion to the right side of~\eqref{unmatched begins}, we will show that the leading term is $-\E[ m\Ir_{2,0}^{(1)}]$. Then, upon substituting $m(z)$ by the deterministic quantity~$\wt m(z)$ and showing that all other terms in the cumulant expansion of the right side of~\eqref{unmatched begins} are negligible, we will get that
\begin{align}\label{for second round needed}
|z+\wt m(z)|\,\big| \E\Ir_{2,0}^{(1)}\big|\le \frac{N^\epsilon}{q_t^2}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2}|P|^{2D-1}\bigg]+\Phiepsi\le 2\Phiepsi\,,
\end{align}
for $N$ sufficiently large. Since $|z+\wt m(z) |\ge 1/6$ uniformly on $\mathcal{E}\supset\mathcal{D}$, as shown in Remark~\ref{rem:stability of wt m}, the lemma directly follows. The main efforts in the proof go into showing that the sub-leading terms in the cumulant expansion of the right side of~\eqref{unmatched begins} are indeed negligible.
For simplicity we abbreviate $\wh I\equiv \Ir_{2,0}^{(1)}$. Then using Lemma~\ref{le stein lemma} and Remark~\ref{remark on other expansions}, we have, for arbitrary $\ell'\in\N$, the cumulant expansion
\begin{align}\label{le second round cumulant expansion}
z \E\Ir_{2,0}^{(1)}=z \E\wh \Ir=\sum_{r'=1}^{\ell'}\sum_{s'=0}^{r'} w_{\wh\Ir_{r',s'}}\E \wh\Ir_{r',s'}+O\left(\frac{N^\epsilon}{q_t^{\ell'}}\right)\,,
\end{align}
with
\begin{align}\label{def wtirs}
{\wh\Ir_{r',s'}\deq {N\kappa_t^{(r'+1)}} {N\kappa_t^{(3)}} \frac{1}{N^3} \sum_{i_1\not=i_2\not=i_3} \big( \partial_{i_2i_3}^{r'-s'}( G_{i_3i_1}G_{i_2i_2} G_{i_1i_1})\big) \big( \partial_{i_2i_3}^{s'}\big( P^{D-1} \ol{P^D} \big) \big)}
\end{align}
and $w_{\wh\Ir_{r',s'}}=\frac{1}{r'!}\binom{r'}{s'}$. Here, we used Corollary~\ref{le corollary for error terms omega} to truncate the series in~\eqref{le second round cumulant expansion} at order $\ell'$. Choosing $\ell'\ge 8D$ the remainder is indeed negligible.
We first focus on $\wh\Ir_{r',0}$. For $r'=1$, we compute
\begin{align}\begin{split}\label{weissenstein}
\E\wh\Ir_{1,0}&=-\frac{\nm^{(3)}}{q_t} \E \bigg[\frac{1}{N^3} \sum_{i_1\not=i_2\not=i_3}G_{i_2i_1}G_{i_3i_3} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D}\bigg]\\ &\qquad- 3\frac{\nm^{(3)}}{q_t} \E \bigg[\frac{1}{N^3} \sum_{i_1\not=i_2\not=i_3} G_{i_2i_3}G_{i_3i_1} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D}\bigg]\\
&\qquad-2\frac{\nm^{(3)}}{q_t} \E \bigg[\frac{1}{N^3} \sum_{i_1\not=i_2\not=i_3} G_{i_1i_2}G_{i_2i_3}G_{i_3i_1}G_{i_2i_2} P^{D-1} \ol{P^D}\bigg]\\
&=:\E\wh\Ir_{1,0}^{(1)}+3\E\wh\Ir_{1,0}^{(2)}+2\E\wh\Ir_{1,0}^{(3)}\,,
\end{split}\end{align}
where we organize the terms according to the off-diagonal Green functions entries. By Lemma~\ref{lem:2 off-diagonal},
\begin{align}\label{weissenstein 2}
|\E\wh\Ir_{1,0}^{(2)}|\le \frac{N^{\epsilon}}{q_t}\E\bigg[\frac{\im m}{N\eta}|P|^{2D-1}\bigg]\le \Phiepsi\,,\qquad |\E\wh\Ir_{1,0}^{(3)}|\le \frac{N^{\epsilon}}{q_t}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{3/2}|P|^{2D-1}\bigg]\le \Phiepsi\,.
\end{align}
Recall $\wt m \equiv \wt m_t(z)$ defined in Proposition \ref{prop:local}. We rewrite $\wh\Ir_{1,0}^{(1)}$ with $\wt m$ as
\begin{align} \begin{split} \label{u1}
z\E\wh\Ir_{1,0}^{(1)}&=- \E \bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2} \wt m G_{i_2i_1} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D} \bigg] \\&\qquad\qquad- \E \bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2} (m-\wt m) G_{i_2i_1} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D} \bigg]+O(\Phiepsi)\,.
\end{split} \end{align}
By Schwarz inequality and the high probability bounds $|G_{kk}|, |G_{ii}| \leq N^{\epsilon/8}$, for $N$ sufficiently large, the second term in \eqref{u1} is bounded as
\begin{align} \begin{split}\label{u1bis}
& \bigg|\E \bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2} (m-\wt m) G_{i_2i_1} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D} \bigg]\bigg| \leq N^{\epsilon/4} \E \bigg[ \frac{1}{N^2} \sum_{i_1\not= i_2} |m-\wt m| |G_{i_1i_2}| |P|^{2D-1} \bigg] \\
&\qquad\leq N^{-\epsilon/4} \E \bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2} |m-\wt m|^2 |P|^{2D-1} \bigg] + N^{3\epsilon/4} \E \bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2} |G_{i_1i_2}|^2 |P|^{2D-1} \bigg] \\
&\qquad= N^{-\epsilon/4} \E \bigg[ |m-\wt m|^2 |P|^{2D-1} \bigg] + N^{3\epsilon/4} \E \bigg[ \frac{\im m}{N\eta} |P|^{2D-1} \bigg]\,.
\end{split} \end{align}
We thus get from~\eqref{weissenstein},~\eqref{weissenstein 2},~\eqref{u1} and~\eqref{u1bis} that
\begin{align}\begin{split}\label{hahama}
z\E\wh \Ir_{1,0}=-\wt m \E \bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2}G_{i_2i_1} G_{i_2i_2} G_{i_1i_1} P^{D-1} \ol{P^D} \bigg] +O(\Phiepsi)= -\E\wt m\Ir_{2,0}^{(1)}+O(\Phiepsi)\,,
\end{split}\end{align}
where we used~\eqref{preumathched begins}.
We remark that in the expansion of $\E\wh \Ir=\E \Ir_{2,0}^{(1)}$ the only term with one off-diagonal entry is $\E\wh\Ir_{2,0}^{(1)}$. All the other terms contain at least two off-diagonal entries.
\begin{remark}\label{remark power counting II}[Power counting II]
Comparing~\eqref{the IR} and~\eqref{le second round cumulant expansion}, we have $\wh\Ir_{r',s'}=(\Ir_{2,0}^{(1)})_{r',s'}$. Consider now the terms with $s'=0$. As in~\eqref{smino} we organize the terms according to the number of off-diagonal Green function entries. For $r'\ge 2$,
\begin{align}\label{smino2}
w_{\wh\Ir_{r',0}}\wh\Ir_{r',0}&=\sum_{l=0}^n w_{\wh\Ir_{r',0}^{(l+1)}}\wh\Ir_{r',0}^{(l+1)}=\sum_{l=0}^n w_{\wh\Ir_{r',0}^{(l+1)}} (\Ir_{2,0}^{(1)})_{r',0}^{(l+1)} \,.
\end{align}
A simple power counting as in Remark~\ref{remark power counting} then directly yields
\begin{align}\label{trivially estimated higher order terms 2}
|\E\wh\Ir_{r',0}^{(1)}|\le\frac{ N^{\epsilon}}{q_t^{r'}}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2}|P|^{2D-1}\bigg]\,,\qquad|\E\wh\Ir_{r',0}^{(l+1)}|\le\frac{ N^{\epsilon}}{q_t^{r'}}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)|P|^{2D-1}\bigg]\,, \qquad(l\ge 1)\,,
\end{align}
for $N$ sufficiently large. Here, we used that each term contains a factor $\kappa_t^{(3)}\kappa_t^{(r'+1)}\le CN^{-2}q_t^{-r'}$. We conclude that all terms in~\eqref{trivially estimated higher order terms 2} with $r'\ge 2$ are negligible, yet we remark that $|\E\wh\Ir_{2,0}^{(1)}|$ is the leading error term in $|\E \Ir_{2,0}^{(1)}|$, which is explicitly listed on the right side of~\eqref{a factor q gained from a round}.
\end{remark}
\begin{remark}\label{remark power counting III}[Power counting III]
Consider the terms $\wh\Ir_{r',s'}$, with $1\le s'\le r'$. For $s'=1$, note that $\partial_{i_2i_3} \big(P^{D-1} \ol{P^D}\big)$ contains two off-diagonal Green function entries. Explicitly,
\begin{align*}\begin{split}
\wh\Ir_{r',1}= & -2(D-1)\frac{N\kappa_t^{(r'+1)}N\kappa_t^{(3)}} {{N^3}} \sum_{i_1\not=i_2\not=i_3} \big(\partial_{i_2i_3}^{r'-1}(G_{i_3i_1}G_{i_2i_2}G_{i_1i_1}) \big)\bigg({\frac{1}{N}}\sum_{i_4=1}^N G_{i_4i_2}G_{i_3i_4}\bigg)P'P^{D-2}\ol{P^D} \\ &-2D\frac{N\kappa_t^{(r'+1)}N\kappa_t^{(3)}} {{N^3}} \sum_{i_1\not=i_2\not=i_3} \big(\partial_{i_2i_3}^{r'-1}(G_{i_3i_1}G_{i_2i_2}G_{i_1i_1}) \big)\bigg({\frac{1}{N}}\sum_{i_4=1}^N \ol{G_{i_4i_2}G_{i_3i_4}}\bigg)\ol{P'} P^{D-1} \ol{P^{D-1}} \,,
\end{split}\end{align*}
where the fresh summation index $i_4$ is generated from $\partial_{i_2i_3} P$. Using Lemma~\ref{lem:2 off-diagonal} we get, for $r'\ge 1$,
\begin{align}\begin{split}\label{siminibo}
|\E\wh\Ir_{r',1}|&\le \frac{N^\epsilon}{q_t^{r'}}\E\bigg[\bigg(\frac{\im m}{N\eta} \bigg)^{3/2}|P'||P|^{2D-2}+\bigg(\frac{\im m}{N\eta} \bigg)^{3/2}|P'||P|^{2D-2} \bigg]\le 2 \Phiepsi\,,
\end{split}\end{align}
for $N$ sufficiently large, where we used that $\partial_{i_2i_3}^{r'-1}(G_{i_3i_1}G_{i_2i_2}G_{i_1i_1})$, $r'\ge 1$, contains at least one off-diagonal Green function entry.
For $2\le s'\le r'$, we first note that, for $N$ sufficiently large,
\begin{align}\begin{split}\label{monday morning}
|\E\wh\Ir_{r',s'}|&\le\frac{N^\epsilon}{q_t^{r'}}\bigg|\E \bigg[\frac{1}{N^3} \sum_{i_1\not=i_2\not=i_3} \big( \partial_{i_2i_3}^{r'-s'} G_{i_3i_1}G_{i_2i_2}G_{i_1i_1} \big) \big( \partial_{i_2i_3}^{s'}\big( P^{D-1} \ol{P^D} \big) \big)\bigg]\bigg|\\
&\le\frac{N^\epsilon}{q_t^{r'}}\E \bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2}\frac{1}{N^2} \sum_{i_2\not=i_3} \big| \partial_{i_2i_3}^{s'}\big( P^{D-1} \ol{P^D} \big) \big|\bigg]\,.
\end{split}\end{align}
Next, since $s'\ge 2$, the partial derivative in $\partial_{i_2i_3}^{s'} \big( P^{D-1} \ol{P^D}\big)$ acts on $P$ and $\ol{P}$ (and on their derivatives) more than once. For example, for $s'=2$,
\begin{align}\begin{split}\nonumber
\partial_{i_1i_2}^2 P^{D-1} &=\frac{4(D-1)(D-2)}{N} \bigg(\sum_{i_3=1}^N G_{i_3i_2}G_{i_1i_3} \bigg)^2(P')^2 P^{D-3}\\ &\qquad + \frac{2(D-1)}{N} \bigg( \sum_{i_3=1}^N G_{i_2i_3}G_{i_3i_1} \bigg)^2 P'' P^{D-2}-\frac{2(D-1)}{N} \sum_{i_3=1}^N \partial_{i_1i_2} \big( G_{i_3i_2}G_{i_1i_3} \big) P' P^{D-2}\,,
\end{split}\end{align}
where $\partial_{i_1i_2}$ acted twice on $P$, respectively $P'$, to produce the first two terms. More generally, for $s'\ge 2$, consider a resulting term containing
\begin{align}\label{fresh indeces term}
P^{D-s_1'} \ol{P^{D-s_2'}} \left( P' \right)^{s_3'} \left( \ol{P'} \right)^{s_4'} \left( P'' \right)^{s_5'}\left( \ol{P''} \right)^{s_6'}\left( P''' \right)^{s_7'}\left( \ol{P'''} \right)^{s_8'}\,,
\end{align}
with $1\le s_1' \le D$, $0\le s_2' \le D$ and $\sum_{n=1}^8s_n'\le s'$. Since $P^{(4)}$ is constant we did not list it. We see that such a term above was generated from $P^{D-1}\overline{P^D}$ by letting the partial derivative $\partial_{i_2i_3}$ act $s_1'-1$-times on~$P$ and~$s_2'$-times on~$\ol P$, which implies that $ s_1'-1\ge s_3'$ and $s_2'\ge s_4'$. If $ s_1'-1>s_3'$, then $\partial_{i_2i_3}$ acted on the derivatives of $P, \ol{P}$ directly $(s_1'-1-s_3')$-times, and a similar argument holds for~$\ol{P'}$. Whenever $\partial_{i_2i_3}$ acted on $P$, $\ol P$ or their derivatives, it generated a term $ 2N^{-1} \sum_{i_l} G_{i_2i_l}G_{i_li_3}$, with $i_l$, $l\ge 3$, a fresh summation index. For each fresh summation index we apply Lemma~\ref{lem:2 off-diagonal} to gain a factor $\frac{\im m}{N\eta}$. The total number of fresh summation indices in a term corresponding to~\eqref{fresh indeces term} is
\begin{align*}
s_1'- + s_2' +(s_1 '- 1- s_3') + (s_2' - s_4')& = 2s_1' + 2s_2' - s_3' - s_4'= 2\tilde s_0-\tilde s-2\,,
\end{align*}
with $\tilde s_0\deq s_1'+s_2'$ and $\tilde s\deq s_3+s_4$ we note this number does not decrease when $\partial_{i_2i_3}$ acts on off-diagonal Green functions entries later. Thus, from~\eqref{monday morning} we conclude, upon using $|G_{i_1i_2}|,|P''(m)|, |P'''(m)|$, $|P^{(4)}(m)| \prec 1$ that, for $2\le s'\le r'$,
\begin{align}\label{monday morning 2}\begin{split}
|\E\wh\Ir_{r',s'}|&\le\frac{N^\epsilon}{q_t^{r'}}\E \bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2}\frac{1}{N^2} \sum_{i_2\not=i_3} \big| \partial_{i_2i_3}^{s'}\big( P^{D-1} \ol{P^D} \big) \big|\bigg]\\
&\le\frac{N^{2\epsilon}}{q_t^{r'}}\sum_{\tilde s_0=2}^{2D}\sum_{\tilde s=1}^{\tilde s_0-2}\E\ \bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2+2\tilde s_0-\tilde s-2}|P' |^{\tilde s }|P|^{2D-\tilde s_0}\bigg]\\
&\qquad\qquad+\frac{N^{2\epsilon}}{q_t^{r'}}\sum_{\tilde s_0=2}^{2D}\E\ \bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2+2\tilde s_0-1}|P' |^{\tilde s_0-1 }|P|^{2D-\tilde s_0}\bigg]\,,
\end{split}\end{align}
for $N$ sufficiently large. Here the last term on the right corresponds to $\wt s=\wt s_0-1$. Thus, we conclude form~\eqref{monday morning 2} and the definition of $\Phiepsi$ in~\eqref{eq:phi} that $\E[\wh\Ir_{r',s'}]$, $2\le s'\le r'$, is negligible.
To sum up, we have established that all terms $\E[\wt\Ir_{r',s'}]$ with $1\le s'\le r'$ are negligible.
\end{remark}
From~\eqref{preumathched begins},~\eqref{le second round cumulant expansion},~\eqref{hahama},~\eqref{trivially estimated higher order terms 2},~\eqref{siminibo} and~\eqref{monday morning 2} we find that
\begin{align*} \begin{split}
&|z+\wt m|\,| \E \Ir_{2,0}^{(1)}|\le\frac{N^\epsilon}{q_t^2}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{1/2}|P|^{2D-1}\bigg]+\Phiepsi\,,
\end{split} \end{align*}
for $N$ sufficiently large. Since $(z+ \wt m)$ is deterministic and $|z+\wt m| > 1/6$, as we showed in Remark~\ref{rem:stability of wt m}, we obtain that $|\E\Ir_{2,0}^{(1)}|\le \Phiepsi$. This concludes the proof of~\eqref{a factor q gained from a round}.
\end{proof}
Summarizing, we showed in~\eqref{e1123} and~\eqref{a factor q gained from a round} that
\begin{align}\label{final bound Ir_20}
|\E\Ir_{2,0}|\le \Phiepsi\,,
\end{align}
for $N$ sufficiently large, \ie all terms in $\E\Ir_{2,0}$ are negligible and the second estimate in~\eqref{le summary equation 1} is proved.
\subsection{Estimate on $\Ir_{r,0}$, $r \geq 4$} \label{sub:e114}
For $r \geq 5$ we use the bounds $|G_{i_1i_1}|,|G_{i_1i_2}|\prec 1$ to get
\begin{align} \begin{split}\label{e115}
|\E\Ir_{r,0}|&=\bigg| N\kappa_t^{(r+1)} \E \bigg[ \frac{1}{N^2}\sum_{i_1 \neq i_2} \big( \partial_{i_1i_2}^r G_{i_2i_1} \big) P^{D-1} \ol{P^D} \bigg] \bigg|\\&\ \leq \frac{N^{\epsilon}}{q_t^4}\E \bigg[ \frac{1}{N^2}\sum_{i_1 \neq i_2} |P|^{2D-1} \bigg] \leq \frac{N^{\epsilon}}{q_t^4} \E [|P|^{2D-1}] \leq \Phiepsi\,,
\end{split}\end{align}
for $N$ sufficiently large. For $r=4$, $\partial_{i_1i_2}^r G_{i_2i_1}$ contains at least one off-diagonal term $G_{i_1i_2}$. Thus
\begin{align} \begin{split} \label{e114}
\bigg| N\kappa_t^{(5)} \E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} \big( \partial_{i_1i_2}^4 G_{ki} \big) P^{D-1} \ol{P^D} \bigg] \bigg| &\leq \frac{N^{\epsilon}}{ q_t^{3}} \E \bigg[ \frac{1}{N^2}\sum_{i_1 \neq i_2} |G_{ik}| |P|^{2D-1} \bigg] \\
&\le \frac{N^\epsilon}{ q_t^{3}}\E\bigg[ \bigg(\frac{\im m}{N\eta}\bigg)^{1/2} |P|^{2D-1} \bigg]\le \Phiepsi\,,
\end{split} \end{align}
for $N$ sufficiently large, where we used Lemma~\ref{lem:2 off-diagonal} to get the last line. We conclude that all terms $\E\Ir_{r,0}$ with $r\geq 4$ are negligible. This proves the fourth estimate in~\eqref{le summary equation 1}.
\subsection{Estimate on $\Ir_{r,s}$, $r\ge 2$, $s\ge 1$} \label{sub:e12}
For $r\ge 2$ and $s=1$, we have
\begin{align*}
\E\Ir_{r,1}= N\kappa_t^{(r+1)}\E\bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2}(\partial_{i_1i_2}^{r-1}G_{i_2i_1}) \partial_{i_1i_2} ( P^{D-1} \ol{P^D} ) \bigg] \,.
\end{align*}
Note that each term in $\E\Ir_{r,1}$, $r\ge 2$, contains at least two off-diagonal Green function entries. For the terms with at least three off-diagonal Green function entries, we use the bound $| G_{i_1i_2}|, |G_{i_1i_1}| \prec 1$ and
\begin{align} \begin{split} \label{e12r1o}
&N\kappa_t^{(r+1)}\E \bigg[ \frac{1}{N^3} \sum_{i_1, i_2,i_3}|G_{i_2i_1}G_{i_1i_3} G_{i_3i_2}| |P'| |P|^{2D-2} \bigg]
\leq N^{\epsilon}\frac{ s^{(r+1)}}{q_t}\E \bigg[ \bigg( \frac{\im m}{N\eta} \bigg)^{3/2} |P'| |P|^{2D-2} \bigg] \\
&\qquad\qquad \leq N^{\epsilon}s^{(r+1)} \E \bigg[ \sqrt{\im m} \bigg( \frac{\im m}{N\eta} \bigg) \bigg( \frac{1}{N\eta} + q_t^{-2} \bigg) |P'| |P|^{2D-2} \bigg]\,,
\end{split} \end{align}
for $N$ sufficiently large, where we used Lemma~\ref{lem:2 off-diagonal}. Note that the right side is negligible since $\im m \prec 1$.
Denoting the terms with two off-diagonal Green function entries in $\E\Ir_{r,1}$ by $\E\Ir_{r,1}^{(2)}$, we have
\begin{align}\label{le Ir 112}\begin{split}
\E\Ir_{r,1}^{(2)}&=N\kappa^{(r+1)}\E\bigg[\frac{2(D-1)}{N^2} \sum_{i_1\neq i_2} G_{i_2i_2}^{r/2} G_{i_1i_1}^{r/2} \Big(\frac{1}{N}\sum_{i_3=1}^N G_{i_2i_3}G_{i_3i_1} \Big)P' P^{D-2} \ol{P^D}\bigg]\\&\qquad +N\kappa^{(r+1)}\E\bigg[\frac{2D}{N^2} \sum_{i_1\neq i_2} G_{i_2i_2}^{r/2} G_{i_1i_1}^{r/2}\Big(\frac{1}{N} \sum_{i_3=1}^N \ol{ G_{i_2i_3}G_{i_3i_1}}\Big) \ol{P'} P^{D-1} \ol{P^{D-1}} \bigg]\,,
\end{split}\end{align}
where $i_3$ is a fresh summation index and where we noted that $r$ is necessarily even in this case.
Lemma~\ref{lem:2 off-diagonal} then give us the upper bound
$$
\big|\E\Ir_{r,1}^{(2)}\big|\le \frac{N^{\epsilon}}{ q_t^{r-1}} \E \bigg[ \frac{\im m}{N\eta} |P'| |P|^{2D-2} \bigg]\,,
$$
which is negligible for $r> 2$. However, for $r=2$, we need to gain an additional factor~$q_t^{-1}$. This can be done as in the proof of Lemma~\ref{le first round lemma q} by considering the off-diagonal entries $G_{i_2i_3}G_{i_3i_1} $, generated from $\partial_{i_1i_2} P(m)$, since the index $i_2$ appears an odd number of times.
\begin{lemma}
For any (small) $\epsilon>0$, we have
\begin{align}\label{the sunday lemmal}
| \E\Ir_{2,1}^{(2)}|\le \frac{N^{\epsilon}}{ q_t^{2}} \E \bigg[ \frac{\im m}{N\eta} |P'| |P|^{2D-2} \bigg]+\Phiepsi\,,
\end{align}
uniformly on $\caD$, for $N$ sufficiently large. In particular, the term $\E\Ir_{2,1}$ is negligible.
\end{lemma}
\begin{proof}
We start with the first term on the right side of~\eqref{le Ir 112}. Using~\eqref{expansion indentity}, we write
\begin{align*}\begin{split}
z N\kappa_t^{(3)}&\E\bigg[\frac{1}{N^3} \sum_{i_1\neq i_2\neq i_3} G_{i_2i_3}G_{i_2i_2}G_{i_1i_1} G_{i_1i_3}P' P^{D-2} \ol{P^D}\bigg]\\ &= N\kappa_t^{(3)}\E\bigg[\frac{1}{N^3} \sum_{i_1\neq i_2\neq i_3\neq i_4} H_{i_2i_4}G_{i_4i_3}G_{i_2i_2}G_{i_1i_1} G_{i_1i_3}P' P^{D-2} \ol{P^D}\bigg]\,.
\end{split}\end{align*}
As in the proof of Lemma~\ref{le first round lemma q}, we now apply the cumulant expansion to the right side. The leading terms of the expansion is
\begin{align}\label{le leading for second q round}
N\kappa_t^{(3)}&\E\bigg[\frac{1}{N^3} \sum_{i_1\neq i_2\neq i_3}m G_{i_2i_3}G_{i_2i_2}G_{i_1i_1} G_{i_1i_3}P' P^{D-2} \ol{P^D}\bigg]\,,
\end{align}
and, thanks to the additional factor of $q_t^{-1}$ from the cumulant $\kappa_t^{(3)}$, all other terms in the cumulant expansion are negligible, as can be checked by power counting as in the proof of Lemma~\ref{le first round lemma q}. Replacing in~\eqref{le leading for second q round}~$m$ by $\widetilde m$, we then get
\begin{align*}
|z+\wt m|\, |N\kappa_t^{(3)}|&\bigg|\E\bigg[\frac{1}{N^3} \sum_{i_1\neq i_2\neq i_3} G_{i_2i_3}G_{i_2i_2}G_{i_1i_1} G_{i_1i_3}P' P^{D-2} \ol{P^D}\bigg]\bigg|\le C\Phiepsi\,,
\end{align*}
for $N$ sufficiently large; \cf~\eqref{for second round needed}. Since $|z+\wt m(z)|\ge 1/6$, $z\in\mathcal{D}$ by Remark~\ref{rem:stability of wt m}, we conclude that the first term on the right side of~\eqref{le Ir 112} is negligible. In the same way one shows that the second term is negligible, too. We leave the details to the reader.
\end{proof}
We hence conclude from~\eqref{e12r1o} and~\eqref{the sunday lemmal} that $\E\Ir_{r,1}$ is negligible for all $r\ge 2$.
Consider next the terms
\begin{align*}
\E\Ir_{r,s}= N\kappa_t^{(r+1)}\E\bigg[ \frac{1}{N^2} \sum_{i_1\not=i_2}(\partial_{i_1i_2}^{r-s}G_{i_2i_1}) \partial_{i_1i_2}^{s} ( P^{D-1} \ol{P^D} ) \bigg] \,,
\end{align*}
with $2\le s\le r$. We proceed in a similar way as in Remark~\ref{remark power counting II}. We note that each term in $\partial_{i_1i_2}^{r-s}G_{i_2i_1}$ contains at least one off-diagonal Green function when $r-s$ is even, yet when $r-s$ is odd there is a term with no off-diagonal Green function entries. Since $s\ge 2$, the partial derivative $\partial_{i_1i_2}^s$ acts on $P$ or~$\ol P$ (or their derivatives) more than once in total; \cf Remark~\ref{remark power counting II}. Consider such a term with
$$
P^{D-s_1} \ol{P^{D-s_2}} \left( P' \right)^{s_3} \left( \ol{P'} \right)^{s_4} \,,
$$
for $ 1\le s_1 \le D$ and $0\le s_2\le D$. Since $P''(m), P'''(m), P^{(4)}(m) \prec 1$ and $P^{(5)}=0$, we do not include derivatives of order two and higher here. We see that such a term was generated from $P^{D-1}\overline{P^D}$ by letting the partial derivative $\partial_{i_1i_2}$ act $(s_1 -1)$-times on~$P$ and $s_2$-times on~$\ol P$, which implies that $s_3 \leq s_1 -1$ and $s_4 \leq s_2$. If $s_3 < s_1 -1$, then $\partial_{i_1i_2}$ acted on $P'$ as well $[(s_1 -1)-s_3]$-times, and a similar argument holds for $\ol{P'}$. Whenever $\partial_{i_1i_2}$ acts on $P$ or~$\ol P$ (or their derivatives), it generates a fresh summation index $i_l$, $l\ge 3$, with a term $ 2N^{-1} \sum_{i_l} G_{i_2i_l}G_{i_li_1}$. The total number of fresh summation indices in this case is
$$
(s_1 -1) + s_2 + [(s_1 -1) - s_3] + [s_2 - s_4] = 2s_1 + 2s_2 - s_3 - s_4 -2\,.
$$
Assume first that $r=s$ so that $ \partial_{i_1i_2}^{r-s}G_{i_2i_1}=G_{i_2i_1}$. Then applying Lemma~\ref{lem:2 off-diagonal} $(2s_1 + 2s_2 - s_3 - s_4 -2)$-times and letting $s_0=s_1+s_2$ and $s'=s_3+s_4$, we obtain an upper bound, $r=s\ge 2$,
\begin{align}\begin{split} \label{e12s-2Kevin}
|\E\Ir_{r,r}|&\le \frac{N^\epsilon}{q_t^{r-1}}\sum_{s_0=2}^{2D}\sum_{s'=1}^{s_0-1}\E\bigg[\bigg(\frac{\im m}{N\eta} \bigg)^{1/2}\bigg(\frac{\im m}{N\eta} \bigg)^{2s_0-s'-2}|P'|^{s'}|P|^{2D-s_0} \bigg] \le \Phiepsi\,,
\end{split}\end{align}
for $N$ sufficiently large, \ie $\E\Ir_{r,r}$, $r\ge 2$, is negligible.
Second, assume that $2\le s<r$. Then applying Lemma~\ref{lem:2 off-diagonal} $(2s_1 + 2s_2 - s_3 - s_4 -2)$-times, we get
\begin{align}\begin{split} \label{e12s-2}
|\E\Ir_{r,s}|&\le \frac{N^\epsilon}{q_t^{r-1}}\sum_{s_0=2}^{2D}\sum_{s'=1}^{s_0-2}\E\bigg[\bigg(\frac{\im m}{N\eta} \bigg)^{2s_0-s'-2}|P'|^{s'}|P|^{2D-s_0} \bigg] \\
&\qquad\qquad +\frac{N^\epsilon}{q_t^{r-1}}\sum_{s_0=2}^{2D}\E\bigg[\bigg(\frac{\im m}{N\eta} \bigg)^{s_0-1}|P'|^{s_0-1}|P|^{2D-s_0} \bigg] \,,
\end{split}\end{align}
for $N$ sufficiently large with $2\le s<r$. In particular, $|\E\Ir_{r,s}|\le \Phiepsi$, $2\le s<r$. In~\eqref{e12s-2} the second term bounds the terms corresponding to $s_0-1=s'$ obtained by acting on $\partial_{i_1i_2}$ exactly $(s_1 -1)$-times on $P$ and $s_2$-times on $\ol P$ but never on their derivatives.
To sum up, we showed that $\E\Ir_{r,s}$ is negligible, for $1\le s<r$. This proves~\eqref{le summary equation 2} for $1\le s<r$.
\subsection{Estimate on $\Ir_{3,0}$} \label{sub:e113}
We first notice that $\Ir_{3,0}$ contains terms with zero, two or four off-diagonal Green function entries and we split accordingly
\begin{align*}
w_{\Ir_{3,0}}\Ir_{3,0}=w_{\Ir_{3,0}^{(0)}}\Ir_{3,0}^{(0)}+w_{\Ir_{3,0}^{(2)}}\Ir_{3,0}^{(2)}+w_{\Ir_{3,0}^{(4)}}\Ir_{3,0}^{(4)}\,.
\end{align*}
When there are two off-diagonal entries, we can use Lemma~\ref{lem:2 off-diagonal} to get the bound
\begin{align} \begin{split} \label{e1132}
|\E\Ir_{3,0}^{(2)}|=\bigg| {N\kappa_t^{(4)}} \E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} G_{i_2i_2} G_{i_1i_1} (G_{i_2i_1})^2 P^{D-1} \ol{P^D} \bigg] \bigg|& \leq \frac{N^{\epsilon}}{ q_t^{2}} \E \bigg[ \frac{\im m}{N\eta} |P|^{2D-1} \bigg]\le \Phiepsi\,,
\end{split} \end{align}
for $N$ sufficiently large. A similar estimate holds for $|\E\Ir_{3,0}^{(2)}|$. The only non-negligible term is $\Ir_{3,0}^{(0)}$.
For $n\in\N$, set
\begin{align}\label{definition of SN}
S_n\equiv S_n(z) \deq \frac{1}{N} \sum_{i=1}^N (G_{ii}(z))^n\,.
\end{align}
By definition $S_1=m$. We remark that $|S_n| \prec 1$, for any fixed $n$, by Proposition~\ref{local semiclrcle law}.
\begin{lemma} \label{le combinatorics lemma} We have
\begin{align} \label{e1130 expansion}
w_{\Ir_{3,0}^{(0)}}\E\Ir_{3,0}^{(0)} =- N\kappa_t^{(4)} \E \Big[S_2^2 P^{D-1} \ol{P^D} \Big] \,.
\end{align}
\end{lemma}
\begin{proof}
Recalling the definition of $\Ir_{r,s}$ in~\eqref{the IR}, we have
\begin{align*}
w_{\Ir_{3,0}}\Ir_{3,0}=\frac{N\kappa_t^{(4)}}{3!} \E \bigg[\frac{1}{N^2} \sum_{i_1 \neq i_2} \big( \partial_{i_1i_2}^3 G_{i_2i_1} \big) P^{D-1} \ol{P^D} \bigg]\,.
\end{align*}
We then easily see that the terms with no off-diagonal entries in $\partial_{i_1i_2}^3 G_{i_2i}$ are of the form
$$
-G_{i_2i_2} G_{i_1i_1} G_{i_2i_2} G_{i_1i_1}\,.
$$
We only need to determine the weight $w_{\Ir_{3,0}^{(0)}}$. With regard to the indices, taking the third derivative corresponds to putting the indices $i_2i_1$ or $i_1i_2$ three times. In that sense, the very first $i_2$ and the very last $i_1$ are from the original $G_{i_2i_1}$. The choice of $i_2i_1$ or $i_1i_2$ must be exact in the sense that the connected indices in the following diagram must have been put at the same time:
$$
i_2\underbrace{i_2 \quad i_1} \underbrace{i_1 \quad i_2} \underbrace{i_2 \quad i_1}i_1\,.
$$
Thus, the only combinatorial factor we have to count is the order of putting the indices. In this case, we have three connected indices, so the number of terms must be $3!=6$. Thus, $w_{\Ir_{3,0}^{(0)}}=1$ and \eqref{e1130 expansion} indeed holds.
\end{proof}
\begin{lemma}\label{le S22 lemma} For any (small) $\epsilon>0$, we have, for all $z\in\mathcal{D}$,
\begin{align}\label{le S22 lemma equation}
N\kappa_t^{(4)}\E \Big[ S_2^2 P^{D-1} \ol{P^D} \Big] =N\kappa_t^{(4)} \E \Big[ m^4 P^{D-1} \ol{P^D} \Big]+ O( \Phiepsi)\,.
\end{align}
\end{lemma}
\begin{proof}
Fix $\epsilon>0$. We first claim that
\begin{align} \label{s21=s22}
N\kappa_t^{(4)}\E \Big[S_2^2 P^{D-1} \ol{P^D} \Big]=N\kappa_t^{(4)} \E \Big[ m^2 S_2 P^{D-1} \ol{P^D} \Big]+O(\Phiepsi)\,.
\end{align}
The idea is to expand the term $\E[zm S_2^2 P^{D-1} \ol{P^D}]$ in two different ways and compare the results. Using the resolvent identity \eqref{expansion indentity} and Lemma~\ref{le stein lemma}, we get
\begin{align} \begin{split} \label{s21}
\E\big[zm S_2^2 P^{D-1} \ol{P^D}\big] &= -\E\big[S_2^2 P^{D-1} \ol{P^D}\big] + \E \bigg[\frac{1}{N} \sum_{i_1\not=i_2} H_{i_1i_2} G_{i_2i_1} S_2^2 P^{D-1} \ol{P^D} \bigg] \\
&= -\E\big[S_2^2 P^{D-1} \ol{P^D}\big] + \sum_{r=1}^{\ell'} \frac{N\kappa_t^{(r+1)}}{r!} \E \bigg[\frac{1}{N^2} \sum_{i_1\not= i_2} \partial_{i_1i_2}^r \Big( G_{i_2i_1} S_2^2 P^{D-1} \ol{P^D} \Big) \bigg]\\ &\qquad\qquad + \E\bigg[\Omega_{\ell'}\bigg(\frac{1}{N}\sum_{i_1\not= i_2} H_{i_1i_2} G_{i_2i_1} S_2^2 P^{D-1} \ol{P^D}\bigg) \bigg]\,,
\end{split} \end{align}
for arbitrary $\ell'\in\N$. Using the resolvent identity~\eqref{expansion indentity} once more, we write
\begin{align*}
z S_2 = \frac{1}{N} \sum_{i=1}^N z G_{ii} G_{ii} = -\frac{1}{N} \sum_{i=1}^N G_{ii} + \frac{1}{N} \sum_{i_1\not= i_2} H_{i_1i_2} G_{i_2i_1} G_{i_1i_1}\,.
\end{align*}
Thus, using Lemma~\ref{le stein lemma}, we also have
\begin{align} \label{s22}
\E[zm S_2^2 P^{D-1} \ol{P^D}] &= -\E[m^2 S_2 P^{D-1} \ol{P^D}] +\E \bigg[ \frac{1}{N} \sum_{i_1 \not=i_2} H_{i_1i_2} G_{i_2i_1} G_{i_1i_1} m S_2 P^{D-1} \ol{P^D} \bigg]\nonumber \\
&= -\E[m^2 S_2 P^{D-1} \ol{P^D}] + \sum_{r=1}^{\ell'} \frac{N\kappa_t^{(r+1)}}{r!} \E \bigg[ \frac{1}{N^2}\sum_{i_1\not=i_2} \partial_{i_1i_2}^r \Big( G_{i_2i_1} G_{i_1i_1} m S_2 P^{D-1} \ol{P^D} \Big) \bigg]\nonumber\\
&\qquad\qquad+ \E\bigg[\Omega_{\ell'}\bigg(\frac{1}{N}\sum_{i_1\not=i_2} H_{i_1i_2} G_{i_2i_1} G_{i_1i_1} m S_2 P^{D-1} \ol{P^D}\bigg) \bigg]\,,
\end{align}
for arbitrary $\ell'\in\N$. By Corollary~\ref{le corollary for error terms omega} and Remark~\ref{remark on other expansions}, the two error terms $\E[\Omega_{\ell'}(\cdot)]$ in~\eqref{s21} and~\eqref{s22} are negligible for $\ell'\ge 8D$.
With the extra factor $N\kappa_t^{(4)}$, we then write
\begin{align}\label{le with the extra factor}\begin{split}
N\kappa_t^{(4)}\E[zm S_2^2 P^{D-1} \ol{P^D}] &= -N\kappa_t^{(4)}\E[S_2^2 P^{D-1} \ol{P^D}] + \sum_{r=1}^{\ell'}\sum_{s=0}^r w_{\scriptsize{\ttilde \Ir_{r,s}}}\E\ttilde \Ir_{r,s}+O(\Phiepsi)\,,\\
N\kappa_t^{(4)} \E[zm S_2^2 P^{D-1} \ol{P^D}] &=- N\kappa_t^{(4)}\E[m^2 S_2 P^{D-1} \ol{P^D}] + \sum_{r=1}^{\ell'}\sum_{s=0}^r w_{\scriptsize{\ttilde[2] \Ir_{r,s}}}\E{\ttilde[2]{ \Ir}}_{r,s}+O(\Phiepsi)\,,
\end{split}\end{align}
with
\begin{align}\begin{split}
\ttilde \Ir_{r,s}&\deq N\kappa_t^{(4)} N\kappa_t^{(r+1)}\frac{1}{N^2}\sum_{i_1\not=i_2}\big(\partial_{i_1i_2}^{r-s}\big( G_{i_2i_1}S_2^2\big)\big)\big(\partial_{i_1i_2}^s \big(P^{D-1}\ol{P^D}\big)\big)\,,\\
{\ttilde[2]{ \Ir}}_{r,s}&\deq N\kappa_t^{(4)} N\kappa_t^{(r+1)}\frac{1}{N^2}\sum_{i_1\not=i_2}\big(\partial_{i_1i_2}^{r-s}\big(G_{i_2i_1}G_{i_1i_1} mS_2\big)\big)\big(\partial_{i_1i_2}^s \big(P^{D-1}\ol{P^D}\big)\big)
\end{split}\end{align}
and $ w_{\scriptsize{\ttilde \Ir_{r,s}}}= w_{\scriptsize{\ttilde[2] \Ir_{r,s}}}=(r!)^{-1}((r-s)!)^{-1}$.
For $r=1$, $s=0$, we find that
\begin{align}\label{le they agree 1}\begin{split}
\E\ttilde \Ir_{1,0}&= -N\kappa_t^{(4)} \E\bigg[\frac{1}{N^2}\sum_{i_1\not=i_2}G_{i_1i_1}G_{i_2i_2}S_2^2 P^{D-1}\ol{P^D}\bigg]+O(\Phiepsi)\\
&=-N\kappa_t^{(4)} \E\Big[m^2S_2^2 P^{D-1}\ol{P^D}\Big]+O(\Phiepsi)\,,
\end{split}\end{align}
and similarly
\begin{align}\label{le they agree 2}\begin{split}
\E\ttilde[2]\Ir_{1,0}&= -N\kappa_t^{(4)} \E\bigg[\frac{1}{N^2}\sum_{i_1\not=i_2} G_{i_1i_1}^2G_{i_2i_2}mS_2P^{D-1}\ol{P^D}\bigg]+O(\Phiepsi)\\
&= -N\kappa_t^{(4)} \E\bigg[m^2S_2^2P^{D-1}\ol{P^D}\bigg]+O(\Phiepsi)\,,
\end{split}\end{align}
where we used~\eqref{definition of SN}. We hence conclude that $\E\ttilde\Ir_{1,0}$ equals $\E\ttilde[2]\Ir_{1,0}$ up to negligible error.
Following the ideas in Subsection~\ref{sub:e112}, we can bound
\begin{align*}
\big|\E\ttilde \Ir_{1,1}\big|&\le\frac{N^\epsilon}{q_t^{2}}\bigg|\E\bigg[\frac{1}{N^2}\sum_{i_1\not=i_2}G_{i_1i_2}S_2^2(\partial_{i_1i_2} P^{D-1}\ol{P^D})\bigg]\bigg|\le \frac{N^\epsilon}{q_t^{2}}\E\bigg[\bigg(\frac{\im m}{N\eta}\bigg)^{3/2}|P'||P|^{2D-2}\bigg]\le \Phiepsi\,,
\end{align*}
and similarly $|\E\ttilde[2] \Ir_{1,1}|\le \Phiepsi$, for $N$ sufficiently large. In fact, for $r\ge 2$, $s\ge 0$ we can use, with small notational modifications the power counting outlined in Remark~\ref{remark power counting II} and Remark~\ref{remark power counting III} to conclude that
\begin{align*}
\E\ttilde\Ir_{r,s}=O(\Phiepsi)\,,\qquad\qquad \E\ttilde[2]\Ir_{r,s}=O( \Phiepsi)\,, \qquad\qquad (r\ge 2\,,s\ge 0)\,.
\end{align*}
Therefore the only non-negligible terms on the right hand side of~\eqref{le with the extra factor} are $N\kappa_t^{(4)}\E[S_2^2 P^{D-1} \ol{P^D}] $, $N\kappa_t^{(4)}\E[m^2 S_2 P^{D-1} \ol{P^D}]$ as well as $\ttilde\Ir_{1,0}$ and $\ttilde[2]\Ir_{1,0}$. Since, by~\eqref{le they agree 1} and~\eqref{le they agree 2}, the latter agree up to negligible error terms, we conclude that the former two must be equal up do negligible error terms. Thus~\eqref{s21=s22} holds.
Next, expanding the term $\E[zm^3 S_2 P^{D-1} \ol{P^D}]$ in two different ways similar to above, we further~get
\begin{align} \label{s22=s23}
N\kappa_t^{(4)}\E \Big[m^2 S_2 P^{D-1} \ol{P^D} \Big]= N\kappa_t^{(4)} \E \Big[ m^4 P^{D-1} \ol{P^D} \Big]+O(\Phiepsi)\,.
\end{align}
Together with~\eqref{s21=s22} this shows~\eqref{le S22 lemma equation} and concludes the proof of the lemma.
\end{proof}
Finally, from~Lemma~\ref{le combinatorics lemma} and Lemma~\ref{le S22 lemma}, we conclude that
\begin{align}\label{le final Ir30}
w_{\Ir_{3,0}}\E[\Ir_{3,0}]=-\frac{\nm_t^{(4)}}{q_t^2} \E \Big[ m^4 P^{D-1} \ol{P^D} \Big]+O(\Phiepsi)\,.
\end{align}
This proves the third estimate in~\eqref{le summary equation 1}.
\begin{proof}[Proof of Lemma~\ref{summary expansions}]
The estimates in~\eqref{le summary equation 1} were obtained in~\eqref{e111},~\eqref{final bound Ir_20},~\eqref{e115},~\eqref{e114} and~\eqref{le final Ir30}. Estimate~\eqref{le summary equation 2} follows from~\eqref{e1211},~\eqref{e12r1o},~\eqref{e12s-2Kevin} and~\eqref{e12s-2}.
\end{proof}
\section{Tracy--Widom limit: Proof of Theorem~\ref{thm tw}} \label{sec:tw}
In this section, we prove the Tracy--Widom limit of the extremal eigenvalues, Theorem~\ref{thm tw}. As we explained in Section~\ref{Tracy-Widom limit and Green function comparison}, the distribution of the largest eigenvalue of $H$ can be obtained by considering the imaginary part of the normalized trace of the Green function $m$ of $H$. For $\eta > 0$, we introduce
\begin{align}\label{le poisson kernel}
\theta_{\eta}(y) = \frac{\eta}{\pi(y^2 + \eta^2)}\,,\qquad\qquad (y\in\R)\,.
\end{align}
Using the functional calculus and the definition of the Green function, we have
\begin{align}
\im m(E+\ii\eta)=\frac{\pi}{ N} \Tr \theta_{\eta}(H-E)\,,\qquad\qquad (z=E+\ii\eta\in\C^+)\,.
\end{align}
We have the following proposition, which corresponds to Corollary~6.2 in \cite{EYY} or Lemma~6.5 of~\cite{EKYY2}.
\begin{proposition} \label{prop:cutoff}
Suppose that $H$ satisfies Assumption~\ref{assumption H} with $\phi > 1/6$. Denote by $\lambda_1^H$ the largest eigenvalue of $H$.
Fix $\epsilon > 0$. Let $E \in \R$ such that $|E-L| \leq N^{-2/3 + \epsilon}$. Let $E_+ \deq L + {2}N^{-2/3 + \epsilon}$ and define $\chi_E \deq \lone_{[E, E_+]}$. Set $\eta_1 \deq N^{-2/3 - 3\epsilon}$ and $\eta_2 \deq N^{-2/3 - 9\epsilon}$. Let $K: \R \to [0, \infty)$ be a smooth function satisfying
\begin{align}
K(x) =
\begin{cases}
1 & \text{ if } |x| < 1/3 \\
0 & \text{ if } |x| > 2/3
\end{cases},
\end{align}
which is a monotone decreasing on $[0, \infty)$. Then, for any~$D > 0$,
\begin{align}
\E \left[ K \left( \Tr (\chi_E \ast \theta_{\eta_2})(H) \right) \right] > \p (\lambda_1^H \leq E-\eta_1) - N^{-D}
\end{align}
and
\begin{align}
\E \left[ K \left( \Tr (\chi_E \ast \theta_{\eta_2})(H) \right) \right] < \p (\lambda_1^H \leq E+\eta_1) + N^{-D}\,,
\end{align}
for $N$ sufficiently large, with $\theta_{\eta_2}$ as in~\eqref{le poisson kernel}.
\end{proposition}
We prove Proposition~\ref{prop:cutoff} in Section~\ref{preliminaries of edge universality}. We move on to the Green function comparison theorem. Let $W^{\GOE}$ be a GOE matrix independent of $H$ with vanishing diagonal entries as introduced in Subsection~\ref{Tracy-Widom limit and Green function comparison} and denote by $m^{\GOE}\equiv m^{W^{\GOE}}$ the normalized trace of its Green function.
\begin{proposition}[Green function comparison] \label{prop:green}
Let $\epsilon>0$ and set $\eta_0 = N^{-2/3 - \epsilon}$. Let $E_1, E_2\in\R$ satisfy $|E_1|, |E_2| \leq N^{-2/3 + \epsilon}$. Let $F : \R \to \R$ be a smooth function satisfying
\begin{align} \label{F bound}
\max_{x\in\R} |F^{(l)}(x)| (|x|+1)^{-C} \leq C\,,\qquad \qquad (l\in\llbracket 1,11\rrbracket)\,.
\end{align}
Then, for any sufficiently small $\epsilon > 0$, there exists $\delta>0$ such that
\begin{align} \label{green_comp}
\bigg| \E F \bigg( N \int_{E_1}^{E_2} \im m(x + L + \ii \eta_0) \,\dd x \bigg) - \E F \bigg( N \int_{E_1}^{E_2} \im m^{\GOE}(x + 2 + \ii \eta_0)\, \dd x \bigg) \bigg| \leq N^{-\delta}\,,
\end{align}
for sufficiently large $N$.
\end{proposition}
Proposition~\ref{prop:green} is proved in Section~\ref{subsection green function comparison}. We are ready to prove Theorem~\ref{thm tw}.
\begin{proof}[Proof of Theorem~\ref{thm tw}]
Fix $\epsilon > 0$ and set $\eta_1 \deq N^{-2/3 - 3\epsilon}$ and $\eta_2 \deq N^{-2/3 - 9\epsilon}$. Consider $E = L + sN^{-2/3}$ with $s \in (-N^{-2/3+\epsilon}, N^{-2/3+\epsilon})$. For any $D > 0$, we find from Proposition~\ref{prop:cutoff} that
\begin{align*}
\p (\lambda_1^H \leq E) < \E K \left( \Tr (\chi_{E+\eta_1} \ast \theta_{\eta_2})(H) \right) + N^{-D}.
\end{align*}
Applying Proposition~\ref{prop:green} with $9\epsilon$ instead of $\epsilon$ and setting $E_1 = E-L +\eta_1$, $E_2 = E_+ -L$, we find that
\begin{align} \begin{split}\nonumber
&\E K \bigg( \Tr (\chi_{E+\eta_1} \ast \theta_{\eta_2})(H) \bigg) = \E K \bigg( \frac{N}{\pi} \int_{E_1}^{E_2} \im m(x + L + \ii \eta_2) \,\dd x \bigg) \\
&\leq \E K \bigg( \frac{N}{\pi} \int_{E_1}^{E_2} \im m^{\GOE}(x + 2 + \ii \eta_2)\, \dd x \bigg) + N^{-\delta} = \E K \bigg( \Tr (\chi_{[E_1+2, E_2+2]} \ast \theta_{\eta_2})(W^{\GOE}) \bigg) + N^{-\delta}
\end{split} \end{align}
for some $\delta > 0$. Hence, applying Proposition~\ref{prop:cutoff} again to the matrix $W^{\GOE}$, we get
\begin{align} \label{eq:TW upper bound}
&\p \left( N^{2/3} (\lambda_1^H -L) \leq s \right) = \p \left(\lambda_1^H \leq E \right) \\
&< \p \left(\lambda_1^{\GOE} \leq E_1 + 2 + \eta_1 \right) + N^{-D} + N^{-\delta} = \p \left( N^{2/3} (\lambda_1^{\GOE} -2) \leq s + 2N^{-3\epsilon} \right) + N^{-D} + N^{-\delta}\,.\nonumber
\end{align}
Similarly, we can also check that
\begin{align} \label{eq:TW lower bound}
\p \left( N^{2/3} (\lambda_1^H -L) \leq s \right) > \p \left( N^{2/3} (\lambda_1^{\GOE} -2) \leq s - 2N^{-3\epsilon} \right) - N^{-D} - N^{-\delta}.
\end{align}
Since the right sides of Equations~\eqref{eq:TW upper bound} and \eqref{eq:TW lower bound} converge both in probability to $F_1(s)$, the $\GOE$ Tracy--Widom distribution, as $N$ tends to infinity we conclude that
\begin{align}
\lim_{N \to \infty} \p \Big( N^{2/3} (\lambda^H_1 -L)\leq s \Big) = F_1 (s)\,.
\end{align}
This proves Theorem~\ref{thm tw}.
\end{proof}
In the rest of this section, we prove Propositions~\ref{prop:cutoff} and \ref{prop:green}
\subsection{Proof of Proposition~\ref{prop:cutoff}}\label{preliminaries of edge universality}
For a given $\epsilon > 0$, we chose $E \in \R$ such that $|E-L| \leq N^{-2/3 + \epsilon}$, $E_+ \deq L + {2}N^{-2/3 + \epsilon}$, $\eta_1 = N^{-2/3 - 3\epsilon}$, and $\eta_2 = N^{-2/3 - 9\epsilon}$. In principle, we could adopt the strategy in the proof of Corollary~6.2 of~\cite{EYY} after proving the optimal rigidity estimate at the edge with the assumption $q \gg N^{1/6}$ and checking that such an optimal bound is required only for the eigenvalues at the edge. However, we introduce a slightly different approach that directly compares $\Tr (\chi_E \ast \theta_{\eta_2})$ and $\Tr \chi_{E-\eta_1}$ by using the local law.
\begin{proof}[Proof of Proposition~\ref{prop:cutoff}]
For an interval $I \subset \R$, let $\caN_I$ be the number of the eigenvalues in $I$, \ie
\begin{align}
\caN_I \deq |\{ \lambda_i : \lambda_i \in I \}|\,.
\end{align}
We compare $(\chi_E \ast \theta_{\eta_2})(\lambda_i)$ and $\chi_{E-\eta_1}(\lambda_i)$ by considering the following cases:
{\it Case 1:} If $x \in [E + \eta_1, E_+ - \eta_1)$, then $\chi_{E-\eta_1}(x) = 1$ and
\begin{align} \begin{split}
(\chi_E \ast \theta_{\eta_2})(x) - 1 &= \frac{1}{\pi} \int_E^{E_+} \frac{\eta_2}{(x-y)^2 + \eta_2^2} \,\dd y -1 = \frac{1}{\pi} \left( \tan^{-1} \frac{E_+ - x}{\eta_2} - \tan^{-1} \frac{E-x}{\eta_2} \right) - 1 \\
&= -\frac{1}{\pi} \left( \tan^{-1} \frac{\eta_2}{E_+ - x} + \tan^{-1} \frac{\eta_2}{x-E} \right) = O \left( \frac{\eta_2}{\eta_1} \right) = O(N^{-6\epsilon})\,.
\end{split} \end{align}
For any $E' \in [E + \eta_1, E_+ - \eta_1)$, with the local law, Proposition~\ref{prop:local}, we can easily see that
\begin{align}\label{the five}
\frac{\caN_{[E'-\eta_1, E'+\eta_1)}}{5N \eta_1} \leq \frac{1}{N} \sum_{i=1}^N \frac{\eta_1}{|\lambda_i -E'|^2 + \eta_1^2} = \im m(E' + \ii \eta_1) \leq \frac{N^{\epsilon/2}}{N\eta_1} + \frac{N^{\epsilon/2}}{q^2}\,,
\end{align}
with high probability, where we used $\im \wt m(E'+\ii\eta_1)\le C\sqrt{\kappa(E')+\eta_1}\ll N^{\epsilon/2}/(N\eta_1)$. Thus, considering at most $[E_+ - (L-N^{-2/3+\epsilon})]/\eta_1 = 3 N^{4\epsilon}$ intervals, we find that
\begin{align}
\caN_{[E + \eta_1, E_+ - \eta_1)} \leq CN^{9\epsilon /2}
\end{align}
and
\begin{align} \label{eq:large x}
\sum_{i: E + \eta_1 < \lambda_i < E_+ - \eta_1} \big( (\chi_E \ast \theta_{\eta_2})(\lambda_i) - \chi_{E-\eta_1} (\lambda_i) \big) \leq N^{-\epsilon}\,,
\end{align}
with high probability.
{\it Case 2:} For $x < E-\eta_1$, choose $k\ge 0$ such that $3^k \eta_1 \leq E-x < 3^{k+1} \eta_1$. Then, $\chi_{E-\eta_1}(x)=0$ and
\begin{align} \begin{split}
(\chi_E \ast \theta_{\eta_2})(x) &= \frac{1}{\pi} \left( \tan^{-1} \frac{E_+ - x}{\eta_2} - \tan^{-1} \frac{E-x}{\eta_2} \right) = \frac{1}{\pi} \left( \tan^{-1} \frac{\eta_2}{E-x} - \tan^{-1} \frac{\eta_2}{E_+ -x} \right) \\
&< \frac{1}{2} \left( \frac{\eta_2}{E-x} -\frac{\eta_2}{E_+ -x} \right) = \frac{1}{2} \frac{\eta_2(E_+ - E)}{(E-x)(E_+ -x)} < 2 N^{-4/3 -8\epsilon} \cdot 3^{-2k} \eta_1^{-2}.
\end{split} \end{align}
Abbreviate $\caN_k = \caN_{(E- 3^{k+1} \eta_1, E - 3^k \eta_1]}$. Consider
\begin{align}
\im m(E- 2 \cdot 3^k \eta_1 + \ii \cdot 3^k \eta_1) = \frac{1}{N} \sum_{i=1}^N \frac{3^k \eta_1}{| \lambda_i - (E- 2 \cdot 3^k \eta_1)|^2 + (3^k \eta_1)^2} > \frac{1}{N} \frac{\caN_k}{2 \cdot 3^k \eta_1}.
\end{align}
With the local law, Proposition~\ref{prop:local}, and the estimate $\im \wt m(x+ \ii y) \sim \sqrt{|x-L| + y}$, we find that
\begin{align}
\im m(E- 2 \cdot 3^k \eta_1 + \ii \cdot 3^k \eta_1) \leq C \sqrt{3^k \eta_1} + \frac{N^{\epsilon/2}}{N \cdot 3^k \eta_1} \leq N^{5\epsilon} \sqrt{3^k \eta_1}
\end{align}
and hence
\begin{align}
\frac{1}{N} \frac{\caN_k}{2 \cdot 3^k \eta_1} < N^{5\epsilon} \sqrt{3^k \eta_1}\,,
\end{align}
with high probability. Thus, with high probability,
\begin{align} \begin{split} \label{eq:small x}
\sum_{i: \lambda_i < E - \eta_1} \big( (\chi_E \ast \theta_{\eta_2})(\lambda_i) - \chi_{E- \eta_1} (\lambda_i) \big) &\leq 2 \sum_{k=0}^{2 \log N} N^{-4/3 -8\epsilon} \cdot 3^{-2k} \eta_1^{-2} \caN_k \\
&\leq 4 N^{-1/3 -3\epsilon} \eta_1^{-1/2} \sum_{k=0}^{\infty} 3^{-k/2} \leq 10 N^{-3\epsilon /2}\,.
\end{split}\end{align}
{\it Case 3:} By Proposition \ref{prop:norm bound} there are with high probability no eigenvalues in $[E_+ - \eta_1, \infty)$.
{\it Case 4:} For $x \in [E-\eta_1, E+\eta_1)$, we use the trivial estimate
\begin{align}\label{le trivial estimate}
(\chi_E \ast \theta_{\eta_2})(x) < 1 = \chi_{E-\eta_1}(x)\,.
\end{align}
Considering the above cases, we find that
\begin{align} \label{chi comparison}
\Tr (\chi_E \ast \theta_{\eta_2})(H) \leq \Tr \chi_{E-\eta_1}(H) + N^{-\epsilon}\,,
\end{align}
with high probability. From the definition of the cutoff $K$ and the fact that $\Tr \chi_{E-\eta_1}(H)$ is an integer,
\begin{align*}
K( \Tr \chi_{E-\eta_1}(H) + N^{-\epsilon} ) = K( \mathcal{N}_{[E-\eta_1,E_+]} )\,.
\end{align*}
Thus, since $K$ is monotone decreasing on $[0, \infty)$, \eqref{chi comparison} implies that
\begin{align*}
K ( \Tr (\chi_E \ast \theta_{\eta_2})(H) ) \geq K ( \Tr \chi_{E-\eta_1}(H) )
\end{align*}
with high probability. After taking expectation, we get
\begin{align*}
\E \left[ K \left( \Tr (\chi_E \ast \theta_{\eta_2})(H) \right) \right] > \p (\lambda_1 \leq E-\eta_1) - N^{-D}\,,
\end{align*}
for any $D>0$. This proves the first part of Proposition~\ref{prop:cutoff}. The second part can also be proved in a similar manner by showing that
\begin{align}
\Tr (\chi_E \ast \theta_{\eta_2})(H) \geq \Tr \chi_{E+\eta_1}(H) - N^{-\epsilon}\,,
\end{align}
applying the cutoff $K$ and taking expectation. In this argument~\eqref{le trivial estimate} gets replaced by
\begin{align}
\chi_{E+\eta_1}(x) = 0 < (\chi_E \ast \theta_{\eta_2})(x)\,,
\end{align}
for $x \in [E-\eta_1, E+\eta_1)$. This proves Proposition~\ref{prop:cutoff}.
\end{proof}
\subsection{Green function comparison: Proof of Proposition~\ref{prop:green}}\label{subsection green function comparison}
We first introduce a lemma that has the analogue role of Lemma~\ref{le stein lemma} in calculations involving $H \equiv H_t$ and $\dot H \equiv \dd H_t / \dd t$.
\begin{lemma} \label{stein interpolation}
Fix $\ell\in \N$ and let $F\in C^{\ell+1}(\R;\C^+)$. Let $Y \equiv Y_0$ be a random variable with finite moments to order $\ell+2$ and let $W$ be a Gaussian random variable independent of $Y$. Assume that $\E[Y] = \E[W] = 0$ and $\E[Y^2] = \E[W^2]$. Define
\begin{align}
Y_t\deq \e{-t/2} Y_0 + \sqrt{1-\e{-t}} W\,,
\end{align}
and let $\dot Y_t\equiv\frac{\dd Y_t}{\dd t}$. Then,
\begin{align}\label{le stein cor}
\E \left[\dot Y_t F(Y_t) \right] = -\frac{1}{2} \sum_{r=2}^\ell \frac{\kappa^{(r+1)}(Y_0)}{r!} \e{-\frac{(r+1)t}{2}} \E \big[ F^{(r)}(Y_t) \big]+\E\big[\Omega_\ell(\dot Y_t F(Y_t))\big]\,,
\end{align}
where $\E$ denotes the expectation with respect to $Y$ and $W$, $\kappa^{(r+1)}(Y)$ denotes the $(r+1)$-th cumulant of $Y$ and~$F^{(r)}$ denotes the $r$-th derivative of the function $F$. The error term~$\Omega_\ell$ in~\eqref{le stein cor} satisfies
\begin{align} \begin{split}
&\big|\E\big[\Omega_\ell(\dot Y F(Y_t))\big]\big| \le C_\ell \E[ |Y_t|^{\ell+2}]\sup_{|x|\le Q}|F^{(\ell+1)}(x)|+ C_\ell \E [|Y_t|^{\ell+2} \lone(|Y_t|>Q)]\sup_{x\in \R} |F^{(\ell+1)}(x)| \,,
\end{split} \end{align}
where $Q\ge 0$ is an arbitrary fixed cutoff and $C_\ell$ satisfies $C_\ell\le \frac{(C\ell)^\ell}{\ell!}$ for some numerical constant~$C$.
\end{lemma}
\begin{proof}
We follow the proof of Corollary 3.1 in~\cite{LP}. First, note that
\begin{align}
\dot Y_t = -\frac{\e{-t/2}}{2} Y + \frac{\e{-t}}{2\sqrt{1-\e{-t}}} W\,.
\end{align}
Thus,
\begin{align*} \begin{split}
&\E \left[ \dot Y_tF(Y_t) \right]= -\frac{\e{-t/2}}{2} \E \left[ Y F \big(\e{-t/2} Y + \sqrt{1-\e{-t}} W \big) \right] + \frac{\e{-t}}{2\sqrt{1-\e{-t}}} \E \left[ W F \big(\e{-t/2} Y + \sqrt{1-\e{-t}} W \big) \right].
\end{split} \end{align*}
Applying Lemma~\ref{le stein lemma} and~\eqref{le SL}, we get~\eqref{le stein cor} since the first two moments of $W$ and $Y$~agree.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:green}]
Fix a (small) $\epsilon>0$. Consider $x \in [E_1, E_2]$. Recall the definition of $H_t$ in~\eqref{eq:A(t)}. For simplicity, let
\begin{align}
G \equiv G_t(x + L_t + \ii \eta_0)\,,\qquad \qquad m \equiv m_t(x + L_t + \ii \eta_0)\,,
\end{align}
with $\eta_0=N^{-2/3-\epsilon}$,
and define
\begin{align}
X\equiv X_t \deq N \int_{E_1}^{E_2} \im m(x + L_t + \ii \eta_0) \,\dd x\,.
\end{align}
Note that $X \prec N^{\epsilon}$ and $|F^{(l)}(X)| \prec N^{C\epsilon}$ for $l \in\llbracket 1,11\rrbracket$. Recall from~\eqref{le L dot} that
\begin{align}
L = 2 + \e{-t} \nm^{(4)} q_t^{-2} + O(\e{-2t}q_t^{-4})\,, \qquad \dot L = -2 \e{-t} \nm^{(4)} q_t^{-2} + O(\e{-2t}q_t^{-4})\,,
\end{align}
where $q_t=\e{t/2}q_0$. Let $z = x + L_t + \ii \eta_0$ and $G \equiv G(z)$. Differentiating $F(X)$ with respect to $t$, we get
\begin{align} \begin{split} \label{time derivative}
\frac{\dd}{\dd t} \E F(X) &= \E \bigg[ F'(X) \frac{\dd X}{\dd t} \bigg] = \E \bigg[ F'(X) \im \int_{E_1}^{E_2} \sum_{i=1}^N \frac{\dd G_{ii}}{\dd t}\, \dd x \bigg] \\
&= \E \bigg[ F'(X) \im \int_{E_1}^{E_2} \bigg( \sum_{i, j, k} \dot H_{jk} \frac{\partial G_{ii}}{\partial H_{jk}} + \dot L \sum_{i, j} G_{ij} G_{ji} \bigg) \dd x \bigg]\,,
\end{split} \end{align}
where by definition
\begin{align}
\dot H_{jk}\equiv \dot{(H_t)}_{jk} = -\frac{1}{2} \e{-t/2} (H_0)_{jk} + \frac{\e{-t}}{2\sqrt{1-\e{-t}}} W_{jk}^{\GOE}\,.
\end{align}
Thus, from Lemma~\ref{stein interpolation}, we find that
\begin{align} \begin{split} \label{green claim'}
&\sum_{i, j, k} \E \left[ \dot H_{jk} F'(X) \frac{\partial G_{ii}}{\partial H_{jk}} \right] = -2 \sum_{i, j, k} \E \left[ \dot H_{jk} F'(X) G_{ij} G_{ki} \right] \\
&= \frac{\e{-t}}{N} \sum_{r=2}^{\ell} \frac{q_t^{-(r-1)} \nm^{(r+1)}}{r!} \sum_i \sum_{j \neq k} \E \left[ \partial_{jk}^r \left( F'(X) G_{ij} G_{ki} \right) \right] + O(N^{1/3 + C\epsilon})
\end{split} \end{align}
for $\ell = 10$, where we abbreviate $\partial_{jk} = \partial/(\partial H_{jk})$. Note that the $O(N^{1/3 + C\epsilon})$ error term in~\eqref{green claim'} is originated from~$\Omega_\ell$ in~\eqref{le stein cor}, which is $O(N^{C\epsilon} N^2 q_t^{-10})$ with $Y = H_{jk}$.
We claim the following lemma.
\begin{lemma} \label{lem:green claim}
For an integer $r \geq 2$, let
\begin{align}
J_r \deq \frac{\e{-t}}{N} \frac{q_t^{-(r-1)} \nm^{(r+1)}}{r!} \sum_{i=1}^N \sum_{j \neq k} \E \left[ \partial_{jk}^r \left( F'(X) G_{ij} G_{ki} \right) \right]\,.
\end{align}
Then,
\begin{align}
J_3 = 2\e{-t} \nm^{(4)} q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{ji} \right] + O(N^{2/3 -\epsilon'})
\end{align}
and, for any $r \neq 3$,
\begin{align}
J_r = O(N^{2/3 -\epsilon'})\,.
\end{align}
\end{lemma}
Assuming that Lemma~\ref{lem:green claim} holds, we obtain that there exists $\epsilon' > 2\epsilon$ such that, for all $t\in[0,6\log N]$,
\begin{align} \label{green claim}
\sum_{i, j, k} \E \left[ \dot H_{jk} F'(X) \frac{\partial G_{ii}}{\partial H_{jk}} \right] = - \dot L \sum_{i, j} \E \left[ G_{ij} G_{ji} F'(X) \right] + O(N^{2/3 -\epsilon'})\,,
\end{align}
which implies that the right side of \eqref{time derivative} is $O(N^{-\epsilon'/2})$. Integrating from $t=0$ to $t=6\log N$, we get
\begin{align*} \begin{split}
&\bigg| \E F \bigg( N \int_{E_1}^{E_2} \im m(x + L + \ii \eta_0) \,\dd x \bigg)_{t=0} - \E F \bigg( N \int_{E_1}^{E_2} \im m(x + L + \ii \eta_0) \,\dd x \bigg)_{t=6\log N} \bigg| \le N^{-\epsilon'/4}\,.
\end{split} \end{align*}
Comparing $\im m|_{t=6\log N}$ and $\im m^{\GOE}$ is trivial; if we let $\lambda_i(6\log N)$ be the $i$-th largest eigenvalue of $H_{6\log N}$ and $\lambda_i^{\GOE}$ the $i$-th largest eigenvalue of $W^{\GOE}$, then $|\lambda_i(6\log N) - \lambda_i^{\GOE}| \prec N^{-3}$, hence
\begin{align}
\left| \im m|_{t=6\log N} - \im m^{\GOE} \right| \prec N^{-5/3}\,.
\end{align}
This proves Proposition~\ref{prop:green}.
\end{proof}
It remains to prove Lemma~\ref{lem:green claim}. The proof uses quite similar ideas as in Section~\ref{sec:stein}, we thus sometimes omit some details and refer to the corresponding paragraph in Section~\ref{sec:stein}.
\begin{proof}[Proof of Lemma~\ref{lem:green claim}]
First, we note that in the definition of $J_r$ we may freely include or exclude the cases $i=j$, $i=k$, or $j=k$ in the summation $\sum_{i, j, k}$, since it contains at least one off-diagonal Green function entry, $G_{ij}$ or $G_{ki}$, and the sizes of such terms are at most of order
$$
N^2 N^{-1} q_t^{-1} q_t^{-1} \ll N^{2/3 - \epsilon'} \e{-t}\,,
$$
for any sufficiently small $\epsilon' > 0$. There may not be any off-diagonal Green function entries when $i=j=k$, but then there is only one summation index, hence we can neglect this case as well.
For the case $r \geq 5$, it is easy to see that $J_r = O(N^{2/3 -\epsilon'})$, since it contains at least two off-diagonal entries in $\partial_{jk}^r \left( F'(X) G_{ij} G_{ki} \right)$ and $|J_r|$ is bounded by
$$
N^3 N^{-1} q_t^{-4} N^{-2/3 +2\epsilon} \ll N^{2/3 - \epsilon'} \e{-2t}\,,
$$
which can be checked using Lemma~\ref{lem:2 off-diagonal} and a simple power counting.
Therefore, we only need to consider the cases $r=2, 3, 4$. In the following subsections, we check each case and complete the proof of Lemma~\ref{lem:green claim}. See Equations~\eqref{g2}, \eqref{g3} and \eqref{g4} below.
\end{proof}
\subsubsection{Proof of Lemma~\ref{lem:green claim} for $r=2$}
We proceed as in Lemma~\ref {le first round lemma q} of Section \ref{sub:e112} and apply the idea of an unmatched index. Observe that
\begin{align} \label{green2}
\partial_{jk}^2 \left( F'(X) G_{ij} G_{ki} \right) = F'(X) \, \partial_{jk}^2 (G_{ij} G_{ki}) + 2 \partial_{jk} F'(X) \, \partial_{jk} (G_{ij} G_{ki}) + (\partial_{jk}^2 F'(X))\,G_{ij} G_{ki} \,.
\end{align}
We first consider the expansion of $\partial_{jk}^2 (G_{ij} G_{ki})$. We can easily estimate the terms with four off-diagonal Green function entries, since, for example,
\begin{align*} \begin{split}
\sum_{i, j, k} \left| \E \left[ F'(X) G_{ij} G_{kj} G_{kj} G_{ki} \right] \right| &\leq N^{C\epsilon} \sum_{i, j, k} |G_{ij} G_{kj} G_{kj} G_{ki}| \leq N^{C\epsilon} \left( \frac{\im m}{N\eta_0} \right)^2 \leq N^{-4/3 +C\epsilon}\,,
\end{split} \end{align*}
where we used Lemma \ref{lem:2 off-diagonal}. Thus, for sufficiently small $\epsilon$ and $\epsilon'$,
\begin{align}\label{le tala a1}
\frac{\e{-t} q_t^{-1}}{N} \sum_{i, j, k} \big|\E \big[ F'(X) G_{ij} G_{kj} G_{kj} G_{ki} \big]\big| \ll N^{2/3 - \epsilon'}\,.
\end{align}
For the terms with three off-diagonal Green function entries, the bound we get from Lemma~\ref{lem:2 off-diagonal} is
$$
q_t^{-1} N^{-1} N^3 N^{C\epsilon} \left( \frac{\im m}{N\eta_0} \right)^{3/2} \sim q_t^{-1} N^{1+C\epsilon}\,,
$$
which is not sufficient. To gain an additional factor of $q_t^{-1}$, which makes the above bound $q_t^{-2} N^{1+C\epsilon} \ll N^{2/3 -\epsilon'}$, we use Lemma~\ref{le stein lemma} to expand in an unmatched index. For example, such a term is of the form
$$
G_{ij} G_{kj} G_{kk} G_{ji}
$$
and we focus on the unmatched index $k$ in $G_{kj}$. Then, multiplying by $z$ and expanding, we get
\begin{align*} \begin{split}
&\frac{q_t^{-1}}{N} \sum_{i, j, k} \E \left[ z F'(X) G_{ij} G_{kj} G_{kk} G_{ji} \right] = \frac{q_t^{-1}}{N} \sum_{i, j, k, n} \E \left[ F'(X) G_{ij} H_{kn} G_{nj} G_{kk} G_{ji} \right] \\
&\qquad\qquad=\frac{q_t^{-1}}{N} \sum_{r'=1}^{\ell} \frac{\kappa^{(r'+1)}}{r'!} \sum_{i, j, k, n} \E \left[ \partial_{kn}^{r'} \left( F'(X) G_{ij} G_{nj} G_{kk} G_{ji} \right) \right] + O(N^{2/3 -\epsilon'})\,,
\end{split} \end{align*}
for $\ell= 10$.
For $r'=1$, we need to consider $\partial_{kn} (F'(X) G_{ij} G_{nj} G_{kk} G_{ji})$. When $\partial_{kn}$ acts on $F'(X)$ it creates a fresh summation index, say $a$, and we get a term
\begin{align*} \begin{split}
&\frac{q_t^{-1}}{N^2} \sum_{i, j, k, n} \E \left[ \left( { \partial_{kn}} F'(X) \right) G_{ij} G_{nj} G_{kk} G_{ji} \right] \\
&\qquad= -\frac{2q_t^{-1}}{N^2} \int_{E_1}^{E_2} \sum_{i, j, k, n, a} \E \big[ G_{ij} G_{nj} G_{kk} G_{ji} F''(X) \im \left( G_{an}(y + L + \ii \eta_0) G_{ka}(y + L + \ii \eta_0) \right) \big] \dd y \\
&\qquad= -\frac{2q_t^{-1}}{N^2} \int_{E_1}^{E_2} \sum_{i, j, k, n, a} \E \big[ G_{ij} G_{nj} G_{kk} G_{ji} F''(X) \im \big( \wt G_{an} \wt G_{ka} \big) \big] \,\dd y\,,
\end{split} \end{align*}
where we abbreviate $\wt G \equiv G(y + L + \ii \eta_0)$. Applying Lemma \ref{lem:2 off-diagonal} to the index $a$ and $\wt G$, we get
\begin{align*}
\sum_{a=1}^N \big| \wt G_{na} \wt G_{ak} \big| \prec N^{-2/3 + 2\epsilon}\,,
\end{align*}
which also shows that
\begin{align} \label{F derivative bound}
|\partial_{kn} F'(X)| \prec N^{-1/3 + C\epsilon}\,.
\end{align}
Applying Lemma~\ref{lem:2 off-diagonal} to the remaining off-diagonal Green function entries, we obtain that
\begin{align} \label{green2 high order}
\frac{1}{q_tN^2} \sum_{i, j, k, n} \big|\E \left[ \left( { \partial_{kn}} F'(X) \right) G_{ij} G_{nj} G_{kk} G_{ji} \right] \big| \leq { q_t^{-1} N^{-2} N^{-1/3 +C\epsilon} N^4 N^{-1 + 3\epsilon} = q_t^{-1} N^{2/3 + C\epsilon}\,.}
\end{align}
If $\partial_{kn}$ acts on $G_{ij} G_{nj} G_{kk} G_{ji}$, then we always get four or more off-diagonal Green function entries with the only exception being
$$
-G_{ij} G_{nn} G_{kj} G_{kk} G_{ji}\,.
$$
To the terms with four or more off-diagonal Green function entries, we apply Lemma \ref{lem:2 off-diagonal} and obtain a bound similar to~\eqref{green2 high order} by power counting. For the term of the exception, we rewrite it as
\begin{align} \begin{split}
&-\frac{q_t^{-1}}{N^2} \sum_{i, j, k, n} \E \left[ F'(X) G_{ij} G_{nn} G_{kj} G_{kk} G_{ji} \right] = -\frac{q_t^{-1}}{N} \sum_{i, j, k} \E \left[ m F'(X) G_{ij} G_{kj} G_{kk} G_{ji} \right] \\
&= -\wt m \frac{q_t^{-1}}{N} \sum_{i, j, k} \E \left[ F'(X) G_{ij} G_{kj} G_{kk} G_{ji} \right] + \frac{q_t^{-1}}{N} \sum_{i, j, k} \E \left[ (\wt m -m) F'(X) G_{ij} G_{kj} G_{kk} G_{ji} \right].
\end{split} \end{align}
Here, the last term is again bounded by $q_t^{-1} N^{2/3 + C\epsilon}$ as we can easily check with Proposition \ref{prop:local} and Lemma \ref{lem:2 off-diagonal}. We thus arrive at
\begin{align} \begin{split}
&\frac{q_t^{-1}}{N}(z + \wt m) \sum_{i, j, k} \E \left[ F'(X) G_{ij} G_{kj} G_{kk} G_{ji} \right] \\
&\qquad\qquad=\frac{q_t^{-1}}{N} \sum_{r'=2}^{\ell} \frac{\kappa^{(r'+1)}}{r'!} \sum_{i, j, k, n} \E \left[ \partial_{kn}^{r'} \left( F'(X) G_{ij} G_{nj} G_{kk} G_{ji} \right) \right] + O(N^{2/3 -\epsilon'})\,.
\end{split} \end{align}
On the right side, the summation is from $r'=2$, hence we have gained a factor $N^{-1} q_t^{-1}$ from $\kappa^{(r'+1)}$ and added a fresh summation index $n$, so the net gain is $q_t^{-1}$. Since $|z + \wt m| \sim 1$, this shows that
\begin{align}
\frac{q_t^{-1}}{N} \sum_{i, j, k} \E \left[ F'(X) G_{ij} G_{kj} G_{kk} G_{ji} \right] = O(N^{2/3 -\epsilon'})\,.
\end{align}
Together with~\eqref{le tala a1}, this takes care of the first term on the right side of~\eqref{green2}.
For the second term on the right side of~\eqref{green2}, we focus on
\begin{align}
\partial_{jk} F'(X) = -\int_{E_1}^{E_2} \sum_{a=1}^N \big[ F''(X) \im \big( \wt G_{ja} \wt G_{ak} \big) \big]\, \dd y
\end{align}
and apply the same argument to the unmatched index $k$ in $\wt G_{ka}$. For the third term, we focus on $G_{ij} G_{ki}$ and again apply the same argument with the index $k$ in $G_{ki}$. We omit the detail.
After estimating all terms accordingly, we eventually get the bound
\begin{align} \label{g2}
\frac{q_t^{-1}}{N} \sum_{i, j, k} \left| \E \left[ \partial_{jk}^2 \left( F'(X) \partial_{jk} G_{ii} \right) \right] \right| = O(N^{2/3 -\epsilon'})\,.
\end{align}
\subsubsection{Proof of Lemma~\ref{lem:green claim} for $r=3$}
We proceed as in Section \ref{sub:e113}. (Note that there will be no unmatched indices for this case.) If $\partial_{jk}$ acts on $F'(X)$ at least once, then that term is bounded by
$$
N^{\epsilon} N^{-1} q_t^{-2} N^3 N^{-1/3 + C\epsilon} N^{-2/3 + 2\epsilon} = q_t^{-2} N^{1+C\epsilon} \ll N^{2/3 - \epsilon'},
$$
where we used \eqref{F derivative bound} and the fact that $G_{ij} G_{ki}$ or $\partial_{jk} (G_{ij} G_{ki})$ contains at least two off-diagonal entries. Moreover, in the expansion of $\partial_{jk}^2 (G_{ij} G_{ki})$, the terms with three or more off-diagonal Green function entries entries can be bounded by
$$
N^{\epsilon} N^{-1} q_t^{-2} N^3 N^{C\epsilon} N^{-1+3\epsilon} = q_t^{-2} N^{1+C\epsilon} \ll N^{2/3 - \epsilon'}\,.
$$
Thus,
\begin{align} \begin{split} \label{g32}
\frac{\e{-t}\nmf q_t^{-2}}{3!N} \sum_{i, j, k} \E \left[ \partial_{jk}^3 \left( F'(X) G_{ij} G_{ki} \right) \right] &= -\frac{4!}{2}\frac{\e{-t} \nmf q_t^{-2}}{3!} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{jj}G_{ji}S_2 \right] \\ &\qquad\qquad+ O(N^{2/3 -\epsilon'})\,,
\end{split} \end{align}
where the combinatorial factor $(4!/2)$ is computed as in Lemma~\ref{le combinatorics lemma} and $S_2$ is as in~\eqref{definition of SN}.
As in Lemma~\ref{le S22 lemma} of Section~\ref{sub:e113}, the first term on right side of~\eqref{g32} is computed by expanding
$$
q_t^{-2} \sum_{i, j} \E \left[ z m S_2 F'(X) G_{ij} G_{jj} G_{ji} \right]
$$
in two different ways, respectively. We can then obtain that
\begin{align} \begin{split} \label{g32'}
q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{jj} G_{ji} S_2 \right] &= q_t^{-2} \sum_{i, j} \E \left[ m^2 F'(X) G_{ij} G_{jj} G_{ji} \right] + O(N^{2/3 -\epsilon'}) \\
&= q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{jj} G_{ji} \right] + O(N^{2/3 -\epsilon'})\,,
\end{split}\end{align}
where we used that $m \equiv m(z) = -1 + O(N^{-1/3 + \epsilon})$ with high probability.
Indeed, since $\widetilde m(L)$, which was denoted by $\tau$ in the proof of Lemma~\ref{lem:w}, satisfies $\widetilde m(L) = -1 + O(\e{-t}q_t^{-2})$ by~\eqref{eq:L_+ estimate}, and since $|\widetilde m(z) -\widetilde m(L)| \sim \sqrt{\kappa+\eta_0} \leq N^{-1/3 + \epsilon}$ by~\eqref{L+ square root}, we have from Proposition~\ref{prop:local} that $m(z) = -1 + O(N^{-1/3 + \epsilon})$ with high probability.
Finally, we consider
\begin{align} \label{g32' expand}
q_t^{-2} \sum_{i, j} \E \left[ z F'(X) G_{ij} G_{jj} G_{ji} \right] = 2 q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{jj} G_{ji} \right] + O(N^{2/3 -\epsilon'})\,.
\end{align}
Expanding the left hand side using~\eqref{expansion indentity}, we also obtain
\begin{align*}
q_t^{-2} \sum_{i, j} \E \left[ z F'(X) G_{ij} G_{jj} G_{ji} \right] = -q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{ji} \right] + q_t^{-2} \sum_{i, j, k} \E \left[ F'(X) H_{jk} G_{ij} G_{kj} G_{ji} \right]\,.
\end{align*}
Applying Lemma~\ref{le stein lemma} to the second term on the right side, we find that most of the terms are $O(N^{2/3 -\epsilon'})$ either due to three (or more) off-diagonal entries, the partial derivative $\partial_{jk}$ acting on $F'(X)$, or higher cumulants. The only term that does not fall into one these categories is
$$
-\frac{q_t^{-2}}{N^2} \sum_{i, j, k} \E \left[ F'(X) G_{ij} G_{kk} G_{jj} G_{ji} \right]\,,
$$
which is generated when $\partial_{jk}$ acts on $G_{kj}$. From this argument, we find that
\begin{align*} \begin{split}
&q_t^{-2} \sum_{i, j} \E \left[ z F'(X) G_{ij} G_{jj} G_{ji} \right] \\
&\qquad= -q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{ji} \right] - q_t^{-2} \sum_{i, j} \E \left[ m F'(X) G_{ij} G_{jj} G_{ji} \right] + O(N^{2/3 -\epsilon'})\,.
\end{split} \end{align*}
Hence, combining it with \eqref{g32' expand} and the fact that $m = -1 + O(N^{-1/3 + \epsilon})$ with high probability, we get
\begin{align*}
q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{jj} G_{ji} \right] = -q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{ji} \right]+ O(N^{2/3 -\epsilon'})\,.
\end{align*}
In combination with~\eqref{g32}~and~\eqref{g32'}, we conclude that
\begin{align} \label{g3}
\frac{\e{-t}}{N} \frac{\nmf q_t^{-2}}{3!} \sum_{i, j, k} \E \left[ \partial_{jk}^3 \left( F'(X) G_{ij} G_{ki} \right) \right] = 2 \e{-t} \nmf q_t^{-2} \sum_{i, j} \E \left[ F'(X) G_{ij} G_{ji} \right] + O(N^{2/3 -\epsilon'})\,.
\end{align}
\subsubsection{Proof of Lemma~\ref{lem:green claim} for $r=4$}
In this case, we estimate the term as in the case $r=2$ and get
\begin{align} \label{g4}
\frac{q_t^{-3}}{N} \sum_{i, j, k} \left| \E \left[ \partial_{jk}^4 \left( F'(X) G_{ij} G_{ki} \right) \right] \right| = O(N^{2/3 -\epsilon'})\,.
\end{align}
We leave the details to the interested reader.
|
\section{Introduction}
Let ${\mathbb F}$ be one of the associative division algebras
over $\mathbb{R}$, namely, the real numbers ${\mathbb R}$, the
complex numbers ${\mathbb C} = \{ a+bi:a,b\in {\mathbb R} \}$, or
the quaternionic numbers ${\mathbb H} = \{a+bi+cj+dk:a,b,c,d\in
{\mathbb R}\}$, with the usual norm and conjugation operations.
For each $n \in {\mathbb N}$, we consider ${\mathbb F}^n$ as a
right inner product space over ${\mathbb F}$ with the usual
multiplication and the standard ${\mathbb F}$-inner product
$\langle|\rangle_{\mathbb F}:{\mathbb F}^n\times {\mathbb F}^n\rightarrow
{\mathbb F}$ given by
$\langle(x_1,\ldots,x_n)|(y_1,\ldots,y_n)\rangle_{\mathbb F}
=\sum_{m=1}^{n}\bar{x}_m y_m$. Let $S({\mathbb F}^n)=\{x\in
{\mathbb F}^n:\langle x|x\rangle_{\mathbb F} = 1\}$ be the standard unit sphere in
${\mathbb F}^n$. An ${\mathbb F}$-vector field on $S({\mathbb
F}^n)$ is a continuous function $v:S({\mathbb F}^n)\rightarrow
{\mathbb F}^n- \{0\}$ such that $\langle x|v(x)\rangle_{\mathbb F}=0$ for each
$x\in S({\mathbb F}^n)$. Given $m$ such ${\mathbb F}$-vector fields
$v_1,\ldots,v_m$ on $S({\mathbb F}^n)$,
we say that they are linearly independent if the vectors $v_1(x),\ldots,v_m(x)$ are linearly independent over ${\mathbb F}$ for all $x\in S({\mathbb F}^n)$.
Let $\rho^{\mathbb F}({\mathbb F}^n)$ denote the maximal number of linearly
independent ${\mathbb F}$-vector fields on $S({\mathbb F}^n)$. The
vector fields on spheres problem has two sides: the first side is
what we call the maximal number problem, namely, the computation
of $\rho^{\mathbb F}({\mathbb F}^n)$, the second side is what we
call the construction problem, namely, the actual construction of
$\rho^{\mathbb F}({\mathbb F}^n)$ linearly independent
${\mathbb F}$-vector fields on $S({\mathbb F}^n)$. The roots of this
classical problem goes back to the hairy ball theorem, the
parallelizability of spheres, and the classification of division
algebras over ${\mathbb R}$. For background materials on this
problem, we refer the reader to \cite{MPS80, Tho69}.
If $n$ is odd, then one can easily show that
$\rho^{\mathbb F}({\mathbb F}^n)=0$. Therefore, unless otherwise indicated,
we will assume that $n$ is an even positive integer.
An explicit formula for computing $\rho^{\mathbb R}({\mathbb R}^n)$ was
given by Adams \cite{Ada62} in 1962, when he showed that
$\rho^{\mathbb R}({\mathbb R}^n) = \rho(n)-1$, where
$\rho(n)$ be the Radon-Hurwitz function.
While there are no explicit formulas for computing
$\rho^{\mathbb C}({\mathbb C}^n)$ and $\rho^{\mathbb H}({\mathbb H}^n)$,
they still can be obtained by using the work of Adams and Walker
\cite{AW65} in 1965 for the complex case, and the work of
Sigrist and Suter \cite{SS73} in 1973 for the quaternionic case.
More direct formulas for computing
$\rho^{\mathbb C}({\mathbb C}^n)$ and $\rho^{\mathbb H}({\mathbb H}^n)$
are given in Section~3 of this paper.
Our main goal in this paper is to draw more attention to the
importance of the construction problem of vector fields on spheres.
There are several known methods that explicitly give
$\rho^{\mathbb R}({\mathbb R}^n)$ linearly independent real vector fields on
$S({\mathbb R}^n)$ (e.g., see \cite{Ogn08, Zve68}).
The situation is completely different with the construction of
complex and quaternionic vector fields; there is no explicit construction
that gives two or more linearly independent complex vector fields on
$S({\mathbb C}^n)$, and there is no known construction that
gives even a single quaternionic vector field on $S({\mathbb H}^n)$.
In addition to their self importance, the actual construction of complex
and quaternionic vector fields on spheres might lead to the solution
of several open problems in the equivairant complex and quaternionic
vector fields on spheres (see \cite{Obi06, Ond01}).
The layout of this paper is as follows.
In Section~2, we present a relationship between real, complex, and
quaternionic standard inner products. We show how such a relationship
leads to a relationship between corresponding real, complex,
and quaternionic vector fields on spheres. More specifically,
we show how $m$ linearly independent quaternionic vector fields can
be used to obtain $2m$ linearly independent complex vector fields,
and how $m$ linearly independent complex vector fields can be used to obtain
$2m$ linearly independent real vector fields.
In Section 3, we provide direct formulas for computing
$\rho^{\mathbb C}({\mathbb C}^n)$ and $\rho^{\mathbb H}({\mathbb H}^n)$, and show that
$\rho^{\mathbb C}({\mathbb C}^{2n}) = 2\rho^{\mathbb H}({\mathbb H}^n)+d$
where $d = 1\mbox{ or } 3$.
In Section~4, we give necessary and sufficient conditions on
linearly independent real (respectively, complex ) vector
fields to be linearly independent complex or quaternionic
(respectively, quaternionic) vector fields.
\section{A relationship between real, complex, and quaternionic vector fields on spheres}
For any positive integer $n$, let
$r_{\mathbb C}:{\mathbb C}^n\rightarrow {\mathbb R}^{2n}$ be the realification function from ${\mathbb C}^n$ to ${\mathbb R}^{2n}$ defined by
$r_{\mathbb C}(a_1+b_1i,\ldots,a_n+b_ni)=(a_1,b_1,\ldots,a_n,b_n)$,
$c_{\mathbb H}:{\mathbb H}^n\rightarrow {\mathbb C}^{2n}$ be the complexification function from ${\mathbb H}^n$ to ${\mathbb C}^{2n}$ defined by
$c_{\mathbb H}(a_1+b_1i+c_1j+d_1k,\ldots,a_n+b_ni+c_nj+d_nk)=
(a_1+b_1i,d_1+c_1i,\ldots,a_n+b_ni,d_n+c_ni)$, and
$r_{\mathbb H}:{\mathbb H}^n\rightarrow {\mathbb R}^{4n}$ be the realification function from ${\mathbb H}^n$ to ${\mathbb R}^{4n}$ defined by
$r_{\mathbb H} = r_{\mathbb C}\circ c_{\mathbb H}$.
For $t\in {\mathbb F}$, let $\alpha_t:{\mathbb F}^n\rightarrow {\mathbb F}^n$
be the right multiplication by $t$ , i.e.,
$\alpha_t(x)=x\cdot t$ for each $x \in {\mathbb F}^n$.
The proof of the following lemma is straightforward.
\begin{lemma} \label{lem:Composition}
Let $s\in {\mathbb C}$ and $t\in {\mathbb H}$. Then
\begin{enumerate}
\item[\textup{(i)}] $c_{\mathbb H}\circ \alpha_s = \alpha_s \circ c_{\mathbb H}$.
\item[\textup{(ii)}] $r_{\mathbb C}\circ \alpha_s \circ c_{\mathbb H}=
r_{\mathbb H}\circ \alpha_s$.
\item[\textup{(iii)}] $r_{\mathbb C}\circ \alpha_s \circ c_{\mathbb H}\circ \alpha_t=
r_{\mathbb H}\circ \alpha_{ts}$.
\end{enumerate}
\end{lemma}
In the following theorem, we give a relationship between real, complex, and
quaternionic standard inner products on ${\mathbb F}^n$.
\begin{theorem} \label{theorem:InnerProduct}
Let $x,y\in {\mathbb C}^n$ and $v,w\in {\mathbb H}^n$. Then
\begin{enumerate}
\item[\textup{(i)}]$\langle x|y\rangle_{\mathbb C}=\langle r_{\mathbb C}(x)|r_{\mathbb C}(y)\rangle_{\mathbb R}
- \langle r_{\mathbb C}(x)|r_{\mathbb C}\circ \alpha_i(y)\rangle_{\mathbb R}i$.
\item[\textup{(ii)}] $\langle v|w\rangle_{\mathbb H}=
\langle c_{\mathbb H}(v)|c_{\mathbb H}(w)\rangle_{\mathbb C}- \langle c_{\mathbb H}(v)|c_{\mathbb H}\circ \alpha_j(w)\rangle_{\mathbb C}j$.
\item[\textup{(iii)}]$\langle v|w\rangle_{\mathbb H}=
\langle r_{\mathbb H}(v)|r_{\mathbb H}(w)\rangle_{\mathbb R}
- \langle r_{\mathbb H}(v)|r_{\mathbb H}\circ \alpha_i(w)\rangle_{\mathbb R}i\\
-\langle r_{\mathbb H}(v)|r_{\mathbb H}\circ \alpha_j(w)\rangle_{\mathbb R}j
-\langle r_{\mathbb H}(v)|r_{\mathbb H}\circ \alpha_k(w)\rangle_{\mathbb R}k$.
\end{enumerate}
\end{theorem}
\textbf{Proof.} Without loss of generality, we can assume that $n = 1$. Then the result follows by using Lemma~{\ref{lem:Composition}} and the definition of
the standard inner products.\\
In the following theorem, we give a relationship between real, complex, and
quaternionic vector fields on spheres.
\begin{theorem} \label{theorem:RelationshipVectorFields}
Let $v:S({\mathbb C}^n)\rightarrow {\mathbb C}^n- \{0\}$ and
$w:S({\mathbb H}^n)\rightarrow {\mathbb H}^n- \{0\}$
be two continuous functions. Then
\begin{enumerate}
\item[\textup{(i)}] $v$ is a complex vector field on $S({\mathbb C}^n)$ if and only if
$r_{\mathbb C}\circ v\circ r_{{\mathbb C}}^{-1}$ and
$r_{\mathbb C}\circ \alpha_i \circ v\circ r_{{\mathbb C}}^{-1}$ are real vector fields on $S({\mathbb R}^{2n})$.
\item[\textup{(ii)}] $w$ is a quaterionic vector field on $S({\mathbb H}^n)$ if and only if
$c_{\mathbb H}\circ w\circ c_{{\mathbb H}}^{-1}$ and
$c_{\mathbb H}\circ \alpha_j \circ w\circ c_{{\mathbb H}}^{-1}$ are complex vector fields on $S({\mathbb C}^{2n})$.
\item[\textup{(iii)}]$w$ is a quaterionic vector field on $S({\mathbb H}^n)$ if and only if
$r_{\mathbb H}\circ w\circ r_{{\mathbb H}}^{-1}$ and
$r_{\mathbb H}\circ \alpha_t \circ w\circ r_{{\mathbb H}}^{-1}$, where
$t\in \{i,j,k\}$, are real vector fields on $S({\mathbb R}^{4n})$.
\end{enumerate}
\end{theorem}
\textbf{Proof.} (i) Let $x\in S({\mathbb R}^{2n})$.
By Theorem~{\ref{theorem:InnerProduct} (i)},
$\langle r_{\mathbb C}^{-1}(x)|v(r_{\mathbb C}^{-1}(x))\rangle_{\mathbb C} = 0$ if and only if
$\langle x|r_{\mathbb C}\circ v\circ r_{{\mathbb C}}^{-1}(x)\rangle_{\mathbb R}= 0$ and
$\langle x|r_{\mathbb C}\circ \alpha_i\circ v\circ r_{{\mathbb C}}^{-1}(x)\rangle_{\mathbb R} = 0$.
The result follows. Similarly, one can prove (ii) and (iii) by
using Theorem~{\ref{theorem:InnerProduct} (ii) and (iii)}.
\begin{example} \textup{\label{exa:ComplexVectorField}
For each $n\geqslant 2$, $v:S({\mathbb C}^n)\rightarrow {\mathbb C}^n- \{0\}$ such
that $v(x_1+ix_2,\ldots,x_{2n-1}+ix_{2n})
= (-x_3+ix_4,x_1-ix_2,\ldots,-x_{2n-1}+ix_{2n},x_{2n-3}-ix_{2n-2})$ is a complex vector field on $S({\mathbb C}^n)$. By Theorem~{\ref{theorem:RelationshipVectorFields} (i)},
$r_{\mathbb C}\circ v\circ r_{{\mathbb C}}^{-1}$ and
$r_{\mathbb C}\circ \alpha_i \circ v\circ r_{{\mathbb C}}^{-1}$ are real vector fields on
$S({\mathbb R}^{2n})$. $r_{\mathbb C}\circ v\circ r_{{\mathbb C}}^{-1}$ is given by
$(x_1, \ldots, x_{2n})\mapsto
(-x_3, x_4, x_1, -x_2, \ldots,-x_{2n-1}, x_{2n}, x_{2n-3}, -x_{2n-2})$
and $r_{\mathbb C}\circ \alpha_i \circ v\circ r_{{\mathbb C}}^{-1}$ is given by
$(x_1,\ldots, x_{2n})\mapsto
(-x_4, -x_3, x_2, x_1, \ldots,-x_{2n}, -x_{2n-1}, x_{2n-2}, x_{2n-3})$.}
\end{example}
\begin{theorem} \label{theorem:LinearlyIndependentVectorFields}
Let $v_1,\ldots,v_m:S({\mathbb C}^n)\rightarrow {\mathbb C}^n- \{0\}$ and
$w_1,\ldots,w_m:S({\mathbb H}^n)\rightarrow {\mathbb H}^n- \{0\}$
be continuous functions. Then
\begin{enumerate}
\item[\textup{(i)}] $v_1,\ldots,v_m$ are linearly independent complex vector field on $S({\mathbb C}^n)$ if and only if
$r_{\mathbb C}\circ v_1\circ r_{{\mathbb C}}^{-1},
r_{\mathbb C}\circ \alpha_i \circ v_1\circ r_{{\mathbb C}}^{-1},
\ldots, r_{\mathbb C}\circ v_m\circ r_{{\mathbb C}}^{-1},
r_{\mathbb C}\circ \alpha_i \circ v_m\circ r_{{\mathbb C}}^{-1}$ are linearly independent real vector fields on $S({\mathbb R}^{2n})$.
\item[\textup{(ii)}] $w_1,\ldots,w_m$ are linearly independent quaternionic vector
field on $S({\mathbb H}^n)$ if and only if
$c_{\mathbb H}\circ w_1\circ c_{{\mathbb H}}^{-1},
c_{\mathbb H}\circ \alpha_j \circ w_1\circ r_{{\mathbb H}}^{-1},
\ldots, c_{\mathbb H}\circ w_m\circ c_{{\mathbb H}}^{-1},
c_{\mathbb H}\circ \alpha_j \circ w_m\circ c_{{\mathbb H}}^{-1}$ are linearly independent complex vector fields on $S({\mathbb C}^{2n})$.
\item[\textup{(iii)}]$w_1,\ldots,w_m$ are linearly independent quaternionic vector
field on $S({\mathbb H}^n)$ if and only if
$r_{\mathbb H}\circ w_1\circ r_{{\mathbb H}}^{-1},
r_{\mathbb H}\circ \alpha_t \circ w_1\circ r_{{\mathbb H}}^{-1},
\ldots, r_{\mathbb H}\circ w_m\circ r_{{\mathbb H}}^{-1},
r_{\mathbb H}\circ \alpha_t \circ w_m\circ r_{{\mathbb H}}^{-1}$, where $t\in\{i,j,k\}$, are linearly independent real vector fields on $S({\mathbb R}^{4n})$.
\end{enumerate}
\end{theorem}
\textbf{Proof.} (i) By Theorem~{\ref{theorem:RelationshipVectorFields} (i)}, $v_1,\ldots,v_m$ are complex vector fields on $S({\mathbb C}^n)$
if and only if $r_{\mathbb C}\circ v_1\circ r_{{\mathbb C}}^{-1},
r_{\mathbb C}\circ \alpha_i \circ v_1\circ r_{{\mathbb C}}^{-1},
\ldots, r_{\mathbb C}\circ v_m\circ r_{{\mathbb C}}^{-1},
r_{\mathbb C}\circ \alpha_i \circ v_m\circ r_{{\mathbb C}}^{-1}$
are real vector fields on $S({\mathbb R}^{2n})$. Let $x\in S({\mathbb R}^{2n}), y=r_{{\mathbb C}}^{-1}(x)$, and $a_1,b_1,\ldots, a_m,b_m\in {\mathbb R}$.
Then $(r_{\mathbb C}\circ v_1\circ r_{{\mathbb C}}^{-1}(x))a_1 +
(r_{\mathbb C}\circ \alpha_i \circ v_1\circ r_{{\mathbb C}}^{-1}(x))b_1 +
\cdots + (r_{\mathbb C}\circ v_m\circ r_{{\mathbb C}}^{-1}(x))a_m +
(r_{\mathbb C}\circ \alpha_i \circ v_m\circ r_{{\mathbb C}}^{-1}(x))b_m = 0$ if and only if $v_1(y)(a_1+b_1i)+\cdots+v_m(y)(a_m+b_mi) = 0$. The result follows.
(ii) and (iii) are similar to (i).
\begin{corollary} \label{corollary:RelationshipNumberOfVectorFields}
$\rho^{\mathbb R}({\mathbb R}^{2n})\geqslant2\rho^{\mathbb C}({\mathbb C}^n)$
and
$\rho^{\mathbb C}({\mathbb C}^{2n})\geqslant 2\rho^{\mathbb H}({\mathbb H}^n)$.
\end{corollary}
\begin{remark} \label{remark:OrthonormalVectorFields}
\textup{A unit ${\mathbb F}$-vector field on $S({\mathbb F}^n)$ is an ${\mathbb F}$-vector field on $S({\mathbb F}^n)$ whose image is a subset of $S({\mathbb F}^n)$. $m$ such unit vector fields $v_1,\ldots,v_m$ are orthonormal if
$\langle v_s(x)|v_t(x)\rangle_{\mathbb F}=\delta_{st}$ for each $s,t\in\{1,\ldots,m\}$ and all $x\in S({\mathbb F}^n)$. Observe that since the Gram-Schmidt orthonormalization process is continuous, then one can convert $m$ linearly independent
${\mathbb F}$-vector fields on $S({\mathbb F}^n)$ to $m$ orthonormal ${\mathbb F}$-vector fields on $S({\mathbb F}^n)$. Consequently, $\rho^{\mathbb F}({\mathbb F}^n)$ is equal to the maximal number of orthonormal ${\mathbb F}$-vector fields on $S({\mathbb F}^n)$.
Also, one can use Theorem~\ref{theorem:InnerProduct} and basic properties of standard
inner products to prove that Theorem~\ref{theorem:RelationshipVectorFields} remains valid if we replace ``vector field'' by ``unit vector field'' , and Theorem~\ref{theorem:LinearlyIndependentVectorFields} remains valid if
we replace ``linearly independent'' by ``orthonormal''.}
\end{remark}
\section{A relationship between $\rho^{\mathbb H}({\mathbb H}^n)$ and $\rho^{\mathbb C}({\mathbb C}^{2n})$}
For a positive integer $m$, let $c_m^{\mathbb F}$ be the
${\mathbb F}$-James number mentioned in \cite{Jam58}. Then $\rho^{\mathbb F}({\mathbb F}^n)$ is the largest $m$ such that $n$ is a multiple of $c_m^{\mathbb F}$.
$c_m^{\mathbb R}$ is computed by Adams in \cite{Ada62} and
is given by $c_m^{\mathbb R}= 2^{f(m)}$ where $f(m)$ is the number of
integers $q$ such that $0<q\leqslant m$ and $q\equiv0,1,2\mbox{ or } 4 \mbox{ mod } 8$.
In \cite{Ada62}, Adams also showed that $\rho^{\mathbb R}({\mathbb R}^n) = 8d+2^c-1$ where
$n = (2a+1)2^{4d+c}$ for some non-negative integers $a, d, c$ with $0\leqslant c \leqslant 3$.
Now, we show how to compute $\rho^{\mathbb C}({\mathbb C}^n),
\rho^{\mathbb H}({\mathbb H}^n)$. For a given prime $p$, let $\nu_p(m)$ be the exponent of $p$ in the prime decomposition of $m$.
$c_m^{\mathbb C}$ is computed by Adams-Walker \cite{AW65}
and is given by $$\nu_p(c_m^{\mathbb C})=\max\{s+\nu_p(s):0\leqslant s \leqslant \lfloor \frac{m}{p-1}\rfloor\}.$$
$c_m^{\mathbb H}$ is computed by Sigrist-Suter \cite{SS73}
and is given by $\nu_2(c_m^{\mathbb H})=\max \{2m+1, 2s+\nu_2(s):0\leqslant s \leqslant m\},$ and for an odd prime $p$,
$$\nu_p(c_m^{\mathbb H})=\max \{s+\nu_p(s):0\leqslant s \leqslant \lfloor \frac{2m}{p-1}\rfloor\}.$$
For $p\leqslant m+1$, let $$S_{m, p}=\{s:\lfloor \frac{m}{p-1}
\rfloor+\nu_p(\lfloor \frac{m}{p-1}\rfloor )-\lfloor \log_p(\frac{m}{p-1})\rfloor \leqslant s \leqslant \lfloor \frac{m}{p-1}\rfloor \},$$
and for $p > m+1$, let $S_{m, p}=\{0\}$.
Notice that if $p\leqslant m+1$, then $|S_{m, p}|= \lfloor \log_p(m/(p-1))\rfloor - \nu_p(\lfloor m/(p-1)\rfloor ) + 1$. In the following lemma, we give refined formulas for computing
$\nu_p(c_m^{\mathbb C})$ and $\nu_p(c_m^{\mathbb H})$.
\begin{lemma} \label{lem:JamesNumbers}
Let $m$ be a positive integer and $p$ be a prime number. Then
\begin{enumerate}
\item[\textup{(i)}] $\nu_p(c_m^{\mathbb C})=\max\{s + \nu_p(s):s\in S_{m,p}\}$.
\item[\textup{(ii)}] $\nu_2(c_m^{\mathbb H})=\max\{2m+1, 2s + \nu_2(s):s\in S_{m,2}\}$, and for an odd prime $p$, $\nu_p(c_m^{\mathbb H})=\max\{s + \nu_p(s):s\in S_{2m,p}\}$.
\end{enumerate}
\end{lemma}
\textbf{Proof.}
\textup{(i)} Let $\nu_p(c_m^{\mathbb C})=u + \nu_p(u)$ for some
$0\leqslant u \leqslant \lfloor m/(p-1)\rfloor$. Then
$\nu_p(u)\leqslant \lfloor \log_pu\rfloor \leqslant \lfloor\log_p(m/(p-1))\rfloor$. So,
$\lfloor m/(p-1)\rfloor+\nu_p(\lfloor m/(p-1)\rfloor )\leqslant u + \nu_p(u)\leqslant u + \lfloor\log_p(m/(p-1))\rfloor$. Hence, $u\geqslant \lfloor m/(p-1)\rfloor+\nu_p(\lfloor m/(p-1)\rfloor - \lfloor\log_p(m/(p-1))\rfloor$.The result follows.
(ii) is similar to (i).\\
Let $p_i$ be the $i$th prime number, i.e.,
$p_1 = 2,p_2 = 3, p_3 = 5, \mbox{etc}$. Let $n = p_1^{t_1}\times \cdots \times p_r^{t_r}\times l$, where $t_i\geqslant 1$ for each $i=1,\ldots, r$ and $\nu_{p_{r+1}}(l) = 0$. For each $i = 1, \ldots, r$, let
$K_{n,i}^{\mathbb C}$ be the largest $m\in\{0,\ldots,p_{r+1}-2\}$ such that
$t_i\geqslant \nu_{p_i}(c_m^{\mathbb C})$, and let $K_{n,i}^{\mathbb H}$
be the largest $m\in\{0,\ldots,(p_{r+1}-3)/2\}$ such that
$t_i\geqslant \nu_{p_i}(c_m^{\mathbb H})$. In the following theorem, we give direct formulas for computing $\rho^{\mathbb C}({\mathbb C}^n)$ and
$\rho^{\mathbb H}({\mathbb H}^n)$.
\begin{theorem} \label{theorem:FormulaComplexQuaternionicVectorFields}
Let $n = p_1^{t_1}\times \cdots \times p_r^{t_r}\times l$, where $t_i\geqslant 1$ for each $i=1,\ldots, r$ and $\nu_{p_{r+1}}(l) = 0$. Then
\begin{enumerate}
\item[\textup{(i)}]$\rho^{\mathbb C}({\mathbb C}^n) =
\min\{K_{n,i}^{\mathbb C}:i=1,\ldots,r\}$.
\item[\textup{(ii)}] $\rho^{\mathbb H}({\mathbb H}^n) =
\min\{K_{n,i}^{\mathbb H}:i=1,\ldots,r\}$.
\end{enumerate}
\end{theorem}
\textbf{Proof. }\textup{(i)} First, since $\nu_{p_{r+1}}(c_{p_{r+1}-1}^{\mathbb C}) = 1$ and $c_i^{\mathbb C}\mid c_{i+1}^{\mathbb C}$
for each $i\geqslant 1$ then $\rho^{\mathbb C}({\mathbb C}^n)\leqslant K_{n,i}^{\mathbb C}$ for each $i=1,\ldots,r$. On the other hand, if
$s = \min\{K_{n,i}^{\mathbb C}:i=1,\ldots,r\}$ then
$t_i\geqslant \nu_{p_i}(c_s^{\mathbb C})$ for each $i=1,\ldots,r$, and hence $\rho^{\mathbb C}({\mathbb C}^n)\geqslant s$. The result follows. (ii) is similar to (i).
From Corollary~\ref{corollary:RelationshipNumberOfVectorFields}, we know that the number of linearly independent complex vector fields on $S({\mathbb C}^n)$ is at most half the number of linearly independent real vector fields on $S({\mathbb R}^{2n})$, and the number of linearly independent quaternionic vector fields on $S({\mathbb H}^{n})$ is at most half the number of linearly independent complex vector fields on $S({\mathbb C}^{2n})$. The main result of this
section is the following theorem , which gives an explicit relationship between
the number of linearly independent complex vector fields on $S({\mathbb C}^{2n})$
and the number of linearly independent quaternionic vector fields on $S({\mathbb H}^{n})$.
\begin{theorem} \label{theorem:RelationshipNumberOfComplexQuaternionicVectorFields}
$\rho^{\mathbb C}({\mathbb C}^{2n}) = 2\rho^{\mathbb H}({\mathbb H}^n)+d$
where $d = 1\mbox{ or } 3$.
\end{theorem}
\textbf{Proof. } Let $\rho^{\mathbb H}({\mathbb H}^n)=m$. Then $m$ is the smallest integer such that $n$ is not a multiple of $c_{m+1}^{\mathbb H}$. From \cite{SS73}, $c_{m+1}^{\mathbb H}$ is either equal to $c_{2m+3}^{\mathbb C}/2$ or to $c_{2m+3}^{\mathbb C}$.
If $c_{m+1}^{\mathbb H}=c_{2m+3}^{\mathbb C}/2$, then $2n$ is not a multiple of
$2c_{m+1}^{\mathbb H}=c_{2m+3}^{\mathbb C}$.
On the other hand, if $c_{m+1}^{\mathbb H}=c_{2m+3}^{\mathbb C}$ then $\nu_2(c_{m+1}^{\mathbb H})= \nu_2(c_{2m+3}^{\mathbb C})= 2m+3$. Hence, $2n$ is not a multiple of $c_{m+2}^{\mathbb H}$, because $c_{m+2}^{\mathbb H}\geqslant c_{m+1}^{\mathbb H}$, $\nu_p(2n)=\nu_p(n)$ for each $p\neq 2$, and $\nu_2(c_{m+2}^{\mathbb H})\geqslant 2m+5$, see Lemma~\ref{lem:JamesNumbers}. But, $c_{m+2}^{\mathbb H}$ is either equal to $c_{2m+5}^{\mathbb C}/2$ or to $c_{2m+5}^{\mathbb C}$. Thus, in all cases, $2n$ is not a multiple of $c_{2m+5}^{\mathbb C}$, and hence
$\rho^{\mathbb C}({\mathbb C}^{2n})< 2m+5$. Now, since $c_{2k+1}^{\mathbb C}=c_{2k}^{\mathbb C}$ for each $k\geqslant 1$, see \cite{AW65}, then $\rho^{\mathbb C}({\mathbb C}^{2n})$ is odd. Consequently, by Corollary~\ref{corollary:RelationshipNumberOfVectorFields}, $\rho^{\mathbb C}({\mathbb C}^{2n})$
is either equal to $2m+1$ or to $2m+3$.\\
In the following table, we give the values of $\rho^{\mathbb R}({\mathbb R}^{4n})$, $\rho^{\mathbb C}({\mathbb C}^{2n})$, and $\rho^{\mathbb H}({\mathbb H}^{n})$ for some values of $n$.
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
$n$&
${1}$&
${2}$&
${4}$&
${6}$&
${12}$&
${24}$&
$1440$ \\
\hline
$\rho^{\mathbb R}({\mathbb R}^{4n})$&
3&
7&
8&
7&
8&
9&
15 \\
\hline
$\rho^{\mathbb C}({\mathbb C}^{2n})$&
1&
1&
1&
1&
3&
3&
5 \\
\hline
$\rho^{\mathbb H}({\mathbb H}^{n})$&
0&
0&
0&
0&
0&
1&
2 \\
\hline
\end{tabular}
\end{table}
\section{Construction of vector fields on spheres}
The construction of $\rho^{\mathbb R}({\mathbb R}^{n})$ linearly independent real vector
fields on $S({\mathbb R}^{n})$ is well understood, for constructions using
Clifford algebras see \cite{Zve68}, and for
constructions using combinatorial methods see the recent work of Ognikyan
\cite{Ogn08}. The situation is completely different with
complex and quaternionic vector fields; there is no explicit constructions
that gives two or more linearly independent complex vector fields on $S({\mathbb C}^{n})$ and there is no known construction that gives even a single
quaternionic vector field on $S({\mathbb H}^{n})$. In fact, the only known
complex vector field on $S({\mathbb C}^{n})$ is the one given
in Example~\ref{exa:ComplexVectorField}.
From Theorem~\ref{theorem:LinearlyIndependentVectorFields}, we know that one can use $m$ linearly independent complex (respectively, quaternionic) vector fields to obtain
$2m$ linearly independent real (respectively, complex) vector
fields. So, it is natural to ask if, in some way, it is possible to use
linearly independent real (respectively, complex) vector fields to build
some linearly independent complex or quaternionic (respectively, quaternionic) vector
fields.
In the following theorem, we give necessary and sufficient conditions on
linearly independent real (respectively, complex) vector fields to be
linearly independent complex or quaternionic (respectively, quaternionic)
vector fields. For simplicity, let $r_{{\mathbb C},i}=r_{\mathbb C}\circ \alpha_i\circ
r_{\mathbb C}^{-1}$, $c_{{\mathbb H},j}=r_{\mathbb C}\circ \alpha_j\circ
r_{\mathbb H}^{-1}$, and for each $t\in \{i,j,k\}$, let $r_{{\mathbb H},t}=r_{\mathbb H}\circ \alpha_t\circ r_{\mathbb H}^{-1}$.
\begin{theorem} \label{theorem:ConstructionOfVectorFields}
\begin{enumerate}
\item[\textup{(i)}] Suppose $u_1,\ldots,u_m$ are linearly independent real vector
fields on $S({\mathbb R}^{2n})$. Then
$r_{\mathbb C}^{-1}\circ u_1 \circ r_{\mathbb C},\ldots,
r_{\mathbb C}^{-1}\circ u_m \circ r_{\mathbb C}$ are linearly independent complex vector field on $S({\mathbb C}^n)$ if and only if
$u_1,\ldots,u_m,r_{{\mathbb C},i}\circ u_1,\ldots,r_{{\mathbb C},i}\circ u_m$ are linearly independent real vector fields on $S({\mathbb R}^{2n})$.
\item[\textup{(ii)}] Suppose $v_1,\ldots,v_m$ are linearly independent complex vector
fields on $S({\mathbb C}^{2n})$. Then
$c_{\mathbb H}^{-1}\circ v_1 \circ c_{\mathbb H},\ldots,
c_{\mathbb H}^{-1}\circ v_m \circ c_{\mathbb H}$ are linearly independent quaternionic vector field on $S({\mathbb H}^n)$ if and only if
$v_1,\ldots,v_m,c_{{\mathbb H},j}\circ v_1,\ldots,c_{{\mathbb H},j}\circ v_m$ are linearly independent complex vector fields on $S({\mathbb C}^{2n})$.
\item[\textup{(iii)}]Suppose $u_1,\ldots,u_m$ are linearly independent real vector
fields on $S({\mathbb R}^{4n})$. Then
$r_{\mathbb H}^{-1}\circ u_1 \circ r_{\mathbb H},\ldots,
r_{\mathbb H}^{-1}\circ u_m \circ r_{\mathbb H}$ are linearly independent quaternionic vector field on $S({\mathbb H}^n)$ if and only if
$u_1,\ldots,u_m,r_{{\mathbb H},t}\circ u_1,\ldots,r_{{\mathbb H},t}\circ u_m$, where
$t\in \{i,j,k\}$, are linearly independent real vector fields on
$S({\mathbb R}^{4n})$.
\end{enumerate}
\end{theorem}
\textbf{Proof.} Follows directly from Theorem~\ref{theorem:LinearlyIndependentVectorFields}.\\
In \cite{Bec72}, Becker solved several important cases of
the equivariant real vector fields problem, both the maximal number and the
construction, on spheres with free group action by using the known Clifford
algebras constructions of the non-equivariant real vector fields on spheres.
As noted in \cite{Obi06} and \cite{Ond01}, both sides of the equivariant complex and quaternionic vector fields problem on spheres with free group action is still completely
open. By using a method similar to that used by Becker, the construction of
the non-equivariant complex and quaternionic vector fields on spheres might
lead to the solution of the equivariant complex and quaternionic vector
fields problem on spheres with free group action.
|
\section{Proofs}
\begin{proof}[Proof of Theorem \ref{thm:chernoff}]
In the theorem presented below, we obtain a Chernoff bound for a random variable $Z$ from an exponential family supported on $\mathbb{R}^d$ and with natural parameter $\beta^*\in \mathbb{R}^d$ and a log-MGF $\Lambda(\eta; \beta^*)$, evaluated at $\eta$.
Theorem \ref{thm:chernoff} will follow as a corollary.
\begin{theorem} \label{thm:gen-chernoff}
If $Z\sim \exp({\beta^*}^T Z-H(\beta^*))$, then for a convex, compact, selective region ${\mathcal{K}} \subset \mathbb{R}^{d}$, the selection probability of the event $\{z:z\in {\mathcal{K}}\}$ is bounded above as
\begin{equation}
\label{sel-exp-cher-thm}
\log\mathbb{P}_{\beta^*} (Z\in {\mathcal{K}})\leq -\inf\limits_{z\in K}\left\{\sup_{\eta} \eta^T z -(H(\beta ^*+ \eta)-H(\beta^*))\right\} = -\inf\limits_{z\in K}\left\{\sup_{\eta} \eta^T z - \Lambda(\eta; \beta^*)\right\}.
\end{equation}
\end{theorem}
To prove Theorem \ref{thm:chernoff}, apply \eqref{sel-exp-cher-thm} to a Gaussian random variable with mean $X^*\beta^*$ and covariance matrix $\sigma^2 I_n$ to get the rate function as:
$$\rho(z)=\inf_{\eta}-\eta^T (z-X^*\beta^*) + \frac{1}{2}\sigma^2\eta^T \eta,$$ by plugging log-MGF of $Y$ evaluated at $\eta$ as $ \Lambda(\eta; \beta^*)= \eta^T X^*\beta^* + \frac{1}{2}\sigma^2 \eta^T \eta$.
\noindent The above infimum occurs at $\eta=\cfrac{(z-X^*\beta^*)}{\sigma^2}$,
yielding the bound
$$\sup_{z\in {{\mathcal{K}}}}- \frac{1}{2\sigma^2}\lVert(z-X^*\beta^*)\lVert^2 =-\inf_{z\in {{\mathcal{K}}}}\frac{1}{2\sigma^2}\lVert(z-X^*\beta^*)\lVert^2,$$
which proves
$$\log\mathbb{P}(Y\in {\mathcal{K}} \lvert \beta^*)\leq -\dfrac{\lVert{X^*\beta^*-{\mathcal{P}}_{{\mathcal{K}}}(X^*\beta^*)}\lVert^{2}}{2\sigma^2}.$$
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:gen-chernoff}]
To prove \eqref{sel-exp-cher-thm},
\begin{align*}
\log \mathbb{P}(Z \in {\mathcal{K}} )
&= \log\mathbb{E}(1_{\{Z\in{\mathcal{K}}\}})\leq \log\mathbb{E}(\exp(\eta^T Z +u))
\end{align*}
for any $\eta\in \mathbb{R}^d$ and for any $u$ satisfying $\eta^T z + u\geq 0\; \forall z\in {\mathcal{K}}$.
\noindent Upon fixing a choice of $u$ as $u_{\text{opt}}(\eta)=\sup_{z\in {\mathcal{K}}}-\eta^T z$, the upper bound can be re-written as:
\begin{align*}
\log \mathbb{P}(Z \in {\mathcal{K}} )
&\leq u_{\text{opt}}+\log\mathbb{E}(\exp(\eta^T Z )\lvert \beta^*) \text{ for all } \eta \in \mathbb{R}^d\\
&= \sup_{z\in {\mathcal{K}}}-\eta^T z + (H(\beta^* + \eta)-H(\beta^*)) \text{ for all } \eta\in \mathbb{R}^d\\
&\leq \inf_{\eta}\left\{\sup_{z\in {\mathcal{K}}}-\eta^T z + (H(\beta^* + \eta)-H(\beta^*))\right\}\\
&= -\sup_{\eta}\left\{\inf_{z\in {\mathcal{K}}}\eta^T z -(H(\beta^* + \eta)-H(\beta^*))\right\}\\
&=-\inf\limits_{z\in K}\left\{\sup_{\eta} \eta^T z -(H(\beta^* + \eta)-H(\beta^*))\right\}\\
&=-\inf\limits_{z\in K}\left\{\sup_{\eta} \eta^T z -\Lambda(\eta; \beta^*)\right\}.
\end{align*}
The third inequality follows by an optimization over $\eta\in \mathbb{R}^d$ and the last one follows from a minimax theorem argument by noting that the function over which the optimization takes place, given by
$$f(z,\eta)=\eta^T z -(H(\beta^* + \eta)-H(\beta^*)).$$
is convex-concave and continuous as a map of $z$ for every $\eta\in \mathbb{R}^d$.
\end{proof}
\medskip
\begin{proof}[Proof of Lemma \ref{lem:map-explicit-gen}]
Differentiating \eqref{eq:map-obj} for $w(z)=w^{\text{bar}}(z)$ with respect to $\beta^*$ gives the estimating equation
$$
X^{*T}{\nabla} \bar{w}^*(X^*\hat{\beta}^*)=X^{*T} y +{\nabla} \log\pi(\hat{\beta}^*).
$$
Now, observing that the maximizer of objective
$$
z^T \mu - \frac{1}{2} \|z\|^2-\sum_{j=1}^m \log (1+1/(b_i - a_i^Tz)),
$$
satisfies
$
{\nabla} \bar{w}(z^*(\mu))=\mu,
$
the statement follows.
\end{proof}
\medskip
\begin{proof}[Proof of Lemma \ref{lem:map-rand-sat}]
The randomized selective MAP satisfies
$$
-y + \frac{\gamma^2}{{1 + \gamma^2}}\hat{\beta}^*+ \cfrac{1}{(1 + \gamma^2) }\nabla \bar{w}^*\left(\cfrac{\hat{\beta}^*}{1 + \gamma^2}\right) - \nabla \log \pi(\hat{\beta}^*) = 0,
$$
where we note that the selection probability is approximated by the optimization problem
$$\bar{w}^*(\mu)=\sup\limits_{z\in \mathbb{R}^n}z^T \mu-\bar{w}(z),$$
with $w(z)$ defined for the polyhedral region $Ay\leq b$. This implies that
\begin{align*}
\nabla \bar{w}^*\left(\cfrac{\hat{\beta}^*}{1 + \gamma^2}\right) &= (1 + \gamma^2)\big(y + \nabla \log \pi(\hat{\beta}^*)\big) - \gamma^2\hat{\beta}^* ,
\end{align*}
which results in estimating equation
$$
\hat{\beta}^* = (1 + \gamma^2) \nabla \bar{w}\left( (1 + \gamma^2)\big(y + \nabla \log \pi(\hat{\beta}^*)\big) - \gamma^2\hat{\beta}^* \right).
$$
\end{proof}
\medskip
\begin{proof}[Proof of Theorem \ref{LDP:approximating:probability}]
It suffices to show
\begin{equation*}
-\inf_{x\in \mathcal{K}}\Lambda^*(x;\beta^*) \leq \liminf_n\frac{1}{n}\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*) \leq \limsup_n \frac{1}{n}\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*) \leq -\inf_{x \in \mathcal{K}}\Lambda^*(x;\beta^*)
\end{equation*}
in order to prove \ref{lim:sel:prob}.
To see the upper bound, we have along similar lines of theorem A.1 in the paper that
\begin{equation*}
\begin{aligned}
\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*) &\leq \log \mathbb{E}(\exp(n\eta^T \bar{Y}_n + u)\lvert \beta^*) \text{ for every } \eta \text{ and } u : n\eta^T x + u \geq 0 \text{ for } x \in \mathcal{K}\\
& \leq n \cdot \sup_{x\in \mathcal{K}}-\eta^T x + \Lambda(\eta;\beta^*)\\
& \leq -n\cdot \sup_{\eta}\{\inf_{x\in \mathcal{K}}\eta^T x - \Lambda(\eta;\beta^*)\}\\
&= -n\cdot\inf_{x\in \mathcal{K}}\{\sup_{\eta}\eta^T x - \Lambda(\eta;\beta^*) \} \text{ by a minimax argument}\\
&= -n\inf_{x\in \mathcal{K}}\Lambda^*(x;\beta^*).
\end{aligned}
\end{equation*}
This completes the proof that
\[\limsup_n \frac{1}{n}\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*) \leq -\inf_{x \in \mathcal{K}}\Lambda^*(x;\beta^*).\]
To prove the lower bound, it suffices to show for every $x \in \mathcal{K}^{o}$ and any $\delta>0$ that
\begin{equation}
\label{lower_bound}
\liminf_n\frac{1}{n}\log \mathbb{P}(\bar{Y}_n \in \mathcal{B}(x,\delta)\lvert \beta^*) \geq -\Lambda^*(x; \beta^*)
\end{equation}
where $ \mathcal{B}(\bar{x},\delta) = \{y: \|y-x\|_{\infty} \leq \delta\}$. This shall prove
\begin{equation}
\liminf_n\frac{1}{n}\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*) \geq -\inf_{x \in \mathcal{K}^{o}}\Lambda^*(x;\beta^*) = -\inf_{x\in \mathcal{K}}\Lambda^*(x;\beta^*).
\end{equation}
For this $x \in \mathcal{K}^{o}$, denote
\[\eta^* = \operatornamewithlimits{argsup}_{\eta} \eta^T x -\Lambda(\eta;\beta^*)\]
which satisfies ${\nabla} \Lambda(\eta^*;\beta^*) = x.$
To prove the above, define an exponentially tilted measure $\tilde{\mathbb{P}}$ as
\[d\tilde{\mathbb{P}}(z)= \exp({\eta^*}^T z - \Lambda(\eta^*)) d\mathbb{P}(z).\]
For $0<\delta_1<\delta$
\begin{equation*}
\begin{aligned}
\mathbb{P}(&\bar{Y}_n \in \mathcal{B}(x,\delta)) \geq \mathbb{P}(\bar{Y}_n \in \mathcal{B}(x,\delta_1))\\
&= \exp(n\Lambda(\eta^*;\beta^*))\int_{ \mathcal{B}(\bar{z},\delta_1)} \exp(-n {\eta^*}^T x)\exp(-{\eta^*}^T\{(z_1 + z_2 +\cdots + z_n)-n x\})d\tilde{\mathbb{P}}(z_1)\cdots d\tilde{\mathbb{P}}(z_n)\\
&= \exp(-n \Lambda^*(x;\beta^*) - n|\eta^*|^T \delta_1)
\end{aligned}
\end{equation*}
Finally, we have \eqref{lower_bound} by letting $\delta_1 \to 0$ in the above lower bound.
\end{proof}
\bigskip
\begin{proof}[Proof of Corollary \ref{LDP:approximate:selective:posterior}]
It follows from \ref{LDP:approximating:probability} that the sequence of true selective posteriors, written as
\[\log\pi_S(\beta^*\lvert\bar{y}_n) =\log \pi(\beta^*) -n\|\bar{y}_n - \beta^*\|^2/2 -\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*)\] can be approximated by
\[\log \pi(\beta^*) -n\|\bar{y}_n - \beta^*\|^2/2 + n \inf_{x\in \mathbb{R}^p}\left\{\|x-\beta^*\|^2/2\right\}.\] Now, it is easy to see that as long as the sequence of convex barrier functions $\frac{1}{n}w_{1/\sqrt{n}}(x) \to I_{\mathcal{K}}(x)$ for all $x \in \mathbb{R}^d$ as $n \to \infty$, the limiting sequence of objectives
\[ \|x-\beta^*\|^2/2 +\frac{1}{n}w_{1/\sqrt{n}}(x)\]
are convex in $x$ and converge to the continuous, convex objective $ \|x-\beta^*\|^2/2 + I_{\mathcal{K}}(x)$, which has a unique minimum. Hence, we also have \[\inf_{x\in \mathbb{R}^d}\left\{\|x-\beta^*\|^2/2 +\frac{1}{n}w_{1/\sqrt{n}}(x)\right\} \to \inf_{x\in \mathbb{R}^d}\left\{\|x-\beta^*\|^2/2 +I_{\mathcal{K}}(x)\right\}\text{ as } n \to \infty.\]
\end{proof}
\bigskip
\begin{proof}[Proof of Theorem \ref{thm:inconsistency}]
By Theorem \ref{thm:map-explicit}, the maximizer of \eqref{soft:max:MLE} is
\[
\hat{\beta}^*_n = \bar{Y}_n - \frac{1}{\sqrt{n}(\sqrt{n}\bar{Y}_n){\sqrt{n}\bar{Y}_n + 1)}}.
\]
Denoting $Z_n = (\sqrt{n}|\beta^*|)\sqrt{n}\bar{Y}_n$ and $b(z)= \log\left(1+\dfrac{1}{z}\right)$ we have
$$
\begin{aligned}
\hat{\beta}^*_n - \beta^* &= \bar{Y}_n - \beta^* - n^{-1/2} \frac{1}{\sqrt{n}\bar{Y}_n{\sqrt{n}\bar{Y}_n + 1)}}\\
& = \bar{Y}_n - \beta^* + n^{-1/2} \nabla b(n^{1/2} \bar{Y}_n)\\
& = \frac{Z_n}{n |\beta^*|}- \beta^* + n^{-1/2}\nabla b\left(\frac{Z_n}{n^{1/2}|\beta^*|} \right)\\
\end{aligned}
$$
For $\beta^*= -\delta<0$, $Z_n = (n^{1/2}|\beta^*|) n^{1/2} \bar{Y}$ converges in law to an exponential
random variable with mean 1. Further note that,
\[z\nabla b(z) \to -K \text{ as } z\downarrow 0 \]
for $K=1$.
Using these two facts, we have $\dfrac{Z_n}{n |\beta^*|}=o_p(1)$ and
\[n^{-1/2}\nabla b\left(\frac{Z_n}{n^{1/2}|\beta^*|} \right) = \left(\frac{|\beta^*|}{Z_n}\right)\frac{Z_n}{n^{1/2}|\beta^*|}\nabla b\left(\frac{Z_n}{n^{1/2}|\beta^*|} \right) = \left(\frac{|\beta^*|}{Z_n}\right) W_n \]
where $W_n \to -K \text{ as } n \to \infty.$ Hence, we can approximate the sequence of random variables $\hat{\beta}^*_n - \beta^*$ in distribution by the random variable \[-\frac{K|\beta^*|}{Z} - \beta^* \text{ where } Z \sim \text{Exp}(1).\]
Set $\beta^* = -\delta$ for $\delta>0$
$$
\begin{aligned}
\mathbb{P}\left(|\hat{\beta}^*_n - \beta^*|>\frac{\delta}{2}\lvert \beta^*\right) &\approx \mathbb{P}\left(\Big\lvert \frac{K|\beta^*|}{Z} + \beta^*\Big \lvert>\frac{\delta}{2} \;\Big\lvert \beta^*\right)\\
& \geq \mathbb{P}\left(-\frac{K|\beta^*|}{Z} - \beta^*>\frac{\delta}{2} \;\Big\lvert \beta^*\right) \\
& = \mathbb{P}\left(Z > -\frac{K|\beta^*|}{\beta^* + \delta/2}\;\Big\lvert \beta^*\right)\\
&= \exp\left(-\frac{K|\beta^*|}{\beta^* + \delta/2}\right) = \exp(2K)>0.
\end{aligned}
$$
\end{proof}
\bigskip
\begin{proof}[Proof of Lemma \ref{strong:convexity}]
$\hat{C}_n(.)$ is a strongly convex as $ \cfrac{\gamma^2}{2(1+\gamma^2)}{\|\beta^*\|}^2$ is strongly convex with index $\gamma^2/(1+\gamma^2)$ and $\bar{w}_n^*(\beta^*)$ is a convex function in $\beta^*$. It is straight forward from here to see that the indices of convexity of the sequence of approximate log partition functions are bounded below by $M= \gamma^2/(1+\gamma^2)>0.$ The same goes for the sequence $-\ell_S^{n}(\beta^*)$ which has a linear term $-\beta^* \bar{y}_n$ added to $\hat{C}_n(\beta^*)$.
\end{proof}
\bigskip
\begin{proof}[Proof of Lemma \ref{identity:convexity}]
The randomized selective MLE $\hat{\beta}^*_n$ satisfies
\[{\nabla}\hat{C}_n(\hat{\beta}^*_n)= \bar{y}_n,\; \text{ that is } \hat{\beta}^*_n = {\nabla}\hat{C}_n^{-1}(\bar{y}_n).\] Thus, we have
\[\|\hat{\beta}^*_n-\beta^*\|^2 = \| {\nabla}\hat{C}_n^{-1}(\bar{y}_n) - \beta^*\|^2 = \|{\nabla}\hat{C}_n^{*}(\bar{y}_n) - {\nabla}\hat{C}_n^{*}({\nabla} \hat{C}_n(\beta^*))\|^2.\]
Lemma \ref{strong:convexity} shows that $\hat{C}_n(\beta^*)$ is strongly convex with indices of convexity $m_n\geq M = \gamma^2/(1+\gamma^2)$ and the proof is complete by using the fact that the convex conjugate of a strongly convex function with index $m_n$ is Lipschitz smooth with Lipschitz index $1/m_n$.
Hence, we have
\[\|\hat{\beta}^*_n-\beta^*\|^2 \leq \cfrac{1}{m_n^2}\|\bar{y}_n- {\nabla} \hat{C}_n(\beta^*)\|^2\leq \cfrac{1}{M^2}\|\bar{y}_n- {\nabla} \hat{C}_n(\beta^*)\|^2.\]
\end{proof}
\bigskip
\begin{proof}[Proof of Theorem \ref{consistency:approximate:MLE}]
For a fixed $\epsilon >0$, Markov's inequality and \ref{identity:convexity} yields
\[\mathbb{P}(\|\hat{\beta}^*_n-\beta^*\|>\epsilon \lvert\; \bar{Y}_n^* \in \mathcal{K}) \leq \dfrac{\mathbb{E}\left[n\|\hat{\beta}^*_n-\beta^*\|^2 \;\Big\lvert \; \bar{Y}_n^* \in \mathcal{K}\right]}{n\epsilon^2}\leq \dfrac{\mathbb{E}\left[n\|\bar{Y}_n-{\nabla} \hat{C}_n(\beta^*)\|^2\;\Big\lvert\;\bar{Y}_n^* \in \mathcal{K}\right]}{n\epsilon^2}.\]
Denoting ${C}_n(\beta^*)$ as the log partition function of the true randomized selective likelihood $f_S^{n}(\beta^*)$, we have
\begin{equation*}
\begin{aligned}
\dfrac{\mathbb{E}\left[n\|\bar{Y}_n-{\nabla} \hat{C}_n(\beta^*)\|^2\;\Big\lvert\;\bar{Y}_n^* \in \mathcal{K}\right]}{n\epsilon^2} &= \dfrac{\mathbb{E}\left[n\|\bar{Y}_n-{\nabla} {C}_n(\beta^*)\|^2\;\Big\lvert\;\bar{Y}_n^* \in \mathcal{K}\right]}{n\epsilon^2} + \dfrac{\|{\nabla} {C}_n(\beta^*)-{\nabla}\hat{C}_n(\beta^*)\|^2}{\epsilon^2}\nonumber\\
&= \dfrac{\mathrm{Var}\left[\sqrt{n}\bar{Y}_n\;\lvert\;\bar{Y}_n^* \in \mathcal{K}\right]}{n\epsilon^2}+ \dfrac{\|{\nabla} {C}_n(\beta^*)-{\nabla}\hat{C}_n(\beta^*)\|^2}{\epsilon^2}\nonumber \\
&\leq \dfrac{\mathrm{Var}\left[\sqrt{n}\bar{Y}_n\right]}{n\epsilon^2} + \dfrac{\|{\nabla} {C}_n(\beta^*)-{\nabla}\hat{C}_n(\beta^*)\|^2}{\epsilon^2}
\end{aligned}
\end{equation*}
The last step uses the fact that variance of a Gaussian random variable reduces when restricted to a convex set (see \cite{kanter1977reduction} for a proof) and thus, the first term converges to $0$ as $n \to \infty$. Convergence of the second term to $0$ follows from a combination of corollary \ref{LDP:approximate:selective:posterior} and the properties of convexity and differentiability $\hat{C}_n(\beta^*)$.
\end{proof}
\bigskip
\begin{proof}[Proof of Theorem \ref{mle:convergence}]
We shall show that
\[\mathbb{P}\left( R_n(\delta) \leq \frac{1}{2}\inf_{s:\|s-\hat{\beta}^*_n\| =\delta}\{\hat{\ell}_S^n(\hat{\beta}^*_n)-\hat\ell_S^n(s)\}\right) \leq \mathbb{P}(\|\hat{\beta}^*_n -\tilde{\beta}^*_n\|\leq \delta)\]
where $R_n(\delta) = \sup_{\|s-\hat{\beta}^*_n\|\leq \delta}|\ell_S^n (s) -\hat{\ell}_S^n(s)|.$
Let
$p_n =\hat{\beta}^*_n +\alpha u $ for a unit vector $u$ and $\alpha>\delta$. Then
\[\ell_S^n(\hat{\beta}^* +\delta u) = \ell_S^n \left(\left(1-\frac{\delta}{\alpha}\right)\hat{\beta}^*_n +\frac{\delta}{\alpha} p_n\right) \geq \left(1-\frac{\delta}{\alpha}\right) \ell_S^n(\hat{\beta}^*_n) +\frac{\delta}{\alpha} \ell_S^n(p_n).\]
\begin{equation*}
\begin{aligned}
\frac{\delta}{\alpha} \{\ell_S^n(\hat{\beta}^*_n)-\ell_S^n(p_n)\} &\geq \ell_S^n(\hat{\beta}^*_n)- \ell_S^n (\hat{\beta}^*_n+ \delta u)\\
&=(\ell_S^n(\hat{\beta}^*_n) - \hat{\ell}_S^n(\hat{\beta}^*_n)) - \{ \ell_S^n (\hat{\beta}^*_n+ \delta u) - \hat{\ell}_S^n(\hat{\beta}^*_n + \delta u)\} +( \hat{\ell}_S^n(\hat{\beta}^*_n )-\hat{\ell}_S^n(\hat{\beta}^*_n+\delta u) ) \\
&\geq \inf_{s:\|s-\hat{\beta}^*_n\| =\delta}\{ \hat{\ell}_S^n(\hat{\beta}^*_n)-\hat\ell_S^n(s)\} -2\cdot\sup_{\|s-\hat{\beta}^*_n\|\leq \delta}|\ell_S^n (s) -\hat{\ell}_S^n(s)|
\end{aligned}
\end{equation*}
We note that the event \[R_n(\delta) \leq \frac{1}{2}\inf_{s:\|s-\hat{\beta}^*_n\| =\delta}\{ \hat{\ell}_S^n(\hat{\beta}^*_n)-\hat\ell_S^n(s)\},\]
implies $\ell_S^n(\hat{\beta}^*_n)-\ell_S^n(p_n)>0 $ for all $p_n =\hat{\beta}^*_n +\alpha u $ and for any $\alpha>\delta$, which means that the maximizer of $\ell_S^n(.)$ given by $\tilde{\beta}^*_n$ lies inside a $\delta$ ball around $\hat{\beta}^*_n $, the maximizer of the pseudo selective posterior sequence.
Thus, we have
\begin{equation*}
\begin{aligned}
\mathbb{P}(\|\hat{\beta}^*_n(\bar{y}_n)- \tilde{\beta}^*_n(\bar{y}_n) \| \geq \delta) \leq \mathbb{P}\left(R_n(\delta) \geq \frac{1}{2}\inf_{s:\|s-\hat{\beta}^*_n\| =\delta}\{\hat{\ell}_S^n(\hat{\beta}^*_n)-\hat\ell_S^n(s)\}\right).
\end{aligned}
\end{equation*}
The above equation holds for any choice of norm $\|.\|$.
To complete the proof, we note from \ref{strong:convexity} that negative of $\hat\ell_S^n$ is strongly convex with index $m_n$ and hence, for any $s$
$$\hat{\ell}_S^n( \hat{\beta}^*_n)-\hat{\ell}_S^n(s) \geq \frac{1}{2}m_n\|s-\hat{\beta}^*_n\|^2, $$
which implies
\[\inf_{s:|s-\hat{\beta}^*_n| =\delta}\{ \hat{\ell}_S^n( \hat{\beta}^*_n)-\hat{\ell}_S^n(s)\} \geq \frac{1}{2}m_n\delta^2\geq \frac{1}{2}M\delta^2. \]
Using the fact that the randomized selective MLE $\hat{\beta}^*_n$ is stochastically bounded under the selective law (follows from Theorem \ref{consistency:approximate:MLE}) and uniform convergence of $R_n(\delta)$ on compact sets (from Remark \ref{unif:convergence})
\[\mathbb{P}(\|\hat{\beta}^*_n(\bar{y}_n)- \tilde{\beta}^*_n(\bar{y}_n) \|\geq \delta) \leq\mathbb{P}\left(R_n(\delta) \geq \frac{1}{2}\inf\limits_{s:\|s-\hat{\beta}^*_n\| =\delta}\{ \hat{\ell}_S^n( \hat{\beta}^*_n)-\hat{\ell}_S^n(s)\}\right)\geq \mathbb{P}\left(R_n(\delta) \geq \frac{1}{2}M\delta^2\right) \to 0.\]
\end{proof}
\section{An Algorithm for Computing ${\mathcal{P}}_{\mathcal{K}}(\mu)$
We present a numerical optimization algorithm that can be implemented to find the projection of mean vector $\mu$ onto the convex selection region ${\mathcal{K}}=\{y: A_Ey\leq b_E\}$ for approximating the Gaussian volume of selection region ${\mathcal{K}}$. The convex problem of interest can be framed as below
$$\min_{\mu:A_E\mu\leq b_E} \frac{1}{2}\lVert(\mu-X_E\beta_E) \lVert^2.$$
We employ the ADMM algorithm \citep{boyd2011distributed}, which augments the term $\frac{\rho}{2}\lVert(A_E\mu-b_E-r) \lVert^2$ to the Lagrangian to solve the optimization problem:
$$\min_{\mu,r,y:r=A_E\mu-b_E}\frac{1}{2}\lVert(\mu-X_E\beta_E) \lVert^2 +y^{T}(A_E\mu-b_E-r))+\frac{\rho}{2}\lVert(A_E\mu-b_E-r) \lVert^2+\mathbbm{1}_{r\leq 0}$$\\
The updates for the ADMM algorithm are given by:
\begin{equation}
\label{label-mu1 update}
\mu^{(k+1)}=(I+\rho A_E^{T}A_E)^{-1}\left(X_E\beta_E+\rho A_E^{T}\left(b_E+r^{(k)}-u^{(k)}\right)\right)
\end{equation}
\begin{equation}
\label{label-r1 update}
r^{(k+1)}=\min\left(A_E\mu^{(k+1)}-b_E+u^{(k)},0\right)
\end{equation}
\begin{equation}
\label{label-gradient1 step}
u^{(k+1)}=u^{(k)}+(A_E\mu^{(k+1)}-b_E-r^{(k+1)})
\end{equation}\\
We can adopt the adaptive Metropolis-Hastings with the selection probability for each fresh draw computed via the above the ADMM updates.
\medskip
\section{Introduction} \label{sec:introduction}
An increasing concern about reproducibility in scientific research has recently brought much attention to the problem of selective inference.
Informally, the term ``selective inference" (sometimes also called {\it data snooping} or {\it adaptive data analysis}) refers to a situation in which the analyst interacts with the data to decide what questions about an underlying population she would like to address.
In this situation, methods which do not take selection into account are generally invalid: the relevant sampling distribution for inference is altered by selection if the same data that was used to choose the question(s) of interest, is used for inference.
For illustration, suppose that an analyst would like to model the relationship between a certain response and a set of potential explanatory variables.
Looking at the data, she includes in the model only the few variables which appear to be most correlated with the response.
If inference about the chosen model ignores selection, it would probably suggest (by construction) that the underlying selected correlations are large, even if all of them are in fact very small.
The problem of correcting for selection effect is not new, and has been formulated in various ways and with different goals in many previous works.
Correspondingly, different approaches to adjusting inference for selection have been suggested, typically with the intention of constructing statistical procedures that have the same (or approximately the same) desirable properties as in the non-selective case.
This includes bias-reducing methods based on an extended definition of Uniform Minimum-Variance Unbiased estimation \citep{robbins1988uv, cohen1989two}, Bayesian approaches \citep{efron2011tweedie} and bootstrapping \citep{simon2013estimating};
as well as methods that work by performing simultaneous inference \citep{berk2013valid}.
Recently, tools from information theory \citep{russo2015controlling} and Differential Privacy \citep{dwork2014preserving, dwork2015generalization} have also been proposed to quantify and control the effect of selection.
Another common approach is to {\it condition} on selection, namely, base inference on the likelihood of the observed data when truncated to the set of all realizations that would result in the analyst posing the same question. A conditional approach has been pursued by several authors that addressed selective inference in so-called large-scale inference problems \citep{efron2012large}.
Underlying such problems is a sequence model, where each observation corresponds to a single parameter, and the parameters are typically not assumed to have any relationship with one another.
Conditional inference for the effects corresponding to the $K$ largest statistics in the sequence model (as described above for $K=1$) was proposed in \citet{reid2014post}.
\citet{zollner2007overcoming} and \citet{zhong2008bias} suggested point estimators and confidence intervals for the parameter of a univariate Gaussian distribution conditional on exceeding a fixed threshold, in the context of genome-wide association studies;
\citet{weinstein2013selection} constructed confidence intervals, and \citet{benjamini2014selective} explored point estimators, for the univariate truncated Gaussian problem with a more flexible (random) choice of cutoff;
\citet{yekutieli2012adjusted} proposed to base inference on the truncated likelihood while incorporating a prior on the parameters; and \citet{simonsohn2014p} proposed a frequentist method to assess effect sizes from the distribution of the p-values corresponding to only the selected.
More recently, the practicability of the conditional approach has been extended considerably from the sequence model to the realm of linear models and GLMs \citep[][among others]{lee2013exact, taylor2013tests, taylor2014exact, lee2014exact, fithian2014optimal, tian2015selective}.
In these works the selection protocol is assumed to partition the sample space into polyhedral (or at least convex) sets, fitting many popular `automatic' model selection procedures such as marginal screening, Lasso, forward-selection etc.
Other forms of selection rules was considered in \citet{loftus2015selectivea, loftus2015selectiveb, yang2016selective}.
The techniques developed in this line of work have practical importance as they allow to carry out exact inference after variable selection, which is one of the most popular situations where the problem of selective inference arises.
In the current work, we propose to give post-selection inference in the linear model by using the truncated likelihood in a Bayesian framework.
As conceptualized by \citet{george2012jsm}, Bayesian inference after selection relies on a model that prepends a prior to the truncated likelihood.
This results in a posterior distribution that is affected by selection, unlike in the usual framework for Bayesian variable selection.
While the ideas have already been proposed, \citeauthor{george2012jsm} have considered only simple selection rules, where computational complications due to the fact that the truncated likelihood replaces the usual likelihood, can be easily overcome.
In the current work we tackle the computational difficulties associated with carrying out selection-adjusted Bayesian inference in a more general setup, namely, when the selection rule partitions the sample space into polyhedral subsets.
The main contribution of the current paper is in offering technical tools that enable to provide selection-adjusted Bayesian inference in the linear model, and, importantly, are scalable to work under fairly general selection schemes.
The crucial element in our methodology is a tractable approximation to the selection probability given in \eqref{eq:adj-factor}, that has several favorable properties:
\begin{itemize}
\item The approximation to \eqref{eq:adj-factor} leads to a convex approximation to the truncated likelihood, which makes it computationally easy to numerically optimize and to analyze
\item The approximation to \eqref{eq:adj-factor} that we offer is shown in Section \ref{sec:asymptotic} to be a large deviations approximation to the exact selection probability
\item We demonstrate that when the corresponding approximate maximum likelihood estimator (MLE) is not a consistent estimator, then applying similar methods to a randomized version of the data can produce a consistent estimator.
\end{itemize}
That the truncated likelihood is involved connects our work to existing frequentist work on post model-selection inference.
However, performing Bayesian inference requires new techniques: the frequentist theory developed so far for the linear model \citep[e.g.,][]{lee2013exact} focuses on hypothesis testing for linear functionals of the parameter, and relies on reducing the problem to a truncated problem free of nuisance parameters.
By contrast, in our framework inference relies on the selection-adjusted posterior, which involves the {\it full} truncated likelihood.
Consequently, even inference for a one-dimensional linear projection of the parameter vector requires to handle nuisance parameters.
Moreover, the frequentist tests and confidence intervals developed in \citep[e.g.,][]{lee2013exact}, based on a univariate pivot statistic, have some optimality properties under a particular generative model for the data \citep[referred to as the ``saturated model'' below; see][]{fithian2014optimal} but not in general (for example, under the selected model they do not).
Our procedures, on the other hand, are not strongly tied to a particular generative model: we can construct them for essentially any generative model by sampling from an approximate posterior distribution (see Section \ref{sec:inference}).
Bayesian inference based on the truncated likelihood has been suggested before by \citet{bayarri1987bayesian} \footnote{we thank Ed George for bringing this reference to our attention. } who used it in situations where the statistician has access to truncated data in the first place;
for example, inference on the distribution of heights in a certain period in the past might have to be based on historical records of the heights of members in the army, for which there was a minimum required height \citep{wachter1982estimating}.
Modeling selection-adjusted inference as a truncation problem, \citet{yekutieli2012adjusted} used the truncated likelihood along with a prior to give Bayesian inference on selected parameters in the sequence model.
The situation considered in the current paper, similarly to the focus in \citet{yekutieli2012adjusted}, is where truncation is imposed by the statistician herself when adaptively choosing the target of inference.
When providing adjusted Bayesian inference, we emphasize the advantages of using a {\it randomized} response for selection, as proposed by \citet{tian2015selective}.
Randomizing before applying selection may seem odd because it can (and often does) result in selecting a different set of variables, but so does the commonly used practice of sample splitting (in fact, if the split is random, then, as pointed out in \citet{tian2015selective}, sample splitting is an example of using a randomized response).
On the up side, randomization is a way to ensure that more information is reserved for the post-selection stage.
This usually results in improved accuracy of post-selection inference procedures, for example shorter credible intervals.
\smallskip
The rest of the paper is organized as follows.
In Section \ref{sec:framework} we describe a framework for Bayesian selection-adjusted inference in the linear model.
We motivate the use of the proposed framework and contrast it with the standard Bayesian variable-selection framework in Section \ref{sec:motivation}.
Section \ref{sec:adj-factor}, the centerpiece of this paper, proposes an approximation to the selection-adjusted likelihood, which we use in Section \ref{sec:inference} to provide adjusted approximate inference for selected parameters in a simulation example.
Section \ref{sec:estimation} is devoted to point estimation, where we study the approximate maximum a-posteriori (MAP) estimator.
We provide some theoretical support for using the proposed approximation in Section \ref{sec:asymptotic} and prove consistency guarantees of the selective MLE, obtained by maximizing the pseudo selection-adjusted likelihood, when selection is based on a randomized response.
The supplement includes an analysis of real data, where we also compare our methods to existing frequentist methods for post-selection inference in the linear model and demonstration of a few useful applications of the pseudo truncated likelihood in frequentist inference.
\section{Bayesian selection-adjusted inference in the linear model} \label{sec:framework}
This section presents a framework for Bayesian adjusted inference after variable selection in the linear model.
\citet{yekutieli2012adjusted} described a general framework for Bayesian adjusted inference, however his framework is aimed more at large-scale inference problems; the case of model selection calls for additional definitions.
The main conceptual difference is that, in the model selection problem, any parameter selected depends (by definition of a parameter) on the generative model, and, as proposed in \citet{fithian2014optimal}, we want to allow the analyst to choose also the generative model after observing the data.
The generative model and the truncating event together determine the truncated likelihood.
We assume that we observe a matrix $X=[X_1\cdots X_p]\in \mathbb{R}^{n\times p}$ and a random $p$-dimensional vector
$$
Y|X\sim F.
$$
$X$ is assumed to be a fixed design matrix, hence we will often write the above just as $Y\sim F$.
Informally, selective inference in the linear model consists of two stages:
\begin{enumerate}[1.]
\item {\bf Selection}. Choose a subset $\widehat{E}(Y) = E\subseteq \{1,...,p\}$ as a function of the data $(X,Y)$
\item { \bf Inference}. Provide inference for
\begin{equation} \label{eq:sel-param}
\begin{aligned}
\beta^E = \operatornamewithlimits{argmin}_{b^E\in \mathbb{R}^{|E|}} \| \mathbb{E}(Y) -X_E b^E \|_2^2 = (X_E^TX_E)^{-1}X_E^T \mathbb{E}(Y)
\end{aligned}
\end{equation}
where $X_E=[X_j: j\in E] \in \mathbb{R}^{n\times |E|}$ is the matrix consisting of only the columns of $X$ with indices in $E$, and is assumed to have full rank.
\end{enumerate}
The target of inference, $\beta^E$, is sometimes referred to as the {\it population} least-squares vector corresponding to the sub-model $E$, and, in words, is the coefficient vector of the best approximation to the mean of $Y$ in the linear space spanned by the columns of $X_E$.
To be formally defined as a parameter, a generative model needs to be specified for $Y$, that is, a family $\mathcal{F}$ of distributions to which the unknown distribution $F$ is assumed to belong.
In contrast to the traditional (non-selective) statistical framework, we allow the statistician to choose a generative model {\it after} she looks at the data, in the same way she chooses the {\it target} of inference only after looking at the data.
The minimum assumptions we make for the generative model are that
\begin{equation} \label{eq:sat-model}
Y\sim N_n(\mu, \Sigma)
\end{equation}
with $\mu \in \mathbb{R}^n$ unknown and $\Sigma$ known.
If the statistician chooses to proceed with \eqref{eq:sat-model} as the generative model, we will say that inference is given under the {\it saturated model}, and in this case
$$
\beta^E = (X^TX)^{-1}X^T \mu.
$$
More generally, we allow the statistician to restrict $\mu$ in \eqref{eq:sat-model} to a linear subspace of $\mathbb{R}^n$, in which case the generative model is the linear model
\begin{equation} \label{eq:lin-model}
Y\sim N_n(X^*\beta^*, \Sigma)
\end{equation}
where $X^*$ may include columns of $X$.
For example, one may take $X^* = X_E$, in which case
$$
\beta^E = (X_E^TX_E)^{-1}X_E^T X_E\beta^* = \beta^*
$$
and we will say that inference is given under the {\it selected model}.
But the generative model does not have to agree with the selected model (or the saturated model): one may take $X^* = X_{E'}$ for $E'\neq E$, including the choice $E' = \{1,...,p\}$, in which case the saturated model obtains; or even include as columns of $X^*$ predictors that were not originally recorded in $X$ but now -- after selecting $E$ -- seem relevant.
In general, then, inference is given for
$$
\beta^E = (X_E^T X_E)^{-1} X_E^T X^*\beta^*.
$$
It is important that the specification of $X^*$ depends on $Y$ only through $E$.
So, for example, it is not allowed to run Lasso again, now for $X$ augmented with some other variables that were not originally recorded in $X$, and let $X^*$ depend on the output.
Indeed, this may lead to bias that is not accounted for.
We remark that allowing to specify the generative model ($X^*$) after obtaining $E$ may reduce the difficulty of specifying a prior for the problem, because only a prior for $\beta^*$ is required rather than a prior on all possible sub-models.
Indeed, in some cases domain knowledge could be used to elicit a prior for the parameters of the generative model.
The question of what generative model to use is, unavoidably, somewhat arbitrary, and the analyst has to make some choice.
It is worth emphasizing, however, that the choice of the generative model can have important consequences not only on the interpretation of the selected parameter, but also on the properties of statistical procedures and on their implementation, as explained in \citet{fithian2014optimal}.
For example, besides power considerations, there is an essential difference between the saturated model, $X^*=I_n$, and the other cases, $\text{rank}(X^*)<n$, in terms of the computational difficulty in constructing optimal (UMPU) tests;
see Section \ref{supp:umpu} of the Supplementary Material.
Throughout the paper we will restrict attention to selection rules\footnote{For some popular selection rules like the Lasso, forward stepwise etc., we condition additionally on the signs of the active coefficients to obtain polyhedral regions} which partition the sample space into convex polyhedral sets,
\begin{equation}\label{eq:poly}
\{ y: A_E y\leq b_E \}
\end{equation}
for corresponding $A_E$ and $b_E$ associated with $E$.
In previous works \citep[][among others]{lee2013exact,taylor2014exact}, it was shown that for various popular `automated' selection algorithms, the set $\{y:\widehat{E}(y)=E\}$ can indeed be written in the form \eqref{eq:poly}.
Inference for $\beta^E$ is based on the {\it selection-adjusted} likelihood
\begin{equation}\label{eq:lik-adj-lin}
\begin{aligned}
f_S(y|\beta^*)
= \frac{f(y|\beta^*)}{\mathbb{P}(A_EY\leq b_E|\beta^*)}I(A_Ey\leq b_E),
\end{aligned}
\end{equation}
where
$$
f(y|\beta^*) = (2\pi)^{-n/2}|\Sigma|^{-1/2} \exp\big(-\frac{1}{2}(y-X^*\beta^*)^T\Sigma^{-1}(y-X^*\beta^*)\big).
$$
\noindent To obtain a Bayesian model we incorporate a prior
$$
\beta^* \sim \pi(\beta^*)
$$
into the model given by \eqref{eq:lik-adj-lin}.
Formally, then, $(\beta^*,Y)$ have a {\it selection-adjusted} joint distribution given by
\begin{equation}\label{eq:joint-adj-lin}
\begin{aligned}
f_S(\beta^*, y) = \pi(\beta^*) f_S(y|\beta^*) = \pi(\beta^*) \frac{f(y|\beta^*)}{\mathbb{P}(A_EY\leq b_E|\beta^*)} I(A_Ey\leq b_E).
\end{aligned}
\end{equation}
\noindent The framework for adjusted Bayesian inference after model selection is no different than an ordinary Bayesian framework, where the joint distribution of the parameter and the data is given by \eqref{eq:joint-adj-lin}.
In particular, if $L(a,\beta^*)$ is a loss function and $\delta_S(y)$ is a rule that maps points in the restricted sample space $\{y: A_E y\leq b_E\}$ to actions, then the adjusted Bayes risk of $\delta$ w.r.t. $\pi$ is
\begin{equation*}
r_S(\pi,\delta) = \iint L(\delta_S(y),\beta^*) f_S(\beta^*,y) \,dy\,d\beta^*.
\end{equation*}
Correspondingly, the Bayes rule minimizes the selection-adjusted posterior expected loss,
\begin{equation*}
\rho_S(\pi,\delta_S,y) = \int L(\delta_S(y),\beta^*) \pi_S(\beta^*|y) \,d\beta^*
\end{equation*}
where
\begin{equation} \label{eq:post-adj-lin}
\pi_S (\beta^*|y) = \frac{f_S(\beta^*,y)}{\int f_S(\beta^*,y)\ d\beta^*} \propto \pi(\beta^*) f_S(y|\beta^*)
\end{equation}
is the {selection-adjusted} posterior distribution.
Note that this can be easily adapted for inference on a function $t(\beta^*)$ by taking $L(a,\beta^*) = L_t(a,t(\beta^*))$ where $L_t$ is a loss function defined on an action set and state set chosen for $t(\beta^*)$.
This is just a formal way of saying that the posterior distribution of $\beta^*$ determines the optimal procedures for any function of $\beta^*$; specifically, this is true for $\beta^E$, the function of $\beta^*$ which inference is required for.
In later sections we base post-model selection inference on the (approximate) selection-adjusted posterior distribution \eqref{eq:post-adj-lin} of $\beta^*$, without necessarily committing to a formal decision theoretic framework.
\section{Motivation and comparison with standard Bayesian model uncertainty framework} \label{sec:motivation}
Our aim in this section is to advocate the use of a truncated likelihood in place of the usual likelihood when providing Bayesian inference after model selection.
Following \citet{yekutieli2012adjusted}, we model Bayesian selection-adjusted inference as a truncation problem.
Namely, the fact that inference for a specific parameter is given only if an associated event in the observation space occurs, elicits a truncated likelihood for this parameter.
As shown by \citet{yekutieli2012adjusted}, the fact that a truncated likelihood is involved, does not immediately imply that selection should alter posterior inference.
In fact, the common perception is that a Bayesian is indifferent to selection: the usual formalism suggests that $\pi(\theta|y, Y\in S) = \pi(\theta|y)$ for $y\in S$, hence no adjustment for selection seems necessary when conditioning on $y$.
\citet{yekutieli2012adjusted} demonstrated that this argument is valid when truncation is applied to the pair $(\theta, Y)\sim \pi(\theta)f(y|\theta)$, but not when truncation is applied to $Y$ conditionally on $\theta\sim \pi$.
To illustrate the difference between the two situations, consider the following example, inspired by \citet{mandel2007statistical}.
\begin{example}\label{ex:coin}
There is an infinite number of bags of coins with different probabilities for a Head.
The composition of each of the bags is the same, and such that if a coin is randomly selected, then the probability of landing on a Head has distribution $\pi$.
We are interested in constructing a posterior interval for $P$, the probability of Head, in the following two experiments.
\smallskip
\noindent
Situation \rom{1}: Draw a coin at random from the first bag, and toss it $10$ times.
If at least 3 Heads observed, report $Y = \text{\# of Heads observed}$.
If not, move on to the second bag, draw a coin at random, and toss it $n$ times.
If at least 3 Heads observed, report $Y = \text{\# of Heads observed}$.
If not, move on to the third bag, and so on until the first time that 3 Heads or more are observed.
Report $Y = \text{\# of Heads observed}$, and let $P$ be the (unobserved) probability of Head for the last coin drawn.
\smallskip
\noindent Situation \rom{2}: Draw a coin at random from the first bag.
Let $P$ be the (unobserved) probability of Head for the coin drawn.
Toss the coin $10$ times.
If at least 3 Heads observed, report $Y = \text{\# of Heads observed}$.
If not, toss the {\it same coin} again $10$ times.
If at least 3 Heads observed, report $Y = \text{\# of Heads observed}$.
If not, toss the {\it same coin} again $10$ times.
Keep going until the first time that 3 Heads or more are observed, and report $Y = \text{\# of Heads observed}$.
\end{example}
\citet{yekutieli2012adjusted} referred to Situation \rom{1} as the ``random" parameter case and to Situation \rom{2} as the ``fixed" parameter case.
In both cases, the number of Heads is reported only if at least 3 Heads are observed;
therefore, the likelihood in both cases is a truncated likelihood, say $f_S(y|p)$, where $f_S(y|p) = f(y|p) I_{\{y\geq 3\}}(y)/\mathbb{P}(Y\geq 3|p)$.
Yet selection matters only in Situation \rom{2}.
Indeed, in this case the selection-adjusted posterior is $\pi_S(p|y) \propto \pi(p) f_S(y|p)$,
whereas in Situation \rom{1}, the selection-adjusted posterior is
$$
\pi_S(p|y) \propto \pi(p|Y\geq 5) f_S(y|p) = \frac{\pi(p)f(Y\geq 5|p)}{\mathbb{P}(Y\geq 5)} \times \frac{f(y|p) I_{\{y\geq 5\}}(y)}{\mathbb{P}(Y\geq 5|p)} \propto \pi(p)f(y|p).
$$
Note that in Situation \rom{1} two sources contribute to ``shrinking" the likelihood: one is the prior $\pi(p)$, and the other is the term $1/\mathbb{P}(Y\geq 5|p)$, that gives a-priori preference to values of $p$ under which the selection event $\{Y\geq 5\}$ is less likely to occur.
As recommended by \citet{yekutieli2012adjusted}, choosing to model $\theta$ as a ``random" parameter or a ``fixed" parameter should depend on the context and be done case by case.
\citet{bayarri1987bayesian} proposed to analyze results reported in a scientific journal that has a policy of publishing only statistically significant findings, when essentially treating the underlying effect size of a published result as a ``fixed" parameter.
Indeed, their perspective reflects a model in which the pair $(\theta_i, Y_i)$ for a published result is generated by sampling $\theta_i\sim \pi$ once, and then sampling $Y_i|\theta_i\sim f(y_i|\theta_i)$ repeatedly, keeping the first $Y_i$ which passes a significance threshold.
We argue that the ``random" parameter model is more suitable for this example.
That is, to model a randomly selected reported result in the journal as being generated by repeatedly sampling a {\it pair} $(\theta_i, Y_i)\sim \pi(\theta_i)f(y_i|\theta_i)$ and keeping the first pair for which $Y_i$ passes the significance threshold.
This model is justifiable because it is natural to think of selection as acting on $\theta_i$ (as well as on $Y_i$, of course) in this case.
In other words, reported results tend to come with larger effect sizes.
Note that the resulting analyses will be substantially different, because the former entails appending the truncated likelihood $f_S(y|\theta) $ to $\pi(\theta)$, whereas the latter entails appending the unadjusted likelihood $f(y|\theta) $ to $\pi(\theta)$.
On the other hand, for inference after model (or variable) selection, we argue that treating the parameter $\beta^*$ as a ``fixed" parameter makes more sense.
From our perspective, the parameter $\beta^*$ and the data $Y$ are not {\it actually} jointly distributed; instead, only the data is really random, while the prior on $\beta$ is used to reflect prior beliefs on an ``unknown constant", to use the words of \citet{yekutieli2012adjusted}.
To try and convince the reader that treating $\beta^*$ as ``fixed" rather than ``random" parameter might be a good idea, we follow Example 2 of \citet{yekutieli2012adjusted} and compare the coverage of credible intervals constructed under a non-informative prior in the linear model, when treating the parameter as ``fixed" and when treating the parameter as ``random".
\begin{example}\label{ex:lasso}
We fix an $n\times p$ design matrix $X$ with $n=100, p=50$, and in each of 100 rounds draw a pair $(\beta, Y)$, where the components of $\beta \in \mathbb{R}^{50\times 1}$ are drawn i.i.d. from
\begin{equation} \label{eq:true-prior}
\pi(\beta) = 0.9\cdot \mathcal{N}(\beta; 0,0.1) + 0.1 \cdot \mathcal{N}(\beta; 0,3),
\end{equation}
a mixture of two zero-mean normal distributions, one with small variance 0.1 and the other with larger variance 3, and
$
Y| \beta \sim N_n(X\beta, I).
$
We run the Lasso with a fixed $\lambda = 1.56$ and denote by $E = \widehat{E}(Y) \subseteq \{1,...,50\}$ the set corresponding to nonzero Lasso estimates.
We then construct adjusted and unadjusted marginal posterior intervals for the components of $\beta^{E} = (X_{E}^T X_{E})^{-1}X_{E}^T X \beta$ assuming a {\it noninformative} prior $\pi(\beta)\propto 1$ instead of the true prior, which is regarded as unknown.
The first interval is constructed using the unadjusted Normal likelihood for $\beta$.
The second interval is constructed using the (approximate) selection-adjusted likelihood and taking $X^* = X_{E}$,
\begin{equation}\label{eq:lik-adj-ex2}
f_S(y|\beta) \propto \frac{\exp\left( -\frac{1}{2}\|y - X_{E}\beta\|_2^2 \right)}{\tilde{\mathbb{P}}(\widehat{E}(Y) = E|\beta)} I_{\{y:\widehat{E}(y) = E\}}(y).
\end{equation}
We also construct randomized selection-adjusted intervals and data-carved selection-adjusted intervals, see Section \ref{sec:inference} for details.
Above, $\tilde{\mathbb{P}}(\widehat{E}(Y) = E|\beta)$ is our barrier approximation to the exact selection-adjusted likelihood, discussed in Section \ref{sec:adj-factor}.
Because the posterior distribution of $\beta^{E}$ under the noninformative prior is $N_{|E|}(\hat{\beta}^{E}, (X_{E}^T X_{E})^{-1})$, where $\hat{\beta}^{E} = (X_{E}^T X_{E})^{-1}X_{E}^T Y$ is the LS estimate for $\beta^{E}$, the equal-tails 0.95 unadjusted marginal posterior interval for $\beta^{E}_j$ is $\hat{\beta}^{E}_j \pm \sigma_j({E}) \cdot z_{1-0.05/2}$ where $\sigma^2_j(E) = [(X_{E}^T X_{E})^{-1}]_{jj}$.
The selection-adjusted marginal credible intervals are computed using posterior sampling, as described in Section \ref{sec:inference}.
\smallskip
Without selection the unadjusted intervals for $\beta_j, j=1,...,50$ are expected to have the property that the (Bayesian) FCR, the expectation of the proportion of noncovering constructed intervals, is controlled at level $0.95$, because these are valid frequentist marginal intervals.
Indeed, with no selection (i.e., when constructing intervals for all $50$ components of $\beta$), the unadjusted intervals had observed FCR of 0.951.
But with selection applied as described above, the observed FCR was only 0.75.
On the other hand, the selection-adjusted posterior intervals had observed FCR of 0.85 -- a considerable improvement over the unadjusted intervals -- despite using a misspecified likelihood (because we computed them using $X^* = X_{E}$ in the generative model, while the true generative model has $X^* = X$) and an approximation instead of the exact term in the denominator of \eqref{eq:lik-adj-ex2}.
With randomization (see Section \ref{subsec:randomization}) the observed FCR increases to 0.94, and with data-carving (see Section \ref{subsec:data-carving}) the observed FCR is 0.942.
The improved robustness of the various selection-adjusted intervals implies that these methods tend to have better frequentist coverage than the unadjusted intervals.
This example also demonstrates the potential benefits of randomization and data-carving, ensuring better frequentist properties of the selection-adjusted intervals.
\end{example}
\section{Approximations to the truncated likelihood} \label{sec:adj-factor}
From now on, let $\|.\|$ denote the $\ell_2$ norm, unless specified otherwise.
To be able to take advantage of Bayesian computational methods, an immediate challenge that presents itself
is to evaluate the selection-adjusted likelihood \eqref{eq:lik-adj-lin} as a function of $\beta^*$.
The difficulty is in computing the {\it adjustment factor}
\begin{equation} \label{eq:adj-factor}
\mathbb{P} (A_E Y\leq b_E\lvert \beta^*)
\end{equation}
as a function of $\beta^*$ for given matrix $A_E$ and vector $b_E$.
Since \eqref{eq:adj-factor} in general has no closed form, we will propose an approximation of the adjustment factor that is computationally tractable while exhibiting similar properties as the exact expression.
We will assume from now on, unless otherwise noted, that $\Sigma=\sigma^2I_n$.
Extending our methods to the case of general (known) covariance is simple, especially since for \eqref{eq:adj-factor} we can assume $\text{cov}(Y) = I_n$ without loss of generality;
indeed,
$$
\mathbb{P}(AY\leq b\lvert \beta^*)= \mathbb{P}(\tilde{A} \tilde{Y}\leq b|\beta^*)
$$
with $\tilde{A}=A\Sigma^{1/2}$, $\tilde{Y}=\Sigma^{-1/2}Y\sim N_n(\Sigma^{-1/2}X^*\beta^*,I_n)$.
\medskip
Before delving into the details, we would like to convince the reader that evaluating the adjustment factor is essential.
The approach relied upon in recent frequentist work on exact post-selection inference, was able to bypass handling the adjustment factor directly.
We explain why this technique is not suitable in our case.
Consider inference under the saturated model (i.e., $X^*=I_n; \; \beta^* \in \mathbb{R}^n$), $Y\sim N_n(\beta^*,I_n)$, and suppose that we want to test a null hypothesis regarding a one-dimensional projection of $\beta^*$, say $\nu = \eta^T\beta^*$, conditional on $Y\in \{y:Ay\leq b\}$.
Decomposing
$$Y = (\eta^TY)\cfrac{\eta}{\|\eta\|^2} + Z$$
where $Z = (I-P_{\eta})Y$ is the projection of $Y$ orthogonally to $\eta$,
a crucial observation is that the conditional distribution of $\eta^T Y$ given that $Y\in \{y:Ay\leq b\}$ {\it and} $Z=z$, is just the distribution of $\eta^T Y$ truncated to the interval $[\gamma^-(z),\gamma^+(z)]$ for known functions $\gamma^+,\gamma^-$ associated with $A,b$.
Importantly, this distribution depends only on the scalar parameter of interest, $\nu = \eta^T\beta^*$.
This result is key to construction of frequentist tests (and, by duality, of confidence intervals) which are valid conditionally, and hence also unconditionally, on $Z$.
In fact, \citet{fithian2014optimal} show that under the saturated model, Uniformly Most Powerful Unbiased (UMPU) tests for testing a two-sided hypothesis are of this form; if instead of a saturated, the model used for inference has $X^*$ with $\text{rank}(X^*)<n$, constructing optimal test again requires to handle terms like \eqref{eq:adj-factor}; this is generally the case when inference under the {\it selected} model is of interest (in Section \ref{supp:umpu} of the supplement we propose to employ our methods for this task).
\smallskip
The situation is therefore congenial when performing selective hypothesis testing.
But it is more complicated when general likelihood-based inference is considered, estimation being a good example.
To see that the conditioning to eliminate nuisance parameters does not lead to a reduction in the Bayesian setting, we may write the selection-adjusted likelihood as
$$
\begin{aligned}
f_S(\eta^T y,z|\beta^*) = f_S(\eta^Ty|z,\beta^*)f_S(z|\beta^*) = g_{\nu, \sigma^2}^{[\gamma^-(z),\gamma^+(z)]}(\eta^Ty) f_S(z|\beta^*).
\end{aligned}
$$
where $g_{\nu, \sigma^2}^{[\gamma^-(z),\gamma^+(z)]}$ is the density of a $N(\nu, \sigma^2)$ variable truncated to $[\gamma^-(z),\gamma^+(z)]$.
For fixed $z$, if $f_S(z|\beta^*)$, the conditional density of a lower dimensional projection of $Y$, did not depend on $\nu$, then computing the posterior density of $\nu$ would conveniently reduce to computing the posterior under a univariate truncated normal likelihood, a case that was studied extensively by \citet{yekutieli2012adjusted}.
Unfortunately, $f_S(z|\beta^*)$ generally {\it does} depend on $\nu$, and so conditioning on $z$ is not particularly helpful.
\smallskip
Now that we have motivated the need for it, we turn to propose a workable approximation for the adjustment factor.
Note that at this point the discussion revolves around the selection-adjusted likelihood only, therefore any of the approximations that we will suggest for the truncated likelihood can be used for frequentist instead of Bayesian inference.
\subsection{A first attempt} \label{sec:chernoff}
The quantity of interest is the probability \eqref{eq:adj-factor} under $Y\sim N_n(X^*\beta^*,\sigma^2I)$.
We already noted that the exact expression is in general a complicated function of $\beta^*$; still, it is clear that this function should take on larger values when the mean $X^*\beta^*$ lies closer to the polytope ${\mathcal{K}} = \{y:A_E y\leq b_E\}$.
It would therefore be sensible to consider the approximation
\begin{equation} \label{eq:chernoff-prob}
\mathbb{P} (A_E y\leq b_E\lvert \beta^*) \approx \exp\left(-\dfrac{\lVert{X^*\beta^*-P_{{\mathcal{K}}}(X^*\beta^*)}\lVert^{2}}{2\sigma^2}\right)
\end{equation}
where $P_{{\mathcal{K}}}(X^*\beta^*)$ is the projection of $X^*\beta^*$ onto ${\mathcal{K}}$.
Indeed, the following theorem shows that the right-hand side of \eqref{eq:chernoff-prob} can be derived as an upper bound for the exact term by generalizing a Chernoff bound.
In fact, the right hand side of \eqref{eq:chernoff-prob} is a large deviations approximation to the left hand side.
We discuss this in detail in Section \ref{sec:asymptotic}.
\begin{theorem}[Chernoff approximation] \label{thm:chernoff}
Let $Y\sim N_n(X^*\beta^*,\sigma^2 I_n)$ and let ${\mathcal{K}} \subseteq \mathbb{R}^n$ be convex and compact.
Then
$$\mathbb{P}(Y\in {\mathcal{K}} \lvert \beta^*)\leq \exp\left(-\dfrac{\lVert{X^*\beta^*-{\mathcal{P}}_{{\mathcal{K}}}(X^*\beta^*)}\lVert^{2}}{2\sigma^2}\right)$$
where
$${\mathcal{P}}_{{\mathcal{K}}}(\mu) = \inf\{ \|\mu-z\|^2: z\in {\mathcal{K}} \}$$
is the projection of mean vector $\mu= X^*\beta^*$ onto ${\mathcal{K}}$.
\end{theorem}
\begin{remark}[Relaxing compactness]
The polyhedral set corresponding to many popular variable selection rules (as in our examples) is not compact.
While compactness is required for our proof, one should be able to remove this condition by considering, for a fixed $\pi$, a sufficiently large compact, convex subset of ${\mathcal{K}}$ that would work for all $\beta^*$ in a bounded set of probability close to 1 under $\pi$.
In any case, the purpose of Theorem \ref{thm:chernoff} is only to motivate the proposed approximation.
\end{remark}
\begin{remark}
While the theorem implies that the Chernoff approximation is sensible for general convex sets, the computation of the approximation is less burdensome when the set ${\mathcal{K}}$ is a polyhedral set: finding ${\mathcal{P}}_{{\mathcal{K}}}(\mu)$ is then a convex optimization problem, and there are available tools to solve it.
Specifically, in the appendix we describe how to employ an algorithm based on ADMM \citep{boyd2011distributed} to numerically solve the optimization problem.
\end{remark}
\subsection{Approximation with a barrier function} \label{sec:barrier}
While the Chernoff approximation \eqref{eq:chernoff-prob} exhibits desirable properties, it may be too crude for certain purposes.
Indeed, it incurs no penalty when $X^*\beta^*$ is inside the selection region, even though the exact probability \eqref{eq:adj-factor} is not constant on $\{\beta^*: X^*\beta^* \in {\mathcal{K}}\}$.
In Section \ref{sec:estimation} we discuss point estimation and show that using the approximation \eqref{eq:chernoff-prob} produces no shrinkage, while we {\it do} expect the point estimates to have the property of shrinking near the selection boundary.
This motivates looking for a refined approximation for the adjustment factor.
In general, consider an approximation of the form
\begin{equation} \label{eq:adj-factor-approx}
\mathbb{P} (A_E y\leq b_E\lvert \beta^*) \approx \exp( -h_{\sigma}(X^*\beta^*) ).
\end{equation}
Then the approximation in \eqref{eq:chernoff-prob} has
\begin{equation}\label{eq:chernoff}
h_{\sigma}(X^*\beta^*) = \frac{1}{2\sigma^2} \lVert{X^*\beta^*-P_{{\mathcal{K}}}(X^*\beta^*)}\lVert^{2} =
\inf_{z\in {\mathcal{K}}} \frac{1}{2\sigma^2}\| X^*\beta^* - z \|^2.
\end{equation}
Note that this can be rewritten as
\begin{equation} \label{eq:chernoff-alt}
\inf_{z\in \mathbb{R}^n} \left\{ \frac{1}{2\sigma^2} \| X^*\beta^* - z \|^2 + I_{\mathcal{K}}(z) \right\}
\end{equation}
where
\begin{equation}\label{eq:indicator}
I_{\mathcal{K}}(z) = \begin{cases}
0 & z \in {\mathcal{K}} \\
\infty & \text{otherwise}
\end{cases}
\end{equation}
is an indicator function for the set ${\mathcal{K}}$.
More generally, consider
\begin{equation} \label{eq:neg-lok-prob-w}
h_{\sigma}(X^*\beta^*) = \inf_{z\in \mathbb{R}^n} \left\{ \frac{1}{2\sigma^2} \| X^*\beta^* - z \|^2 + w_{\sigma}(z) \right\}
\end{equation}
where $w_{\sigma}(z)$ is a non-negative function that should, informally, reflect a preference for values of $z$ farther away from the boundary and {inside} the selection region; in other words, $w_{\sigma}(z)$ should take on smaller values for such $z$.
Throughout, when $\sigma^2=1$, we write simply $w(z)$ instead of $w_{1}(z)$ and $h(z)$ instead of $h_1(z)$.
Refining the choice in \eqref{eq:indicator}, we propose to substitute for $w_{\sigma}(z)$ a barrier function, \\
\begin{equation} \label{eq:barrier}
w^{\text{bar}}_{\sigma}(z) =
\begin{cases}
\sum_{i=1}^{m}\log\left(1+\dfrac{\sigma}{b_i-a_i^T z}\right) & z \in {\mathcal{K}} \\
\infty & \text{otherwise}
\end{cases}.
\end{equation}
Correspondingly, \eqref{eq:adj-factor} will be approximated as
$$
\exp( -h_{\sigma}^{\text{bar}}(X^*\beta^*))
$$
with
\begin{equation}\label{eq:adj-factor-barrier}
h_{\sigma}^{\text{bar}}(\beta^*) = \inf_{z\in \mathbb{R}^n} \left\{\frac{1}{2\sigma^2} \| X^*\beta^* - z \|^2 + w^{\text{bar}}_{\sigma}(z) \right\}.
\end{equation}
Note that as $z\in {\mathcal{K}}$ approaches the boundary of the selection region,
$\sum_{i=1}^{m}\log\left(1+\sigma/(b_i-a_i^T z)\right)$
will tend to $\infty$, which is typical of a barrier function.
Hence, for fixed $X^*\beta^*$, \eqref{eq:adj-factor-barrier} encourages $z$ to lie deeper inside the selection region, where $b_i-a_i^T z$ take on larger values.
Another advantage of using \eqref{eq:adj-factor-barrier} instead of \eqref{eq:chernoff-alt} is that the optimization problem is now not constrained.
The left panel of Figure \ref{fig:univar} displays the two approximations to the adjustment factor as well as the exact expression (using Mills' approximation for large $|\mu|$) for the truncated univariate example.
It can be seen that the Barrier approximation is more accurate than the Chernoff approximation.
As we will see in Section \ref{sec:estimation}, the modification above will also have a desirable shrinking effect on our point estimates. Corollary \ref{LDP:approximate:selective:posterior} in Section \ref{sec:asymptotic} says that the barrier modification also yields a large deviation approximation to the adjusted likelihood as long as the barrier function (with scale $1/n$) converges pointwise to the discrete indicator penalty $I_{\mathcal{K}}$.
\begin{figure}[!htb]
\centering
\begin{minipage}[t]{0.45\textwidth}
\includegraphics[scale=.45]{Weight_chernoff_barrier_uni.png}
\end{minipage}
\quad
\begin{minipage}[t]{0.45\textwidth}
\includegraphics[scale=.45]{MLE_true_approx_rand.png}
\end{minipage}
\caption{
{LEFT:
Approximations to $\log \mathbb{P}(A_E Y\leq b_E|\beta^*)$ in the univariate case, $Y\sim N(\beta^*,1)$.
Here $\{A_E Y\leq b_E\}=\{Y>0\}$.
The Barrier approximation is closer to the exact function.
}
\newline
{RIGHT:
Selective MLE with Barrier approximation for the univariate example, $Y\sim N(\beta^*,1)$ truncated to $\{y>0\}$.
The solid blue line is the approximate selective MLE under randomized response, $Y^* = Y + W$ truncated to $\{Y^*>0\}$ (note that this estimator is defined for $y\in \mathbb{R}$).
All other estimates are for the standard (non-randomized) setting.
The straight line is the identity, corresponding to the unadjusted estimate. }
}
\label{fig:univar}
\end{figure}
\section{Inference based on the selection-adjusted posterior} \label{sec:inference}
Equipped with a tractable approximation to the adjustment factor, we can now employ generic Bayesian computational methods to give selection-adjusted Bayesian inference for the parameters of the model.
In general, we will use a diffuse prior for $\beta^*$ and provide selection-adjusted inference for scalar functions of the parameter, $t(\beta^*),\ t:\mathbb{R}^k \to \mathbb{R}$.
An example is $\beta^E_j = e_j^T (X_E^T X_E)^{-1} X_E^T X^*\beta^*$.
When providing inference based on draws from the (approximate) selection-adjusted posterior, we use the selected model, $X^* = X_E$, as the generative model.
In principle, if the general model $y \sim N(X^*\beta^*, \sigma^2I)$ is preferred, one would draw posterior samples of $\beta^*$ and then compute the projection $\beta^E = (X_E^T X_E)^{-1}X_E^T X^*\beta^*$, for each sampled value of $\beta^*$.
We use a Langevin random walk to obtain a sample of size $n_S$ from the (approximate) selection-adjusted posterior,
with the $(K+1)$-th iterate drawn as
\[\beta^*_{(K+1)} = \beta^*_{(K)} +\gamma {\nabla}\log\pi_S(\beta^*_{(K)}\lvert y) + \sqrt{2\gamma}\epsilon_{(K)}\]
where $\epsilon_{(K)}$ for $K=1,2,\cdots, n_S$ are independent draws from a centered gaussian with unit variance and $\gamma$ is a suitably chosen step size. This sampler takes a noisy step along the gradient of the log-posterior, with no accept-reject step unlike the usual Metropolis Hastings (MH) algorithm.
Hence, at each draw of the sampler, we compute the gradient of the log selection-adjusted posterior
based on the optimizer in the optimization problem associated with \eqref{eq:adj-factor-barrier}, see \eqref{grad:log:posterior}.
Alternatively, we could employ a MH algorithm to compute $h_{\sigma}^{\text{bar}}(X^*\beta^*)$; that would require the value of optimization problem in \eqref{eq:adj-factor-barrier} for computing the acceptance ratio each time.
When approximating the adjustment factor, \eqref{eq:chernoff} could be used instead of \eqref{eq:adj-factor-barrier};
however, the previous section suggests that the barrier approximation produces more accurate results.
\subsection{Inference employing Data-carving}\label{subsec:data-carving}
\citet{fithian2014optimal} proposed a general scheme, {\it data-carving}, to increase precision of post-selection inference.
In that scheme a small portion of the data is reserved for the inference stage, i.e., not used for selection; while the entire data is used for inference conditional on selection.
Thus, as opposed to data splitting, which ignores the hold-out portion of the data completely at the inference stage, data carving discards only the information in the hold-out portion that is captured by the random variable $1_{A_q}(Y)$.
\citet{fithian2014optimal} show that, as expected, selective tests based on data-carving can be substantially more powerful than selective tests based on data splitting.
Data carving can certainly be incorporated into our Bayesian framework.
Let the data be
$$
Y=
\begin{bmatrix}
Y^{(1)}\\
Y^{(2)}
\end{bmatrix}
\sim
N\left(
\begin{bmatrix}
X^{*(1)}\beta^*\\
X^{*(2)}\beta^*
\end{bmatrix},
\sigma^2
\begin{bmatrix}
I_{n_1} & 0\\
0 & I_{n_2}
\end{bmatrix}
\right).
$$
The selection event is $\{A_E^{(1)} Y^{(1)}\leq b_m\}$ where $A_E^{(1)} = A_E S,\ S = [I_{n_1} \ 0_{(n-n_1)\times n_1}]^T$.
Since the selection event is determined by $Y^{(1)}$ only, the selection-adjusted joint distribution of the parameter and the data is
\begin{equation}\label{eq:joint-carv}
f_S(\beta^*,y^{(1)},y^{(2)}) = \pi(\beta^*) f_S(y^{(1)}|\beta^*) f(y^{(2)}|\beta^*) .
\end{equation}
Note that, rearranging the terms on the right hand side and using Bayes rule, we have
$$
f_S(\beta^*,y^{(1)},y^{(2)}) \propto \pi(\beta^*|y^{(2)}) f_S(y^{(1)}|\beta^*).
$$
That is, with data carving, the posterior distribution can be interpreted as arising from updating first the distribution of $\beta^*$ based on the untruncated likelihood of $Y^{(2)}$ only, and using this as the updated prior for $\beta^*$ in a selective problem involving only $Y^{(1)}$.
\subsection{Inference employing randomization} \label{subsec:randomization}
Data-carving may reduce the length of confidence intervals because it guarantees that the information for the parameter in the hold-out portion, is not affected by selection.
Another general strategy for ensuring that enough information is reserved for inference, suggested by \citet{tian2015selective}, is to introduce user-generated randomness in the selection stage.
Indeed, randomization tends to smooth out the effect of truncation, which results in more {\it leftover information} (introduced in \citealt{fithian2014optimal}, this is the expectation of the conditional fisher information given the indicator $1_{A_q}(Y)$).
In fact, as they pointed out, data-splitting and data-carving that use random splitting, are particular schemes for selective inference with a randomized response.
\citet{tian2015selective} proposed a specific {\it Gaussian randomization} scheme, in which the observed data is still the vector $y$, but truncation is applied to $Y^* = Y + W$ where $W\sim N_n(0,\gamma^2I)$ and independent of $Y$.
If the selection region remains $\{y^*:A_Ey^*\leq b_E\}$, then the selection-adjusted joint distribution of $(\beta^*,y,w)$ is now
\begin{equation*}
f_S(\beta^*,y,w) = \pi(\beta^*) \frac{f(y,w|\beta^*)}{\mathbb{P}(A_E(Y+W)\leq b_E|\beta^*)} I(A_E(y+w)\leq b_E).
\end{equation*}
Integrating out $w$ gives the selection-adjusted joint distribution of $\beta^*$ and the observed data $y$ as
\begin{equation} \label{eq:joint-sel-rand}
\begin{aligned}
f_S(\beta^*,y) &= \pi(\beta^*) \frac{f(y|\beta^*)}{\mathbb{P}(A_E(Y+W)\leq b_E|\beta^*)} \mathbb{P}(A_E(y+W)\leq b_E) \\
&\propto \pi(\beta^*) \frac{f(y|\beta^*)}{\mathbb{P}(A_E(Y+W)\leq b_E|\beta^*)}.
\end{aligned}
\end{equation}
In the last step we used the fact that $W$ is independent of $Y$ and its distribution does not depend on $\beta^*$.
To provide adjusted Bayesian inference in this situation, we can proceed as in the non-randomized setup after we obtain an approximation for $\mathbb{P}(A_E(Y+W)\leq b_E|\beta^*)$;
note that this is equal to $\mathbb{P}(A_E\tilde{Y}\leq b_E|\beta^*)$ where $\tilde{Y} \stackrel{d}{=} Y+W\sim N_n(X^*\beta^*,(\sigma^2+\gamma^2)I)$, hence reduces to the computation of the adjustment factor when $\sigma^2$ is replaced by $\sigma^2+\gamma^2$.
\addtocounter{example}{-1}
\begin{example}[continued] \label{ex:lasso-cont}
We return to Example 2 of Section \ref{sec:motivation} and obtain credible intervals and point estimates for the components of $\hat{\beta}^{E}$.
Recall that $Y|\beta \sim N_n(X\beta, I)$ but we assume the selected model $Y|\beta^E \sim N_{|E|}(X_E\beta^E, I)$ when providing inference.
Also, a noninformative prior is used for $\beta^E$ instead of the true prior \eqref{eq:true-prior}.
Figure \ref{fig:inference} displays selection-adjusted as well as unadjusted credible intervals and point estimates for a particular realization $Y=y$: the left panel is for the non-randomized and the right panel is for the randomized setup.
In applying the randomized methods, selection is applied to a perturbed version of $y$, $Y^* = y + W$ with $W\sim N_n(0,\gamma^2I),\ \gamma^2 = 0.1$.
This resulted in selecting a different set $E$ (from the figure we can tell that $|E| = 6$ without randomization, and $|E| = 7$ with randomized response).
The adjusted intervals are considerably longer than the unadjusted intervals, and are usually not symmetric about the estimate (black filled circles in the figure).
The figure also shows that randomization can contribute significantly to shortening the credible intervals, which illustrates the gain in power with randomization.
The adjusted-MAP point estimates, discussed in the next section,
seem to shrink the estimates $\hat{\beta}^{E}_j$ toward zero in the left panel of the figure.
the adjusted mean point estimates and the equi-tailed adjusted credible intervals are computed using the barrier approximation and posterior sampling.
\begin{figure}[H]
\captionsetup{skip=-10pt}
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=1\linewidth,keepaspectratio]{non_randomized_inference_lasso_fcr.png}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=1\linewidth,keepaspectratio]{randomized_inference_lasso_fcr.png}
\end{subfigure}
\caption{Selection-adjusted credible intervals and point estimates for selection via Lasso with fixed $\lambda=1.56$.
The left panel is for the non-randomized setting and right panel corresponds to the randomized setting.
Vertical bars are the barrier approximate credible intervals based on samples from the posterior using a Langevin random walk.
Selection-adjusted point estimates -- approximate MLE and approximate posterior mean -- are marked on the bars.
}
\label{fig:inference}
\end{figure}
\end{example}
\section{Point estimates} \label{sec:estimation}
In this section we explore point estimates for functions of the parameters $\beta^*$ in the generative model.
We can obtain an approximation for the posterior mean estimate by sampling from the approximate posterior as suggested in the previous section.
We do implement this estimate when comparing different methods in Example \ref{ex:lasso-cont} above, but the focus in this section is on studying the {\it mode} of the (approximate) selection-adjusted posterior.
We refer to the latter as the selective MAP (or just the MAP) estimate.
A main reason for concentrating on this estimate is that, by addressing the MAP, we make the discussion applicable also to the selective MLE (maximum-likelihood estimator), by considering a constant prior.
We would like to point out that frequentist point estimation based on the full truncated likelihood (in particular, the MLE of $\beta^*$) is not addressed fully in the recent work on exact post-model selection inference, though there has been some work on pseudo-maximum likelihood estimators in \cite{reid2014post}, as well as earlier work on two stage estimators \cite{cohen1989two,zhong2008bias,zollner2007overcoming}.
Whereas implementation of the posterior mean estimate requires sampling from the approximate posterior, we will show that the problem of finding the approximate MAP is a convex optimization problem; this further adds to the appeal of the barrier approximation.
\smallskip
Recall that we proposed to approximate the selection-adjusted posterior in Section \ref{sec:adj-factor} by
\begin{equation} \label{eq:adj-post-approx}
\pi_S(\beta^*\lvert y) \propto \; \pi(\beta^*)\dfrac{\exp\{-\lVert y-X^*\beta^* \lVert^2/(2\sigma^2)\}}{\exp(-h_{\sigma}(X^*\beta^*))}
\end{equation}
where the function $h_{\sigma}(\mu)$ depends on $A_E,b_E$ (and $X^*,\sigma^2$) and has the general form
\begin{equation*}
h_{\sigma}(\mu) = \inf_{z\in \mathbb{R}^n} \left\{\frac{1}{2\sigma^2} \| \mu - z \|^2 + w_{\sigma}(z) \right\},
\end{equation*}
and where we considered two specific choices for $w_{\sigma}(z)$ in \eqref{eq:indicator} and \eqref{eq:barrier}.
Fixing a choice of $w_{\sigma}(z)$, we will refer to a maximizer of \eqref{eq:adj-post-approx} or, equivalently, a minimizer of
\begin{equation}\label{eq:neg-log-post-approx}
\left( \frac{1}{2\sigma^2}\|y-X^*\beta^*\|^2 -h_{\sigma}(X^*\beta^*) \right)- \log \pi(\beta^*),
\end{equation}
as an approximate MAP estimator.
In the remainder of this section we assume without loss of generality that $\sigma^2=1$, i.e., $Y\sim N_n(X^*\beta^*,I)$.
The following lemma asserts that finding an approximate MAP amounts to minimizing a convex objective.
\begin{lemma}[Convexity of MAP (MLE) problem with general approximation] \label{lem:map-convex}
Let $\pi(\beta^*)$ be a log-concave prior.
For a nonnegative function $w(z)$, finding the minimizer of \eqref{eq:neg-log-post-approx} in $\beta^*$ is a convex optimization problem.
\end{lemma}
\begin{proof}
With $\sigma^2=1$, minimizing \eqref{eq:neg-log-post-approx} is equivalent to minimizing
$$
\frac{1}{2} \|X^*\beta^*\|^2 - h(X^*\beta^*) - \beta^{*T}(X^{*T}y) - \log \pi(\beta^*).
$$
Now
$$
\begin{aligned}
\frac{1}{2} \|X^*\beta^*\|^2 - h(\beta^*)
&= \frac{1}{2} \|X^*\beta^*\|^2 - \inf_{z \in \mathbb{R}^n} \left\{\frac{1}{2} \|X^*\beta^*-z\|^2
+ w(z) \right\}\\
&= \sup_{z \in \mathbb{R}^n} \frac{1}{2}\|X^*\beta^*\|^2 - \frac{1}{2} \|X^*\beta^* - z\|^2_2 -w(z) \\
&= \sup_{z \in \mathbb{R}^n} z^T(X^*\beta^*) - \frac{1}{2} \|z\|^2-w(z) \\
&= \bar{w}^*(X^*\beta^*)
\end{aligned}
$$
where $\bar{w}^*$ is the convex conjugate of the function of
$
\bar{w}(z) = w(z) + \frac{1}{2} \|z\|^2.
$
Hence a MAP estimate minimizes
\begin{equation} \label{eq:map-obj}
\bar{w}^*(X^*\beta^*) - \beta^{*T}(X^{*T}y) - \log \pi(\beta^*)
\end{equation}
which is
convex.
\end{proof}
\begin{remark} We quickly note from the above calculations that the gradient of the logarithm of the approximate selection adjusted posterior with a prior $\pi$ on $\beta^*$ can be computed as
\begin{align}
{\nabla}\log \pi_S(\beta^*\lvert y) &= X^{*T}(y-{\nabla}\bar{w}^*(X^*\beta^*))+ {\nabla}\log \pi(\beta^*)\nonumber \\
&= X^{*T}(y-z^*(X^*\beta^*))+ {\nabla}\log \pi(\beta^*)\numberthis \label{grad:log:posterior}
\end{align}
where $z^*(\mu) =\operatornamewithlimits{argsup}_{z \in \mathbb{R}^n} z^T\mu - \frac{1}{2} \|z\|^2-w(z)$, the optimizer of the approximation. We use the above gradient of the log-posterior when drawing samples using a Langevin sampler.
\end{remark}
\medskip
\noindent Next we show that, at least under the saturated model, the minimizer of \eqref{eq:map-obj} has an explicit form.
\begin{theorem} \label{thm:map-explicit}
Under $Y \sim N_n(\beta^*, I)$, the approximate selective MAP satisfie
\begin{equation} \label{eq:map-est-eq}
\hat{\beta}^* = \nabla \bar{w}(y + \nabla \log \pi(\hat{\beta}^*))
\end{equation}
with $\bar{w}(z) = w(z) + \frac{1}{2} \|z\|^2_2$.
In particular, taking $\pi(\beta^*)\propto 1$, the selective MLE is given by
$$
\hat{\beta}^* = y + \nabla w(y)
$$
\end{theorem}
\begin{proof}
Substituting $X^* = I_n$ into \eqref{eq:map-obj} and taking derivative w.r.t. $\beta^*$, the selective MAP satisfies
$$
-y + \nabla \bar{w}^*(\hat{\beta}^*) - \nabla \log \pi(\hat{\beta}^*) = 0
$$
which implies that
$\hat{\beta}^* = \nabla \bar{w}^{*-1}(y + \nabla \log \pi(\hat{\beta}^*))
= \nabla \bar{w}(y + \nabla \log \pi(\hat{\beta}^*))$
where
$$
\nabla \bar{w}(z) = z + \nabla w(z).
$$
Now the selective MLE is the selective MAP with $\pi(\mu) = 1$, in which case $\log \pi(\hat{\mu}) = 0$ and the above reduces to
$$
\hat{\beta}^* = \nabla \bar{w}(y) = y + \nabla w(y).
$$
\end{proof}
To implement the MAP estimate we need to specify a function $w(z)$.
The following example shows that choosing $w(z)=I_{{\mathcal{K}}}(z)$, the first suggestion in Section \ref{sec:adj-factor}, produces estimates that lack an important feature that the exact MAP has.
After that, we will see that this undesirable situation is mitigated by replacing the indicator function with $w^{\text{bar}}(z)$.
\begin{example} \label{ex:univar-map}
$Y\sim N(\beta^*,1)$.
For the selection event $\{y:y>0\}$, the negative log of the selection-adjusted likelihood is
$
\frac{1}{2}(y-\beta^*)^2 + \log(1-\Phi(-\beta^*))
$
and the exact selective MLE can be easily seen to satisfy
$$
y = \hat{\beta}^* + \frac{\phi(\hat{\beta}^*)}{1-\Phi(-\hat{\beta}^*)}.
$$
The selective MLE exhibits shrinkage of the unadjusted MLE, $y$, for $y$ values close to the selection boundary; see the solid black line in the right panel of Figure \ref{fig:univar}.
For any choice of $w(z)$, the approximate selective MLE solves
\begin{equation}\label{eq:univar-map-obj}
\minimize_{\beta^*} \frac{1}{2 } \left({ \beta^*}^2 - \mu y \right) - h(\beta^*)
\end{equation}
with
$
h(\beta^*)=\inf_{z\in \mathbb{R}}\frac{1}{2} (\beta^*-z)^2+ w(z).
$
If $w(z)=I_{{\mathcal{K}}}(z)$, then by Theorem \ref{thm:map-explicit}, the approximate selective MLE is
$$\hat{\beta}^* = y + \nabla I_{{\mathcal{K}}}(y) = y,$$
noting that $y\in {\mathcal{K}}$.
Hence, for that choice of $w(z)$ the approximate MLE coincides with the unadjusted MLE
(In fact, the preceding argument can be extended to the general regression situation, observing that
the minimizer of the approximate selective likelihood restricted to $\{\beta^*: X^*\beta^*\in {\mathcal{K}}\}$, is the usual least squares estimate $\hat{\beta}^*_{LS}=(X^{*T} X^*)^{-1}X^{*T}y$, whenever $X^* {\hat{\beta}^*}_{LS}\in \text{int}({\mathcal{K}})$).
The barrier function defined in \eqref{eq:barrier} has the property that bigger penalty is applied to values $z\in {\mathcal{K}}$ closer to the selection boundary.
In our univariate case ($\sigma=1$) the barrier function is just $w^{\text{bar}}(z) = \log (1+ 1/z )$ for $z>0$ ($w^{\text{bar}}(z) =\infty$ otherwise).
By Theorem \ref{thm:map-explicit}, the corresponding approximate selective MAP is ($y>0$)
\begin{equation} \label{eq:univar-map-barrier}
\hat{\beta}^* = y + \nabla w_{\text{bar}}(y) = y - \frac{1}{y(y+1)},
\end{equation}
which indeed shrinks the unadjusted MLE towards zero.
The broken line in the right panel of Figure \ref{fig:univar} shows the estimator \eqref{eq:univar-map-barrier}, obtained with the barrier function.
We can see that it follows the true selective MLE closely, shrinking to $-\infty$ near the selection boundary while reverting to the unadjusted MLE as the observation lies deeper inside the selection region.
\end{example}
Theorem \ref{thm:map-explicit} applies to the saturated model.
In that case, the approximate selective MLE with $w(z)=w^{\text{bar}}(z)$ has a simple, closed form.
For the more general linear model we offer the following lemma, which is proved in the appendix.
\begin{lemma} \label{lem:map-explicit-gen}
Suppose that $Y\sim N_n(X^*\beta^*,I)$ and the selection event is of the form ${\mathcal{K}}= \{Ay\leq b\} = \{y: a_i^Ty-b_i\leq 0,\ i=1,...,m\}$. The selective MAP satisfies the estimating equation
\begin{equation*}
X^{*T}{\nabla} \bar{w}^*(X^*\hat{\beta}^*) = X^{*T} z^*(X^*\hat{\beta}^*)
= X^{*T} y +{\nabla} \log\pi(\hat{\beta}^*)
\end{equation*}
and the selective MLE satisfies
\begin{equation*}
{\nabla} \bar{w}^*(X^*\hat{\beta}^*)=z^*(X^*\hat{\beta}^*)=y+v_0
\end{equation*}
for $v_0\in {\mathcal{N}}(X^{*T} )$, the null space of $X^{*T}$, with
$$z^*(\mu)={\nabla} \bar{w}^{-1}(\mu)=\operatornamewithlimits{argsup}_{z \in \mathbb{R}^n} z^T \mu - \frac{1}{2} \|z\|^2-\sum_{j=1}^m \log (1+1/(b_i - a_i^Tz)).$$
\end{lemma}
\subsection{Selective MAP (MLE) with randomization} \label{subsec:mle-rand}
The MAP estimator for the randomized situation with additive, independent gaussian noise $W\sim N_n(0,\gamma^2 I)$ maximizes \eqref{eq:joint-sel-rand} with respect to $\beta^*$.
By the comment in the last paragraph of Section \ref{subsec:randomization}, the approximate MAP for any $w$ minimizes \eqref{eq:neg-log-post-approx} only with $h(X^*\beta^*)$ replaced by (remember that $\sigma^2=1$)
\begin{equation*}
\tilde{h}(X^*\beta^*) = \inf_{z\in \mathbb{R}^n} \left\{\frac{1}{2(1+\gamma^2)} \| X^*\beta^* - z \|^2 + w_{(1+\gamma^2)^{1/2}}(z) \right\}.
\end{equation*}
We can look for a maximizer by numerically optimizing the objective, relying on numerical methods to evaluate $\tilde{h}(X^*\beta^*)$ for a candidate $\beta^*$.
The optimization problem is no longer equivalent to \eqref{eq:map-obj} because the variances used in the (unadjusted) likelihood and in $\tilde{h}$ are not the same.
Still, with our approximation, finding a MAP (MLE) is again a convex problem when a log-concave prior is used.
Indeed, the negative of the logarithm of the approximate selection adjusted likelihood is
\begin{equation*}
\begin{aligned}
&-yX^*\beta^*+ \frac{\gamma^2}{2(1+\gamma^2)} \|X^*\beta^*\|^2_2 + \sup_{z \in \mathbb{R}^n} z^T\frac{X^*\beta^*}{1+\gamma^2} -\left\{\frac{\|z\|^2}{2(1+\gamma^2)}
+ w_{(1+\gamma^2)^{1/2}}(z) \right\} \\
&=-yX^*\beta^* + \frac{\gamma^2}{2(1+\gamma^2)} \|X^*\beta^*\|^2_2 + \bar{w}^*\left(\frac{X^*\beta^*}{1+\gamma^2}\right),
\end{aligned}
\end{equation*}
with
\begin{equation}
\label{bar:w:randomized}
\bar{w}(z) = \frac{\|z\|^2}{2(1+\gamma^2)}+ w_{(1+\gamma^2)^{1/2}}(z).
\end{equation}
Under the saturated model we can characterize the approximate MAP for the randomized setup by an estimating equation in the following lemma. The proof is seen trivially by taking the derivative of the logarithm of the selection-adjusted posterior.
\begin{lemma}[Approximate selective randomized MAP under the saturated model] \label{lem:map-rand-sat}
Suppose that $Y\sim N_n(\beta^*, I)$ and let $\pi(\beta^*)$ be a log-concave prior.
Let $\hat{\beta}^*$ be a minimizer of
$$
\frac{1}{2}\|y-\beta^* \|^2 -\tilde{h}(\beta^*) - \log \pi(\beta^*)
$$
with $\tilde{h}$ defined above.
Then
$$
\hat{\beta}^* = (1 + \gamma^2) \nabla \bar{w}\left( (1 + \gamma^2)\big(y + \nabla \log \pi(\hat{\beta}^*)\big) - \gamma^2\hat{\beta}^* \right)
$$
with $\bar{w}$ given by \eqref{bar:w:randomized}. In particular, the estimating equation for an approximate selective MLE is obtained by setting $\nabla \log \pi(\hat{\mu}) = 0$ above.
\end{lemma}
\addtocounter{example}{-1}
\begin{example}[continued] \label{ex:univar-map-rand}
$Y\sim (\beta^*,1)$ and $W\sim N(0,\gamma^2)$, independent.
Also $\mu\sim \pi$ where $\pi$ is log-concave.
The observation is $Y$ truncated to $\{Y+W > 0\}$.
We consider the approximate selective MLE, $\hat{\beta}^*$, corresponding to
$$w^{\text{bar}}_{\sqrt{1+\gamma^2}}(z) =
\begin{cases}
\log\left(1+\dfrac{\sqrt{1+\gamma^2}}{z}\right) & z \in {\mathcal{K}} \\
\infty & \text{otherwise}
\end{cases}.
$$
By Lemma \ref{lem:map-rand-sat}, this is given by
$$
\hat{\beta}^* = (1 + \gamma^2) \left( \frac{z}{1+\gamma^2} - \frac{\sqrt{1+\gamma^2}}{z(z+\sqrt{1+\gamma^2})} \right) \bigg|_{z = Z_{\text{obs}}} \text{ with } Z_{\text{obs}}=(1 + \gamma^2) Y -\gamma^2\hat{\beta}^*.$$
\medskip
\noindent The right panel of Figure \ref{fig:univar} shows the randomized approximate MLE versus the approximate MLE (and other estimators) without randomization.
Note that with randomization, $y$ takes on values between $\pm \infty$ on the selection event (because $y^*$, not $y$, is restricted to be positive).
Since the effect of truncation is attenuated by randomization, leaving more information for inference, shrinkage near the selective boundary is much more moderate than in the nonrandomized version.
\end{example}
Similarly to the relation between Theorem \ref{thm:map-explicit} and Lemma \ref{lem:map-explicit-gen}, there is an easy extension of \ref{lem:map-rand-sat} to the more general linear regression scenario in the randomized setting. The more general estimating equation for the randomized MAP $\hat{\beta}^*$ when $Y\sim N(X^*\beta^*,I_n)$ and $Y^* = Y + W$ where $W\sim N_n(0,\gamma^2 I_n)$ independent of $Y$ can be written as
\begin{equation*}
\label{est:equation:randomized}
\frac{1}{(1 + \gamma^2)}X^{*T} {\nabla} w^*\left(\dfrac{X^*\hat{\beta}^*}{(1 + \gamma^2)}\right)=X^{*T}y+{\nabla}\log \pi(\hat{\beta}^*)-\frac{\gamma^2}{1+\gamma^2}X^{*T}X^*\hat{\beta}^*
\end{equation*}
where $${\nabla} \bar{w}^*(\mu)=z^*(\mu)=\operatornamewithlimits{argsup}_{z\in \mathbb{R}^n} z^T\mu-\frac{\|z\|^2}{2(1+\gamma^2)}-w_{(1+\gamma^2)^{1/2}}(z).$$
\section{Asymptotic guarantees under approximation} \label{sec:asymptotic}
In this section we study asymptotic properties related to the barrier approximation to the adjustment factor, in an attempt to provide some theoretical support for using it.
Specifically, we show first that for i.i.d. variables (from an arbitrary distribution) the barrier approximation to the adjustment factor has the correct limiting exponential decay rate under so-called non-local alternatives.
The second main result of this section states that the randomized selective MLE of Section \ref{subsec:mle-rand}, which employs the barrier approximation, is a consistent estimator for the mean under non-local alternatives.
Other common estimators lack consistency properties under the nonrandomized selective law; the unadjusted sample mean is trivially inconsistent, and we show that the selective MLE is also inconsistent for non-local alternatives.
This again highlights the potential advantages in using randomization in selection.
The first result we state is a standard result in large deviations theory.
For $d=1$ the theorem below is known as Cramer's theorem, see \cite{dembo2010large}.
The result for higher dimensions uses similar arguments; except that the upper bound in that case requires a minimax result to interchange the supremum and infimum.
Although there is nothing novel in Theorem \ref{LDP:approximating:probability}, we provide a proof in the Appendix, which is adapted from \cite{dembo2010large}.
\begin{theorem} \label{LDP:approximating:probability}
Let $Y_1,Y_2,\cdots, Y_n \stackrel{i.i.d. }{\sim} F$ with support on $\mathbb{R}^{d}$, $d$ fixed such that
$F\in \mathcal{F}=\{ F : \mathbb{E}_{F}(Y)= \beta^* \in \mathbb{R}^d, \mathbb{E}_{F}(Y-\beta^*)^2= I_d \}.$ For any convex set $\mathcal{K}\subseteq \mathbb{R}$ when $d=1$, and for any convex and compact set $\mathcal{K}\subseteq \mathbb{R}^d$ when $d>1$, we have
\begin{equation}
\label{lim:sel:prob}
\lim_n \frac{1}{n}\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*) = -\inf_{x \in \mathcal{K}}\Lambda^*(x;\beta^*)
\end{equation}
where
$$
\Lambda^*(x;\beta^*) = \sup_{\eta} \eta^T x - \log\mathbb{E}[(\exp(\eta^T Y_1)\lvert \beta^*]
$$
is the convex conjugate of the log of the moment generating function $\Lambda(\eta;\beta^*) = \log\mathbb{E}[\exp(\eta^T Y_1)\lvert \beta^*]$ of random variable $Y_1$.
\end{theorem}
As a consequence we have the following corollary that says that the sequence of selection-adjusted posterior based on a suitable barrier approximation indeed converges to the corresponding (true) selection-adjusted posterior.
\begin{corollary}
\label{LDP:approximate:selective:posterior}
Let $Y_1, Y_2,\cdots, Y_n \sim N_d(\beta^*, I)$ be random $d$-dimensional Normal vectors.
Consider the selection event $\{\bar{Y}_n \in {\mathcal{K}}\}$ for a convex (and compact for $d>1$) set ${\mathcal{K}}\subseteq \mathbb{R}^d$.
Denote by
$$
\log\pi_S(\beta^*\lvert\bar{y}_n) =\log \pi(\beta^*) -n\|\bar{y}_n - \beta^*\|^2/2 -\log \mathbb{P}(\bar{Y}_n \in \mathcal{K}\lvert \beta^*)
$$
the corresponding (true) selection-adjusted posterior, and by
$$
\log\hat{\pi}_S(\beta^*\lvert \bar{y}_n) = \log \pi(\beta^*) -n\|\bar{y}_n - \beta^*\|^2/2 + n \inf_{x\in \mathbb{R}^d}\left\{\|x-\beta^*\|^2/2 +\frac{1}{n}w_{1/\sqrt{n}} (x)\right\}
$$
a corresponding {approximate} selection-adjusted posterior, where $w_{1/\sqrt{n}}(\cdot)$ is a (convex) barrier function associated with ${\mathcal{K}}$.
Then
\[\lim_n \frac{1}{n} \left\{\log\pi_S(\beta^*\lvert\bar{y}_n)- \log\hat{\pi}_S(\beta^*\lvert \bar{y}_n)\right\} \to 0 \text{ as } n \to \infty\]
for all $\beta^* \in \mathbb{R}^d$ whenever $\frac{1}{n}w_{1/\sqrt{n}}(\cdot)$ converges to $I_{\mathcal{K}}(\cdot)$ pointwise as $n \to \infty$.
\end{corollary}
\begin{remark} The results in this section extend to the case of general regression with random design,
\[(x_i, y_i) \stackrel{i.i.d.}{\sim} F,\; (x_i, y_i) \in \mathbb{R}^p\times \mathbb{R}\] with a fixed p.
In that case, the selection probability typically takes on the form
\[\mathbb{P}(A_E\sqrt{n}\bar{T}_n \geq b_E)= \mathbb{P}(A_E\sqrt{n}(\bar{T}_n-\beta^*)\geq b_E -\sqrt{n} A_E \beta^*)\]
for a statistic $\bar{T}_n= \bar{T}_n(X,Y)\in \mathbb{R}^p$; a polytope defined by $A_E \in \mathbb{R}^{m\times p}, b_E\in \mathbb{R}^{m}$; and a population parameter $\beta^* = \mathbb{E}(\bar{T}_n)$.
The statistic $\bar{T}_n(X,Y)$ is usually asymptotically Gaussian and the central limit theorem can be again used when $\sqrt{n}\beta^* = O(1)$ to obtain the selection probability as $n\to \infty$.
An example is best linear prediction, where the object of inference is
\[\beta_E^* = \mathbb{E}_{F}(X_E^T X_E)^{-1}\mathbb{E}_{F}(X_E^T Y),\]
where Lasso selects the active set of coefficients $E$.
The event of selecting set $E$ with corresponding signs $z_E$ can be described as an affine event of the above form with
\[
\bar{T}_n = \begin{pmatrix}\hat{\beta}_E \\ \frac{1}{n} X_{-E}^T(y - X_E \hat{\beta}_E) \end{pmatrix}
\]
for $\hat{\beta}_E = (X_E^T X_E)^{-1}X_E^TY$.
The statistic $\bar{T}_n$ is asyptotically Gaussian with mean
\[\begin{pmatrix}\beta_E^* \\ \mathbb{E}_{F}(X_{i,-E}^T(y_i - X_{i,E}\beta_E^*) \end{pmatrix}\] and a finite covariance matrix $\Sigma /n$.
A similar example is discussed in the randomized setup in \cite{tian2015selective} with the logistic loss.
For non-local alternatives, namely, when $\|b_E -\sqrt{n} A_E \beta^*\|\to \infty$ as $n \to \infty$, we are back to the univariate problem illustrated above. The proposed approximation gives that
\[\log \mathbb{P}(\bar{T}_n \in \mathcal{K}\lvert \beta^*) \sim -n\inf_{x\in \mathcal{K}}\left\{\Lambda^*(x,\beta^*)+\frac{1}{n}w_{1/\sqrt{n}} (x)\right\}\]
for
\[\Lambda^*(x;\beta^*) = \sup_{\eta} \eta^T x - \log\mathbb{E}[(\exp(\eta^T T)\lvert \beta^*]\text{ and } \mathcal{K} = \{z: A_E z \geq 0\}.\]
In terms of notation, it would make sense to switch to $\bar{T}_n$ to denote the sample mean of an asymptotically Gaussian statistic, but we state all results for $\bar{Y}_n\in \mathbb{R}^d$ in order to avoid introducing additional notation.
\end{remark}
We now shift the focus to the {\it randomized} approximate selective MLE discussed in Section \ref{subsec:mle-rand} and show that it is a consistent estimator for $\beta^*$ under non-local alternatives.
Before proceeding, it is important to note that the {\it nonrandomized} approximate selective MLE based on the barrier approximation \eqref{eq:barrier} is generally not consistent for non-local alternatives:
\begin{theorem} \label{thm:inconsistency}
Let $Y_i \sim N(\beta^*, 1)$ be i.i.d. random variables and consider inference for $\beta^*$ conditionally on $\{ \sqrt{n} \bar{Y}_n >0 \}$.
Then the maximizer $\hat{\beta}^*_n$ of the approximate selection-adjusted log likelihood,
\begin{equation} \label{soft:max:MLE}
\ell_S^n(\beta^*) = -n(\bar{y}_n -\beta^*)^2 + n \inf_{x \in \mathbb{R}}\left\{\frac{(x-\beta^*)^2}{2} + \frac{1}{n}\log\left(1+\frac{1}{\sqrt{n}\bar{y}_n}\right)\right\},
\end{equation}
does not converge in probability to $\beta^*$ when $\beta^*= -\delta<0$.
More generally, $\hat{\beta}^*_n$ is inconsistent for the sequence of parameters $\beta^*_n$ under the asymptotics $\sqrt{n}\beta^*_n \to -\infty$.
\end{theorem}
\begin{remark}
Theorem \ref{thm:inconsistency} is stated for the barrier approximation \eqref{eq:barrier} for clarity of exposition.
The approximate selective MLE is not consistent as long as the function $w(\cdot)$ satisfies
\[x \nabla w(x) \to C \text{ as } x \downarrow 0.\]
\end{remark}
\noindent A remark is also in order regarding the unadjusted estimator $\bar{Y}_n$.
\citet{tian2015selective} showed that there are heavier tailed randomization schemes under which $\bar{Y}_n$ is a consistent estimator for non-local alternatives.
We note that this is not true for Gaussian randomization.
Indeed, let $\beta^*<0$ be fixed (with $n$) and the selection event be $\{\sqrt{n} \bar{Y}_n + W>0\}$, where $W\sim N(0,\gamma^2)$ independently of $\bar{Y}_n$.
Then it can be shown that, for large $n$, the selective distribution of $\sqrt{n}({\bar{Y}_n -\beta^*})$ is approximately $\mathcal{N}(-\sqrt{n}\beta^*/(1+\gamma^2), (1+1/\gamma^2)^{-1})$, and hence $\bar{Y}_n$ cannot be consistent for $\beta^*$.
\medskip
Denote by
\begin{equation}
\label{true:randomized:sel:likelihood}
\ell_S^{n}(\beta^*) = \log f_S^{n}(\bar{y}_n\lvert\beta^*)= -n\|\bar{y}_n -\beta^*\|^2/2 - \log\mathbb{P}(\bar{Y}_n^* \in \mathcal{K}\lvert \beta^*),
\end{equation}
the (exact) selection-adjusted posterior that corresponds to applying selection to $\bar{Y}_n^* \sim \mathcal{N}(\beta^*, 1+\gamma^2)$, the randomized sample mean.
Denote also by
\begin{equation}
\label{randomized:sel:likelihood}
\hat\ell_S^{n}(\beta^*) = \log\hat f_S^{n}(\bar{y}_n\lvert\beta^*)= -n\|\bar{y}_n -\beta^*\|^2/2 + n \inf_{x\in \mathbb{R}^d}\left\{\dfrac{\|x-\beta^*\|^2}{2(1+\gamma^2)} + \frac{1}{n}w_{\sqrt{1+\gamma^2}/\sqrt{n}} (x).\right\}
\end{equation}
the corresponding {approximate} selection-adjusted posterior.
We make a few crucial observations about the approximate sequence of likelihoods $\ell_S^{n}(\beta^*) $ before stating the main result.
\begin{lemma}[Strong convexity] \label{strong:convexity}
The sequence of approximate log partition functions
\begin{equation}
\label{approx:partition:function}
\hat{C}_n(\beta^*) = \cfrac{\gamma^2{\|\beta^*\|}^2}{2(1+\gamma^2)} + \bar{w}_n^*(\beta^*)
\end{equation}
where $\bar{w}_n^*(.)$ is the convex conjugate of
\[\bar{w}_n(x) = \|x\|^2/2(1+\gamma^2) + \frac{1}{n}w_{\sqrt{1+\gamma^2}/\sqrt{n}}(x)\]
and hence, the negative of logarithms of approximate selection adjusted likelihood $\hat f_S^{n}(\beta^*)$
are strongly convex with indices of convexity $m_n$ lower bounded by $M= \gamma^2/(1+\gamma^2)$.
\end{lemma}
\begin{remark}[Uniform convergence on compact sets] \label{unif:convergence}
The approximation $\hat{\ell}_S^{n}(\beta^*)$ is continuous in $\beta^*$ and so is its limit.
Hence the sequence of scaled differences $n^{-1}(\hat{\ell}_S^{n}(\beta^*) - {\ell}_S^{n}(\beta^*))$ converges uniformly on a compact subset $\Theta$ of the parameter space.
\end{remark}
\begin{lemma}
\label{identity:convexity}
If $\hat{C}_n$ is the approximate log partition function in \eqref{approx:partition:function}, then the approximate randomized selective MLE $\hat{\beta}^*_n$ that maximizes \eqref{randomized:sel:likelihood} satisfies the following inequality
\[\|\hat{\beta}^*_n -\beta^*\|^2\leq \cfrac{(\gamma^2+1)^2}{\gamma^4}\|\bar{y}_n -{\nabla} \hat{C}_n(\beta^*)\|^2 \text{ for all } \beta^*\in \mathbb{R}^p.\]
\end{lemma}
\smallskip
We are now ready to prove consistency of the approximate randomized selective MLE.
\begin{theorem}
\label{consistency:approximate:MLE}
For a convex (and compact for $d>1$) set $\mathcal{K}\subset \mathbb{R}^d$, the approximate randomized selective MLE $\beta^*_n$ based on maximizing
$\hat{\ell}_S^{n}(\beta^*)$ in \eqref{randomized:sel:likelihood} is consistent for parameter $\beta^*$ under the selective law \eqref{true:randomized:sel:likelihood}.
\end{theorem}
\bigskip
\noindent To complete the picture, we prove that the approximate randomized selective MLE converges to the exact randomized selective MLE.
\begin{theorem}
\label{mle:convergence}
If $\hat{\beta}^*_n$ denotes the approximate randomized selective MLE, obtained as the maximizer of $\hat{\ell}_S^{n}(\beta^*)$ in \eqref{randomized:sel:likelihood} and $\tilde{\beta}_n^*$ is the selective MLE, which maximizes $\ell_S^{n}(\beta^*)$ in \eqref{true:randomized:sel:likelihood}, then for any $\delta>0$,
\[\mathbb{P}(\|\hat{\beta}^*_n -\tilde{\beta}^*_n\|>\delta\lvert \;\bar{Y}^*_n \in \mathcal{K}) \to 0 \text{ as } n \to \infty. \]
\end{theorem}
\section{Discussion}\label{sec:discussion}
In the current work we adopted the perspective in which post-model selection inference is modeled as a truncation problem.
Hence, inference regarding any specific parameter should be valid under the conditional distribution of the data give that this parameter was selected.
When model selection is done by a known formal procedure specified before observing the data, the conditional approach aims for the exact distribution of the data; therefore, it may lead to more precise inference as compared to methods which offer guarantees irrespective of the selection procedure \citep[for example,][]{berk2013valid, dwork2015generalization}; the latter have their own advantages, especially the fact that they do not depend on the selection algorithm.
A recent line of work \citep[][for example]{lee2013exact, taylor2013tests, taylor2014exact, lee2014exact} has contributed to the feasibility of conditional inference after variable selection in generalized linear models; these contributions are essentially based on the observation that classical theory for exponential families applies also to the truncated distribution \citep{fithian2014optimal}.
However, it is mainly testing (and confidence intervals) that is addressed by this theory and pursued in recent works;
other questions remain open.
For example, this theory has little implication on how to construct ``good" point estimators for the regression coefficients after model selection.
Since, by conditioning, the starting point of the inference stage is {\it after} selection, it makes as much sense to use a prior for the parameters of the selected model as it does in the non-adaptive situation.
The Bayesian approach has the advantage that any kind of inference (estimation or hypothesis testing, for example) is based on the posterior distribution of the parameters given the data:
in that sense, Bayesian inference is more straightforward.
We would like to summarize and point out some further advantages of selection-adjusted Bayesian approach:
\begin{enumerate
\item Even after substituting our approximation (or any other) to the adjustment factor, frequentist inference based on the approximate truncated likelihood is challenging, to say the least.
The conditions for usual asymptotic theory (say, for the likelihood ratio statistic) to hold need to be checked carefully.
On the other hand, when taking a Bayesian approach we can resort to posterior sampling techniques without relying on asymptotics.
As a concrete example consider constructing confidence sets for two coefficients or more after selection; the implementation of UMPU tests in \citet{fithian2014optimal} is considerably more difficult because the problem now does not reduce to constructing a CI for a univariate truncated normal.
By contrast, extending the procedures in Section \ref{sec:inference} to construct credible sets instead of intervals is straightforward.
\item With the Bayesian approach we can easily obtain approximate posterior predictive intervals under the selected model.
Hence, if $\tilde{Y}|\beta^* \sim N(\tilde{x}^T\beta^*,\sigma^2)$ is independent of $Y\in \mathbb{R}^n$, then
$$
f_S(\tilde{y}|y) = \int f_S(\tilde{y},\beta^*|y) d\beta^* = \int f(\tilde{y}|\beta^*) \pi_S(\beta^*|y)d\beta^*
$$
(the unadjusted likelihood is used for $\tilde{y}$).
As usual, this is the marginal distribution of $\tilde{Y}$ with respect to the updated distribution of the parameter of the model, only that observed data $Y$ updates the prior via the {\it selection-adjusted} posterior.
Sampling from the predictive distribution is trivial after obtaining (per the prescription in Section \ref{sec:inference}, for example) a sample from the approximate posterior distribution .
\item The often inconvenient assumption of known noise level can be replaced with a prior distribution on $\sigma^2$.
Our methods can be extended to apply to generalized linear models, and hence can in principle accommodate the situation where there is a prior on $\sigma^2$.
A non-informative prior may be preferred in this situation.
\end{enumerate}
A main message which we tried to convey, then, is that methods suggested recently for conditional inference in the linear model are not all-encompassing; and when existing tools fail to provide conclusive answers, the Bayesian approach pursued in the current paper can be useful.
There are certainly directions for further research.
It would be interesting to study the performance of the approximate point estimators of section \ref{sec:estimation} in finite samples.
An important question is what benchmark can be used for the mean squared error of an estimator.
From a frequentist perspective, there is usually no unbiased estimator, and the Cramer-Rao bound is not applicable (even if we knew how to compute it) in the small-sample regime.
From a Bayesian viewpoint, the minimum Bayes risk -- the expectation of the conditional variance -- usually cannot be computed analytically either, but tractable lower bounds on it (which would automatically bound the Bayes risk of any other estimator) may exist.
Lastly, other suggestions for an approximation to the truncated likelihood may be considered, and it would be interesting to see how sensitive the adjusted posterior is to that choice.
\medskip
\textbf{Acknowledgement.} Providing Bayesian adjusted inference after variable selection was previously proposed by Daniel Yekutieli and Edward George and presented at the 2012 Joint Statistical Meetings in San Diego.
To a large extent, the framework we proposed in Section \ref{sec:framework} relies on their ideas, which we greatly acknowledge.
Our interest re-arose with the recent developments on exact post-selection inference in the linear model.
We are thankful to Daniel and Ed for helpful conversations and to Ed for pointing out the reference to \citet{bayarri1987bayesian}.
We would also like to sincerely thank and acknowledge Chiara Sabatti for the long discussions and for her suggestions along the way, which greatly improved presentation helped to form our perspective on the problem of Bayesian post-model selection inference.
\medskip
\textbf{Supplementary Material.} The supplement contains an analysis of real data and a section on using our methods in frequentist inference.
Specifically, we utilize our methods to derive an approximate uniformly minimum variance unbiased estimator for the saturated model; and describe how our methods can be applied to approximate a uniformly most powerful unbiased test when the computation does not reduce to constructing tests for a univariate truncated normal variable.
\section{Supplement}
\subsection{A real data example} \label{supp:real-data}
We give Bayesian post-selection inference to the publicly available diabetes data set from \cite{efron2004least}.
The data set consists of $10$ standardized measurements from $442$ samples.
We apply the Lasso at a fixed value of the tuning parameter, $\lambda = \hat{\sigma}\mathbb{E}(\|X^T \epsilon \|_{\infty})$ where $\epsilon\sim N_n(0,I)$ \citep[as proposed in][]{negahban2009unified}, to select a subset of predictors.
\cite{lee2014exact} have also analyzed this data set and constructed confidence intervals based on the univariate truncated Gaussian pivot in the {\it saturated} model.
For the above choice of tuning parameter, the Lasso selects the set of variables $E = \{\text{sex, bmi, bp, s1, s3, s5, s6}\}$.
We run selection again, first with randomized Gaussian response, adding an independent variable $W\sim N(0,\gamma^2 I)$ to the response; and second with a random $80\%$ split of the data to construct data-carving intervals.
We obtain credible intervals and point estimates -- selective posterior mean and MAP -- for $\beta^{{E}} = \mathbb{E}[(X_E^T X_E)^{-1} X_E^T Y]$ based on a non-informative prior and the {\it selected} model as the generative model.
For comparison, we also plot the intervals of \cite{lee2014exact}, even though the models for inference are different.
The randomized and data-carving intervals are comparable in length to the unadjusted ones, and much shorter than the proposed Bayesian adjusted intervals and the frequentist confidence intervals of \cite{lee2014exact}.
Another advantage of the proposed adjusted intervals (randomized or not), is that, utilizing posterior sampling, we are able construct them for any generative model.
By contrast, for the confidence intervals of \cite{lee2014exact} it is important to assume the saturated model, because in that case a simple univariate pivot exists.
\begin{figure}[H]
\captionsetup{skip=-10pt}
\centering
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=1\linewidth,keepaspectratio]{diabetes_intervals.png}
\end{subfigure}%
~
\begin{subfigure}[t]{0.5\textwidth}
\centering
\includegraphics[width=1\linewidth,keepaspectratio]{diabetes_point_estimates.png}
\end{subfigure}
\caption{Selection-adjusted inference after selection via Lasso: the left panel gives Bayesian adjusted intervals for $\beta^{{E}} = \mathbb{E}[(X_E^T X_E)^{-1} X_E^T Y]$ under the selected model and a non-informative prior.
Vertical bars are the Barrier approximate adjusted credible intervals from Section 5 of the main paper.
The right panel shows the point estimates: MLE and mean based on (the approximate) selection adjusted posterior are marked on the bars.
}
\label{fig:inference}
\end{figure}
\subsection{Using our approximation in frequentist inference} \label{sec:approx-freq}
Since our approximation serves as a surrogate to the selection-adjusted {\it likelihood}, we can use it to provide approximate inference from an entirely frequentist approach;
as we did in Section
7
when considering the approximate maximum-likelihood (ML) estimate.
In this section we discuss a few other cases where our barrier approximation is advantageous in frequentist computations.
We consider frequentist procedures suggested in \citet{tian2015selective} and in \citet{fithian2014optimal} that involve approximation of intractable integrals; we demonstrate how our methods can be used instead of generating random samples to approximate these integrals, and thereby reduce computational effort.
In fact, our approximation performs well in recovering the densities against which these integrals are taken;
hence, the density produced using our approximation can also serve as an effective choice for a proposal function when implementing MCMC techniques.
\subsubsection{Selective UMVUE under randomized response}
Consider inference under the saturated model, $Y\sim N_n(\beta^*, \sigma^2I)$.
In general, an unbiased estimator for $\beta^*$ conditional on $Y\in \{y:A_Ey\leq b_E\}$ does not exist.
However, \citet{tian2015selective} show that if selection is based on a randomized response, a simple unbiased estimator exists, which can be Rao-Blackwellized to produce an Unbiased Minimum-Variance Estimator (UMVUE) as suggested in \citet{fithian2014optimal}.
Indeed, let $Y^* = Y + W$ where $W\sim N_n(0,\gamma^2I)$ is an independent randomization term, and let the selection event be $\{(y,w): A_E(y+w)\leq b_E\}$.
It is trivial to verify that, marginally,
$$
\hat{\beta}^* = Y - \frac{\sigma^2}{\gamma^2}W
$$
is independent of $Y^*=Y+W$ and unbiased for $\beta^*$.
Therefore, it is unbiased for $\beta^*$ also conditionally on $A_EY^*\leq b_E$.
$Y$ is of course a sufficient statistic for $\beta^*$ conditionally on selection, hence, invoking a Rao-Blackwell argument, the selective UMVUE is $T(Y)$ where
\begin{equation} \label{eq:umvue}
T(y) = \mathbb{E} \left(Y-\frac{\sigma^2}{\gamma^2}W\Big\lvert y, A_E(Y+W)\leq b_E\right)=y-\frac{\sigma^2}{\gamma^2}\mathbb{E} (W\lvert A_E(y+W)\leq b_E).
\end{equation}
Thus, Rao-Blackwellization reduces to computing $\mathbb{E} (W\lvert A_E(y+W)\leq b_E)$, which \citet[Section 3.2]{tian2015selective} suggest to numerically approximate by a Monte-Carlo estimate.
Specifically, Hit-and-Run Gibbs sampling can be used to simulate draws from the conditional distribution of $W$ truncated to $A_EW\leq b_E-A_Ey$.
Using our approximation to the adjustment factor we can offer an alternative approximation to the UMVUE without sampling, as shown by the following theorem.
\begin{theorem}[Approximate selective UMVUE in saturated model] \label{thm:umvue}
Suppose that $Y\sim N(\beta^*,\sigma^2 I_n)$ and let $Y^* = Y + W$ where $W\sim N_n(0,\gamma^2I)$ independent of $Y$.
Then the approximate selective UMVUE for $\mu$ conditionally on $A_EY^*\leq b_E$ is given by
$$
\tilde{T}(y) = \left(1+\frac{\sigma^2}{\gamma^2}\right)y-\frac{\sigma^2}{\gamma^2} z^*(y)
$$
where
\begin{equation}
\label{opt-sat}
z^*(y)=\arg\max_{z}\left\{\dfrac{z^T y}{\gamma^2} -\left(\dfrac{\lVert z\lVert ^2}{2\gamma^2}+\sum_{i=1}^{m}\log\left(1+\cfrac{\gamma}{(b_i- a_i^T z)}\right)\right)\right\}.
\end{equation}
\end{theorem}
\begin{proof}
We have
$$
\begin{aligned}
\mathbb{E}\left(Y-\frac{\sigma^2}{\gamma^2}W\Big\lvert y, A_E(W+y)\leq b_E\right)&=y-\frac{\sigma^2}{\gamma^2}\mathbb{E}(W\lvert y, A_E(y+W)\leq b_E)\\
&= \left(1+\frac{\sigma^2}{\gamma^2}\right)y-\frac{\sigma^2}{\gamma^2}\mathbb{E}(W+y \lvert y, A_E(W+y)\leq b_E)\\
&=\left(1+\frac{\sigma^2}{\gamma^2}\right)y-\frac{\sigma^2}{\gamma^2}\mathbb{E}(R\lvert y, A_ER \leq b_E)
\end{aligned}
$$
where $R=(W+y)$.
The task is now to approximate the conditional expectation in the last equation.
Since $R \sim N_n(\alpha, \gamma^2 I_n)$ with $\alpha=y$, the conditional distribution of $R$ given $A_ER\leq b_E$, belongs to an exponential family with density
$$
f(r\lvert A_ER\leq b_E) \propto \exp\left(-\lVert r-\alpha\lVert^2/2\gamma^2 \right)I{(A_Er\leq b_E)}.
$$
The corresponding log-partition function is
$${A}(\eta)=-\gamma^2 \frac{\eta^T\eta}{2}-\log {\mathbb{P}}(A_ER\leq b_E\lvert \eta),
$$
where $\eta=\alpha/\gamma^2$ is a re-parametrization of the natural parameter.
With the barrier approximation for negative of $\log {\mathbb{P}}(A_ER\leq b_E)$ plugged in,
we approximate the log-partition function $A(\eta)$ as
$$\sup_{z\in \mathbb{R}^n}z^T\eta -\left(\dfrac{\lVert z\lVert^2}{2\gamma^2}+\sum_{i=1}^m \log(1+\gamma/(b_i- a_i^Tz))\right)=g^*(\eta),$$
the convex conjugate of $g(z)=\dfrac{\lVert z\lVert^2}{2\gamma^2}+\sum_{i=1}^m \log\left(1+\cfrac{\gamma}{(b_i- a_i^Tz)}\right)$.
Then
$$
\begin{aligned}
\mathbb{E}\left(y-\frac{\sigma^2}{\gamma^2}W\Big\lvert y, A_E(W+y)\leq b_E\right) &\approx \left(1+\frac{\sigma^2}{\gamma^2}\right)y-\frac{\sigma^2}{\gamma^2}\dfrac{\partial{A(\eta)}}{\partial{\eta}}\Bigg\lvert_{y/\gamma^2}\\
&= \left(1+\frac{\sigma^2}{\gamma^2}\right)y-\frac{\sigma^2}{\gamma^2} {\nabla} g^*(\eta)\lvert_{y/\gamma^2}\\
&= \left(1+\frac{\sigma^2}{\gamma^2}\right)y-\frac{\sigma^2}{\gamma^2}({\nabla} g)^{-1}(\eta)\lvert_{y/\gamma^2}\\
&= \left(1+\frac{\sigma^2}{\gamma^2}\right)y-\frac{\sigma^2}{\gamma^2} z^*(y).
\end{aligned}
$$
The final equation follows from observing that
$$z^*(y)=\arg\max_{z}\left\{\dfrac{z^T y}{\gamma^2} -\left(\dfrac{\lVert z\lVert ^2}{2\gamma^2}+\sum_{i=1}^{m}\log\left(1+\cfrac{\gamma}{(b_i- a_i^T z)}\right)\right)\right\}$$
satisfies
$$
y/\gamma^2={\nabla} g(z^*(y)),
$$
which implies
$$
z^*(y)=({\nabla} g)^{-1}(y/\gamma^2)={\nabla} g^*(y/\gamma^2).
$$
\end{proof}
\addtocounter{example}{+1}
\begin{example}[continued] \label{ex:univar-umvue}
The setup is the randomized setup of Example
2
of Subsection 7.1.
By Theorem \ref{thm:umvue}, the approximate selective UMVUE is
$$\hat{\beta}^*\approx \left(1+\frac{1}{\gamma^2}\right)y-\frac{1}{\gamma^2}z^*(y),\text{ where}$$
$$
z^*(y)=\arg\max_{z}\left\{\dfrac{z^T y}{\gamma^2} -\left(\dfrac{\lVert z\lVert ^2}{2\gamma^2}+\log(1+\gamma/z)\right)\right\}.
$$
Figure \ref{fig:umvue} shows how closely the approximate UMVUE mimics the exact UMVUE in the univariate example.
\begin{figure}[]
\centering
\includegraphics[scale=.5]{UMVUE_true_approx.png}
\caption{Approximate UMVUE for the univariate example with randomization, $Y\sim N(\beta^*,1)$ truncated to $\{Y+W>0\}$. }
\label{fig:umvue}
\end{figure}
\end{example}
The statement in Theorem \ref{thm:umvue} is for the saturated model.
\subsubsection{Approximate UMPU tests under the selected model} \label{supp:umpu}
Consider inference for $\beta^*_j$ under a generative model, $Y\sim N(X^*\beta^*, \sigma^2 I)$, and conditionally on $\{\widehat{E}(Y)=E\}$, equivalently, $\{A_EY\leq b_E\}$. To present this section without notational challenges, let $X^*= X_{\tilde{E}}, \beta^* =\beta_{\tilde{E}}$ for $\tilde{E}\subset \{1,2,\cdots, p\}.$ The special case of $\tilde{E} =E$ corresponding to $X^*= X_E$ leads to inference in the selected model.
For known $\sigma^2$, it follows from the general theory in \citet{fithian2014optimal} that Uniformly Most Powerful Unbiased (UMPU) tests (and confidence intervals) are based on the conditional law
\begin{equation} \label{eq:law-cond}
{{\mathcal{L}}}( X_{j.\tilde{E}}^TY |A_EY \leq b_E, X_{\tilde{E} \setminus j}^TY)= {{\mathcal{L}}}({\eta}^T Y |A_EY \leq b_E, X_{\tilde{E} \setminus j}^T Y) ,
\end{equation}
with
$$
\eta=\dfrac{X_{j\cdot \tilde{E}}}{\lVert X_{j\cdot \tilde{E}}\lVert^2}, \ \ \ \ \ X_{j\cdot \tilde{E}}=R_{\tilde{E}\setminus j}X_j
$$
and $R_{\tilde{E}\setminus j} = (I-P_{\tilde{E}\setminus j})$ and $P_{\tilde{E}\setminus j} = X_{\tilde{E}\setminus j}(X_{\tilde{E}\setminus j}^T X_{\tilde{E}\setminus j})^{-1}X_{\tilde{E}\setminus j}^T$ is the projection matrix onto the column space of $X_{\tilde{E}\setminus j}$. In particular, this distribution depends only on $\beta_{j.\tilde{E}}$, the $j$-th parameter in the model $\tilde{E}$.
Denoting $T={\eta}^T Y$, we observe that, unconditionally,
$
T\sim N (\beta_{j.\tilde{E}},\sigma^2 /\lVert X_{j\cdot \tilde{E}}\lVert^2).
$
Finally, noting that
\begin{equation} \label{eq:decomp-y}
y= P_{\tilde{E}\setminus j} y + (t/{\lVert \eta\lVert^2} )\eta + P_{\tilde{E}}^\perp y
\end{equation}
where $P_{\tilde{E}}^\perp = (I-P_{\tilde{E}})$, we conclude that
$$
{{\mathcal{L}}}({\eta}^T Y |A_E Y \leq b_E, X_{\tilde{E} \setminus j}^Ty)\equiv {{\mathcal{L}}}(T\lvert A_E\{(T/{\lVert \eta\lVert^2})\eta + P_{\tilde{E}}^\perp Y\}\leq \tilde{b},P_{\tilde{E}\setminus j} y)
$$
with
$$
\tilde{b}=b-A P_{\tilde{E}\setminus j} y.
$$
This law belongs to a one parameter exponential family with density proportional to
\begin{equation}
\label{exp family}
\exp\left\{-{\lVert X_{j\cdot \tilde{E}}\lVert^2(t-\beta_{j.\tilde{E}})^2}/{2\sigma^2}\right\}\times \int_{\{v: A_E\{(t/{\lVert \eta\lVert^2})\eta + v\}\leq \tilde{b}\}}\exp(-v^T \tilde{\Sigma} _V ^{-1} v/2)dv
\end{equation}
where
$$
V = (I-P_{\tilde{E}})Y \sim N_n(0,{\Sigma}_V),\ \ \ \ \ \ \ {\Sigma}_V = \sigma^2(I-P_{\tilde{E}}).
$$
As
$$
\int_{\{v: A_E\{(t/{\lVert \eta\lVert^2})\eta + v\}\leq \tilde{b}\}}\exp(-v^T {\Sigma} _V ^{-1} v/2)dv
\ \propto \ \mathbb{P}(A_E V\leq b'(t))
$$
where $ b'(z)=\tilde{b}-(z/{\lVert \eta\lVert^2})A_E \eta$, we can use our approximation to compute the integral.
In this way will be able to compute the approximate density of $T$ on a grid of values on the real line, and use this discretized version to obtain the approximate UMPU tests described in general in Theorem 5 of \citet{fithian2014optimal}.
\bibliographystyle{plainnat}
|
\section{Introduction}
Given jointly independent $m$ samples from $P=(p_1,\cdots,p_S)$ and $n$ samples from $Q=(q_1,\cdots,q_S)$ over some \emph{unknown} common alphabet of size $S$, consider the problem of estimating a functional of the distribution of the following form:
\begin{align}
F(P,Q) = \sum_{i=1}^S f(p_i,q_i)
\end{align}
where $f: A\to \mathbb{R}$ is a continuous function with some $A\subset [0,1]^2$. Note that by allowing $f$ to solely depend on $p$, this problem generalizes the functional estimation problem considered in \cite{Jiao--Venkat--Han--Weissman2015minimax}. Among the most fundamental of such functionals is the $f$-divergence \cite{csisz1967information}
\begin{align}
D_f(P\|Q) = \int f\left(\frac{dP}{dQ}\right)dQ = \sum_{i=1}^S f\left(\frac{p_i}{q_i}\right)q_i
\end{align}
for some convex function $f$ with $f(1)=0$. The $f$-divergence serves as the fundamental information contained in binary statistical models \cite{liese2007statistical} and enjoys numerable applications in information theory \cite{Cover--Thomas2006} and statistics \cite{Tsybakov2008}.
Among many $f$-divergences, we focus on the estimation problem of the Kullback--Leibler (KL) divergence (with $f(t)=t\ln t-t+1$) in this paper, and the general approach naturally extends to the Hellinger distance and $\chi^2$-divergence. The KL divergence is an important measure of the discrepancy between two discrete distributions $P=(p_1,\cdots,p_S)$ and $Q=(q_1,\cdots,q_S)$, defined as \cite{kullback1951information}
\begin{align}
D(P\|Q) = \begin{cases}
\sum_{i=1}^{S} p_i\ln\frac{p_i}{q_i}& \text{if }P\ll Q,\\
+\infty& \text{otherwise,}
\end{cases}
\end{align}
where $P\ll Q$ denotes that the absolute continuity of $P$ with respect to $Q$. Like the entropy and mutual information \cite{Shannon1948}, the KL divergence is a key information theoretic measure arising naturally in data compression \cite{catoni2004statistical}, communications \cite{csiszar2011information}, probability theory \cite{sanov1958probability}, statistics \cite{kullback1997information}, optimization \cite{dempster1977maximum}, machine learning \cite{bishop2006pattern, kingma2013auto}, and many other disciplines. Throughout the paper we use the squared error loss, i.e., the risk function for any estimator $\hat{D}$ is defined as
\begin{align}
L(\hat{D};P,Q) \triangleq \mathbb{E}_{(P,Q)} |\hat{D}-D(P\|Q)|^2.
\end{align}
The maximum risk of an estimator $\hat{D}$, and the minimax risk in estimating $D(P\|Q)$ are defined as
\begin{align}
R_{\text{maximum}}(\hat{D};\mathcal{U}) &\triangleq \sup_{(P,Q)\in\mathcal{U}} L(\hat{D};P,Q),\\
R_{\text{minimax}}(\mathcal{U}) &\triangleq \inf_{\hat{D}}\sup_{(P,Q)\in\mathcal{U}} L(\hat{D};P,Q)
\end{align}
respectively, where $\mathcal{U}$ is a given collection of probability measures $(P,Q)$, and the infimum is taken over all possible estimators $\hat{D}$. We aim to obtain the minimax risk $R_{\text{minimax}}(\mathcal{U})$ for some properly chosen $\mathcal{U}$.
Notations: for non-negative sequences $a_\gamma,b_\gamma$, we use the notation $a_\gamma \lesssim b_\gamma$ to denote that there exists a universal constant $C$ such that $\sup_{\gamma } \frac{a_\gamma}{b_\gamma} \leq C$, and $a_\gamma \gtrsim b_\gamma$ is equivalent to $b_\gamma \lesssim a_\gamma$. Notation $a_\gamma \asymp b_\gamma$ is equivalent to $a_\gamma \lesssim b_\gamma$ and $b_\gamma \lesssim a_\gamma$. Notation $a_\gamma \gg b_\gamma$ means that $\liminf_\gamma \frac{a_\gamma}{b_\gamma} = \infty$, and $a_\gamma \ll b_\gamma$ is equivalent to $b_\gamma \gg a_\gamma$. We write $a\wedge b=\min\{a,b\}$ and $a\vee b=\max\{a,b\}$. Moreover, $\mathsf{poly}_n^d$ denotes the set of all $d$-variate polynomials of degree no more than $n$, and $E_n[f;I]$ denotes the distance of the function $f$ to the space $\mathsf{poly}_n^d$ in the uniform norm $\|\cdot\|_{\infty,I}$ on $I\subset \mathbb{R}^d$. All logarithms are in the natural base.
\subsection{Background and main results}
There have been several attempts to estimate the KL divergence for the continuous case, see~\cite{Wang--Kulkarni--Verdu2005divergence,Lee--Park2006estimation,Gretton--Borgwardt--Rasch--Scholkopf--Smola2006kernel,Perez2008kullback,Wang--Kulkarni--Verdu2009divergence,Nguyen--Wainwright--Jordan2010estimating} and references therein. These approaches usually do not operate in the minimax framework, and focus on consistency but not rates of convergence, unless strong smoothness conditions on the densities are imposed to achieve the parametric rate (i.e., $\Theta(n^{-1})$ in mean squared error). In the discrete setting, \cite{Cai--Kulkarni--Verdu2006universal} and \cite{Zhang--Grabchak2014nonparametric} proved consistency of some specific estimators without arguing minimax optimality. We note that in the discrete case, if the alphabet size $S$ is fixed and the number of samples $m,n$ go to infinity, the standard H\'{a}jek--Le Cam theory of classical asymptotics shows that the plug-in approach is asymptotically efficient \cite[Thm. 8.11, Lemma 8.14]{Vandervaart2000}. The key challenge we face in the discrete setting is the regime where the support size $S$ can be comparable to or even larger than the number of observations $m,n$, which classical analyses do not address.
Now we consider the estimation of KL divergence between discrete distributions in a large-alphabet setting. For the choice of $\mathcal{U}$, it may appear natural to allow $P$ to be any distribution which is absolutely continuous with respect to $Q$ with alphabet size $S$, i.e.,
\begin{align}
\mathcal{U}_S=\{(P,Q):P,Q\in\mathcal{M}_S,P\ll Q\}
\end{align}
where $\mathcal{M}_S$ denotes the set of all probability measures with support size $S$. However, in this case, it turns out to be impossible to estimate the KL divergence in the minimax sense, i.e., $R_{\text{minimax}}(\mathcal{U}_S)=\infty$ for any configuration $(S,m,n)$ with $S\ge 2$. Intuitively, this is because that the observation from the Multinomial model depends continuously on $(P,Q)$ while the KL divergence does not at extremal points. A rigorous statement and proof of this result is given in Lemma \ref{lem.minimax_unbounded} of the Appendix.
It seems natural then to consider an alternative uncertainty set with bounded likelihood ratio:
\begin{align}
\mathcal{U}_{S,u(S)} = \{(P,Q): P,Q\in\mathcal{M}_S, \frac{p_i}{q_i}\le u(S), \forall i\}
\end{align}
where $u(S)\ge 1$ is an upper bound on the likelihood ratio. Since $u(S)=1$ results in the trivial case where $D(P\|Q)\equiv 0$, throughout we will assume that $u(S)\ge c$ for some constant $c>1$.
The main result of this paper is as follows.
\begin{theorem}\label{thm.kl_divergence}
For $m\gtrsim S/\ln S$, $n\gtrsim Su(S)/\ln S$, $u(S)\gtrsim (\ln S)^2$ and $\ln S\gtrsim \ln(m\vee n)$, we have
\begin{align}
R_{\text{\rm minimax}}(\mathcal{U}_{S,u(S)}) \asymp \left(\frac{S}{m\ln m}+\frac{Su(S)}{n\ln n}\right)^2 + \frac{(\ln u(S))^2}{m} + \frac{u(S)}{n}.
\end{align}
Furthermore, our estimator $\hat{D}_{\text{\rm A}}$ in Section \ref{sec.construction} achieves this bound under the Poisson sampling model, and is adaptive in the sense that it does not require knowledge of $S$ or $u(S)$.
\end{theorem}
The following corollary is a direct consequence of Theorem \ref{thm.kl_divergence}. Note that $\ln S\gtrsim \ln n$ and $n\gtrsim Su(S)/\ln S$ have already implied that $\ln S\gtrsim\ln u(S)$, and thus $\ln(Su(S))\asymp \ln S$.
\begin{corollary}\label{cor.sample_complexity}
For our KL divergence estimator, the maximum mean squared error vanishes provided that $m\gg S/\ln S$ and $n\gg Su(S)/\ln S$. Moreover, if $m\lesssim S/\ln S$ or $n\lesssim Su(S)/\ln S$, then the maximum risk of any estimator for KL divergence is bounded away from zero.
\end{corollary}
Next we consider the plug-in approach in the context of minimax rate-optimality. Since it is possible that $\hat{p}_i>0$ and $\hat{q}_i=0$ for some $i\in\{1,\cdots,S\}$, where $P_m=(\hat{p}_1,\cdots,\hat{p}_S),Q_n=(\hat{q}_1,\cdots,\hat{q}_S)$ are the respective empirical probability distributions, the direct plug-in estimate $D(P_m\|Q_n)$ may be infinity with positive probability. Hence, we use the following modification of the direct plug-in approach: when we observe that $\hat{p}_i>0$ and $\hat{q}_i=0$, since naturally $\hat{q}_i$ is an integral multiple of $1/n$, we manually change the value of $\hat{q}_i$ to the closed lattice $1/n$ of zero. More precisely, we define
\begin{align}
Q_n' = \left(\frac{1}{n}\vee\hat{q}_1,\cdots,\frac{1}{n}\vee\hat{q}_S\right)
\end{align}
and use the estimator $D(P_m\|Q_n')$ to estimate the KL divergence. Note that $Q_n'$ may not be a probability distribution (in which case $D(P_m\|Q_n')$ is extended in the obvious way). The performance of this modified plug-in approach is summarized in the following theorem.
\begin{theorem}\label{thm.mle}
Under the Poisson sampling model, the modified plug-in estimator $D(P_m\|Q_n')$ satisfies
\begin{align}\label{eq.mle_upper}
R_{\text{\rm maximum}}(D(P_m\|Q_n');\mathcal{U}_{S,u(S)})\lesssim \left(\frac{S}{m}+\frac{Su(S)}{n}\right)^2 + \frac{(\ln u(S))^2}{m} + \frac{u(S)}{n}.
\end{align}
Moreover, for $m\ge 15S$ and $n\ge 4Su(S)$, we have
\begin{align}\label{eq.error_mle}
R_{\text{\rm maximum}}(D(P_m\|Q_n');\mathcal{U}_{S,u(S)})\gtrsim \left(\frac{S}{m}+\frac{Su(S)}{n}\right)^2 +\frac{(\ln u(S))^2}{m} + \frac{u(S)}{n}.
\end{align}
\end{theorem}
The following corollary on the minimum sample complexity is immediate.
\begin{corollary}\label{cor.sample_complexity_mle}
The worst-case mean squared error of the modified plug-in estimator $D(P_m\|Q_n')$ vanishes if and only if $m\gg S\vee(\ln u(S))^2$ and $n\gg Su(S)$.
\end{corollary}
Hence, compared with the mean squared error or the minimum sample complexity of the modified plug-in approach, the optimal estimator enjoys a logarithmic improvement. Note that $(\ln u(S))^2\lesssim (\ln S)^2\ll S$ is negligible under the condition in Theorem \ref{thm.kl_divergence}, so there is no counterpart of $m\gg (\ln u(S))^2$ in Corollary \ref{cor.sample_complexity}. Specifically, the performance of the optimal estimator with $(m,n)$ samples is essentially that of the plug-in approach with $(m\ln m,n\ln n)$ samples, which is another manifestation of the \emph{effective sample size enlargement} phenomenon~\cite{Jiao--Venkat--Han--Weissman2015minimax,jiao2016minimax}. Note that in the KL divergence example, the modified plug-in estimator $D(P_m\|Q_n')$ essentially exploits the plug-in idea.
After our submission of this work to arXiv, an independent study of the same problem was presented in ISIT 2016 \cite{bu2016estimation} without the construction of the optimal estimator, which was added to the full version \cite{bu2016estimation_journal} that appeared later on arXiv. Specifically, the main result (i.e., Theorem \ref{thm.kl_divergence}) was also obtained in \cite{bu2016estimation_journal}, while there are some differences. First, our estimator is agnostic to both the support size $S$ and the upper bound $u(S)$ on the likelihood ratio, while the estimator in \cite{bu2016estimation_journal} requires both. Second, as for Theorem \ref{thm.mle}, there is an unnecessary additional term $\frac{(\ln S)^2}{m}$ in the upper bound \eqref{eq.mle_upper} of the plug-in approach in \cite{bu2016estimation_journal}, though there is a minor difference between our choices of the plug-in estimator. Third, and most significant, \cite{bu2016estimation_journal} is dedicated exclusively to the KL divergence case, while in our paper we propose a general approximation-based methodology for the estimation of a wide class of functionals, with the estimation of KL divergence serving as the main example for concrete illustration of the concepts. As additional examples, following the general recipe in the next subsection and the later analysis, the result on estimating the $L_1$ distance in \cite{jiao2016minimax} can be recovered, and for the Hellinger distance and the $\chi^2$-divergence \cite{csisz1967information}
\begin{align}
H^2(P,Q) &\triangleq \frac{1}{2}\sum_{i=1}^S \left(\sqrt{p_i}-\sqrt{q_i}\right)^2\\
\chi^2(P,Q) &\triangleq \begin{cases}
\sum_{i=1}^S \frac{p_i^2}{q_i} - 1 & \text{if }P\ll Q,\\
+\infty & \text{otherwise}
\end{cases}
\end{align}
we can similarly obtain the following results on the optimal estimation rates in the large-alphabet setting.
\begin{theorem}\label{thm.hellinger}
For $m\wedge n\gtrsim S/\ln S$ and $\ln S\gtrsim \ln(m\vee n)$, for Hellinger distance we have
\begin{align}
\inf_{\hat{T}}\sup_{P\in\mathcal{M}_S, Q\in\mathcal{M}_S} \mathbb{E}_{(P,Q)}\left(\hat{T}-H^2(P,Q)\right)^2\asymp \frac{S}{(m\wedge n)\ln (m\wedge n)} + \frac{1}{m\wedge n}
\end{align}
and the estimator in Section \ref{sec.others} achieves this bound without the knowledge of $S$ under the Poisson sampling model.
\end{theorem}
\begin{theorem}\label{thm.chi_squared}
For $n\gtrsim \frac{S(u(S))^2}{\ln S}, u(S)\gtrsim (\ln S)^2$ and $\ln S\gtrsim\ln(m\vee n)$, for $\chi^2$-divergence we have
\begin{align}
\inf_{\hat{T}}\sup_{(P,Q)\in\mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left(\hat{T}-\chi^2(P,Q)\right)^2\asymp \left(\frac{S(u(S))^2}{n\ln n}\right)^2 + \frac{(u(S))^2}{m} + \frac{(u(S))^3}{n}
\end{align}
and the estimator in Section \ref{sec.others} achieves this bound without the knowledge of either $S$ or $u(S)$ under the Poisson sampling model.
\end{theorem}
The following corollaries on the minimum sample complexities follow directly from the previous theorems.
\begin{corollary}\label{cor.hellinger}
For Hellinger distance over $(P,Q)\in \mathcal{M}_S\times \mathcal{M}_S$, there exists an estimator with a vanishing maximum mean squared error if and only if $m\wedge n\gg \frac{S}{\ln S}$.
\end{corollary}
\begin{corollary}\label{cor.chi_squared}
For $\chi^2$-divergence over $(P,Q)\in\mathcal{U}_{S,u(S)}$, there exists an estimator with a vanishing maximum mean squared error if and only if $m\gg (u(S))^2$ and $n\gg \frac{S(u(S))^2}{\ln S}\vee (u(S))^3$.
\end{corollary}
\subsection{Approximation: the general recipe}
Estimation of KL divergence belongs to a large family of functional estimation problems: consider estimating the functional $G(\theta)$ of a parameter $\theta \in \Theta \subset \mathbb{R}^p$ for an experiment $\{P_\theta: \theta \in \Theta\}$. There has been a recent wave of study on functional estimation of high dimensional parameters, e.g., the scaled $\ell_1$ norm $\frac{1}{n}\sum_{i=1}^{n}|\theta_i|$ in the Gaussian model \cite{Cai--Low2011}, the Shannon entropy $\sum_{i=1}^S-p_i\ln p_i$ \cite{Valiant--Valiant2011power,Valiant--Valiant2013estimating,Wu--Yang2014minimax,Jiao--Venkat--Han--Weissman2015minimax}, the mutual information\cite{Jiao--Venkat--Han--Weissman2015minimax}, the power sum function $\sum_{i=1}^{S}p_i^\alpha$ \cite{Jiao--Venkat--Han--Weissman2015minimax}, the R\'{e}nyi entropy $\frac{\ln \sum_{i=1}^S p_i^\alpha}{1-\alpha}$\cite{Acharya--Orlitsky--Suresh--Tyagi2014complexity} and the $\ell_1$ distance $\sum_{i=1}^S |p_i-q_i|$\cite{jiao2016minimax} in Multinomial and Poisson models. Moreover, the effective sample size enlargement phenomenon holds in all these examples: the performance of the minimax estimators with $n$ samples is essentially that of the plug-in approach with $n\ln n$ samples.
The optimal estimators in the previous examples all follow the general methodology of \emph{Approximation} proposed in~\cite{Jiao--Venkat--Han--Weissman2015minimax}: suppose $\hat{\theta}_n$ is a consistent estimator for $\theta$, where $n$ is the number of observations. Suppose the functional $G(\theta)$ is analytic\footnote{A function $f$ is analytic at a point $x_0$ if and only if its Taylor series about $x_0$ converges to $f$ in some neighborhood of $x_0$.} everywhere except at $\theta \in \Theta_0$. A natural estimator for $G(\theta)$ is $G(\hat{\theta}_n)$, and we know from classical asymptotics \cite[Lemma 8.14]{Vandervaart2000} that given the benign LAN (Local Asymptotic Normality) condition~\cite{Vandervaart2000}, $G(\hat{\theta}_n)$ is asymptotically efficient for $G(\theta)$ for $\theta \notin \Theta_0$ if $\hat{\theta}_n$ is asymptotically efficient for $\theta$. In the estimation of functionals of discrete distributions, $\Theta$ is the $S$-dimensional probability simplex, and a natural candidate for $\hat{\theta}_n$ is the empirical distribution, which is unbiased for any $\theta\in\Theta$. Then the following two-step procedure is conducted in estimating $G(\theta)$.
\begin{enumerate}
\item \textbf{Classify the Regime}: Compute $\hat{\theta}_n$, and declare that we are in the ``non-smooth'' regime if $\hat{\theta}_n$ is ``close'' enough to $\Theta_0$. Otherwise declare we are in the ``smooth'' regime;
\item {\bf Estimate}:
\begin{itemize}
\item If $\hat{\theta}_n$ falls in the ``smooth'' regime, use an estimator ``similar'' to $G(\hat{\theta}_n)$ to estimate $G(\theta)$;
\item If $\hat{\theta}_n$ falls in the ``non-smooth'' regime, replace the functional $G(\theta)$ in the ``non-smooth'' regime by an approximation $G_{\text{appr}}(\theta)$ (another functional) which can be estimated without bias, then apply an unbiased estimator for the functional $G_{\text{appr}}(\theta)$.
\end{itemize}
\end{enumerate}
Simple as it may sound, this methodology has a few drawbacks and ambiguities. In our recent work \cite{jiao2016minimax}, we applied this general recipe to the estimation of $\ell_1$ distance between two discrete distributions, where this recipe proves to be inadequate. In the estimation of the $\ell_1$ distance, a \emph{bivariate} function $f(x,y)=|x-y|$ which is non-analytic in a \emph{segment} was considered, which is completely different from the previous studies \cite{Valiant--Valiant2011power,Valiant--Valiant2013estimating,Wu--Yang2014minimax,Jiao--Venkat--Han--Weissman2015minimax,Acharya--Orlitsky--Suresh--Tyagi2014complexity} where a univariate function analytic everywhere except a point is always taken into consideration. In particular, two more topics, i.e., \emph{multivariate approximation} and \emph{localization via confidence sets}, were introduced and used.
\begin{question}\label{Q1}
What if the domain of $\hat{\theta}_n$ is different from (usually larger than) $\Theta$, the domain of $\theta$?
\end{question}
\begin{question}\label{Q2}
How to determine the ``non-smooth" regime? What is its size?
\end{question}
\begin{question}\label{Q3}
If $\hat{\theta}_n$ falls in the ``non-smooth" regime, in which region should $G_{\textrm{appr}}(\theta)$ be a good approximation of $G(\theta)$ (e.g., the whole domain $\Theta$, or a proper neighborhood of $\hat{\theta}_n$)?
\end{question}
\begin{question}\label{Q4}
If $\hat{\theta}_n$ falls in the ``smooth" regime, how to construct an estimator ``similar" to $G(\hat{\theta}_n)$?
\end{question}
Other questions, such as what type/degree of approximation $G_{\textrm{appr}}(\theta)$ should be used, were answered in more detail in \cite{Jiao--Venkat--Han--Weissman2015minimax}. Among these questions, Question \ref{Q1} is a relatively new one, where the estimation of KL divergence is the second example so far for which it has arisen, where the first example on estimating the support size of a discrete distribution \cite{wu2015chebyshev} did not explicitly propose and answer this question. Question \ref{Q2} and \ref{Q3} were partially addressed in \cite{Jiao--Venkat--Han--Weissman2015minimax} and \cite{jiao2016minimax}, but the answer to Question \ref{Q2} changes in view of Question \ref{Q1}, and further elaborations are also necessary for Question \ref{Q3}. As for Question \ref{Q4}, the previous approaches can only handle order-one bias correction, while for some problems bias correction with an arbitrary order is proved to be necessary \cite{Han--Jiao--Mukherjee--Weissman2016optimal}. Before answering these questions, we begin with a formal definition of \emph{confidence set} in statistical experiments, which is motivated by \cite{jiao2016minimax}.
\begin{definition}[Confidence set]\label{def.localization}
Consider a statistical model $(P_\theta)_{\theta\in\Theta}$ and an estimator $\hat{\theta}\in \hat{\Theta}$ of $\theta$, where $\Theta\subset \hat{\Theta}$. For $r\in [0,1]$, a confidence set of significance level $r$ is a collection of sets $\{U(x)\}_{x\in \hat{\Theta}}$, where $U(x)\subset \Theta$ for any $x\in \hat{\Theta}$, and\footnote{In standard terminology in statistical testing, this is also the confidence set of level $1-r$.}
\begin{align}
\sup_{\theta\in \Theta}\mathbb{P}_\theta(\theta\notin U(\hat{\theta})) \le r.
\end{align}
Moreover, every confidence set of significance level $r$ can also induce a reverse confidence set $\{V(y)\}_{y\in\Theta}$ of significance level $r$, where $V(y)\triangleq \{x\in \hat{\Theta}: y\in U(x)\}\subset \hat{\Theta}$ for any $y\in\Theta$, and
\begin{align}
\sup_{\theta\in\Theta} \mathbb{P}_\theta(\hat{\theta}\notin V(\theta))\le r.
\end{align}
\end{definition}
Intuitively, if $\{U(x)\}_{x\in \hat{\Theta}}$ is a confidence set of significance level $r$, then after observing $\hat{\theta}$ we can conclude that $\theta\in U(\hat{\theta})$ with error probability at most $r$. More precisely, for any $\theta\in\Theta$, with probability at least $1-r$, we can get back to $\theta$ based on $U(\cdot)$ after observing $\hat{\theta}$. Conversely, with probability at least $1-r$, we can also restrict $\hat{\theta}$ in the region $V(\theta)$. In other words, the true parameter $\theta$ is localized at $U(\hat{\theta})$, and the observation $\hat{\theta}$ is localized at $V(\theta)$, from which the name \emph{localization via confidence sets} originates. Note that confidence set of any level exists for any statistical model $(P_\theta)_{\theta\in\Theta}$ and estimator $\hat{\theta}$, since $U(x)\equiv \Theta$ is always a feasible confidence set of level zero (and then $V(y)\equiv \hat{\Theta}$). In practice, we seek confidence sets which are as small as possible. We also remark that, apart from the confidence set used in traditional hypothesis testing where $r$ is usually chosen to be a fixed constant (e.g., $0.01$), here we allow $r$ to decay with $n$, e.g., $r_n\asymp n^{-A}$ with some constant $A>0$. For example, in the Binomial model $n\hat{p}\sim \mathsf{B}(n,p)$ with $\hat{\Theta}=[0,1]$ and any $\Theta\subset \hat{\Theta}$, for $r_n\asymp n^{-A}$, by measure concentration (cf. Lemma \ref{lemma.poissontail} in Appendix \ref{appen.aux}) the collection $\{U(x)\}_{x\in [0,1]}$ with
\begin{align}
U(x) = \Theta \cap \begin{cases}
[0, \frac{c_1\ln n}{n}] & \text{if } x\le \frac{c_1\ln n}{n},\\
[x-\sqrt{\frac{c_1x\ln n}{n}}, x+\sqrt{\frac{c_1x\ln n}{n}}] & \text{if } \frac{c_1\ln n}{n} < x \le 1
\end{cases}
\end{align}
is a confidence set of significance level $r_n$ assuming the universal constant $c_1>0$ is large enough, and the induced reverse confidence set is contained in
\begin{align}
V(y) = [0,1]\cap \begin{cases}
[0,\frac{4c_1\ln n}{n}] & \text{if } y\le \frac{2c_1\ln n}{n},\\
[y-\sqrt{\frac{2c_1y\ln n}{n}}, y+\sqrt{\frac{2c_1y\ln n}{n}}] & \text{if } \frac{2c_1\ln n}{n} <y\le 1
\end{cases}
\end{align}
which is of a similar structure. Figure \ref{fig.localization} gives a pictorial illustration of both the confidence set and the reverse confidence set in 2D Binomial and Gaussian models, respectively.
\begin{figure}[!htbp]
\centering
\includegraphics{localization.pdf}
\caption{Pictorial illustration of confidence set $U(\hat{\theta})$ and reverse confidence set $V(\theta)$ in 2D Binomial (left panel) and Gaussian (right panel) models. In the Binomial model, $n(\hat{p}_1,\hat{p}_2)\sim \mathsf{B}(n,p_1)\times \mathsf{B}(n,p_2)$ with $(p_1,p_2)\in \Theta$, $(\hat{p}_1,\hat{p}_2)\in \hat{\Theta}$ and $\Theta\subset \hat{\Theta}$. In the Gaussian model, $\hat{\theta}\sim \mathcal{N}(\theta,\sigma^2I_2)$ with $\theta\in \Theta$, $\hat{\theta}\in\hat{\Theta}$ and $\Theta=\hat{\Theta}$.}
\label{fig.localization}
\end{figure}
Now we provide answers to these questions with the help of localization via confidence sets.
\begin{enumerate}
\item Question \ref{Q1}: When we consider the non-analytic region of $G(\cdot)$, we should always stick to the domain of $\hat{\theta}_n$ instead of that of the true parameter $\theta$ (for the existence of $G(\hat{\theta}_n)$, here we assume that $G(\cdot)$ is well-defined on the $\hat{\Theta}\supset \Theta$, where $\hat{\Theta}$ is the domain of $\hat{\theta}_n$). In fact, we should distinguish the ``smooth" (resp. ``non-smooth") regime of $\theta$ and that of $\hat{\theta}_n$: we determine the corresponding regimes of $\theta$ first, and then localize $\theta$ using $\hat{\theta}_n$ since $\theta$ cannot be observed. Hence, in the first step, to make the plug-in approach $G(\hat{\theta}_n)$ work for the estimation of $G(\theta)$, it must be ensured that with high probability $\hat{\theta}_n$ does not fall into the non-analytic region of $G(\cdot)$, which is defined over $\hat{\Theta}$ instead of $\Theta$. As a result, the non-analytic region of $G(\cdot)$ over the domain of $\hat{\theta}_n$ is the correct region to consider.
\item Question \ref{Q2}: We first determine the ``smooth" regime $\Theta_{\textrm{s}}$ of $\theta$. Let $\hat{\Theta}_0\subset \hat{\Theta}$ be the non-analytic region of $G(\cdot)$ over $\hat{\Theta}$. By the previous answer to Question \ref{Q1}, $\Theta_{\textrm{s}}$ should be set to
\begin{align}\label{eq.theta_smooth}
\Theta_{\textrm{s}} \triangleq \{\theta\in\Theta: \mathbb{P}_\theta(\hat{\theta}_n\in \hat{\Theta}_0) \le r_n\}
\end{align}
where the convergence rate $r_n$ (e.g., $r_n\asymp n^{-A}$ for some constant $A>0$) depends on the specific problem. Usually $r_n$ can be of any negligible order compared to the minimax risk of the estimation problem. With the help of localization via confidence sets, we can just set $\Theta_{\textrm{s}}=\Theta-\cup_{x\in \hat{\Theta}_0}U(x)$ for any confidence set $\{U(x)\}_{x\in \hat{\Theta}}$ of significance level $r_n$. In fact, if $\theta\in \Theta_{\textrm{s}}$ and $\hat{\theta}_n\in \hat{\Theta}_0$, we have $U(\hat{\theta}_n)\subset \cup_{x\in \hat{\Theta}_0}U(x)=\Theta-\Theta_{\textrm{s}}$ and thus $\theta \notin U(\hat{\theta}_n)$. As a result, by definition of the confidence set we have
\begin{align}
\sup_{\theta\in \Theta_{\textrm{s}}} \mathbb{P}_\theta(\hat{\theta}_n\in \hat{\Theta}_0) \le \sup_{\theta\in \Theta_{\textrm{s}}} \mathbb{P}_\theta(\theta\notin U(\hat{\theta}_n))\le \sup_{\theta\in \Theta} \mathbb{P}_\theta(\theta\notin U(\hat{\theta}_n)) \le r_n
\end{align}
as desired. By taking complement we obtain the ``non-smooth" regime $\Theta_{\textrm{ns}}\triangleq \Theta-\Theta_{\textrm{s}}$ of $\theta$.
Since we cannot observe $\theta$, we need to determine the ``smooth" regime $\hat{\Theta}_{\textrm{\rm s}}$ based on $\hat{\theta}_n$ rather than $\theta$. A natural choice is given by confidence set: $\hat{\Theta}_{\textrm{\rm s}}\triangleq \{\hat{\theta}_n\in \hat{\Theta}: \Theta_{\textrm{s}}\supset U(\hat{\theta}_n)\}$, i.e., $\hat{\Theta}_{\textrm{\rm s}}$ contains all observations whose confidence set for the true parameter falls into the ``smooth" regime $\Theta_{\textrm{s}}$. Likewise, we can define $\hat{\Theta}_{\textrm{\rm ns}}\triangleq \{\hat{\theta}_n\in\hat{\Theta}: \Theta_{\textrm{ns}}\supset U(\hat{\theta}_n)\}$ for the ``non-smooth" regime based on $\hat{\theta}_n$. Since $\Theta_{\textrm{s}}\cap \Theta_{\textrm{ns}}=\emptyset$, it can be easily seen that $\hat{\Theta}_{\textrm{\rm s}}\cap \hat{\Theta}_{\textrm{\rm ns}}=\emptyset$ as well, but one problem is that $\hat{\Theta}_{\textrm{\rm ns}} \cup \hat{\Theta}_{\textrm{\rm s}} \subsetneqq \hat{\Theta}$, i.e., some observation $\hat{\theta}_n$ is attributed to neither the ``non-smooth" regime nor the ``smooth" regime.
To solve this problem, we should expand $\Theta_{\textrm{s}}$ and $\Theta_{\textrm{ns}}$ a little bit to ensure that $\hat{\Theta}_{\textrm{\rm s}}$ and $\hat{\Theta}_{\textrm{\rm ns}}$ form a partition of $\hat{\Theta}$. In fact, this expansion can be done in many statistical models with satisfactory measure concentration properties (e.g., in Multinomial, Poisson and Gaussian models). Specifically, for some proper $r_n^{(1)}\ge r_n^{(2)}$ of order both negligible to that of the minimax risk, there exists confidence sets $\{U_1(x)\}_{x\in \hat{\Theta}}$ and $\{U_2(x)\}_{x\in \hat{\Theta}}$ of significance level $r_n^{(1)}$ and $r_n^{(2)}$, respectively, such that
\begin{align}
\Theta_{\textrm{s}}^{(1)} &\triangleq \Theta - \cup_{x\in\hat{\Theta}_0} U_1(x)\label{eq.para_smooth}\\
\Theta_{\textrm{ns}}^{(2)}&\triangleq \cup_{x\in \hat{\Theta}_0} U_2(x)\label{eq.para_nonsmooth}\\
\hat{\Theta}_{\textrm{\rm s}} &\triangleq \{\hat{\theta}_n\in \hat{\Theta}: \Theta_{\textrm{s}}^{(1)}\supset U_1(\hat{\theta}_n)\}\label{eq.est_smooth}\\
\hat{\Theta}_{\textrm{\rm ns}} &\triangleq \{\hat{\theta}_n\in \hat{\Theta}: \Theta_{\textrm{ns}}^{(2)}\supset U_2(\hat{\theta}_n)\}\label{eq.est_nonsmooth}
\end{align}
satisfy that $\hat{\Theta}_{\textrm{\rm s}}\cup \hat{\Theta}_{\textrm{\rm ns}}=\hat{\Theta}$ (by passing through subsets it does not matter if $\hat{\Theta}_{\textrm{\rm s}}\cap \hat{\Theta}_{\textrm{\rm ns}}\neq\emptyset$). Note that in this case we must have $\Theta_{\textrm{s}}^{(1)}\cap \Theta_{\textrm{ns}}^{(2)}\neq \emptyset$, i.e., there exists some $\theta$ which belongs to both the ``smooth" regime and the ``non-smooth" regime.
The interpretation of this approach is as follows. If the true parameter $\theta$ falls in the ``smooth" regime $\Theta_{\textrm{s}}^{(1)}$, then the plug-in approach will work; conversely, if the true parameter $\theta$ falls in the ``non-smooth" regime $\Theta_{\textrm{ns}}^{(2)}$, then the approximation idea will work. Then $\Theta_{\textrm{s}}^{(1)}\cap \Theta_{\textrm{ns}}^{(2)}\neq \emptyset$ implies that there exists an \emph{intermediate} regime such that both the plug-in approach and the approximation approach work when $\theta$ falls into this regime. This intermediate regime is unnecessary when we are given the partial information whether $\theta\in \Theta_{\textrm{s}}^{(1)}$ or $\theta\in \Theta_{\textrm{ns}}^{(2)}$, but it becomes important when we need to infer this partial information based on $\hat{\theta}_n$. Our target is as follows: if the true parameter $\theta$ does not fall in the ``smooth" (resp. ``non-smooth") regime, then with high probability we will also declare based on $\hat{\theta}_n$ that we are not in the ``smooth" (resp. ``non-smooth") regime. Mathematically, with high probability, $\theta\in \Theta-\Theta_{\textrm{s}}^{(1)}$ implies $\hat{\theta}_n\in \hat{\Theta}_{\textrm{\rm ns}}$, and $\theta\in \Theta-\Theta_{\textrm{ns}}^{(2)}$ implies $\hat{\theta}_n\in \hat{\Theta}_{\textrm{\rm s}}$. Note that if $\theta\in \Theta_{\textrm{s}}^{(1)}\cap \Theta_{\textrm{ns}}^{(2)}$ falls in the intermediate regime, either $\hat{\theta}_n\in \hat{\Theta}_{\textrm{\rm s}}$ or $\hat{\theta}_n\in \hat{\Theta}_{\textrm{\rm ns}}$ suffices for our estimator to perform well. The key fact is that this target is fulfilled by the definition of confidence sets: if $\theta\in \Theta-\Theta_{\textrm{s}}^{(1)}$ and $\hat{\theta}_n\notin \hat{\Theta}_{\textrm{\rm ns}}$, we have $\hat{\theta}_n\in \hat{\Theta}_{\textrm{\rm s}}$, and by definition of $\hat{\Theta}_{\textrm{\rm s}}$ we have $U_1(\hat{\theta}_n)\subset \Theta_{\textrm{s}}^{(1)}$, which implies $\theta\notin U_1(\hat{\theta}_n)$. As a result,
\begin{align}
\sup_{\theta\in \Theta-\Theta_{\textrm{s}}^{(1)}} \mathbb{P}_\theta(\hat{\theta}_n\notin \hat{\Theta}_{\textrm{\rm ns}}) \le \sup_{\theta\in \Theta-\Theta_{\textrm{s}}^{(1)}} \mathbb{P}_\theta(\theta\notin U_1(\hat{\theta}_n)) \le \sup_{\theta\in \Theta} \mathbb{P}_\theta(\theta\notin U_1(\hat{\theta}_n))\le r_n^{(1)}
\end{align}
and similarly $\sup_{\theta\in \Theta-\Theta_{\textrm{ns}}^{(2)}} \mathbb{P}_\theta(\hat{\theta}_n\notin \hat{\Theta}_{\textrm{\rm s}}) \le r_n^{(2)}$. Hence, we successfully localize $\theta$ via confidence sets based on $\hat{\theta}_n$ such that the true parameter $\theta$ is very likely to belong to the declared regime based on $\hat{\theta}_n$.
A pictorial illustration of this idea is shown in Figure \ref{fig.Q2}.
\begin{figure}[!htbp]
\centering
\begin{tikzpicture}[xscale=5, yscale=4]
\draw (0,0) rectangle (2,1.6);
\node [below right] at (0, 1.6) {$\hat{\Theta}=\Theta$};
\draw (1,0.8) ellipse (0.2 and 0.25);
\node at (1, 0.8) {$\hat{\Theta}_0$};
\draw [dashed] (1,0.8) ellipse (0.3 and 0.5);
\node at (1.02, 0.5) {$\Theta-\Theta_\textrm{\rm s}^{(1)}$};
\node at (1.0, 0.4) {I};
\node at (1.25, 0.35) {II};
\node at (1.5, 0.3) {III};
\node at (1.83, 0.2) {IV};
\node at (1.85, 0.3) {$\Theta - \Theta_{\textrm{\rm ns}}^{(2)}$};
\draw [dashed] (1,0.8) ellipse (0.9 and 0.7);
\filldraw (1.335, 1.17) circle (0.3pt);
\node [right] at (1.335, 1.17) {$\hat{\theta}_1$};
\draw [dashed] (1.335, 1.17) ellipse (0.38 and 0.2);
\node [below] at (1.335, 1.15) {$U_1(\hat{\theta}_1)$};
\filldraw (0.455, 1.24) circle (0.3pt);
\node [right] at (0.455, 1.24) {$\hat{\theta}_2$};
\draw [dashed] (0.455, 1.24) ellipse (0.37 and 0.2);
\node [above] at (0.455, 1.26) {$U_2(\hat{\theta}_2)$};
\draw [thick] (1,0.8) ellipse (0.6 and 0.6);
\node at (1.45, 0.8) {$\hat{\Theta}_{\textrm{\rm ns}}$};
\node at (1.75, 0.8) {$\hat{\Theta}_\textrm{\rm s}$};
\end{tikzpicture}
\caption{Pictorial explanation of the ``smooth" and ``non-smooth" regimes based on $\theta$ and $\hat{\theta}_n$, respectively. In the above figure, we have $\Theta_\textrm{\rm s}^{(1)}=\text{II}\cup\text{III}\cup\text{IV}$, $\Theta_\textrm{\rm ns}^{(2)}=\text{I}\cup\text{II}\cup\text{III}$, $\hat{\Theta}_{\textrm{\rm s}}=\text{III}\cup\text{IV}$, and $\hat{\Theta}_{\textrm{\rm ns}}=\text{I}\cup\text{II}$. In particular, $\Theta_\textrm{\rm s}^{(1)}\cap \Theta_\textrm{\rm ns}^{(2)}=\text{II}\cup\text{III}$ is the intermediate regime, where both the plug-in and the approximation approach are performing well. For $\hat{\theta}_1\in\hat{\Theta}_\textrm{\rm ns}$, we have $U_1(\hat{\theta}_1)\subset \Theta_\textrm{\rm ns}^{(2)}$; for $\hat{\theta}_2\in\hat{\Theta}_\textrm{\rm s}$, we have $U_2(\hat{\theta}_2)\subset \Theta_\textrm{\rm s}^{(1)}$.}
\label{fig.Q2}
\end{figure}
\item Question \ref{Q3}: Given a confidence set $\{U(x)\}_{x\in\hat{\Theta}}$ of a satisfactory significance level $r_n$, after observing $\hat{\theta}_n\in \hat{\Theta}_{\textrm{\rm ns}}$ we can always set the approximation region to be $U(\hat{\theta}_n)$. Note that $U(\hat{\theta}_n)\subset \Theta_{\textrm{ns}}$ by definition, and in fact $U(\hat{\theta}_n)$ can be considerably smaller than $\Theta_{\textrm{ns}}$, which makes it a desirable regime to approximate over rather than $\Theta_\textrm{\rm ns}$ and is proved to be necessary in \cite{jiao2016minimax}. The reason why $U(\hat{\theta}_n)$ is sufficient is as follows: by definition of confidence sets we have $\sup_{\theta\in\Theta} \mathbb{P}_\theta(\theta\in U(\hat{\theta}_n))\le r_n$, hence with probability at least $1-r_n$, the approximation region $U(\hat{\theta}_n)$ based on $\hat{\theta}_n$ covers $\theta$, which allows us to operate as if $\theta$ is conditioned to be inside $U(\hat{\theta})$. Note that in order to obtain a good approximation performance, we need to find a confidence set $\{U(x)\}_{x\in\hat{\Theta}}$ as small as possible, which depends on the statistical model.
\item Question \ref{Q4}: there has been a long history of correcting the bias of the MLE based on Taylor expansion. For example, in entropy estimation, one of the earliest investigations on reducing the bias of MLE in entropy estimation is due to Miller \cite{Miller1955}. Interestingly, it was already observed in 1969 by Carlton \cite{Carlton1969bias} that Miller's bias correction formula should only be applied when $p\gg 1/n$, which is automatically satisfied when $p$ belongs to the ``smooth" regime $[\frac{\ln n}{n},1]$ defined in \cite{Jiao--Venkat--Han--Weissman2015minimax}. As a result, in the ``smooth" regime, Miller's idea was used in \cite{Jiao--Venkat--Han--Weissman2015minimax}. In our generalization of the ``smooth" regime, by definition of $\Theta_{\textrm{s}}$, $G^{(T)}(\hat{\theta}_n)$ remains bounded with high probability for any order $T>0$ and $\theta\in\Theta_{\textrm{s}}$. Hence, it shows that Miller's bias-correction approach based on Taylor expansion can also be used in general. However, Miller's approach fails when high-order bias correction is desired, or equivalently, when $T$ is large. To see why it is the case, we take a look at the procedure considered in \cite{Jiao--Venkat--Han--Weissman2015minimax}. For Binomial random variable $X\sim\mathsf{B}(n,p)$, denote the empirical frequency by $\hat{p}=\frac{X}{n}$. Then it follows from Taylor's theorem that
\begin{align}
\mathbb{E} f(\hat{p}) - f(p) = \frac{1}{2}f''(p)\mathsf{Var}_p(\hat{p}) + O(\frac{1}{n^2}) = \frac{p(1-p)}{2n}f''(p) + O(\frac{1}{n^2})
\end{align}
where $f''(p)$ is the second-order derivative of $f$ at $p$. Hence, the bias-corrected estimator in \cite{Jiao--Venkat--Han--Weissman2015minimax} was proposed as follows:
\begin{align}
f^c(\hat{p}) = f(\hat{p}) - \frac{f''(\hat{p})\hat{p}(1-\hat{p})}{2n}.
\end{align}
However, the plug-in approach was still used for the bias-correction term in the previous estimator, which should be further corrected based on Taylor expansion again in order to achieve higher-order bias correction. Continuing this approach, the further correction still suffers from the same problem and additional corrections need to be done, and so on and so forth. As a result, the previous bias-correction fails to be generalized to high-order corrections, and a successful bias-correction approach should avoid employing the plug-in approach for bias-correction terms.
One way to avoid the plug-in approach is as follows: instead of doing Taylor expansion of $G(\hat{\theta}_n)$ near $\theta$, we employ Taylor expansion of $G(\theta)$ near $\hat{\theta}_n$ as
\begin{align}
G(\theta) \approx \sum_{k=0}^T \frac{G^{(k)}(\hat{\theta}_n)}{k!}(\theta-\hat{\theta}_n)^k.
\end{align}
The advantage is that, now $G^{(k)}(\hat{\theta}_n)$ is by definition an unbiased estimator of $\mathbb{E}_\theta[G^{(k)}(\hat{\theta}_n)]$. However, the unknown $\theta$ in the RHS still prevents us from using this estimator explicitly. Fortunately, this difficulty can be overcome by the standard sample splitting approach: we split samples to obtain independent $\hat{\theta}_n^{(1)}$ and $\hat{\theta}_n^{(2)}$, both of which follow the same class of distribution (with possibly different parameters) as $\hat{\theta}_n$. We remark that sample splitting can be employed for divisible distributions, including Multinomial, Poisson and Gaussian models \cite{nemirovski2000topics}, and it will be discussed in detail at the beginning of Section \ref{sec.construction} for Poisson models. Now our bias-corrected estimator is
\begin{align}\label{eq.bias_cor}
\hat{G}_{\textrm{s}}(\hat{\theta}_n) = \sum_{k=0}^T \frac{G^{(k)}(\hat{\theta}_n^{(1)})}{k!}\sum_{j=0}^k \binom{k}{j}S_j(\hat{\theta}_n^{(2)})(-\hat{\theta}_n^{(1)})^{k-j}
\end{align}
where $S_j(\hat{\theta}_n^{(2)})$ is an unbiased estimator of $\theta^j$ (which usually exists). Now it is straightforward to show that
\begin{align}
\mathbb{E}[\hat{G}_{\textrm{s}}(\hat{\theta}_n)] - G(\theta) = \mathbb{E}_\theta\left[\sum_{k=0}^T \frac{G^{(k)}(\hat{\theta}_n^{(1)})}{k!}(\theta-\hat{\theta}_n^{(1)})^k - G(\theta) \right]= O\left(\mathbb{E}_\theta\left|\frac{G^{(T+1)}(\xi)}{(T+1)!}(\hat{\theta}_n^{(1)}-\theta)^{T+1}\right|\right)
\end{align}
i.e., the estimator in \eqref{eq.bias_cor} achieves the bias-correction of any desired order. Although in many scenarios (those in \cite{Jiao--Venkat--Han--Weissman2015minimax,Cai--Low2011,Valiant--Valiant2011power,Valiant--Valiant2013estimating,Wu--Yang2014minimax,jiao2016minimax}) no bias correction or only order-one bias correction is required in the ``smooth" regime, bias correction of an arbitrary order turns out to be crucial in our recent work on the estimation of nonparametric functionals \cite{Han--Jiao--Mukherjee--Weissman2016optimal}. We also conjecture that this approach is crucial for the construction of the minimax rate-optimal estimator for the R\'{e}nyi entropy in the large alphabet setting, which \cite{Acharya--Orlitsky--Suresh--Tyagi2014complexity} did not address completely.
\end{enumerate}
The answers to these questions shed light on the detailed implementation of the general recipe and give rise to the important concept of localization via confidence sets, which leads us to propose a refined two-step approach. As before, denote by $\hat{\Theta}\supset \Theta$ the set containing all possible values of the estimator $\hat{\theta}_n$, and by $\hat{\Theta}_0\subset \hat{\Theta}$ the set on which $G(\cdot)$ is non-analytic. Let $\{U(x)\}_{x\in\hat{\Theta}}$ be a satisfactory confidence set.
\begin{enumerate}
\item \textbf{Classify the Regime}:
\begin{itemize}
\item For the true parameter $\theta$, declare that $\theta$ is in the ``non-smooth" regime if $\theta$ is ``close" enough to $\hat{\Theta}_0$ in terms of localization via confidence sets (cf. (\ref{eq.para_nonsmooth})). Otherwise declare $\theta$ is in the ``smooth" regime (cf. (\ref{eq.para_smooth}));
\item Compute $\hat{\theta}_n$, and declare that we are in the ``non-smooth'' regime if the confidence set of $\hat{\theta}_n$ falls into the ``non-smooth" regime of $\theta$ (cf. (\ref{eq.est_nonsmooth})). Otherwise declare we are in the ``smooth'' regime (cf. (\ref{eq.est_smooth}));
\end{itemize}
\item {\bf Estimate}:
\begin{itemize}
\item If $\hat{\theta}_n$ falls in the ``smooth'' regime, use an estimator ``similar'' to $G(\hat{\theta}_n)$ to estimate $G(\theta)$;
\item If $\hat{\theta}_n$ falls in the ``non-smooth'' regime, replace the functional $G(\theta)$ in the ``non-smooth'' regime by an approximation $G_{\text{appr}}(\theta)$ (another functional which well approximates $G(\theta)$ on $U(\hat{\theta}_n)$) which can be estimated without bias, then apply an unbiased estimator for the functional $G_{\text{appr}}(\theta)$.
\end{itemize}
\end{enumerate}
In this paper, we follow the refined recipe for the construction of our optimal estimator in estimating several divergences between discrete distributions, including the KL divergence, Hellinger distance and $\chi^2$-divergence, where only the KL divergence will be discussed in detail. Moreover, in the estimation of KL divergence, we will encounter a new phenomenon, i.e., \emph{multivariate approximation in polytopes}, which is a highly non-trivial topic in approximation theory, and will also propose a general tool to analyze the risk of the bias-corrected plug-in approach with the help of localization via confidence sets.
The rest of this paper is organized as follows. We first analyze the performance of the modified plug-in estimator and prove Theorem \ref{thm.mle} in Section \ref{sec.mle}. In Section \ref{sec.construction}, we first follow the general recipe to explicitly construct our estimator for the KL divergence step by step, and show that it essentially achieves the bound in Theorem \ref{thm.kl_divergence}. Then we adopt and adapt some tricks to construct another estimator which is rate-optimal, adaptive and easier to implement. The minimax lower bound for estimating the KL divergence is proved in Section \ref{sec.lowerbound}. For the Hellinger distance and the $\chi^2$-divergence, we sketch the construction of the respective minimax rate-optimal estimators in Section \ref{sec.others}. Conclusions are drawn in Section \ref{sec.conclusion}, and complete proofs of the remaining theorems and lemmas are provided in the appendices. The Matlab code of estimating KL divergence has been released on \url{http://www.stanford.edu/~tsachy/index_hjw}.
\section{Performance of the modified plug-in approach}\label{sec.mle}
In this section, we give the upper bound and the lower bound of the worst-case mean squared error via the modified plug-in approach, i.e., we prove Theorem \ref{thm.mle}. Throughout our analysis, we utilize the Poisson sampling model, i.e., each component $X_i$ (resp. $Y_i$) in the histogram $\mathbf{X}$ (resp. $\mathbf{Y}$) has distribution $\mathsf{Poi}(mp_i)$ (resp. $\mathsf{Poi}(nq_i)$), and all coordinates of $\mathbf{X}$ (resp. $\mathbf{Y}$) are independent. In other words, instead of drawing fixed sample sizes $m$ and $n$, there are i.i.d. samples from distributions $P,Q$ of sizes $M\sim\mathsf{Poi}(m)$ and $N\sim\mathsf{Poi}(n)$, respectively. Consequently, the observed number of occurrences of each symbol are independent \cite[Theorem 5.6]{Mitzenmacher--Upfal2005probability}. We note that the Poisson sampling model is essentially the same as the Multinomial model, and their minimax risks are related via Lemma \ref{lemma.poissonmultinomial} in Appendix \ref{appen.aux}.
\subsection{Proof of upper bounds}
Recall that the empirical distribution $Q_n$ has been modified to
\begin{align}
Q_n' = \left(\frac{1}{n}\vee\hat{q}_1,\cdots,\frac{1}{n}\vee\hat{q}_S\right)
\end{align}
the modified plug-in estimator $D(P_m\|Q_n')$ is not the exact plug-in approach. However, it can be observed that this quantity
\begin{align}
D(P_m\|Q_n') = \sum_{i=1}^S \left[\hat{p}_i\ln \hat{p}_i - \hat{p}_i\ln \left(\frac{1}{n}\vee\hat{q}_i\right) \right]
\end{align}
is close to the following natural plug-in estimator
\begin{align}
D_1(P_m\|Q_n) = \sum_{i=1}^S \left[\hat{p}_i\ln \hat{p}_i - \hat{p}_ig_n(\hat{q}_i) \right],
\end{align}
where
\begin{align}
g_n(q) \triangleq \begin{cases}
-(1+\ln n) + nq & \text{if }0\le q< \frac{1}{n},\\
\ln q& \text{if }\frac{1}{n}\le q\le 1.
\end{cases}
\end{align}
In view of this fact, we can apply the general approximation-based method in \cite{Jiao--Venkat--Weissman2014MLE} to analyze the performance of the plug-in approach.
By construction it is obvious that $g_n(q)$ is continuously differentiable on $[0,1]$, which coincides with $g(q)=\ln q$ on $[\frac{1}{n},1]$. Moreover, since $\hat{q}_i$ is a multiple of $\frac{1}{n}$, $g_n(\hat{q}_i)$ only differs from $\ln(\frac{1}{n}\vee\hat{q}_i)$ at $\hat{q}_i=0$ by $|-(1+\ln n)-\ln(1/n)|=1$. Hence, we may consider the performance of the plug-in estimator $\hat{p}_i(\ln \hat{p}_i-g_n(\hat{q}_i))$ in estimating $p(\ln p-g_n(q))$, which is summarized in the following lemma.
\begin{lemma}\label{lem.mle}
Let $n\hat{p}\sim\mathsf{Poi}(mp)$ and $n\hat{q}\sim\mathsf{Poi}(nq)$ be independent, and $p\le u(S)q$. Then we have
\begin{align}
|\mathbb{E}[\hat{p}(\ln\hat{p}-g_n(\hat{q}))] - p(\ln p-g_n(q))| &\le \frac{30u(S)}{n}+\frac{5\ln 2}{m},\\
\mathsf{Var}(\hat{p}(\ln\hat{p}-g_n(\hat{q})) &\le \frac{51}{m^2} + \frac{2}{m}\left(p+p(\ln u(S))^2+\frac{4q}{e^2}+\frac{4u(S)}{en}\right) + \frac{700u(S)}{n}\left(p+\frac{1}{m}\right).
\end{align}
In particular,
\begin{align}
|\mathbb{E}[\hat{p}(\ln\hat{p}-g_n(\hat{q}))] - p(\ln p-g_n(q))| &\lesssim \frac{u(S)}{n} + \frac{1}{m},\\
\mathsf{Var}(\hat{p}(\ln \hat{p}-g_n(\hat{q}))) &\lesssim \frac{1}{m^2} + \frac{pu(S)}{n} + \frac{u(S)}{mn} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m}.
\end{align}
\end{lemma}
Hence, by Lemma \ref{lem.mle}, we conclude that
\begin{align}
\left|\mathbb{E} D_1(P_m\|Q_n) - D_1(P\|Q)\right| &\le \sum_{i=1}^S |\mathbb{E}[\hat{p_i}(\ln \hat{p}_i-g_n(\hat{q_i}))] - p_i(\ln p_i-g_n(q_i))|\\
&\lesssim \sum_{i=1}^S \frac{u(S)}{n} + \frac{1}{m}\\
&= \frac{Su(S)}{n} + \frac{S}{m}
\end{align}
and
\begin{align}
\mathsf{Var}(D_1(P_m\|Q_n)) &= \sum_{i=1}^S \mathsf{Var}(\hat{p}_ig_n(\hat{q}_i))\\
&\lesssim \sum_{i=1}^S \frac{1}{m^2}+\frac{p_iu(S)}{n} + \frac{u(S)}{mn} + \frac{p_i(1+\ln u(S))^2}{m} + \frac{q_i}{m}\\
&\lesssim \frac{S}{m^2} + \frac{u(S)}{n} + \frac{Su(S)}{mn} + \frac{(1+\ln u(S))^2}{m}\\
&\lesssim \frac{S}{m^2}+\frac{u(S)}{n} + \left(\frac{(Su(S))^2}{n^2} + \frac{1}{m^2}\right) + \frac{(\ln u(S))^2}{m}\\
&\lesssim \frac{S}{m^2}+\frac{u(S)}{n} + \frac{(Su(S))^2}{n^2} + \frac{(\ln u(S))^2}{m}.
\end{align}
Combining these two inequalities yields, for any $(P,Q)\in \mathcal{U}_{S,u(S)}$,
\begin{align}
\mathbb{E}\left(D_1(P_m\|Q_n)-D_1(P\|Q)\right)^2 &=\left|\mathbb{E} D_1(P_m\|Q_n) - D_1(P\|Q)\right|^2+\mathsf{Var}(D_1(P_m\|Q_n))\\
&\lesssim \left(\frac{Su(S)}{n}+\frac{S}{m}\right)^2 + \frac{u(S)}{n} + \frac{(\ln u(S))^2}{m}.
\end{align}
To prove Theorem \ref{thm.mle}, it remains to compute the difference between $D$ and $D_1$. By the definition of $g_n(\cdot)$, we have
\begin{align}
\mathbb{E}|D_1(P_m\|Q_n) - D(P_m\|Q_n')|^2 &= \mathbb{E}\left|\sum_{i=1}^S\hat{p}_i\mathbbm{1}(\hat{q}_i=0)\right|^2\\
&\le S\sum_{i=1}^S \mathbb{E}|\hat{p}_i\mathbbm{1}(\hat{q}_i=0)|^2 \\
&= S\sum_{i=1}^S \left(p_i^2+\frac{p_i}{m}\right)e^{-nq_i}\\
&\le S\sum_{i=1}^S \left((u(S))^2 q_i^2e^{-nq_i} + \frac{u(S)}{m}q_ie^{-nq_i}\right)\\
&\le S\sum_{i=1}^S \left((u(S))^2 \left(\frac{2}{en}\right)^2 + \frac{u(S)}{enm}\right)\\
&\lesssim \left(\frac{Su(S)}{n}\right)^2 + \frac{S^2u(S)}{mn}\\
&\lesssim \left(\frac{Su(S)}{n}+\frac{S}{m}\right)^2
\end{align}
where we have used the fact that
\begin{align}
\sup_{x\in [0,1]} x^ke^{-nx} = \left(\frac{k}{en}\right)^k.
\end{align}
Moreover, for any $(P,Q)\in\mathcal{U}_{S,u(S)}$,
\begin{align}
|D_1(P\|Q)-D(P\|Q)| &\le \sum_{i=1}^S p_i|-(\ln n+1)+nq_i-\ln q_i|\mathbbm{1}(q_i<\frac{1}{n})\\
&\le \sum_{i=1}^S p_i(1-\ln(nq_i))\mathbbm{1}(q_i<\frac{1}{n})\\
&\le u(S)\cdot \sum_{i=1}^S q_i(1-\ln(nq_i))\mathbbm{1}(q_i<\frac{1}{n})\\
&\lesssim \frac{Su(S)}{n}
\end{align}
where we have used that $\sup_{q\in [0,1/n]}q(1-\ln(nq))=1/n$. Hence, by the triangle inequality, for any $(P,Q)\in\mathcal{U}_{S,u(S)}$, we have
\begin{align}
&\mathbb{E}\left(D(P_m\|Q_n')-D(P\|Q)\right)^2 \nonumber\\
&\qquad \le 3\left(\mathbb{E}\left(D_1(P_m\|Q_n)-D_1(P\|Q)\right)^2+\mathbb{E}\left(D_1(P_m\|Q_n)-D(P_m\|Q_n')\right)^2+|D_1(P\|Q)-D(P\|Q)|^2\right)\\
&\qquad\lesssim \left(\frac{Su(S)}{n}+\frac{S}{m}\right)^2 + \frac{u(S)}{n} + \frac{(\ln u(S))^2}{m}
\end{align}
which completes the proof of the upper bound in Theorem \ref{thm.mle}.
\subsection{Proof of lower bounds}
By the bias-variance decomposition of the mean squared error, to prove that the squared term in Theorem \ref{thm.mle} serves as a lower bound, it suffices to find some $(P,Q)\in \mathcal{U}_{S,u(S)}$ such that
\begin{align}\label{eq.mle_lower_bias}
|\mathbb{E}_{(P,Q)} D(P_m\|Q_n') - D(P\|Q)|&\gtrsim \frac{Su(S)}{n}+\frac{S}{m}.
\end{align}
Note that here we prove this inequality based on the Multinomial model, and then obtain the result for the Poisson sampling model via Lemma \ref{lemma.poissonmultinomial}. The construction of $(P,Q)$ is as follows: $P=(\frac{1}{S},\cdots,\frac{1}{S})$ is the uniform distribution, and $Q=(\frac{1}{Su(S)},\cdots,\frac{1}{Su(S)},1-\frac{S-1}{Su(S)})$ is near-uniform. We first recall from \cite{Jiao--Venkat--Weissman2014MLE} that, if $m\ge 15S$, we have
\begin{align}\label{eq.mle_lower_first}
\sum_{i=1}^S [\mathbb{E}[\hat{p}_i\ln \hat{p}_i]-p_i\ln p_i] \ge \frac{S-1}{2m} - \frac{S^2}{20m^2} - \frac{1}{12m^2}.
\end{align}
Next we give a lower bound for the term $\mathbb{E}(-\ln (\hat{q}_i\vee\frac{1}{n}))-(-\ln q_i)$ for $q_i\ge \frac{4}{n}$. We shall use the following lemma for the approximation error of the Bernstein polynomial, which corresponds to the bias in the Multinomial model. Define the Bernstein operator $B_n$ as follows:
\begin{align}
B_n[f](x) = \sum_{i=0}^n \binom{n}{i}x^i(1-x)^{n-i}\cdot f\left(\frac{i}{n}\right), \qquad f\in C[0,1].
\end{align}
\begin{lemma}\label{lem.bernstein}
\cite{Braess--Sauer2004} Let $k\ge 4$ be an even integer. Suppose that the $k$-th derivative of $f$ satisfies that $f^{(k)}\le 0$ on $(0,1)$, and $Q_{k-1}$ is the Taylor polynomial of order $k-1$ of $f$ at some point $x_0$. Then for $x\in [0,1]$,
\begin{align}
f(x) - B_n[f](x) \ge Q_{k-1}(x) - B_n[Q_{k-1}](x).
\end{align}
\end{lemma}
Since our modification of $\ln(\cdot)$ is not even differentiable, Lemma \ref{lem.bernstein} cannot be applied directly. However, we can consider the following function instead:
\begin{align}
h_n(x) = \begin{cases}
-\ln n+n(x-\frac{1}{n})-\frac{n^2}{2}(x-\frac{1}{n})^2+\frac{n^3}{3}(x-\frac{1}{n})^3-\frac{n^4}{4}(x-\frac{1}{n})^4& \text{if }0\le x<\frac{1}{n},\\
\ln x& \text{if }\frac{1}{n}\le x\le 1.
\end{cases}
\end{align}
By construction it is obvious that $h_n(x)\in C^4[0,1]$ which coincides with $\ln x$ on $[\frac{1}{n},1]$. Moreover, $h_n(\hat{q})$ only differs from $\ln(\frac{1}{n}\vee\hat{q}) $ at zero by $|h_n(0)+\ln n|= \frac{25}{12}$. Since $h_n^{(4)}(x)\le 0$, Lemma \ref{lem.bernstein} can be applied here to yield the following lemma.
\begin{lemma}\label{lem.lower_mle}
For $\frac{4}{n}\le x\le 1$, we have
\begin{align}
h_n(x) - B_n[h_n](x) \ge \frac{(1-x)((n+4)x-2)}{n^2x^2} > 0.
\end{align}
\end{lemma}
Since our assumption $n\ge 4Su(S)$ ensures that for our choice of $Q$, $q_i\ge \frac{4}{n}$ for any $i$. Hence, by Lemma \ref{lem.lower_mle} and the concavity of $h_n(\cdot)$, we have
\begin{align}
\sum_{i=1}^S [p_i\ln q_i - \mathbb{E}[\hat{p}_ih_n(\hat{q}_i)]]&\ge \sum_{i=1}^{S-1} \frac{1}{S}\left[\ln q_i - \mathbb{E}[h_n(\hat{q}_i)]\right] \\
&\ge \frac{S-1}{S}\cdot \frac{(Su(S))^2}{n^2}\left(1-\frac{1}{Su(S)}\right)\left(\frac{n+4}{Su(S)}-2\right)\\
&\ge \frac{S-1}{S}\cdot \frac{(Su(S))^2}{n^2}\left(1-\frac{1}{Su(S)}\right)\cdot \frac{n}{2Su(S)}\\
&= \frac{(S-1)u(S)}{2n}\left(1-\frac{1}{Su(S)}\right).
\end{align}
Now note that
\begin{align}
\sum_{i=1}^S |\mathbb{E}[\hat{p}_ih_n(\hat{q}_i)] - \mathbb{E}[\hat{p}_i\ln (\frac{1}{n}\vee\hat{q}_i)]|&\le \frac{25}{12}\sum_{i=1}^S p_i\cdot \mathbb{P}(\hat{q}_i=0) \\
&= \frac{25}{12}\sum_{i=1}^S p_i(1-q_i)^n \\
&\le \frac{25}{12}u(S)\sum_{i=1}^S q_i(1-q_i)^n \\
&\le \frac{25Su(S)}{3ne^4}
\end{align}
where we have used the fact that
\begin{align}
\sup_{x\in [\frac{4}{n},1]} x(1-x)^n = \frac{4}{n}\left(1-\frac{4}{n}\right)^n \le \frac{4}{ne^4}.
\end{align}
A combination of these two inequalities yields
\begin{align}
\sum_{i=1}^S \left[\mathbb{E}[-\hat{p}_i\ln (\frac{1}{n}\vee\hat{q}_i)] - (-p_i\ln q_i)\right] &\ge \sum_{i=1}^S [p_i\ln q_i - \mathbb{E}[\hat{p}_ih_n(\hat{q}_i)]] - \sum_{i=1}^S |\mathbb{E}[\hat{p}_ih_n(\hat{q}_i)] - \mathbb{E}[\hat{p}_i\ln (\frac{1}{n}\vee\hat{q}_i)]|\\\label{eq.mle_lower_second}
&\ge \frac{(S-1)u(S)}{2n}\left(1-\frac{1}{Su(S)}\right) -\frac{25Su(S)}{3ne^4}.
\end{align}
Hence, when $m\ge 15S$ and $n\ge 4Su(S)$, combining (\ref{eq.mle_lower_first}) and (\ref{eq.mle_lower_second}) gives
\begin{align}
|\mathbb{E}_{(P,Q)} D(P_m\|Q_n') - D(P\|Q)| &\ge \mathbb{E}_{(P,Q)} D(P_m\|Q_n') - D(P\|Q) \\
&= \sum_{i=1}^S [\mathbb{E}[\hat{p}_i\ln \hat{p}_i]-p_i\ln p_i] + \sum_{i=1}^S \left[\mathbb{E}[-\hat{p}_i\ln (\frac{1}{n}\vee\hat{q}_i)] - (-p_i\ln q_i)\right]\\
&\ge \frac{S-1}{2m} - \frac{S^2}{20m^2} - \frac{1}{12m^2} + \frac{(S-1)u(S)}{2n}\left(1-\frac{1}{Su(S)}\right) -\frac{25Su(S)}{3ne^4}
\end{align}
which gives (\ref{eq.mle_lower_bias}), as desired.
For the remaining terms, we remark that
\begin{align}
\sup_{(P,Q)\in\mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left(\hat{D}-D(P\|Q)\right)^2 \gtrsim \frac{(\ln u(S))^2}{m} + \frac{u(S)}{n}
\end{align}
holds for any estimator $\hat{D}$ (and thus for the modified plug-in estimator $D(P_m\|Q_n')$), and we postpone the proof to Section \ref{sec.lowerbound}. Now the proof of Theorem \ref{thm.mle} is complete.
\section{Construction of the Optimal Estimator}\label{sec.construction}
We stay with the Poisson sampling model in this section. For simplicity of analysis, we conduct the classical ``splitting'' operation \cite{Tsybakov2013aggregation} on the Poisson random vector $\mathbf{X}$, and obtain three independent identically distributed random vectors $\mathbf{X}_j = [X_{1,j},X_{2,j},\ldots,X_{S,j}]^T, j\in\{1,2,3\}$, such that each component $X_{i,j}$ in $\mathbf{X}_j$ has distribution $\mathsf{Poi}(mp_i/3)$, and all coordinates in $\mathbf{X}_j$ are independent. For each coordinate $i$, the splitting process generates a random sequence $\{T_{ik}\}_{k=1}^{X_i}$ such that $\{T_{ik}\}_{k=1}^{X_i}|\mathbf{X} \sim \mathsf{Multinomial}(X_i; (1/3,1/3,1/3))$, and assign $X_{i,j}=\sum_{k=1}^{X_i} \mathbbm{1}(T_{ik}=j)$ for $j\in\{1,2,3\}$. All the random variables $\{ \{T_{ik}\}_{k=1}^{X_i}:1\leq i\leq S\}$ are conditionally independent given our observation $\mathbf{X}$. The splitting operation is similarly conducted for the Poisson random vector $\mathbf{Y}$.
For simplicity, we re-define $m/3$ as $m$ and $n/3$ as $n$, and denote
\begin{equation}
\hat{p}_{i,j} = \frac{X_{i,j}}{m}, \quad\hat{q}_{i,j} = \frac{Y_{i,j}}{n}, \qquad i\in\{1,2,\cdots,S\},j\in\{1,2,3\}.
\end{equation}
We remark that the ``splitting'' operation is not necessary in implementation. We also note that for independent random variables $(X,Y)$ such that $nX \sim \mathsf{Poi}(mp), nY\sim \mathsf{Poi}(nq)$,
\begin{equation}\label{eq.monomial_unbiased}
\mathbb{E} \prod_{r = 0}^{k-1} \left( X - \frac{r}{m} \right)\prod_{s=0}^{l-1} \left( Y - \frac{s}{n} \right) = p^kq^l,
\end{equation}
for any $k,l \in \mathbb{N}$. For a proof of this fact we refer to Withers~\cite[Example 2.8]{Withers1987}.
\subsection{Estimator construction}
Now we apply our general recipe to construct the estimator. Note that
\begin{align}
D(P\|Q) = \sum_{i=1}^S p_i\ln\frac{p_i}{q_i} = \sum_{i=1}^S p_i\ln p_i - \sum_{i=1}^S p_i\ln q_i = - H(P) - \sum_{i=1}^S p_i\ln q_i
\end{align}
where $H(P)=\sum_{i=1}^S -p_i\ln p_i$ is the entropy function. Hence, the optimal estimator $\hat{H}$ for entropy \cite{Valiant--Valiant2011power,Valiant--Valiant2013estimating,Wu--Yang2014minimax,Jiao--Venkat--Han--Weissman2015minimax} can be used here and it remains to estimate the cross entropy $\sum_{i=1}^S p_i\ln q_i$, i.e., our target is the bivariate function $f(p,q)=p\ln q$.
We first classify the regime. For the bivariate function $f(p,q)=p\ln q$, the entire parameter set is $\Theta=\{(p,q)\in[0,1]^2: p\le u(S)q\}$, and the function is analytic everywhere except for the point $\Theta_0=\{(0,0)\}$. For all possible values of the estimator $(\hat{p},\hat{q})$, we have $\hat{\Theta}=[0,1]^2$, and non-analytic points are $\hat{\Theta}_0=[0,1]\times \{0\}$. For the confidence set of this two-dimensional Poisson model $(m\hat{p},n\hat{q})\sim \mathsf{Poi}(mp)\times \mathsf{Poi}(nq)$, we can set $r_n\asymp n^{-A}$ for some universal constant $A>0$ and use
\begin{align}
U(x,y) = \Theta \cap \left(\begin{cases}
[0, \frac{c\ln m}{m}] & \text{if } x\le \frac{c\ln m}{m},\\
[x-\sqrt{\frac{cx\ln m}{m}}, x+\sqrt{\frac{cx\ln m}{m}}] & \text{if } \frac{c\ln m}{m} < x \le 1.
\end{cases} \times \begin{cases}
[0, \frac{c\ln n}{n}] & \text{if } y\le \frac{c\ln n}{n},\\
[y-\sqrt{\frac{cy\ln n}{n}}, y+\sqrt{\frac{cy\ln n}{n}}] & \text{if } \frac{c\ln n}{n} < y \le 1.
\end{cases} \right)
\end{align}
for some constant $c>0$. Hence, by choosing $c=c_1/2$ and $c=2c_1$ respectively in (\ref{eq.para_smooth}) and (\ref{eq.para_nonsmooth}) for some universal constant $c_1>0$ to be specified later, we get the ``smooth" and ``non-smooth" regimes for $(p,q)$ as (for brevity we omit the superscripts in \eqref{eq.para_smooth} and \eqref{eq.para_nonsmooth})
\begin{align}
\Theta_{\textrm{s}} &= \Theta \cap \left([0,1]\times [\frac{c_1\ln n}{2n},1]\right) = \left\{ (p,q)\in [0,1]^2: \frac{c_1\ln n}{2n}\le q\le 1, p\le u(S)q \right\}\\
\Theta_{\textrm{ns}} &= \Theta \cap \left([0,1]\times [0,\frac{2c_1\ln n}{n}]\right) = \left\{ (p,q)\in [0,1]^2: 0\le q\le\frac{2c_1\ln n}{n}, p\le u(S)q \right\}.
\end{align}
Further, by (\ref{eq.est_smooth}) and (\ref{eq.est_nonsmooth}), the ultimate ``smooth" and ``non-smooth" regimes are given by
\begin{align}
\hat{\Theta}_{\textrm{\rm s}} = [0,1] \times [\frac{c_1\ln n}{n},1],\qquad \hat{\Theta}_{\textrm{\rm ns}} = [0,1] \times [0,\frac{c_1\ln n}{n}]
\end{align}
i.e., we are in the ``non-smooth" regime if $\hat{q}\le \frac{c_1\ln n}{n}$, and are in the ``smooth" regime otherwise.
Next we construct the estimator in each regime. First, if we are in the ``smooth" regime, our bias-corrected estimator \eqref{eq.bias_cor} of order $T=3$ becomes
\begin{align}
T^{(3)}(\hat{q}_1,\hat{q}_2) &= \sum_{k=0}^3 \frac{g^{(k)}(\hat{q}_1)}{k!}\sum_{j=0}^k \binom{k}{j} S_j(\hat{q}_2)(-\hat{q}_1)^{k-j}\\ \label{eq.plugin_taylor}
&= \ln \hat{q}_1 + \frac{\hat{q}_2-\hat{q}_1}{\hat{q}_1} - \frac{(\hat{q}_2-\hat{q}_1)^2}{2\hat{q}_1^2} + \frac{3\hat{q}_2}{2n\hat{q}_1^2} + \frac{(\hat{q}_2-\hat{q}_1)^3}{3\hat{q}_1^3} - \frac{\hat{q}_2^2}{n\hat{q}_1^3} + \frac{2\hat{q}_2}{n^2\hat{q}_1^3}
\end{align}
for estimating $g(q)=\ln q$. Note that in the Poisson model $n\hat{q}_2\sim \mathsf{Poi}(nq)$, by \eqref{eq.monomial_unbiased} we have $S_j(\hat{q}_2)=\prod_{k=0}^{j-1}(\hat{q}_2-\frac{k}{n})$. Then for estimating $f(p,q)=p\ln q$ in the ``smooth" regime, our estimator becomes
\begin{align}\label{eq.T_s}
T_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) =\hat{p}_1\cdot T^{(3)}(\hat{q}_1,\hat{q}_2) = \hat{p}_1(\ln \hat{q}_1 + \frac{\hat{q}_2-\hat{q}_1}{\hat{q}_1} - \frac{(\hat{q}_2-\hat{q}_1)^2}{2\hat{q}_1^2} + \frac{3\hat{q}_2}{2n\hat{q}_1^2} + \frac{(\hat{q}_2-\hat{q}_1)^3}{3\hat{q}_1^3} - \frac{\hat{q}_2^2}{n\hat{q}_1^3} + \frac{2\hat{q}_2}{n^2\hat{q}_1^3}).
\end{align}
To ensure that $T_\textrm{\rm s}$ is well-defined, it suffices to define an additional value of $T_\textrm{\rm s}$ (e.g., zero) when $\hat{q}_1=0$. Note that sample splitting here is only used for the simplicity of analysis, and it is indeed not necessary in implementation. We can also replace $\hat{p}_1$ with $\frac{\hat{p}_1+\hat{p}_2}{2}$ here to further reduce the variance.
Now consider the case where we are in the ``non-smooth" regime, i.e., $\hat{q}\le \frac{c_1\ln n}{n}$. By our general recipe, we should approximate $f(p,q)=p\ln q$ in the approximation region given by the confidence set
\begin{align}
U(\hat{p},\hat{q}) = \Theta \cap \left( \begin{cases}
[0, \frac{2c_1\ln m}{m}] & \text{if } \hat{p}\le \frac{c_1\ln m}{m},\\
[\hat{p}-\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}}, \hat{p}+\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}}] & \text{if } \frac{c_1\ln m}{m} < \hat{p} \le 1.
\end{cases} \times [0,\frac{4c_1\ln n}{n}]
\right).
\end{align}
As a result, we further distinguish the ``non-smooth" regime into two sub-regimes depending on $\hat{p}\le \frac{c_1\ln m}{m}$ or not, which by localization via confidence sets is essentially equivalent to $p\le \frac{c_1\ln m}{m}$ or not.
If $\hat{p}>\frac{c_1\ln m}{m}$, the approximation region is given by
\begin{align}
&\left\{ (p,q)\in [0,1]^2: \hat{p}-\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}} \le p \le \hat{p}+\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}}, 0\le q\le \frac{4c_1\ln n}{n}, p\le u(S)q \right\} \nonumber \\
&\qquad \subset \left[\hat{p}-\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}},\hat{p}+\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}}\right]\times \left[\frac{1}{u(S)}\left(\hat{p}-\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}}\right),\frac{4c_1\ln n}{n}\right]
\end{align}
where the latter is a rectangle. Since $q$ cannot hit zero in this approximation regime, and $f(p,q)=p\ln q$ is a product of $p$ and $\ln q$, we can consider the best polynomial approximation of $\ln q$ in this regime. As a result, in this regime, we use the approximation-based estimator
\begin{align}\label{eq.T_ns_I}
T_{\textrm{\rm ns},\text{I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) = \mathbbm{1}(\hat{p}_2\ge \frac{c_1\ln m}{3m})\cdot \sum_{k=0}^{K} g_{K,k}(\hat{p}_2)\cdot \hat{p}_1\prod_{j=0}^{k-1} \left(\hat{q}_1 - \frac{j}{n}\right)
\end{align}
where
\begin{align}
\sum_{k=0}^K g_{K,k}(\hat{p})z^k = \arg\min_{P\in \mathsf{poly}_K^1}\max_{z\in [\frac{1}{u(S)}(\hat{p}-\frac{1}{2}\sqrt{\frac{c_1\hat{p}\ln m}{m}}),\frac{4c_1\ln n}{n}]} |\ln z- P(z)|
\end{align}
is the best 1D order-$K$ polynomial approximation of $g(q)=\ln q$, where $K=c_2\ln n$ with universal constant $c_2>0$ to be specified later. Note that $\hat{p}_2\ge \frac{c_1\ln m}{3m}$ ensures that the 1D approximation interval does not contain zero and is thus valid. We call this regime as ``non-smooth" regime I.
If $\hat{p}\le \frac{c_1\ln m}{m}$, the approximation region is given by
\begin{align}
R = \left\{ (p,q)\in [0,1]^2: 0\le p\le \frac{2c_1\ln m}{m}, 0\le q\le \frac{4c_1\ln n}{n}, p\le u(S)q \right\}.
\end{align}
Since $q$ may be zero in $R$, the usual best 1D polynomial approximation of $g(q)=\ln q$ over this region does not work, and the best 2D polynomial approximation of $f(p,q)=p\ln q$ should be employed here. Hence, in this regime the approximation-based estimator is
\begin{align}\label{eq.T_ns_II}
T_{\textrm{\rm ns},\text{II}}(\hat{p}_1,\hat{q}_1) = \sum_{k,l\ge 0, 0<k+l\le K} h_{K,k,l}\prod_{i=0}^{k-1}\left(\hat{p}_1 - \frac{i}{m}\right)\prod_{j=0}^{l-1}\left(\hat{q}_1-\frac{j}{n}\right)
\end{align}
where
\begin{align}
\sum_{k,l\ge 0, k+l\le K} h_{K,k,l}w^kz^l = \arg\min_{P\in \mathsf{poly}_K^2}\max_{(w,z)\in R} |w\ln z- P(w,z)|
\end{align}
is the best 2D order-$K$ polynomial approximation of $f(p,q)=p\ln q$ in $R$, where $K=c_2\ln n$. Note that the condition $k+l>0$ in the summation ensures that the estimator is zero for unseen symbols. We call this regime as ``non-smooth" regime II.
In summary, we have the following estimator construction for $\sum_{i=1}^S p_i\ln q_i$.
\begin{construction}
Conduct three-fold sample splitting to obtain i.i.d. samples $(\hat{p}_{i,1},\hat{p}_{i,2},\hat{p}_{i,3})$ and $(\hat{q}_{i,1},\hat{q}_{i,2},\hat{q}_{i,3})$. The estimator $\hat{D}'$ for the cross entropy $\sum_{i=1}^S p_i\ln q_i$ is constructed as follows:
\begin{align}\label{eq.est_general_recipe}
\hat{D}' &= \sum_{i=1}^S \left[\left( \tilde{T}_{\textrm{\rm ns},\text{\rm I}}(\hat{p}_{i,1},\hat{q}_{i,1};\hat{p}_{i,2},\hat{q}_{i,2}) \mathbbm{1}(\hat{p}_{i,3}>\frac{c_1\ln m}{m}) + \tilde{T}_{\textrm{\rm ns},\text{\rm II}}(\hat{p}_{i,1},\hat{q}_{i,1}) \mathbbm{1}(\hat{p}_{i,3}\le\frac{c_1\ln m}{m}) \right)\mathbbm{1}(\hat{q}_{i,3}\le \frac{c_1\ln n}{n})\right.\nonumber\\
&\left.\qquad\qquad + \tilde{T}_\textrm{\rm s}(\hat{p}_{i,1},\hat{q}_{i,1};\hat{p}_{i,2},\hat{q}_{i,2})\mathbbm{1}(\hat{q}_{i,3}> \frac{c_1\ln n}{n})\right],
\end{align}
where
\begin{align}
\tilde{T}_{\textrm{\rm ns},\text{\rm I}}(x,y;x',y') &\triangleq (T_{\textrm{\rm ns},\text{\rm I}}(x,y;x',y')\wedge 1)\vee (-1)\\
\tilde{T}_{\textrm{\rm ns},\text{\rm II}}(x,y) &\triangleq (T_{\textrm{\rm ns},\text{\rm II}}(x,y;x',y')\wedge 1)\vee (-1)\\ \label{eq.est_smooth_tilde}
\tilde{T}_{\textrm{\rm s}}(x,y;x',y') &\triangleq T_\textrm{\rm s}(x,y;x',y')\cdot \mathbbm{1}(y\neq 0)
\end{align}
and $T_{\textrm{\rm ns},\text{\rm I}}, T_{\textrm{\rm ns},\text{\rm II}}, T_\textrm{\rm s}$ are given by (\ref{eq.T_ns_I}), (\ref{eq.T_ns_II}) and (\ref{eq.T_s}), respectively. A pictorial illustration of three regimes and our estimator is displayed in Figure \ref{fig.three_regimes}.
For the estimation of entropy, we essentially follow the estimator in \cite{Jiao--Venkat--Han--Weissman2015minimax}. Specifically, let $L_H(x)$ be the lower part function defined in \cite{Jiao--Venkat--Han--Weissman2015minimax}, and $U_H(x)$ be defined as
\begin{align}\label{eq.U_H}
U_H(x) \triangleq -x\ln x + \frac{1}{2m}
\end{align}
which gets rid of the interpolation function compared with the upper part function defined in \cite{Jiao--Venkat--Han--Weissman2015minimax}. Then the entropy estimator is defined as
\begin{align}\label{eq.upper_part}
\hat{H} = \sum_{i=1}^S \left[U_H(\hat{p}_{i,1})\mathbbm{1}(\hat{p}_{i,3}>\frac{c_1\ln m}{m}) + L_H(\hat{p}_{i,1})\mathbbm{1}(\hat{p}_{i,3}\le \frac{c_1\ln m}{m}) \right].
\end{align}
Finally, the overall estimator $\hat{D}$ for $D(P\|Q)$ is defined as
\begin{align}
\hat{D} = -\hat{D}' - \hat{H}
\end{align}
where $c_1,c_2>0$ are suitably chosen universal constants.
\end{construction}
\begin{figure}
\centering
\begin{tikzpicture}[xscale=9,yscale=8]
\draw [<->, help lines] (0.6,1.7) -- (0.6,0.6) -- (2.0,0.6);
\draw (0.6,0.6) -- (1.6,1.6) -- (1.95,1.6);
\draw (0.63,0.75) circle[radius=0.1pt];
\fill (0.63,0.75) circle[radius=0.1pt];
\draw (1.7,1.4) circle[radius=0.1pt];
\fill (1.7,1.4) circle[radius=0.1pt];
\draw (1.2,1.1) circle[radius=0.1pt];
\fill (1.2,1.1) circle[radius=0.1pt];
\node [left] at (0.6,0.6) {$0$};
\draw [dashed] (1.4, 0.6) -- (1.4,1.6);
\draw [dashed] (0.6, 0.8) -- (1.4,0.8);
\draw [dashed] (0.6, 1.6) -- (1.6,1.6);
\draw [dashed] (1.6, 0.6) -- (1.6,1.6);
\draw [dashed] (0.9, 0.6) -- (0.9,0.9);
\draw [dashed] (0.6, 1.3) -- (1.4,1.3);
\draw [dashed] (0.6, 1.1) -- (1.2,1.1);
\draw [dashed] (0.6, 0.9) -- (1.4,0.9);
\node [right] at (1.7,1.4) {$(\hat{q}_1,\hat{p}_1)$};
\node [right] at (1.2,1.1) {$(\hat{q}_2,\hat{p}_2)$};
\node [right] at (0.63,0.75) {$(\hat{q}_3,\hat{p}_3)$};
\node [below] at (1.4,0.6) {$\frac{c_1\ln n}{n}$};
\node [below] at (1.6,0.6) {$\frac{1}{u(S)}$};
\node [below] at (0.9,0.6) {$\frac{1}{u(S)}\left(\hat{p}_2-\frac{1}{2}\sqrt{\frac{c_1\ln m}{m}\cdot \hat{p}_2}\right)$};
\node [left] at (0.6,0.8) {$\frac{c_1\ln m}{m}$};
\node [left] at (0.6,1.6) {$1$};
\node [left] at (0.6,1.1) {$\hat{p}_2$};
\node [left] at (0.6,0.9) {$\hat{p}_2-\frac{1}{2}\sqrt{\frac{c_1\ln m}{m}\cdot \hat{p}_2}$};
\node [left] at (0.6,1.3) {$\hat{p}_2+\frac{1}{2}\sqrt{\frac{c_1\ln m}{m}\cdot \hat{p}_2}$};
\node [above] at (1.7, 1.1) {``Smooth" regime:};
\node [above] at (1.7, 1.04) {plug-in approach with};
\node [above] at (1.7, 1.0) {order-three bias correction};
\node [above] at (1.0, 0.7) {``Non-smooth" regime II: };
\node [above] at (1.0, 0.65) {unbiased estimate of best 2D};
\node [above] at (1.0, 0.6) {polynomial approximation of $p\ln q$};
\node [above] at (1.2, 1.0) {``Non-smooth" regime I: };
\node [above] at (1.2, 0.95) {$\hat{p}_2\times$ unbiased estimate of best};
\node [above] at (1.2, 0.9) {1D polynomial approximation of $\ln q$};
\node [right] at (2.0,0.6) {$q$};
\node [above] at (0.6,1.7) {$p$};
\end{tikzpicture}
\caption{Pictorial explanation of three regimes and our estimator for $\sum_{i=1}^S p_i\ln q_i$. The point $(\hat{q}_1,\hat{p}_1)$ falls in the ``smooth" regime, $(\hat{q}_2,\hat{p}_2)$ falls in the ``non-smooth" regime I, and $(\hat{q}_3,\hat{p}_3)$ falls in the ``non-smooth" regime II.}
\label{fig.three_regimes}
\end{figure}
\subsection{Estimator analysis}
In this subsection we will prove that the estimator constructed above achieves the minimax rate in Theorem \ref{thm.kl_divergence}. Recall that the mean squared error of any estimator $\hat{D}$ in estimating $D(P\|Q)$ can be decomposed into the squared bias and the variance as follows:
\begin{align}
\mathbb{E}_{(P,Q)}(\hat{D}-D(P\|Q))^2 = |\mathsf{Bias}(\hat{D})|^2 + \mathsf{Var}(\hat{D})
\end{align}
where the bias and the variance are defined as
\begin{align}
\mathsf{Bias}(\hat{D}) &\triangleq \mathbb{E}_{(P,Q)}\hat{D} - D(P\|Q)\\
\mathsf{Var}(\hat{D}) &\triangleq \mathbb{E}_{(P,Q)}(\hat{D}-\mathbb{E}_{(P,Q)}\hat{D})^2
\end{align}
respectively. Hence, it suffices to analyze the bias and the variance in these three regimes.
\subsubsection{``Smooth" regime}
First we consider the ``smooth" regime where the true parameter $(p,q)$ belongs to $\Theta_\textrm{\rm s}$, i.e., $q>\frac{c_1\ln n}{2n}$. In this regime, the estimator we employ is the plug-in approach whose bias is corrected by Taylor expansion, e.g., (\ref{eq.plugin_taylor}). Recall that the bias of our bias-corrected plug-in estimator can be expressed as
\begin{align}
|\mathbb{E}_q T^{(r)}(\hat{q}_1,\hat{q}_2) - g(q)| &= \left|\mathbb{E}_q\sum_{k=0}^r \frac{g^{(k)}(\hat{q})}{k!}(q-\hat{q})^k-g(q)\right| \le \mathbb{E}_q \left| \frac{g^{(r+1)}(\xi)}{(r+1)!}(q-\hat{q})^{k+1}\right|
\end{align}
where $n\hat{q}\sim \mathsf{Poi}(nq)$, and $\xi\in[q\wedge\hat{q},q\vee\hat{q}]$ depends on $\hat{q}$. For smooth $g(\cdot)$ with $|g^{(r+1)}(\xi)|$ bounded everywhere, it suffices to consider $\max_\xi|g^{(r+1)}(\xi)|$, the upper bound of the $(r+1)$-th derivative. However, the reason why the plug-in approach and its bias-corrected version are both strictly suboptimal for the estimation of non-smooth functionals (e.g., the empirical entropy \cite{Jiao--Venkat--Weissman2014MLE}) is that the functional $g(\cdot)$ may have non-analytic points where the high-order derivatives may be unbounded. Hence, a direct application of the Taylor expansion does not work for a general non-smooth $g(\cdot)$. However, now we are at the ``smooth" regime (i.e., $(p,q)\in \Theta_\textrm{\rm s}$), by our general recipe we know that with high probability $\hat{q}$ will not fall into the non-analytic region $\hat{\Theta}_0$ of $g(\cdot)$, thus $g(\cdot)$ is sufficiently smooth on the segment connecting $\hat{q}$ and $q$, and $\max_{q\wedge\hat{q}\le \xi\le q\vee\hat{q} }|g^{(r+1)}(\xi)|$ can be well controlled. In other words, the bias can be upper bounded with the help of localization via confidence sets.
Motivated by the previous insights, we begin with the following general lemma.
\begin{lemma}\label{lem.plugin_general}
Assume that the estimator in \eqref{eq.bias_cor} is well-defined, $\mathbb{E}_\theta\hat{\theta}_n^{(1)}=\theta$, and let $\{V(\theta)\}_{\theta\in\Theta}$ be a reverse confidence set of level $1-\delta$ with $\delta\in(0,1)$. Further suppose that for any $k=0,\cdots,r$, the function $H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})$ coincides with
\begin{align}
G_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)}) \triangleq \frac{G^{(k)}(\hat{\theta}_n^{(1)})}{k!}\sum_{j=0}^k \binom{k}{j} S_j(\hat{\theta}_n^{(2)})(-\hat{\theta}_n^{(1)})^{k-j}
\end{align}
whenever $\hat{\theta}_n^{(1)}\in V(\theta)$, and define $H(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\triangleq \sum_{k=0}^r H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})$. Then
\begin{align}
|\mathbb{E}_\theta H(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)}) - G(\theta)| \le \frac{\mathbb{E}_\theta|\hat{\theta}_n^{(1)}-\theta|^{r+1}}{(r+1)!}\cdot \sup_{\hat{\theta}\in V(\theta)}|G^{(r+1)}(\hat{\theta})| + \delta\cdot\left(|G(\theta)|+\sup_{\theta_1,\theta_2\in\hat{\Theta}}|H(\theta_1,\theta_2)|\right)
\end{align}
and for any $k=0,1,\cdots,r$,
\begin{align}
&\mathsf{Var}_\theta(H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})) \le A_r\left(\delta\cdot \sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_k(\theta_1,\theta_2)|+ \sum_{j=0}^k \left(\mathbb{E}_\theta(\hat{\theta}_n^{(1)}-\theta)^2\cdot\sup_{\hat{\theta}\in V(\theta)}|G_{k,j}'(\hat{\theta})|^2 + \delta|G_{k,j}(\theta)|^2\right)\cdot \mathbb{E}_\theta S_{k-j}^2(\hat{\theta}_n^{(2)})\right.\nonumber\\
&\qquad + \left.\sum_{j=0}^{k-1}\left(|G_{k,j}(\theta)| + \mathbb{E}_\theta(\hat{\theta}_n^{(1)}-\theta)^2\cdot \sup_{\hat{\theta}\in V(\theta)}|G_{k,j}''(\hat{\theta})|+\delta (|\theta|+\sup_{\hat{\theta}\in\hat{\Theta}}|\hat{\theta}|)\cdot|G_{k,j}'(\theta)|\right)^2\cdot \mathsf{Var}_\theta (S_{k-j}(\hat{\theta}_n^{(2)}))\right)\\
& |\mathbb{E}_\theta H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})| \le \begin{cases}
\mathbb{E}_\theta (\hat{\theta}_n^{(1)}-\theta)^2\cdot \sup_{\hat{\theta}\in V(\theta)}|G''(\hat{\theta})| + \delta\cdot \left(\sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_1(\theta_1,\theta_2)|+|G'(\theta)|(|\theta|+\sup_{\hat{\theta}\in\hat{\Theta}}|\hat{\theta}|)\right) & \text{if } k=1,\\
\frac{\mathbb{E}_\theta|\hat{\theta}_n^{(1)}-\theta|^k}{k!}\cdot \sup_{\hat{\theta}\in V(\theta)}|G^{(k)}(\hat{\theta})| + \delta\cdot \sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_k(\theta_1,\theta_2)| & \text{if }k\ge 2.
\end{cases}
\end{align}
where $G_{k,j}(x)\triangleq x^jG^{(k)}(x)$, and $A_r>0$ is a universal constant which depends on $r$ only.
\end{lemma}
It can be seen from the previous lemma that the upper bounds of both the bias and the variance are very easy to compute, for we only need to calculate the finite-order derivatives of $G(\cdot)$ and the moments of some usually well-behaved estimators (i.e., $S_j(\hat{\theta}_n^{(2)})$). Moreover, with the help of localization via confidence sets, all bounds only depend on the local behavior of function $G(\cdot)$ (so we only require that $H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})$ coincide with $G_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})$ when $\hat{\theta}_n^{(1)}\in V(\theta)$) plus a negligible term corresponding to the event that $\hat{\theta}_n^{(1)}\notin V(\theta)$. Note that this was the major difficult part in the analysis of the bias-corrected plug-in estimator in \cite{Jiao--Venkat--Han--Weissman2015minimax}, whose proof is quite lengthy (over four pages in the proof of Lemma 2) and requires the explicit construction of the interpolation function in estimator construction. Note that in Lemma \ref{lem.plugin_general} we implicitly use the following ``interpolation" idea: we essentially condition on the event that $\hat{\theta}_n\in V(\theta)$, which is similar as we ``interpolate" the function $G(\hat{\theta}_n)$ using the rectangle window $\mathbbm{1}(\hat{\theta}_n\in V(\theta))$ so as to prevent $\max_{\xi}|G^{(r+1)}(\xi)|$ becoming infinity. Note that this interpolation is done only in the analysis but not in the construction of our estimator, and thus we remark that the explicit interpolation in \cite{Jiao--Venkat--Han--Weissman2015minimax} is indeed unnecessary given the implicit interpolation by localization via confidence sets. Following this idea, although $U_H(\cdot)$ in \eqref{eq.U_H} does not follow the same idea of our bias correction \eqref{eq.bias_cor}, the result in \cite{Jiao--Venkat--Han--Weissman2015minimax} can still be easily recovered without explicit interpolation:
\begin{lemma}\label{lem.U_H}
Let $p\ge \frac{c_1\ln m}{2m}$, and $m\hat{p}\sim \mathsf{Poi}(mp)$. If $c_1\ln m\ge 4$, the following inequalities hold:
\begin{align}
|\mathbb{E} U_H(\hat{p}) + p\ln p|&\le \frac{6}{c_1m\ln m} + 4m^{-c_1/24+2}\\
\mathsf{Var}(U_H(\hat{p}))&\le A_0\left(\frac{p(2-\ln p)^2}{m} + 4m^{-c_1/24}\right)
\end{align}
where $A_0$ is the universal constant appearing in Lemma \ref{lem.plugin_general}.
\end{lemma}
Now we apply Lemma \ref{lem.plugin_general} to analyze the estimation performance of the bias-corrected plug-in estimator $T^{(3)}(\hat{q}_1,\hat{q}_2)$ in \eqref{eq.plugin_taylor}. In the Poisson model $n\hat{q}\sim \mathsf{Poi}(nq)$ with $q\ge \frac{c_1\ln n}{2n}$, a natural reverse confidence set is given by
\begin{align}
V(q) = \left[q - \frac{1}{2}\sqrt{\frac{c_1q\ln n}{2n}},q + \frac{1}{2}\sqrt{\frac{c_1q\ln n}{2n}} \right],\qquad q\ge \frac{c_1\ln n}{2n}.
\end{align}
By Lemma \ref{lemma.poissontail} we know that this reverse confidence set has level $1-\delta$ with
\begin{align}\label{eq.localize_level}
\delta\le \sup_{q\ge \frac{c_1\ln n}{2n}}\mathbb{P}_q(\hat{q}\notin V(q)) \le \exp\left(-\frac{1}{3}\left(\frac{1}{2}\sqrt{\frac{c_1\ln n}{2nq}}\right)^2\cdot nq\right) + \exp\left(-\frac{1}{2}\left(\frac{1}{2}\sqrt{\frac{c_1\ln n}{2nq}}\right)^2\cdot nq\right)\le 2n^{-c_1/24}
\end{align}
which can decay faster than any polynomial rate provided that $c_1$ is large enough. In this special case we can simplify the expressions in Lemma \ref{lem.plugin_general}.
\begin{lemma}\label{lem.plugin_upper}
Let $q\ge \frac{c_1\ln n}{2n}, (n\hat{q}_1,n\hat{q}_2)\sim\mathsf{Poi}(nq)\times\mathsf{Poi}(nq)$ and $c_1\ln n\ge 2$, and $g(\cdot)$ be an $(r+1)$ times differentiable function on $[\frac{c_1\ln n}{4n},1]$. Suppose that for any $k=0,\cdots,r$, the function $h_k(\hat{q}_1,\hat{q}_2)$ coincides with
\begin{align}
h_k(\hat{q}_1,\hat{q}_2) \triangleq \frac{g^{(k)}(\hat{q}_1)}{k!}\sum_{j=0}^k \binom{k}{j} S_j(\hat{q}_2)(-\hat{q}_1)^{k-j}
\end{align}
whenever $\hat{q}_1\in V(q)$, and define $h(\hat{q}_1,\hat{q}_2)\triangleq \sum_{k=0}^r h_k(\hat{q}_1,\hat{q}_2)$. Then there exists a universal constant $B_r>0$ depending on $r$ only such that
\begin{align}
|\mathbb{E} h(\hat{q}_1,\hat{q}_2)-g(q)| \le B_r\left(\left(\frac{q}{n}\right)^{\frac{r+1}{2}}\cdot \sup_{\xi\in [q/2,2q]}|g^{(r+1)}(\xi)| + n^{-c_1/24}\cdot \left(|g(q)|+\sup_{q_1,q_2}|h(q_1,q_2)|\right)\right)
\end{align}
and for any $k\le r$,
\begin{align}
\mathsf{Var}(h_k(\hat{q}_1,\hat{q}_2))&\le B_r\left(n^{-c_1/24}\cdot \sup_{q_1,q_2}|h_k(q_1,q_2)| + \sum_{j=0}^k q^{2(k-j)}\left(\frac{q}{n}\sup_{\xi\in[q/2,2q]}|g_{k,j}'(\xi)|^2+n^{-c_1/24}|g_{k,j}(q)|^2 \right)\right.\nonumber\\
&\left.\qquad + \sum_{j=0}^{k-1} \frac{q^{2(k-j)-1}}{n}\left(|g_{k,j}(q)|+\frac{q}{n}\sup_{\xi\in[q/2,2q]}|g_{k,j}''(\xi)|+n^{-c_1/24}|g_{k,j}'(q)|\right)^2
\right), \qquad k\ge 0\\
|\mathbb{E} h_k(\hat{q}_1,\hat{q}_2)| &\le B_r\left(\left(\frac{q}{n}\right)^{\frac{k\vee 2}{2}} \cdot\sup_{\xi\in[q/2,2q]}|g^{(k\vee 2)}(\xi)| + n^{-c_1/24}\cdot(|g'(q)|+\sup_{q_1,q_2}|h_k(q_1,q_2)|)\right), \qquad k\ge 1
\end{align}
where $g_{k,j}(q)\triangleq q^jg^{(k)}(q)$.
\end{lemma}
Note that due to the nice property of the Poisson model, the previous lemma greatly simplifies the expression involving $S_{j}(\hat{q})$, the unbiased estimate of the monomial functions. Moreover, if $g^{(s)}(q)$ becomes a power function of $q$ for some $s$, all summands in Lemma \ref{lem.plugin_upper} with $k\ge s$ will have the same order of magnitude and can thus be merged into one term. An interesting observation is that, if we change $r$ to $r+1$ in this case, the order of the bias of our bias-corrected estimator is multiplied by $\frac{1}{\sqrt{nq}}$, which is at most of the order $\frac{1}{\sqrt{\ln n}}$ since $q\ge \frac{c_1\ln n}{2n}$. Hence, by continuing this bias correction approach, we can improve the bias of the plug-in approach by any desired logarithmic multiplicative factor.
Next we apply Lemma \ref{lem.plugin_upper} to the bias-corrected estimator $\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)$ in the smooth regime, where $q\ge \frac{c_1\ln n}{2n}$ and $g(q)=\ln q$ satisfied the previous property (i.e., $g'(q)=1/q$ is a power of $q$).
\begin{lemma}\label{lem.T_s_I}
Let $p\in [0,1], q\ge \frac{c_1\ln n}{2n}, p\le u(S)q$, and Poisson random variables $m(\hat{p}_1,\hat{p}_2)\sim \mathsf{Poi}(mp)\times \mathsf{Poi}(mp)$, $n(\hat{q}_1,\hat{q}_2)\sim \mathsf{Poi}(nq)\times \mathsf{Poi}(nq)$ be independent. Moreover, let $c_1\ln m\ge 4$ and $c_1\ln n\ge 4$.
If $p\le \frac{2c_1\ln m}{m}$ is small, we have
\begin{align}
|\mathbb{E} [\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + L_H(\hat{p}_{1})]+p\ln(p/q)| &\le \frac{192B_3u(S)}{c_1n\ln n} + 6B_3pn^{-c_1/24+3} + \frac{C}{m\ln m} + \frac{(2c_1\ln m)^4}{m^{2-8c_2\ln 2}}\\
\mathsf{Var}(\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)+L_H(\hat{p}_1))&\le \frac{2(2c_1\ln m)^4}{m^{2-8c_2\ln 2}} + 1600B_3\left(p^2+\frac{p}{m}\right)\left(\frac{125}{nq}+n^{-c_1/24+3}\right)\nonumber\\
&\qquad \qquad +\frac{p}{m}\left(B_3\left(48+6n^{-c_1/24+3}\right)-\ln q\right)^2
\end{align}
where $B_3, C$ are universal constants given in Lemma \ref{lem.plugin_upper} and \cite[Lemma 4]{Jiao--Venkat--Han--Weissman2015minimax}, respectively. If $p\ge\frac{c_1\ln m}{2m}$ is large, we have
\begin{align}
|\mathbb{E} [\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + U_H(\hat{p}_{1})]+p\ln(p/q)| &\le \frac{192B_3u(S)}{c_1n\ln n} + 6B_3pn^{-c_1/24+3} + \frac{6}{c_1m\ln m} + 4m^{-c_1/24+2}\\
\mathsf{Var}(\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)+U_H(\hat{p}_1))&\le \frac{204}{m^2} + \frac{8}{m}\left(p+p(\ln u(S))^2+\frac{4q}{e^2}+\frac{4u(S)}{en}\right) \nonumber\\
&\qquad\qquad + \frac{2800u(S)}{n}\left(p+\frac{1}{m}\right) +900B_3\left(p^2+\frac{p}{m}\right)\left(\frac{125}{nq}+n^{-c_1/24+3}\right)\nonumber \\
&\qquad\qquad + 4(1+\ln n)^2\cdot n^{-c_1/2}+\frac{9B_3^2 p}{m}\left(48+3n^{-c_1/24+3}\right)^2.
\end{align}
In particular, if $c_1>96$ and $8c_2\ln 2<\epsilon\in (0,1)$, the previous bounds imply that for $p\le \frac{2c_1\ln m}{m}$,
\begin{align}
|\mathbb{E} [\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + L_H(\hat{p}_{1})]+p\ln(p/q)| &\lesssim \frac{u(S)}{n\ln n} + \frac{1}{m\ln m}\\
\mathsf{Var}(\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + L_H(\hat{p}_1)) &\lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{u(S)}{mn} + \frac{pu(S)}{n}
\end{align}
and for $p\ge \frac{c_1\ln m}{2m}$,
\begin{align}
|\mathbb{E} [\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + U_H(\hat{p}_{1})]+p\ln(p/q)| &\lesssim \frac{u(S)}{n\ln n} + \frac{1}{m\ln m}\\
\mathsf{Var}(\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + U_H(\hat{p}_1)) &\lesssim \frac{1}{m^2} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m} + \frac{u(S)}{mn} + \frac{pu(S)}{n}.
\end{align}
\end{lemma}
Note that the variance bound given by Lemma \ref{lem.T_s_I} is a non-asymptotic result whose order coincides with that given by classical asymptotics, where the asymptotic variance is the leading term and can be obtained easily via the delta method \cite{Vandervaart2000}. Now we use Lemma \ref{lem.T_s_I} to analyze the property of the overall estimator
\begin{align}\label{eq.T_s_overall}
\overline{T}_\textrm{\rm s}(\hat{p},\hat{q}) = \tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + L_H(\hat{p}_1)\mathbbm{1}(\hat{p}_3\le \frac{c_1\ln m}{m}) + U_H(\hat{p}_1)\mathbbm{1}(\hat{p}_3> \frac{c_1\ln m}{m})
\end{align}
in the ``smooth" regime $q\ge \frac{c_1\ln n}{2n}$, where $\hat{p},\hat{q}$ are the vector representations of $\hat{p}_1,\hat{p}_2,\hat{p}_3$ and $\hat{q}_1,\hat{q}_2,\hat{q}_3$, respectively.
\begin{lemma}\label{lem.T_s}
Let $p\in [0,1], q\ge \frac{2c_1\ln n}{n}$, $p\le u(S)q$, and $m\hat{p}=m(\hat{p}_1,\hat{p}_2,\hat{p}_3)\sim \mathsf{Poi}(mp)^3, n\hat{q}=n(\hat{q}_1,\hat{q}_2,\hat{q}_3)\sim \mathsf{Poi}(nq)^3$ be independent. Moreover, let $c_1>96$ and $8c_2\ln 2<\epsilon\in(0,1)$, we have
\begin{align}
|\mathbb{E}\overline{T}_\textrm{\rm s}(\hat{p},\hat{q}) + p\ln(p/q)|&\lesssim \frac{1}{m\ln m}+\frac{u(S)}{n\ln n}\\
\mathsf{Var}(\overline{T}_\textrm{\rm s}(\hat{p},\hat{q}))&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m} + \frac{u(S)}{mn} + \frac{pu(S)}{n} + \left(\frac{u(S)}{n\ln n}\right)^2.
\end{align}
\end{lemma}
\subsubsection{``Non-smooth" regime I}
Next we consider the ``non-smooth" regime I where $p\ge \frac{c_1\ln m}{2m}, q\le \frac{2c_1\ln n}{n}$. By construction of the approximation-based estimator $T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)$, when $\hat{p}_2\ge \frac{c_1\ln m}{3m}$ and the approximation region
\begin{align}
\left[ \frac{1}{u(S)}\left(\hat{p}_2 - \frac{1}{2}\sqrt{\frac{c_1\hat{p}_2\ln m}{m}}\right) , \frac{4c_1\ln n}{n}\right]
\end{align}
contains the true parameter $q$, the bias of this estimator is essentially the product of $p$ and the best polynomial approximation error of $\ln q$ in the previous approximation region. This approximation error can be easily obtained, for the 1D polynomial approximation is well-understood (Lemma \ref{lem.approx_log} gives an upper bound for the approximation error of $\ln x$). Moreover, note that the previous event occurs if $\frac{c_1\ln m}{3m}\le \hat{p}_2\le p+\frac{1}{2}\sqrt{\frac{c_1\hat{p}_2\ln m}{m}}$, which holds with overwhelming probability by confidence sets. As for the variance, it suffices to bound the variance of each term of the form $\prod_{l=0}^{s-1} (Y-\frac{l}{n})$, where $nY\sim \mathsf{Poi}(nq)$. Complicated as it may seem, the present authors showed in \cite{jiao2016minimax} that the variance has an explicit expression in Poisson models, which is the so-called Charlier polynomial \cite{Peccati--Taqqu2011some}.
Hence, we have good tools for the analysis of both the bias and the variance of $\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)$, which is presented in the following lemma.
\begin{lemma}\label{lem.T_ns_1}
Let $\frac{c_1\ln m}{2m}\le p\le u(S)q, q\le \frac{2c_1\ln n}{n}$, and Poisson random variables $m(\hat{p}_1,\hat{p}_2)\sim \mathsf{Poi}(mp)\times \mathsf{Poi}(mp)$, $n(\hat{q}_1,\hat{q}_2)\sim \mathsf{Poi}(nq)\times \mathsf{Poi}(nq)$ be independent. If $c_1\ge 2c_2$ and $c_2\ln n\ge 1$, we have
\begin{align}
|\mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)] - p\ln q|&\le C_{\ln }pW\left(\frac{42c_1}{c_2^2}\cdot \frac{u(S)}{pn\ln n}\right) + 2m^{-c_1/36}(1-p\ln q) \nonumber\\
&\qquad\qquad +\frac{16c_1u(S)\ln n}{n^{1-11c_2\ln 2}}\left(\frac{2c_1u(S)\ln n}{n}+\frac{1}{m}\right)\left(42C_{\ln}^2 + (c_2\ln n)^2(C_{\ln}^2 + (\ln n)^2)\right) \\
\mathsf{Var}(\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2))&\le \frac{16c_1u(S)\ln n}{n^{1-11c_2\ln 2}}\left(\frac{2c_1u(S)\ln n}{n}+\frac{1}{m}\right)\left(42C_{\ln}^2 + (c_2\ln n)^2(C_{\ln}^2 + (\ln n)^2)\right) + 2m^{-c_1/36}
\end{align}
where $W(\cdot)$ and $C_{\ln}>0$ are given in Lemma \ref{lem.approx_log}. In particular, by Lemma \ref{lem.U_H}, if $c_1>72$ and $11c_2\ln 2<\epsilon\in(0,1)$, we have
\begin{align}
|\mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)+U_H(\hat{p}_1)]+p\ln(p/q)|&\lesssim \frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n}) + \frac{u(S)(\ln n)^5}{n^{1-\epsilon}}\left(\frac{u(S)\ln n}{n} + \frac{1}{m}\right)\\
\mathsf{Var}(\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)+U_H(\hat{p}_1))&\lesssim \frac{p(1+(\ln p)^2)}{m} + \frac{1}{m^3} + \frac{u(S)(\ln n)^5}{n^{1-\epsilon}}\left(\frac{u(S)\ln n}{n} + \frac{1}{m}\right)\\
&\lesssim \frac{u(S)(\ln n)^3}{mn} + \frac{p}{m} + \frac{1}{m^3} + \frac{u(S)(\ln n)^5}{n^{1-\epsilon}}\left(\frac{u(S)\ln n}{n} + \frac{1}{m}\right).
\end{align}
\end{lemma}
Here $pW(\frac{u(S)}{pn\ln n})$ corresponds to the polynomial approximation error, which will become the leading term in the bias.
\subsubsection{``Non-smooth" regime II}
Now we consider the ``non-smooth" regime II where $p\le \frac{2c_1\ln m}{m}, q\le \frac{2c_1\ln n}{n}$ and $p\le u(S)q$. By the estimator construction, it is necessary to deal with the best 2D polynomial approximation of $p\ln q$ in
\begin{align}\label{eq.2D_approx_region}
R = \left\{ (p,q)\in [0,1]^2: 0\le p\le \frac{2c_1\ln m}{m}, 0\le q\le \frac{4c_1\ln n}{n}, p\le u(S)q \right\}.
\end{align}
We emphasize that the polynomial approximation in general multivariate case is extremely complicated. Rice~\cite{Rice1963tchebycheff} wrote:
\begin{center}
\vspace{5pt}
\parbox{0.9\textwidth}{~~\emph{ ``The theory of Chebyshev approximation (a.k.a. best approximation) for functions
of one real variable has been understood for some time and is quite elegant. For about fifty years attempts have been made to generalize this theory to functions of several variables. These attempts have failed because of the lack
of uniqueness of best approximations to functions of more than one variable.
"}}
\vspace{5pt}
\end{center}
We also show in \cite{jiao2016minimax} that the non-uniqueness can cause serious trouble: some polynomial that achieves the best approximation error cannot be used in our general methodology in functional estimation. What if we relax the requirement of computing the best approximation in the multivariate case, and merely analyze the best approximation rate (i.e., the best approximation error up to a multiplicative constant)? That turns out also to be extremely difficult. Ditzian and Totik~\cite[Chap. 12]{Ditzian--Totik1987} obtained the error rate estimate on simple polytopes\footnote{A simple polytope in $\mathbb{R}^d$ is a polytope such that each vertex has $d$ edges. }, balls, and spheres, and it remained open until very recently Totik~\cite{Totik2013polynomial} generalized the results to general polytopes. For results in balls and spheres, the readers are referred to Dai and Xu~\cite{Dai--Xu2013approximation}. We still know little about regimes beyond polytopes, balls, and spheres.
Complicated as the general multivariate case is, it is still possible to solve our problem since the approximation region $R$ in (\ref{eq.2D_approx_region}) is a convex polytope. Now we review the general theory of polynomial approximation in convex polytopes \cite{Totik2013polynomial}. In $\mathbb{R}^d$ we call a closed set $K\subset \mathbb{R}^d$ a convex polytope if it is the convex hull of finitely many points. Let $x\in K$ and $e\in\mathbb{R}^d$ be a direction (i.e., a unit vector). For continuous function $f$ on $K$, we define the $r$-th symmetric difference in the direction $e$ as
\begin{align}
\Delta_{he}^rf(x) = \sum_{k=0}^r (-1)^k\binom{r}{k}f\left(x + (\frac{r}{2}-k)he \right)
\end{align}
with the understanding that $\Delta_{he}^rf(x)=0$ if $x+\frac{r}{2}he$ or $x-\frac{r}{2}he$ does not belong to $K$. Moreover, letting the line $l_{e,x}$ through $x$ with direction $e$ intersects $K$ at point $A_{e,x}, B_{e,x}$, we define the normalized distance as
\begin{align}
\tilde{d}_K(e,x) = \sqrt{\|x - A_{e,x}\|_2\cdot \|x - B_{e,x}\|_2}.
\end{align}
Denoting by $S_{d-1}$ the set of all unit vectors in $\mathbb{R}^d$, we define the Ditzian--Totik modulus of smoothness as follows:
\begin{align}\label{eq.DT_modulus_general}
\omega_K^r(f,t) = \sup_{e\in S^{d-1}}\sup_{x\in K}\sup_{h\le t} |\Delta_{h\tilde{d}_K(e,x)e}^r f(x)|.
\end{align}
The significance of this quantity is presented in the following lemma.
\begin{lemma}\cite{Totik2013polynomial}\label{lem.DT_modulus}
Let $K\subset \mathbb{R}^d$ be a $d$-dimensional convex polytope and $r=1,2,\cdots$. Then, for $n\ge rd$, we have
\begin{align}
E_n[f;K] &\le M\omega_K^r(f,\frac{1}{n})\\
\frac{M}{n^r}\sum_{k=0}^n (k+1)^{r-1}E_k[f;K] &\ge \omega_K^r(f,\frac{1}{n})
\end{align}
where the constant $M>0$ only depends on $r$ and $d$.
\end{lemma}
Hence, Lemma \ref{lem.DT_modulus} shows that once we compute the Ditzian--Totik modulus of smoothness $\omega_K^r(f,1/n)$ for $f$, we immediately obtain an upper bound for the best polynomial approximation error. Moreover, Lemma \ref{lem.DT_modulus} also shows that this is essentially also the lower bound. In our case, $K=R$, and $f(p,q)=p\ln q$. Choosing $r=2$, Lemma \ref{lem.approx_2D} gives an upper bound for the Ditzian--Totik modulus of smoothness. Moreover, tracing back to the proof for the simple polytope case in \cite{Ditzian--Totik1987}, it suffices to take the supremum in (\ref{eq.DT_modulus_general}) over all directions $e$ which are parallel to some edge of the simple polytope $K$, which makes the evaluation of $\omega_K^r(f,t)$ much simpler.
Now we are in position to bound the bias and the variance of $\tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)$, which are summarized in the following lemma.
\begin{lemma}\label{lem.T_ns_II}
Let $p\le \frac{2c_1\ln m}{m}\wedge u(S)q, q\le \frac{2c_1\ln n}{n}$, and Poisson random variables $m(\hat{p}_1,\hat{p}_2)\sim \mathsf{Poi}(mp)\times \mathsf{Poi}(mp)$, $n(\hat{q}_1,\hat{q}_2)\sim \mathsf{Poi}(nq)\times \mathsf{Poi}(nq)$ be independent. If $c_1\ge 2c_2$ and $c_2\ln n\ge 1$, we have
\begin{align}
|\mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)]-p\ln q| &\le \frac{2MC_0}{c_2^2}\frac{u(S)}{n\ln n} +\frac{2^{13}c_1^2c_2^4(u(S))^2(\ln n)^4}{n^{2-26c_2\ln 2}}\left(1+\left(1+\frac{\ln n}{\ln m}\right)^{2c_2\ln n}+\left(\frac{n(\ln m+\ln n)}{mu(S)\ln n}\right)^{2c_2\ln n} \right)\\
\mathsf{Var}(\tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)) &\le \frac{2^{13}c_1^2c_2^4(u(S))^2(\ln n)^4}{n^{2-26c_2\ln 2}}\left(1+\left(1+\frac{\ln n}{\ln m}\right)^{2c_2\ln n}+\left(\frac{n(\ln m+\ln n)}{mu(S)\ln n}\right)^{2c_2\ln n} \right).
\end{align}
where the universal constant $M>0$ is given by Lemma \ref{lem.DT_modulus} (with $r=d=2$), and the constant $C_0$ which only depends on $c_1$ is given by Lemma \ref{lem.approx_2D}. In particular, by \cite[Lemma 5]{Jiao--Venkat--Han--Weissman2015minimax}, if $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$, there exists some universal constant $B>0$ such that
\begin{align}
|\mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)+L_H(\hat{p}_1)]+p\ln(p/q)| &\lesssim \frac{1}{m\ln m} + \frac{u(S)}{n\ln n} + \frac{(u(S))^2}{n^{2-c_2B}}\\
\mathsf{Var}(\tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) + L_H(\hat{p}_1) ) &\lesssim \frac{1}{m^{2-c_2B}} + \frac{(u(S))^2}{n^{2-c_2B}}.
\end{align}
\end{lemma}
We remark that the condition $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$ can be removed later in the construction of the adaptive estimator. In fact, the reason why we need this condition here is that we use an \emph{arbitrary} best 2D polynomial approximation, which is not unique in general. This point is very subtle: as was shown by the present authors in \cite{jiao2016minimax}, not all polynomials which can achieve the best uniform approximation error can be used to construct the rate-optimal estimator. Actually, we will show in the next subsection that a special approximating polynomial can achieve the same rate without this condition. Moreover, a more careful design of the approximating polynomial should require different degrees on $p$ and $q$ (instead of fixing a total degree $K=c_2\ln n$), but it has yet been unknown to approximation theorists that how to analyze the corresponding approximation error in general polytopes.
\subsubsection{Overall performance}
Now we analyze the performance of the entire estimator $\hat{D}'$. For simplicity, we define
\begin{align}
T_{\textrm{\rm ns}}(\hat{p}_i,\hat{q}_i) &\triangleq (\tilde{T}_{\textrm{\rm ns},\text{I}}(\hat{p}_{i,1},\hat{q}_{i,1};\hat{p}_{i,2},\hat{q}_{i,2}) +U_H(\hat{p}_{i,1}))\mathbbm{1}(\hat{p}_{i,3}>\frac{c_1\ln m}{m}) + (\tilde{T}_{\textrm{\rm ns},\text{II}}(\hat{p}_{i,1},\hat{q}_{i,1}) + L_H(\hat{p}_{i,1})) \mathbbm{1}(\hat{p}_{i,3}\le\frac{c_1\ln m}{m}) \\
\xi(\hat{p}_i,\hat{q}_i) &\triangleq T_{\textrm{\rm ns}}(\hat{p}_i,\hat{q}_i)\mathbbm{1}(\hat{q}_{i,3}\le \frac{c_1\ln n}{n}) + \overline{T}_\textrm{\rm s}(\hat{p}_i,\hat{q}_i)\mathbbm{1}(\hat{q}_{i,3}> \frac{c_1\ln n}{n})
\end{align}
where $\overline{T}_\textrm{\rm s}$ is given in (\ref{eq.T_s_overall}), and $\hat{p}_i=(\hat{p}_{i,1},\hat{p}_{i,2},\hat{p}_{i,3})$ is the vector representation of the independent components, and similarly for $\hat{q}_i$. Based on the current notations, we have
\begin{align}
\hat{D} = -\sum_{i=1}^S \xi(\hat{p}_i,\hat{q}_i)
\end{align}
and by the independence between different symbols, we have
\begin{align}
|\mathsf{Bias}(\hat{D})| &\le \sum_{i=1}^S |\mathsf{Bias}(\xi(\hat{p}_i,\hat{q}_i))|\\
\mathsf{Var}(\hat{D}) &= \sum_{i=1}^S \mathsf{Var}(\xi(\hat{p}_i,\hat{q}_i)).
\end{align}
Hence, it suffices to analyze the bias and the variance of each $\xi(\hat{p}_i,\hat{q}_i)$ separately and then add them all. Based on Lemma \ref{lem.T_ns_1} and Lemma \ref{lem.T_ns_II}, the next lemma first analyzes the bias and the variance of $T_{\textrm{\rm ns}}(\hat{p}_i,\hat{q}_i)$.
\begin{lemma}\label{lem.overall_T_ns}
Let $0\le p\le 1\wedge u(S)q, 0\le q\le \frac{2c_1\ln n}{n}$, and $m\hat{p}=m(\hat{p}_1,\hat{p}_2,\hat{p}_3)\sim \mathsf{Poi}(mp)^{3}$ and $n\hat{q}=n(\hat{q}_1,\hat{q}_2,\hat{q}_3)\sim \mathsf{Poi}(nq)^3$ be independent. Moreover, we assume that $n\gtrsim u(S)$, $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$, $c_1>72$ and $c_2(B\vee 11\ln 2)<\epsilon\in(0,1)$ (where $B$ is given in Lemma \ref{lem.T_ns_II}). Then,
\begin{enumerate}
\item when $p\le \frac{c_1\ln m}{2m}$,
\begin{align}
|\mathsf{Bias}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))| &\lesssim \frac{1}{m\ln m} + \frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}},\\
\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) & \lesssim \frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-\epsilon}}.
\end{align}
\item when $\frac{c_1\ln m}{2m} < p < \frac{2c_1\ln m}{m}$,
\begin{align}
|\mathsf{Bias}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))| &\lesssim \frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}},\\
\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) & \lesssim \frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-2\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} + [pW(\frac{u(S)}{pn\ln n})]^2.
\end{align}
\item when $p\ge\frac{2c_1\ln m}{m}$,
\begin{align}
|\mathsf{Bias}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))| &\lesssim \frac{1}{m\ln m}+ pW(\frac{u(S)}{pn\ln n})+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}},\\
\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) & \lesssim \frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-2\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}.
\end{align}
\end{enumerate}
\end{lemma}
Note that in Lemma \ref{lem.overall_T_ns}, the condition $n\gtrsim u(S)$ is a natural requirement for the consistency of the optimal estimator in view of Theorem \ref{thm.kl_divergence}, and $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$ is the additional condition in Lemma \ref{lem.T_ns_II}. Now based on Lemma \ref{lem.T_s} and Lemma \ref{lem.overall_T_ns}, we are about to analyze the bias and the variance of $\xi(\hat{p},\hat{q})$.
\begin{lemma}\label{lem.overall_xi}
Let $0\le p\le 1\wedge u(S)q, 0\le q\le 1$, and $m\hat{p}=m(\hat{p}_1,\hat{p}_2,\hat{p}_3)\sim \mathsf{Poi}(mp)^3$ and $n\hat{q}=n(\hat{q}_1,\hat{q}_2,\hat{q}_3)\sim \mathsf{Poi}(nq)^3$ be independent. Moreover, we assume that $n\gtrsim u(S)$, $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$, $c_1>96$ and $c_2(B\vee 11\ln 2)<\epsilon\in(0,1/2)$ (where $B$ is given in Lemma \ref{lem.T_ns_II}). Then,
\begin{enumerate}
\item when $q\le \frac{c_1\ln n}{2n}$,
\begin{align}
|\mathsf{Bias}(\xi(\hat{p},\hat{q}))| &\lesssim \begin{cases}
\frac{1}{m\ln m} + \frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} & p\le \frac{c_1\ln m}{2m},\\
\frac{1}{m\ln m}+pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n} + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & \frac{c_1\ln m}{2m}<p<\frac{2c_1\ln m}{m},\\
\frac{1}{m\ln m}+pW(\frac{u(S)}{pn\ln n}) + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & p\ge \frac{2c_1\ln m}{m}.
\end{cases}\\
\mathsf{Var}(\xi(\hat{p},\hat{q})) & \lesssim \begin{cases}
\frac{1}{m^{2-\epsilon}}+\frac{(u(S))^2}{n^{2-\epsilon}} & p\le \frac{c_1\ln m}{2m},\\
\frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-2\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} + [pW(\frac{u(S)}{pn\ln n})]^2 & \frac{c_1\ln m}{2m}<p<\frac{2c_1\ln m}{m},\\
\frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & p\ge \frac{2c_1\ln m}{m}.
\end{cases}
\end{align}
\item when $\frac{c_1\ln n}{2n} < q < \frac{2c_1\ln n}{n}$,
\begin{align}
|\mathsf{Bias}(\xi(\hat{p},\hat{q}))| &\lesssim \begin{cases}
\frac{1}{m\ln m}+\frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} & p\le \frac{c_1\ln m}{2m},\\
\frac{1}{m\ln m}+pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & p\ge \frac{c_1\ln m}{2m}.
\end{cases}\\
\mathsf{Var}(\xi(\hat{p},\hat{q})) & \lesssim \frac{1}{m^2} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m}+ \frac{pu(S)}{n} + \frac{(u(S))^2}{n^{2-2\epsilon}} +\begin{cases}
\frac{u(S)}{mn} & p\le \frac{c_1\ln m}{2m},\\
\frac{(u(S))^2}{n^{2-2\epsilon}} + [pW(\frac{u(S)}{pn\ln n})]^2 & p>\frac{c_1\ln m}{2m}.
\end{cases}
\end{align}
\item when $q\ge\frac{2c_1\ln n}{n}$,
\begin{align}
|\mathsf{Bias}(\xi(\hat{p},\hat{q}))| &\lesssim \frac{1}{m\ln m}+\frac{u(S)}{n\ln n},\\
\mathsf{Var}(\xi(\hat{p},\hat{q})) & \lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m}+ \frac{u(S)}{mn} + \frac{pu(S)}{n} + (\frac{u(S)}{n\ln n})^2.
\end{align}
\end{enumerate}
\end{lemma}
Based on Lemma \ref{lem.overall_xi}, we can analyze the total bias and variance of our estimator $\hat{D}$. By differentiation, we have
\begin{align}
pW(\frac{u(S)}{pn\ln n}) \le \frac{u(S)}{en\ln n}
\end{align}
and the maximum is attained at $p=\frac{u(S)}{en\ln n}$. Hence,
\begin{align}
|\mathsf{Bias}(\hat{D})| &\le \sum_{i=1}^S |\mathsf{Bias}(\xi(\hat{p}_i,\hat{q}_i))| \lesssim \frac{S}{m\ln m}+\frac{Su(S)}{n\ln n} + \frac{S(u(S))^2}{n^{2-2\epsilon}} + \frac{Su(S)}{mn^{1-\epsilon}}\\
\mathsf{Var}(\hat{D}) &= \sum_{i=1}^S \mathsf{Var}(\xi(\hat{p}_i,\hat{q}_i)) \lesssim \frac{S}{m^{2-\epsilon}} + \frac{S(u(S))^2}{n^{2-2\epsilon}} + \frac{Su(S)}{mn^{1-2\epsilon}} + \frac{(1+\ln u(S))^2}{m} + \frac{u(S)}{n}.
\end{align}
If we further require that $\ln S\gtrsim \ln(m\vee n)$ and $n\gtrsim Su(S)/\ln S$, the previous results can be further upper bounded as (let $\epsilon\to 0$)
\begin{align}
|\mathsf{Bias}(\hat{D})| &\lesssim \frac{S}{m\ln m} + \frac{Su(S)}{n\ln n} + (\frac{Su(S)}{n\ln n} )^2 + \left(\frac{S}{m^2} + \frac{S(u(S))^2}{n^{2-2\epsilon}}\right)\\
&\lesssim \frac{S}{m\ln m} + \frac{Su(S)}{n\ln n} + (\frac{Su(S)}{n\ln n} )^2 + \left(\frac{S}{m^2} + (\frac{Su(S)}{n\ln n})^2\right)\\
&\lesssim \frac{S}{m\ln m} + \frac{Su(S)}{n\ln n}\\
\mathsf{Var}(\hat{D}) &\lesssim (\frac{S}{m\ln m})^2 + (\frac{Su(S)}{n\ln n} )^2 + \left(\frac{S}{m^2} + \frac{S(u(S))^2}{n^{2-4\epsilon}}\right) + \frac{(1+\ln u(S))^2}{m} + \frac{u(S)}{n}\\
&\lesssim (\frac{S}{m\ln m})^2 + (\frac{Su(S)}{n\ln n} )^2 + \frac{(\ln u(S))^2}{m} + \frac{u(S)}{n}.
\end{align}
Hence we come to the following theorem.
\begin{theorem}\label{thm.general_recipe}
Let $m\gtrsim \frac{S}{\ln S}, n\gtrsim \frac{Su(S)}{\ln S}, \ln S\gtrsim \ln(m\vee n)$ and $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$. Then for our estimator $\hat{D}$ in (\ref{eq.est_general_recipe}) constructed from the general recipe, we have
\begin{align}
\sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left(\hat{D} - D(P\|Q)\right)^2 \lesssim \left(\frac{S}{m\ln m}+\frac{Su(S)}{n\ln n}\right)^2 + \frac{(\ln u(S))^2}{m} + \frac{u(S)}{n}.
\end{align}
Moreover, $\hat{D}$ does not require the knowledge of the support size $S$.
\end{theorem}
\subsection{An adaptive estimator}
So far we have obtained an essentially minimax estimator via our general recipe. However, since this estimator is purely obtained from the general method, it is not surprising that it is also subject to some disadvantages. Firstly, in the estimator we do not specify the explicit form of the best 2D polynomial approximation in the ``non-smooth" regime II. Although the best 1D polynomial approximation is unique and can be efficiently obtained via the Remez algorithm \cite{Remez1934determination}, which has been efficiently implemented in Matlab \cite{Trefethen--chebfunv5}, the best 2D polynomial approximation is not unique and hard to compute. Moreover, as what we have remarked before, the non-uniqueness forces us to add an unnecessary condition $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$ in Lemma \ref{lem.T_ns_II} and thus in Theorem \ref{thm.general_recipe}. Secondly, although the estimator does not require the knowledge of the support size $S$ (we remove the constant term in the polynomial approximation), but it requires the upper bound on the likelihood ratio $u(S)$ (in the design of ``non-smooth" regime I). In practice, we wish to obtain an adaptive estimator which achieves the minimax rate and is agnostic to both $S$ and $u(S)$. Thirdly, for the estimator construction in the ``non-smooth" regime I, the approximating polynomial depends on the empirical probabilities (i.e., we cannot store the polynomials in advance), which incurs large computational complexity.
To resolve these issues, we need to apply some tricks to explicitly construct an approximating polynomial for $f(p,q)=p\ln q$ in the ``non-smooth" regime, i.e., $q\le \frac{2c_1\ln n}{n}$. We first suppose that there exists a 1D polynomial $T(q)$ in $q$ with degree $\lesssim \ln n$ such that $pT(q)$ has the desired approximation property for $f(p,q)=p\ln q$ in the entire ``non-smooth" regime, i.e., we need not to distinguish ``non-smooth" regimes I and II. We remark that either the 1D approximation or only one approximation on the entire ``non-smooth" regime is not always doable in general. For example, for estimating the $\ell_1$ distance, it has been shown in \cite{jiao2016minimax} that not only any single approximation in the entire ``non-smooth" regime will always fail to give the correct order of the approximation error, but any 1D polynomial approximation of $|p-q|$ in $p-q$ will also not work when both $p$ and $q$ are small. Nevertheless, this ambitious target can be achieved in our special example. Motivated by Lemma \ref{lem.T_ns_1} and Lemma \ref{lem.T_ns_II}, the correct order of the approximation error is $\frac{u(S)}{n\ln n}$, i.e., $T(q)$ should satisfy that
\begin{align}
\sup_{0\le q\le \frac{2c_1\ln n}{n}}|pT(q) - p\ln q| \lesssim \frac{u(S)}{n\ln n}.
\end{align}
Since $p\le u(S)q$, it suffices to have
\begin{align}
|T(q) - \ln q|\lesssim \frac{1}{qn\ln n}, \qquad \forall q\in (0,\frac{2c_1\ln n}{n}]
\end{align}
i.e., to find a 1D polynomial approximation which satisfies the desired pointwise bound. However, it is easy to show that there exists some polynomial $T_0(q)$ on $[0,\frac{2c_1\ln n}{n}]$ such that $\deg(T_0)\lesssim \ln n$ and
\begin{align}
\sup_{0\le q\le \frac{2c_1\ln n}{n}}|T_0(q) - q\ln q| \lesssim \frac{1}{n\ln n}.
\end{align}
Hence, if we remove the constant term of $T_0(q)$ and define $T(q)=T_0(q)/q$, then $T(q)$ will have the desired property.
Motivated by the previous observations, we can construct an explicit estimator as follows:
\begin{construction}
Conduct three-fold sample splitting to obtain i.i.d. samples $(\hat{p}_{i,1},\hat{p}_{i,2},\hat{p}_{i,3})$ and $(\hat{q}_{i,1},\hat{q}_{i,2},\hat{q}_{i,3})$. The adaptive estimator $\hat{D}_{\textrm{A}}$ for the KL divergence $D(P\|Q)$ is
\begin{align}\label{eq.est_adaptive}
\hat{D}_{\textrm{A}} &= -\sum_{i=1}^S \left[L_H(\hat{p}_{i,1})\mathbbm{1}(\hat{p}_{i,3}\le \frac{c_1\ln m}{m})+U_H(\hat{p}_{i,1})\mathbbm{1}(\hat{p}_{i,3}>\frac{c_1\ln m}{m})\right.\nonumber\\
&\qquad\qquad\left.+\tilde{T}_{\textrm{\rm ns}}(\hat{p}_{i,1},\hat{q}_{i,1})\mathbbm{1}(\hat{q}_{i,3}\le \frac{c_1\ln n}{n}) + \tilde{T}_\textrm{\rm s}(\hat{p}_{i,1},\hat{q}_{i,1};\hat{p}_{i,2},\hat{q}_{i,2})\mathbbm{1}(\hat{q}_{i,3}> \frac{c_1\ln n}{n})\right]
\end{align}
where $\tilde{T}_\textrm{\rm s}$ is given by (\ref{eq.est_smooth_tilde}), $L_H(x), U_H(x)$ are given by \cite{Jiao--Venkat--Han--Weissman2015minimax} and (\ref{eq.upper_part}), respectively, and
\begin{align}
\tilde{T}_\textrm{\rm ns}(x,y) &= (T_\textrm{\rm ns}(x,y)\wedge 1)\vee(-1)\\
T_\textrm{\rm ns}(x,y) &= \sum_{k=0}^K g_{K,k+1}\left(\frac{2c_1\ln n}{n}\right)^{-k}\cdot x\prod_{l=0}^{k-1}\left(y-\frac{l}{n}\right)
\end{align}
where the coefficients $\{g_{K,k}\}_{k=1}^{K+1}$ are given by the best polynomial approximation of $x\ln x$ as follows:
\begin{align}
g_{K,1} &= r_{K,1} + \ln \left(\frac{2c_1\ln n}{n}\right), \qquad g_{K,k} = r_{K,k}, \quad 2\le k\le K+1\\
\sum_{k=0}^{K+1} r_{K,k}x^k &= \arg\min_{P\in \mathsf{poly}_{K+1}^1} \sup_{x\in [0,1]} |P(x) - x\ln x|.
\end{align}
The parameters $c_1,c_2>0$ are suitably chosen universal constants. A pictorial illustration of $\hat{D}_{\textrm{A}}$ is displayed in Fig. \ref{fig.adaptive}.
\end{construction}
Recall that the entropy estimator in \cite{Jiao--Venkat--Han--Weissman2015minimax} does not require the knowledge of $S$, we conclude that $\hat{D}_{\textrm{A}}$ always sets zero to unseen symbols and does not depend on $u(S)$. In other words, the estimator $\hat{D}$ for $D(P\|Q)$ is agnostic to both $S$ and $u(S)$, and is thus adaptive. Moreover, the estimator $\hat{D}_{\textrm{A}}$ is easy to implement in practice with near-linear computational complexity, and the coefficients $\{g_{K,k}\}_{k=1}^{K+1}$ can be obtained \emph{offline} via the Remez algorithm before observing any samples.
\begin{figure}
\centering
\begin{tikzpicture}[xscale=9,yscale=8]
\draw [<->, help lines] (0.6,1.7) -- (0.6,0.6) -- (2.0,0.6);
\draw (0.6,0.6) -- (1.6,1.6) -- (1.95,1.6);
\node [left] at (0.6,0.6) {$0$};
\draw [dashed] (1.4, 0.6) -- (1.4,1.6);
\draw [dashed] (0.6, 1.6) -- (1.6,1.6);
\node [below] at (1.4,0.6) {$\frac{c_1\ln n}{n}$};
\node [left] at (0.6,1.6) {$1$};
\node [above] at (1.7, 1.1) {``Smooth" regime:};
\node [above] at (1.7, 1.04) {plug-in approach with};
\node [above] at (1.7, 1.0) {order-three bias correction};
\node [above] at (1.0, 1.1) {``Non-smooth" regime: };
\node [above] at (1.0, 1.05) {$\hat{p}\times$ unbiased estimate of a special};
\node [above] at (1.0, 1.0) {1D polynomial approximation of $\ln q$};
\node [right] at (2.0,0.6) {$q$};
\node [above] at (0.6,1.7) {$p$};
\end{tikzpicture}
\caption{Pictorial explanation of our adaptive estimator for $\sum_{i=1}^S p_i\ln q_i$.}
\label{fig.adaptive}
\end{figure}
Now we analyze the performance of $\tilde{T}_\textrm{\rm ns}(\hat{p},\hat{q})$ when $q\le \frac{2c_1\ln n}{n}$.
\begin{lemma}\label{lem.adaptive}
Let $0\le p\le u(S)q, 0\le q\le \frac{2c_1\ln n}{n}$, and $m\hat{p}\sim \mathsf{Poi}(mp), n\hat{q}\sim \mathsf{Poi}(nq)$ be independent random variables. If $c_1\ge 2c_2$ and $c_2\ln n\ge 1$, we have
\begin{align}
|\mathbb{E}[\tilde{T}_\textrm{\rm ns}(\hat{p},\hat{q})]-p\ln q|&\le \frac{Cu(S)}{n\ln n} + \frac{16c_2^2}{n^{11c_2\ln 2}}(C+2c_1(\ln n)^3)^2\cdot \left(\frac{2u(S)}{c_1mn\ln n}+\frac{4(u(S))^2}{n^2}\right)\\
\mathsf{Var}(\tilde{T}_\textrm{\rm ns}(\hat{p},\hat{q})) &\le \frac{16c_2^2}{n^{11c_2\ln 2}}(C+2c_1(\ln n)^3)^2\cdot \left(\frac{2u(S)}{c_1mn\ln n}+\frac{4(u(S))^2}{n^2}\right).
\end{align}
where $C>0$ is a constant which only depends on $c_1$ and $c_2$. In particular, if $11c_2\ln 2<\epsilon \in(0,1)$, by \cite[Lemma 5]{Jiao--Venkat--Han--Weissman2015minimax} we have
\begin{align}
|\mathbb{E}\tilde{T}_\textrm{\rm ns}(\hat{p},\hat{q})-p\ln q| &\lesssim \frac{u(S)}{n\ln n} + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}\\
\mathsf{Var}(\tilde{T}_\textrm{\rm ns}(\hat{p},\hat{q})) &\lesssim \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}.
\end{align}
\end{lemma}
Note that in Lemma \ref{lem.adaptive} we have removed the condition $\frac{n}{\ln n}\lesssim \frac{mu(S)}{\ln m}$ in Lemma \ref{lem.T_ns_II}. Moreover, since the upper bounds of the bias and variance of $\tilde{T}_\textrm{\rm ns}$ presented in Lemma \ref{lem.adaptive} are no worse than those in Lemma \ref{lem.T_ns_1} and Lemma \ref{lem.T_ns_II}, by the same argument in Lemma \ref{lem.overall_T_ns} and Lemma \ref{lem.overall_xi} we conclude that the adaptive estimator $\hat{D}_{\textrm{A}}$ is rate-optimal, and thereby satisfies Theorem \ref{thm.kl_divergence}.
\section{Minimax Lower Bound}\label{sec.lowerbound}
In this section, we prove the minimax lower bounds presented in Theorem \ref{thm.kl_divergence}. There are two main lemmas that we employ towards the proof of the minimax lower bound. The first is the Le Cam two-point method, which helps to prove the minimax lower bound corresponding to the variance, or equivalently, the classical asymptotics. Suppose we observe a random vector ${\bf Z} \in (\mathcal{Z},\mathcal{A})$ which has distribution $P_\theta$ where $\theta \in \Theta$. Let $\theta_0$ and $\theta_1$ be two elements of $\Theta$. Let $\hat{T} = \hat{T}({\bf Z})$ be an arbitrary estimator of a function $T(\theta)$ based on $\bf Z$. Le Cam's two-point method gives the following general minimax lower bound.
\begin{lemma}\label{lem.twopoint}
\cite[Sec. 2.4.2]{Tsybakov2008} The following inequality holds:
\begin{align}
\inf_{\hat{T}} \sup_{\theta \in \Theta} \mathbb{P}_\theta\left( |\hat{T} - T(\theta)| \geq \frac{|T(\theta_1)-T(\theta_0)|}{2} \right) \geq
\frac{1}{4}\exp\left(-D\left(P_{\theta_1}\|P_{\theta_0}\right)\right).
\end{align}
\end{lemma}
The second lemma is the so-called method of two fuzzy hypotheses presented in Tsybakov \cite{Tsybakov2008}. Suppose we observe a random vector ${\bf Z} \in (\mathcal{Z},\mathcal{A})$ which has distribution $P_\theta$ where $\theta \in \Theta$. Let $\sigma_0$ and $\sigma_1$ be two prior distributions supported on $\Theta$. Write $F_i$ for the marginal distribution of $\mathbf{Z}$ when the prior is $\sigma_i$ for $i = 0,1$.
Let $\hat{T} = \hat{T}({\bf Z})$ be an arbitrary estimator of a function $T(\theta)$ based on $\bf Z$. We have the following general minimax lower bound.
\begin{lemma}\cite[Thm. 2.15]{Tsybakov2008} \label{lemma.tsybakov}
Given the setting above, suppose there exist $\zeta\in \mathbb{R}, s>0, 0\leq \beta_0,\beta_1 <1$ such that
\begin{align}
\sigma_0(\theta: T(\theta) \leq \zeta -s) & \geq 1-\beta_0 \\
\sigma_1(\theta: T(\theta) \geq \zeta + s) & \geq 1-\beta_1.
\end{align}
If $V(F_1,F_0) \leq \eta<1$, then
\begin{equation}
\inf_{\hat{T}} \sup_{\theta \in \Theta} \mathbb{P}_\theta\left( |\hat{T} - T(\theta)| \geq s \right) \geq \frac{1-\eta - \beta_0 - \beta_1}{2},
\end{equation}
where $F_i,i = 0,1$ are the marginal distributions of $\mathbf{Z}$ when the priors are $\sigma_i,i = 0,1$, respectively.
\end{lemma}
Here $V(P,Q)$ is the total variation distance between two probability measures $P,Q$ on the measurable space $(\mathcal{Z},\mathcal{A})$. Concretely, we have
\begin{equation}
V(P,Q) \triangleq \sup_{A\in \mathcal{A}} | P(A) - Q(A) | = \frac{1}{2} \int |p-q| d\nu,
\end{equation}
where $p = \frac{dP}{d\nu}, q = \frac{dQ}{d\nu}$, and $\nu$ is a dominating measure so that $P \ll \nu, Q \ll \nu$.
By the proof of the achievability results in previous sections, we observe that $(\frac{S}{m\ln m}+\frac{Su(S)}{n\ln n})^2$ corresponds to the squared bias term, and $\frac{(\ln S)^2}{m}+\frac{u(S)}{n}$ corresponds to the variance term. In the sequel we will also prove the minimax lower bound for the squared bias term and the variance term separately.
\subsection{Minimax lower bound for the ``variance"}
First we prove that when $n\gtrsim u(S)$, we have
\begin{align}\label{eq.lower_var_n}
\inf_{\hat{D}} \sup_{(P,Q) \in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left( \hat{D}-D(P\|Q)\right)^2 \gtrsim \frac{u(S)}{n}.
\end{align}
Fix $P=(\frac{1}{2(S-1)},\cdots,\frac{1}{2(S-1)},\frac{1}{2})$. Applying Lemma \ref{lem.twopoint} to our Poisson sampling model $n\hat{q}_i\sim\mathsf{Poi}(nq_i),1\le i\le S$, we know that for feasible $\theta_1=P\times Q_1=(p_1,\cdots,p_S)\times (q_1,\cdots,q_S), \theta_0=P\times Q_0=(p_1,\cdots,p_S)\times (q_1',\cdots,q_S')$ with $\theta_0,\theta_1\in \mathcal{U}_{S,u(S)}$,
\begin{align}
D\left(P_{\theta_1}\|P_{\theta_0}\right)
&= \sum_{i=1}^S D\left( \mathsf{Poi}(mp_i)\times \mathsf{Poi}(nq_i)\|\mathsf{Poi}(mp_i)\times \mathsf{Poi}(nq_i')\right)\\
&= \sum_{i=1}^S D(\mathsf{Poi}(nq_i)\| \mathsf{Poi}(nq_i'))\\
&= \sum_{i=1}^S \sum_{k=0}^\infty \mathbb{P}\left(\mathsf{Poi}(nq_i)=k\right)\cdot \left[k\ln\frac{q_i}{q_i'}-n(q_i-q_i')\right] \\
&= \sum_{i=1}^S np_i\ln\frac{q_i}{q_i'} - n\sum_{i=1}^S(q_i-q_i')\\
& = nD(Q_1\|Q_0),
\end{align}
then Markov's inequality yields
\begin{align}
\inf_{\hat{D}} \sup_{(P,Q) \in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left( \hat{D}-D(P\|Q)\right)^2
&\ge \frac{|D(\theta_1)-D(\theta_0)|^2}{4}\times \nonumber \\
& \qquad\qquad
\inf_{\hat{D}} \sup_{P \in \mathcal{M}_S} \mathbb{P}\left( |\hat{D}-D(P\|Q)| \ge \frac{|D(\theta_1)-D(\theta_0)|}{2}\right)\\
&\ge \frac{|D(P\|Q_1)-D(P\|Q_0)|^2}{16}\exp\left(-nD(Q_1\|Q_0)\right)
\end{align}
where we are operating under the Poisson sampling model.
Fix $\epsilon\in(0,1/2)$ to be specified later. Letting
\begin{align}
Q_1 & = \left(\frac{1}{(S-1)u(S)},\cdots,\frac{1}{(S-1)u(S)},1-\frac{1}{u(S)}\right), \\
Q_0 & = \left(\frac{1+\epsilon}{(S-1)u(S)},\frac{1-\epsilon}{(S-1)u(S)},\cdots,\frac{1+\epsilon}{(S-1)u(S)},\frac{1-\epsilon}{(S-1)u(S)},1-\frac{1}{u(S)}\right),
\end{align}
where without loss of generality we assume that $S$ is an odd integer. Direct computation yields
\begin{align}
D(Q_1\|Q_0) & = -\frac{1}{2u(S)}\ln(1+\epsilon) -\frac{1}{2u(S)}\ln(1-\epsilon) \\
& = -\frac{1}{2u(S)}\ln(1-\epsilon^2)\\
&\le \frac{\epsilon^2}{u(S)},
\end{align}
and
\begin{align}
|D(P\|Q_1)-D(P\|Q_0)|
&\ge -\frac{1}{4}\ln(1+\epsilon)-\frac{1}{4}\ln(1-\epsilon) \\
&= -\frac{1}{4}\ln(1-\epsilon^2) \\
& \ge \frac{\epsilon^2}{4}.
\end{align}
Hence, by choosing $\epsilon=\sqrt{\frac{u(S)}{n}}$, we know that
\begin{align}
\inf_{\hat{D}} \sup_{(P,Q) \in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left( \hat{D}-D(P\|Q)\right)^2 \ge \frac{u(S)}{256en}
\end{align}
under the Poisson sampling model. The result for the multinomial case can be obtained via Lemma \ref{lemma.poissonmultinomial}.
Next we apply Lemma \ref{lem.twopoint} to show that
\begin{align}\label{eq.lower_var_m}
\inf_{\hat{D}} \sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left(\hat{D} - D(P\|Q)\right)^2 \gtrsim \frac{(\ln u(S))^2}{m}.
\end{align}
Now fix $Q=(\frac{1}{2(S-1)u(S)},\cdots,\frac{1}{2(S-1)u(S)},1-\frac{1}{2u(S)})$, and consider
\begin{align}
P_1&=\left(\frac{1}{2(S-1)},\cdots,\frac{1}{2(S-1)},\frac{1}{2}\right)\\
P_0&=\left(\frac{1-\epsilon}{2(S-1)},\cdots,\frac{1-\epsilon}{2(S-1)},\frac{1+\epsilon}{2}\right)
\end{align}
with $\epsilon\in(0,\frac{1}{2})$ to be specified later. By the same argument, for $\theta_1=P_1\times Q\in \mathcal{U}_{S,u(S)}, \theta_0=P_0\times Q\in\mathcal{U}_{S,u(S)}$, we have
\begin{align}
\inf_{\hat{D}} \sup_{(P,Q) \in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left( \hat{D}-D(P\|Q)\right)^2 \ge \frac{|D(P_1\|Q)-D(P_0\|Q)|^2}{16}\exp\left(-mD(P_1\|P_0)\right)
\end{align}
under the Poisson sampling model. It is straightforward to compute that
\begin{align}
D(P_1\|P_0) = \frac{1}{2}\ln\frac{1}{1-\epsilon} + \frac{1}{2}\ln \frac{1}{1+\epsilon} = -\frac{1}{2}\ln (1-\epsilon^2) \le \frac{\epsilon^2}{4}
\end{align}
and
\begin{align}
|D(P_1\|Q) - D(P_0\|Q)| &= \left| \frac{1}{2}\ln u(S) + \frac{1}{2}\ln \frac{u(S)}{2u(S)-1} - \frac{1-\epsilon}{2}(\ln u(S) + \ln(1-\epsilon)) - \frac{1+\epsilon}{2}\ln \frac{(1+\epsilon)u(S)}{2u(S)-1}\right|\\
&= \left| \frac{\epsilon}{2}\ln u(S) - \frac{1+\epsilon}{2}\ln(1+\epsilon) - \frac{1-\epsilon}{2}\ln(1-\epsilon) - \frac{\epsilon}{2}\ln\frac{u(S)}{2u(S)-1} \right|\\
&\ge \frac{\epsilon \ln u(S)}{2}.
\end{align}
Combining these inequalities and setting $\epsilon=m^{-\frac{1}{2}}$ completes the proof of (\ref{eq.lower_var_m}).
\subsection{Minimax lower bound for the ``squared bias"}
We employ Lemma \ref{lemma.tsybakov} to prove the minimax lower bounds corresponding to the squared bias terms. First we show that when $m\gtrsim \frac{S}{\ln S}$ and $u(S)\gtrsim (\ln S)^2, \ln S\gtrsim \ln m$, we have
\begin{align}\label{eq.lower_bias_m}
\inf_{\hat{D}} \sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left(\hat{D} - D(P\|Q)\right)^2 \gtrsim \left(\frac{S}{m\ln m}\right)^2.
\end{align}
In fact, by choosing $Q$ to be the uniform distribution, the estimation of KL divergence reduces to the estimation of entropy of discrete distribution $P$, subject to an additional constraint $p_i\le u(S)q_i = \frac{u(S)}{S}$ for all $i=1,\cdots,S$. Since in the proof of the minimax lower bound in \cite{Wu--Yang2014minimax}, all $p_i$ satisfy that $p_i\lesssim \frac{\ln m}{m}$, and our assumption implies
\begin{align}
p_i \lesssim \frac{\ln m}{m} \lesssim \frac{(\ln S)^2}{S}\lesssim \frac{u(S)}{S}
\end{align}
i.e., the additional condition is automatically satisfied. Hence, we can operate as if we do not have the additional condition, and \cite{Wu--Yang2014minimax} gives
\begin{align}
\inf_{\hat{D}} \sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left(\hat{D} - D(P\|Q)\right)^2 \gtrsim \left(\frac{S}{m\ln S}\right)^2\asymp \left(\frac{S}{m\ln m}\right)^2.
\end{align}
where we have used the condition $m\gtrsim \frac{S}{\ln S}$ and $\ln S\gtrsim \ln m$ to give $\ln S\asymp \ln m$ here.
Now we are about the prove that when $m\gtrsim \frac{S}{\ln S}, n\gtrsim \frac{Su(S)}{\ln S}$ and $u(S)\gtrsim (\ln S)^2, \ln S\gtrsim \ln n$, we have
\begin{align}\label{eq.lower_bias_n}
\inf_{\hat{D}} \sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left(\hat{D} - D(P\|Q)\right)^2 \gtrsim \left(\frac{Su(S)}{n\ln n}\right)^2.
\end{align}
We begin with a lemma to construct two measures with matching moments and large difference on the functional value, which corresponds to the duality between function space and measure space.
\begin{lemma}\label{lem.measure}
\cite[Lemma 10, Lemma 12]{Jiao--Venkat--Han--Weissman2015minimax} For any bounded interval $I\subset \mathbb{R}$, positive integer $K>0$ and continuous function $f$ on $I$, there exist two probability measures $\nu_0$ and $\nu_1$ supported on $I$ such that
\begin{enumerate}
\item $\int t^{l} \nu_1(dt) = \int t^{l} \nu_0(dt)$, for all $l=0,1,2,\cdots,L$;
\item $\int f(t) \nu_1(dt) - \int f(t) \nu_0(dt) = 2E_K[f;I]$.
\end{enumerate}
Recall that $E_K[f;I]$ is the distance in the uniform norm on $I$ from the function $f(x)$ to the space spanned by $\{1,x,\cdots,x^K\}$.
\end{lemma}
Based on Lemma \ref{lem.measure}, we choose $I=[\frac{1}{n\ln n},\frac{d_1\ln n}{n}]$, $K=d_2\ln n$ with universal constants $d_1,d_2>0$ to be specified later, and $f(x)=\ln x$. The following lemma presents a lower bound of the approximation error $E_K[f;I]$.
\begin{lemma}\label{lem.approx_log_lower}
\cite{Wu--Yang2014minimax} For $K=d_2\ln n$, $I=[\frac{1}{n\ln n},\frac{d_1\ln n}{n}]$ and $f(x)=\ln x$, we have
\begin{align}
E_K[f;I] \ge c'
\end{align}
where the constant $c'>0$ only depends on $d_1,d_2$.
\end{lemma}
Define $\mu\triangleq \int t\nu_1(dt)=\int t\nu_0(dt)$, by construction we have $\mu\le \frac{d_1\ln n}{n}$. Now the two fuzzy hypotheses $\sigma_0,\sigma_1$ in Lemma \ref{lemma.tsybakov} are constructed as follows: for each $i=0,1$, $\sigma_i$ fixes
\begin{align}
P = \left( \frac{u(S)}{n\ln n}, \cdots, \frac{u(S)}{n\ln n}, 1-\frac{(S-1)u(S)}{n\ln n} \right)
\end{align}
and assigns $\nu_i^{S-1}$ to the vector $(q_1,\cdots,q_{S-1})$, and fixes $q_S=1-(S-1)\mu$. Note that by assumption,
\begin{align}
S\mu \lesssim \frac{S\ln n}{n} \lesssim \frac{(\ln S)^2}{u(S)}\lesssim 1
\end{align}
thus $q_S$ takes positive value and is thus valid under proper parameter configurations. Moreover, it is straightforward to verify that $(P,Q)\in \mathcal{U}_{S,u(S)}$ with probability one under $\sigma_i$. Since under $\sigma_i$, $Q$ may not form a probability measure, we consider the set of approximate probability vectors
\begin{align}
\mathcal{U}_{S,u(S)}(\epsilon) \triangleq \left\{ (P,Q): P\in \mathcal{M}_S, |\sum_{i=1}^S q_i - 1|<\epsilon, p_i\le u(S)q_i, \forall i \right\}
\end{align}
with parameter $\epsilon\in(0,1)$ to be specified later. Further define the minimax under the Poisson sampling model for estimating $D(P\|Q)$ with $(P,Q)\in\mathcal{U}_{S,u(S)}(\epsilon)$ as
\begin{align}
R_P(S,m,n,u(S),\epsilon) \triangleq \inf_{\hat{D}}\sup_{(P,Q)\in \mathcal{U}_{S,u(S)}(\epsilon)}\mathbb{E}_{(P,Q)}\left(\hat{D}-D(P\|Q)\right)^2.
\end{align}
The equivalence between $R_P(S,m,n,u(S),\epsilon)$ and $R(S,m,n,u(S))$ defined in (\ref{eq:minimaxrisk.multi}) is established in the following lemma.
\begin{lemma}\label{lem.equivalence_approx}
For any $S,m,n\in\mathbb{N}$ and $\epsilon\in(0,1/2)$, we have
\begin{align}
R(S,m,n,u(S)) \ge \frac{1}{2}R_P(S,2m,2n,\frac{u(S)}{1+\epsilon},\epsilon) - (\ln u(S))^2\left(\exp(-\frac{m}{4}) + \exp(-\frac{n}{4})\right) - 2\epsilon^2.
\end{align}
\end{lemma}
Then condition $\sigma_i$ on the event
\begin{align}
E_i \triangleq \mathcal{U}_{S,u(S)}(\epsilon) \cap \left\{ (P,Q): \frac{1}{S}\left|\sum_{j=1}^S (\ln q_j - \mathbb{E}_{\sigma_i}\ln q_j)\right| \le \frac{E_K[f;I]}{2} \right\}, \qquad i=0,1
\end{align}
and define the conditional probability distribution as
\begin{align}
\pi_i(\cdot) \triangleq \frac{\sigma_i(\cdot \cap E_i)}{\sigma_i(E_i)}, \qquad i=0,1.
\end{align}
By setting
\begin{align}
d_1 = 1,\qquad d_2=10e,\qquad \epsilon=\frac{S}{n\ln n}
\end{align}
we have
\begin{align}
\sigma_i(\mathcal{U}_{S,u(S)}(\epsilon)^c) = \sigma_i\left(\left|\sum_{j=1}^S q_j-1\right|\ge \epsilon\right)\le \frac{1}{\epsilon^2}\sum_{j=1}^{S-1} \mathsf{Var}_{\nu_i}(q_j) \le \frac{1}{\epsilon^2}\sum_{j=1}^{S-1} \mathbb{E}_{\nu_i}(q_j^2) \le \frac{S}{\epsilon^2}\left(\frac{d_1\ln n}{n}\right)^2 = \frac{d_1^2(\ln n)^4}{S} \to 0
\end{align}
and by Lemma \ref{lem.approx_log_lower},
\begin{align}
\sigma_i\left( \frac{1}{S}\left|\sum_{j=1}^S (\ln q_j - \mathbb{E}_{\sigma_i}\ln q_j)\right| \ge \frac{E_K[f;I]}{2} \right)\le \frac{4}{(c'S)^2}\sum_{j=1}^{S-1} \mathsf{Var}_{\nu_i}(\ln q_j) \le \frac{4}{(c')^2}\cdot \frac{(\ln (n\ln n))^2}{S} \lesssim \frac{(\ln S)^2}{S} \to 0.
\end{align}
Hence, by the union bound,
\begin{align}
\sigma_i(E_i^c) \le \sigma_i(\mathcal{U}_{S,u(S)}(\epsilon)^c) + \sigma_i\left( \frac{1}{S}\left|\sum_{j=1}^S (\ln q_j - \mathbb{E}_{\sigma_i}\ln q_j)\right| \ge \frac{E_K[f;I]}{2} \right) \to 0.
\end{align}
Denote by $F_i,G_i$ the marginal probability under prior $\pi_i$ and $\sigma_i$, respectively, for each $i=0,1$. Now by the triangle inequality and \cite[Lemma 11]{Jiao--Venkat--Han--Weissman2015minimax}, we have
\begin{align}
V(F_0,F_1)&\le V(F_0,G_0) + V(G_0,G_1) + V(G_1,F_1)\\
&\le \sigma_0(E_0^c) + V(G_0,G_1) + \sigma_1(E_1^c)\\
&\le \sigma_0(E_0^c) + \frac{S}{n^6} + \sigma_1(E_1^c)\\
&\to 0.
\end{align}
Moreover, by the definition of $\pi_i$, the first two conditions of Lemma \ref{lemma.tsybakov} hold with $\beta_0=\beta_1=0$ for
\begin{align}
\zeta &= H(P) - \frac{(S-1)u(S)}{n\ln n}\cdot \frac{\mathbb{E}_{\nu_0}(\ln q)+\mathbb{E}_{\nu_1}(\ln q)}{2} - \left(1-\frac{(S-1)u(S)}{n\ln n}\right)\ln (1-(S-1)\mu)\\
s &= \frac{(S-1)u(S)}{n\ln n}\cdot \frac{\mathbb{E}_{\nu_1}(\ln q)-\mathbb{E}_{\nu_0}(\ln q)-2\cdot \frac{E_K[f;I]}{2}}{2} = \frac{(S-1)u(S)}{n\ln n}\cdot \frac{E_K[f;I]}{2} \gtrsim \frac{Su(S)}{n\ln n}.
\end{align}
Hence, by Lemma \ref{lemma.tsybakov}, we conclude that
\begin{align}
R_P(S,m,n,u(S),\epsilon) \gtrsim s^2 \gtrsim \left(\frac{Su(S)}{n\ln n}\right)^2
\end{align}
and the desired bound (\ref{eq.lower_bias_n}) follows from Lemma \ref{lem.equivalence_approx}.
Hence, the combination of (\ref{eq.lower_var_n}), (\ref{eq.lower_var_m}), (\ref{eq.lower_bias_m}) and (\ref{eq.lower_bias_n}) yields that when $u(S)\gtrsim (\ln S)^2, n\gtrsim \frac{Su(S)}{\ln S}, m\gtrsim \frac{S}{\ln S}$ and $\ln S\gtrsim \ln(m\vee n)$, we have
\begin{align}
\inf_{\hat{D}} \sup_{(P,Q) \in \mathcal{U}_{S,u(S)}} \mathbb{E}_{(P,Q)}\left( \hat{D}-D(P\|Q)\right)^2 \gtrsim \left(\frac{Su(S)}{n\ln n}\right)^2+ \left(\frac{S}{m\ln m}\right)^2 + \frac{(\ln u(S))^2}{m} + \frac{u(S)}{n}
\end{align}
and the proof of Theorem \ref{thm.kl_divergence} is complete.
\section{Optimal Estimators for Hellinger Distance and $\chi^2$-Divergence}\label{sec.others}
Having analyzed the minimax rate-optimal estimator for the KL divergence thoroughly, in this section we apply the general recipe to other divergence functions, i.e., the Hellinger distance and the $\chi^2$-divergence. Specifically, we explicitly construct the minimax rate-optimal estimators for the Hellinger distance and the $\chi^2$-divergence, and sketch the proof of the achievability part in Theorem \ref{thm.hellinger} and \ref{thm.chi_squared}. For brevity, we omit the complete proof and remark that it can be obtained in a similar fashion as in the analysis of the KL divergence.
\subsection{Optimal estimator for the Hellinger distance}
For the Hellinger distance, the bivariate function of interest is $f(p,q)=(\sqrt{p}-\sqrt{q})^2$. We first classify the regime. In this case, $\Theta=\hat{\Theta}=[0,1]^2$, and the non-analytic regime is
\begin{align}
\hat{\Theta}_0 = \{(p,q)\in [0,1]^2: p=0 \text{ or } q=0\}.
\end{align}
Based on the confidence sets in Poisson models, the ``smooth" and ``non-smooth" regimes for $(p,q)$ can obtained via \eqref{eq.para_smooth} and \eqref{eq.para_nonsmooth} as
\begin{align}
\Theta_\textrm{\rm s} &= ([0,1] \times [\frac{c_1\ln n}{2n}, 1]) \cup ([\frac{c_1\ln m}{2m}, 1]\times [0,1])\\
\Theta_\textrm{\rm ns} &= ([0,1] \times [0,\frac{2c_1\ln n}{n}]) \cup ([0,\frac{2c_1\ln m}{m}]\times [0,1])
\end{align}
where $c_1>0$ is some universal constant. Further, by \eqref{eq.est_smooth} and \eqref{eq.est_nonsmooth}, we obtain the ``smooth" and ``non-smooth" regimes based on observations $(\hat{p},\hat{q})$ as
\begin{align}
\hat{\Theta}_\textrm{\rm s} &= ([0,1] \times [\frac{c_1\ln n}{n}, 1]) \cup ([\frac{c_1\ln m}{m}, 1]\times [0,1])\\
\hat{\Theta}_\textrm{\rm ns} &= ([0,1] \times [0,\frac{c_1\ln n}{n}]) \cup ([0,\frac{c_1\ln m}{m}]\times [0,1]).
\end{align}
Next we estimate the quantity $f(p,q)=(\sqrt{p}-\sqrt{q})^2$ in each regime. In the ``smooth" regime where $p\ge \frac{c_1\ln m}{m}$ and $q\ge \frac{c_1\ln n}{n}$, we simply employ the plug-in approach with no bias correction:
\begin{align}
T_\textrm{\rm s}(\hat{p},\hat{q}) = (\sqrt{\hat{p}}-\sqrt{\hat{q}})^2.
\end{align}
In the ``non-smooth" regime, by symmetry it suffices to consider the case where $p\le \frac{c_1\ln m}{m}$. Now we need to find a proper polynomial $P(p,q)$ to approximate $f(p,q)=(\sqrt{p}-\sqrt{q})^2=p+q-2\sqrt{pq}$. Recall that the degree of the approximating polynomial $P(p,q)$ is determined by the bias-variance tradeoff, and we will have the following result after careful analysis:
\begin{itemize}
\item when $\hat{q}\ge \frac{c_1\ln n}{n}$, the resulting polynomial $P(p,q)$ should be of degree $c_2\ln m$ on $p$ and of degree $0$ on $q$;
\item when $\hat{q}< \frac{c_1\ln n}{n}$, the resulting polynomial $P(p,q)$ should be of degree $c_2\ln m$ on $p$ and of degree $c_2\ln n$ on $q$.
\end{itemize}
Now we explicitly give the expression of $P(p,q)$ in both cases in terms of the following best approximating polynomial of $\sqrt{x}$:
\begin{align}\label{eq.poly_sqrt}
Q_K(x) = \sum_{k=0}^K a_{K,k}x^k \triangleq \arg\min_{Q\in \mathsf{poly}_K^1}\max_{z\in [0,1]}|Q(z)-\sqrt{z}|.
\end{align}
Recall that we use the sample splitting technique to determine the approximation region and approximate the functional, respectively, i.e., the approximation region is based on $(\hat{p}_2,\hat{q}_2)$, while the polynomial is evaluated using $(\hat{p}_1,\hat{q}_1)$. Hence,
\begin{itemize}
\item when $\hat{q}\ge \frac{c_1\ln n}{n}$, the approximation region is $[0,\frac{2c_1\ln m}{m}]\times [\hat{q}_2-\sqrt{\frac{c_1\hat{q}_2\ln n}{n}},\hat{q}_2+\sqrt{\frac{c_1\hat{q}_2\ln n}{n}}]$, and by the degree requirements of $P(p,q)$, a natural choice is $P(p,q)= \sqrt{\Delta_m}Q_{K_m}(\frac{p}{\Delta_m})\cdot \sqrt{\hat{q}_2}$ with $K_m=c_2\ln m, \Delta_m=\frac{2c_1\ln m}{m}$;
\item when $\hat{q}< \frac{c_1\ln n}{n}$, the approximation region is $[0,\frac{2c_1\ln m}{m}]\times [0,\frac{2c_1\ln n}{n}]$, and by the degree requirements of $P(p,q)$, a natural choice is $P(p,q)= \sqrt{\Delta_m}Q_{K_m}(\frac{p}{\Delta_m})\cdot \sqrt{\Delta_n}Q_{K_n}(\frac{q}{\Delta_n})$ with $K_m=c_2\ln m, \Delta_m=\frac{2c_1\ln m}{m}$ and $K_n=c_2\ln n, \Delta_n=\frac{2c_1\ln n}{n}$.
\end{itemize}
In summary, the estimator is constructed as follows.
\begin{construction}
Conduct three-fold sample splitting to obtain i.i.d. samples $(\hat{p}_{i,1},\hat{p}_{i,2},\hat{p}_{i,3})$ and $(\hat{q}_{i,1},\hat{q}_{i,2},\hat{q}_{i,3})$. The estimator $\hat{T}$ for the Hellinger distance $H^2(P,Q)$ is given by
\begin{align}
\hat{T} = 1 - \sum_{i=1}^S\left(\sqrt{\hat{p}_{i,2}}\mathbbm{1}(\hat{p}_{i,3}\ge \frac{c_1\ln m}{m})+\tilde{R}_m(\hat{p}_{i,1})\mathbbm{1}(\hat{p}_{i,3}< \frac{c_1\ln m}{m})\right)\left(\sqrt{\hat{q}_{i,2}}\mathbbm{1}(\hat{q}_{i,3}\ge \frac{c_1\ln n}{n})+\tilde{R}_n(\hat{q}_{i,1})\mathbbm{1}(\hat{q}_{i,3}< \frac{c_1\ln n}{n})\right)
\end{align}
where for $l=m,n$
\begin{align}
\tilde{R}_l(x) &\triangleq (R_l(x)\wedge 1)\vee(-1)\\
R_l(x) &\triangleq \sum_{k=1}^{c_2\ln l} a_{c_2\ln l,k}\left(\frac{2c_1\ln l}{l}\right)^{-k+\frac{1}{2}}\prod_{j=0}^{k-1}(x-\frac{j}{l})
\end{align}
coefficients $\{a_{K,k}\}$ are given by \eqref{eq.poly_sqrt}, and $c_1,c_2>0$ are some suitably chosen universal constants.
\end{construction}
The previous estimator does not require the knowledge of $S$ since it assigns zero to unseen symbols (we remove the constant term in the expression of $R_l(x)$). Note that since the Hellinger distance $H^2(P,Q)=1-\sum_{i=1}^S\sqrt{p_iq_i}$ enjoys a natural separation in its variables $(p_i,q_i)$, the pair $(p_i,q_i)$ is also separated in our resulting estimator. Moreover, in our estimator construction we can also merge $(\hat{p}_{i,1},\hat{p}_{i,2})$ and $(\hat{q}_{i,1},\hat{q}_{i,2})$ to result in a two-fold sample splitting.
Now we analyze the bias of the previous estimator. When $p<\frac{2c_1\ln m}{m}$ is small, by \cite[Lemma 17]{Jiao--Venkat--Han--Weissman2015minimax}, we know that
\begin{align}
|\mathbb{E} R_m(\hat{p}) - \sqrt{p}| = |\sqrt{\Delta_m}Q_{K_m}(\frac{p}{\Delta_m})-\sqrt{p}| \lesssim \frac{\sqrt{\Delta_m}}{K_m^2} \lesssim \frac{1}{\sqrt{m\ln m}}.
\end{align}
When $p\ge \frac{2c_1\ln m}{m}$ is large, by Lemma \ref{lem.plugin_upper} with $r=0$ we know that
\begin{align}
|\mathbb{E} \sqrt{\hat{p}}-\sqrt{p}|\lesssim \frac{p}{m}\cdot \sup_{\xi\in [p/2,2p]}\left|\frac{d^2\sqrt{\xi}}{d\xi^2} \right|\lesssim \frac{1}{m\sqrt{p}}\lesssim \frac{1}{\sqrt{m\ln m}}.
\end{align}
Hence, the total bias of $\hat{T}$ can be upper bounded as
\begin{align}
|\mathsf{Bias}(\hat{T})|\lesssim \sum_{i=1}^S \left(\frac{\sqrt{q_i}}{\sqrt{m\ln m}}+\frac{\sqrt{p_i}}{\sqrt{n\ln n}}\right)\le \sqrt{\frac{S}{m\ln m}} + \sqrt{\frac{S}{n\ln n}} \asymp \sqrt{\frac{S}{(m\wedge n)\ln(m\wedge n)}}
\end{align}
as shown in Theorem \ref{thm.hellinger}. The variance can also be obtained in a similar fashion, and we omit the details.
\subsection{Optimal estimator for the $\chi^2$-divergence}
For the $\chi^2$-divergence $\chi^2(P,Q)$ over $\mathcal{U}_{S,u(S)}$, the bivariate function of interest is $f(p,q)=\frac{p^2}{q}$. We first classify the regime. Since this function shares very similar analytic properties as the function $p\ln q$ used in the KL divergence case, and our parameter set $\mathcal{U}_{S,u(S)}$ remains the same, here the ``smooth" and ``non-smooth" regimes are also given by
\begin{align}
\hat{\Theta}_\textrm{\rm s} &= [0,1] \times [\frac{c_1\ln n}{n},1]\\
\hat{\Theta}_\textrm{\rm ns} &= [0,1] \times [0,\frac{c_1\ln n}{n}]
\end{align}
with some universal constant $c_1>0$.
Next we estimate $f(p,q)=\frac{p^2}{q}$ in each regime. In the ``smooth" regime where $q\ge \frac{c_1\ln n}{n}$, we seek to correct the bias of the plug-in estimator $1/\hat{q}$ in estimating $1/q$. Based on our general bias correction technique \eqref{eq.bias_cor}, we simply use the following order-three bias correction:
\begin{align}
\hat{T}^{(3)}(\hat{q}_1,\hat{q}_2) = \frac{1}{\hat{q}_1} - \frac{\hat{q}_2-\hat{q}_1}{\hat{q}_1^2} + \frac{(\hat{q}_2-\hat{q}_1)^2}{\hat{q}_1^3} - \frac{4\hat{q}_2}{n\hat{q}_1^3} - \frac{(\hat{q}_2-\hat{q}_1)^3}{\hat{q}_1^4} + \frac{3\hat{q}_2^2}{n\hat{q}_1^4} - \frac{6\hat{q}_2}{n^2\hat{q}_1^4}.
\end{align}
Since $p^2$ admits an unbiased estimate $\hat{p}(\hat{p}-\frac{1}{m})$ in the Poisson model $m\hat{p}\sim\mathsf{Poi}(mp)$, the overall estimator in the ``smooth" regime is given by
\begin{align}
\hat{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) &= \hat{p}_1\left(\hat{p}_1-\frac{1}{m}\right)\cdot \hat{T}^{(3)}(\hat{q}_1,\hat{q}_2) \cdot\mathbbm{1}(\hat{q}_1\neq 0)\\\label{eq.chi_smooth}
&= \hat{p}_1\left(\hat{p}_1-\frac{1}{m}\right)\cdot\left(\frac{1}{\hat{q}_1} - \frac{\hat{q}_2-\hat{q}_1}{\hat{q}_1^2} + \frac{(\hat{q}_2-\hat{q}_1)^2}{\hat{q}_1^3} - \frac{4\hat{q}_2}{n\hat{q}_1^3} - \frac{(\hat{q}_2-\hat{q}_1)^3}{\hat{q}_1^4} + \frac{3\hat{q}_2^2}{n\hat{q}_1^4} - \frac{6\hat{q}_2}{n^2\hat{q}_1^4}\right)\cdot\mathbbm{1}(\hat{q}_1\neq 0).
\end{align}
In the ``non-smooth" regime, as is in the KL divergence case, we can further distinguish into ``non-smooth" regime I and ``non-smooth" regime II and employ best polynomial approximation in these regimes, respectively. However, motivated by the adaptive estimator $\hat{D}_{\textrm{A}}$ for the KL divergence where a single polynomial approximation is enough, we wonder whether or not it is also the case in the estimation of $\chi^2$-divergence. Specifically, we seek a polynomial $Q_K(q)$ of degree $K$ on $[0,\Delta_n]$ such that the following quantity
\begin{align}
\sup_{q\in (0,\Delta_n]} q^2\left|\frac{1}{q}-Q_K(q)\right|
\end{align}
is as small as possible, where $\Delta_n=\frac{2c_1\ln n}{n}$. In other words, we seek to approximate the linear function $q$ using $q^2,\cdots,q^{K+2}$ on $[0,\Delta_n]$. Fortunately, this task can be done with the help of the Chebyshev polynomial, and we summarize the result in the following lemma.
\begin{lemma}\label{lem.chebyshev_poly}
Let $T_K(x)=\cos(K\arccos x)$ be the degree-$K$ Chebyshev polynomial on $[-1,1]$. Then
\begin{align}
Q_K(x) = (-1)^K \frac{T_{2(K+2)}(\sqrt{x/\Delta_n})-(-1)^{K+2}-2(K+2)^2x/\Delta_n}{2(K+2)^2(x/\Delta_n)^2}
\end{align}
is a degree-$K$ polynomial such that
\begin{align}
\sup_{x\in (0,\Delta_n]} x^2\left|\frac{1}{x}-Q_K(x)\right| = \frac{\Delta_n}{(K+2)^2}.
\end{align}
\end{lemma}
On the other hand, using the Ditzian--Totik modulus of smoothness and Lemma \ref{lem.DT_modulus} (or by Chebyshev's alternating theorem since $\{x^2,x^3,\cdots\}$ satisfies the Haar condition \cite{haar1917minkowskische}), it is not hard to prove that no other degree-$K$ polynomial can achieve an approximation error of order $o(\frac{\Delta_n}{K^2})$. Hence, the polynomial $Q_K$ defined in Lemma \ref{lem.chebyshev_poly} achieves the rate-optimal uniform approximation error.
Motivated by Lemma \ref{lem.chebyshev_poly}, define $\hat{Q}_K(x)$ be another polynomial such that $\mathbb{E}\hat{Q}_K(\hat{q})=Q_K(q)$ for $n\hat{q}\sim\mathsf{Poi}(nq)$. Now in the ``non-smooth" regime, we choose $K=c_2\ln n$, and use the following estimator:
\begin{align}\label{eq.chi_nonsmooth}
\hat{T}_\textrm{\rm ns}(\hat{p}_{1},\hat{q}_1) = \hat{p}_1\left(\hat{p}_1-\frac{1}{m}\right)\cdot \hat{Q}_K(\hat{q}_1).
\end{align}
In summary, we have arrived at the following estimator construction.
\begin{construction}
Conduct three-fold sample splitting to obtain i.i.d. samples $(\hat{p}_{i,1},\hat{p}_{i,2},\hat{p}_{i,3})$ and $(\hat{q}_{i,1},\hat{q}_{i,2},\hat{q}_{i,3})$. The estimator $\hat{T}$ for $\chi^2$-divergence $\chi^2(P,Q)$ is given by
\begin{align}
\hat{T} = \sum_{i=1}^S \left(\hat{T}_\textrm{\rm s}(\hat{p}_{i,1},\hat{q}_{i,1};\hat{p}_{i,2},\hat{q}_{i,2})\mathbbm{1}(\hat{q}_{i,3}\ge \frac{c_1\ln n}{n}) + \tilde{T}_\textrm{\rm ns}(\hat{p}_{i,1},\hat{q}_{i,1})\mathbbm{1}(\hat{q}_{i,3}<\frac{c_1\ln n}{n})\right) - 1
\end{align}
where
\begin{align}
\tilde{T}_\textrm{\rm ns}(x,y) \triangleq (\hat{T}_\textrm{\rm ns}(x,y)\wedge 1)\vee(-1)
\end{align}
and $\hat{T}_\textrm{\rm s}(x,y;x',y'), \hat{T}_\textrm{\rm ns}(x,y)$ are given by \eqref{eq.chi_smooth} and \eqref{eq.chi_nonsmooth}, respectively, and $c_1,c_2>0$ are some suitably chosen universal constants.
\end{construction}
By construction, the previous estimator does not require the knowledge of $S$ nor $u(S)$, and is thus adaptive. For the analysis of its performance, when $q<\frac{c_1\ln n}{n}$ is small, with the help of Lemma \ref{lem.chebyshev_poly} we know that
\begin{align}
\left|\mathbb{E} \hat{T}_\textrm{\rm ns}(\hat{p},\hat{q}) - \frac{p^2}{q}\right| =\left|p^2Q(q) - \frac{p^2}{q}\right|\le (u(S))^2q^2\left|Q(q)-\frac{1}{q}\right|\lesssim \frac{(u(S))^2}{n\ln n}.
\end{align}
Moreover, for $q\ge\frac{c_1\ln n}{n}$ is large, applying Lemma \ref{lem.plugin_upper} with $r=3$ yields
\begin{align}
\left|\mathbb{E} \hat{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) - \frac{p^2}{q}\right|=p^2\left|\mathbb{E} \hat{T}^{(3)}(\hat{q}_1,\hat{q}_2)-\frac{1}{q}\right|\lesssim p^2\cdot (\frac{q}{n})^2\sup_{\xi\in[q/2,2q]}\left|\frac{d^4(\xi^{-1})}{d\xi^4}\right|\lesssim \frac{p^2}{n^2q^3}\lesssim \frac{(u(S))^2}{n^2q}\lesssim \frac{(u(S))^2}{n\ln n}.
\end{align}
Hence, the total bias of the previous estimator can be upper bounded as
\begin{align}
|\mathsf{Bias}(\hat{T})| \lesssim \sum_{i=1}^S \frac{(u(S))^2}{n\ln n} = \frac{S(u(S))^2}{n\ln n}
\end{align}
which coincides with the term in Theorem \ref{thm.chi_squared}. The variance can be dealt with analogously, and we omit the lengthy proofs.
\section{Conclusions and Future Work}\label{sec.conclusion}
We proposed a general and detailed methodology for the construction of minimax rate-optimal estimators for low-dimensional functionals of high-dimensional parameters, especially when the functional of interest is non-smooth in some part of its domain. We elaborate on the insights of \cite{Jiao--Venkat--Han--Weissman2015minimax} which shows that the bias is the dominating term in the estimation of functionals and approximation is the key for an efficient bias reduction, and find an interesting interplay between the functional itself and the statistical model. Specifically, we show that the ``smooth" and ``non-smooth" regimes are determined by both the non-analytic region of the underlying functional which is only related to the smoothness of the functional, and the confidence sets given by concentration of measures which solely depend on the statistical model. Moreover, in the ``non-smooth" regime, the approximation region is determined by the confidence sets, while the approximation error is determined by the smoothness of the functional in this region. Our general recipe is based on the interplay between these two factors, and successfully yields the minimax rate-optimal estimators for various divergences including KL divergence, Hellinger distance and $\chi^2$-divergence.
We have also explored the ideas behind the polynomial approximation and the plug-in approach in bias reduction. For polynomial approximation, the uniform approximation error corresponds to the bias of the resulting estimator, and thus the best approximating polynomials are usually used. We remark that it is a highly non-trivial task and remains open in general to obtain and analyze the best polynomial approximation error for multivariate functionals, while for some special cases (e.g., general polytopes, balls and spheres) there are powerful tools from approximation theory. The plug-in approach corrects the bias with the help of high-order Taylor expansions, which only works for the region where the functional is analytic. For bias correction of the plug-in approach, in this paper we propose a general unbiased estimator of the Taylor series up to an arbitrary order.
Following \cite{Jiao--Venkat--Han--Weissman2015minimax}, this paper presents another second step towards a general theory of functional estimation. Despite our progress, the interplay between the smoothness of the functional and the statistical model has yet to be completely revealed, and the choice of the approximating polynomial in the ``non-smooth" regime has thus far required functional-specific ``tricks''. An ambitious but worthy goal is to establish a general explanation of the effective sample size enlargement phenomenon in the parametric case, and to find the counterpart in the estimation of non-smooth nonparametric functionals beyond the insights provided by \cite{Lepski--Nemirovski--Spokoiny1999estimation,Han--Jiao--Mukherjee--Weissman2016optimal}.
\appendices
\section{Auxiliary Lemmas}\label{appen.aux}
We first prove that the worst-case mean squared error of any estimator is infinity if we allow to choose any $P$ which is absolutely continuous with respect to $Q$.
\begin{lemma}\label{lem.minimax_unbounded}
Let $\mathcal{U}_S=\{(P,Q): P,Q\in\mathcal{M}_S, P\ll Q\}$. Then for any configuration $(S,m,n)$ with $S\ge 2$, we have
\begin{align}
R_{\text{\rm minimax}}(\mathcal{U}_S) = \infty.
\end{align}
\end{lemma}
The next lemma relates the minimax risk under the Poisson sampling model and that under the Multinomial model. We define the minimax risk for Multinomial model with $(m,n)$ observations with $(P,Q)\in \mathcal{U}_{S,u(S)}$ for estimating the KL divergence $D(P\|Q)$ as
\begin{equation}\label{eq:minimaxrisk.multi}
R(S,m,n,u(S)) \triangleq \inf_{\hat{D}} \sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{\mathrm{Multinomial}} \left( \hat{D} - D(P\|Q) \right)^2,
\end{equation}
and the counterpart for the Poisson sampling model as
\begin{equation}
R_P(S,m,n,u(S)) \triangleq \inf_{\hat{D}} \sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{\mathrm{Poisson}} \left( \hat{D} - D(P\|Q) \right)^2.
\end{equation}
\begin{lemma}\label{lemma.poissonmultinomial}
The minimax risks under the Poisson sampling model and the Multinomial model are related via the following inequalities:
\begin{equation}
R_P(S,2m,2n,u(S)) - (\ln u(S))^2\left(\exp(-\frac{m}{4})+\exp(-\frac{n}{4})\right) \leq R(S,m,n,u(S)) \leq 4R_P(S,\frac{m}{2},\frac{n}{2},u(S)).
\end{equation}
\end{lemma}
The next lemma gives the approximation properties of $\ln x$.
\begin{lemma}\label{lem.approx_log}
There exists a universal constant $C_{\ln}>0$ such that for any $0<a<b$,
\begin{align}
E_n[\ln x;[a,b]] \le C_{\ln}W(\frac{b}{an^2}) \equiv C_{\ln}\cdot \begin{cases}
\frac{b}{ean^2} & b\le ean^2,\\
\ln(\frac{b}{an^2}) & b>ean^2.
\end{cases}
\end{align}
\end{lemma}
\begin{lemma}\label{lem.approx_2D}
For $f(p,q)=p\ln q$ and the region $R$ given in (\ref{eq.2D_approx_region}), there exists a universal constant $C_0$ only depending on $c_1$ such that
\begin{align}
\omega_R^2(f,\frac{1}{K}) \le C_0\cdot \frac{u(S)\ln n}{K^2n}.
\end{align}
\end{lemma}
The following lemma gives an upper bound for the second moment of the unbiased estimate of $(p-q)^j$ in Poisson model.
\begin{lemma}\cite{Jiao--Han--Weissman2015divergence}\label{lemma.middlevariancebound}
Suppose $nX \sim \mathsf{Poi}(np),p\geq 0, q\geq 0$. Then, the estimator
\begin{align}
g_{j,q}(X) \triangleq \sum_{k = 0}^j {j \choose k} (-q)^{j-k} \prod_{h = 0}^{k-1} \left( X - \frac{h}{n} \right)
\end{align}
is the unique unbiased estimator for $(p-q)^j,j\in \mathbb{N}$, and its second moment is given by
\begin{align}
\mathbb{E}[g_{j,q}(X)^2] & = \sum_{k = 0}^j {j \choose k}^2 (p-q)^{2(j-k)} \frac{p^k k!}{n^k} \\
& = j! \left( \frac{p}{n} \right)^j L_j\left( - \frac{n(p-q)^2}{p} \right)\quad \text{assuming }p>0,
\end{align}
where $L_m(x)$ stands for the Laguerre polynomial with order $m$, which is defined as:
\begin{align}
L_m(x) = \sum_{k = 0}^m {m \choose k} \frac{(-x)^k}{k!}
\end{align}
If $M \geq \frac{n(p-q)^2}{p}\vee j $, we have
\begin{align}
\mathbb{E}[g_{j,q}(X)^2] \leq \left( \frac{2Mp}{n}\right)^j.
\end{align}
\end{lemma}
In order to bound the coefficients of best polynomial approximations, we need the following result by Qazi and Rahman\cite[Thm. E]{Qazi--Rahman2007} on the maximal coefficients of polynomials on a finite interval.
\begin{lemma}\label{lemma.chebyshev}
Let $p_n(x) = \sum_{\nu=0}^n a_\nu x^\nu$ be a polynomial of degree at most $n$ such that $|p_n(x)|\leq 1$ for $x\in [-1,1]$. Then, $|a_{n-2\mu}|$ is bounded above by the modulus of the corresponding coefficient of $T_n$ for $\mu = 0,1,\ldots,\lfloor n/2 \rfloor$, and $|a_{n-1-2\mu}|$ is bounded above by the modulus of the corresponding coefficient of $T_{n-1}$ for $\mu = 0,1,\ldots,\lfloor (n-1)/2 \rfloor$. Here $T_n(x)$ is the $n$-th Chebyshev polynomials of the first kind.
\end{lemma}
Moreover, it is shown in Cai and Low\cite[Lemma 2]{Cai--Low2011} that all of the coefficients of Chebyshev polynomial $T_{2m}(x), m\in \mathbb{Z}_+$ are upper bounded by $2^{3m}$. Hence, we can obtain the following result when the approximation interval is not centered at zero.
\begin{lemma}\label{lem.polycoeff}
Let $p_n(x) = \sum_{\nu=0}^n a_\nu x^\nu$ be a polynomial of degree at most $n$ such that $|p_n(x)|\leq A$ for $x\in [a,b]$, where $a+b\neq 0$. Then
\begin{align}
|a_\nu| \le 2^{7n/2}A\left|\frac{a+b}{2}\right|^{-\nu}\left(\left|\frac{b+a}{b-a}\right|^n +1 \right), \qquad \nu=0,\cdots,n.
\end{align}
\end{lemma}
The following lemma gives some tail bounds for Poisson and Binomial random variables.
\begin{lemma}\cite[Exercise 4.7]{mitzenmacher2005probability}\label{lemma.poissontail}
If $X\sim \mathsf{Poi}(\lambda)$ or $X\sim \mathsf{B}(n,\frac{\lambda}{n})$, then for any $\delta>0$, we have
\begin{align}
\mathbb{P}(X \geq (1+\delta) \lambda) & \leq \left( \frac{e^\delta}{(1+\delta)^{1+\delta}} \right)^\lambda \le e^{-\delta^2\lambda/3}\vee e^{-\delta\lambda/3} \\
\mathbb{P}(X \leq (1-\delta)\lambda) & \leq \left( \frac{e^{-\delta}}{(1-\delta)^{1-\delta}} \right)^\lambda \leq e^{-\delta^2 \lambda/2}.
\end{align}
\end{lemma}
The following lemmas deal with the upper bound of the variance in different scenarios.
\begin{lemma}\label{lem.varprod}
For independent random variables $X,Y$ with finite second moment, we have
\begin{align}
\mathsf{Var}(XY) = (\mathbb{E} Y)^2\mathsf{Var}(X) + (\mathbb{E} X)^2 \mathsf{Var}(Y) + \mathsf{Var}(X)\mathsf{Var}(Y).
\end{align}
\end{lemma}
\begin{lemma}\cite[Lemma 4]{Cai--Low2011}\label{lemma.cailow1}
Suppose $\mathbbm{1}(A)$ is an indicator random variable independent of $X$ and $Y$, then
\begin{align}
\mathsf{Var}(X \mathbbm{1}(A) + Y \mathbbm{1}(A^c))= \mathsf{Var}(X) \mathbb{P}(A) + \mathsf{Var}(Y) \mathbb{P}(A^c) + (\mathbb{E} X - \mathbb{E} Y)^2 \mathbb{P}(A) \mathbb{P}(A^c).
\end{align}
\end{lemma}
\begin{lemma}\cite[Lemma 5]{Cai--Low2011}\label{lemma.cailow2}
For any two random variables $X$ and $Y$,
\begin{equation}
\mathsf{Var}(X\wedge Y) \leq \mathsf{Var}(X) + \mathsf{Var}(Y).
\end{equation}
In particular, for any random variable $X$ and any constant $C$,
\begin{equation}
\mathsf{Var}(X\wedge C) \leq \mathsf{Var}(X).
\end{equation}
\end{lemma}
\section{Proof of Main Lemmas}
\subsection{Proof of Lemma \ref{lem.mle}}
First we give an upper bound of $\mathsf{Var}(g_n(\hat{q}))$ for $n\hat{q}\sim\mathsf{Poi}(nq)$. Note that $g_n(q)$ is continuously differentiable on $[0,1]$, we have
\begin{align}
\mathsf{Var}(g_n(\hat{q})) &\le \mathbb{E}(g_n(\hat{q})-g_n(q))^2\\
&= \mathbb{E}(g_n(\hat{q})-g_n(q))^2\mathbbm{1}(\hat{q}\ge \frac{q}{2}) + \mathbb{E}(g_n(\hat{q})-g_n(q))^2\mathbbm{1}(\hat{q}< \frac{q}{2})\\
&= \mathbb{E}[g_n'(\xi_1)]^2(\hat{q}-q)^2\mathbbm{1}(\hat{q}\ge \frac{q}{2}) + \mathbb{E}[g_n'(\xi_2)]^2(\hat{q}-q)^2\mathbbm{1}(\hat{q}< \frac{q}{2})\\
&\le \sup_{\xi_1\in [q/2,1]} |g_n'(\xi_1)|^2 \cdot \mathbb{E}(\hat{q}-q)^2 + \sup_{\xi_2\in [0,1]}|g_n'(\xi_2)|\cdot q^2\mathbb{P}(\hat{q}<\frac{q}{2})\\
&\le \frac{4}{q^2}\cdot \frac{q}{n} + n^2\cdot q^2e^{-nq/8}\\
&\le \frac{4}{nq} + \frac{n^2}{q}\cdot \left(\frac{24}{en}\right)^3 \\
&\le \frac{700}{nq}
\end{align}
where in the previous steps we have used Lemma \ref{lemma.poissontail} and the fact
\begin{align}\label{eq.poly_exp_bound}
q^ke^{-cnq} \le \left(\frac{k}{ecn}\right)^k
\end{align}
for any $q\in [0,1]$.
Now we are ready to bound the bias. By independence and the triangle inequality, we have
\begin{align}
|\mathbb{E}[\hat{p}g_n(\hat{q})]-pg_n(q)| &= p|\mathbb{E}[g_n(\hat{q})-g_n(q)]|\\
&\le u(S)q|\mathbb{E}[g_n(\hat{q})-g_n(q)]|\\
&\le u(S)\left(|\mathbb{E}[\hat{q}g_n(\hat{q})]-qg_n(q)| + |\mathbb{E}[(\hat{q}-q)g_n(\hat{q})]|\right).
\end{align}
We bound these two terms separately. For the first term, it can be obtained similar to \cite{Jiao--Venkat--Weissman2014MLE} (via the second-order Ditzian--Totik modulus of smoothness defined in \cite{Ditzian--Totik1987}) that
\begin{align}
|\mathbb{E}[\hat{q}g_n(\hat{q})]-qg_n(q)| \le \frac{5\ln 2}{n}
\end{align}
for any $q\in [0,1]$. For the second term, first note that $\mathbb{E}[\hat{q}]=q$, we have
\begin{align}
|\mathbb{E}[(\hat{q}-q)g_n(\hat{q})]| = |\mathbb{E}[(\hat{q}-q)(g_n(\hat{q}) - \mathbb{E} g_n(\hat{q}))]|.
\end{align}
Hence, by the Cauchy-Schwartz inequality and the previous bound on $\mathsf{Var}(g_n(\hat{q}))$, we have
\begin{align}
|\mathbb{E}[(\hat{q}-q)g_n(\hat{q})]|^2 &\le \mathbb{E}|\hat{q}-q|^2 \cdot \mathbb{E}|g_n(\hat{q})-\mathbb{E} g_n(\hat{q})|^2\\
&= \mathbb{E}|\hat{q}-q|^2 \cdot \mathsf{Var}(g_n(\hat{q}))\\
&\le \frac{q}{n}\cdot \frac{700}{nq} \\
&= \frac{700}{n^2}.
\end{align}
A combination of these two inequalities yields the bias bound
\begin{align}
|\mathbb{E}[\hat{p}g_n(\hat{q})]-pg_n(q)| \le u(S)\left(\frac{5\ln 2}{n} + \frac{\sqrt{700}}{n}\right) \le \frac{30u(S)}{n}
\end{align}
which together with $|\mathbb{E}[\hat{p}\ln\hat{p}]-p\ln p|\le \frac{5\ln 2}{m}$ in \cite{Jiao--Venkat--Weissman2014MLE} yields the desired bias bound.
Next we bound the variance as follows:
\begin{align}
\mathsf{Var}(\hat{p}(\ln \hat{p}-g_n(\hat{q}))) &\le \mathbb{E}[\hat{p}(\ln \hat{p}-g_n(\hat{q}))-p(\ln p-g_n(q))]^2\\
&\le 3\left(\mathbb{E}[\hat{p}(\ln\hat{p}-\ln p)]^2 + \mathbb{E}[\hat{p}(g_n(\hat{q})-g_n(q))]^2 + \mathbb{E}[(\hat{p}-p)(\ln p-g_n(q))]^2\right)\\
&\equiv 3(A_1+A_2+A_3)
\end{align}
We bound $A_1,A_2,A_3$ separately. To bound $A_1$, we further decompose $A_1$ as
\begin{align}
A_1 &= \mathbb{E}[\hat{p}(\ln\hat{p}-\ln p)]^2\mathbbm{1}(\hat{p}\le p) + \mathbb{E}[\hat{p}(\ln\hat{p}-\ln p)]^2\mathbbm{1}(\hat{p}>p)\mathbbm{1}(p\ge \frac{1}{m}) + \mathbb{E}[\hat{p}(\ln\hat{p}-\ln p)]^2\mathbbm{1}(\hat{p}>p)\mathbbm{1}(p< \frac{1}{m})\\
&\equiv B_1 + B_2 + B_3
\end{align}
where
\begin{align}
B_1 &\le \mathbb{E}\left[\sup_{\xi\ge \hat{p}} |\frac{\hat{p}}{\xi}|^2 \cdot|\hat{p}-p|^2\mathbbm{1}(\hat{p}\le p)\right]\le \mathbb{E}|\hat{p}-p|^2 = \frac{p}{m}\\
B_2 &\le \mathbb{E}\left[\sup_{\xi\ge p} |\frac{\hat{p}}{\xi}|^2 \cdot|\hat{p}-p|^2\mathbbm{1}(\hat{p}> p)\right]\mathbbm{1}(p\ge \frac{1}{m})\le \frac{\mathbb{E}[\hat{p}^2(\hat{p}-p)^2]}{p^2}\mathbbm{1}(p\ge \frac{1}{m})\\
& = \left(\frac{1}{m^3p}+\frac{5}{m^2}+\frac{p}{m}\right)\mathbbm{1}(p\ge \frac{1}{m}) \le \frac{6}{m^2} + \frac{p}{m}.
\end{align}
Upper bounding $B_3$ requires more delicate analysis. First note that by differentiation with respect to $p$, for any $k\ge 1$ we have
\begin{align}
\sup_{p\le \frac{1}{m}}\frac{(mp)^k}{k!}\cdot \frac{k^2}{m^2}(\ln \frac{k}{mp})^2 &\le \sup_{p\le \frac{1}{m}}\frac{(mp)^k}{k!}\cdot \frac{k^2}{m^2}(2+\ln \frac{k}{mp})^2 \\
&\le \frac{(2+\ln k)^2}{k!}\cdot \frac{k^2}{m^2}.
\end{align}
Hence, expanding the expectation of $B_3$ yields
\begin{align}
B_3 = \sum_{k=1}^\infty e^{-mp}\frac{(mp)^k}{k!}\cdot \frac{k^2}{m^2}(\ln \frac{k}{mp})^2\mathbbm{1}(p<\frac{1}{m})\le \frac{1}{m^2}\sum_{k=1}^\infty \frac{k^2(2+\ln k)^2}{k!} < \frac{45}{m^2}
\end{align}
where the infinite sum converges to
\begin{align}
\sum_{k=1}^\infty \frac{k^2(2+\ln k)^2}{k!} \approx 44.17 < 45.
\end{align}
Hence, $A_1$ can be upper bounded as
\begin{align}
A_1 = B_1+B_2+B_3\le \frac{51}{m^2} + \frac{2p}{m}.
\end{align}
As for $A_2$, since we have proved that $\mathbb{E}(g_n(\hat{q})-g_n(q))^2\le \frac{700}{nq}$, by independence we have
\begin{align}
A_2 \le \mathbb{E}(\hat{p}^2)\cdot \frac{700}{nq} = \left(p^2+\frac{p}{m}\right)\cdot \frac{700}{nq} \le \frac{700u(S)}{n}\left(p+\frac{1}{m}\right).
\end{align}
For $A_3$, it is clear that
\begin{align}
A_3 = \frac{p}{m}(\ln p-g_n(q))^2 \le \frac{2p}{m}\left((\ln \frac{p}{q})^2 + (g_n(q)-\ln q)^2\right)\le \frac{2}{m}\left(p(\ln u(S))^2 + \frac{4q}{e^2} + \frac{4u(S)}{en}\right)
\end{align}
where we have used the fact that for $p\le u(S)q$,
\begin{align}
p(\ln \frac{p}{q})^2 &\le p(\ln u(S))^2\vee \frac{4q}{e^2} \le p(\ln u(S))^2 + \frac{4q}{e^2}\\
p(g_n(q)-\ln q)^2 &\le p(1-\ln(nq))^2\mathbbm{1}(q<\frac{1}{n})\le u(S)q(1-\ln(nq))^2\mathbbm{1}(q<\frac{1}{n})\le \frac{4u(S)}{en}.
\end{align}
A combination of the upper bounds of $A_1,A_2,A_3$ yields
\begin{align}
\mathsf{Var}(\hat{p}(\ln\hat{p}-g_n(\hat{q}))) \le \frac{51}{m^2} + \frac{2}{m}\left(p+p(\ln u(S))^2+\frac{4q}{e^2}+\frac{4u(S)}{en}\right) + \frac{700u(S)}{n}\left(p+\frac{1}{m}\right).
\end{align}
The proof is complete. \qed
\subsection{Proof of Lemma \ref{lem.lower_mle}}
Braess and Sauer \cite[Prop. 4]{Braess--Sauer2004} showed the following equalities for the Bernstein polynomials:
\begin{align}
B_n[(x-x_0)^2](x_0) &= \frac{x_0(1-x_0)}{n}\\
B_n[(x-x_0)^3](x_0) &= \frac{x_0(1-x_0)(1-2x_0)}{n^2}.
\end{align}
Hence, choosing $x_0=x$, we have
\begin{align}
Q_3(x) - B_n[Q_3](x) = \frac{1}{x^2} \cdot \frac{x(1-x)}{n} - \frac{2}{x^3}\cdot \frac{x(1-x)(1-2x)}{n^2} = \frac{(1-x)((n+4)x-2)}{n^2x^2}.
\end{align}
Then the desired inequality is a direct result of Lemma \ref{lem.bernstein}. \qed
\subsection{Proof of Lemma \ref{lem.plugin_general}}
For the first statement, define the remainder term of the Taylor expansion as
\begin{align}
R(\hat{\theta}_n^{(1)}) \triangleq \sum_{k=0}^r \frac{G^{(k)}(\hat{\theta}_n^{(1)})}{k!}(\theta-\hat{\theta}_n^{(1)})^k - G(\theta)
\end{align}
and denote by $E$ the event that $\hat{\theta}_n^{(1)}\in V(\theta)$. By the definition of reverse confidence sets, $\mathbb{P}_\theta(E^c)\le \delta$, and
\begin{align}
|\mathbb{E}_\theta H(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)}) - G(\theta)| &\le |\mathbb{E}_\theta (H(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)}) - G(\theta))\mathbbm{1}(E)| + |\mathbb{E}_\theta (H(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)}) - G(\theta))\mathbbm{1}(E^c)| \\
&\le |\mathbb{E}_\theta (\hat{G}^{(r)}(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)}) - G(\theta))\mathbbm{1}(E)| + \delta\cdot \left(|G(\theta)|+\sup_{\theta_1,\theta_2\in \hat{\Theta}} |H(\theta_1,\theta_2)|\right)\\
&= |\mathbb{E}_\theta R(\hat{\theta}_n^{(1)})\mathbbm{1}(E)| + \delta\cdot \left(|G(\theta)|+\sup_{\theta_1,\theta_2\in \hat{\Theta}} |H(\theta_1,\theta_2)|\right)\\
&\le \frac{\mathbb{E}_\theta[|\hat{\theta}_n^{(1)}-\theta|^{r+1}\mathbbm{1}(E)]}{(r+1)!}\cdot \sup_{\hat{\theta}\in V(\theta)}|G^{(r+1)}(\hat{\theta})| + \delta\cdot \left(|G(\theta)|+\sup_{\theta_1,\theta_2\in \hat{\Theta}} |H(\theta_1,\theta_2)|\right)\\
&\le \frac{\mathbb{E}_\theta|\hat{\theta}_n^{(1)}-\theta|^{r+1}}{(r+1)!}\cdot \sup_{\hat{\theta}\in V(\theta)}|G^{(r+1)}(\hat{\theta})| + \delta\cdot \left(|G(\theta)|+\sup_{\theta_1,\theta_2\in \hat{\Theta}} |H(\theta_1,\theta_2)|\right).
\end{align}
As for the variance of $H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})$ with $k\ge 0$, we first note by triangle inequality that
\begin{align}
\mathsf{Var}_\theta(H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})) &\le 2\mathsf{Var}_\theta(H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\mathbbm{1}(E)) + 2\mathsf{Var}_\theta(H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\mathbbm{1}(E^c))\\
&\le 2\mathsf{Var}_\theta(G_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\mathbbm{1}(E)) + 2\delta\cdot \sup_{\theta_1,\theta_2\in \hat{\Theta}} |H_k(\theta_1,\theta_2)|^2.
\end{align}
Hence, it suffices to upper bound $\mathsf{Var}_\theta(G_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\mathbbm{1}(E))$. Note that $G_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})$ is a linear combination of terms of the form $G^{(k)}(\hat{\theta}_n^{(1)})(\hat{\theta}_n^{(1)})^{k-j}S_j(\hat{\theta}_n^{(2)})$ with $0\le j\le k$, we employ the triangle inequality again to reduce the problem of bounding the total variance to bounding the variance of individual terms. By independence and Lemma \ref{lem.varprod}, it further suffices to upper bound $|\mathbb{E}_\theta G^{(k)}(\hat{\theta}_n^{(1)})(\hat{\theta}_n^{(1)})^{k-j}\mathbbm{1}(E)|$ and $\mathsf{Var}_\theta(G^{(k)}(\hat{\theta}_n^{(1)})(\hat{\theta}_n^{(1)})^{k-j}\mathbbm{1}(E))$, respectively. In fact, defining $G_{k,j}(\theta)=\theta^jG^{(k)}(\theta)$, by Taylor expansion again we have
\begin{align}
|\mathbb{E}_\theta G_{k,j}(\hat{\theta}_n^{(1)})\mathbbm{1}(E)| &\le |\mathbb{E}_\theta (G_{k,j}(\hat{\theta}_n^{(1)})-G_{k,j}(\theta))\mathbbm{1}(E)| + |G_{k,j}(\theta)| \\
&= |\mathbb{E}_\theta (G_{k,j}'(\theta)(\hat{\theta}_n^{(1)}-\theta)+\frac{1}{2}G_{k,j}''(\xi)(\hat{\theta}_n^{(1)}-\theta)^2)\mathbbm{1}(E)| + |G_{k,j}(\theta)|\\
&\le \frac{1}{2}|\mathbb{E}_\theta G_{k,j}''(\xi)(\hat{\theta}_n^{(1)}-\theta)^2\mathbbm{1}(E)| + |G_{k,j}(\theta)| + \mathbb{P}(E^c)\cdot \left(|G_{k,j}'(\theta)|(|\theta|+\sup_{\hat{\theta}\in\hat{\Theta}}|\hat{\theta}|)\right)\\
&\le \frac{\mathbb{E}_\theta(\hat{\theta}_n^{(1)}-\theta)^2}{2}\cdot \sup_{\hat{\theta}\in V(\theta)}|G_{k,j}''(\hat{\theta})| + |G_{k,j}(\theta)| + \delta\cdot \left(|G_{k,j}'(\theta)|(|\theta|+\sup_{\hat{\theta}\in\hat{\Theta}}|\hat{\theta}|)\right)
\end{align}
and
\begin{align}
\mathsf{Var}_\theta(G_{k,j}(\hat{\theta}_n^{(1)})\mathbbm{1}(E)) &\le \mathbb{E}_\theta(G_{k,j}(\hat{\theta}_n^{(1)})\mathbbm{1}(E)-G_{k,j}(\theta))^2\\
&\le 2\mathbb{E}_\theta((G_{k,j}(\hat{\theta}_n^{(1)})-G_{k,j}(\theta))\mathbbm{1}(E))^2 + 2|G_{k,j}(\theta)|^2\cdot \mathbb{P}(E^c)\\
&\le 2\mathbb{E}_\theta(G_{k,j}'(\xi)(\hat{\theta}_n^{(1)}-\theta)\mathbbm{1}(E))^2 + 2\delta\cdot |G_{k,j}(\theta)|^2\\
&\le 2\mathbb{E}_\theta(\hat{\theta}_n^{(1)}-\theta)^2\cdot \sup_{\hat{\theta}\in V(\theta)}|G_{k,j}'(\hat{\theta})|^2 + 2\delta\cdot |G_{k,j}(\theta)|^2
\end{align}
which establishes the desired variance bound.
Finally it remains to bound the quantity $|\mathbb{E}_\theta H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})|$ for $k\ge 1$. If $k\ge 2$, as above, by triangle inequality, we conclude that
\begin{align}
|\mathbb{E}_\theta H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})| &\le |\mathbb{E}_\theta H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\mathbbm{1}(E)| + |\mathbb{E}_\theta H_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\mathbbm{1}(E^c)| \\
&\le |\mathbb{E}_\theta G_k(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})\mathbbm{1}(E)| + \delta\cdot \sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_k(\theta_1,\theta_2)| \\
&= \frac{1}{k!}|\mathbb{E}_\theta G^{(k)}(\hat{\theta}_n^{(1)})(\theta-\hat{\theta}_n^{(1)})^k\mathbbm{1}(E)| + \delta\cdot \sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_k(\theta_1,\theta_2)|\\
&\le \frac{\mathbb{E}_\theta|\hat{\theta}_n^{(1)}-\theta|^k}{k!}\cdot \sup_{\hat{\theta}\in V(\theta)}|G^{(k)}(\hat{\theta})| + \delta\cdot \sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_k(\theta_1,\theta_2)|.
\end{align}
For $k=1$, we further note that $\mathbb{E}_\theta\hat{\theta}_n^{(1)}=\theta$, and conduct order-one Taylor expansion to yield
\begin{align}
|\mathbb{E}_\theta H_1(\hat{\theta}_n^{(1)},\hat{\theta}_n^{(2)})| &\le |\mathbb{E}_\theta G'(\hat{\theta}_n^{(1)})(\theta-\hat{\theta}_n^{(1)})\mathbbm{1}(E)| + \delta\cdot \sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_1(\theta_1,\theta_2)|\\
&= |\mathbb{E}_\theta (G'(\theta)+(\hat{\theta}_n^{(1)}-\theta)G''(\xi))(\theta-\hat{\theta}_n^{(1)})\mathbbm{1}(E)| + \delta\cdot \sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_1(\theta_1,\theta_2)|\\
&\le \mathbb{E}_\theta (\hat{\theta}_n^{(1)}-\theta)^2\cdot \sup_{\hat{\theta}\in V(\theta)}|G''(\hat{\theta})| + \delta\cdot \left(\sup_{\theta_1,\theta_2\in\hat{\Theta}}|H_1(\theta_1,\theta_2)|+|G'(\theta)|(|\theta|+\sup_{\hat{\theta}\in\hat{\Theta}}|\hat{\theta}|)\right)
\end{align}
as desired. \qed
\subsection{Proof of Lemma \ref{lem.U_H}}
Replacing $n$ by $m$, we adopt the notations of $V(p)$ and $\delta$ in \eqref{eq.localize_level}. Denote by $E$ the event $\hat{p}\in V(p)$, we have
\begin{align}
|\mathbb{E} U_H(\hat{p}) + p\ln p| &= \left|\mathbb{E}\left(-\hat{p}\ln\hat{p}+\frac{1}{2m}\right)+p\ln p\right|\\
&\le \left| \mathbb{E}\left(-\hat{p}\ln\hat{p}-(1+\ln p)(\hat{p}-p)+\frac{(\hat{p}-p)^2}{2p}-\frac{(\hat{p}-p)^3}{6p^2}\right)+p\ln p \right| + \frac{|\mathbb{E}(\hat{p}-p)^3|}{6p^2}\\
&\le \left| \mathbb{E}\left(-\hat{p}\ln\hat{p}-(1+\ln p)(\hat{p}-p)+\frac{(\hat{p}-p)^2}{2p}-\frac{(\hat{p}-p)^3}{6p^2}+p\ln p \right)\mathbbm{1}(E)\right| + \nonumber\\
&\qquad \mathbb{P}(E^c)\cdot\sup_{\hat{p}\in [0,1]}\left(-\hat{p}\ln\hat{p}-(1+\ln p)(\hat{p}-p)+\frac{(\hat{p}-p)^2}{2p}-\frac{(\hat{p}-p)^3}{6p^2}+p\ln p\right)+ \frac{|\mathbb{E}(\hat{p}-p)^3|}{6p^2}\\
&\le \frac{\mathbb{E}|\hat{p}-p|^4}{24}\sup_{x\in V(p)}|(-x\ln x)^{(4)}| + 2m^{-c_1/24}\cdot \left(\frac{1}{e}-p\ln p+1-\ln p+\frac{1}{2p}+\frac{1}{6p^2}\right) + \frac{1}{6pm^2}\\
&\le \frac{p+3mp^2}{24m^3}\sup_{x\in V(p)}\frac{2}{x^3} + 4m^{-c_1/24+2}+ \frac{1}{6pm^2}
\end{align}
where we have used $mp\ge c_1\ln m/2\ge 2$. Since any $x\in V(p)$ satisfies
\begin{align}
x\ge p-\frac{1}{2}\sqrt{\frac{c_1p\ln m}{2m}}\ge \frac{p}{2}
\end{align}
by the previous inequality we have
\begin{align}
|\mathbb{E} U_H(\hat{p}) + p\ln p|&\le \frac{p+3mp^2}{24m^3}\cdot\frac{2}{(p/2)^3} + 4m^{-c_1/24+2}+ \frac{1}{6pm^2}\\
&\le \frac{4mp^2}{24m^3}\cdot\frac{2}{(p/2)^3} + 4m^{-c_1/24+2}+ \frac{1}{6pm^2}\\
&\le \frac{3}{m^2p} + 4m^{-c_1/24+2}\\
&\le \frac{6}{c_1m\ln m} + 4m^{-c_1/24+2}
\end{align}
as desired.
As for the variance, since the constant bias correcting term does not affect variance, applying Lemma \ref{lem.plugin_general} with $k=0$ yields
\begin{align}
\mathsf{Var}(U_H(\hat{p})) &\le A_0\left(2m^{-c_1/24}\cdot 1+\frac{p}{m}(1-\ln (p/2))^2 + 2m^{-c_1/24}\cdot (p\ln p)^2\right)\\
&\le A_0\left(\frac{p(2-\ln p)^2}{m} + 4m^{-c_1/24}\right).
\end{align}
\qed
\subsection{Proof of Lemma \ref{lem.plugin_upper}}
The only non-trivial part in deducing the third inequality from Lemma \ref{lem.plugin_general} is to prove that for any $k\in \mathbb{N}$ and $q\ge \frac{1}{n}$,
\begin{align}
\mathbb{E}_q |\hat{q}_1-q|^k \lesssim \left(\frac{q}{n}\right)^{\frac{k}{2}}.
\end{align}
In fact, since $\mathbb{E}_q \exp(s\hat{q}_1)=\exp(nq(e^{s/n}-1))$, we have $\mathbb{E}_q \exp(s(\hat{q}_1-q))=\exp(nq(e^{s/n}-1-s/n))$. Hence, by comparing the coefficient of $s^k$ at both sides of
\begin{align}
\sum_{k=0}^\infty \frac{\mathbb{E}_q (\hat{q}_1-q)^k}{k!}s^k = \sum_{k=0}^\infty \frac{(nq)^k}{k!}\left(\sum_{j=2}^\infty \frac{1}{j!}(\frac{s}{n})^j\right)^k
\end{align}
yields that for even $k$, $\mathbb{E}_q(\hat{q}_1-q)^k$ can be expressed as
\begin{align}
\mathbb{E}_q(\hat{q}_1-q)^k = \sum_{j=0}^{\frac{k}{2}} a_{k,j}\frac{q^j}{n^{k-j}}
\end{align}
for some coefficients $\{a_{k,j}\}_{j=0}^{k/2}$. Now for even $k$, the desired inequality follows from the assumption $q\ge n^{-1}$. For odd $k$, Cauchy-Schwartz inequality yields
\begin{align}
\mathbb{E}_q |\hat{q}_1-q|^k \le \left(\mathbb{E}_q |\hat{q}_1-q|^{k-1}\right)^{\frac{1}{2}}\left(\mathbb{E}_q |\hat{q}_1-q|^{k+1}\right)^{\frac{1}{2}}\lesssim \left(\frac{q}{n}\right)^{\frac{k-1}{2}\cdot \frac{1}{2}+\frac{k+1}{2}\cdot \frac{1}{2}} = \left(\frac{q}{n}\right)^{\frac{k}{2}}.
\end{align}
Hence, the third inequality follows. The first inequality also follows from this fact, Lemma \ref{lem.plugin_general}, $\delta\le 2n^{-c_1/24}$ and $V(q)\subset [\frac{q}{2},2q]$.
Now it remains to deduce the second inequality from Lemma \ref{lem.plugin_general}. By \eqref{eq.monomial_unbiased}, we know that $S_{j}(\hat{q}_2)$ is a linear combination of $\frac{\hat{q}_2^{j-i}}{n^{i}}$ with constant coefficients and $i=0,1,\cdots,j$. By the triangle inequality for the variance, it suffices to upper bound the variance of each individual term $\frac{\hat{q}_2^{j-i}}{n^{i}}$. Using the same approach based on moment generating function, we conclude that for any $k\ge 0$ and $q\ge n^{-1}$,
\begin{align}
\mathsf{Var}_q(\hat{q}_2^k) = \mathbb{E}_q(\hat{q}_2^{2k}) - (\mathbb{E}_q(\hat{q}_2^k))^2 \lesssim \sum_{j=0}^{2k-1} \frac{q^j}{n^{2k-j}}\lesssim \frac{q^{2k-1}}{n}.
\end{align}
As a result, for $0\le i\le j$,
\begin{align}
\mathsf{Var}_q(\frac{\hat{q}_2^{j-i}}{n^{i}}) \lesssim \frac{q^{2j-2i-1}}{n^{2i+1}}\le \frac{q^{2j-1}}{n}
\end{align}
and thus $\mathsf{Var}_q(S_j(\hat{q}_2))\lesssim \frac{q^{2j-1}}{n}$. Finally it suffices to note that for $q\ge \frac{c_1\ln n}{2n}\ge n^{-1}$,
\begin{align}
\mathbb{E}_q[S_j^2(\hat{q}_2)] = (\mathbb{E}_q S_j(\hat{q}_2))^2 + \mathsf{Var}_q(S_j(\hat{q}_2)) \lesssim q^{2j} + \frac{q^{2j-1}}{n} \lesssim q^{2j}
\end{align}
which completes the proof of the second inequality. \qed
\subsection{Proof of Lemma \ref{lem.T_s_I}}
To invoke Lemma \ref{lem.plugin_upper}, we remark that $g(q)=\ln q$, $r=3$, and
\begin{align}
h_0(\hat{q}_1,\hat{q}_2) &= \ln\hat{q}_1 \cdot \mathbbm{1}(\hat{q}_1\neq 0) \\
h_1(\hat{q}_1,\hat{q}_2) &= \frac{\hat{q}_2-\hat{q}_1}{\hat{q}_1}\cdot \mathbbm{1}(\hat{q}_1\neq 0) \\
h_2(\hat{q}_1,\hat{q}_2) &= \left(-\frac{(\hat{q}_2-\hat{q}_1)^2}{2\hat{q}_1^2}+\frac{\hat{q}_2}{2n\hat{q}_1^2}\right)\cdot \mathbbm{1}(\hat{q}_1\neq 0)\\
h_3(\hat{q}_1,\hat{q}_2) &= \left(\frac{(\hat{q}_2-\hat{q}_1)^3}{3\hat{q}_1^3}+\frac{\hat{q}_2}{n\hat{q}_1^2}-\frac{\hat{q}_2^2}{n\hat{q}_1^3}+\frac{2\hat{q}_2}{n^2\hat{q}_1^3}\right)\cdot \mathbbm{1}(\hat{q}_1\neq 0).
\end{align}
Noting that $\sup_{\xi\in [q/2,2q]}|g^{(k)}(\xi)|=|g^{(k)}(q/2)|$ for any $k\ge 1$ and $q\ge \frac{c_1\ln n}{2n}$, we have
\begin{align}
|\mathbb{E} \hat{T}^{(3)}(\hat{q}_1,\hat{q}_2)\cdot\mathbbm{1}(\hat{q}_1\neq 0)-\ln q| &\le B_3\left(\frac{q^2}{n^2}\cdot 6(\frac{q}{2})^{-4} + n^{-c_1/24}\cdot\left(-\ln q+\ln n+n+n^2+2n^3\right)\right)\\
&\le B_3\left(\frac{96}{n^2q^2}+6n^{-c_1/24+3}\right)\\
&\le B_3\left(\frac{192}{c_1nq\ln n}+6n^{-c_1/24+3}\right)
\end{align}
and thus by independence,
\begin{align}
|\mathbb{E}[\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)]-p\ln q| &= p|\mathbb{E} \hat{T}^{(3)}(\hat{q}_1,\hat{q}_2)\cdot\mathbbm{1}(\hat{q}_1\neq 0) - \ln q| \\
&\le pB_3\left(\frac{192}{c_1nq\ln n}+6n^{-c_1/24+3}\right) \\
&\le \frac{192B_3u(S)}{c_1n\ln n} + 6B_3pn^{-c_1/24+3}.
\end{align}
Moreover, for any $0\le k\le 3$,
\begin{align}
\mathsf{Var}(h_k(\hat{q}_1,\hat{q}_2)) &\le B_3\left(n^{-c_1/24}\cdot n^3+\sum_{j=0}^k q^{2(k-j)}\left(\frac{q}{n}\cdot (\frac{24}{q^{k-j+1}})^2 + n^{-c_1/24}\cdot q^{2(j-k)}(-\ln q)\right)\nonumber\right.\\
&\left. \qquad\qquad + \sum_{j=0}^{k-1} \frac{q^{2(k-j)-1}}{n}\left(q^{j-k}+\frac{q}{n}\cdot 20(\frac{q}{2})^{j-k-2}+4n^{-c_1/24}q^{j-k-1}\right)^2\right)\\
&\le B_3\left(\frac{576(k+1)}{nq}+\frac{2k}{nq}\left(1+\frac{80}{nq}\right)^2+n^{-c_1/24}\left(n^3+(k+1)\ln n+32kn^2\right)\right)\\
&\le B_3\left(\frac{12500}{nq}+100n^{-c_1/24+3}\right)
\end{align}
where in the last step we have used the fact that $nq\ge 2$. For $1\le k\le 3$, we also have
\begin{align}
|\mathbb{E} h_k(\hat{q}_1,\hat{q}_2)| &\le B_3\left((\frac{q}{n})^{\frac{k\vee 2}{2}}\cdot 6(\frac{q}{2})^{-(k\vee 2)}+n^{-c_1/24}\cdot(q^{-1}+2n^3)\right)\\
&\le B_3\left(48+3n^{-c_1/24+3}\right).
\end{align}
Now we are about to bound the bias and the variance for small $p$ and large $p$, respectively. If $p\le \frac{2c_1\ln m}{m}$, first note that $L_H(x)=S_{K,H}(x)\wedge 1$ with $S_{K,H}(x)$ defined in \cite{Jiao--Venkat--Han--Weissman2015minimax}. It was shown in \cite[Lemma 4]{Jiao--Venkat--Han--Weissman2015minimax} that
\begin{align}\label{eq.bias_S_KH}
|\mathbb{E} S_{K,H}(\hat{p}_1)+p\ln p|&\le \frac{C}{m\ln m}\\\label{eq.var_S_KH}
\mathbb{E} S_{K,H}^2(\hat{p}_1) &\le m^{8c_2\ln 2}\frac{(2c_1\ln m)^4}{m^2}
\end{align}
where we note that the constant $c_1$ in \cite{Jiao--Venkat--Han--Weissman2015minimax} corresponds to the constant $c_1/2$ in our paper. Then applying Lemma \ref{lemma.cailow2}, we have
\begin{align}
\mathsf{Var} (L_H(\hat{p}_1))\le \mathsf{Var}(S_{K,H}(\hat{p}_1)) \le \mathbb{E} S_{K,H}^2(\hat{p}_1) \le \frac{(2c_1\ln m)^4}{m^{2-8c_2\ln 2}}
\end{align}
and thus
\begin{align}
|\mathbb{E} L_H(\hat{p}_1) + p\ln p| &\le |\mathbb{E} S_{K,H}(\hat{p}_1) + p\ln p| + \mathbb{E}|S_{K,H}(\hat{p}_1)|\mathbbm{1}(S_{K,H}(\hat{p}_1)\ge 1)\\
&\le |\mathbb{E} S_{K,H}(\hat{p}_1) + p\ln p| + \mathbb{E}|S_{K,H}(\hat{p}_1)|^2\\
&\le \frac{C}{m\ln m} + \frac{(2c_1\ln m)^4}{m^{2-8c_2\ln 2}}.
\end{align}
Hence, the total bias can be upper bounded as
\begin{align}
|\mathbb{E} [\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)+L_H(\hat{p}_1)] + p\ln (p/q)|&\le |\mathbb{E}[\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)]-p\ln q| + |\mathbb{E} L_H(\hat{p}_1) + p\ln p|\\
&\le \frac{192B_3u(S)}{c_1n\ln n} + 6B_3pn^{-c_1/24+3} + \frac{C}{m\ln m} + \frac{(2c_1\ln m)^4}{m^{2-8c_2\ln 2}}.
\end{align}
As for the total variance, Lemma \ref{lem.varprod} can be used here to obtain
\begin{align}
\mathsf{Var}(\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2))&= \mathbb{E}[\hat{p}_1^2]\cdot\mathsf{Var}(\sum_{k=0}^3 h_k(\hat{q}_1,\hat{q}_2)) + \mathsf{Var}(\hat{p}_1)\cdot (\mathbb{E} \hat{T}^{(3)}(\hat{q}_1,\hat{q}_2)\cdot \mathbbm{1}(\hat{q}_1\neq 0))^2\\
&\le \left(p^2+\frac{p}{m}\right)\cdot 16B_3\left(\frac{12500}{nq}+100n^{-c_1/24+3}\right) + \frac{p}{m}\left(B_3\left(\frac{192}{c_1nq\ln n}+6n^{-c_1/24+3}\right)-\ln q\right)^2\\
&\le 1600B_3\left(p^2+\frac{p}{m}\right)\left(\frac{125}{nq}+n^{-c_1/24+3}\right) + \frac{p}{m}\left(B_3\left(48+6n^{-c_1/24+3}\right)-\ln q\right)^2.
\end{align}
Now the desired variance bound follows from the triangle inequality $\mathsf{Var}(X+Y)\le 2(\mathsf{Var}(X)+\mathsf{Var}(Y))$.
If $p\ge \frac{c_1\ln m}{2m}$, by Lemma \ref{lem.U_H} and the triangle inequality we have
\begin{align}
|\mathbb{E} [\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)+U_H(\hat{p}_1)] + p\ln (p/q)|&\le |\mathbb{E}[\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)]-p\ln q| + |\mathbb{E} U_H(\hat{p}_1) + p\ln p|\\
&\le \frac{192B_3u(S)}{c_1n\ln n} + 6B_3pn^{-c_1/24+3} + \frac{6}{c_1m\ln m} + 4m^{-c_1/24+2}
\end{align}
which is the desired bias bound. As for the variance, we have
\begin{align}
\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + U_H(\hat{p}_{1}) = -\hat{p}_1\ln\frac{\hat{p}_1}{\hat{q}_1}\cdot \mathbbm{1}(\hat{q}_1\neq 0) + \hat{p}_1\cdot\sum_{k=1}^3 h_k(\hat{q}_1,\hat{q}_2)+ \frac{1}{2m}
\end{align}
and the triangle inequality gives
\begin{align}
\mathsf{Var}(\tilde{T}_\textrm{\rm s}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + U_H(\hat{p}_{1})) &\le 2\mathsf{Var}\left(\hat{p}_1\ln\frac{\hat{p}_1}{\hat{q}_1}\cdot \mathbbm{1}(\hat{q}_1\neq 0)\right) + 2\mathsf{Var}\left( \hat{p}_1\cdot\sum_{k=1}^3 h_k(\hat{q}_1,\hat{q}_2) \right)\\
&\equiv 2(B_1+B_2).
\end{align}
Now we bound these two terms separately. For $B_1$, recall that Lemma \ref{lem.mle} gives
\begin{align}
\mathsf{Var}(\hat{p}_1(\ln\hat{p}_1-g_n(\hat{q}_1))) \le \frac{51}{m^2} + \frac{2}{m}\left(p+p(\ln u(S))^2+\frac{4q}{e^2}+\frac{4u(S)}{en}\right) + \frac{700u(S)}{n}\left(p+\frac{1}{m}\right)
\end{align}
and the difference between these two quantities is upper bounded by
\begin{align}
\mathbb{E}[\hat{p}_1(g_n(\hat{q}_1)-\ln \hat{q}_1\cdot \mathbbm{1}(\hat{q}_1\neq 0))]^2 &\le (1+\ln n)^2\cdot \mathbb{P}(\hat{q}_1=0)\\
&\le (1+\ln n)^2\cdot n^{-c_1/2}.
\end{align}
Hence, by triangle inequality again, we have
\begin{align}
B_1 &\le 2\mathsf{Var}(\hat{p}_1(\ln\hat{p}_1-g_n(\hat{q}_1))) + 2\mathbb{E}[\hat{p}_1(g_n(\hat{q}_1)-\ln \hat{q}_1\cdot \mathbbm{1}(\hat{q}_1\neq 0))]^2\\
&\le \frac{102}{m^2} + \frac{4}{m}\left(p+p(\ln u(S))^2+\frac{4q}{e^2}+\frac{4u(S)}{en}\right) + \frac{1400u(S)}{n}\left(p+\frac{1}{m}\right) + 2(1+\ln n)^2\cdot n^{-c_1/2}.
\end{align}
As for $B_2$, Lemma \ref{lem.varprod} is employed to obtain
\begin{align}
B_2&= \mathbb{E}[\hat{p}_1^2]\cdot\mathsf{Var}(\sum_{k=1}^3 h_k(\hat{q}_1,\hat{q}_2)) + \mathsf{Var}(\hat{p}_1)\cdot (\mathbb{E} \sum_{k=1}^3 h_k(\hat{q}_1,\hat{q}_2))^2\\
&\le \left(p^2+\frac{p}{m}\right)\cdot 9B_3\left(\frac{12500}{nq}+100n^{-c_1/24+3}\right) + \frac{p}{m}\left(3B_3\left(48+3n^{-c_1/24+3}\right)\right)^2\\
&= 900B_3\left(p^2+\frac{p}{m}\right)\left(\frac{125}{nq}+n^{-c_1/24+3}\right) + \frac{9B_3^2p}{m}\left(48+3n^{-c_1/24+3}\right)^2.
\end{align}
The desired variance bound then follows from the upper bounds of $B_1$ and $B_2$.
For the rest of the results, the only non-trivial observation is that when $p\le \frac{2c_1\ln m}{m}$, we have
\begin{align}
p(\ln q)^2 &\le p(-\ln p +\ln u(S))^2\\
& \le 2p(\ln p)^2 + 2p(\ln u(S))^2\\
&\le \frac{4c_1\ln m}{m}\left(\ln \frac{m}{2c_1\ln m}\right)^2 + 2p(\ln u(S))^2\\
&\le \frac{4c_1(\ln m)^3}{m} + 2p(\ln u(S))^2
\end{align}
since $p\le u(S)q$ and $2c_1\ln m\ge 8$.\qed
\subsection{Proof of Lemma \ref{lem.T_s}}
For simplicity, we define
\begin{align}
\overline{T}_{\textrm{\rm s},\textrm{\rm I}}(\hat{p},\hat{q}) &\triangleq \tilde{T}_{\textrm{\rm s}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + U_H(\hat{p}_1)\\
\overline{T}_{\textrm{\rm s},\textrm{\rm II}}(\hat{p},\hat{q}) &\triangleq \tilde{T}_{\textrm{\rm s}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + L_H(\hat{p}_1)
\end{align}
then
\begin{align}
\overline{T}_\textrm{\rm s}(\hat{p},\hat{q}) = \overline{T}_{\textrm{\rm s},\textrm{\rm I}}(\hat{p},\hat{q})\mathbbm{1}(\hat{p}_3> \frac{c_1\ln m}{m}) + \overline{T}_{\textrm{\rm s},\textrm{\rm II}}(\hat{p},\hat{q})\mathbbm{1}(\hat{p}_3\le \frac{c_1\ln m}{m}).
\end{align}
By Lemma \ref{lem.T_s_I}, the bias can be upper bounded as
\begin{align}
|\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})+\ln(p/q)| &\le |\mathbb{E} \overline{T}_{\textrm{\rm s},\textrm{\rm I}}(\hat{p},\hat{q}) + p\ln (p/q)| + |\mathbb{E} \overline{T}_{\textrm{\rm s},\textrm{\rm II}}(\hat{p},\hat{q}) + p\ln(p/q)|\\
&\lesssim \frac{1}{m\ln m} + \frac{u(S)}{n\ln n}.
\end{align}
As for the variance, by Lemma \ref{lemma.cailow1} we have
\begin{align}
\mathsf{Var}(\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})) &\le \mathsf{Var}(\overline{T}_{\textrm{\rm s},\textrm{\rm I}}(\hat{p},\hat{q})) + \mathsf{Var}(\overline{T}_{\textrm{\rm s},\textrm{\rm II}}(\hat{p},\hat{q}) ) + (\mathbb{E}\overline{T}_{\textrm{\rm s},\textrm{\rm I}}(\hat{p},\hat{q})-\mathbb{E} \overline{T}_{\textrm{\rm s},\textrm{\rm II}}(\hat{p},\hat{q}) )^2\\
&\le \mathsf{Var}(\overline{T}_{\textrm{\rm s},\textrm{\rm I}}(\hat{p},\hat{q})) + \mathsf{Var}(\overline{T}_{\textrm{\rm s},\textrm{\rm II}}(\hat{p},\hat{q}) ) + 2 |\mathbb{E}\overline{T}_{\textrm{\rm s},\textrm{\rm I}}(\hat{p},\hat{q})+p\ln(p/q)|^2 + 2|\mathbb{E}\overline{T}_{\textrm{\rm s},\textrm{\rm II}}(\hat{p},\hat{q})+p\ln(p/q)|^2\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m} + \frac{u(S)}{mn} + \frac{pu(S)}{n} + \left(\frac{1}{m\ln m}+\frac{u(S)}{n\ln n}\right)^2\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m} + \frac{u(S)}{mn} + \frac{pu(S)}{n} + \left(\frac{1}{m\ln m}\right)^2+\left(\frac{u(S)}{n\ln n}\right)^2\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m} + \frac{u(S)}{mn} + \frac{pu(S)}{n} + \left(\frac{u(S)}{n\ln n}\right)^2
\end{align}
as desired. \qed
\subsection{Proof of Lemma \ref{lem.T_ns_1}}
Denote by $E$ the event that $\frac{2p}{3} \le \hat{p}_2\le p+\frac{1}{2}\sqrt{\frac{c_1\hat{p}_2\ln m}{m}}$, then by Lemma \ref{lemma.poissontail} we have
\begin{align}
\mathbb{P}(E^c) &\le \mathbb{P}(\hat{p}_2 < \frac{2p}{3}) + \mathbb{P}(\hat{p}_2 > p+\frac{1}{2}\sqrt{\frac{c_1\hat{p}_2\ln m}{m}})\\
&\le \mathbb{P}(\hat{p}_2 <p - \sqrt{\frac{c_1p\ln m}{18m}}) + \mathbb{P}(\hat{p}_2 > p+\frac{1}{2}\sqrt{\frac{c_1p\ln m}{m}})\\
&\le \exp\left( -\frac{1}{2}\left(\sqrt{\frac{c_1\ln m}{18mp}} \right)^2\cdot mp \right) + \exp\left( -\frac{1}{3}\left(\sqrt{\frac{c_1\ln m}{4mp}} \right)^2\cdot mp \right)\\
&\le 2e^{-c_1m/36}.
\end{align}
Note that conditioning on the event $E$, we have $\hat{p}_2\ge \frac{c_1\ln m}{3m}$, and
\begin{align}\label{eq.inequality_chain}
0 < \frac{2p}{21u(S)}\le \frac{1}{u(S)}\left( \hat{p}_2 - \frac{\sqrt{3}}{2}\hat{p}_2 \right) \le \frac{1}{u(S)}\left( \hat{p}_2 - \frac{1}{2}\sqrt{\frac{c_1\hat{p}_2\ln m}{m}} \right) \le \frac{p}{u(S)}\le q \le \frac{2c_1\ln n}{n}
\end{align}
i.e., the approximation region contains $q$.
We first analyze the variance:
\begin{align}
\mathsf{Var}(\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)) &\le \mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2]\\
&= \mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2\mathbbm{1}(E)] + \mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2\mathbbm{1}(E^c)] \\\label{eq.var_decomp}
&\le \mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2\mathbbm{1}(E)] + 1\cdot \mathbb{P}(E^c).
\end{align}
Note that conditioning on $E$, we have
\begin{align}
T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) = \hat{p}_1\cdot \sum_{k=0}^K g_{K,k}(\hat{p}_2)\prod_{l=0}^{k-1}(\hat{q}_2-\frac{l}{n}).
\end{align}
By construction, $\sum_{k=0}^K g_{K,k}(\hat{p}_2)x^k$ is the best polynomial approximation of $\ln q$ on $R$, where
\begin{align}
\left[\frac{2p}{21u(S)},\frac{4c_1\ln n}{n}\right]\supset R = \left[ \frac{1}{u(S)}\left(\hat{p}_2-\frac{1}{2}\sqrt{\frac{c_1\hat{p}_2\ln m}{m}} \right), \frac{4c_1\ln n}{n}\right] \supset \left[\frac{2c_1\ln n}{n},\frac{4c_1\ln n}{n}\right]
\end{align}
where we have used (\ref{eq.inequality_chain}) here. Since $W(\cdot)$ in Lemma \ref{lem.approx_log} is an increasing function, conditioning on $E$ the approximation error can be upper bounded as
\begin{align}\label{eq.approx_err}
\sup_{x\in R} \left|\sum_{k=0}^Kg_{K,k}(\hat{p}_2)x^k - \ln x\right| \le C_{\ln }W\left(\frac{4c_1\ln n/n}{(c_2\ln n)^2\cdot (2p/21u(S))}\right) = C_{\ln }W\left(\frac{42c_1}{c_2^2}\cdot \frac{u(S)}{pn\ln n}\right).
\end{align}
Note that $W(x)\le 1\vee \ln x$, we conclude that for any $x\in [\frac{2c_1\ln n}{n},\frac{4c_1\ln n}{n}]\subset R$, we have
\begin{align}
\left|\sum_{k=0}^Kg_{K,k}(\hat{p}_2)x^k\right| &\le \left|\sum_{k=0}^Kg_{K,k}(\hat{p}_2)x^k - \ln x\right| + \ln \frac{n}{2c_1\ln n} \\
&\le C_{\ln }\left(\ln\left(\frac{42c_1}{c_2^2}\cdot \frac{u(S)}{pn\ln n}\right)\vee 1\right)+ \ln n \equiv A.
\end{align}
Now we are about to apply Lemma \ref{lem.polycoeff} to bound each coefficient $|g_{K,k}(\hat{p}_2)|$. Lemma \ref{lem.polycoeff} yields
\begin{align}
|g_{K,k}(\hat{p}_2)| \le 2^{7K/2}A\cdot \left(\frac{3c_1\ln n}{n}\right)^{-k}(3^K+1) \le 2^{11K/2}A\cdot \left(\frac{3c_1\ln n}{n}\right)^{-k}
\end{align}
for any $k=0,1,\cdots,K=c_2\ln n$ conditioning on $E$. Hence, by the triangle inequality,
\begin{align}
\mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2\mathbbm{1}(E)] &= \mathbb{E}(\hat{p}_1^2) \cdot \mathbb{E}\left[ \left(\sum_{k=0}^K g_{K,k}(\hat{p}_2)\prod_{l=0}^{k-1}(\hat{q}_2-\frac{l}{n}) \right)^2 \mathbbm{1}(E)\right] \\
&\le \mathbb{E}(\hat{p}_1^2) \cdot (K+1)\sum_{k=0}^K \left(2^{11K/2}A\cdot \left(\frac{3c_1\ln n}{n}\right)^{-k}\right)^2 \mathbb{E}\left[\prod_{l=0}^{k-1} (\hat{q}_2-\frac{l}{n})^2\right]\\
& = 2^{11K}(K+1)A^2\left(p^2+\frac{p}{m}\right)\cdot \sum_{k=0}^K \left(\frac{3c_1\ln n}{n}\right)^{-2k} \mathbb{E}\left[\prod_{l=0}^{k-1} (\hat{q}_2-\frac{l}{n})^2\right].
\end{align}
To evaluate the expectation, Lemma \ref{lemma.middlevariancebound} with $q=0$ is applied here to yield
\begin{align}
\mathbb{E}\left[\prod_{l=0}^{k-1} (\hat{q}_2-\frac{l}{n})^2\right] \le \left(\frac{2q(k\vee nq)}{n}\right)^k \le \left(\frac{4c_1q\ln n}{n}\right)^k\le \left(\frac{3c_1\ln n}{n}\right)^{2k}
\end{align}
thus
\begin{align}
\mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2\mathbbm{1}(E)] \le 2^{11K}(K+1)^2A^2\left(p^2+\frac{p}{m}\right)\le 2^{11K+2}K^2A^2\left(p^2+\frac{p}{m}\right).
\end{align}
By differentiation it is easy to show
\begin{align}
p\left[\ln\left(\frac{42c_1}{c_2^2}\cdot \frac{u(S)}{pn\ln n}\right)\vee 1\right]^2 \le \frac{84c_1}{c_2^2}\cdot \frac{u(S)}{n\ln n}\vee p \le \frac{84c_1}{c_2^2}\cdot \frac{u(S)}{n\ln n} + p
\end{align}
and note that $p\le u(S)q\le \frac{2c_1u(S)\ln n}{n}$, we have
\begin{align}
\mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2\mathbbm{1}(E)] &\le 2^{11K+3}K^2\left(p+\frac{1}{m}\right)\left(\frac{84c_1C_{\ln}^2}{c_2^2}\cdot \frac{u(S)}{n\ln n} + C_{\ln}^2p + p(\ln n)^2\right)\\
&\le 2^{11K+3}K^2\left(\frac{2c_1u(S)\ln n}{n}+\frac{1}{m}\right)\left(\frac{84c_1C_{\ln}^2}{c_2^2}\cdot \frac{u(S)}{n\ln n} + (C_{\ln}^2 + (\ln n)^2)\frac{2c_1u(S)\ln n}{n}\right)\\\label{eq.var_bound_upper}
&= \frac{16c_1u(S)\ln n}{n^{1-11c_2\ln 2}}\left(\frac{2c_1u(S)\ln n}{n}+\frac{1}{m}\right)\left(42C_{\ln}^2 + (c_2\ln n)^2(C_{\ln}^2 + (\ln n)^2)\right)
\end{align}
which together with (\ref{eq.var_decomp}) is the variance bound.
Now we start to analyze the bias of $\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)$. By triangle inequality,
\begin{align}
\left|\mathbb{E}\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) - p\ln q \right|
&\le \left|\mathbb{E}\left(\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) - p\ln q \right)\mathbbm{1}(E)\right| + \left|\mathbb{E}\left(\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) - p\ln q \right)\mathbbm{1}(E^c)\right|\\
&\le \left|\mathbb{E}\left({T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) - p\ln q \right)\mathbbm{1}(E)\right| + \left|\mathbb{E}\left(\tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) - T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)\right)\mathbbm{1}(E)\right| \nonumber\\
&\qquad\qquad + 2m^{-c_1/36}(1-p\ln q)\\
&\le \left|\mathbb{E}\left({T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) - p\ln q \right)\mathbbm{1}(E)\right| + \mathbb{E}\left|T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)\mathbbm{1}(|T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)|\ge 1)\mathbbm{1}(E)\right| \nonumber\\
&\qquad\qquad + 2m^{-c_1/36}(1-p\ln q)\\
&\equiv A_1+A_2+2m^{-c_1/36}(1-p\ln q).
\end{align}
Now we bound $A_1$ and $A_2$ separately. For $A_1$, since conditioning on $E$, the approximation region contains $q$, by (\ref{eq.approx_err}) we get
\begin{align}
A_1 \le p\mathbb{E} \left|\left(\sum_{k=0}^Kg_{K,k}(\hat{p}_2)q^k - \ln q\right)\mathbbm{1}(E)\right| \le C_{\ln }pW\left(\frac{42c_1}{c_2^2}\cdot \frac{u(S)}{pn\ln n}\right).
\end{align}
As for $A_2$, since for any random variable $X$ with finite second moment, we have
\begin{align}\label{eq.truncation}
\mathbb{E}[|X|\mathbbm{1}(|X|\ge 1)] \le \mathbb{E}[|X|^2\mathbbm{1}(|X|\ge 1)]\le \mathbb{E}[X^2],
\end{align}
by (\ref{eq.var_bound_upper}) we get
\begin{align}
A_2 &\le \mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)^2\mathbbm{1}(E)] \\
&\le \frac{16c_1u(S)\ln n}{n^{1-11c_2\ln 2}}\left(\frac{2c_1u(S)\ln n}{n}+\frac{1}{m}\right)\left(42C_{\ln}^2 + (c_2\ln n)^2(C_{\ln}^2 + (\ln n)^2)\right).
\end{align}
Now combining $A_1$ and $A_2$ completes the proof. \qed
\subsection{Proof of Lemma \ref{lem.T_ns_II}}
First we bound the variance of $T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)$. Recall that
\begin{align}
T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) = \sum_{k,l\ge 0,k+l\le K}h_{K,k,l}\prod_{i=0}^{k-1}\left(\hat{p}_1-\frac{i}{m} \right)\prod_{j=0}^{l-1}\left(\hat{q}_1-\frac{j}{n}\right).
\end{align}
We first bound the coefficients $|h_{K,k,l}|$. It is straightforward to see that
\begin{align}
\sup_{(x,y)\in R} \left|\sum_{k,l\ge 0,k+l\le K}h_{K,k,l}x^ky^l\right| \le 2\sup_{(x,y)\in R}|x\ln y| \le \frac{8c_1u(S)\ln n}{n}\ln \left(\frac{n}{4c_1\ln n}\right)
\le \frac{8c_1u(S)(\ln n)^2}{n} \equiv A.
\end{align}
We distinguish into two cases.
\subsubsection{Case I}
If $\frac{\ln m}{mu(S)}\le \frac{\ln n}{n}$, we have
\begin{align}
R \supset \left[ 0,\frac{2c_1\ln m}{m} \right] \times \left[ \frac{2c_1\ln n}{n},\frac{4c_1\ln n}{n} \right] \triangleq R_1.
\end{align}
Hence, for any $(x,y)\in R_1$, we have
\begin{align}
\left| \sum_{k=0}^K \left(\sum_{l=0}^{K-k} h_{K,k,l}y^l \right) x^k \right| \le A.
\end{align}
By Lemma \ref{lem.polycoeff}, we conclude that for any $y\in [\frac{2c_1\ln n}{n}, \frac{4c_1\ln n}{n}]$,
\begin{align}
\left|\sum_{l=0}^{K-k} h_{K,k,l}y^l \right|\le 2^{7K/2+1}A\cdot \left(\frac{c_1\ln m}{m}\right)^{-k}, \qquad k=0,1,\cdots,K.
\end{align}
Using Lemma \ref{lem.polycoeff} again, we have
\begin{align}
|h_{K,k,l}|&\le 2^{7(K-k)/2}\cdot 2^{7K/2+1}A \left(\frac{c_1\ln m}{m}\right)^{-k}\cdot \left(\frac{3c_1\ln n}{n}\right)^{-l}(3^{K-k}+1)\\
&\le 2^{9K+1} A\left(\frac{c_1\ln m}{m}\right)^{-k} \left(\frac{3c_1\ln n}{n}\right)^{-l}, \qquad \forall k,l\ge 0,k+l\le K.
\end{align}
\subsubsection{Case II}
If $\frac{\ln m}{mu(S)}>\frac{\ln n}{n}$, define $t\triangleq p/q$, we have
\begin{align}
R\supset \left\{ (t,q): 0\le t\le u(S), 0\le q\le \frac{2c_1\ln n}{n} \right\}\triangleq R_2
\end{align}
and for any $(t,q)\in R_2$, we have
\begin{align}
\left| \sum_{k=0}^K \left(\sum_{l=0}^{L-k} h_{K,k,l}q^{k+l}\right) t^k \right| = \left| \sum_{k,l\ge 0,k+l\le K} h_{K,k,l}(qt)^kq^l \right| \le A.
\end{align}
By Lemma \ref{lem.polycoeff}, we conclude that for any $q\in [0,\frac{c_1\ln n}{n}]$,
\begin{align}
\left|\sum_{l=0}^{L-k} h_{K,k,l}q^{k+l}\right| \le 2^{7K/2+1}A\cdot \left(\frac{u(S)}{2}\right)^{-k}.
\end{align}
By Lemma \ref{lem.polycoeff} again, we have
\begin{align}
|h_{K,k,l}| &\le 2^{7K/2+1}\cdot 2^{7K/2+1}A\left(\frac{u(S)}{2}\right)^{-k}\cdot \left(\frac{c_1\ln n}{n}\right)^{-(k+l)}\\
&= 2^{7K+2}A\left(\frac{c_1u(S)\ln n}{2n}\right)^{-k}\left(\frac{c_1\ln n}{n}\right)^{-l}, \qquad \forall k,l\ge 0,k+l\le n.
\end{align}
Hence, combining these two cases yields
\begin{align}
|h_{K,k,l}|\le 2^{9K+1}A\left[\left(\frac{c_1u(S)\ln n}{2n}\right)^{-k} + \left(\frac{c_1\ln m}{m}\right)^{-k}\right]\left(\frac{c_1\ln n}{n}\right)^{-l}.
\end{align}
Moreover, by Lemma \ref{lemma.middlevariancebound}, we have
\begin{align}
\mathbb{E}\left[\prod_{i=0}^{k-1} (\hat{p}_1-\frac{i}{m})^2\right] &\le \left(\frac{2p(k\vee mp)}{m}\right)^k \le \left(\frac{4c_1p(\ln n+\ln m)}{m}\right)^k\le \left(\frac{3c_1(\ln m+\ln n)}{m}\right)^{2k}\\
\mathbb{E}\left[\prod_{j=0}^{l-1} (\hat{q}_1-\frac{j}{n})^2\right] &\le \left(\frac{2q(l\vee nq)}{n}\right)^l \le \left(\frac{8c_1q\ln n}{n}\right)^l\le \left(\frac{6c_1\ln n}{n}\right)^{2l}.
\end{align}
Now by the triangle inequality and previous inequalities, we have
\begin{align}
\mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)^2]
&\le (K+1)^2\sum_{k,l\ge 0,0<k+l\le K} |h_{K,k,l}|^2 \mathbb{E}\left[\prod_{i=0}^{k-1}\left(\hat{p}_1-\frac{i}{m} \right)^2 \right]\mathbb{E}\left[\prod_{j=0}^{l-1}\left(\hat{q}_1-\frac{j}{n} \right)^2 \right]\\
&\le 2^{18K+3}(K+1)^2A^2\sum_{k,l\ge 0,0<k+l\le K}6^{2l}\left(3^{2k}\left(1+\frac{\ln n}{\ln m}\right)^{2k} + 12^{2k}\left(\frac{n(\ln m+\ln n)}{mu(S)\ln n}\right)^{2k}\right)\\
&\le 2^{18K+3}12^{2K}(K+1)^4A^2\left(1+\left(1+\frac{\ln n}{\ln m}\right)^{2K}+\left(\frac{n(\ln m+\ln n)}{mu(S)\ln n}\right)^{2K} \right)\\
&\le 2^{26K+7}K^4A^2\left(1+\left(1+\frac{\ln n}{\ln m}\right)^{2K}+\left(\frac{n(\ln m+\ln n)}{mu(S)\ln n}\right)^{2K} \right).
\end{align}
Hence, by Lemma \ref{lemma.cailow2} we get
\begin{align}
\mathsf{Var}(\tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)) &\le \mathsf{Var}({T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)) \le \mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)^2]\\
& \le 2^{26K+7}K^4A^2\left(1+\left(1+\frac{\ln n}{\ln m}\right)^{2K}+\left(\frac{n(\ln m+\ln n)}{mu(S)\ln n}\right)^{2K} \right)
\end{align}
which is the desired variance bound.
As for the bias, Lemma \ref{lem.DT_modulus} and Lemma \ref{lem.approx_2D} give
\begin{align}
|\mathbb{E} T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) - p\ln q| &\le E_K[p\ln q;R] + |h_{K,0,0}|\\
&\le 2E_K[p\ln q;R] \le 2M\omega_R^2(p\ln q,\frac{1}{K})
\le \frac{2MC_0}{c_2^2}\cdot \frac{u(S)}{n\ln n}
\end{align}
where $|h_{K,0,0}|\le E_K[p\ln q;R]$ is obtained by setting $(p,q)=(0,0)\in R$ in
\begin{align}
\sup_{(x,y)\in R}\left|\sum_{k,l\ge 0,k+l\le K} h_{K,k,l}x^ky^l - x\ln y \right| \le E_K[p\ln q;R].
\end{align}
Hence, by triangle inequality and (\ref{eq.truncation}), we get
\begin{align}
|\mathbb{E}[\tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) ]-p\ln q| &\le |\mathbb{E} T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) - p\ln q| + \mathbb{E}|T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)\mathbbm{1}(|T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)|\ge 1)|\\
&\le \frac{2MC_0}{c_2^2}\cdot \frac{u(S)}{n\ln n} + \mathbb{E}[T_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)^2]\\
&\le \frac{2MC_0}{c_2^2}\cdot \frac{u(S)}{n\ln n} +2^{26K+7}K^4A^2\left(1+\left(1+\frac{\ln n}{\ln m}\right)^{2K}+\left(\frac{n(\ln m+\ln n)}{mu(S)\ln n}\right)^{2K} \right)
\end{align}
as desired. \qed
\subsection{Proof of Lemma \ref{lem.overall_T_ns}}
We distinguish into three cases based on different values of $p$. For simplicity, we define
\begin{align}
\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) &\triangleq \tilde{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + L_H(\hat{p}_1)\\
\overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) &\triangleq \tilde{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) + L_H(\hat{p}_1).
\end{align}
\subsubsection{Case I}
We first consider the case where $p\le \frac{c_1\ln m}{2m}$. By the triangle inequality, the bias can be decomposed into
\begin{align}
|\mathsf{Bias}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))| &\le |\mathbb{E} \overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) + p\ln (p/q)| + (\mathbb{E}|\overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)| + \mathbb{E}|\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)| )\mathbb{P}(\hat{p}_{3}\ge \frac{c_1\ln m}{m})\\
&\le |\mathbb{E} \overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) + p\ln (p/q)| + 4\cdot \left(e/4\right)^{\frac{c_1\ln m}{2}}\\
&\lesssim \frac{1}{m\ln m} + \frac{u(S)}{n\ln n} + \frac{(u(S))^2}{n^{2-c_2B}} + m^{-\frac{c_1}{2}\ln(4/e)}\\
&\lesssim \frac{1}{m\ln m} + \frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}}
\end{align}
where we have used Lemma \ref{lem.T_ns_II} and Lemma \ref{lemma.poissontail} here. Similarly, by Lemma \ref{lemma.cailow1}, the variance can be upper bounded as
\begin{align}
\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))&\le \mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)) + \mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2))\mathbb{P}(\hat{p}_3\ge \frac{c_1\ln m}{m}) + |\mathbb{E}\overline{T}_{\textrm{\rm ns},\textrm{\rm II}} - \mathbb{E}\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}|^2\mathbb{P}(\hat{p}_3\ge \frac{c_1\ln m}{m})\\
&\le \mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)) + 2^2\cdot \left(e/4\right)^{\frac{c_1\ln m}{2}} + 4^2\cdot \left(e/4\right)^{\frac{c_1\ln m}{2}}\\
&\lesssim \frac{1}{m^{2-c_2B}} + \frac{(u(S))^2}{n^{2-c_2B}} + m^{-\frac{c_1}{2}\ln(4/e)}\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-\epsilon}}.
\end{align}
\subsubsection{Case II}
Next we consider the case where $\frac{c_1\ln m}{2m}<p<\frac{2c_1\ln m}{m}$. By Lemma \ref{lem.T_ns_1} and Lemma \ref{lem.T_ns_II}, the bias can be upper bounded as
\begin{align}
|\mathsf{Bias}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))| &\le |\mathbb{E} \overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + p\ln (p/q)| + |\mathbb{E} \overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) + p\ln(p/q)|\\
&\lesssim \frac{1}{m\ln m}+pW(\frac{u(S)}{pn\ln n}) + \frac{u(S)(\ln n)^5}{n^{1-11c_2\ln 2}}\left(\frac{u(S)\ln n}{n}+\frac{1}{m}\right) + \frac{1}{m\ln m}+\frac{u(S)}{n\ln n} + \frac{(u(S))^2}{n^{2-c_2B}}\\
&\lesssim \frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n} + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}.
\end{align}
The variance is obtained by Lemma \ref{lemma.cailow1} as follows:
\begin{align}
\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) &\le \mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)) + \mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) ) + (\mathbb{E}\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)-\mathbb{E} \overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1) )^2\\
&\lesssim \frac{u(S)(\ln n)^3}{mn} + \frac{p}{m} + \frac{1}{m^3} + \frac{u(S)(\ln n)^5}{n^{1-11c_2\ln 2}}\left(\frac{u(S)\ln n}{n}+\frac{1}{m}\right) + \frac{(\ln m)^4}{m^{2-c_2B}}+ \frac{u(S)}{n^{2-c_2B}} \nonumber\\
&\qquad\qquad + \left( \frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n} + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}\right)^2\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{(u(S))^4}{n^{4-2\epsilon}}+ \frac{(u(S))^2}{n^{2-2\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} + [pW(\frac{u(S)}{pn\ln n})]^2\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-2\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} + [pW(\frac{u(S)}{pn\ln n})]^2
\end{align}
where in the last step we have used that $n\gtrsim u(S)$.
\subsubsection{Case III}
Finally we consider the case where $p\ge \frac{2c_1\ln m}{m}$. By Lemma \ref{lem.T_ns_1} and Lemma \ref{lemma.poissontail},
\begin{align}
|\mathsf{Bias}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))| &\le |\mathbb{E} \overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + p\ln(p/q)| + (\mathbb{E}|\overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1)| + \mathbb{E}|\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)| )\mathbb{P}(\hat{p}_{3}\le \frac{c_1\ln m}{m})\\
&\le |\mathbb{E} \overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + p\ln(p/q)| + 4\cdot e^{-\frac{c_1\ln m}{2}}\\
&\lesssim \frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+ \frac{u(S)(\ln n)^5}{n^{1-11c_2\ln 2}}\left(\frac{u(S)\ln n}{n}+\frac{1}{m}\right) + m^{-c_1/2}\\
&\lesssim \frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}
\end{align}
and the variance is given by Lemma \ref{lemma.cailow1} that
\begin{align}
\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q}))&\le \mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + \mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm II}}(\hat{p}_1,\hat{q}_1))\mathbb{P}(\hat{p}_3\le \frac{c_1\ln m}{m}) + |\mathbb{E}\overline{T}_{\textrm{\rm ns},\textrm{\rm II}} - \mathbb{E}\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}|^2\mathbb{P}(\hat{p}_3\le \frac{c_1\ln m}{m})\\
&\le\mathsf{Var}(\overline{T}_{\textrm{\rm ns},\textrm{\rm I}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + 2^2\cdot e^{-\frac{c_1\ln m}{2}} + 4^2\cdot e^{-\frac{c_1\ln m}{2}}\\
&\lesssim \frac{u(S)(\ln n)^3}{mn} + \frac{p}{m} + \frac{1}{m^3} + \frac{u(S)(\ln n)^5}{n^{1-11c_2\ln 2}}\left(\frac{u(S)\ln n}{n}+\frac{1}{m}\right) + m^{-c_1/2}\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}.
\end{align}
A combination of these three cases completes the proof. \qed
\subsection{Proof of Lemma \ref{lem.overall_xi}}
As in the proof of Lemma \ref{lem.overall_T_ns}, we also distinguish into three cases.
\subsubsection{Case I}
We first consider the case where $q\le \frac{c_1\ln n}{2n}$. By Lemma \ref{lem.overall_T_ns} and Lemma \ref{lemma.poissontail},
\begin{align}
|\mathsf{Bias}(\xi(\hat{p},\hat{q}))| &\le |\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q})+p\ln(p/q)| + (\mathbb{E}|T_{\textrm{\rm ns}}(\hat{p},\hat{q})| + \mathbb{E}|\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})|)\mathbb{P}(\hat{q}_{3}\ge \frac{c_1\ln n}{n}) \\
&\le |\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q})+p\ln(p/q)| + (2+\ln n+5n^3+1+\frac{1}{2n}+1)\cdot (e/4)^{\frac{c_1\ln n}{2}}\\
&\lesssim \begin{cases}
\frac{1}{m\ln m} + \frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}}& p\le \frac{c_1\ln m}{2m},\\
\frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & \frac{c_1\ln m}{2m}<p<\frac{2c_1\ln m}{m},\\
\frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & p\ge \frac{2c_1\ln m}{m}.
\end{cases}
\end{align}
where we have used the fact that
\begin{align}
|\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})|\le |\tilde{T}_{\textrm{\rm s}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2)| + |U_H(\hat{p}_1)| + |L_H(\hat{p}_1)| \le 5n^3+\ln n+1+\frac{1}{2n} + 1
\end{align}
here. Similarly, by Lemma \ref{lemma.cailow1}, the variance can be upper bounded as
\begin{align}
\mathsf{Var}(\xi(\hat{p},\hat{q})) &\le \mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) + \left[\mathsf{Var}(\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})) + (\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q}) - \mathbb{E}\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q}))^2\right]\mathbb{P}(\hat{q}_3\ge \frac{c_1\ln n}{n})\\
&\lesssim \mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) + \left[\frac{1}{m^{2-\epsilon}}+\frac{p(1+\ln u(S))^2}{m} + \frac{q}{m} + \frac{u(S)}{mn} + \frac{pu(S)}{n} + (\frac{u(S)}{n\ln n})^2\right.\nonumber\\
&\qquad\qquad \left.+ 2^2 + (5n^3+\ln n+1+\frac{1}{2n}+1)^2\right]\cdot (e/4)^{\frac{c_1\ln n}{2}}\\
&\lesssim \begin{cases}
\frac{1}{m^{2-\epsilon}}+ \frac{(u(S))^2}{n^{2-\epsilon}} & p\le \frac{c_1\ln m}{2m},\\
\frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-2\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} + [pW(\frac{u(S)}{pn\ln n})]^2 & \frac{c_1\ln m}{2m}<p<\frac{2c_1\ln m}{m},\\
\frac{1}{m^{2-\epsilon}} + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}}& p\ge \frac{2c_1\ln m}{m}.
\end{cases}
\end{align}
\subsubsection{Case II}
Next we consider the case where $\frac{c_1\ln n}{2n}<q<\frac{2c_1\ln n}{n}$. By Lemma \ref{lem.T_s} and Lemma \ref{lem.overall_T_ns},
\begin{align}
|\mathsf{Bias}(\xi(\hat{p},\hat{q}))| &\le |\mathbb{E} \overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})+p\ln (p/q)| + |\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q})+p\ln (p/q)|\\
&\lesssim \frac{1}{m\ln m} + \frac{u(S)}{n\ln n} + \begin{cases}
\frac{1}{m\ln m} + \frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} & p\le \frac{c_1\ln m}{2m},\\
\frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & \frac{c_1\ln m}{2m}<p<\frac{2c_1\ln m}{m},\\
\frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n}) + \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & p\ge \frac{2c_1\ln m}{m}.
\end{cases}\\
&\lesssim \begin{cases}
\frac{1}{m\ln m} + \frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} & p\le \frac{c_1\ln m}{2m},\\
\frac{1}{m\ln m} + pW(\frac{u(S)}{pn\ln n})+\frac{u(S)}{n\ln n}+ \frac{(u(S))^2}{n^{2-\epsilon}} + \frac{u(S)}{mn^{1-\epsilon}} & p\ge \frac{c_1\ln m}{2m}.
\end{cases}
\end{align}
As for the variance, Lemma \ref{lemma.cailow1} is used here to yield
\begin{align}
\mathsf{Var}(\xi(\hat{p},\hat{q}))&\le \mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) + \mathsf{Var}(\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})) + |\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q}) - \mathbb{E}\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})|^2\\
&\lesssim \mathsf{Var}(T_\textrm{\rm ns}(\hat{p},\hat{q})) + \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m}+ \frac{u(S)}{mn} + \frac{pu(S)}{n} + (\frac{u(S)}{n\ln n})^2 \nonumber\\
&\qquad\qquad + |\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q})+p\ln (p/q)|^2 + |\mathbb{E}\overline{T}_{\textrm{\rm s}}(\hat{p}_1,\hat{q}_1;\hat{p}_2,\hat{q}_2) + p\ln(p/q)|^2\\
&\lesssim \frac{1}{m^2} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m} + \frac{pu(S)}{n} + \frac{(u(S))^2}{n^{2-2\epsilon}}+\begin{cases}
\frac{u(S)}{mn} & p\le \frac{c_1\ln m}{2m},\\
\frac{u(S)}{mn^{1-2\epsilon}} + [pW(\frac{u(S)}{pn\ln n})]^2& p>\frac{c_1\ln m}{2m}.
\end{cases}
\end{align}
where we have used the fact that $n\gtrsim u(S)$ here.
\subsubsection{Case III}
Finally we come to the case where $q\ge\frac{2c_1\ln n}{n}$. By Lemma \ref{lem.T_s} and Lemma \ref{lemma.poissontail},
\begin{align}
|\mathsf{Bias}(\xi(\hat{p},\hat{q}))| &\le |\mathbb{E} \overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})+p\ln(p/q)| + (\mathbb{E}|T_{\textrm{\rm ns}}(\hat{p},\hat{q})| + \mathbb{E}|\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})|)\mathbb{P}(\hat{q}_{3}\le \frac{c_1\ln n}{n}) \\
&\lesssim \frac{1}{m\ln m}+\frac{u(S)}{n\ln n} + (2+5n^3+\ln n+1+\frac{1}{2n}+1)e^{-\frac{c_1\ln n}{2n}}\\
&\lesssim \frac{1}{m\ln m} + \frac{u(S)}{n\ln n}.
\end{align}
The variance bound is obtained in a similar fashion via Lemma \ref{lemma.cailow1}:
\begin{align}
\mathsf{Var}(\xi(\hat{p},\hat{q})) &\le \mathsf{Var}(\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q})) + \left[\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) + (\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q}) - \mathbb{E}\overline{T}_{\textrm{\rm s}}(\hat{p},\hat{q}))^2\right]\mathbb{P}(\hat{q}_3\le \frac{c_1\ln n}{n})\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m}+ \frac{u(S)}{mn} + \frac{pu(S)}{n} + (\frac{u(S)}{n\ln n})^2 \nonumber\\
&\qquad\qquad + \left[2^2 + 2^2 + \left(5n^3+\ln n+1+\frac{1}{2n}+1\right)^2\right]\cdot e^{-\frac{c_1\ln n}{2n}}\\
&\lesssim \frac{1}{m^{2-\epsilon}} + \frac{p(1+\ln u(S))^2}{m} + \frac{q}{m}+ \frac{u(S)}{mn} + \frac{pu(S)}{n} + (\frac{u(S)}{n\ln n})^2.
\end{align}
Combining these three cases yields the desired result. \qed
\subsection{Proof of Lemma \ref{lem.adaptive}}
As before, we first analyze the variance. By \cite[Lemma 20]{Jiao--Venkat--Han--Weissman2015minimax}, we know that there exists some constant $C>0$ such that for any $x\in [0,\frac{2c_1\ln n}{n}]$,
\begin{align}
\left| \sum_{k=1}^{K+1} g_{K,k}\left(\frac{2c_1\ln n}{n}\right)^{1-k}x^k-x\ln x \right| \le \frac{C}{n\ln n}.
\end{align}
By triangle inequality, for any $x\in [0,\frac{2c_1\ln n}{n}]$,
\begin{align}
\left| \sum_{k=1}^{K+1} g_{K,k}\left(\frac{2c_1\ln n}{n}\right)^{1-k}x^k\right| &\le \left| \sum_{k=1}^{K+1} g_{K,k}\left(\frac{2c_1\ln n}{n}\right)^{1-k}x^k-x\ln x \right| + |x\ln x| \\
&\le \frac{C}{n\ln n} + \frac{2c_1\ln n}{n} \cdot \ln\frac{n}{2c_1\ln n} \\
&\le \frac{C}{n\ln n} + \frac{2c_1(\ln n)^2}{n} \equiv A.
\end{align}
As a result, by Lemma \ref{lem.polycoeff}, for any $k=1,\cdots,K+1$, we have
\begin{align}
|g_{K,k}|\left(\frac{2c_1\ln n}{n}\right)^{1-k} \le 2^{7K/2+1}A\cdot \left(\frac{c_1\ln n}{n}\right)^{-k}.
\end{align}
Hence, by triangle inequality again and Lemma \ref{lemma.middlevariancebound}, we have
\begin{align}
\mathsf{Var}(T_{\textrm{\rm ns}}(\hat{p},\hat{q})) &\le \mathbb{E}[T_{\textrm{\rm ns}}(\hat{p},\hat{q})^2] \\
&\le (K+1)\sum_{k=0}^K |g_{K,k+1}|^2\left(\frac{2c_1\ln n}{n}\right)^{-2k}\cdot \mathbb{E}[\hat{p}^2]\mathbb{E}\left[\prod_{l=0}^{k-1}(\hat{q}-\frac{l}{n})^2\right]\\
&\le 2^{7K+2}(K+1)A^2\sum_{k=0}^K \left(\frac{c_1\ln n}{n}\right)^{-2(k+1)}\cdot \left(\frac{p}{m}+p^2\right)\left(\frac{4qc_1\ln n}{n}\right)^k\\
&\le 2^{11K+2}(K+1)A^2\sum_{k=0}^K \left(\frac{c_1\ln n}{n}\right)^{-2}\cdot \left(\frac{p}{m}+p^2\right)\\
&\le 2^{11K+2}(K+1)^2\frac{(C+2c_1(\ln n)^3)^2}{(\ln n)^2}\cdot \left(\frac{2u(S)}{c_1mn\ln n}+\frac{4(u(S))^2}{n^2}\right)\\
&\le 2^{11K+4}c_2^2(C+2c_1(\ln n)^3)^2\cdot \left(\frac{2u(S)}{c_1mn\ln n}+\frac{4(u(S))^2}{n^2}\right)
\end{align}
and by Lemma \ref{lemma.cailow2}, we have
\begin{align}
\mathsf{Var}(\tilde{T}_\textrm{\rm ns}(\hat{p},\hat{q})) \le \mathsf{Var}(T_\textrm{\rm ns}(\hat{p},\hat{q}))\le 2^{11K+4}c_2^2(C+2c_1(\ln n)^3)^2\cdot \left(\frac{2u(S)}{c_1mn\ln n}+\frac{4(u(S))^2}{n^2}\right).
\end{align}
As for the bias, by construction we have
\begin{align}
\left|\mathbb{E} T_\textrm{\rm ns}(\hat{p},\hat{q}) - p\ln q\right| &= p\left|\sum_{k=0}^{K} g_{K,k+1}\left(\frac{2c_1\ln n}{n}\right)^{-k}q^k - \ln q\right| \\
&\le u(S)q\left|\sum_{k=1}^{K+1} g_{K,k}\left(\frac{2c_1\ln n}{n}\right)^{1-k}q^{k-1} - \ln q\right| \\
&= u(S)\left|\sum_{k=1}^{K+1} g_{K,k}\left(\frac{2c_1\ln n}{n}\right)^{1-k}q^{k} - q\ln q\right|\\
&\le \frac{Cu(S)}{n\ln n}.
\end{align}
Hence, by triangle inequality and (\ref{eq.truncation}), we get
\begin{align}
|\mathsf{Bias}(\tilde{T}_{\textrm{\rm ns}}(\hat{p},\hat{q}) )| &\le |\mathbb{E} T_{\textrm{\rm ns}}(\hat{p},\hat{q}) - p\ln q| + \mathbb{E}|T_{\textrm{\rm ns}}(\hat{p},\hat{q})\mathbbm{1}(|T_{\textrm{\rm ns}}(\hat{p},\hat{q})|\ge 1)|\\
&\le \frac{Cu(S)}{n\ln n} + \mathbb{E}[T_{\textrm{\rm ns}}(\hat{p}_1,\hat{q}_1)^2]\\
&\le \frac{Cu(S)}{n\ln n} + 2^{11K+4}c_2^2(C+2c_1(\ln n)^3)^2\cdot \left(\frac{2u(S)}{c_1mn\ln n}+\frac{4(u(S))^2}{n^2}\right)
\end{align}
as desired. \qed
\subsection{Proof of Lemma \ref{lem.equivalence_approx}}
Fix $\delta>0$. Let $\hat{D}(\mathbf{X},\mathbf{Y})$ be a near-minimax estimator of $D(P\|Q)$ under the multinomial model with an upper bound $(1+\epsilon)u(S)$ on the likelihood ratio. Note that the estimator $\hat{D}$ obtains the sample sizes $m,n$ from observations. By definition, we have
\begin{align}
\sup_{(P,Q)\in \mathcal{U}_{S,u(S)}} \mathbb{E}_{\mathrm{Multinomial}}\left(\hat{D} - D(P\|Q)\right)^2 \le R(S,m,n,(1+\epsilon)u(S)) + \delta.
\end{align}
Now given $(P,Q)\in\mathcal{U}_{S,u(S)}(\epsilon)$, let $\mathbf{X}=[X_1,\cdots,X_S]^T$ and $\mathbf{Y}=[Y_1,\cdots,Y_S]^T$ with $(X_i,Y_i)\sim \mathsf{Poi}(mp_i)\times \mathsf{Poi}(nq_i)$. Write $m'=\sum_{i=1}^S X_i$ and $n'=\sum_{i=1}^S Y_i$, we use the estimator $\hat{D}(\mathbf{X},\mathbf{Y})$ to estimate $D(P\|Q)$.
Note that conditioned on $m'=M$, we have $\mathbf{X}\sim \mathsf{Multinomial}(M,\frac{P}{\sum_{i=1}^S p_i})$, and similarly for $\mathbf{Y}$. Moreover, $(P,Q)\in\mathcal{U}_{S,u(S)}(\epsilon)$ implies that $(P,\frac{Q}{\sum_{i=1}^S q_i})\in \mathcal{U}_{S,(1+\epsilon)u(S)}$ by construction. By the triangle inequality we have
\begin{align}
\frac{1}{2}R_P(S,m,n,u(S),\epsilon) &\le \frac{1}{2}\mathbb{E}_{(P,Q)}\left( \hat{D}-D(P\|Q) \right)^2 \\
&\le \mathbb{E}_{(P,Q)}\left( \hat{D}-D\left(P\|\frac{Q}{\sum_{i=1}^S q_i}\right) \right)^2 + \left(D\left(P\|\frac{Q}{\sum_{i=1}^S q_i}\right) - D(P\|Q)\right)^2\\
&\le \sum_{k,l=0}^\infty \mathbb{E}_{(P,Q)}\left(\left. \hat{D}-D\left(P\|\frac{Q}{\sum_{i=1}^S q_i} \right)\right| m'=k,n'=l \right)^2\mathbb{P}(m'=k)\mathbb{P}(n'=l) + (\ln \sum_{i=1}^S q_i)^2\\
&\le \sum_{k,l=0}^\infty \left(R(S,k,l,(1+\epsilon)u(S)) + \delta\right)\mathbb{P}(m'=k)\mathbb{P}(n'=l) + 2\epsilon^2\\
&= \sum_{k\ge \frac{m}{2},l\ge \frac{(1-\epsilon)n}{2}} R(S,k,l,(1+\epsilon)u(S))\mathbb{P}(m'=k)\mathbb{P}(n'=l) \nonumber\\
&\qquad\qquad + (\ln[(1+\epsilon)u(S)])^2(\mathbb{P}(m'<\frac{m}{2})+\mathbb{P}(n'<\frac{(1-\epsilon)n}{2}))+\delta + 2\epsilon^2\\
&\le R(S,\frac{m}{2},\frac{(1-\epsilon)n}{2},(1+\epsilon)u(S)) + (\ln[(1+\epsilon) u(S)])^2\left(\exp(-\frac{m}{8}) + \exp(-\frac{(1-\epsilon)n}{8})\right) + \delta + 2\epsilon^2
\end{align}
where we have used Lemma \ref{lemma.poissontail}. Then the result follows from the arbitrariness of $\delta$. \qed
\subsection{Proof of Lemma \ref{lem.chebyshev_poly}}
The properties of Chebyshev polynomials were well studied in \cite{mason2002chebyshev}. In particular, the Chebyshev polynomial $T_{2(K+2)}(x)$ is an even function and takes the form
\begin{align}
T_{2(K+2)}(x) = S_{2K}(x)x^4 - 2(-1)^K(K+2)^2x^2 + (-1)^K
\end{align}
for some even polynomial $S_{2K}(x)$ of degree $2K$. Since $|T_{2(K+2)}(x)|\le 1$ for any $x\in [-1,1]$, by triangle inequality
\begin{align}
\left|(-1)^K \frac{S_{2K}(x)x^4}{2(K+2)^2} - x^2\right| \le \frac{|T_{2(K+2)}(x)|}{2(K+2)^2} + \frac{1}{2(K+2)^2} \le \frac{1}{(K+2)^2}.
\end{align}
Now the desired result follows from the variable substitution $y=\Delta_n x^2\in [0,\Delta_n]$. \qed
\section{Proof of Auxiliary Lemmas}
\subsection{Proof of Lemma \ref{lem.minimax_unbounded}}
Fix $m,n$ and an arbitrary (possibly randomized) estimator $\hat{D}$. Denote by $\mu$ the (possibly randomized) decision made by $\hat{D}$ conditioning on the event $E$ where $m$ first symbols and no other symbols from $P$, and $n$ second symbols and no other symbols from $Q$ are observed. Note that $\mu$ is a probability measure on $\mathbb{R}$. Choose $P=(1,0,\cdots,0)$ and $Q=(\delta,1-\delta,0,\cdots,0)$ with $\delta\in(0,1)$ to be specified in the sequel, then $P\ll Q$. Hence, with probability at least $(1-\delta)^n$, the event $E$ holds, and thus
\begin{align}
\sup_{(P,Q)\in\mathcal{U}_S} \mathbb{E}_{(P,Q)}\left(\hat{D}-D(P\|Q)\right)^2\ge (1-\delta)^n\cdot \int_{\mathbb{R}} (a-\ln(1/\delta))^2\mu(da).
\end{align}
As a result, denote by $a_{1/2}=\inf\{a\in\mathbb{R}: \mu((-\infty,a])\ge \frac{1}{2}\}$ a median of $\mu$, choosing $\delta=\frac{1}{2}\wedge\exp(-a_{1/2}-M)$ for any $M>0$ yields
\begin{align}
\sup_{(P,Q)\in\mathcal{U}_S} \mathbb{E}_{(P,Q)}\left(\hat{D}-D(P\|Q)\right)^2\ge (1-\frac{1}{2})^n\cdot \int_{(-\infty,a_{1/2}]} (a-\ln(1/\delta))^2\mu(da) \ge \frac{M^2}{2^{n+1}}.
\end{align}
Letting $M\to\infty$ yields the desired result. \qed
\subsection{Proof of Lemma \ref{lemma.poissonmultinomial}}
Similar to the proof of \cite[Lemma 16]{Jiao--Venkat--Han--Weissman2015minimax}, we can show that
\begin{align}
R_P(S,m,n,u(S),\pi) = \sum_{k,l=0}^\infty R(S,k,l,u(S),\pi)\mathbb{P}(\mathsf{Poi}(m)=k)\mathbb{P}(\mathsf{Poi}(n)=l)
\end{align}
where $R(\cdot,\cdot,\cdot,\cdot,\pi)$ and $R_P(\cdot,\cdot,\cdot,\cdot,\pi)$ represent the Bayes error given prior $\pi$ under the Multinomial model and the Poisson sampling model, respectively. On one hand, we have
\begin{align}
R_P(S,m,n,u(S),\pi) &\ge \sum_{0\le k\le 2m, 0\le l\le 2n}R(S,k,l,u(S),\pi)\mathbb{P}(\mathsf{Poi}(m)=k)\mathbb{P}(\mathsf{Poi}(n)=l)\\
&\ge R(S,2m,2n,u(S),\pi)\mathbb{P}(\mathsf{Poi}(m)\le 2m)\mathbb{P}(\mathsf{Poi}(n)\le 2n)\\
&\ge \frac{1}{4}R(S,2m,2n,u(S),\pi)
\end{align}
where we have used the Markov inequality to get $\mathbb{P}(\mathsf{Poi}(m)\le 2m)\ge \frac{1}{2}$ and $\mathbb{P}(\mathsf{Poi}(n)\le 2n)\ge \frac{1}{2}$. On the other hand, note that $D(P\|Q)\le \ln u(S)$ whenever $(P,Q)\in\mathcal{U}_{S,u(S)}$, by the Poisson tail bound in Lemma \ref{lemma.poissontail} we also have
\begin{align}
R_P(S,m,n,u(S),\pi)&\le \sum_{k>m/2, l> n/2}R(S,k,l,u(S),\pi)\mathbb{P}(\mathsf{Poi}(m)=k)\mathbb{P}(\mathsf{Poi}(n)=l) \nonumber\\
&\qquad\qquad + (\ln u(S))^2\left(\mathbb{P}(\mathsf{Poi}(m)\le \frac{m}{2}) + \mathbb{P}(\mathsf{Poi}(n)\le \frac{n}{2})\right)\\
&\le \sum_{k>m/2, l> n/2}R(S,\frac{m}{2},\frac{n}{2},u(S),\pi)\mathbb{P}(\mathsf{Poi}(m)=k)\mathbb{P}(\mathsf{Poi}(n)=l) + (\ln u(S))^2\left(\exp(-\frac{m}{8})+\exp(-\frac{n}{8})\right)\\
&\le R(S,\frac{m}{2},\frac{n}{2},u(S),\pi) + (\ln u(S))^2\left(\exp(-\frac{m}{8})+\exp(-\frac{n}{8})\right).
\end{align}
By the minimax theorem \cite{Wald1950statistical}, taking supremum over all priors $\pi$ yields the desired result.\qed
\subsection{Proof of Lemma \ref{lem.approx_log}}
We apply the general approximation theory on convex polytopes to our one-dimensional case where $[a,b]$ is an interval. Note that by polynomial scaling,
\begin{align}
E_n[\ln x;[a,b]] = E_n[\ln \left((b-a)x+a\right); [0,1]] = E_n[\ln (x+\frac{a}{b-a}); [0,1]]
\end{align}
it suffices to consider the function $h(x)=\ln(x+\Delta)$ defined on $[0,1]$, where $\Delta=\frac{a}{b-a}>0$. In this case, the second-order Ditzian--Totik modulus of smoothness in (\ref{eq.DT_modulus_general}) is reduced to \cite{Ditzian--Totik1987}
\begin{align}\label{eq.DT_modulus_int}
\omega_{[0,1]}^2(f,t) = \omega_\varphi^2(f,t) \triangleq \sup\left\{ \left|f(u)+f(v)-2f\left(\frac{u+v}{2}\right)\right|: u,v\in [0,1], |u-v|\le 2t\varphi(\frac{u+v}{2}) \right\}
\end{align}
where $\varphi(x)=\sqrt{x(1-x)}$. For the evaluation of $\omega_\varphi^2(h,t)$ for $t\in [0,1]$, we write $u=r+s,v=r-s$ with $u,v\in [0,1]$ and $0\le s\le t\varphi(r)$ in the definition. Then for $\Delta>t^2$, by Taylor expansion we have
\begin{align}\label{eq.taylor_exp}
\left|h(r+s)+h(r-s)-2h(r)\right| \le 2\sum_{k=1}^\infty \frac{|h^{(2k)}(r)|}{(2k)!}s^{2k} = \sum_{k=1}^\infty\frac{1}{k}\cdot \frac{s^{2k}}{(r+\Delta)^{2k}} \le \sum_{k=1}^\infty \frac{1}{k}\left(\frac{t^{2}r(1-r)}{(r+\Delta)^{2}}\right)^k.
\end{align}
By differentiation, it is easy to show that the maximum of $\frac{r(1-r)}{(r+\Delta)^2}$ is attained at $r=\frac{\Delta}{2\Delta+1}$, and the corresponding maximum is $\frac{1}{4\Delta(1+\Delta)}$. Hence, by (\ref{eq.taylor_exp}) we have
\begin{align}
\left|h(r+s)+h(r-s)-2h(r)\right| \le \sum_{k=1}^\infty \left(\frac{t^2}{4\Delta(\Delta+1)}\right)^k\le \frac{t^2}{4\Delta(\Delta+1)}\sum_{k=1}^\infty \left(\frac{1}{4}\right)^{k-1} = \frac{t^2}{3\Delta(\Delta+1)} \le \frac{t^2}{3\Delta}
\end{align}
i.e., we conclude that $\omega_\varphi^2(h,t)\le \frac{t^2}{3\Delta}$ when $\Delta>t^2$.
For $\Delta\le t^2$, the concavity of $\ln(\cdot)$ yields that the maximum-achieving pair $(u,v)$ must satisfy one of the following: (1) $u=1$; (2) $v=0$; (3) $s=t\varphi(r)$.
We start with the case where $s=t\varphi(r)$, then $\frac{t^2}{1+t^2}\le r\le \frac{1}{1+t^2}$. In this case (\ref{eq.taylor_exp}) still holds, but now the maximum of $\frac{r(1-r)}{(r+\Delta)^2}$ is attained at $r=\frac{t^2}{1+t^2}$, and the corresponding inequality becomes
\begin{align}
\left|h(r+s)+h(r-s)-2h(r)\right| &\le \sum_{k=1}^\infty \frac{1}{k}\left(\frac{t^4}{(t^2+(1+t^2)\Delta)^2}\right)^k
\le \sum_{k=1}^\infty \frac{1}{k}\left(\frac{t^4}{(t^2+\Delta)^2}\right)^k\\
&=-\ln \left(1 - \frac{t^4}{(t^2+\Delta)^2} \right) = \ln\left(\frac{(t^2+\Delta)^2}{\Delta(\Delta+2t^2)} \right)\le \ln\left(\frac{2t^2+\Delta}{\Delta}\right).
\end{align}
Note that this inequality only requires $r\ge \frac{t^2}{1+t^2}$, hence it also holds for the case where $u=1$.
Now we are only left with the case where $v=0$, then $s=r$ and $r\le \frac{t^2}{1+t^2}$. As a result,
\begin{align}
\left|h(r+s)+h(r-s)-2h(r)\right| &= 2\ln(r+\Delta) - \ln(2r+\Delta) - \ln(\Delta) \\
&= \ln\left(1+ \frac{r^2}{\Delta(2r+\Delta)}\right)\le \ln\left(1+ \frac{r}{2\Delta}\right)\le \ln\left(1+\frac{t^2}{2\Delta}\right).
\end{align}
In summary, for $\Delta\le t^2$ we can obtain that
\begin{align}
\omega_\varphi^2(h,t) \le \ln\left(1+\frac{2t^2}{\Delta}\right)
\end{align}
and Lemma \ref{lem.approx_log} follows from the previous upper bounds on $\omega_\varphi^2(h,n^{-1})$ and Lemma \ref{lem.DT_modulus}. \qed
\subsection{Proof of Lemma \ref{lem.approx_2D}}
It suffices to prove the claim for $\omega_T^2(f,1/K)$, where $T$ is the following triangle containing $R$:
\begin{align}
T\triangleq \left\{ (x,y): 0\le y\le \frac{2c_1\ln n}{n}, 0\le x\le u(S)y \right\}.
\end{align}
Denote by $E_1,E_2,E_3$ three edges of this triangle (excluding endpoints):
\begin{align}
E_1 &\triangleq \{(x,y): 0< x < \frac{2c_1u(S)\ln n}{n}, y=\frac{2c_1\ln n}{n}\}\\
E_2 &\triangleq \{(x,y): x = 0, 0< y< \frac{2c_1\ln n}{n}\}\\
E_3 &\triangleq \{(x,y): x = u(S)y, 0< y< \frac{2c_1\ln n}{n}\}.
\end{align}
Since $|\Delta_{h\tilde{d}_T(e,x,y)e}^2f(x,y)|$ is a continuous function with respect to $(x,y,e,h)\in T\times S^{1}\times [0,1]$, which is a compact set, we can assume that $(x_0,y_0,e_0,h_0)$ achieves the supremum. Let $A,B$ be the intersection of the line $\ell$ passing through $(x_0,y_0)$ with direction $e_0$ with the triangle $T$.
If either $A$ or $B$ belongs to $E_1$, say, $B\in E_1$, then for sufficiently small $\epsilon>0$, the line connecting $A$ and $(x_0+ \epsilon,y_0)$ (resp. $(x_0-\epsilon,y_0)$) intersects $E_1$ with direction $e_1$ (resp. $e_2$). Hence, by the similarity relation in geometry, the $y$-coordinates of $(x_0,y_0)+h_0\tilde{d}_T(e_0,x_0,y_0)$ and $(x_0+\epsilon,y_0)+h_0\tilde{d}_T(e_1,x_0+\epsilon,y_0)$ are equal, and similarly for others. Hence, by linearity of $f(x,y)=x\ln y$ in $x$,
\begin{align}
2|\Delta_{h_0\tilde{d}_T(e_0,x_0,y_0)e_0}^2f(x_0,y_0)| &=|\Delta_{h_0\tilde{d}_T(e_1,x_0+\epsilon,y_0)e_1}^2f(x_0+\epsilon,y_0) + \Delta_{h_0\tilde{d}_T(e_2,x_0-\epsilon,y_0)e_2}^2f(x_0-\epsilon,y_0)|\\
&\le |\Delta_{h_0\tilde{d}_T(e_1,x_0+\epsilon,y_0)e_1}^2f(x_0+\epsilon,y_0)| + |\Delta_{h_0\tilde{d}_T(e_2,x_0-\epsilon,y_0)e_2}^2f(x_0-\epsilon,y_0)|
\end{align}
i.e., we can always perturb $(x_0,y_0,e_0)$ such that $\ell$ does not intersect $E_1$.
Now we assume that $A=(0,y_1), B=(u(S)y_2,y_2)$ with $y_1,y_2\in [0,\frac{2c_1\ln n}{n}]$. If $y_2=y_1$, then $f(x,y)$ is linear on $\ell$, and $\omega_T^2(f,1/K)$ is zero. If $y_2\neq y_1$, the function on $\ell$ becomes
\begin{align}
h(y) = \frac{y-y_1}{y_2-y_1}y_2u(S)\ln y,\qquad y\in [u,v]
\end{align}
where $u\triangleq y_1\wedge y_2$ and $v\triangleq y_1\vee y_2$. Hence, by the sub-additivity of the Ditzian--Totik modulus of smoothness, for $t\in [0,1]$ we have
\begin{align}
\omega_T^2(f,t) &= \omega_{[u,v]}^2(h,t) \le \frac{y_2u(S)}{|y_2-y_1|}\cdot\omega_{[u,v]}^2(y\ln y,t) + \frac{y_1y_2u(S)}{|y_2-y_1|}\cdot\omega_{[u,v]}^2 (\ln y, t)\\
&\le u(S)v\left[\frac{\omega_{[u,v]}^2(y\ln y,t)}{v-u} + \frac{u}{v-u}\cdot\omega_{[u,v]}^2 (\ln y, t)\right]
\end{align}
By the proof of Lemma \ref{lem.approx_log} (where we have used $\ln (1+x)\le x$ when $\Delta\le t^2$), we have
\begin{align}
\omega_{[u,v]}^2 (\ln y, t) \le \frac{3t^2(v-u)}{u}.
\end{align}
As for $\omega_{[u,v]}^2(y\ln y,t)$, we distinguish into two cases. If $v\le 2u$, by Taylor expansion we have
\begin{align}
\omega_{[u,v]}^2(y\ln y,t)\le \sup_{y\in [u,v]}|(y\ln y)''|\cdot (v-u)^2t^2 = \frac{(v-u)^2t^2}{u} \le (v-u)t^2.
\end{align}
Otherwise, if $v>2u$, since it has been shown in \cite{Jiao--Venkat--Weissman2014MLE} that $\omega_{[0,1]}^2(y\ln y,t) = \frac{2t^2\ln 2}{1+t^2}$, by scaling $[u,v]\mapsto [u/v,1]$ we have
\begin{align}
\omega_{[u,v]}^2(y\ln y,t)\le v\cdot \omega_{[u/v,1]}^2(y\ln y,t)\le v\cdot \omega_{[0,1]}^2(y\ln y,t)\le 2\ln 2\cdot vt^2 \le 4\ln 2\cdot (v-u)t^2.
\end{align}
A combination of the previous three inequalities yields
\begin{align}
\omega_T^2(f,t) &\le u(S)v\left[\frac{4\ln 2\cdot (v-u)t^2}{v-u} + \frac{u}{v-u}\cdot \frac{3t^2(v-u)}{u}\right] \\
&= (3+4\ln 2)u(S)vt^2\\
&\le \frac{(6+8\ln 2)c_1u(S)\ln n}{n}\cdot t^2
\end{align}
where we have used $v=y_1\vee y_2\le \frac{2c_1\ln n}{n}$. Now the desired result follows directly by choosing $t=1/K^2$ and $C_0=(6+8\ln 2)c_1$. \qed
\subsection{Proof of Lemma \ref{lem.polycoeff}}
First we assume that $b=t+1, a=t-1, t\neq 0$, and write
\begin{align}
p_n(x) = \sum_{\nu=0}^n a_\nu x^\nu = \sum_{\nu=0}^{n} b_\nu(x-t)^\nu.
\end{align}
By Lemma \ref{lemma.chebyshev} and the related discussions, we know that $|b_\nu|\le 2^{3n/2}$ for any $\nu=0,\cdots,n$. Comparing coefficients yields
\begin{align}
|a_\nu| &= \left|\sum_{\mu=\nu}^{n} \binom{\mu}{\nu} (-t)^{\mu-\nu}b_\mu\right|\\
& \le 2^{3n/2}|t|^{-\nu} \sum_{\mu=\nu}^{n} \binom{\mu}{\nu} |t|^{\mu} \\
&\le 2^{3n/2}|t|^{-\nu} 2^n(n+1)(|t|^n+1) \\
&\le 2^{7n/2}|t|^{-\nu}(|t|^n+1).
\end{align}
In the general case, the desired result is obtained by scaling.
\qed
\subsection{Proof of Lemma \ref{lem.varprod}}
It is straightforward to show
\begin{align}
\mathsf{Var}(XY) &= \mathbb{E}[(XY)^2] - (\mathbb{E}[XY])^2\\
&= \mathbb{E}[X^2]\mathbb{E}[Y^2] - (\mathbb{E} X)^2(\mathbb{E} Y)^2\\
&= (\mathsf{Var}(X)+(\mathbb{E} X)^2)(\mathsf{Var}(Y)+(\mathbb{E} Y)^2) -(\mathbb{E} X)^2(\mathbb{E} Y)^2\\
&= \mathsf{Var}(X)(\mathbb{E} Y)^2 + \mathsf{Var}(Y)(\mathbb{E} X)^2 + \mathsf{Var}(X)\mathsf{Var}(Y)
\end{align}
as desired.\qed
\bibliographystyle{IEEEtran}
|
\subsection*{#1}}
\newcommand{\customsubsubsection}[1]{\subsubsection*{#1}}
\pgfdeclarelayer{foreground}
\pgfdeclarelayer{background}
\pgfsetlayers{main,foreground,background}
\makeatletter
\pgfkeys{%
/tikz/on layer/.code={
\pgfonlayer{#1}\begingroup
\aftergroup\endpgfonlayer
\aftergroup\endgroup
},
/tikz/node on layer/.code={
\gdef\node@@on@layer{%
\setbox\tikz@tempbox=\hbox\bgroup\pgfonlayer{#1}\unhbox\tikz@tempbox\endpgfonlayer\egroup}
\aftergroup\node@on@layer
},
/tikz/end node on layer/.code={
\endpgfonlayer\endgroup\endgroup
}
}
\def\node@on@layer{\aftergroup\node@@on@layer}
\makeatother\def0.7pt{0.7pt}
\tikzstyle{oldmorphism}=[minimum width=30pt, minimum height=16pt, draw, font=\small, inner sep=0pt, fill=white, line width=0.7pt]
\tikzstyle{cross}=[preaction={draw=white, -, line width=10pt}]
\tikzstyle{braid}=[double=black, line width=3*0.7pt, double distance=0.7pt, white]
\tikzstyle{string}=[line width=0.7pt]
\tikzstyle{scalar}=[circle, inner sep=0pt, minimum width=15pt, draw, line width=0.7pt]
\tikzstyle{dot}=[circle, draw=black, fill=black!25, inner sep=.4ex, line width=0.7pt, node on layer=foreground]
\tikzstyle{blackdot}=[circle, draw=black, fill=black!35, inner sep=.5ex, line width=0.7pt, node on layer=foreground]
\tikzstyle{whitedot}=[circle, draw=black, fill=white, inner sep=.5ex, line width=0.7pt, node on layer=foreground]
\tikzstyle{reddot}=[circle, draw=black, fill=red, inner sep=.4ex, line width=0.7pt, node on layer=foreground]
\tikzstyle{bluedot}=[circle, draw=black, fill=blue, inner sep=.4ex, line width=0.7pt, node on layer=foreground]
\tikzstyle{greendot}=[circle, draw=black, fill=green, inner sep=.4ex, line width=0.7pt, node on layer=foreground]
\tikzstyle{lsdot}=[circle, draw=black, fill=white, inner sep=.4ex, line width=0.7pt, node on layer=foreground, path picture={\draw (path picture bounding box.south east) -- (path picture bounding box.north west) (path picture bounding box.south west) -- (path picture bounding box.north east);}]
\tikzstyle{lssdot}=[circle, draw=black, fill=white, inner sep=.4ex, line width=0.7pt, node on layer=foreground, path picture={\draw (path picture bounding box.south) -- (path picture bounding box.north) (path picture bounding box.west) -- (path picture bounding box.east);}]
\tikzstyle{mixedmorphism}=[morphism, minimum width=30pt, minimum height=16pt, draw, font=\small, inner sep=0pt, fill=white, line width=0.7pt,rounded corners=1ex]
\tikzstyle{thick}=[line width=0.7pt]
\tikzstyle{tiny}=[font=\tiny]
\tikzset{arrow/.style={decoration={
markings,
mark=at position #1 with \arrow{thickarrow}},
postaction=decorate}
}
\tikzset{reverse arrow/.style={decoration={
markings,
mark=at position #1 with \arrow{reversethickarrow}},
postaction=decorate}
}
\newif\ifblack\pgfkeys{/tikz/black/.is if=black}
\newif\ifwedge\pgfkeys{/tikz/wedge/.is if=wedge}
\newif\ifvflip\pgfkeys{/tikz/vflip/.is if=vflip}
\newif\ifhflip\pgfkeys{/tikz/hflip/.is if=hflip}
\newif\ifhvflip\pgfkeys{/tikz/hvflip/.is if=hvflip}
\newif\ifconnectnw\pgfkeys{/tikz/connect nw/.is if=connectnw}
\newif\ifconnectne\pgfkeys{/tikz/connect ne/.is if=connectne}
\newif\ifconnectsw\pgfkeys{/tikz/connect sw/.is if=connectsw}
\newif\ifconnectse\pgfkeys{/tikz/connect se/.is if=connectse}
\newif\ifconnectn\pgfkeys{/tikz/connect n/.is if=connectn}
\newif\ifconnects\pgfkeys{/tikz/connect s/.is if=connects}
\newif\ifconnectnwf\pgfkeys{/tikz/connect nw >/.is if=connectnwf}
\newif\ifconnectnef\pgfkeys{/tikz/connect ne >/.is if=connectnef}
\newif\ifconnectswf\pgfkeys{/tikz/connect sw >/.is if=connectswf}
\newif\ifconnectsef\pgfkeys{/tikz/connect se >/.is if=connectsef}
\newif\ifconnectnf\pgfkeys{/tikz/connect n >/.is if=connectnf}
\newif\ifconnectsf\pgfkeys{/tikz/connect s >/.is if=connectsf}
\newif\ifconnectnwr\pgfkeys{/tikz/connect nw </.is if=connectnwr}
\newif\ifconnectner\pgfkeys{/tikz/connect ne </.is if=connectner}
\newif\ifconnectswr\pgfkeys{/tikz/connect sw </.is if=connectswr}
\newif\ifconnectser\pgfkeys{/tikz/connect se </.is if=connectser}
\newif\ifconnectnr\pgfkeys{/tikz/connect n </.is if=connectnr}
\newif\ifconnectsr\pgfkeys{/tikz/connect s </.is if=connectsr}
\tikzset{keylengthnw/.initial=\connectheight}
\tikzset{keylengthn/.initial =\connectheight}
\tikzset{keylengthne/.initial=\connectheight}
\tikzset{keylengthsw/.initial=\connectheight}
\tikzset{keylengths/.initial =\connectheight}
\tikzset{keylengthse/.initial=\connectheight}
\tikzset{connect nw length/.style={connect nw=true, keylengthnw={#1}}}
\tikzset{connect n length/.style ={connect n =true, keylengthn ={#1}}}
\tikzset{connect ne length/.style={connect ne=true, keylengthne={#1}}}
\tikzset{connect sw length/.style={connect sw=true, keylengthsw={#1}}}
\tikzset{connect s length/.style ={connect s =true, keylengths ={#1}}}
\tikzset{connect se length/.style={connect se=true, keylengthse={#1}}}
\tikzset{connect nw < length/.style={connect nw <=true, keylengthnw={#1}}}
\tikzset{connect n < length/.style ={connect n <=true, keylengthn ={#1}}}
\tikzset{connect ne < length/.style={connect ne <=true, keylengthne={#1}}}
\tikzset{connect sw < length/.style={connect sw <=true, keylengthnw={#1}}}
\tikzset{connect s < length/.style ={connect s <=true, keylengths ={#1}}}
\tikzset{connect se < length/.style={connect se <=true, keylengthse={#1}}}
\tikzset{connect nw > length/.style={connect nw >=true, keylengthnw={#1}}}
\tikzset{connect n > length/.style ={connect n >=true, keylengthn ={#1}}}
\tikzset{connect ne > length/.style={connect ne >=true, keylengthne={#1}}}
\tikzset{connect sw > length/.style={connect sw >=true, keylengthsw={#1}}}
\tikzset{connect s > length/.style ={connect s >=true, keylengths ={#1}}}
\tikzset{connect se > length/.style={connect se >=true, keylengthse={#1}}}
\newlength\morphismheight
\setlength\morphismheight{0.6cm}
\newlength\minimummorphismwidth
\setlength\minimummorphismwidth{0.6cm}
\newlength\stateheight
\setlength\stateheight{0.6cm}
\newlength\minimumstatewidth
\setlength\minimumstatewidth{0.89cm}
\newlength\connectheight
\setlength\connectheight{0.5cm}
\tikzset{width/.initial=\minimummorphismwidth}
\makeatletter
\pgfarrowsdeclare{thickarrow}{thickarrow}
{
\pgfutil@tempdima=-0.84pt%
\advance\pgfutil@tempdima by-1.3\pgflinewidth%
\pgfutil@tempdimb=-1.7pt%
\advance\pgfutil@tempdimb by.625\pgflinewidth%
\pgfarrowsleftextend{+\pgfutil@tempdima}
\pgfarrowsrightextend{+\pgfutil@tempdimb}
}
{
\pgfmathparse{\pgfgetarrowoptions{thickarrow}}%
\pgfsetlinewidth{1.25 pt}
\pgfutil@tempdima=0.28pt%
\advance\pgfutil@tempdima by.3\pgflinewidth%
\pgfsetlinewidth{0.8\pgflinewidth}
\pgfsetdash{}{+0pt}
\pgfsetroundcap
\pgfsetroundjoin
\pgfpathmoveto{\pgfqpoint{-3\pgfutil@tempdima}{4\pgfutil@tempdima}}
\pgfpathcurveto
{\pgfqpoint{-2.75\pgfutil@tempdima}{2.5\pgfutil@tempdima}}
{\pgfqpoint{0pt}{0.25\pgfutil@tempdima}}
{\pgfqpoint{0.75\pgfutil@tempdima}{0pt}}
\pgfpathcurveto
{\pgfqpoint{0pt}{-0.25\pgfutil@tempdima}}
{\pgfqpoint{-2.75\pgfutil@tempdima}{-2.5\pgfutil@tempdima}}
{\pgfqpoint{-3\pgfutil@tempdima}{-4\pgfutil@tempdima}}
\pgfusepathqstroke
}
\pgfarrowsdeclare{reversethickarrow}{reversethickarrow}
{
\pgfutil@tempdima=-0.84pt%
\advance\pgfutil@tempdima by-1.3\pgflinewidth%
\pgfutil@tempdimb=0.2pt%
\advance\pgfutil@tempdimb by.625\pgflinewidth%
\pgfarrowsleftextend{+\pgfutil@tempdima}
\pgfarrowsrightextend{+\pgfutil@tempdimb}
}
{
\pgftransformxscale{-1}
\pgfmathparse{\pgfgetarrowoptions{thickarrow}}%
\ifpgfmathunitsdeclared%
\pgfmathparse{\pgfmathresult pt}%
\els
\pgfmathparse{\pgfmathresult*\pgflinewidth}%
\fi%
\let0.7pt=\pgfmathresult
\pgfsetlinewidth{1.25 pt}
\pgfutil@tempdima=0.28pt%
\advance\pgfutil@tempdima by.3\pgflinewidth%
\pgfsetlinewidth{0.8\pgflinewidth}
\pgfsetdash{}{+0pt}
\pgfsetroundcap
\pgfsetroundjoin
\pgfpathmoveto{\pgfqpoint{-3\pgfutil@tempdima}{4\pgfutil@tempdima}}
\pgfpathcurveto
{\pgfqpoint{-2.75\pgfutil@tempdima}{2.5\pgfutil@tempdima}}
{\pgfqpoint{0pt}{0.25\pgfutil@tempdima}}
{\pgfqpoint{0.75\pgfutil@tempdima}{0pt}}
\pgfpathcurveto
{\pgfqpoint{0pt}{-0.25\pgfutil@tempdima}}
{\pgfqpoint{-2.75\pgfutil@tempdima}{-2.5\pgfutil@tempdima}}
{\pgfqpoint{-3\pgfutil@tempdima}{-4\pgfutil@tempdima}}
\pgfusepathqstroke
}
\makeatother
\makeatletter
\pgfdeclareshape{ground}
{
\savedanchor\centerpoint
{
\pgf@x=0pt
\pgf@y=0pt
}
\anchor{center}{\centerpoint}
\anchorborder{\centerpoint}
\anchor{north}
{
\pgf@x=0pt
\pgf@y=0.5*0.33*\stateheight
}
\anchor{south}
{
\pgf@x=0pt
\pgf@y=0pt
}
\saveddimen\overallwidth
{
\pgfkeysgetvalue{/pgf/minimum width}{\minwidth}
\pgf@x=\minimumstatewidth
\ifdim\pgf@x<\minwidth
\pgf@x=\minwidth
\fi
}
\backgroundpath
{
\begin{pgfonlayer}{foreground}
\pgfsetstrokecolor{black}
\pgfsetcolor{gray}
\pgfsetlinewidth{1.25pt}
\ifhflip
\pgftransformyscale{-1}
\fi
\pgftransformscale{0.5}
\pgfpathmoveto{\pgfpoint{-0.5*\overallwidth}{0}}
\pgfpathlineto{\pgfpoint{0.5*\overallwidth}{0}}
\pgfpathmoveto{\pgfpoint{-0.33*\overallwidth}{0.33*\stateheight}}
\pgfpathlineto{\pgfpoint{0.33*\overallwidth}{0.33*\stateheight}}
\pgfpathmoveto{\pgfpoint{-0.16*\overallwidth}{0.66*\stateheight}}
\pgfpathlineto{\pgfpoint{0.16*\overallwidth}{0.66*\stateheight}}
\pgfpathmoveto{\pgfpoint{-0.02*\overallwidth}{\stateheight}}
\pgfpathlineto{\pgfpoint{0.02*\overallwidth}{\stateheight}}
\end{pgfonlayer}
}
}
\tikzset{forward arrow style/.style={every to/.style, decoration={
markings,
mark=at position 0.5 with \arrow{thickarrow}},
postaction=decorate}}
\tikzset{reverse arrow style/.style={every to/.style, decoration={
markings,
mark=at position 0.5 with \arrow{reversethickarrow}},
postaction=decorate}}
\pgfdeclareshape{morphism}
{
\savedanchor\centerpoint
{
\pgf@x=0pt
\pgf@y=0pt
}
\anchor{center}{\centerpoint}
\anchorborder{\centerpoint}
\saveddimen\savedlengthnw
{
\pgfkeysgetvalue{/tikz/keylengthnw}{\len}
\pgf@x=\len
}
\saveddimen\savedlengthn
{
\pgfkeysgetvalue{/tikz/keylengthn}{\len}
\pgf@x=\len
}
\saveddimen\savedlengthne
{
\pgfkeysgetvalue{/tikz/keylengthne}{\len}
\pgf@x=\len
}
\saveddimen\savedlengthsw
{
\pgfkeysgetvalue{/tikz/keylengthsw}{\len}
\pgf@x=\len
}
\saveddimen\savedlengths
{
\pgfkeysgetvalue{/tikz/keylengths}{\len}
\pgf@x=\len
}
\saveddimen\savedlengthse
{
\pgfkeysgetvalue{/tikz/keylengthse}{\len}
\pgf@x=\len
}
\saveddimen\overallwidth
{
\pgfkeysgetvalue{/tikz/width}{\minwidth}
\pgf@x=\wd\pgfnodeparttextbox
\ifdim\pgf@x<\minwidth
\pgf@x=\minwidth
\fi
}
\savedanchor{\upperrightcorner}
{
\pgf@y=.5\ht\pgfnodeparttextbox
\advance\pgf@y by -.5\dp\pgfnodeparttextbox
\pgf@x=.5\wd\pgfnodeparttextbox
}
\anchor{north}
{
\pgf@x=0pt
\pgf@y=0.5\morphismheight
}
\anchor{north east}
{
\pgf@x=\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=0.5\morphismheight
}
\anchor{east}
{
\pgf@x=\overallwidth
\divide \pgf@x by 2
\advance \pgf@x by 5pt
\pgf@y=0pt
}
\anchor{west}
{
\pgf@x=-\overallwidth
\divide \pgf@x by 2
\advance \pgf@x by -5pt
\pgf@y=0pt
}
\anchor{north west}
{
\pgf@x=-\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=0.5\morphismheight
}
\anchor{connect nw}
{
\pgf@x=-\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=0.5\morphismheight
\advance\pgf@y by \savedlengthnw
}
\anchor{connect ne}
{
\pgf@x=\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=0.5\morphismheight
\advance\pgf@y by \savedlengthne
}
\anchor{connect sw}
{
\pgf@x=-\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=-0.5\morphismheight
\advance\pgf@y by -\savedlengthsw
}
\anchor{connect se}
{
\pgf@x=\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=-0.5\morphismheight
\advance\pgf@y by -\savedlengthse
}
\anchor{connect n}
{
\pgf@x=0pt
\pgf@y=0.5\morphismheight
\advance\pgf@y by \savedlengthn
}
\anchor{connect s}
{
\pgf@x=0pt
\pgf@y=-0.5\morphismheight
\advance\pgf@y by -\savedlengths
}
\anchor{south east}
{
\pgf@x=\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=-0.5\morphismheight
}
\anchor{south west}
{
\pgf@x=-\overallwidth
\multiply \pgf@x by 2
\divide \pgf@x by 5
\pgf@y=-0.5\morphismheight
}
\anchor{south}
{
\pgf@x=0pt
\pgf@y=-0.5\morphismheight
}
\anchor{text}
{
\upperrightcorner
\pgf@x=-\pgf@x
\pgf@y=-\pgf@y
}
\backgroundpath
{
\pgfsetstrokecolor{black}
\pgfsetlinewidth{0.7pt}
\begin{scope}
\ifhflip
\pgftransformyscale{-1}
\fi
\ifvflip
\pgftransformxscale{-1}
\fi
\ifhvflip
\pgftransformxscale{-1}
\pgftransformyscale{-1}
\fi
\pgfpathmoveto{\pgfpoint
{-0.5*\overallwidth-5pt}
{0.5*\morphismheight}}
\pgfpathlineto{\pgfpoint
{0.5*\overallwidth+5pt}
{0.5*\morphismheight}}
\ifwedge
\pgfpathlineto{\pgfpoint
{0.5*\overallwidth + 15pt}
{-0.5*\morphismheight}}
\else
\pgfpathlineto{\pgfpoint
{0.5*\overallwidth + 5pt}
{-0.5*\morphismheight}}
\fi
\pgfpathlineto{\pgfpoint
{-0.5*\overallwidth-5pt}
{-0.5*\morphismheight}}
\pgfpathclose
\pgfusepath{stroke}
\end{scope}
\ifconnectnw
\pgfpathmoveto{\pgfpoint
{-0.4*\overallwidth}
{0.5*\morphismheight}}
\pgfpathlineto{\pgfpoint
{-0.4*\overallwidth}
{0.5*\morphismheight+\savedlengthnw}}
\pgfusepath{stroke}
\fi
\ifconnectne
\pgfpathmoveto{\pgfpoint
{0.4*\overallwidth}
{0.5*\morphismheight}}
\pgfpathlineto{\pgfpoint
{0.4*\overallwidth}
{0.5*\morphismheight+\savedlengthne}}
\pgfusepath{stroke}
\fi
\ifconnectsw
\pgfpathmoveto{\pgfpoint
{-0.4*\overallwidth}
{-0.5*\morphismheight}}
\pgfpathlineto{\pgfpoint
{-0.4*\overallwidth}
{-0.5*\morphismheight-\savedlengthsw}}
\pgfusepath{stroke}
\fi
\ifconnectse
\pgfpathmoveto{\pgfpoint
{0.4*\overallwidth}
{-0.5*\morphismheight}}
\pgfpathlineto{\pgfpoint
{0.4*\overallwidth}
{-0.5*\morphismheight-\savedlengthse}}
\pgfusepath{stroke}
\fi
\ifconnectn
\pgfpathmoveto{\pgfpoint
{0pt}
{0.5*\morphismheight}}
\pgfpathlineto{\pgfpoint
{0pt}
{0.5*\morphismheight+\savedlengthn}}
\pgfusepath{stroke}
\fi
\ifconnects
\pgfpathmoveto{\pgfpoint
{0pt}
{-0.5*\morphismheight}}
\pgfpathlineto{\pgfpoint
{0pt}
{-0.5*\morphismheight-\savedlengths}}
\pgfusepath{stroke}
\fi
\ifconnectnwf
\draw [forward arrow style] (-0.4*\overallwidth,0.5*\morphismheight)
to (-0.4*\overallwidth,0.5*\morphismheight+\savedlengthnw);
\fi
\ifconnectnef
\draw [forward arrow style] (0.4*\overallwidth,0.5*\morphismheight)
to (0.4*\overallwidth,0.5*\morphismheight+\savedlengthne);
\fi
\ifconnectswf
\draw [forward arrow style] (-0.4*\overallwidth,-0.5*\morphismheight-\savedlengthsw)
to (-0.4*\overallwidth,-0.5*\morphismheight);
\fi
\ifconnectsef
\draw [forward arrow style] (0.4*\overallwidth,-0.5*\morphismheight-\savedlengthse)
to (0.4*\overallwidth,-0.5*\morphismheight);
\fi
\ifconnectnf
\draw [forward arrow style] (0,0.5*\morphismheight)
to (0,0.5*\morphismheight+\savedlengthn);
\fi
\ifconnectsf
\draw [forward arrow style] (0,-0.5*\morphismheight-\savedlengths)
to (0,-0.5*\morphismheight);
\fi
\ifconnectnwr
\draw [reverse arrow style] (-0.4*\overallwidth,0.5*\morphismheight)
to (-0.4*\overallwidth,0.5*\morphismheight+\savedlengthnw);
\fi
\ifconnectner
\draw [reverse arrow style] (0.4*\overallwidth,0.5*\morphismheight)
to (0.4*\overallwidth,0.5*\morphismheight+\savedlengthne);
\fi
\ifconnectswr
\draw [reverse arrow style] (-0.4*\overallwidth,-0.5*\morphismheight-\savedlengthsw)
to (-0.4*\overallwidth,-0.5*\morphismheight);
\fi
\ifconnectser
\draw [reverse arrow style] (0.4*\overallwidth,-0.5*\morphismheight-\savedlengthse)
to (0.4*\overallwidth,-0.5*\morphismheight);
\fi
\ifconnectnr
\draw [reverse arrow style] (0,0.5*\morphismheight)
to (0,0.5*\morphismheight+\savedlengthn);
\fi
\ifconnectsr
\draw [reverse arrow style] (0,-0.5*\morphismheight-\savedlengths)
to (0,-0.5*\morphismheight);
\fi
}
}
\pgfdeclareshape{swish right}
{
\savedanchor\centerpoint
{
\pgf@x=0pt
\pgf@y=0pt
}
\anchor{center}{\centerpoint}
\anchorborder{\centerpoint}
\anchor{north}
{
\pgf@x=\minimummorphismwidth
\divide\pgf@x by 5
\pgf@y=\morphismheight
\divide\pgf@y by 2
\advance\pgf@y by \connectheight
}
\anchor{south}
{
\pgf@x=-\minimummorphismwidth
\divide\pgf@x by 5
\pgf@y=-\morphismheight
\divide\pgf@y by 2
\advance\pgf@y by -\connectheight
}
\backgroundpath
{
\pgfsetstrokecolor{black}
\pgfsetlinewidth{0.7pt}
\pgfpathmoveto{\pgfpoint
{-0.2*\minimummorphismwidth}
{-0.5*\morphismheight-\connectheight}}
\pgfpathcurveto
{\pgfpoint{-0.2*\minimummorphismwidth}{0pt}}
{\pgfpoint{0.2*\minimummorphismwidth}{0pt}}
{\pgfpoint
{0.2*\minimummorphismwidth}
{0.5*\morphismheight+\connectheight}}
\pgfusepath{stroke}
}
}
\pgfdeclareshape{swish left}
{
\savedanchor\centerpoint
{
\pgf@x=0pt
\pgf@y=0pt
}
\anchor{center}{\centerpoint}
\anchorborder{\centerpoint}
\anchor{north}
{
\pgf@x=-\minimummorphismwidth
\divide\pgf@x by 5
\pgf@y=\morphismheight
\divide\pgf@y by 2
\advance\pgf@y by \connectheight
}
\anchor{south}
{
\pgf@x=\minimummorphismwidth
\divide\pgf@x by 5
\pgf@y=-\morphismheight
\divide\pgf@y by 2
\advance\pgf@y by -\connectheight
}
\backgroundpath
{
\pgfsetstrokecolor{black}
\pgfsetlinewidth{0.7pt}
\pgfpathmoveto{\pgfpoint
{0.2*\minimummorphismwidth}
{-0.5*\morphismheight-\connectheight}}
\pgfpathcurveto
{\pgfpoint{0.2*\minimummorphismwidth}{0pt}}
{\pgfpoint{-0.2*\minimummorphismwidth}{0pt}}
{\pgfpoint
{-0.2*\minimummorphismwidth}
{0.5*\morphismheight+\connectheight}}
\pgfusepath{stroke}
}
}
\pgfdeclareshape{state}
{
\savedanchor\centerpoint
{
\pgf@x=0pt
\pgf@y=0pt
}
\anchor{center}{\centerpoint}
\anchorborder{\centerpoint}
\saveddimen\overallwidth
{
\pgf@x=3\wd\pgfnodeparttextbox
\ifdim\pgf@x<\minimumstatewidth
\pgf@x=\minimumstatewidth
\fi
}
\savedanchor{\upperrightcorner}
{
\pgf@x=.5\wd\pgfnodeparttextbox
\pgf@y=.5\ht\pgfnodeparttextbox
\advance\pgf@y by -.5\dp\pgfnodeparttextbox
}
\anchor{A}
{
\pgf@x=-\overallwidth
\divide\pgf@x by 4
\pgf@y=0pt
}
\anchor{B}
{
\pgf@x=\overallwidth
\divide\pgf@x by 4
\pgf@y=0pt
}
\anchor{text}
{
\upperrightcorner
\pgf@x=-\pgf@x
\ifhflip
\pgf@y=-\pgf@y
\advance\pgf@y by 0.4\stateheight
\else
\pgf@y=-\pgf@y
\advance\pgf@y by -0.4\stateheight
\fi
}
\backgroundpath
{
\pgfsetstrokecolor{black}
\pgfsetlinewidth{0.7pt}
\pgfpathmoveto{\pgfpoint{-0.5*\overallwidth}{0}}
\pgfpathlineto{\pgfpoint{0.5*\overallwidth}{0}}
\ifhflip
\pgfpathlineto{\pgfpoint{0}{\stateheight}}
\else
\pgfpathlineto{\pgfpoint{0}{-\stateheight}}
\fi
\pgfpathclose
\ifblack
\pgfsetfillcolor{black!35}
\pgfusepath{fill,stroke}
\else
\pgfusepath{stroke}
\fi
}
}
\makeatother
\newcommand{\tinycomult}[1][dot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node[#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-3pt}}}}
\newcommand{\tinycomultls}[1][whitedot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node[ls,scale=0.9,#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-3pt}}}}
\newcommand{\tinycomultagg}[1][whitedot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node[agg,scale=0.9,#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-3pt}}}}
\newcommand{\tinycounit}[1][dot]{
\smash{\raisebox{-2pt}{\ensuremath{\hspace{-3pt}\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node (1) at (0,1) {};
\node[#1, inner sep=1.5pt] (d) at (0,0.55) {};
\draw (0.center) to (d.center);
\end{pic}
\hspace{-1pt}}}}}
\newcommand{\tinycounitls}[1][whitedot]{
\smash{\raisebox{-2pt}{\ensuremath{\hspace{-3pt}\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node (1) at (0,1) {};
\node[ls,scale=0.9,#1, inner sep=1.5pt] (d) at (0,0.55) {};
\draw (0.center) to (d.center);
\end{pic}
\hspace{-1pt}}}}}
\newcommand{\tinymult}[1][dot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node[#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-3pt}}}}
\newcommand{\tinyleftdivide}[1][dot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node[#1,scale=0.9, inner sep=1.5pt] (1) at (0,0.55) {};
\node[blackdot,scale=0.5] (2) at (-0.7,1) {};
\node (4) at (-1.3,0){};
\node (3) at (0.5,1.2) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=right, out looseness=1] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\draw[out=up,in=left,looseness=1.5] (4.center) to (2.center);
\end{pic}
}\hspace{-1pt}}}}
\newcommand{\tinyspider}[1][dot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string,yscale=1]
\node (0) at (0,0) {};
\node[#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\node (4) at (-0.5,0) {};
\node (5) at (0.5,0) {};
\draw (4.center) to [out=up, in=left, out looseness=1.5](1.center);
\draw (5.center) to [out=up, in=right, out looseness=1.5](1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\draw[densely dotted] (-0.365,0.1) to (0.5,0.1){};
\draw[densely dotted] (-0.365,0.9) to (0.5,0.9){};
\end{pic}
}\hspace{-3pt}}}}
\newcommand{\tinymultk}[1][whitedot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node[#1, inner sep=1.5pt,label={[xshift=-0.1cm,yshift=-0.1cm]0:{\tiny$k$}}] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-3pt}}}}
\newcommand{\tinymultls}[1][whitedot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node[ls,scale=0.9,#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-1pt}}}}
\newcommand{\tinymultagg}[1][whitedot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node[agg,scale=0.9,#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-1pt}}}}
\newcommand{\tinyunit}[1][dot]{
\smash{\raisebox{-2pt}{\ensuremath{\hspace{-3pt}\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node (1) at (0,1) {};
\node[#1, inner sep=1.5pt] (d) at (0,0.55) {};
\draw (0.center) to (d.center);
\end{pic}
\hspace{-1pt}}}}}
\newcommand{\tinyunitls}[1][whitedot]{
\smash{\raisebox{-2pt}{\ensuremath{\hspace{-3pt}\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node (1) at (0,1) {};
\node[ls,scale=0.9,#1, inner sep=1.5pt] (d) at (0,0.55) {};
\draw (0.center) to (d.center);
\end{pic}
\hspace{-1pt}}}}}
\newcommand{\tinyunitagg}[1][whitedot]{
\smash{\raisebox{-2pt}{\ensuremath{\hspace{-3pt}\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node (1) at (0,1) {};
\node[agg,scale=0.9,#1, inner sep=1.5pt] (d) at (0,0.55) {};
\draw (0.center) to (d.center);
\end{pic}
\hspace{-1pt}}}}}
\newcommand{\tinyid}{\raisebox{-2pt}{\ensuremath{\hspace{-4pt}\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node (1) at (0,1) {};
\draw[string] (0.center) to (1.center);
\end{pic}
\hspace{-2pt}}}}
\newcommand{\tinyhandle}[1][dot]{\raisebox{-2pt}{\ensuremath{\hspace{-3pt}\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node[dot, inner sep=1.0pt] (1) at (0,0.3) {};
\node[dot, inner sep=1.0pt] (2) at (0,0.7) {};
\node (3) at (0,1) {};
\draw (0.center) to (1.center);
\draw (2.center) to (3.center);
\draw[in=180, out=180, looseness=2] (1.center) to (2.center);
\draw[in=0, out=0, looseness=2] (1.center) to (2.center);
\end{pic}\hspace{-1pt}}}}
\newcommand{\tinypants}{\smash{\raisebox{-2pt}{\hspace{-2pt}\ensuremath{\begin{pic}[scale=0.2,string]
\draw (0,0) to [out=up, in=down] (1,2);
\draw (1,0) to [out=up, in=up, looseness=2] (2,0);
\draw (3,0) to[out=up, in=down] (2,2);
\end{pic}}}}}
\newcommand{\tinycup}{\smash{\raisebox{0pt}{\hspace{-2pt}\ensuremath{\begin{pic}[scale=0.2,string]
\draw (0,0) to[out=-90,in=-90,looseness=1.5] (1.5,0);
\end{pic}}}}}
\newcommand{\tinycomultdb}[1][dot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node[#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\node (0') at (0.75,0) {};
\node[#1, inner sep=1.5pt] (1') at (0.75,0.55) {};
\node (2') at (0.25,1) {};
\node (3') at (1.25,1) {};
\draw (0'.center) to (1'.center);
\draw (1'.center) to [out=left, in=down, out looseness=1.5] (2'.center);
\draw (1'.center) to [out=right, in=down, out looseness=1.5] (3'.center);
\end{pic}
}\hspace{-3pt}}}}
\newcommand{\tinymultlss}[1][whitedot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string,yscale=-1]
\node (0) at (0,0) {};
\node[ls',scale=0.65,#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-1pt}}}}
\newcommand{\tinycomultlss}[1][whitedot]{
\smash{\raisebox{-2pt}{\hspace{-5pt}\ensuremath{\begin{pic}[scale=0.4,string]
\node (0) at (0,0) {};
\node[ls',scale=0.65,#1, inner sep=1.5pt] (1) at (0,0.55) {};
\node (2) at (-0.5,1) {};
\node (3) at (0.5,1) {};
\draw (0.center) to (1.center);
\draw (1.center) to [out=left, in=down, out looseness=1.5] (2.center);
\draw (1.center) to [out=right, in=down, out looseness=1.5] (3.center);
\end{pic}
}\hspace{-3pt}}}}
\renewcommand\dag{\ensuremath{\dagger}}
\def{\textstyle \bigotimes}{{\textstyle \bigotimes}}
\section{#1}}
\newcommand\id[1][]{\ensuremath{\mathrm{id}_{#1}}}
\def{\textstyle \bigotimes}{{\textstyle \bigotimes}}
\newcommand\sxleftarrow[1]{\xleftarrow{\smash{#1}}}
\newcommand\sxrightarrow[1]{\xrightarrow{\smash{#1}}}
\newcommand\xto[1]{\xrightarrow{#1}}
\newcommand\sxto[1]{\sxrightarrow{#1}}
\renewcommand{\-}[0]{\nobreakdash-\hspace{0pt}}
\DeclareMathOperator{\Ob}{Ob}
\newcommand\C{\ensuremath{\mathbb{C}}}
\newcommand{\inprod}[2]{\ensuremath{\langle #1\hspace{0.5pt}|\hspace{0.5pt}#2 \rangle}}
\newcommand{\defined}{\ensuremath{\mathop{\downarrow}}}
\newcommand\Hom{\ensuremath{\textrm{Hom}}}
\newcommand\pdag{{\phantom{\dagger}}}
\DeclareMathOperator{\CP}{CP}
\DeclareMathOperator{\CPq}{CP_q}
\DeclareMathOperator{\CPc}{CP_c}
\DeclareMathOperator{\CPstar}{CP^*}
\newcommand{\proj}{\ensuremath{p}}
\newcommand{\inj}{\ensuremath{i}}
\DeclareMathOperator{\Prob}{Prob}
\newcommand{\ie}{\textit{i.e.}\xspace}
\newcommand{\eg}{\textit{e.g.}\xspace}
\newcommand\ud{\ensuremath{\mathrm{d}}}
\usepackage{color}
\newcommand{\todo}[1]{{\color{red}#1}}
\def\arraystretch{1.0}
\newcommand\grid[1]{\ensuremath{\def\arraystretch{1.4}\begin{array}{|c|c|c|c|c|c|c|c|c|}\hline#1\\\hline\end{array}}}
\newcommand\diag{\mathrm{diag}}
\newcommand\inv{{-1}}
\newcommand\I{\ensuremath{\mathbb I}}
\newcommand\M{\ensuremath{\mathcal M}}
\newcommand\F{\ensuremath{\mathcal F}}
\newcommand\super[2]{\stackrel{\makebox[0pt]{\smash{\tiny #1}}}{#2}}
\newcommand\T{\ensuremath{\mathrm T}}
\newcommand\bra[1]{\langle #1|}
\newcommand\ket[1]{{|} #1 \rangle}
\newcommand\braket[2]{\langle #1 | #2 \rangle}
\newcommand\ketbra[2]{|#1 \rangle \hspace{-1pt} \langle #2 |}
\setlength\marginparwidth{0.9in}
\setlength\marginparsep{0.05in}
\newcounter{jamiecomment}
\setcounter{jamiecomment}{1}
\newcommand\JVcomm[1]{\ensuremath{{}^{\color{red}\thejamiecomment}}\marginpar{\color{red}\tiny\raggedright \thejamiecomment: #1}\stepcounter{jamiecomment}}
\usepackage{amsbsy}
\usepackage[normalem]{ulem}
\allowdisplaybreaks[1]
\input{book-header}
\def\titlerunning{Constructing Mutually Unbiased Bases from Quantum Latin Squares}
\def\authorrunning{Benjamin Musto}
\begin{document}
\title{Constructing Mutually Unbiased Bases \\ from Quantum Latin Squares}
\author{Benjamin Musto
\institute{Department of Computer Science\\ University of Oxford}
\email{<EMAIL>}
}
\date{\today}
\maketitle
\begin{abstract}
We introduce \textit{orthogonal quantum Latin squares}, which restrict to traditional orthogonal Latin squares, and investigate their application in quantum information science. We use quantum Latin squares to build maximally entangled bases, and show how mutually unbiased maximally entangled bases can be constructed in square dimension from orthogonal quantum Latin squares. We also compare our construction to an existing construction due to Beth and Wocjan~\cite{ortho1} and show that ours is strictly more general. \end{abstract}
\section{Introduction}
In this paper we introduce a notion of orthogonality between \textit{quantum Latin squares} (QLSs)~\cite{mypaper1}, mathematical objects which generalise \textit{Latin squares}. We use this concept to give a new construction of \textit{maximally entangled mutually unbiased bases} (MUBs), extending existing known techniques for Latin squares~\cite{ortho1,ortho2}. In addition we prove that our construction can produce bases that are unobtainable by existing methods~\cite{ortho1,ortho2}.
We also introduce the concept of \textit{mutually weak orthogonal quantum Latin squares }(MOQLS) which generalise \textit{mutually orthogonal Latin squares} (MOLS), about which a significant body of research exists in connection with quantum information, and particularly pertaining to the connection between MOLS and MUBs~\cite{klapp2,paterek,cao2016more}.
Mutually unbiased bases are of fundamental importance to quantum information, as they capture the physical notion of complementary observables, quantities that cannot be simultaneously measured. Entanglement is one of the central phenomena of quantum theory that is at the foundation of quantum information and computation.
The results presented in this paper were developed using the graphical calculus of categorical quantum mechanics (CQM), and we have made use of it where we believe it elucidates some detail. For those unfamiliar with CQM, there is a short introduction of the concepts necessary to understand this paper in Appendix~\ref{apx:cqm}; for a thorough introduction please refer to~\cite{qcs, abramskycoecke2004, surveycategoricalquantummechanics}. Everything that we present here is in the category $\cat{FHilb}$ of finite Hilbert spaces and linear maps, but could be interpreted in any monoidal category such as $\cat{Rel}$ with \textit{quantum-like} properties, which have been extensively researched as quantum toy theories.
We start with a definition of quantum Latin squares.
\begin{definition}
\label{def:qls}
A \textit{quantum Latin square of order $n$} is an $n \times n$ array of elements of the Hilbert space $\C^n$, such that every row and every column is an orthonormal basis.
\end{definition}
\begin{example}\label{ex:qls}
Here is an example of a quantum Latin square given in terms of the computational basis states $\ket i$ for $i\in \{0,...,9\}$, and the following states:
\begin{minipage}[t]{6.2cm}
\begin{align}\label{eq:a}
\ket a&:=\frac{1}{\sqrt{3}}(\ket 3+ \ket 4 + i\ket 5)\\ \label{eq:b}
\ket b&:=\frac{1}{\sqrt{6}}(2\ket 3- \ket 4 + i\ket 5)\\ \label{eq:c}
\ket c&:=\frac{1}{\sqrt{14}}(-2i\ket 3-i\ket 4 + 3\ket 5)
\end{align}
\end{minipage}
\begin{minipage}[t]{7cm}
\begin{align}\label{eq:abg}
\ket \alpha &:=\frac{1}{\sqrt{3}}(\ket 0+ \ket 1 + \ket 2)
\\ \label{eq:abg2}
\ket \beta &:=\frac{1}{\sqrt{3}}(\ket 0 +e^{\frac{2 \pi i}{3}} \ket 1 + e^{\frac{-2 \pi i}{3}}\ket 2 )\\ \label{eq:abg3}
\ket \gamma &:=\frac{1}{\sqrt{3}}(\ket 0 +e^{\frac{-2 \pi i}{3}} \ket 1 + e^{\frac{2 \pi i}{3}}\ket 2 )
\end{align}
\end{minipage}
\hspace{-67pt}\begin{equation*}\grid{
\ket{0} & \ket{2} & \ket{1} & \ket{3} & \ket{5} & \ket{4} & \ket {6} & \ket {8} & \ket 7
\\ \hline
\ket 2
& \ket{1}
& \ket{0}
& \ket 5 & \ket 4 & \ket 3 & \ket 8 & \ket 7 & \ket 6
\\ \hline
\ket{1}
& \ket{0}
& \ket{2}
& \ket{4} & \ket{3} & \ket 5 & \ket 7 & \ket 6 & \ket 8
\\ \hline
\ket{6} & \ket{8} & \ket{7} & \ket{0} & \ket 2 & \ket 1 & \ket 3 & \ket 5 & \ket 4
\\ \hline
\ket 8 & \ket 7 & \ket 6 & \ket 2 & \ket 1 & \ket 0 & \ket 5 & \ket 4 & \ket 3 \\ \hline
\ket 7 & \ket 6 & \ket 8 & \ket 1 & \ket 0 & \ket 2 & \ket 4 & \ket 3 & \ket 5 \\ \hline
\ket a & \ket c & \ket b & \ket 6 & \ket 8 & \ket 7 & \ket \alpha & \ket \gamma & \ket \beta \\ \hline
\ket c & \ket b&\ket a &\ket 8&\ket 7&\ket 6&\ket \gamma&\ket \beta&\ket \alpha \\ \hline
\ket b & \ket a&\ket c &\ket 7&\ket 6&\ket 8&\ket \beta&\ket \alpha&\ket \gamma}
\end{equation*}
\end{example}
\noindent It can be checked that every row and every column is an orthonormal basis.
\begin{definition}[Latin square] A \textit{Latin square} is a QLS with entries that all come from the computational basis. For those who are familiar with the traditional definition, it is recovered by mapping each computational basis state to a different symbol.\end{definition}
The main result of this paper is a construction of mutually unbiased maximally entangled bases from orthogonal QLSs. We now define the necessary concepts.
\begin{definition}[Mutually unbiased bases]
Two orthonormal bases $\ket {a_i}$ and $\ket {b_j}$ for a Hilbert space $\mathcal H $ of dimension $n$ are \textit{mutually unbiased} when, for all $i,j \,\in \, \, \{0 ,...,n-1\}$~\cite{bandyopadhyay}:
\begin{equation}\label{def:mub}
|\inprod{a_i}{b_j}|^2 \, = \, \frac{1}{n}
\end{equation}
\end{definition}
\begin{definition}[Maximally entangled state]
A \textit{maximally entangled state} of a bipartite system is a state $\ket \psi$ of a product Hilbert space $\mathcal H_A \otimes \mathcal H_B$ with dim($\mathcal H_B)=n,$ such that the partial trace over one of the systems of its density operator $\rho_{AB}=\ketbra{\psi}{\psi}$ is proportional to the identity. i.e~\cite{medefinition}.
\begin{equation}
\rho_A :=\sum_{k=0}(\id[A] \otimes \bra k) \rho_{AB} (\id[A] \otimes \ket{k})=\frac{1}{n}\id[A \otimes B]
\end{equation}
\end{definition}
\begin{remark}For the Hilbert space $\mathcal H \otimes \mathcal H$ with dim$(\mathcal H)=n$, all maximally entangled states are of the following form, where $U$ is a unitary linear map and $\tinyspider$ is the classical structure (see Appendix~\ref{apx:cqm}) corresponding to the orthonormal basis $\ket k$~\cite{vedral}:
\begin{equation}\label{mes}\ket U :=\frac{1}{\sqrt{n}}\sum_{k=0}^{n-1}\ket k \otimes U\ket k \qquad \text{ or equivalently }\qquad \ket U :=\frac{1}{\sqrt{n}}\begin{pic}
\node[blackdot] (b) {};
\node (1) at (-0.5,1) {};
\node (2) at (0.5,1) {};
\node[morphism,scale=0.7] (U) at (0.5,0.5) {$U$};
\draw[string,in=left,out=down,looseness=1.3] (1.center) to (b.center);
\draw[string,in=up,out=down,looseness=1.3] (2.center) to (U.north);
\draw[string,in=down,out=right,looseness=1.1] (b.center) to (U.south);
\end{pic}\end{equation}
\end{remark}
\begin{definition}[Maximally entangled basis]\label{def:meb}
A \textit{maximally entangled basis} (MEB) for a bipartite system represented by a tensor product Hilbert space $\mathcal H \otimes \mathcal H$, is an orthonormal basis such that each basis state is maximally entangled. \end{definition}
Two MEBs
$\mathcal A:=\ket{A_i}$ and $\mathcal B:=\ket{B_i}$ are equivalent when there exist unitaries $U$ and $V$ and complex numbers of modulus $1$, $c_i$ such that:
\begin{equation}\label{eq:eqimeb}
\begin{pic}
\node[state,xscale=1.5] {$A_i$};
\node(1) at (-0.5,1) {};
\node(2) at (0.5,1) {};
\draw (1.center) to +(0,-1);
\draw (2.center) to +(0,-1);
\end{pic}
\quad = \quad
c_i
\begin{pic}
\node[state,xscale=1.5] {$B_i$};
\node(1)[morphism,scale=0.7] at (-0.5,0.5){$U$};
\node(2)[morphism,scale=0.7] at (0.5,0.5){$V$};
\node at (-0.5,1) {};
\node at (0.5,1) {};
\draw (-0.5,0) to (1.south);
\draw (0.5,0) to (2.south);
\draw (-0.5,1) to (1.north);
\draw (0.5,1) to (2.north);
\end{pic}
\end{equation}
In Section~\ref{sec:main} we introduce our main result, the most general construction of mutually unbiased bases of the three presented in this paper. We introduce orthogonal quantum Latin squares and show how they can be used to construct MUBs, and we construct an explicit example. In Section~\ref{sec:left} we start with traditional orthogonality of Latin squares and then show that the definition of orthogonality for QLSs in Section~\ref{sec:main} generalises it. In Section~\ref{sec:bw} we present Beth and Wocjan's construction for MUBs in square dimension, and show that ours is strictly more general. In Section~\ref{sec:dual} we explain the correspondence between unitary error bases and maximally entangled bases and introduce mutually unbiased error bases. Finally in Section~\ref{sec:moqls} we introduce mutually weak orthogonal quantum Latin squares, which generalise mutually orthogonal Latin squares.
\subsection*{Acknowledgements} The author is grateful to Dominic Verdon and Jamie Vicary for useful discussions, and to EPSRC for financial support.
\section{New construction for square dimension MUBs}\label{sec:main}
In this section we introduce the main result of this paper, a new construction for mutually unbiased maximally entangled bases. In order to formulate our construction we introduce \textit{weak orthogonal quantum Latin squares} which, as we will show in Section~\ref{sec:left}, reduce to traditional orthogonal Latin squares. It will be useful to introduce some notation for quantum Latin squares. Given a QLS $\mathcal Q$ we will denote the vector appearing in the $i^{th}$ column of the $j^{th}$ row as $\ket{Q_{ij}}$.
Before the main result it will be requisite to define generalised Hadamards.
\begin{definition}[Hadamard, see \cite{hadamard}, Definition~2.1]
A \textit{Hadamard matrix of order $n$} is an $n \times n$ matrix $H$ with the following properties for all $i,j$, which we write in both matrix and index form:
\begin{align}
\label{had1}
|H_{ij}| &=1
&
H_{ij} ^\pdag H_{ij} ^* &= 1
\\
\label{had2}
H \circ H^\dag &= n \,\mathbb{I}_n
& \textstyle \sum_p H_{ip} ^\pdag H^* _{jp} &= n \, \delta _{ij}
\\
\label{had3}
H ^\dag \circ H &= n \,\mathbb{I}_n
& \textstyle \sum_p H ^*_{pi} H _{pj} ^\pdag &= n \, \delta _{ij}
\end{align}
\end{definition}
We now introduce a method for constructing MEBs given as input a family of Hadamards and a quantum Latin square. This construction is in fact dual to the quantum shift-and-multiply method for constructing unitary error bases~\cite{mypaper1}, as we will explain in Section~\ref{sec:dual}.
\begin{definition}[Quantum Latin square maximally entangled basis]\label{def:qlsmeb}
Given a quantum Latin square $\mathcal Q$ and a family of Hadamards $H_j$, a \textit{quantum Latin square maximally entangled basis} $B(\mathcal Q, H_j)$ is defined as follows:
\begin{equation}\label{eq:qlsmeb}
\mathcal A := \left\{ A_{ij} = \frac{1}{\sqrt{n}}\sum_{k=0}^{n-1}\ket k \otimes \ketbra{Q_{kj}}{k}H_j\ket i \text{ such that } i,j \in \{ 0,..,n-1\} \right\}
\end{equation}
\end{definition}
\begin{lemma}\label{lem:qlsmeb}
Quantum Latin square maximally entangled bases are maximally entangled bases.
\end{lemma}
\begin{proof}
This MEB construction is the dual of the quantum shift-and-multiply basis construction, for a proof of the correctness of that construction see~\cite[Theorem 19]{mypaper1}.
\end{proof}\begin{definition}[Weak orthogonal quantum Latin squares]\label{def:woqls} Given a pair of QLSs $\mathcal P$ and $\mathcal Q$ with vector entries $\ket{P_{ij}}$ and $\ket{Q_{ij}}$ respectively, they are \textit{weak orthogonal} when for all $i,j \in \{0,..,n-1\}$, there exists unique $t \in \{0,...,n-1 \}$ such that:
\begin{equation}\label{eq:woqls}
\sum_{k=0}^{n-1}\ket k \braket{Q_{ki}}{P_{kj}}=\ket t
\end{equation}
In words: if we take any row from $\mathcal P$ and any row from $\mathcal Q$ and compute the componentwise inner product of their vector entries, the resulting $n$ numbers will always be $n-1$ zeros and a single $1$. If the $1$ appears in the $t^{th}$ column then the output state of the linear map above will be $\ket t.$
\end{definition}
\begin{example}\label{ex:orthoqls}
We present a pair of $9 \times 9$ weak orthogonal quantum Latin squares, the first is the QLS from Example~\ref{ex:qls}. Again let $\ket i$, $i\in \{0,...,9\}$ be the computational basis states and define the states $\ket a ,\ket b ,\ket c ,\ket \alpha,\ket \beta$ and $\ket \gamma$ as in Equations~\eqref{eq:a}~\eqref{eq:b}~\eqref{eq:c}~\eqref{eq:abg}~\eqref{eq:abg2} and~\eqref{eq:abg3}. We define the following pair of QLSs:
\begin{align}
\mathcal P &:= \small\arraycolsep=3pt \grid{
\ket{0} & \ket{2} & \ket{1} & \ket{3} & \ket{5} & \ket{4} & \ket {6} & \ket {8} & \ket 7
\\ \hline
\ket 2
& \ket{1}
& \ket{0}
& \ket 5 & \ket 4 & \ket 3 & \ket 8 & \ket 7 & \ket 6
\\ \hline
\ket{1}
& \ket{0}
& \ket{2}
& \ket{4} & \ket{3} & \ket 5 & \ket 7 & \ket 6 & \ket 8
\\ \hline
\ket{6} & \ket{8} & \ket{7} & \ket{0} & \ket 2 & \ket 1 & \ket 3 & \ket 5 & \ket 4
\\ \hline
\ket 8 & \ket 7 & \ket 6 & \ket 2 & \ket 1 & \ket 0 & \ket 5 & \ket 4 & \ket 3 \\ \hline
\ket 7 & \ket 6 & \ket 8 & \ket 1 & \ket 0 & \ket 2 & \ket 4 & \ket 3 & \ket 5 \\ \hline
\ket a & \ket c & \ket b & \ket 6 & \ket 8 & \ket 7 & \ket \alpha & \ket \gamma & \ket \beta \\ \hline
\ket c & \ket b&\ket a &\ket 8&\ket 7&\ket 6&\ket \gamma&\ket \beta&\ket \alpha \\ \hline
\ket b & \ket a&\ket c &\ket 7&\ket 6&\ket 8&\ket \beta&\ket \alpha&\ket \gamma}
&
\mathcal Q &:= \small\arraycolsep=3pt\grid{
\ket{0} & \ket{1} & \ket{2} & \ket{6} & \ket{7} & \ket{8} & \ket {3} & \ket {4} & \ket 5
\\ \hline
\ket 2
& \ket{0}
& \ket{1}
& \ket 8 & \ket 6 & \ket 7 & \ket 5 & \ket 3 & \ket 4
\\ \hline
\ket{1}
& \ket{2}
& \ket{0}
& \ket{7} & \ket{8} & \ket 6 & \ket 4 & \ket 5 & \ket 3
\\ \hline
\ket{a} & \ket{b} & \ket{c} & \ket{0} & \ket 1 & \ket 2 & \ket 6 & \ket 7 & \ket 8
\\ \hline
\ket c & \ket a & \ket b & \ket 2 & \ket 0 & \ket 1 & \ket 8 & \ket 6 & \ket 7 \\ \hline
\ket b & \ket c & \ket a & \ket 1 & \ket 2 & \ket 0 & \ket 7 & \ket 8 & \ket 6 \\ \hline
\ket 6 & \ket 7 & \ket 8 & \ket 3 & \ket 4 & \ket 5 & \ket \alpha & \ket \beta & \ket \gamma \\ \hline
\ket 8 & \ket 6&\ket 7 &\ket 5&\ket 3&\ket 4&\ket \gamma&\ket \alpha&\ket \beta \\ \hline
\ket 7 & \ket 8&\ket 6 &\ket 4&\ket 5&\ket 3&\ket \beta&\ket \gamma&\ket \alpha}
\end{align}
It can be checked that if we take any row from $\mathcal P$ and any row from $\mathcal Q$ and compute the componentwise inner product of their vector entries, the resulting $n$ numbers will always be $n-1$ zeros and a single $1$.
\end{example}
\begin{theorem}\label{thm:main}
Given two indexed families of $n$ Hadamards $H_k$ and $G_j$both of size $n \times n$, and a pair of $n \times n$ weak orthogonal quantum Latin squares $\mathcal P$ and $\mathcal Q$, the bases $B(\mathcal Q, H_k)$ and $B(\mathcal P, G_j)$ are mutually unbiased.
\end{theorem}
\begin{proof}
See Appendix~\ref{apxprf}.
\end{proof}
\begin{example}\label{ex:qlsmub}
Given as input $\mathcal P$ and $\mathcal Q$ from Example~\ref{ex:orthoqls} and the Hadamard ${H=H_0=H_1=...=H_{n-1}}=G_0=...=G_{n-1}$ defined below with $\omega:=e^{2\pi i/3}$ we have constructed a pair of maximally entangled mutually unbiased bases $\mathcal A$ and $\mathcal B$ for the Hilbert space $\C^9 \otimes \C^9$.
\begin{equation}
H:=
\begin{pmatrix}
1 & 1 & 1 & 1 & 1 &1&1&1&1\\
1 & \omega & \omega^2 & 1 & \omega & \omega^2&1&\omega&\omega^2\\
1&\omega^2 &\omega&1&\omega^2 &\omega&1&\omega^2 &\omega\\
1&1&1&\omega&\omega&\omega&\omega^2&\omega^2&\omega^2\\
1&1&1&\omega&\omega&\omega&\omega^2&\omega^2&\omega^2\\
1&\omega^2 &\omega&\omega&1 &\omega^2&\omega^2 &\omega&1\\
1&1&1&\omega^2 &\omega^2 &\omega^2 &\omega&\omega&\omega\\
1&\omega&\omega^2&\omega^2&1&\omega&\omega&\omega^2&1\\
1&\omega^2 &\omega&\omega^2 &\omega&1&\omega&1 &\omega^2
\end{pmatrix}
\end{equation}
A sample of the $162$ basis states of $\mathcal A$ and $\mathcal B$ with some calculations showing mutual unbiasedness (see Definition~\ref{def:mub}) can be found in Appendix~\ref{apx:6561}. We have performed inner product calculations for all $6561$ combinations of states from $\mathcal A$ and $\mathcal B$ and can confirm that they are mutually unbiased.
\end{example}
\section{Weak orthogonality and Latin square conjugates}\label{sec:left}
In this section we explain how weak orthogonality for QLSs restricts to orthogonality for Latin squares, and why this is the natural generalisation of orthogonality for QLSs. We start with the traditional definition of orthogonality.\begin{definition}[Orthogonal Latin squares]\label{def:ols}Given a pair of Latin squares $A$ and $B$ of equal size, we take each computational basis state from $A$ and form the ordered pair with the state from $B$ corresponding to the same position in the grid. $A$ and $B$ are \textit{orthogonal} when this procedure gives us all possible pairs of computational basis states~\cite{mann}.
\end{definition}
\noindent This definition does not lend itself to generalisation to QLSs since we may now have more than $n^2$ possible ordered pairs, but we can take an alternative approach. We characterise orthogonality in the following way:
\begin{lemma}
Latin squares $A$ and $B$ are orthogonal if and only if the following linear map $P$ is a permutation of basis states:
\begin{equation}\label{eq:orthols}
P:=\sum_{i=0}^{n-1}\sum_{j=0}^{n-1}\sum_{k=0}^{n-1}\ket i \ket j \bra{A_{ij}} \braket{k}{B_{ij}}\bra k
\end{equation}
\end{lemma}
\begin{proof}
We now rearrange the equation defining the linear map $P$:
\begin{align*}P&:=\sum_{i}\sum_{j}\sum_{k}\ket i \ket j \bra{A_{ij}} \braket{k}{B_{ij}}\bra k\\&=\sum_{i}\sum_{j}\sum_{k}\ket i \ket j \bra{A_{ij}} \braket{B_{ij}}{k}\bra k \\ &=\sum_{i=0}^{n-1}\sum_{j=0}^{n-1}\ket i \ket j \bra{A_{ij}} \bra{B_{ij}}\sum_k \ket{k}\bra k\\ &=\sum_{i=0}^{n-1}\sum_{j=0}^{n-1}\ket i \ket j \bra{A_{ij}} \bra{B_{ij}}
\end{align*}
The second equality above holds because all $\ket{B_{ij}}$ and $\ket k$ are real valued vectors, and so $\braket{k}{B_{ij}}=\overline{\braket{k}{B_{ij}}}=\braket{B_{ij}}{k}$. The third equality is just a rearranging of terms. The last equality holds by virtue of $\sum_k \ketbra{k}{k}$ being the resolution of the identity. The linear map $P$ takes in the state $\ket p \ket q$ and outputs a superposition of all the states $\ket i \ket j$ such that $\ket{A_{ij}}=\ket p$ \textit{and} $\ket{B_{ij}}=\ket q$, or outputs $0$ if no such $i,j$ exist. $P$ is a permutation if and only if for all inputs $p,q$ there exists unique $i,j$ such that $\ket{A_{ij}}=\ket p$ \textit{and} $\ket{B_{ij}}=\ket q$, i.e. $A$ and $B$ are orthogonal Latin squares.
\end{proof}
We now have a condition that we can apply to quantum Latin squares. However, for QLSs $A$ and $B$ this turns out to preclude superpositions, thus making $A$ and $B$ Latin squares.
\begin{lemma}
Given a pair of quantum Latin squares, if they obey equation~\eqref{eq:orthols}, then they are Latin squares.
\end{lemma}
\begin{proof}Let $A$ and $B$ be QLSs such that the linear map $P$ as defined above is a permutation of basis states. Then the adjoint of $P$, $P^\dag= \sum_{i}\sum_{j}\sum_{k}\ket{ A_{ij}} \ket k \bra{i} \braket{{B_{ij}}}{k}\bra j$ must also be a permutation of basis states. We input computational basis states $p$ and $q$ into $P^\dag$
\begin{align*}
P^\dag(\ket p \ket q)&=\sum_k\ket{A_{pq}}\ket k \braket{{B_{pq}}}{k}\\&=\sum_k\ket{A_{pq}}\ket k\overline{ \braket{k}{{B_{pq}}}}\\&=\sum_k\ket{A_{pq}}\ket k{ \braket{k}{\overline{{B_{pq}}}}}\\&=\ket{A_{pq}}\left [\sum_k\ket k \bra{k} \right ]\overline{\ket{{B_{pq}}}}\\&=\ket{A_{pq}}\overline{\ket{B_{pq}}}
\end{align*}
The second equality is due to the fact that the inner product is Hermitian, the third equality is due to $\ket k$ being real valued for all $k$, the fourth equality is an algebraic rearrangement and the final equality is a resolution of the identity. If $P^\dag$ above is a permutation of basis states, then for all $p,q \in \{0,...,n-1\}$, $\ket{A_{pq}}$ and $\overline{\ket{B_{pq}}}$ must be computational basis states. Thus $A$ and $B$ are Latin squares.
\end{proof}
In order to define orthogonality for QLSs we will now make a (very) brief detour into quasigroup theory.
Latin squares can be thought of as the multiplication (Cayley) table for finite order quasigroups~\cite{smith} on the computational basis states. Let $*$ be the binary operation given by a Latin square. The fact that each state appears exactly once in each row and each column means that knowledge of any two of $a, b$ and $c$ in the equation ${a*b=c}$ uniquely determines the third. This means we can canonically define the binary operation $\backslash$ , read as \textit{left divide}, such that $a*b=c \Rightarrow a \backslash c=b$. This new binary operation defines a new quasigroup and therefore a new Latin square called the \textit{left conjugate Latin square} (it can easily be checked that this does indeed give a Latin square)~\cite{smith}. The map that takes a Latin square and gives the left conjugate $L \super{$\backslash$}\longrightarrow L'$, is in fact involutive so we can recover $L$ from $L'$ by applying the map again. We will see a nice graphical characterisation of this fact below. The map $L \super{$\backslash$}\longrightarrow L'$ is a bijection on the set of all Latin squares.
\begin{definition}[Left orthogonality]\label{def:lols}
Given a pair of Latin squares they are \textit{left orthogonal} when their left conjugates are orthogonal.
\end{definition}
\begin{remark}We could equally well talk about the right conjugate given by right divide and define right orthogonality. In this paper we only make use of left orthogonality.
\end{remark}
Since $L \super{$\backslash$}\longrightarrow L'$ is a bijection as mentioned above, the set of orthogonal Latin squares and left orthogonal Latin squares are isomorphic. Left orthogonality is in fact the property that we have generalised to QLSs in Definition~\ref{def:woqls}.
To proceed further it will be useful to introduce some diagrams (see Appendix~\ref{apx:cqm}). Let \tinymult[lsdot] be a Latin square and $\tinyspider$ be the classical structure corresponding to the computational basis. Then the left divide map has the following form:
\begin{equation}\label{eq:leftdivide2}
\begin{pic}
\node[lsdot] (ls) {};
\draw[string,in=right,out=up] (0.25,-0.5) to (ls.center);
\draw[string,in=left,out=up] (-0.25,-0.5) to (ls.center);
\draw[string] (ls.center) to (0,0.5);
\end{pic}
\quad \super{$\backslash$}\longrightarrow \quad
\begin{pic}
\node[lsdot] (ls) {};
\node[blackdot,scale=0.8] (b) at (-0.35,0.25) {};
\draw[string] (ls.center) to (0,-0.5);
\draw[string,in=right,out=down] (0.25,0.5) to (ls.center);
\draw[string,out=left,in=right,looseness=1.7] (ls.center) to (b.center);
\draw[string,in=left,out=up] (-0.7,-0.5) to (b.center);
\end{pic}
\end{equation}\label{eq:leftdivide}
The fact that $\backslash$ is an involution can be verified using the snake equation:
\begin{equation}\label{eq:idem}
\begin{pic}
\node[lsdot] (ls) {};
\draw[string,in=right,out=up] (0.25,-0.5) to (ls.center);
\draw[string,in=left,out=up] (-0.25,-0.5) to (ls.center);
\draw[string] (ls.center) to (0,0.5);
\end{pic}
\quad \super{$\backslash$}\longrightarrow \quad
\begin{pic}
\node[lsdot] (ls) {};
\node[blackdot,scale=0.8] (b) at (-0.35,0.25) {};
\draw[string] (ls.center) to (0,-0.5);
\draw[string,in=right,out=down] (0.25,0.5) to (ls.center);
\draw[string,out=left,in=right,looseness=1.7] (ls.center) to (b.center);
\draw[string,in=left,out=up] (-0.7,-0.5) to (b.center);
\end{pic}
\quad \super{$\backslash$}\longrightarrow \quad
\begin{pic}
\node[lsdot] (ls) {};
\node[blackdot,scale=0.8] (b) at (-0.35,-0.25) {};
\node[blackdot,scale=0.8] (b2) at (-0.7,0){};
\draw[string] (ls.center) to (0,0.5);
\draw[string,in=right,out=up] (0.25,-0.5) to (ls.center);
\draw[string,out=left,in=right,looseness=1.7] (ls.center) to (b.center);
\draw[string,in=left,out=right,looseness=1.7] (b2.center) to (b.center);
\draw[string,in=left,out=up] (-0.95,-0.5) to (b2.center);
\end{pic}
\quad \super{\eqref{eq:sm}}= \quad
\begin{pic}
\node[lsdot] (ls) {};
\draw[string,in=right,out=up] (0.25,-0.5) to (ls.center);
\draw[string,in=left,out=up] (-0.25,-0.5) to (ls.center);
\draw[string] (ls.center) to (0,0.5);
\end{pic}
\end{equation}
For Latin squares $A=\tinymult[lsdot]$ and $B=\tinymult[lssdot]$, equation~\eqref{eq:orthols} can be expressed diagramatically as follows:
\begin{equation}\label{eq:orthoperm}
\begin{pic}
\node[morphism,scale=2](p) {$P$};
\draw[string] (0.475,0.6) to (0.475,1.2);
\draw[string] (-0.475,0.6) to (-0.475,1.2);
\draw[string] (0.475,-0.6) to (0.475,-1.2);
\draw[string] (-0.475,-0.6) to (-0.475,-1.2);\end{pic}
\quad := \quad
\begin{pic}
\node[lsdot] (ls) at (0,-0.25){};
\node[lssdot] (lss) at (0,1){};
\node[blackdot,scale=0.8] (b1) at(-0.5,0.375){};
\node[blackdot,scale=0.8] (b2) at (0.5,0.375){};
\node[blackdot,scale=0.8] (b3) at (0.4,1.3){};
\draw[string, in=down,out=left] (ls.center) to (b1.center);
\draw[string, in=down,out=right] (ls.center) to (b2.center);
\draw[string, in=right,out=up] (b2.center) to (lss.center);
\draw[string, in=up,out=left] (lss.center) to (b1.center);
\draw[string, in=left,out=up] (lss.center) to (b3.center);
\node (1) at (0,-0.8){};
\draw[string] (1.center) to (ls.center);
\node (2) at (-1,1.8){};
\draw[string] (2) to (b1.center);
\node (3) at (1,1.8){};
\draw[string] (3) to (b2.center);
\node (4) at (1.25,-0.8){};
\draw[string,in=right,out=up] (4.center) to (b3.center);
\end{pic}
\quad
\text{ is a permutation}
\end{equation}
We now substitute in the left conjugates of Latin squares $A$ and $B$, $\tinymult[lsdot]\super{$\backslash$}\longrightarrow \tinyleftdivide[lsdot]$ and $\tinymult[lssdot]\super{$\backslash$}\longrightarrow \tinyleftdivide[lssdot]$ to obtain a linear map $P'$ which must be a permutation of basis states for $A$ and $B$ to be left orthogonal.
The condition that $A$ and $B$ are left orthogonal is thus equivalent to the following statement:
\begin{equation}\label{eq:lolsb}
\begin{pic}
\node[morphism,scale=2](p) {$P'$};
\draw[string] (0.475,0.6) to (0.475,1.2);
\draw[string] (-0.475,0.6) to (-0.475,1.2);
\draw[string] (0.475,-0.6) to (0.475,-1.2);
\draw[string] (-0.475,-0.6) to (-0.475,-1.2);
\end{pic}
\quad := \quad
\begin{pic}
\node[lsdot] (ls) at (0,-0.25){};
\node[lssdot] (lss) at (0,1){};
\node[blackdot,scale=0.8] (b1) at(-0.75,0.375){};
\node[blackdot,scale=0.8] (b2) at (0,0.375){};
\node[blackdot,scale=0.8] (b3) at (0.4,1.3){};
\node[blackdot,scale=0.8] (b5) at (-0.375,1.25){};
\node[blackdot,scale=0.8] (b4) at (-0.375,-0.5){};
\draw[string,out=left,in=up] (b5.center) to (b1.center);
\draw[string, in=right,out=left] (ls.center) to (b4.center);
\draw[string,out=left,in=down] (b4.center) to (b1.center);
\draw[string, in=down,out=up] (ls.center) to (b2.center);
\draw[string, in=down,out=up] (b2.center) to (lss.center);
\draw[string, in=right,out=left] (lss.center) to (b5.center);
\draw[string, in=left,out=right,looseness=1.5] (lss.center) to (b3.center);
\node (1) at (0.25,-0.8){};
\draw[string,in=right,out=up] (1.center) to (ls.center);
\node (2) at (-1,1.8){};
\draw[string] (2.center) to (b1.center);
\node (3) at (1,1.8){};
\draw[string] (3.center) to (b2.center);
\node (4) at (1.25,-0.8){};
\draw[string,in=right,out=up] (4.center) to (b3.center);
\end{pic}
\quad \super{\eqref{eq:sm}}= \quad
\begin{pic}
\node[lsdot] (ls) at (0,-0.25){};
\node[lssdot] (lss) at (0,1){};
\node (b1) at(-0.75,0.375){};
\node[blackdot,scale=0.8] (b2) at (0,0.375){};
\node[blackdot,scale=0.8] (b3) at (0.4,1.3){};
\node[blackdot,scale=0.8] (b5) at (-0.375,1.25){};
\node[blackdot,scale=0.8] (b4) at (-0.375,-0.5){};
\draw[string,out=left,in=up] (b5.center) to (b1.center);
\draw[string, in=right,out=left] (ls.center) to (b4.center);
\draw[string,out=left,in=down] (b4.center) to (b1.center);
\draw[string, in=down,out=up] (ls.center) to (b2.center);
\draw[string, in=down,out=up] (b2.center) to (lss.center);
\draw[string, in=right,out=left] (lss.center) to (b5.center);
\draw[string, in=left,out=right,looseness=1.5] (lss.center) to (b3.center);
\node (1) at (0.25,-0.8){};
\draw[string,in=right,out=up] (1.center) to (ls.center);
\node (2) at (-0.375,1.8){};
\draw[string] (2.center) to (b5.center);
\node (3) at (1,1.8){};
\draw[string] (3.center) to (b2.center);
\node (4) at (1.25,-0.8){};
\draw[string,in=right,out=up] (4.center) to (b3.center);
\end{pic}
\text{ is a permutation}
\end{equation}
In words: first we input two states $i$ and $j$ and then compute the component-wise inner products of the $i^{th}$ row of $A$ and the $j^{th}$ row of $B$. There must be one unique column, say $s,$ such that $\braket{B_{sj}}{A_{si}}=1$ with $\braket{B_{rj}}{A_{ri}}=0$ for all $r$ not equal to $s$. We then output $s$ on the left and $\ket{A_{si}}$ on the right. The set of output states $s \otimes \ket{A_{si}}$ must be every possible combination of computational basis states.
We can interpret this for QLSs but again we encounter the same difficulty. \begin{lemma}
Every pair of left orthogonal QLSs are Latin squares.
\end{lemma}
\begin{proof}
For a contradiction assume that $A$ and $B$ are left orthogonal QLSs that are not Latin squares. There is some vector entry in $A$ that is not a computational basis state say $\ket{A_{pq}}$. For $P'$ as defined in Equation~\eqref{eq:lolsb} to be a permutation, $\ket{A_{pq}}$ cannot be the output on the right for any input $q,j$. This means that no row of $B$ has the complex conjugate of $\ket{A_{pq}}$ as its $p^{th}$ column entry. But each row of $B$ must have one column entry that is the complex conjugate of the corresponding column entry of the $q^{th}$ row of $A.$ Thus at least two of the rows of $B$\ have the same vector in the same column. This violates the rule that $B$ is a QLS and thus gives a contradiction. Therefore $A$ must be a Latin square. Reversing the roles, we find that $B$ must be a Latin square too (left orthogonality, like orthogonality is a symmetric relation).
\end{proof}
The condition must therefore be weakened if we want to define a property that non-Latin square QLSs can satisfy. One approach is to delete the output from the right hand wire and require that the linear map thus obtained be a function on the computational basis states. This is in fact the \textit{weak orthogonality} property of Definition~\ref{def:woqls}. This condition turns out to be strong enough to give rise to interesting and useful properties such as using QLSs to build mutually unbiased MEBs (see Theorem \ref{thm:main}), yet weak enough so that pairs of Latin squares are weak orthogonal if and only if they are orthogonal.
Diagrammatically Definition~\ref{def:woqls} becomes the following:
\begin{equation}\label{eq:lols}
\begin{pic}
\node[whitedot,scale=2](p) {$f$};
\draw[string] (p.center) to (0,1.2);
\draw[string] (0.475,0) to (0.475,-1.2);
\draw[string] (-0.475,0) to (-0.475,-1.2);
\end{pic}
\quad := \quad
\begin{pic}
\node[lsdot] (ls) at (0,-0.25){};
\node[lssdot] (lss) at (0,1){};
\node (b1) at(-0.75,0.375){};
\node[blackdot,scale=0.8] (b2) at (0,0.375){};
\node[blackdot,scale=0.8] (b3) at (0.4,1.3){};
\node[blackdot,scale=0.8] (b5) at (-0.375,1.25){};
\node[blackdot,scale=0.8] (b4) at (-0.375,-0.5){};
\draw[string,out=left,in=up] (b5.center) to (b1.center);
\draw[string, in=right,out=left] (ls.center) to (b4.center);
\draw[string,out=left,in=down] (b4.center) to (b1.center);
\draw[string, in=down,out=up] (ls.center) to (b2.center);
\draw[string, in=down,out=up] (b2.center) to (lss.center);
\draw[string, in=right,out=left] (lss.center) to (b5.center);
\draw[string, in=left,out=right,looseness=1.5] (lss.center) to (b3.center);
\node (1) at (0.25,-0.8){};
\draw[string,in=right,out=up] (1.center) to (ls.center);
\node (2) at (-0.375,1.8){};
\draw[string] (2.center) to (b5.center);
\node[blackdot,scale=0.8] (3) at (0.8,1.5){};
\draw[string] (3.center) to (b2.center);
\node (4) at (1.25,-0.8){};
\draw[string,in=right,out=up] (4.center) to (b3.center);
\end{pic}
\quad \super{\eqref{eq:sm}}= \quad
\begin{pic}
\node[lsdot] (ls) at (0,-0.25){};
\node[lssdot] (lss) at (0,1){};
\node (b1) at(-0.75,0.375){};
\node[blackdot,scale=0.8] (b3) at (0.4,1.3){};
\node[blackdot,scale=0.8] (b5) at (-0.375,1.25){};
\node[blackdot,scale=0.8] (b4) at (-0.375,-0.5){};
\draw[string,out=left,in=up] (b5.center) to (b1.center);
\draw[string, in=right,out=left] (ls.center) to (b4.center);
\draw[string,out=left,in=down] (b4.center) to (b1.center);
\draw[string, in=down,out=up] (ls.center) to (b2.center);
\draw[string, in=down,out=up] (b2.center) to (lss.center);
\draw[string, in=right,out=left] (lss.center) to (b5.center);
\draw[string, in=left,out=right,looseness=1.5] (lss.center) to (b3.center);
\node (1) at (0.25,-0.8){};
\draw[string,in=right,out=up] (1.center) to (ls.center);
\node (2) at (-0.375,1.8){};
\draw[string] (2.center) to (b5.center);
\node (3) at (1,1.8){};
\node (4) at (1.25,-0.8){};
\draw[string,in=right,out=up] (4.center) to (b3.center);
\end{pic}
\text{ is a function}
\end{equation}
\begin{lemma}\label{lem:rest}
Given a pair of Latin squares, $A$ and $B$ the following are equivalent:
\begin{itemize}
\item $A$ and $B$ are weak orthogonal (see Definition~\ref{def:woqls}). \item $A$ and $B$ are left orthogonal (see Definition~\ref{def:lols}).
\end{itemize}
\end{lemma}
\begin{proof}
If $A$ and $B$ are left orthogonal then $P'$, as defined in Equation~\eqref{eq:lolsb}, is a permutation of basis states, which clearly implies the weaker condition that $f$ as defined in Equation~\eqref{eq:lols} is a function. For the other implication let $A$ and $B$ be weak orthogonal Latin squares. Consider the $p^{th}$ columns of $A$ and $B.$ They both contain all $n$ computational basis states and there must therefore exist values of $i$ and $j$ for all $q\in \{0,...,n-1\}$ such that $\ket{A_{pi}}=\ket{B_{pj}}=\ket q$. So for column $p$ there exist $i,j$ such that $P'(\ket i \otimes \ket j)=\ket p \otimes\ket q$ for all $q$. This is true for all rows $q$, so $P'$ is a permutation.
\end{proof}
\begin{remark}
We defined weak orthogonality from left orthogonality by setting the requirement that the linear map $P'$ (see Equation~\eqref{eq:lolsb})with the right hand output deleted needs to be a function on the basis states, rather than requiring $P'$ itself to be a permutation of the basis states. We could have tried to weaken orthogonality directly by requiring that $P$ (see Equation~\eqref{eq:orthoperm}) with the right hand output deleted be a function on basis states. However, it turns out that this would still preclude non-Latin square QLSs.
\end{remark}
\section{Beth and Wocjan's MUB construction}\label{sec:bw}
In their 2004 paper~\cite{ortho1} Beth and Wocjan gave a construction for a pair of mutually unbiased bases of a Hilbert space $\mathcal H$ of square dimension $s=n^2$, given as input a pair of $n \times n$ orthogonal Latin squares and an $n \times n$ Hadamard matrix which was later put in explicit Latin square form by Wehner and Winter~\cite{ortho1,ortho2}.
The construction takes each Latin square together with the Hadamard and produces an MEB of dimension $n^2$. The fact that the Latin squares are orthogonal is then shown to entail that these two bases are mutually unbiased. I will refer to this MEB construction as the Left Beth-Wocjan maximally entangled basis (LBW MEB) construction\footnote{The construction presented here is technically the construction given by taking the left conjugate of the Latin square $L$ first and then applying the construction defined by Beth and Wocjan. Since taking the left conjugate gives us a bijection (see Equation~\eqref{eq:leftdivide}) on the set of Latin squares the MEBs obtainable are not affected by this.}.
\begin{definition}[Left Beth-Wocjan maximally entangled basis]\label{def:bw}
Given an $n \times n$ Latin square $L$ and an $n\times n$ Hadamard $H$, then $\mathcal B$ as defined below is a \textit{Left Beth-Wocjan maximally entangled basis} (LBW MEB).~\footnote{~\label{ftn}The definition below is slightly different to the one given by Beth and Wocjan even taking into account the use of the left conjugate Latin square. However, when the input is a Latin square the two constructions agree precisely. }\begin{equation}\label{eq:bwmeb}
\mathcal B := \left\{ B_{ij} = \frac{1}{\sqrt{n}}\sum_{k,p=0}^{n-1}\ket {k,p}H_{ik}\braket{L_{kp}} j \text{ such that } i,j \in \{ 0,..,n-1\} \right\}
\end{equation}
\end{definition}
The graphical calculus gives a good notation with which to compare LBW MEBs to QLS MEBs (see Definition~\ref{def:qlsmeb}).
\begin{lemma}\label{lem:subset}
Under the restriction to Latin squares and to having a single fixed Hadamard the QLS MEBs are the same as LBW MEBs.
\end{lemma}
\begin{proof}We construct an LBW MEB $B_{ij}$ and a QLS MEB $C_{ij}$ from the latin square $L=\tinymult[lsdot]$ and Hadamard $H$.
\vspace{20pt}
\hspace{0pt}\begin{tabular}[b]{m{6cm} m{7cm}}
\textit{\textbf{Left Beth-Wocjan MEB}} &
\textit{\textbf{Quantum Latin square MEB}} \\
\begin{equation*}\label{eq:bwdiag}\vspace{15pt}B_{ij}:=
\frac{1}{\sqrt{n}} \begin{pic}
\node (in) at (-0.5,4) {};
\node (H)[morphism,wedge,scale=0.5] at(0.25,1.5) {H};
\node (b1)[blackdot,scale=1] at (0.25,2) {};
\node (b2)[lsdot,scale=1.2] at (1,3) {};
\node (i)[state,black,scale=0.5] at(0.25,1) {$i$};
\node (0b) at (0.5,2) {};
\node (j)[state,black,scale=0.5] at(1.5,2) {$j$};
\node (out) at (1,4) {};
\draw[string,out=270,in=180,looseness=0.75] (in.center) to (b1.center); \draw (H.north) to (b1.center);
\draw[string,out=0,in=180] (b1.center) to (b2.center);
\draw[string,out=90,in=0] (j.center) to (b2.center);
\draw[string] (b2.center) to (out.center);
\draw[string] (i) to (H.south);
\end{pic}\end{equation*} &
\begin{equation*}\label{eq:qlsdiag}C_{ij}:=
\frac{1}{\sqrt{n}}\begin{pic}
\node (in) at (-0.5,4) {};
\node (H)[morphism,wedge,scale=0.5] at(0.25,1.5) {H};
\node (b1)[blackdot,scale=1] at (0.25,2) {};
\node (b2)[lsdot,scale=1.2] at (1,3) {};
\node (i)[state,black,scale=0.5] at(0.25,1) {$i$};
\node (0b) at (0.5,2) {};
\node (j)[state,black,scale=0.5] at(1.5,2) {$j$};
\node (out) at (1,4) {};
\draw[string,out=270,in=180,looseness=0.75] (in.center) to (b1.center); \draw (H.north) to (b1.center);
\draw[string,out=0,in=180] (b1.center) to (b2.center);
\draw[string,out=90,in=0] (j.center) to (b2.center);
\draw[string] (b2.center) to (out.center);
\draw[string] (i) to (H.south);
\end{pic}\end{equation*}
\\
\end{tabular}
We see that the diagrams are the same.\end{proof}
\begin{theorem}\label{thm:bw}
Given a pair of $n \times n$ left\footnote{In their paper Beth and Wocjan use orthogonal Latin squares, but since we defined their MEB construction on the left conjugate the \textit{left} becomes necessary here.} orthogonal Latin squares and an $n \times n$ Hadamard, construct two LBW MEBs using each Latin square with the Hadamard. The bases are mutually unbiased.
\end{theorem}
\begin{lemma}
The construction of MUBs in Theorem~\ref{thm:main} restricts to the construction of Theorem~\ref{thm:bw}, under the restriction of the QLS to a Latin square and the two families of Hadamards to a single fixed Hadamard.
\end{lemma}
\begin{proof}
Follows directly from Lemma~\ref{lem:subset}.
\end{proof}
The following corollary gives a construction for MUBs in square dimension that is more general than the LBW MUB construction but not as general as our main construction. \begin{corollary}\label{corl}
Given two indexed families of $n$ Hadamards $H_k$ and $G_j$ both of size $n \times n$, and a pair of $n \times n$ left orthogonal Latin squares $\mathcal P$ and $\mathcal Q$, the bases $B(\mathcal P, H_k)$ and $B(\mathcal Q, G_j)$ are mutually unbiased.
\end{corollary}
So our new construction generalises Beth and Wocjan's in two directions, having two arbitrary families of Hadamards rather than a single fixed Hadamard and quantum Latin squares
rather than Latin squares. The next theorem shows, by explicit example, that the generalisation is strict.
\begin{theorem}\label{thm:ineq}
The pair of mutually unbiased MEBs from Example~\ref{ex:qlsmub} are inequivalent to any MEBs obtainable by the LBW MEB construction.
\end{theorem}
\begin{proof}
It will be sufficient to prove that one of our MEBs is inequivalent to any obtainable by the LBW MEB construction. Since equivalence of MEBs is the same as equivalence of UEBs we will take the dual approach here (see Section~\ref{sec:dual} below) and prove that the UEB arising from QLS $\mathcal P$ and Hadamard $H$ in Example~\ref{ex:qlsmub}, which we will refer to as $X$, is inequivalent to any LBW UEB.
We will proceed along the same lines as~\cite[Corollary 31]{mypaper1}. Note that LBW UEBs are a restriction to a single fixed Hadamard of shift-and-multiply UEBs. Thus by~\cite[Proposition 30]{mypaper1}, LBW UEBs are \textit{monomial} (meaning each unitary matrix of the basis is the product of a diagonal matrix and a permutation matrix).
Suppose for a contradiction that $X$ is equivalent to a monomial basis. The first matrix of $X$ is as follows:
\begin{equation*}
X_{00}=
\begin{pmatrix}
1&0&0&0&0&0&0&0&0\\
0&0&1&0&0&0&0&0&0\\
0&1&0&0&0&0&0&0&0\\
0&0&0&1&0&0&0&0&0\\
0&0&0&0&0&1&0&0&0\\
0&0&0&0&1&0&0&0&0\\
0&0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&0&1\\
0&0&0&0&0&0&0&1&0
\end{pmatrix}
\end{equation*}
$X_{00}$ is self adjoint. We obtain the equivalent UEB $X'$ by composing all the matrices of $X$ on the right by $X_{00}$. Thus $X'_{00}=\id[9]$.
Now $X'$ contains the identity and is equivalent to a monomial basis so by~\cite[Proposition 26]{mypaper1} $X'$ is \textit{simultaneously monomializable.} (See~\cite[Definition 25]{mypaper1} .
The least common multiple of $\{1,2,3,4,5,6,7,8,9\}$ is $\mu_9 = 2520$; thus by~\cite[Proposition 28]{mypaper1} the $2520^{th}$ powers of the elements of $X$ will commute. Now let $\omega=e^{2\pi i/3}$and consider $X'_{06}$ (left) and $X'_{07}$ (right) below:
\begin{equation*}
\begin{pmatrix}
0&0&0&0&0&0& \frac{1}{\sqrt{3}}&\frac{1}{\sqrt{3}}&\frac{1}{\sqrt{3}}\\
0&0&0&0&0&0& \frac{1}{\sqrt{3}}&\frac{\omega^2}{\sqrt{3}}&\frac{\omega}{\sqrt{3}}\\
0&0&0&0&0&0& \frac{1}{\sqrt{3}}&\frac{\omega}{\sqrt{3}}&\frac{\omega^2}{\sqrt{3}}\\
\frac{1}{\sqrt{3}}&-i\sqrt{\frac{2}{7}}&\sqrt{\frac{2}{3}}&0&0&0&0&0&0\\
\frac{1}{\sqrt{3}}&\frac{-i}{\sqrt{14}}&\frac{-1}{\sqrt{6}}&0&0&0&0&0&0\\
\frac{i}{\sqrt{3}}&\frac{3}{\sqrt{14}}&\frac{i}{\sqrt{6}}&0&0&0&0&0&0\\
0&0&0&1&0&0&0&0&0\\
0&0&0&0&0&1&0&0&0\\
0&0&0&0&1&0&0&0&0
\end{pmatrix}
,
\begin{pmatrix}
0&0&0&0&0&0& \frac{1}{\sqrt{3}}&\frac{1}{\sqrt{3}}&\frac{1}{\sqrt{3}}\\
0&0&0&0&0&0& \frac{\omega^2}{\sqrt{3}}&\frac{\omega}{\sqrt{3}}&\frac{1}{\sqrt{3}}\\
0&0&0&0&0&0& \frac{\omega}{\sqrt{3}}&\frac{\omega^2}{\sqrt{3}}&\frac{1}{\sqrt{3}}\\
-i\sqrt{\frac{2}{7}}&\sqrt{\frac{2}{3}}&\frac{1}{\sqrt{3}}&0&0&0&0&0&0\\
\frac{-i}{\sqrt{14}}&\frac{-1}{\sqrt{6}}&\frac{1}{\sqrt{3}}&0&0&0&0&0&0\\
\frac{3}{\sqrt{14}}&\frac{i}{\sqrt{6}}&\frac{i}{\sqrt{3}}&0&0&0&0&0&0\\
0&0&0&1&0&0&0&0&0\\
0&0&0&0&0&1&0&0&0\\
0&0&0&0&1&0&0&0&0
\end{pmatrix}
\end{equation*}
For a contradiction we now compute the first column first row entry of the commutator:
\begin{align*}K&:=(X'_{06})^{2520}(X'_{07})^{2520}-(X'_{07})^{2520}(X'_{06})^{2520} \\ \bra 0 K&\ket 0\approx-0.0219 + 0.0252i
\neq 0
\end{align*}
Thus $X'$ and therefore $X$ is not equivalent to any monomial basis, and in particular any LBW MEB.
\end{proof}
\section{Mutually unbiased error bases}\label{sec:dual}
Unitary error bases (UEBs) are the mathematical data necessary for protocols such as dense coding and teleportation as well as having important applications to quantum error correction. In this section we explain how the results of this paper can also be described in terms of UEBs via the correspondence between maximally entangled bases in square dimension and UEBs by introducing the natural concept of mutually unbiased UEBs.
\begin{definition}[Unitary error basis]
A \textit{unitary error basis} on an $n$-dimensional Hilbert space is a family of $n^{2}$ unitary matrices $U_i$, each of size $n \times n$, such that~\cite{klapp}:
\begin{equation}\label{eqdef:ueb}
\tr (U^{\dag}_i \circ U_{j})=\delta_{ij}n
\end{equation}
\end{definition}
Via state-process duality a bijection exists between UEBs and MEBs (See Definition~\ref{def:meb})~\cite{ghosh}. The correspondence is particularly clear diagrammatically.
Given a UEB, $\mathcal A:=\{U_{i}|0<i \leq n^2\}$ and the computational basis $\tinyspider$, we have the corresponding MEB, $\mathcal B :=\{\ket{U_{i}}|0<i \leq n^2\}$ defined as follows (see~\cite{werner2000} Lemma 2):
\begin{equation}\label{uebmeb}
U_i := \quad
\begin{pic}
\node[morphism,scale=0.7] (U) {$U_i$};
\draw[string] (0,-0.5) to (U.south);
\draw[string] (0,0.5) to (U.north);
\end{pic}
\quad \leadsto \quad
\frac{1}{\sqrt{n}}\begin{pic}
\node[blackdot] (b) {};
\node (1) at (-0.5,1) {};
\node (2) at (0.5,1) {};
\node[morphism,scale=0.7] (U) at (0.5,0.5) {$U_i$};
\draw[string,in=left,out=down,looseness=1.3] (1.center) to (b.center);
\draw[string,in=up,out=down,looseness=1.3] (2.center) to (U.north);
\draw[string,in=down,out=right,looseness=1.1] (b.center) to (U.south);
\end{pic}
\quad =:\ket{U_i}
\end{equation}
By Equation~\eqref{mes} the condition that the matrices $U_i$ are unitary means that the states $\ket{U_i}$ are maximally entangled. Under this duality equivalence of MEBs as described by Equation~\ref{eq:eqimeb}, becomes the usual notion of equivalence for UEBs. The fact that the states on the right hand side of Equation~\eqref{uebmeb} are orthonormal follows directly from Equation~\eqref{eqdef:ueb} as follows:
\begin{equation}\label{eq:dual}
\braket{U_i}{U_j}\super{\eqref{uebmeb}}=
\frac{1}{n}\begin{pic}
\node[blackdot] (b) at (0,-1) {};
\node (1) at (-0.5,0) {};
\node (2) at (0.5,0) {};
\node[morphism,scale=0.7] (U) at (0.5,-0.5) {$U_i$};
\draw[string,in=left,out=down,looseness=1.3] (1.center) to (b.center);
\draw[string,in=up,out=down,looseness=1.3] (2.center) to (U.north);
\draw[string,in=down,out=right,looseness=1.1] (b.center) to (U.south);
\node[blackdot] (b1) at (0,0.75) {};
\node (a1) at (-0.5,0) {};
\node (a2) at (0.5,0) {};
\node[morphism,scale=0.7] (aU) at (0.5,0.25) {$U^\dagger_j$};
\draw[string,in=left,out=up,looseness=1.3] (a1.center) to (b1.center);
\draw[string,in=down,out=up,looseness=1.3] (a2.center) to (aU.south);
\draw[string,in=up,out=right,looseness=1.1] (b1.center) to (aU.north);
\end{pic}
=\frac{1}{n}\tr (U^{\dag}_i \circ U_{j})\super{\eqref{eqdef:ueb}}=\delta_{ij}
\end{equation}
In this paper the dual MEB constructions of two of the main constructions for UEBs were used. As mentioned above Lemma~\ref{lem:qlsmeb} the QLS MEB\ of that lemma is the dual of the quantum shift-and-multiply error bases of this author's paper with Jamie Vicary~\cite{mypaper1}. The MEB used in Corollary~\ref{corl} is the shift-and-multiply basis introduced by Werner~\cite{werner2001all}. Thus the LBW MEB construction described in Definition~\ref{thm:bw} gives us a family of UEBs strictly contained within Werners construction.
The duality of MEBs and UEBs makes it natural to talk about mutually unbiased unitary error bases. \begin{definition}[Mutually unbiased error bases]
A pair of unitary error bases over a Hilbert space $\mathcal H$ of dimension $n$, ${\mathcal A=\{U_{i}|i \in \{ 0,...,n-1\}\}}$ and ${\mathcal B=\{V_j|j \in \{ 0,...,n-1\}\}}$ are \textit{mutually unbiased} when the following equation holds for all $i,j$:
\begin{equation}
|\tr (U^{\dag}_i \circ V_j)| ^2= \frac{1}{n}
\end{equation}
\end{definition}
We had two choices in defining mutually unbiased UEBs above, we used the inner product of Equation~\eqref{eqdef:ueb} to interpret Equation~\eqref{def:mub} of Definition~\ref{def:mub} directly but we could have defined mutually unbiased UEBs to be UEBs with corresponding MEBs that are mutually unbiased. Fortunately it does not matter as they are equivalent by a similar argument to Equation~\eqref{eq:dual}.
This definition brings up the question of what it may mean for two teleportation protocols to be mutually unbiased, or what kind of error correction could be performed by a pair of mutually unbiased error bases.
The main result of this paper can now be interpreted as a construction for a pair of mutually unbiased unitary error bases from a pair of weak orthogonal quantum Latin squares.
\section{Mutually orthogonal quantum Latin squares}\label{sec:moqls}
In this section we introduce the concept of families of orthogonal quantum Latin squares. In their 2004 paper Beth and Wocjan~\cite{ortho1} introduced the construction of square dimensional MUBs from orthogonal Latin squares as described in Section~\ref{sec:bw}. They used this construction to improve the known lower bounds for maximal sets of pairwise mutually unbiased bases. A set of \textit{mutually orthogonal Latin squares} (MOLs) is a set of two or more Latin squares that are pairwise orthogonal. Beth and Wocjan use their construction on a set of $w$ MOLs of size $n\times n$ and give $w+2$ MUBs for dimension $n^2$. The extra two MUBs come from the two squares of vectors (which do not satisfy the axioms to be Latin squares, or even quantum Latin squares) described below:~\footnote{Note that due to the presentation of Beth and Wocjan's construction in Section~\ref{sec:bw}, in which we start by taking the left-conjugate, the left conjugate map must also be applied to these squares of vectors to recover the ones used by Beth and Wocjan. In addition the second square here only gives a basis using the original Beth-Wocjan method and not the altered version given by definition~\ref{def:bw} (See footnote~\ref{ftn}).}
\begin{itemize}
\item The first is the $n \times n$ grid with the $i^{th}$ row consisting of the repeated entry $\ket i$ for every column.
\item The second is the $n \times n$ grid with
$\sum_k^{n-1}\ket k$ as every diagonal entry and $0$s elsewhere.
\end{itemize}
Some thought reveals that although they are not Latin squares, these two squares are left orthogonal to every $n \times n$ Latin square and to each other. Note that the bases obtained from these extra two however are not maximally entangled. The following definition is a natural extension of the concept of sets of MOLs.
\begin{definition}[Mutually weak orthogonal quantum Latin squares]
A set of $w$ quantum Latin squares are \textit{Mutually weak orthogonal quantum Latin squares} (MOQLs) when they are pairwise weak orthogonal.
\end{definition}
There are no generalisations of the two squares of vectors described above that would be weak orthogonal to every QLS. However, with a particular set of MOQLs, an analogue of the first vector square above can be found by considering the subspaces spanned by the non-computational basis states. As an example we present a square of vectors that is weak orthogonal to both of the pair of weak orthogonal QLSs from Example~\ref{ex:orthoqls}. Again let $\ket i$, $i\in \{0,...,9\}$ be the computational basis states and define the states $\ket a , \ket b , \ket c,\ket \alpha,\ket \beta$ and $\ket \gamma$ as in Equations~\eqref{eq:a}~\eqref{eq:b}~\eqref{eq:c}~\eqref{eq:abg}~\eqref{eq:abg2} and~\eqref{eq:abg3}. We define the following square of vectors:
\begin{equation*}\small\arraycolsep=3pt \grid{
\ket{0} & \ket{0} & \ket{0} & \ket{0} & \ket{0} & \ket{0} & \ket{\alpha} & \ket{\alpha} & \ket{\alpha}
\\ \hline
\ket 1
& \ket{1}
& \ket{1} & \ket{1} & \ket{1} & \ket{1} & \ket{\beta} & \ket{\beta} & \ket{\beta}
\\ \hline
\ket{2}
& \ket{2}
& \ket{2}
& \ket{2} & \ket{2} & \ket 2 & \ket \gamma & \ket \gamma & \ket \gamma
\\ \hline
\ket{a} & \ket{a} & \ket{a} & \ket{3} & \ket 3 & \ket 3 & \ket 3 & \ket 3 & \ket 3
\\ \hline
\ket b & \ket b & \ket b & \ket 4 & \ket 4 & \ket 4 & \ket 4 & \ket 4 & \ket 4 \\ \hline
\ket c & \ket c & \ket c & \ket 5 & \ket 5 & \ket 5 & \ket 5 & \ket 5 & \ket 5 \\ \hline
\ket 6 & \ket 6 & \ket 6 & \ket 6 & \ket 6 & \ket 6 & \ket 6 & \ket 6 & \ket 6 \\ \hline
\ket 7 & \ket 7 &\ket 7 &\ket 7 &\ket 7 &\ket 7 &\ket 7 &\ket 7 &\ket 7 \\ \hline
\ket 8 & \ket 8 &\ket 8 &\ket 8 &\ket 8 &\ket 8 &\ket 8 &\ket 8 &\ket 8}
\end{equation*}
It can be checked that this square is weak orthogonal to $\mathcal P$ and $\mathcal Q$ in Example~\ref{ex:orthoqls}. It is also weak orthogonal to any QLS weak orthogonal to $\mathcal P$ or $\mathcal Q$. To see this consider that any two weak orthogonal QLSs must have columns that are permutations of each other.
This example relies on the \textit{block-like} structure of the QLSs in question. Any family of MOQLS\ having a similar structure will admit a similar square of vectors. It is unknown whether all QLSs are of this form, but to the authors knowledge none have been found yet that do not have this structure up to equivalence.
The lower bound for the number of MOQLS in dimension $n$ must be at least the lower bound for the number of MOLS, more research is required to say any more than that at this stage.
\section{Conclusion}
In our 2015 paper~\cite{mypaper1} the author together with Jamie Vicary introduced the quantum combinatorial objects of quantum Latin squares and gave a construction of UEBs using them. In this paper we have built upon that work by introducing mutually orthogonal quantum Latin squares which generalise mutually orthogonal Latin squares, which have been used extensively to derive results in quantum information. As an application we have given a construction for mutually unbiased bases in square dimension which gives MUBs that are inequivalent to those that can be constructed by any known method. There is the potential for improved bounds on maximal families of MUBs in composite dimensions using the main result of this paper.
\bibliographystyle{eptcs}
|
\section{Introduction}
In the last decade and a half, a quite effective tool is emerging for solving the
Bethe-Salpeter equation (BSE)\cite{SB_PR84_51} directly in Minkowski space, i.e.
avoiding to look for solutions in the Euclidean space by exploiting the Wick rotation
\cite{WICK_54}. The novel approach is based on the Nakanishi integral representation
(NIR) of the $n$-leg transition amplitudes, that was proposed long time
ago~\cite{nak63,nak69,nak71}. In particular, the Bethe-Salpeter (BS) amplitude can be formally written
like the NIR for the $3$-leg amplitude, namely a proper folding of an unknown Nakanishi
weight function and a denominator that contains the analytic
structure \cite{KusPRD95,KusPRD97,SauPRD03,carbonell1,carbonell2,carbonell3,FSV1,FSV2}. To
be practical, let us quickly mention the main features of the NIR for the BS amplitude: (i)
the weight function is real and smooth for bound states
(see \cite{FSV3} for the zero-energy scattering states), and (ii) it depends upon
real variables, of which one is non-compact and the others are compact; (iii) the
denominator must depend only upon the independent scalars that can be constructed
from the external momenta. Assuming the validity of the NIR for the bound-state case,
and taking advantage of the above mentioned features, one can exactly project onto the
null-plane (see, e.g. \cite{SalPRC00,FreFBS00}) the BS
amplitude integrating over the Light-front (LF) variable $k^-=k^0-k^3$ ($k^{+}= k^{0}+ k^{3}$
and ${\bf k}_{\perp}\equiv \{k^1,k^2 \} $) and formally obtain the so-called
LF {\em valence wave function} (cf. \cite{CK_rev,Brodrev,FSV1}), i.e. the amplitude of the component
with the lowest number of constituents when the LF Fock expansion of the
interacting-system state is considered. Remarkably, within the NIR approach, the LF
valence wave function is given by a non-singular integral involving the Nakanishi
weight function. This suggests to integrate on $k^-$ both sides of the BSE,
getting an integral equation for the Nakanishi weight function.
If there exist solutions for this integral equation (this validates a posteriori the
previous assumption), then the BS amplitudes of bound states can be reconstructed.
In particular, when the above procedure is applied to the BSE with the irreducible kernel in
ladder approximation, a generalized eigen-equation for the Nakanishi weight function is obtained
(see, e.g., Refs. \cite{carbonell1,FSV2} for the LF case), while for
the cross-ladder case one has to deal with a non-linear eigen-problem (see Ref. \cite{carbonell2}).
Notice that in Refs. \cite{KusPRD95,KusPRD97} solutions of the scalar BSE in ladder approximation
have been obtained by using (i) standard variables (and not the LF ones), and (ii)
exploiting the uniqueness of the Nakanishi weight function.
Aim of the present work is to carefully study both spectrum and 3D structure of the bound states,
obtained by solving the ladder BSE for a system composed by two massive scalars interacting through
a massive scalar. Such an investigation is a natural extension of the previous analysis of only
the ground state \cite{FSV2}. In particular, the structure is studied by means of the 3D
representation of the LF valence component, both in momentum and impact-parameter (IP) spaces.
One of the motivations for starting a detailed analysis of the non-perturbative features of an interacting
system in momentum and IP spaces (see, e.g.\cite{BurIJMP03} for an introduction) is given
by the increasing interest on this topic in hadronic physics, where the valence component plays an
important role in determining the dynamical properties of hadrons. For instance, the valence
component is an important dynamical ingredient for evaluating parton transverse-momentum distributions,
which depend upon both the Bjorken momentum fraction $x$ and the transverse components of parton
momentum~\cite{DiePRep03,BarPRep02}, or parton density distributions in IP
space, that can be related to the generalized parton distributions (see, e.g., Ref.~\cite{DiePRep03}).
Our paper is organized as follows.
In Sec. \ref{MSBSE}, we quickly introduce the general formalism (see, e.g.,
Refs. \cite{FSV1,FSV2,FSV3} for more details) and we present
a comparison between Minkowski and Euclidean results for the eigenvalues of the relevant
integral equation. In Sec. \ref{LFWF}, the
valence LF wave function and the corresponding density distributions,
evaluated both in transverse-momentum space
and impact-parameter one, are discussed, showing our numerical results
for the available spectrum together with some interesting formal outcomes of our analysis.
In Sec. \ref{END}, conclusions are drawn and some perspectives presented.
\section{Minkowski space solutions of the Bethe-Salpeter equation}
\label{MSBSE}
Let us recall the general formalism we have adopted to solve the BSE in Minkowski space.
As it is well known the BSE in momentum space for
a relativistic bound state is given by the following homogeneous integral equation
\begin{equation}\label{bse}
\Phi(k,p)= G_{0}^{(12)}(k,p) \int \frac{d^4 k'}{(2\pi)^4} {\rm i}
K(k,k';p)\Phi(k',p) \, ,
\end{equation}
where ${\rm i} \,K(k,k';p)$ is the interaction kernel that contains all two-body irreducible diagrams,
$p^\mu$ is the total momentum with the bound state mass given by $M=\sqrt{p^2}$.
In the present approach we do not consider the self-energy contribution, so that
$G_{0}^{(12)}(k,p)$ is the product of two free propagators,
\begin{equation}\label{free-prop}
G_{0}^{(12)}(k,p)=\frac{\rm i}{\left[(p/2+k)^2 -m^2 +{\rm i}\epsilon\right]} \frac{{\rm i}}{\left[(p/2-k)^2 -m^2 +{\rm i}\epsilon\right]}
,\end{equation}
with $m$ the constituent mass.
The BS amplitude for an $s-$wave state solution of Eq. (\ref{bse})
can be written in terms of NIR as \cite{carbonell1,FSV1,FSV2}
\begin{equation}\label{NIR-bs}
\Phi(k,p)=-{\rm i} \int_{-1}^{1} dz' \int_{0}^{\infty} d \gamma' \frac{g(\gamma',z';\kappa^2)}
{[\gamma'+\kappa^2-k^2- p \cdot kz'-{\rm i}\epsilon]^3},
\end{equation}
where $\kappa^2\equiv m^2-M^2/4$.
By substituting (\ref{NIR-bs}) into (\ref{bse}) and integrating over $k^{-}$ on both sides, one can
obtain the following generalized integral equation for the Nakanishi weight function
(for details see Refs.~\cite{carbonell1,FSV1,FSV2}):
{\small
\begin{equation}\label{nakie}
\hspace{-0.2cm}\int_0^{\infty}\hspace{-0.2cm}d\gamma'
\left\{\hspace{-0.1cm}
\frac{g( \gamma', z; \kappa^2)}
{[ \gamma' + \gamma + z^2 m^2 + \left( 1 - z^2 \right) \kappa^2]^{2} }-
\hspace{-0.2cm}
\int_{-1}^{1}\hspace{-0.2cm}dz' V^{LF}(z,z',\gamma, \gamma') g(\gamma',z';\kappa^2)\right\}
=0
,\end{equation}\hspace{-0.2cm}
}
where {\small
\begin{equation} \label{val1}
\int_0^{\infty}d\gamma'~\frac{g(\gamma',z;\kappa^2)}
{[\gamma'+\gamma +z^2 m^2 + \left( 1 - z^2 \right) \kappa^2]^2}=p^+
\int {dk^- \over 2 \pi} \Phi(k,p)=
{\frac{\sqrt{2}\, \psi(\xi,{\bf k}_\perp)} {\xi(1-\xi)}},
\end{equation}
}
with $\gamma=|{\bf k}_\perp|^2$, $\xi=(1 -z)/2$ and $\psi(\xi,{\bf k}_\perp)$ is the {\em valence
light-front wave function} (the factor $\sqrt{2}$ comes from the symmetry of the problem; see, for example,
\cite{FSV1}). In Eq. (\ref{nakie})
$V^{LF}$ is the Nakanishi kernel given in terms of the BS 4D kernel, by
{\small \begin{equation}
V^{LF}(z,z',\gamma, \gamma')\equiv {\rm i} p^{+} \int_{-\infty}^{\infty} \frac{dk^{-}}{2 \pi} G_{0}^{(12)}(k,p) \int \frac{d^4 k'}{(2 \pi)^4}
\frac{{\rm i}K(k,k';p)}{[k'^2 + p\cdot k'z'-\gamma'-\kappa^2 + {\rm i}\epsilon ]^3}
.\end{equation}}
In this work we adopt the ladder approximation for the BS kernel:
\begin{equation}
\label{ladder-ker}
{\rm i}\,K^{(Ld)}(k,k')=\frac{{\rm i}\,(-{\rm i} \,g)^2}{ (k-k')^2-\mu^2+{\rm i}\,\epsilon} =
-{\rm i}~\frac{\alpha~(16 \pi m^2 )}{ (k-k')^2-\mu^2+{\rm i}\,\epsilon}
\, ,
\end{equation}
where $\alpha=g^2/(16 \pi m^2)$ and $\mu$ is the exchanged-scalar mass.
According to \cite{FSV2}, we have solved Eq. (\ref{nakie}) by using a basis function
expansion of the Nakanishi weight function, composed by Laguerre polynomials
$L_j(a\gamma)$ (with $j=0, 1, N_g$) for describing the $\gamma$-dependence (where
$a$ is an appropriate parameter, as discussed in \cite{FSV2}) and even
Gegenbauer polynomials $C^{(5/2)}_{2\ell}(z)$ for the $z$ one (with $2\ell=0,2,...,2 N_z$).
More specifically, for the $\gamma$-dependence we use an expansion in terms of the
functions ${\cal L}_j(\gamma) \equiv \sqrt{a} L_j(a\gamma) e^{-a\gamma/2}$, where
$\int_0^\infty d\gamma {\cal L}_i(\gamma){\cal L}_j(\gamma) = \delta_{ij}$.
The expansion in Gegenbauer polynomials is given in terms of the functions
$G_\ell(z)\equiv 3\sqrt{(2\ell)!\left(2\ell+\frac52\right)/\Gamma(2\ell+5)}(1-z^2)C_{2\ell}^{5/2}(z)$,
where $\int_{-1}^{1}dzG_\ell(z)G_{\ell'}(z)=\delta_{\ell\ell'}$.
This last choice is dictated by the
symmetry property of the Nakanishi weight function
$g(\gamma,z;\kappa^2)=g(\gamma,-z;\kappa^2)$, that is requested by the bosonic nature of the
adopted constituents. It should be recalled that a definite statistical property of the BS amplitude
avoids the so-called abnormal solutions of BSE, namely the ones
with negative norm \cite{nak63,nak69,KusPRD97}, that are associated with
excitations in relative time of the bound states (see Refs.~\cite{alkofer,desplanques} for a
more recent discussion of the issue).
Finally, the $z^2$ dependence of $g(\gamma,z;\kappa^2)$ entails a symmetry of
the valence wave function, namely $\psi(\xi,{\bf k}_\perp)=\psi(1-\xi,{\bf k}_\perp)$.
In our numerical approach, accurate convergence was achieved for the
ground state by using 14 Laguerre ($N_g=13$) and 10 Gegenbauer ($N_z=9$)
polynomials. For the excited states, the convergence was reached with
26 Laguerre and 10 Gegenbauer polynomials.
After introducing the basis function expansion and the ladder
approximation Eq. (\ref{ladder-ker}), Eq. (\ref{nakie})
turns into the matrix form of a generalized eigenvalue problem.
In particular, one can symbolically write
\begin{equation}\label{symbnak}
{\mathcal B}(M)\,g\, =\,\alpha {\mathcal A}(M)\,g ,
\end{equation}
where $g$ is the eigenvector. Differently from the familiar non-relativistic case, in the
eigen-equation (\ref{symbnak}) the role of eigenvalue is played by
the coupling constant $\alpha$, while
the mass of the system $M$ is a parameter that can be assigned, after fixing the
exchanged-scalar mass $\mu$. In the standard way
of analyzing the BSE within the NIR framework \cite{KusPRD95,KusPRD97,SauPRD03,carbonell1,carbonell2,carbonell3,FSV1,FSV2},
one introduces the binding energy as
\begin{equation}
B = 2m -M,
\end{equation}
which constrains the range of $B/m$ to the interval between $0$ and $2$, i.e. $0\le M/m\le 2$,
avoiding in this way the well-known instability of the $\phi^3$ model (see Ref.
\cite{gbaym}). In our NIR studies of the ladder BSE, Eq. (\ref{symbnak}) has a pivotal role.
First, after fixing $m$, $\mu$ and $B_{gr}$, it yields the coupling constant of the ground
state, i.e. the smallest value of the coupling constant that we call $\alpha_{gr}$; secondly
it allows one to calculate the spectrum. Indeed, once we have found the
coupling constant $\alpha_{gr}$, one can find the excited state with respect to
$B_{gr}$ by slightly changing Eq. (\ref{symbnak}), as follows
\begin{equation}\label{symbnak1}
\lambda ~g= {1\over \alpha_{gr}}~{\mathcal B}(M)^{-1} {\mathcal A}(M)g
~~.\end{equation}
In other words, after fixing $m$, $\mu$ and $\alpha_{gr}$, we search for
values $M=2m-B ~>~ M_{gr}$ that produce eigenvalues
$\lambda=1$ (as trivially seen, for $M=M_{gr}$ one has $\lambda=1$).
\subsection{Comparing Minkowski and Euclidean eigenvalues}
In order to check the reliability of the computed masses for the excited states,
we provide a comparison between the results of our calculations, obtained in the Minkowski
space within the NIR, with those one can evaluate in the Euclidean space.
In Table \ref{comparison}, we show the binding energies, in unit of the constituent mass $m$, for
the first, $B(1)/m$, and the second, $B(2)/m$, excited states, corresponding to a
ground state $B(0)/m=1.9$ and different values of $\mu$. The choice of such a large binding energy is
motivated by the fact that strongly-bound states should be affected by large relativistic effects.
First, we have verified that the values of $\alpha_{gr}$ for the binding
energy of the ground state of $B(0)/m=1.9$, obtained with Euclidean- and Minkowski-space
calculations are the same within our numerical accuracy. Then, we have computed the excited
state energies given in Table \ref{comparison}, achieving a very satisfactory agreement
between the results evaluated in the two spaces. As a remark on the numerical procedure,
it should be pointed out that for values of $\mu/m$ smaller than 0.05 or
$B/m\,< \,0.01$ the convergence is quite slow, and it is needed an extrapolation of the
results with respect to $N_g$, in order to accurately determine the eigenvalues.
It is also important to show the behavior of energy ratios, $B(n)/B(0)$ with $n\geq0$, for small $B/m$
and $\mu \rightarrow 0$. For a bound state composed for two spinless bosons exchanging a massless
scalar boson, the corresponding relativistic expression
in lower orders of $\alpha$, as derived in \cite{1973-Feldman}, is
\begin{equation}
B(n)=\frac{m}{4} \frac{\alpha^2}{(n+1)^2}\left[1+\frac{4\alpha}{\pi}{\rm ln}\alpha \right]
+ ... \;\;\; (n \geq 0) .
\label{nrB}
\end{equation}
The first term is the non-relativistic limit.
As verified in Fig.~\ref{ratio}, { where $B(1)/B(0)$ and $B(2)/B(0)$
are shown for small values of $\mu/m$, the energy ratios } are consistent
with the non-relativistic limit. { Moreover, the agreement for $\mu\to0$ between
the relativistic and non-relativistic eigenvalues is observed only
for small values of $n$. Indeed, as binding energies increase relativistic
effects become larger and larger}.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.5]{bs01-ratioenergies.pdf}
\end{center}
\caption{
Energy ratios $B(n)/B(0)$ vs $\mu/m$ for the first (solid line with bullets)
and second (dashed line with triangles) excited states. The symbols on the lines are the
values obtained through Eq. (\ref{symbnak1}), while the circle ($n=$1) and the triangle ($n=$2),
at the origin, represent the corresponding non-relativistic limits, given by
Eq.~(\ref{nrB}) with $ B_{nr}(0)/m=0.25\alpha^2 $.}
\label{ratio}
\end{figure}
\begin{table}[h]
\begin{center}
\caption{ Comparison of the spectra obtained in the Euclidean space in the Minkowski
one, by varying $\mu/m$ and, consequently, $\alpha_{gr}$, but taken fixed
the value of the ground-state
binding energy to $B(0)/m=1.9$.}\label{comparison}
\begin{tabular}{cccc}
\hline\hline
$(\mu/m , \alpha_{gr})$ & & \textbf{Euclidean} &
\textbf{Minkowski} \\ \hline \hline
\multicolumn{1}{c}{\vspace{-0.4cm} } & $B(1)/m$ & 0.258 & 0.259 \\
\multicolumn{1}{c}{\vspace{-0.4cm} (0.05, 6.324)} &&& \\
\multicolumn{1}{c}{ } & $B(2)/m$ & 0.090 & 0.090 \\ \hline
\multicolumn{1}{c}{\vspace{-0.4cm}} & $B(1)/m$ & 0.220 & 0.221 \\
\multicolumn{1}{c}{\vspace{-0.4cm} (0.1, 6.437)} &&& \\
\multicolumn{1}{c}{ } & $B(2)/m$ & 0.051 & 0.050 \\ \hline
\multicolumn{1}{c}{\vspace{0cm}} (0.5, 8.047)~~ & $B(1)/m$ & 0.0082& 0.0082 \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\section{ Valence light-front wave function and momentum distributions}
\label{LFWF}
It is attractive to perform numerical comparisons of dynamical quantities
that in perspective could be useful for an experimental investigation of actual
interacting systems. In view of this, from
the valence LF wave function introduced in Eq. (\ref{val1}) (see the next
subsection for the numerical results), one can define both the probability
distribution to find a
constituent with LF longitudinal fraction $\xi=p^+_i/P^+$, given by
\begin{equation}
\varphi(\xi)= {1 \over 2 (2 \pi)^3}~ {1\over \xi(1-\xi)}~ \int
d^2{\bf k}_\perp~
\Bigl [\psi(\xi,{\bf k}_\perp)\Bigr]^2 \, ,
\label{phixi}\end{equation}
and the probability distribution to find a
constituent with LF transverse momentum $k_\perp=|{\bf k}_\perp|$, that reads
\begin{equation}
{\mathcal P}(k_\perp)= {1 \over 4(2 \pi)^3}~ \int_0^1 {d\xi\over \xi(1-\xi)}~
\int_0^{2\pi} d\theta~
\Bigl [\psi(\xi,{\bf k}_\perp)\Bigr]^2 \, .
\label{probgam}
\end{equation}
Both LF distributions are normalized to the probability of the valence component,
once the BS amplitude itself
is properly normalized (see Ref. \cite{lurie} for a general discussion and Ref. \cite{FSV2}
for the application within the NIR). Such a probability yields
the probability to find the valence contribution in the LF Fock expansion of the interacting
two-scalar state (see, e.g., \cite{Brodrev,FSV1,dae}). As a matter of fact, one has
\begin{equation}
P_{val}= ~ \int_{0}^1 d\xi \, \varphi(\xi)\, = \int_{0}^{\infty} dk_\perp~ {\mathcal P}(k_\perp).
\label{pval}\end{equation}
Notice that ${\bf k}_\perp$ can be associated with the intrinsic transverse momentum,
in the frame where ${\bf p}_\perp=0$, which is allowed by the covariance of our
description.
Although we have discussed the issue of the proper normalization of the valence state,
in what follows we are interested on the overall 3D structure of the valence wave function, and therefore we have
simply adopted an arbitrary normalization.
\subsection{Momentum space valence wave function for excited states }
{In Figs.~\ref{lf-g} and ~\ref{lf-g1}}, we present the LF wave function of the { first (left panels)
and second (right panels)} excited states, corresponding
to the case $\mu/m$=0.1, $\alpha_{gr}$=6.437, $B(1)/m$=0.22 and $B(2)/m$=0.05
(see Table \ref{comparison}). As clearly shown, the wave function displays the typical
feature of the first and second excited states, i.e. one and two \textit{nodes}, respectively.
By a direct inspection of the corresponding panels for the first excited state
in Fig.~\ref{lf-g} { and for the second excited state
in Fig.~\ref{lf-g1}}, one observes that, in the plane
($\xi\, ,\, k_\perp/m$), the node structure is present
for $(k_\perp/m)^2 < 1$ and $\xi < 0.75 $, and it is symmetric with respect to $\xi=1/2$.
In particular, the node structure moves toward $\xi=1/2$
as $k_\perp$ increases. Such a behavior can be naively expected when Cartesian three-momenta
are adopted. As a matter of fact, the relation between Cartesian and LF components is
\begin{equation}
{\bf k}^2=\frac{k^2_\perp+m^2}{4\xi(1-\xi)}-m^2 .
\label{cmom}
\end{equation}
If we assume a dependence upon ${\bf k}^2$ for the excited-state valence wave function
(instead of the actual dependence upon
$\xi$
and $k_\perp$, separately), i.e. the same dependence found
in phenomenological valence wave functions widely
adopted for describing ground states (as the one exploited in the discussion of the nucleon form factors in Ref.
\cite{Brodrev}), then the behaviors shown in
{Figs. \ref{lf-g} and \ref{lf-g1}, for the node structures and asymptotic behaviors of the states,}
become quite reasonable.
Indeed, the assumed excited-state valence wave functions have to display a node at a fixed value for
${\bf k}^2$, and therefore according to { Fig.~(\ref{lf-g})},
for increasing $k_\perp$ the variable $\xi$ is constrained to approach
$1/2$ (i.e. the maximal value of $\xi(1-\xi)$), in order to take (almost) constant
${\bf k}^2$.
In conclusion, the correlation between the LF components,
$\xi$ and $k_\perp$, in determining the node position can be largely explained
by the rotational invariance of the phenomenological wave functions,
if they depend upon ${\bf k}^2$. Notably, our calculations, genuinely in Minkowski space, actually confirm
the overall expectation, based on a simple phenomenological Ansatz, that takes into account
the rotational invariance. It should be reminded that, within a LF framework, the rotational invariance
can be fully recovered only if the whole Fock expansion is considered~\cite{Brodrev}. We just add that
the second node present in the right panel of Fig. \ref{lf-g} is hard to be seen given the scale of the plot.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.33]{bs02a-wfxi.pdf}
\includegraphics[scale=0.33]{bs02b-wfz_2nd.pdf}
\end{center}
\caption{
{The valence wave functions vs $\xi$ with fixed values of $(k_\perp/m)^2$, for the first (left panel) and second (right panel) excited states,
with $B(1)/m = 0.22$ and $B(2)/m = 0.05$, respectively, obtained from (\ref{symbnak1}) with $\mu/m= 0.1$ and $\alpha$=6.437. }
}
\label{lf-g}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.33]{bs03a-wfkp.pdf}
\includegraphics[scale=0.33]{bs03b-wfg_2nd.pdf}
\end{center}
\caption{
{The valence wave functions vs $(k_\perp/m)^2$ with fixed values of $\xi$, for the first (left panel) and second (right panel) excited states,
with $B(1)/m = 0.22$ and $B(2)/m = 0.05$, respectively, obtained from (\ref{symbnak1}) with $\mu/m= 0.1$ and $\alpha$=6.437. }
\label{lf-g1}
}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.33]{bs04a-powerlaw.pdf}
\includegraphics[scale=0.33]{bs04b-powerlaw_2nd.pdf}
\end{center}
\caption{
The asymptotic $k_\perp$ behaviors of the first (left frame) and second (right frame) excited states
are shown, using the same label convention as given in Fig.~\ref{lf-g1}.
}
\label{lf-g2}
\end{figure}
From Eq. (\ref{val1}), one can obtain the asymptotic behaviors, $k_\perp\rightarrow \infty$, of LF valence wave function for
bound states, which are given by
\begin{equation}\label{asymptot}
\psi(\xi,{\bf k}_\perp)\rightarrow {k_\perp^{-4}} \,{C(\xi)}\, .
\end{equation}
Such behavior is explicitly shown in Fig.~\ref{lf-g2} for the first two excited states
(for ground-state, see Ref.~\cite{FSV2}).
It should be emphasized that independently of the value of $\xi$ the wave function is damped as $\sim k_\perp^4$.
We close this subsection by discussing the equivalence of the transverse-momentum amplitudes in
Minkowski and Euclidean spaces~\cite{SalPRC00}, respectively defined as
\begin{eqnarray}\label{tma}
\phi^T_{M}(\mathbf{k}_\perp)&\equiv&\int dk^0dk^3\Phi(k,p)= \frac12\int
dk^+dk^-\Phi(k,p)\;\; {\rm and}\\
\phi^T_E(\mathbf{k}_\perp)&\equiv&{\rm i}\int dk^0_Edk^3\Phi_E(k_E,p),
\label{tea} \end{eqnarray}
where $\Phi_E(k_E,p)$ is obtained from $\Phi(k,p)$ after applying the Wick rotation with
$k^0\to {\rm i} k^0_E$.
Notably, within NIR, one can easily prove that
$\phi^T_{M}(\mathbf{k}_\perp)=\phi^T_E(\mathbf{k}_\perp)$, since one can exploit the explicit
expression of the analytic dependence of the BS amplitude, as given in Eq. (\ref{NIR-bs}).
As a matter of fact,
choosing the rest frame, one can straightforwardly see that
the zeros of the Nakanishi denominator in the complex plane of $k_0$ are given by:
\begin{equation}
k_0=-\frac{M\,z'}{2}\pm\sqrt{\frac{M^2\,z'^2}{4}+\gamma' +\kappa^2+k_3^2+k_\perp^2-{\rm i} \epsilon} \, ,
\end{equation}
with $z'$ $\in[-1,1]$ and $\gamma'$ $\in[0,\infty]$. Therefore,
Eq. (\ref{NIR-bs}), as a function of complex $k_0$, has two cuts,
with branch-points at
\begin{equation}
k_{0\pm}^{\text{b}}=
\pm \left(\frac{M}{2}-\sqrt{\frac{M^2}{4}+\kappa^2}\right) \, .
\end{equation}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.7]{bs05-complexrot.pdf}
\end{center}
\caption{
Analytic structure of the BS amplitude in the complex plane of $k_0$,
showing the left- and right-hand cuts with the corresponding branch points
$k_{0\pm}^{\text{b}}$. The rotation path of the $k_0$-integration contour is
also shown for the transverse amplitude (\ref{tma}). }
\label{comprot}
\end{figure}
Recalling that $\kappa^2$ is positive { for bound states},
one can show that at the branch-point $ k_{0+}^{\text{b}}$ a cut
starts in the upper half-plane for
$\text{Re}\,k_0<0$, while at the branch-point $ k_{0-}^{\text{b}}$ the cut
is placed in the lower half-plane for
positive values of $\text{Re}\,k_0$, as shown in Fig. \ref{comprot}.
If, in Eq. (\ref{tma}), where the Minkowski space is adopted, one
considers the integration variable $k_0$ as a complex one, i.e. $k_0=|k_0|
e^{i\theta}$, and rotates the angle $\theta$
up to $90^o$, no singularities are crossed (cf. Fig. \ref{comprot}).
Furthermore, assuming that the BS amplitude drops out fast enough for
large complex $|k_0|$, the Cauchy theorem holds and the
Wick rotation \cite{WICK_54} can be applied for computing the transverse
amplitude. Namely, one can adopt a new integration path, along a purely
imaginary $k_0$, without dealing with any singular integrals. Consequently,
the Minkowski and Euclidean transverse amplitudes, given
by Eqs.~(\ref{tma}) and (\ref{tea}) are formally equivalent.
The quantitative comparison for the cases $\mu/m$=0.1 and 0.5,
with $\alpha_{gr}$ taken from Table \ref{comparison}
(recall that one has always $B(0)/m=1.9$), is presented in Fig. \ref{td}, showing
a very good agreement between the transverse amplitudes, within the accuracy of our numerical
approaches. It is worth noticing that such an equivalence gives an
additional confidence in NIR, since it should be emphasized that the Euclidean
solutions of BSE are not obtained by assuming the NIR for BS amplitudes. Therefore the
comparison in Fig. \ref{td} should be considered as a further check
of the reliability of the NIR itself, at least at the ladder level, besides the passed tests
for both eigenvalues \cite{KusPRD95,KusPRD97,carbonell1,FSV2} and scattering
lengths \cite{FSV3}.
Moreover, Fig. \ref{td} illustrates nice and general features of the transverse amplitudes,
that appear when the binding energies change. As a
matter of fact, the position of the node in the first
excited state moves toward smaller values of $k_\perp$
as the binding energies decreases, i.e.
from the left panel
($B(1)/m=0.22$)
to the right panel ($B(1)/m=0.0082$). Analogously, the amplitudes themselves
decrease more
quickly in momentum space. Both features can be explained by the increase
of size of the
bound state
when the binding energy decreases.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.33]{bs06a-phiEM4-mu01}
\includegraphics[scale=0.33]{bs06b-phiEM4-mu05}
\end{center}
\caption{Transverse momentum amplitudes $s-$wave states,
in Euclidean and Minkowski spaces, vs $k_\perp$, for both ground-
and first-excited states, and two values of $\mu/m$ and $\alpha_{gr}$ (as indicated in the insets). The amplitudes
$\phi_E^T$ and $\phi_M^T$, arbitrarily normalized to 1 at the origin, are not easily distinguishable.}
\label{td}
\end{figure}
\subsection{ Valence LF wave function in the impact-parameter space}
The transverse charge densities have been thoroughly discussed by Miller in Ref.~\cite{MilARNP10},
in close relation to the elastic electromagnetic form factor. Indeed, the transverse charge density
allows one to properly generalize the well-known non-relativstic relation between form factor
and density to a relativistic framework. As a matter of fact, it turns out that for a composite bosonic state,
the form factor $F(Q^2=-q^2)$ can be written as
\begin{equation}
F(Q^2)=\int d^2{\bf b} \,\rho({\bf b})\, \e^{-{\rm i} {\bf b}\cdot {\bf q}_\perp}
\, ,
\end{equation}
where
(i) the momentum transfer $q^\mu$ is evaluated in the Breit frame with $q^+=0$,
(ii) $Q^2={\bf q}_\perp^2$,
(iii) ${\bf b}$ belongs to the transverse plane, called IP space \cite{BurIJMP03}, and
(iv) $\rho({\bf b})$ is the IP density. It has to be pointed out that the IP density
is the sum of contributions from all the LF amplitudes of the Fock expansion
of the interacting-system state, such that
\begin{equation}
\rho({\bf b})=\rho_{\text{val}}({\bf b})+\,\,\text{higher Fock states densities}\,\cdots
.\end{equation}
The valence term is defined through the valence wave function in the IP space,
$\phi(\xi,{\bf b})$, as follows
\begin{equation} \label{rhoval}
\rho_{\text{val}}({\bf b})=\frac{1}{4\pi}
\int^1_0 \frac{d\xi}{\xi (1-\xi)^3} \,| \phi(\xi,{\bf b}/(1-\xi))|^2
\, .
\end{equation}
with normalization (cf. Eq. (\ref{pval}))
$\int d^2{\bf b}~\rho_{\text{val}}({\bf b})=P_{val}$. In Eq. (\ref{rhoval}), the IP-space valence wave
function is the 2D Fourier transform of $\psi(\xi,{\bf k_\perp})$, given by
\begin{equation} \label{phibgm}
\phi(\xi,{\bf b})=
\int\frac{ d^2{\bf k_\perp}}{(2\pi)^2}\; \psi(\xi,{\bf k_\perp})
{\rm e}^{{\rm i}{\bf k_
\perp}\cdot {\bf b}}
,\end{equation}
where {$\phi(\xi,{\bf b})$ results to be symmetric with respect to $1-2\xi$, as a consequence of the
already discussed symmetry of $g(\gamma,z;\kappa^2)$ under the transformation $z\to -z$.
Moreover, one can deduce the general behavior for large transverse separations, $b=|{\bf b}|$, as
illustrated in what follows. By performing the 2D Fourier transformation, the IP-space valence wave
function can be written within NIR for the $s-$wave state as follows
\begin{equation}
\phi(\xi,b)= \frac{\xi(1-\xi)}{4\pi\sqrt{2}}~F(\xi,b)
,\label{valbt}
\end{equation}
where
\begin{equation}\label{f(b,z)}
F(\xi,b)=\int^\infty_0 d\gamma~J_0(b \sqrt{\gamma})
\int_0^{\infty}d\gamma'~{g(\gamma',1-2\xi;\kappa^2)
\over [\gamma+\gamma' +\kappa^2+(1/2-\xi)^2M^2]^2}\, ,
\end{equation}
with $J_n(x)$ the integer-order Bessel function for $n=0$.
Also the integration over
$\gamma$ can be analytically carried out, leading to
\begin{equation}
F(\xi,b)
= b\,\int_0^{\infty}d\gamma~g(\gamma,1-2\xi;\kappa^2)
\frac{K_1\left(b\,\sqrt{\gamma +\kappa^2+(1/2-\xi)^2M^2}\right)}
{\sqrt{\gamma +\kappa^2+(1/2-\xi)^2M^2}},
\label{f1(b,z)}
\end{equation}
where $K_1(x)$ is the modified Bessel function of the second kind.
The function $F(\xi,b)$ exponentially drops out for $b \to \infty$.
Such a behavior can be
understood by a close analysis of Eq. (\ref{f1(b,z)}).}
First, from the physically-motivated request~\cite{MilARNP10} that $\phi(\xi,b)$ is
finite for $ b\to 0$ (see also \cite{FSV2}), such that
\begin{equation}
\phi(\xi,b=0)=
{\xi~(1-\xi)\over 4\pi \sqrt{2}}\int_0^{\infty}d\gamma~\frac{
g(\gamma,1-2\xi;\kappa^2)}
{\gamma+\kappa^2+(1/2-\xi)^2M^2} \, < \infty~~,
\label{valbt1}
\end{equation}
one can deduce that $g(\gamma,1-2\xi;\kappa^2)$ must vanish for
$\gamma \to \infty$. Therefore, the relevant interval of $\gamma$ in the
integral (\ref{f1(b,z)}) can be taken effectively finite.
Exploiting such an observation,
one can extract the driving exponential fall-off of $F(\xi,b)$
in the asymptotic limit $b \rightarrow \infty$.
In this limit $K_{1}(x)$ reads:
\begin{equation}\label{k1asymp}
K_{1}(x)|_{x\to\infty}\to\left( \frac{\pi}{2\,x} \right)^{\frac{1}{2}} \e^{-x} \, .
\end{equation}
The leading exponential behavior in the integral (\ref{f1(b,z)})
comes from values of $\e^{-b\,\sqrt{\gamma+\kappa^2+(\xi-1/2)^2M^2 }}$
[as seen from Eq.~\ref{k1asymp})] with
$\gamma$ close to $0$. Therefore,
\begin{equation}\label{f2(b,z)}
F(\xi,b)|_{b\to\infty}\to~\e^{-b\,\sqrt{\kappa^2+(\xi-1/2)^2M^2}}\,f(\xi,b) ,
\end{equation}
where the exponential fall-off is singled out and the reduced function
$f(\xi,b)$ should decrease more smoothly
for large values of $b$. It has to be pointed out that an exponential fall-off
is expected for bound states, since it is generated by
short range interactions, in analogy with the behavior found for the
non-relativistic two-dimensional case.
The above feature has been investigated by actually calculating $F(\xi,b)$, and in
turn $f(\xi,b)$, for both ground and first-excited states. In
Fig.~\ref{fig:impactpara2D},
$f(\xi,b)$ is presented for the ground (left) and
first-excited (right) states. In both cases, we have $\mu/m=0.1$
and $\alpha_{gr}$=6.437.
It is worth noting in the right panel the nice
node structure of the excited state in the whole $\{1-2\xi,b\}$ plane.
The tail of the function $F(\xi,b)$ with respect to $b$ is studied in more
detail through $f(\xi,b)$, putting in evidence that the steep fall-off of
the valence wave function, which is largely taken into account
by the leading exponential term, included in the definition (\ref{f2(b,z)}). This suggests
at most a polynomial behavior in $f(\xi,b)$, which is clearly seen in the figure for $b <15/m$.
\begin{figure}[thb]
\begin{center}
\includegraphics[scale=.13]{bs07a-asynground.pdf}
\includegraphics[scale=.13]{bs07b-asynexcited.pdf}
\end{center}
\caption{
The valence functions $f(\xi,b)$ in the impact parameter space. Left panel: the ground state, corresponding to
$B(0)$=$1.9m$, $\mu$=$0.1m$ and $\alpha_{gr}$=$6.437$.
Right panel: first-excited state, corresponding to $B(1)$=$0.22m$, $\mu$=$0.1 m$ and
$\alpha_{gr}$=6.437~.}\label{fig:impactpara2D}
\end{figure}
\section{Conclusions and Perspectives}
\label{END}
We have investigated both spectrum and excited states of the scalar Bethe-Salpeter equation, in ladder
approximation, by getting, for the first time, solutions directly in Minkowski space, within the
Nakanishi integral representation of the BS amplitude. A basic ingredient of our approach is the
exact projection of the BSE onto the null-plane (see, e.g. \cite{SalPRC00,FreFBS00}), that allows one to
master in a simple and very effective way the singularities typical of the BS formalism.
We have considered an $s$-wave interacting system composed by two massive scalars and interacting
through a massive scalar, extending the study of the ground state performed in Ref. \cite{FSV2} (see Ref.
\cite{carbonell1} for an analogous study within the explicitly-covariant LF approach), and
carefully analyzing the valence wave function both in Minkowski and impact-parameter spaces.
Within the numerical accuracy of our approach, we have found a finite
number of excited states for non-zero exchanged-scalar mass, and we have successfully
compared our results with the corresponding ones
obtained in the Euclidean space, where obviously NIR is not assumed. A detailed study of the
valence wave function structure has been carried out in the plane $(\xi,k_\perp)$, showing the
expected node structure of the first and second excited states. Furthermore,
our investigation, both analytic and quantitative, of the transverse-momentum amplitude has allowed
to remarkably show the
equivalence of the quantity evaluated both in Euclidean and Minkowski spaces. This further strengthens
the reliability of the approach based on NIR for solving the BSE, already applied to
fermionic systems \cite{carbonell3}, kernels beyond the ladder one \cite{carbonell2} and in the
zero-energy scattering case \cite{FSV3}.
Finally, we also explored the asymptotic properties of the impact-parameter space valence wave function
for large transverse distances, where an exponential fall-off was singled out
(similar to the non-relativistic case in the 3D Euclidean space) and quantitatively tested for
the excited states.
In perspective, the present study encourages the extension of the approach based on the NIR to excited
states of actual physical systems, as well as to explore results obtained for other dynamical quantities within a
wider and deeper comparison between Minkowski and Euclidean calculations.
{\it Acknowledgments.}
We thank the Brazilian agencies Coordena\c c\~ao de Aperfei\c coamento de Pessoal de N\'ivel Superior
(CAPES), Conselho Nacional de Desenvolvimento Cient\'ifico e Tecnol\'ogico (CNPq) and
Funda\c c\~ao de Amparo \`a Pesquisa do Estado de S\~ao Paulo (FAPESP) for partial support.
MV and GS acknowledge the warm hospitality of
the Instituto Tecnol\'ogico de Aeron\'autica, S\~ao Jos\'e dos Campos where part of this work was performed.
|
\section{Introduction}
With the development of high-density data storage technologies, while the codes are defined as usual over some discrete symbol alphabet, their reading from the channel is performed as overlapping pairs of symbols. A channel whose outputs are overlapping pairs of symbols is called a symbol-pair channel. A pair-error is defined as a pair-read in which one or more of the symbols are read in error. The design of codes to protect efficiently against a certain number of pair-errors is significant.
Cassuto and Blaum first studied codes that protect against pair-errors in \cite{CB}, as well as pair-error correctability conditions, code construction and decoding, and lower and upper bounds on code sizes. Later, Cassuto and Litsyn \cite{CL} gave algebraic cyclic code constructions of symbol-pair codes and asymptotic bounds on code rates. They also showed the existence of pair-error codes with rates strictly higher than those of the codes in the Hamming metric with the same relative distance. Yaakobi et al.\ proposed efficient decoding algorithms for cyclic symbol-pair codes in \cite{YBS,YBH16}.
Chee et al.\ in \cite{CJKWY} established a Singleton-type bound on symbol-pair codes and constructed infinite families of symbol-pair codes that meet the Singleton-type bound, which are called maximum distance separable symbol-pair codes or MDS symbol-pair codes for short. The construction of MDS symbol-pair codes is interesting since the codes have the best pair-error correcting capability for fixed length and dimension. The authors in \cite{CJKWY} made use of interleaving and graph theoretic concepts as well as combinatorial configurations to construct MDS symbol-pair codes. Kai et al.\ \cite{KZL} constructed MDS symbol-pair codes from cyclic and constacyclic codes.
Classical MDS codes are MDS symbol-pair codes \cite{CJKWY} and other known families of MDS $(n,d)_{q}$ symbol-pair codes are shown in Table \ref{tab}.
\begin{table}[h]\caption{Known families of MDS symbol-pair codes\label{tab}}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
$d$ & $q$ & $n$ & Reference \\\hline
$2$, $3$ & $q\geq2$ & $n\geq2$ &\cite{CJKWY} \\\hline
$4$& $q\geq 2$ & $n\geq 2$ & \cite{CJKWY} \\\hline
& even prime power &$n\leq q+2$&\cite{CJKWY}\\\cline{2-4}
& odd prime &$5\leq n\leq 2q+3$&\cite{CJKWY}\\\cline{2-4}
$5$& prime power &$n|q^{2}-1$, $n>q+1$&\cite{KZL}\\\cline{2-4}
& prime power &$n=q^{2}+q+1$&\cite{KZL}\\\cline{2-4}
& prime power, $q\equiv 1\pmod 3$ &$n=\frac{q^{2}+q+1}{3}$&\cite{KZL}\\\hline
\multirow{2}{*}{6}&prime power &$n=q^{2}+1$&\cite{KZL}\\\cline{2-4}
&odd prime power& $n=\frac{q^{2}+1}{2}$ &\cite{KZL}\\\hline
$7$ &odd prime&$n=8$&\cite{CJKWY}\\
\hline
\end{tabular}
\end{center}
\end{table}
In this paper, we present new constructions of linear MDS symbol-pair codes over the finite field $\mathbb{F}_{q}$ and obtain the following three new families:
\begin{enumerate}
\item there exists a linear MDS $(n,5)_q$ symbol-pair code if and only if $5\le n \le q^2+q+1$;
\item there exists a linear MDS $(n,6)_{q}$ symbol-pair code for $q\geq3$ and $6\leq n\leq q^{2}+1$;
\item there exists a linear MDS $(n,d+2)_{q}$ symbol-pair code for general $n,d$ satisfying $7\le d+2\leq n\le q+\lfloor 2\sqrt{q}\rfloor+\delta(q)-3$, where
\[
\delta(q)=\left\{
\begin{array}{ll}
$0$, & \hbox{if $q=p^a,\,a\ge 3$, $a$ odd and $p\,|\,\lfloor 2\sqrt{q}\rfloor$;} \\
$1$, & \hbox{otherwise.}
\end{array}
\right.
\]
\end{enumerate}
Compared with the known MDS symbol-pair codes, the MDS symbol-pair codes constructed in this paper provide a much larger range of parameters.
This paper is organized as follows. Basic notations and definitions are given in Section \ref{sec2}. In Section \ref{sec3}, we construct MDS symbol-pair codes with pair-distance $5$. And in Section \ref{sec4} we derive MDS symbol-pair codes with pair-distance $6$ from projective geometry. In Section \ref{sec5}, by using elliptic curves, we give the construction of MDS symbol-pair codes for any pair-distance satisfying certain conditions. Section \ref{sec6} concludes the paper.
\section{Preliminaries}\label{sec2}
Let $\Sigma$ be the alphabet consisting of $q$ elements. Each element in $\Sigma$ is called a symbol. For a vector ${\bf u}=(u_{0},u_{1},\cdots,u_{n-1})$ in $\Sigma^{n}$, we define the symbol-pair read vector of ${\bf u}$ as
$$\pi({\bf u})=((u_{0},u_{1}),(u_{1},u_{2}),\cdots,(u_{n-1},u_{0})).$$
Throughout this paper, let $q$ be a prime power and $\mathbb{F}_{q}$ be the finite field containing $q$ elements. We will focus on vectors over $\mathbb{F}_{q}$, so $\Sigma=\mathbb{F}_{q}$.
It is obvious that each vector ${\bf u}$ in $\mathbb{F}_{q}^{n}$ has a unique symbol-pair read vector $\pi({\bf u})$ in $(\mathbb{F}_{q} \times \mathbb{F}_{q})^{n}$. For two vectors ${\bf u}$, ${\bf v}$ in $\mathbb{F}_{q}^{n}$, the pair-distance between ${\bf u}$ and ${\bf v}$ is defined as
$$d_{p}({\bf u},{\bf v}):=|\lbrace0\le i\le n-1:(u_{i},u_{i+1})\ne (v_{i},v_{i+1})\rbrace|,$$
where the subscripts are reduced modulo $n$. And for any vector ${\bf u}$ in $\mathbb{F}_{q}^{n}$, the pair-weight of ${\bf u}$ is defined as
$$w_{p}({\bf u})=|\lbrace0\le i\le n-1:(u_{i},u_{i+1})\ne (0,0)\rbrace|,$$ where the subscripts are reduced modulo $n$.
The following relationship between the pair-distance and the Hamming distance was shown in \cite{CB}.
\begin{proposition}\label{prop1}
Let ${\bf u},{\bf v}\in \mathbb{F}_{q}^n$ be such that $0<d_{H}({\bf u},{\bf v})<n$, where $d_{H}$ denotes the Hamming distance, we have $$d_{H}({\bf u},{\bf v})+1\leq d_{p}({\bf u},{\bf v})\leq 2d_{H}({\bf u},{\bf v}).$$
\end{proposition}
Meanwhile, the following relationship between the pair-distance and the pair-weight holds.
\begin{proposition}\label{prop2}
For all ${\bf u},{\bf v}\in \mathbb{F}_{q}^n$, $d_{p}({\bf u},{\bf v})=w_{p}({\bf u}-{\bf v}).$
\end{proposition}
A code $\mathcal{C}$ over $\mathbb{F}_{q}$ of length $n$ is a nonempty subset of $\mathbb{F}_{q}^{n}$ and the elements of $\mathcal{C}$ are called codewords. The minimum pair-distance of $\mathcal{C}$ is defined as
$$d_{p}(\mathcal{C})=\min\lbrace{d_{p}(\bf u},{\bf v})\mid{\bf u},{\bf v}\in \mathcal{C},{\bf u}\neq {\bf v}\rbrace,$$
and the size of $\mathcal{C}$ is the number of codewords it contains. In general, a code $\mathcal{C}$ over $\mathbb{F}_{q}$ of length $n$, size $M$ and minimum pair-distance $d$ is called an $(n,M,d)_{q}$ symbol-pair code. Besides, if $\mathcal{C}$ is a subspace of $\mathbb{F}_{q}^{n}$, then $\mathcal{C}$ is called a linear symbol-pair code. When $\mathcal{C}$ is a linear code, the minimum pair-distance of $\mathcal{C}$ is the smallest pair-weight of nonzero codewords of $\mathcal{C}$. And in this paper we consider linear symbol-pair codes over $\mathbb{F}_{q}$
The minimum pair-distance $d$ is an important parameter in determining the error-correcting capability of $\mathcal{C}$. Thus it is significant to find symbol-pair codes of fixed length $n$ with pair-distance $d$ as large as possible. In \cite{CJKWY}, the authors proved the following Singleton-type bound.
\begin{theorem}[Singleton bound]
Let $q\geq 2$ and $2\leq d\leq n$. If $\mathcal{C}$ is an $(n,M,d)_{q}$ symbol-pair code, then $M\leq q^{n-d+2}$.
\end{theorem}
A symbol-pair code achieving the Singleton bound is a maximum distance separable (MDS) symbol-pair code. An MDS $(n,M,d)_{q}$ symbol-pair code is simply called an MDS $(n,d)_{q}$ symbol-pair code.
In \cite{KZL}, the authors presented the following theorem.
\begin{theorem}\label{thmcite}
Let $\mathcal{C}$ be an $[n,n-d_{H},d_{H}]$ linear code over $\mathbb{F}_{q}$. If the pair-distance $d\geq d_{H}+2$, then $\mathcal{C}$ is an MDS $(n,d_{H}+2)_{q}$ symbol-pair code.
\end{theorem}
Now we are ready to give a sufficient condition for the existence of linear MDS symbol-pair codes in the following theorem.
\begin{theorem}\label{thmAMDS}
There exists a linear MDS $(n,d_{H}+2)_{q}$ symbol-pair code $\mathcal{C}$ if there exists a matrix with $d_{H}$ rows and $n\ge d_{H}+2\ge 4$ columns over $\mathbb{F}_{q}$, denoted by $H=[H_{0},H_{1},\cdots,H_{n-1}]$, where $H_{i}$ $(0\leq i\leq n-1)$ is the $i$-th column of $H$, satisfying:
\begin{itemize}
\item[1.] any $d_{H}-1$ columns of $H$ are linearly independent;
\item[2.] there exist $d_{H}$ linearly dependent columns;
\item[3.] any $d_{H}$ cyclically consecutive columns are linearly independent, i.e., $H_{i},H_{i+1},\cdots,H_{i+d-1}$ are linearly independent for $0\leq i\leq n-1$, where the subscripts are reduced modulo $n$.
\end{itemize}
\end{theorem}
\begin{proof}
Let $\mathcal{C}$ be the linear code with parity check matrix $H$. The first two conditions indicate that $\mathcal{C}$ is an $[n,n-d_{H},d_{H}]$ linear code with size $q^{n-d_{H}}$. Consider any codeword $c\in \mathcal{C}$ with $d_{H}$ nonzero coordinates. From Propositions \ref{prop1}, \ref{prop2} and the third condition, we can see that the $d_{H}$ nonzero coordinates are not in cyclically consecutive positions, and thus $w_{p}(c)\ge d_{H}+2$. For any other codeword $c'\in \mathcal{C}$, we must have the Hamming weight $w_{H}(c')\ge d_{H}+1$ and $w_{p}(c')\ge d_{H}+2$. Hence the pair-distance $d\ge d_{H}+2$ and $\mathcal{C}$ is an MDS $(n,d_{H}+2)_{q}$ symbol-pair code.
\end{proof}
\section{MDS symbol-pair codes with pair-distance 5}\label{sec3}
We construct MDS $(n,5)_{q}$ symbol-pair codes in this section. According to Theorem \ref{thmAMDS}, what we need is to construct a matrix $H$ with $3$ rows and $n$ columns over $\mathbb{F}_{q}$ satisfying the following conditions:
\begin{itemize}
\item[1.] any two columns of $H$ are linearly independent;
\item[2.] there exist three linearly dependent columns;
\item[3.] any three cyclically consecutive columns are linearly independent.
\end{itemize}
\begin{lemma} \label{MDS5}
A linear MDS $(n,5)_{q}$ symbol-pair code, where $q$ is a prime power, exists only if the length $n$ ranges from $5$ to $q^2+q+1$.
\end{lemma}
\begin{proof}
From Proposition \ref{prop1}, we know that a symbol-pair code with the minimum pair-distance $d=5$ must have the minimum Hamming distance $d_{H}\ge 3$. Thus the parity check matrix of the code must satisfy the first condition above and the conclusion follows.
\end{proof}
In this section we aim to show the existence of MDS $(n,5)_{q}$ symbol-pair codes for every $5\le n\le q^2+q+1$. We first describe how to construct a full matrix $H(q)$ of size $3\times (q^2+q+1)$ and then we mention how to adjust $H(q)$ to get a matrix $H(q;n)$ of size $3\times n$ for any $n$, $5\le n \le q^2+q+1$. Choose the column vectors of $H(q)$ from the following $q^2+q+1$ vectors: $\{(0,0,1)^{\textup{T}}$, $(0,1,c)^{\textup{T}}$ for $c\in \mathbb{F}_{q}$, $(1,a,b)^{\textup{T}}$ for $a,b\in\mathbb{F}_{q}\}$. In this way the first two conditions above are guaranteed, and we only need to order these vectors in a proper way to meet the third condition.
First we deal with the case when $q$ is odd. Denote the elements in $\mathbb{F}_{q}$ in an arbitrary order $\{x_0,x_1,\dots,x_{q-1}\}$. As a preparatory step, we partition the $q^2$ vectors of the form $\{(1,a,b)^{\textup{T}},a,b\in\mathbb{F}_{q}\}$ into $q$ disjoint blocks $B_i=\{(1,a,a^2+x_i)^{\textup{T}},a\in\mathbb{F}_{q}\}$ for $0\le i <q$. We give an order of the vectors within $B_i$. Set the first vector to be $(1,x_i,x_i^2+x_i)^{\textup{T}}$, the next to be $(1,x_{i+1},x_{i+1}^2+x_i)^{\textup{T}}$, and then the next to be $(1,x_{i+2},x_{i+2}^2+x_i)^{\textup{T}}\dots$ until finally the vector $(1,x_{i+q-1},x_{i+q-1}^2+x_i)^{\textup{T}}$, where subscripts are reduced modulo $q$. That is,
\begin{equation*}
B_i=
\left[
\begin{array}{ccccc}
1 & 1 & 1 & \cdots & 1\\
x_i & x_{i+1} & x_{i+2} & \cdots & x_{i+q-1} \\
x_i^2+x_i & x_{i+1}^2+x_i & x_{i+2}^2+x_i & \cdots & x_{i+q-1}^2+x_i
\end{array}
\right].
\end{equation*}
Then we construct the matrix $H(q)$ as follows. List all the blocks $B_i$ defined above in the reverse order of their subscripts: $B_{q-1}$, $B_{q-2},\dots$, $B_1$, $B_0$. Between any pair of consecutive blocks $B_{i+1}$ and $B_{i}$, insert a vector $(0,1,2x_{i})^{\textup{T}}$. Note that the pair of $B_0$ and $B_{q-1}$ is also considered, and the vector $(0,1,2x_{q-1})^{\textup{T}}$ should be inserted between them, which is further restricted to be the first column of $H(q)$. Finally the vector $(0,0,1)^{\textup{T}}$ could be placed anywhere and we just set it as the last column. That is,
\begin{equation*}
H(q)=
\left[
\begin{array}{ccccccccccccccc}
0 & & 0 & & 0 & & \dots & &0 & & \cdots & & 0 & & 0\\
1 & B_{q-1} & 1 & B_{q-2} & 1 & B_{q-3} & \dots & B_{i+1} &1 &B_i & \cdots & B_1 & 1 & B_0 & 0\\
2x_{q-1} & & 2x_{q-2} & & 2x_{q-3} & & \dots & &2x_i & & \cdots & & 2x_0 & & 1
\end{array}
\right].
\end{equation*}
\begin{proposition} \label{independent}
Every three cyclically consecutive columns of $H(q)$ are linearly independent over $\mathbb{F}_q$.
\end{proposition}
\begin{proof}
For three consecutive columns within a block $B_i$, $0\leq i\leq q-1$, we have
\begin{equation*}\left|\begin{array}{ccc}1 & 1 & 1\\x_{a-1} & x_a & x_{a+1}\\ x_{a-1}^2+x_i & x_a^2+x_i & x_{a+1}^2+x_i \end{array}\right| = \left|\begin{array}{ccc}1 & 1 & 1\\x_{a-1} & x_a & x_{a+1}\\ x_{a-1}^2 & x_a^2 & x_{a+1}^2 \end{array}\right| =(x_{a-1}-x_a)(x_{a}-x_{a+1})(x_{a+1}-x_{a-1})\ne 0.\end{equation*}
For three consecutive columns with a vector $(0,1,2x_j)^{\textup{T}}$ in the middle, we have
\begin{equation*}\left|\begin{array}{ccc}1 & 0 & 1\\x_j & 1 & x_j\\ x_j^2+x_{j+1} & 2x_j & x_j^2+x_{j} \end{array}\right| = \left|\begin{array}{ccc}1 & 0 & 0\\x_j & 1 & 0\\ x_j^2+x_{j+1} & 2x_j & x_{j}-x_{j+1} \end{array}\right| = x_{j}-x_{j+1} \ne0.\end{equation*}
For three consecutive columns containing a vector $(0,1,2x_j)^{\textup{T}}$, which is not in the middle, we have either
\begin{equation}\label{check1}\left|\begin{array}{ccc}0 & 1 & 1\\1 & x_j & x_{j+1}\\ 2x_j & x_j^2+x_{j} & x_{j+1}^2+x_{j} \end{array}\right| = -(x_j-x_{j+1})^2 \ne0, \end{equation}
or
\begin{equation}\label{check2}\left|\begin{array}{ccc}1 & 1 & 0\\x_{j-1} & x_j & 1\\ x_{j-1}^2+x_{j+1} & x_j^2+x_{j+1} & 2x_j \end{array}\right| = (x_j-x_{j-1})^2 \ne0.\end{equation}
Finally, it is easy to see that every three consecutive columns in $H(q)$ containing the vector $(0,0,1)^{\textup{T}}$ are linearly independent over $\mathbb{F}_q$.
\end{proof}
We now focus on the case when $q$ is even and $q\neq 2,4$. The general outline is similar. Let $\omega$ be a primitive element in $\mathbb{F}_q$. Denote the elements in $\mathbb{F}_{q}$ in an arbitrary order $\{x_0,x_1,\dots,x_{q-1}\}$, with the only constraint that the first several elements are preset to be $x_0=0$, $x_1=1$, $x_2=\omega$, $x_3=\omega^2$, $x_4=\omega+1$, $x_5=\omega^2+\omega$. First define the blocks $B_i$ in the same way as above and list all the blocks $B_i$ in the reverse order of their subscripts: $B_{q-1},B_{q-2},\dots,B_1,B_0$. Now we need to find out which vector of the form $(0,1,y)^{\textup{T}}$ can be inserted between the blocks $B_{j+1}$ and $B_{j}$. Recall the proof of Proposition \ref{independent}. It can be checked that the choice of the value $y$ only affects equations (\ref{check1}) and (\ref{check2}). So for the validity of that proof we only require that
\begin{equation*}\left|\begin{array}{ccc}0 & 1 & 1\\1 & x_j & x_{j+1}\\ y & x_j^2+x_{j} & x_{j+1}^2+x_{j} \end{array}\right| = (x_{j+1}-x_j)(y-x_j-x_{j+1}) \ne0,\end{equation*}
and
\begin{equation*}\left|\begin{array}{ccc}1 & 1 & 0\\x_{j-1} & x_j & 1\\ x_{j-1}^2+x_{j+1} & x_j^2+x_{j+1} & y \end{array}\right| = (x_j-x_{j-1})(y-x_j-x_{j-1}) \ne0.\end{equation*}
That is, $y$ could be any value except for $x_j+x_{j-1}$ and $x_j+x_{j+1}$. An explicit insertion scheme seems hard to be expressed in an easy form, however, we can show that a proper insertion scheme surely exists. Construct a bipartite graph. The first part of the vertices corresponds to $\mathbb{F}_{q}$. The second part of the vertices is the set $\{L_j:0\le j < q\}$, where the symbol $L_j$ indicates the location between the blocks $B_{j+1}$ and $B_j$. $y\in\mathbb{F}_{q}$ is connected to $L_j$ if and only if the vector $(0,1,y)^{\textup{T}}$ could be inserted in the location $L_j$, i.e. $y\neq x_j+x_{j-1}$ and $y\neq x_j+x_{j+1}$. A perfect matching in this bipartite graph corresponds to a proper insertion scheme.
Following the analysis above, we can find that the degree of every vertex in the second part is exactly $q-2$. Recall that we have preset $x_0=0$, $x_1=1$, $x_2=\omega$, $x_3=\omega^2$, $x_4=\omega+1$, $x_5=\omega^2+\omega$. Thus we have:
$\bullet$ $L_1$ is connected to every $y\in\mathbb{F}_{q}$ except for $1$ and $\omega+1$;
$\bullet$ $L_2$ is connected to every $y\in\mathbb{F}_{q}$ except for $\omega+1$ and $\omega^2+\omega$;
$\bullet$ $L_3$ is connected to every $y\in\mathbb{F}_{q}$ except for $\omega^2+\omega$ and $\omega^2+\omega+1$; and
$\bullet$ $L_4$ is connected to every $y\in\mathbb{F}_{q}$ except for $\omega^2+\omega+1$ and $\omega^2+1$.
So, even only among these four vertices, we can deduce that every $y\in\mathbb{F}_{q}$ is connected to at least two of them. So we have
$\bullet$ the neighbourhood of every no more than $q-2$ vertices from the second part is of size at least $q-2$;
$\bullet$ the neighbourhood of every $q-1$ or $q$ vertices from the second part is of size $q$.
Therefore the famous Hall's theorem \cite{Hall} guarantees a perfect matching in this bipartite graph, which corresponds to a proper insertion scheme.
However, the case $q=4$ is listed as a separated case since the framework above using Hall's theorem would fail. To follow a similar framework, the order within a block needs some slight modifications and then a proper insertion scheme comes along. We shall just list the desired $3\times21$ matrix $H(4)$ instead of tedious explanations.
{\scriptsize
\begin{equation*}
H(4)=
\left[
\begin{array}{ccccccccccccccccccccc}
0 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 0 \\
1 & 0 & 1 & \omega & \omega+1 & 1 & \omega+1 & \omega & 1 & 0 & 1 & 0 & \omega+1 & \omega & 1 & 1 & 1 & \omega & \omega+1 & 0 & 0 \\
0 & 0 & 1 & \omega+1 & \omega & \omega+1 & \omega+1 & \omega & 0 & 1 & \omega & \omega & 0 & 1 & \omega+1 & 1 & \omega & 0 & 1 & \omega+1 & 1
\end{array}
\right].
\end{equation*}
}
Up till now we have constructed the matrix $H(q)$ of size $3\times (q^2+q+1)$ for every prime power $q\ge3$. Next we discuss how to adjust $H(q)$ to get a $3\times n$ matrix $H(q;n)$ for every $n$, $5\le n\le q^2+q+1$. Denote $n=\alpha(q+1)+\beta$, where $0\le \beta \le q$. There are certainly lots of methods to get such a desired matrix and we offer one as follows.
$\bullet$ If $\beta\neq 2$, select the first $n-1$ columns of $H(q)$, then add the vector $(0,0,1)^{\textup{T}}$.
$\bullet$ If $\beta=2$, select the first $n-1$ columns of $H(q)$, then insert the vector $(0,0,1)^{\textup{T}}$ as the new third column.
The case $\beta=2$ is separated since if we still abide by the first rule then we will come across a triple of the form $\{(0,1,x)^{\textup{T}},(0,0,1)^{\textup{T}},(0,1,y)^{\textup{T}}\}$ which is certainly not independent.
The validity of the construction of the $3\times n$ matrix can be easily inferred from Proposition \ref{independent} plus some further simple checks on those triples containing the vector $(0,0,1)^{\textup{T}}$, and the two triples of the form $\{(0,1,a)^{\textup{T}},(0,1,b)^{\textup{T}},(1,c,d)^{\textup{T}}\}$ (in the $\beta=2$ case).
As illustrative examples, for $q=5$ we list the following matrices: the full matrix $H(5)$ of size $3\times 31$, the adjusted matrix $H(5;13)$ (corresponding to $\beta\neq 2$) and $H(5;14)$ (corresponding to $\beta=2$).
{\scriptsize
\begin{equation*}
H(5)=
\left[
\begin{array}{ccccccccccccccccccccccccccccccc}
0 & 1 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 1 & 0 \\
1 & 4 & 0 & 1 & 2 & 3 & 1 & 3 & 4 & 0 & 1 & 2 & 1 & 2 & 3 & 4 & 0 & 1 & 1 & 1 & 2 & 3 & 4 & 0 & 1 & 0 & 1 & 2 & 3 & 4 & 0 \\
3 & 0 & 4 & 0 & 3 & 3 & 1 & 2 & 4 & 3 & 4 & 2 & 4 & 1 & 1 & 3 & 2 & 3 & 2 & 2 & 0 & 0 & 2 & 1 & 0 & 0 & 1 & 4 & 4 & 1 & 1
\end{array}
\right],
\end{equation*}}
{\scriptsize
\begin{equation*}
H(5;13)=
\left[
\begin{array}{ccccccccccccc}
0 & 1 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 1 & 0 \\
1 & 4 & 0 & 1 & 2 & 3 & 1 & 3 & 4 & 0 & 1 & 2 & 0 \\
3 & 0 & 4 & 0 & 3 & 3 & 1 & 2 & 4 & 3 & 4 & 2 & 1
\end{array}
\right],
H(5;14)=
\left[
\begin{array}{cccccccccccccc}
0 & 1 & 0 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 1 & 1 & 0 \\
1 & 4 & 0 & 0 & 1 & 2 & 3 & 1 & 3 & 4 & 0 & 1 & 2 & 1 \\
3 & 0 & 1 & 4 & 0 & 3 & 3 & 1 & 2 & 4 & 3 & 4 & 2 & 4
\end{array}
\right].
\end{equation*}}
Finally, for the case $q=2$, we list the matrices $H(2)$, $H(2;5)$, $H(2;6)$ as follows.
\begin{equation*}
H(2)=
\left[
\begin{array}{ccccccc}
1 & 0 & 0 & 1 & 0 & 1 & 1 \\
0 & 1 & 0 & 1 & 1 & 1 & 0 \\
0 & 0 & 1 & 0 & 1 & 1 & 1
\end{array}
\right],
H(2;5)=
\left[
\begin{array}{ccccc}
1 & 0 & 0 & 1 & 1 \\
0 & 1 & 0 & 1 & 0 \\
0 & 0 & 1 & 1 & 1
\end{array}
\right],
H(2;6)=
\left[
\begin{array}{cccccc}
1 & 0 & 0 & 1 & 0 & 1 \\
0 & 1 & 0 & 1 & 1 & 0 \\
0 & 0 & 1 & 1 & 1 & 1
\end{array}
\right].
\end{equation*}
So far we have finished the construction of MDS $(n,5)_q$ symbol-pair codes for any prime power $q\ge2$ and $5\le n \le q^2+q+1$. The construction, together with Lemma \ref{MDS5}, leads to the following theorem.
\begin{theorem}
There exists a linear MDS $(n,5)_q$ symbol-pair code, where $q$ is a prime power, if and only if the length $n$ ranges from $5$ to $q^2+q+1$.
\end{theorem}
\section{MDS symbol-pair codes from projective geometry}\label{sec4}
Let $V(r+1,q)$ be a vector space of rank $r+1$ over $\mathbb{F}_{q}$. The projective space $PG(r,q)$ is the geometry whose points, lines, planes, $\cdots$, hyperplanes are the subspaces of $V(r+1,q)$ of rank $1,2,3,\cdots,r$, respectively. The dimension of a subspace of $PG(r,q)$ is one less than the rank of a subspace of $V(r+1,q)$.
Label each point of $PG(r,q)$ as $\langle(a_{0},a_{1},\cdots,a_{r})\rangle$, the subspace spanned by a nonzero vector $(a_{0},a_{1},\cdots,a_{r})$, where $a_{i}\in \mathbb{F}_{q}$ for $0\leq i\leq r$. Since these coordinates are defined only up to multiplication by a nonzero scalar $\lambda\in \mathbb{F}_{q}$ (here $\langle(\lambda a_{0},\lambda a_{1},\cdots,\lambda a_{r})\rangle=\langle(a_{0},a_{1},\cdots,a_{r})\rangle$), we refer to $a_{0},a_{1},\cdots,a_{r}$ as homogeneous coordinates. Thus, there are a total of $(q^{r+1}-1)/(q-1)$ points in $PG(r,q)$. For an integer $r\geq 2$, if we choose $n\geq r+3$ points in $PG(r,q)$ and regard them as column vectors of a matrix $H$, then from Theorem \ref{thmAMDS} we have the following theorem.
\begin{theorem}\label{themcon}
There exists a linear MDS $(n,r+3)_{q}$ symbol-pair code if there exists a set $\mathcal{S}$ of $n\geq r+3\ge 5$ points of $PG(r,q)$ satisfying the following conditions:
\begin{itemize}
\item[1.] any $r$ points from $\mathcal{S}$ generate a hyperplane in $PG(r,q)$;
\item[2.] there exist $r+1$ points in $\mathcal{S}$ lying on a hyperplane;
\item[3.] if the $n$ points are ordered, say $\mathcal{P}_{0},\mathcal{P}_{1},\cdots,\mathcal{P}_{n-1}$, then any $r+1$ cyclically consecutive points do not lie on a hyperplane, i.e., $\mathcal{P}_{i},\mathcal{P}_{i+1},\cdots,\mathcal{P}_{i+r}$, where the subscripts are reduced modulo $n$, do not lie on a hyperplane for $0\leq i\leq n-1$.
\end{itemize}
\end{theorem}
Here we consider the case $r=3$.
\begin{definition}[]
A set $\mathcal{O}$ of points of $PG(3,q)$ is called an ovoid provided it satisfies the following conditions:
\begin{itemize}
\item[1.] each line meets $\mathcal{O}$ in at most two points;
\item[2.] through each point of $\mathcal{O}$ there are $q+1$ lines, each of which meets $\mathcal{O}$ in exactly one point, and all of them lie on a plane.
\end{itemize}
\end{definition}
The following two lemmas can be found in \cite{P09}.
\begin{lemma}
Each ovoid has $q^{2}+1$ points.
\end{lemma}
\begin{lemma}\label{lemPG1}
Each plane meets $\mathcal{O}$ either in one point or in $q+1$ points.
\end{lemma}
We can easily derive the following lemma.
\begin{lemma}\label{thmPG2}
For an ovoid $\mathcal{O}$ in $PG(3,q)$, there exist $q+1$ planes, each of which contains $q+1$ points in $\mathcal{O}$. Moreover, these planes intersect in a common line in $\mathcal{O}$ and cover all points of $\mathcal{O}$.
\end{lemma}
\begin{proof}
Fix two arbitrary points $A,B \in \mathcal{O}$, and then choose a point $P$ from $\mathcal{O}\setminus \lbrace A,B\rbrace$. By Lemma \ref{lemPG1}, the plane formed by $A,B,P$, which we denote by $ABP$, must meet $\mathcal{O}$ in $q+1$ points. Next, choose a point $Q\in\mathcal{O}$ which is not on $ABP$. Then, again, we get
a plane $ABQ$ which also meets $\mathcal{O}$ in $q+1$ points. If we continue in this way, we can get $q+1$ planes, each of which contains $q+1$ points of $\mathcal{O}$. These planes intersect in a common line which meets $\mathcal{O}$ in the points $A,B$.
\end{proof}
We can now state our construction.
\begin{theorem}\label{thm1}
Let $q\geq5$ be an odd prime power. Then there exist linear MDS $(n,6)_{q}$ symbol-pair codes for all $n$, $6\leq n\leq q^{2}+1$.
\end{theorem}
\begin{proof} Let $\mathcal{O}$ be an ovoid in $PG(3,q)$ and $\pi_{0},\pi_{1},\cdots,\pi_{q}$ be the planes described in Lemma \ref{thmPG2}. Moreover, let the intersection of $\pi_{0},\pi_{1},\cdots,\pi_{q}$ meets $\mathcal{O}$ in the points $A$ and $B$. For convenience, denote the plane formed by points $P,Q,R$ by $PQR$ and denote the set of the points lying in a set, say $\Omega$, but not on the plane $PQR$ by $\Omega\setminus PQR$. For four ordered points $P,Q,R,S$, we say $S$ is a \emph{proper} point if $S$ does not lie on the plane $PQR$. In other words, we say $S$ is a \emph{proper} point if $S$ does not lie on the plane formed by the three points ordered right ahead of it.
We now consider the three conditions stated in Theorem \ref{themcon}. It is clear that, for the points of $\mathcal{O}$, the first condition is inherently satisfied and the second condition can be easily satisfied. Thus, the points of $\mathcal{O}$ simply need to be ordered such that any four cyclically consecutive points do not lie on a plane. To attain this goal, we discuss it in two parts. First we order $n$ ($6\leq n\leq q^{2}+1$) points of $\mathcal{O}$ as $\mathcal{P}_{0},\cdots,\mathcal{P}_{n-1}$ and make sure that any four consecutive points do not lie on a plane, i.e., $\mathcal{P}_{i},\mathcal{P}_{i+1},\mathcal{P}_{i+2},\mathcal{P}_{i+3}$ do not lie on a plane for $0\leq i\leq n-4$. On this basis, we then adjust the order to make sure that any four cyclically consecutive points do not lie on a plane, i.e., $\mathcal{P}_{i},\mathcal{P}_{i+1},\mathcal{P}_{i+2},\mathcal{P}_{i+3}$ do not lie on a plane for $0\leq i\leq n-1$.
\begin{figure}[h]
\centering
\begin{picture}(200,160)
\thicklines
\multiput(20,115)(0,-15){2}{\circle*{2}}
\multiput(20,30)(0,15){4}{\circle*{2}}
\multiput(15,88)(5,0){3}{\circle*{1.2}}
\multiput(-25,70)(5,0){3}{\circle*{2}}
\multiput(245,70)(5,0){3}{\circle*{2}}
\put(20,0){$\alpha$}\put(80,0){$\beta$}\put(140,0){$\gamma$}\put(200,0){$\delta$}
\multiput(80,115)(0,-15){2}{\circle*{2}}
\multiput(80,30)(0,15){4}{\circle*{2}}
\multiput(75,88)(5,0){3}{\circle*{1.2}}
\put(80,145){\circle*{2}}
\put(140,145){\circle*{2}}
\put(83,143){\small $A$}
\put(143,143){\small $B$}
\multiput(140,115)(0,-15){2}{\circle*{2}}
\multiput(200,115)(0,-15){2}{\circle*{2}}
\multiput(135,88)(5,0){3}{\circle*{1.2}}
\multiput(195,88)(5,0){3}{\circle*{1.2}}
\multiput(25,70)(60,0){4}{\oval(50,120)}
\put(23,112){\small $P_{1}$}
\put(23,97){\small $P_{2}$}
\put(23,72){\small $P_{q-4}$}
\put(23,57){\small $P_{q-3}$}
\put(23,42){\small $P_{q-2}$}
\put(23,27){\small $P_{q-1}$}
\put(83,112){\small $Q_{1}$}
\put(83,97){\small $Q_{2}$}
\put(83,72){\small $Q_{q-4}$}
\put(83,57){\small $Q_{q-3}$}
\put(83,42){\small $Q_{q-2}$}
\put(83,27){\small $Q_{q-1}$}
\put(143,112){\small $R_{1}$}
\put(143,97){\small $R_{2}$}
\put(203,112){\small $S_{1}$}
\put(203,97){\small $S_{2}$}
\end{picture}
\caption {The sets $\pi_{i}\setminus \lbrace A,B\rbrace$ when $q$ is an odd prime power\label{fig1}.}
\end{figure}
First, let $\alpha,\beta,\gamma$ and $\delta$ denote the sets $\pi_{0}\setminus \lbrace A,B\rbrace,\pi_{1}\setminus \lbrace A,B\rbrace,\pi_{2}\setminus \lbrace A,B\rbrace,\pi_{3}\setminus \lbrace A,B\rbrace$ respectively, as illustrated in Figure \ref{fig1}. Let $A,B$ be the first and the second points. Choose an arbitrary point $P_{1}$ from $\alpha$ to be the third and an arbitrary point $Q_{1}$ from $\beta$ to be the fourth. It is obvious that $A,B,P_{1},Q_{1}$ do not lie on a plane. Next, choose $P_{2}\in\alpha\setminus BP_{1}Q_{1}$ to be the fifth and $Q_{2}\in\beta\setminus P_{1}P_{2}Q_{1}$ to be the sixth. Two planes intersect in a line and a line meets $\mathcal{O}$ in at most two points. Thus, we can continue in this way, i.e., take \emph{proper} points from $\alpha$ and $\beta$ in turn, until only one point remains in $\alpha$.
Now suppose this has been done so that the point $P_{q-1}$ remains, i.e., we have ordered the points as $A,B,P_{1},Q_{1},\cdots,P_{q-2},Q_{q-2}$. Then we have that $P_{q-4},Q_{q-4},P_{q-3},Q_{q-3}$ do not lie on a plane, nor do $Q_{q-4},P_{q-3},Q_{q-3},P_{q-2}$, and nor do $P_{q-3},Q_{q-3},P_{q-2},Q_{q-2}$. Next we order the two points $P_{q-1}$ and $Q_{q-1}$. We consider the following three cases:
{\bf Case 1:} $P_{q-1}\notin P_{q-2}Q_{q-3}Q_{q-2}$, $Q_{q-1}\notin P_{q-2}P_{q-1}Q_{q-2}$.
Note that this situation is ideal. Let the order be $P_{q-4},Q_{q-4},P_{q-3},Q_{q-3},P_{q-2},Q_{q-2},P_{q-1},Q_{q-1}$.
{\bf Case 2:} $P_{q-1}\notin P_{q-2}Q_{q-3}Q_{q-2}$, but $Q_{q-1}\in P_{q-2}P_{q-1}Q_{q-2}$.
Change the order to be $P_{q-4},Q_{q-4},P_{q-3},P_{q-2},Q_{q-3},Q_{q-2},P_{q-1},Q_{q-1}$.
{\bf Case 3:} $P_{q-1}\in P_{q-2}Q_{q-3}Q_{q-2}$.
Change the order to be $P_{q-4},Q_{q-4},P_{q-3},Q_{q-3},P_{q-2},Q_{q-2},Q_{q-1},P_{q-1}$.
Next, we find a \emph{proper} point $R_{1}\in\gamma$ to be the next point, as well as \emph{proper} points $S_{1}\in\delta$ and $R_{2}\in\gamma$. Then order the remaining points in $\gamma$ and $\delta$ just as what we have done for the points in $\alpha$ and $\beta$. Repeat the procedure until we have covered $n$ ($6\leq n\leq q^{2}+1$) points in $\mathcal{O}$. By now, we have got $n$ ordered points such that any four consecutive points do not lie on a plane.
Note that we have finished our first part. Denote the last four points by $W,X,Y$ and $Z$. To make sure that any four cyclically consecutive points do not lie on a plane, we still need to ensure that $X,Y,Z,A$ do not lie on a plane, nor do $Y,Z,A,B$ and nor do $Z,A,B,P_{1}$. We discuss in the following cases.
{\bf Case a:} $X,Y,Z$ and $A$ lie on a plane.
This happens only when $X\in\pi_{i}$, $Y\in\pi_{i+1}$ and $Z\in\pi_{i+2}$, for some $i$, $0\leq i\leq q-2$. For example, $P_{q-1},Q_{q-1}$ and $R_{1}$ in Figure \ref{fig1}. Otherwise, we always have exactly two of $X,Y,Z$ belonging to the same set $\pi_{j}\setminus\lbrace A,B\rbrace$, which ensures $X,Y,Z,A$ do not lie on a plane.
Note that $WXY$ intersects $\pi_{i+2}$ in at most two points and also $XYA$ intersects $\pi_{i+2}$ in at most two points, one of which is the point $A$. Thus, in this case, we find a new point $Z'$ in $\pi_{i+2}\setminus\lbrace A,B\rbrace$, not lying on planes $WXY$ and $XYA$ to be the new last point. We can always do this since there are totally $q+1\geq 6$ points on $\pi_{i+2}$.
{\bf Case b:} $Y,Z,A$ and $B$ lie on a plane.
This happens when the last two points lie in the same $\pi_{i}\setminus\lbrace A,B\rbrace$, which occurs in Cases $2$ and $3$ above. Note that $\alpha$ and $\beta$ can be any $\pi_{i}\setminus\lbrace A,B\rbrace$ and $\pi_{i+1}\setminus\lbrace A,B\rbrace$ respectively for $i=0,2,4,\cdots,q-1$ in the following discussion. In Case $2$, if the last three points are $Q_{q-4}$, $P_{q-3}$ and $P_{q-2}$, then we replace them by $Q_{q-4}$, $P_{q-3}$ and $Q_{q-3}$. If the last three points are $P_{q-2}$, $Q_{q-3}$ and $Q_{q-2}$,
then we replace them by $Q_{q-3}$, $P_{q-2}$ and $Q_{q-2}$. In Case $3$, if the last three points are $P_{q-2},Q_{q-2}$ and $Q_{q-1}$, then we replace them by $P_{q-2},P_{q-1}$ and $Q_{q-1}$.
{\bf Case c:} $Z,A,B$ and $P_{1}$ lie on a plane.
This happens when $Z$ lies in $\pi_{0}\setminus\lbrace A,B\rbrace$, i.e., $7\leq n\leq 2q-1$ and $n$ is odd.
In this case, after choosing the first three points $A,B,P_{1}$, we choose \emph{proper} points from $\pi_{2}\setminus\lbrace A,B\rbrace$ and $\pi_{3}\setminus\lbrace A,B\rbrace$ in turn.
\end{proof}
\begin{remark}
We use the condition that points $P_{q-4}$,$P_{q-3}$ and $P_{q-2}$ are on the same plane $\pi_{i}$, $0\le i\le q$, in Theorem {\rm \ref{thm1}}. Thus we exclude the case when $q=3$ since there are not enough points on each plane $\pi_{i}$. We give the MDS symbol-pair codes directly for $q=3$. There exists a linear MDS $(n,6)_{3}$ symbol-pair code, $n\in\lbrace 6,7,8,9,10\rbrace$, whose parity check matrix is formed by the first $n$ columns of the matrix
$$\left[
\begin{array}{cccccccccc}
0&1&1&1&1&1&1&1&1&1\\
1&0&1&2&1&2&2&1&2&1\\
0&0&1&0&2&0&2&2&1&1\\
0&0&1&1&2&2&1&0&2&0\\
\end{array}
\right].$$
\end{remark}
\begin{theorem}\label{thm2}
Let $q\geq 8$ be an even prime power. Then there exist linear MDS $(n,6)_{q}$ symbol-pair codes for all $n$, $6\leq n\leq q^{2}+1$.
\end{theorem}
\begin{proof}
Let the notations be defined as in Theorem \ref{thm1}. Note that the case when $q$ is even is different from that when $q$ is odd due to there being an odd number of planes. For $6\leq n\leq q^{2}-q+2$, we can order $n$ points on $\pi_{0},\pi_{1},\cdots,\pi_{q-1}$ just as in Theorem \ref{thm1} since the number of planes is even. The key step of this proof is to put the remaining $q-1$ points in order. To attain this goal, we first order all the points of the first three planes, and then we can just proceed as the case when $q$ is odd.
Let $\alpha,\beta,\gamma,\delta,\zeta$ denote the sets $\pi_{0}\setminus \lbrace A,B\rbrace,\pi_{1}\setminus \lbrace A,B\rbrace,\pi_{2}\setminus \lbrace A,B\rbrace,\pi_{3}\setminus \lbrace A,B\rbrace,\pi_{4}\setminus \lbrace A,B\rbrace$ respectively, as illustrated in Figure \ref{fig2}. Again, let $A$ and $B$ be the first two points and choose arbitrary $P_{1}$ and $Q_{1}$ from $\alpha$ and $\beta$ respectively. Choose the next point $R_{1}\in\gamma\setminus BP_{1}Q_{1}$, and then $P_{2}\in\alpha \setminus P_{1}Q_{1}R_{1}$ and $Q_{2}\in\beta\setminus P_{2}Q_{1}R_{1}$, i.e., take \emph{proper} points from $\alpha,\beta$ and $\gamma$ in turn. We can continue in this way until only one point remains in $\alpha$.
Suppose this has been done so that $P_{q-1}$ remains, i.e., we have ordered the points as $A,B,P_{1},Q_{1}$,\\$R_{1},\cdots,P_{q-2},Q_{q-2},R_{q-2}$. Note that the intersection of two planes meets $\mathcal{O}$ in at most two points. We can always find a point $S_{1}$ in $\delta$ that does not lie on the planes $P_{q-2}Q_{q-2}R_{q-2}$ and $P_{q-1}Q_{q-2}R_{q-2}$ since the two planes intersect $\delta$ in at most four points and there are $q-1\geq 7$ points in $\delta$. Similarly, we can find $T_{1}\in \zeta$ not lying on planes $P_{q-1}R_{q-2}S_{1}$ and $P_{q-1}Q_{q-1}S_{1}$, $S_{2}\in \delta$ not lying on planes $P_{q-1}Q_{q-1}T_{1}$ and $Q_{q-1}R_{q-1}T_{1}$. Next find $T_{2}\in \zeta\setminus Q_{q-1}R_{q-1}S_{2}$, $S_{3}\in \delta \setminus R_{q-1}S_{2}T_{2}$ and $T_{3}\in \zeta\setminus R_{q-1}S_{3}T_{2}$. Let the order of points be $P_{q-2},Q_{q-2},R_{q-2},S_{1},P_{q-1},T_{1},Q_{q-1},S_{2},R_{q-1},T_{2},S_{3},T_{3}$. Note that we have ordered all the points in $\alpha,\beta$ and $\gamma$, and any four consecutive points do not lie on a plane. There are an even number of planes left. We can then simply proceed as in Theorem \ref{thm1}, and also the similar discussion follows that of Theorem \ref{thm1}.
\begin{figure}[h]
\begin{center}
\begin{picture}(290,140)
\thicklines
\put(20,0){$\alpha$}\put(80,0){$\beta$}\put(140,0){$\gamma$}\put(200,0){$\delta$}\put(260,0){$\zeta$}
\multiput(25,60)(60,0){5}{\oval(50,90)}
\multiput(-25,58)(5,0){3}{\circle*{2}}
\multiput(305,58)(5,0){3}{\circle*{2}}
\multiput(20,90)(0,-15){2}{\circle*{2}}
\multiput(20,30)(0,15){2}{\circle*{2}}
\multiput(15,58)(5,0){3}{\circle*{1.2}}
\multiput(80,90)(0,-15){2}{\circle*{2}}
\multiput(80,30)(0,15){2}{\circle*{2}}
\multiput(75,58)(5,0){3}{\circle*{1.2}}
\multiput(140,90)(0,-15){2}{\circle*{2}}
\multiput(140,30)(0,15){2}{\circle*{2}}
\multiput(135,58)(5,0){3}{\circle*{1.2}}
\put(120,120){\circle*{2}}
\put(180,120){\circle*{2}}
\put(123,118){\small $A$}
\put(183,118){\small $B$}
\multiput(200,90)(0,-15){3}{\circle*{2}}
\multiput(260,90)(0,-15){3}{\circle*{2}}
\multiput(195,48)(5,0){3}{\circle*{1.2}}
\multiput(255,48)(5,0){3}{\circle*{1.2}}
\put(23,87){\small $P_{1}$}
\put(23,72){\small $P_{2}$}
\put(23,42){\small $P_{q-2}$}
\put(23,27){\small $P_{q-1}$}
\put(83,87){\small $Q_{1}$}
\put(83,72){\small $Q_{2}$}
\put(83,42){\small $Q_{q-2}$}
\put(83,27){\small $Q_{q-1}$}
\put(143,87){\small $R_{1}$}
\put(143,72){\small $R_{2}$}
\put(143,42){\small $R_{q-2}$}
\put(143,27){\small $R_{q-1}$}
\put(203,87){\small $S_{1}$}
\put(203,72){\small $S_{2}$}
\put(203,57){\small $S_{3}$}
\put(263,87){\small $T_{1}$}
\put(263,72){\small $T_{2}$}
\put(263,57){\small $T_{3}$}
\end{picture}
\end{center}\caption{The sets $\pi_{i}\setminus \lbrace A,B\rbrace$ when $q$ is an even prime power\label{fig2}.}
\end{figure}
\end{proof}
\begin{remark}
When $q=4$, since there are only five points on each plane $\pi_{i}$, $0\le i\le q$, we discuss it as a special case. Denote the primitive element of $\mathbb{F}_{4}$ as $w$. Then there is a linear MDS $(n,6)_{4}$ symbol-pair code, $n\in\lbrace6,8,9,10,11,12,13,14,15,16,17\rbrace$, and its parity check matrix is formed by the first $n$ columns of the matrix
$$\left[
\begin{array}{ccccccccccccccccc}
0&1&1&1&1 &1&1 &1& 1&1&1& 1&1& 1& 1& 1& 1\\
1&0&1&w&1+w&1&w &1+w&w&w&1+w&1&w& 1& 1+w& 1+w&1\\
0&0&1&0&w &0&1+w&0& 1&1&w& w&w+1&w& 1+w& 1+w&1\\
0&0&0&1&0 &w&0 &1+w&1&w&1& w&w& 1+w&1+w& 1 &1+w\\
\end{array}
\right].
$$
There exists a linear MDS $(7,6)$ symbol-pair code with parity check matrix
$$\left[
\begin{array}{ccccccc}
0&1&1&1&1 &1&1\\
1&0&1&w&1+w&1&w\\
0&0&1&0&w &0&1\\
0&0&0&1&0 &w&w\\
\end{array}
\right].
$$
\end{remark}
Summing up the above, we can conclude the following theorem.
\begin{theorem}
For any prime power $q$, $q\geq 3$, and any integer $n$, $6 \leq n\leq q^{2}+1$, there exists a linear MDS $(n,6)_{q}$ symbol-pair
code.
\end{theorem}
\begin{remark}
Compare the cases $r=2$ and $r=3$ in Theorem {\rm \ref{themcon}}, if we consider the set of all the points instead of the ovoid, all the lines through a fixed point instead of the planes described in Lemma {\rm \ref{thmPG2}}, then we can also get linear MDS $(n,5)_{q}$ symbol-pair codes for $5\leq n\leq q^{2}+q+1$ with $q$ being a prime power in a similar way. Thus, this method deserves further investigation for larger $r$, which may derive MDS symbol-pair codes with larger pair-distance.
\end{remark}
\section{MDS symbol-pair codes from elliptic curves}\label{sec5}
The previous two sections construct MDS symbol-pair codes with pair-distance $5$ and $6$. In this section, we give a construction of MDS symbol-pair codes with general pair-distance ($\ge 7$) from elliptic curve codes. We first briefly review some facts about elliptic curve codes.
Let $E/\f{q}$ be an elliptic curve over $\f{q}$
with function field $\f{q}(E)$. Let $E(\f{q})$ be the set of all $\f{q}$-rational points on $E$.
Suppose $D=\{P_{1},P_{2},\cdots,P_{n}\}$ is a proper subset of rational points $E(\f{q})$, and $G$ is a divisor of degree $k$
($2g-2<k<n$) with $\mathrm{Supp}(G)\cap D=\emptyset$.
Without any confusion, we also write $D=P_{1}+P_{2}+\cdots+P_{n}$.
Denote by $\mathscr{L}(G)$ the $\f{q}$-vector space of all
rational functions $f\in \f{q}(E)$ with the principal divisor
$\mathrm{div}(f)\geqslant -G$, together with the zero function
(\cite{Stichtenoth}).
The functional AG code $C_{\mathscr{L}}(D, G)$ is defined to be the image of the following evaluation map:
\[
ev: \mathscr{L}(G)\rightarrow \f{q}^{n};\, f\mapsto
(f(P_{1}),f(P_{2}),\cdots,f(P_{n})).\enspace
\]
It is well-known that $C_{\mathscr{L}}(D, G)$ is a linear code with parameters $[n,k,d_{H}]$, where the minimum Hamming distance $d_{H}$ has two choices:
\[
d_{H}=n-k,\,\mbox{ or }\, d_{H}=n-k+1.\
\]
A linear $[n,k,d_{H}]$ code is called an MDS code if $d_{H}=n-k+1$ and is called an almost MDS code if $d_{H}=n-k$.
Suppose $O$ is one of the $\f{q}$-rational
points on $E$. The set of rational points $E(\f{q})$ forms an abelian group
with zero element $O$ (for the definition of the sum of any two
points, we refer to \cite{Silverman09}), and it is isomorphic to the
Picard group
$\mathrm{div}^o(E)/\mathrm{Prin}(\f{q}(E))$, where $\mathrm{Prin}(\f{q}(E))$
is the subgroup consisting of all principal divisors.
Denote by $\oplus$ and $\ominus$ the additive and minus operator in the group $E(\f{q})$, respectively.
\begin{proposition}[\cite{Cheng08,ZFW14}]\label{ECC}
Let $E$ be an elliptic curve over $\f{q}$ with an $\f{q}$-rational point $O$,
$D=\{P_{1},P_{2},\cdots,P_{n}\}$ a subset of $E(\f{q})$ such that
$O\notin D$ and let $G=kO$ ($0<k<n$). Endow $E(\f{q})$ a group structure with the zero element $O$.
Denote by
\[
N(k,O,D)=|\{S\subset D\,:\,|S|=k,\,\oplus_{P\in S}P=O\}|.
\]
Then the AG code $C_{\mathscr{L}}(D, G)$ has the minimum Hamming distance $d_H=n-k+1$ if and only if
\[
N(k,O,D)= 0.
\]
And the minimum Hamming distance $d_H=n-k$ if and only if
\[
N(k,O,D)>0.
\]
\end{proposition}
\begin{proof}
We have already seen that the minimum distance of $C_{\mathscr{L}}(D, G)$ has two choices: $n-k$, $n-k+1$. So $C_{\mathscr{L}}(D, G)$ is not MDS, i.e., $d=n-k$ if and only if there is a function $f\in\mathscr{L}(G)$ such that the evaluation $ev(f)$ has weight $n-k$. This is equivalent to that $f$ has $k$ zeros in $D$, say $P_{i_1}, \cdots, P_{i_k}$. That is
\[
\mathrm{div}(f)\geq -(k-1)O-P+(P_{i_1}+\cdots+P_{i_k}),
\]
which is equivalent to
\[
\mathrm{div}(f)=-(k-1)O-P+(P_{i_1}+\cdots+P_{i_k}).
\]
The existence of such an $f$ is equivalent to saying
\[
P_{i_1}\oplus\cdots\oplus P_{i_k}=P.
\]
Namely, $N(k,P,D)> 0.$
It follows that the AG code $C_{\mathscr{L}}(D, G)$ has the minimum Hamming distance $n-k+1$ if and only if
$ N(k,P,D)=0.\ $
\end{proof}
We restrict to the case $n> q+1$, since for every length $n\le q+1$, MDS symbol-pair codes of length $n$ can be constructed from Reed-Solomon codes. In this case, the minimum Hamming distance $d_H$ of elliptic curve codes is related to the main conjecture of MDS codes which was affirmed for elliptic curve codes~\cite{LWZ15,Mun92}.
\begin{proposition}[\cite{LWZ15,Mun92}]\label{ECSSP}
Let $C_{\mathscr{L}}(D, G)$ be the elliptic curve code constructed in Proposition {\rm \ref{ECC}} with length $n> q+1$. Then the subset sum problem always has solutions, i.e.,
\[
N(k,O,D)>0.
\]
And hence, elliptic curve codes with length $n> q+1$ have deterministic minimum Hamming distance $d_H=n-k$.
\end{proposition}
That is, elliptic curve codes with length $n> q+1$ are almost MDS codes.
To obtain long codes from elliptic curves, we need the following two well-known results of elliptic curves over finite fields.
\begin{lemma}[Hasse-Weil Bound~\cite{Silverman09}]\label{HasseWeil}
Let $E$ be an elliptic curve over $\f{q}$. Then the number of $\f{q}$-rational points on $E$ is bounded by
\[
|E(\f{q})|\le q+\lfloor 2\sqrt{q}\rfloor+1.
\]
\end{lemma}
\begin{lemma}[Hasse-Deuring~\cite{Deuring41}]\label{HasseDeuring}
The maximal number $N(\f{q})$ of $\f{q}$-rational points on $E$, where $E$ runs over all elliptic curves over $\f{q}$, is
\[
N(\f{q})=\left\{
\begin{array}{ll}
q+\lfloor 2\sqrt{q}\rfloor, &\hbox{if $q=p^a,\,a\ge 3$, $a$ odd and $p|\lfloor 2\sqrt{q}\rfloor$};\\
q+\lfloor 2\sqrt{q}\rfloor+1, &\hbox{otherwise.}
\end{array}
\right.
\]
\end{lemma}
Denote by
\[
\delta(q)=\left\{
\begin{array}{ll}
0, & \hbox{if $q=p^a,\,a\ge 3$, $a$ odd and $p\,|\,\lfloor 2\sqrt{q}\rfloor$;} \\
1, & \hbox{otherwise.}
\end{array}
\right.
\]
To construct an MDS symbol-pair code from classical error-correcting codes with large minimum Hamming distance, the key step is to find a way of ordering the coordinates. For general codes, this step seems very difficult. In the rest of this paper, we deal with the case of elliptic curve codes.
\begin{theorem}
Let $N(\f{q})=q+\lfloor 2\sqrt{q}\rfloor+\delta(q)$. Then for any $7\le d+2\leq n\le N(\f{q})-3$, there exist linear
MDS symbol-pair codes over $\f{q}$ with parameters $(n,d+2)_{q}$.
\end{theorem}
\begin{proof}
The existence of MDS symbol-pair codes with parameters $d+2=n$ follows from~\cite{CJKWY}. Below we only consider the case $7\le d+2< n\le N(\f{q})-3$. By Lemma~\ref{HasseDeuring}, take $E$ to be a maximal elliptic curve over $\f{q}$ with an $\f{q}$-rational point $O$, i.e.,
\[
|E(\f{q})|=N(\f{q}).
\]
Take divisor $G=kO$ in the construction of elliptic curve codes.
Case (I): $N=N(\f{q})$ is odd, then there is no element of order $2$ in $E(\f{q})$. Suppose
$$E(\f{q})=\{P_{1},P_{2},\cdots,P_{N-2}, P_{N-1}, O\},$$
where $P_1\oplus P_2=P_3\oplus P_4=\cdots=P_{N-2}\oplus P_{N-1}=O$.
\begin{enumerate}
\item For odd $d$ and even $n:\,7\le d+2<n\le N-1$, in this case $k=N-1-d$ is odd. Take
\[
D=\{P_{1},P_{2},\cdots,P_{N-2}, P_{N-1}\}.
\]
Then by Proposition~\ref{ECSSP}, there are no $k$ cyclically consecutive points whose sum is $O$. And hence, the elliptic curve code $C_{\mathscr{L}}(D, G)$ is an MDS symbol-pair code with parameters $(N-1,d+2)_{q}$. By deleting pairs $(P_1,P_2),(P_3,P_4)$, etc., we can obtain MDS symbol-pair codes with parameters $(n,d+2)_{q}$, where $n$ runs over all even integers $7\le d+2<n\le N-1$.
\item For even $d$ and odd $n:\,7\le d+2<n\le N-2$, in this case $k=N-2-d$ is odd. Take
\[
D=\{P_{1},P_{2},\cdots,P_{N-2}\}.
\]
Then by Proposition~\ref{ECSSP}, there are no $k$ cyclically consecutive points whose sum is $O$. And hence, the elliptic curve code $C_{\mathscr{L}}(D, G)$ is an MDS symbol-pair code with parameters $(N-2,d+2)_{q}$. By deleting pairs $(P_1,P_2),(P_3,P_4)$, etc., we can obtain MDS symbol-pair codes with parameters $(n,d+2)_{q}$ where $n$ runs over all odd integers $d+2<n\le N-2$.
\item For even $d$ and even $n:\,7\le d+2<n\le N-3$, in this case $k=N-3-d$ is even. Write $N-3=(k+1)s+r$ for some integers $s\ge 1$ and $0\le r\le k$.
Take the pre-evaluation set
\[
D_0=\{P_{1},P_{2},\cdots,P_{N-5},P_{N-4},P_{N-2}\}
\]
and arrange it by the following algorithm:
\textbf{Step 1.} Arrange $D_0$ as following
\begin{equation*}
\begin{array}{rl}
D_1=&\{P_{1},\cdots,P_{k-1},P_{N-5},P_k,\cdots,P_{(s-1)k-1},P_{N-3-s},P_{(s-1)k},\\
&\cdots,P_{sk-1},P_{N-4},P_{sk}, P_{sk+1},\cdots,P_{sk+r-1},P_{N-2}\}.\\
\end{array}
\end{equation*}
After this step, there are no $k$ consecutive points whose sum is $O$ in the sequence
\[
P_{1},\cdots,P_{k-1},P_{N-5},P_k,\cdots,P_{(s-1)k-1},P_{N-3-s},P_{(s-1)k},\cdots,P_{sk-1},P_{N-4},P_{sk}, P_{sk+1},\cdots,P_{sk+r-2}.
\]
But there may be some $k$ cyclically consecutive points whose sum is $O$ in the tail sequence
\[
P_{(s-1)k+r+1},\cdots,P_{sk-1},P_{N-4},P_{sk}, P_{sk+1},\cdots,P_{sk+r-1},P_{N-2}, P_{1},\cdots,P_{k-r-1}.
\]
For instance, $k=6$, $N=19$, by Step~1, we get
\[
D_1={P_1,\cdots,P_5,P_{14},P_6,\cdots,P_{11},P_{15},P_{12},P_{13},P_{17}}.
\]
There are no $6$ consecutive points whose sum is $O$ in the sequence
\[
P_1,\cdots,P_5,P_{14},P_6,\cdots,P_{11},P_{15},P_{12},P_{13}.
\]
But there may be some $6$ cyclically consecutive points whose sum is $O$ in the tail sequence
\[
P_{10},P_{11},P_{15},P_{12},P_{13},P_{17},P_1,P_2.
\]
\textbf{Step 2.} In the case that $r$ is even. It is easy to see that at most one of the following two equalities holds:
\[
P_{(s-1)k+r+2}\oplus\cdots\oplus P_{N-4}\oplus\cdots\oplus P_{N-2}=P_{(s-1)k+r+2}\oplus P_{N-4}\oplus P_{sk+r-1}\oplus P_{N-2}=O,
\]
and
\[
P_{(s-1)k+r+3}\oplus\cdots\oplus P_{N-4}\oplus\cdots\oplus P_{N-2}\oplus P_1= P_{N-4}\oplus P_{sk+r-1}\oplus P_{N-2}\oplus P_1=O.
\]
If the first one holds, then SWITCH $P_{(s-1)k+r+1}$ and $P_{(s-1)k+r+2}$; if the second one holds, then SWITCH $P_1$ and $P_2$; if neither of the two
holds, then do nothing.
For any $i=1,\cdots,\frac{k-r-2}{2}$, similarly at most one of the following two equalities holds:
\[
P_{(s-1)k+r+2i+2}\oplus\cdots\oplus P_{N-4}\oplus\cdots\oplus P_{N-2}\oplus P_{1}\oplus \cdots\oplus P_{2i}=P_{(s-1)k+r+2i+2}\oplus P_{N-4}\oplus P_{sk+r-1}\oplus P_{N-2}=O,
\]
and
\[
P_{(s-1)k+r+2i+1}\oplus\cdots\oplus P_{N-4}\oplus\cdots\oplus P_{N-2}\oplus P_{1}\oplus \cdots\oplus P_{2i-1}= P_{N-4}\oplus P_{sk+r-1}\oplus P_{N-2}\oplus P_{2i+1}=O.
\]
If the first one holds, then SWITCH $P_{(s-1)k+r+2i+1}$ and $P_{(s-1)k+r+2i+2}$; if the second one holds, then SWITCH $P_{2i+1}$ and $P_{2i+2}$; if neither of the two
holds, then do nothing.
In the case that $r$ is odd, the algorithm is the same as the even case, check the sum of $k$ cyclically consecutive points and do the corresponding SWITCH operation.
Continue the above example, if
\[
P_{10}\oplus P_{11}\oplus P_{15}\oplus P_{12}\oplus P_{13}\oplus P_{17}=P_{10}\oplus P_{15}\oplus P_{13}\oplus P_{17}=O,
\]
then SWITCH $P_{9}$ and $P_{10}$; and in this case, it is immediate that
\[
P_{11}\oplus P_{15}\oplus P_{12}\oplus P_{13}\oplus P_{17}\oplus P_1=P_{15}\oplus P_{13}\oplus P_{17}\oplus P_1\neq O,
\]
so we do not need to reorder $P_1$ and $P_2$, and so on.
Using the above algorithm to rearrange the evaluation set to get a newly arranged evaluation set $D$, by Proposition~\ref{ECSSP}, there are no $k$ cyclically consecutive points whose sum is $O$. And hence, the elliptic curve code $C_{\mathscr{L}}(D, G)$ is an MDS symbol-pair code with parameters $(N-3,d+2)_{q}$. So, similarly as above, by deleting pairs from the pre-evaluation set, we can obtain MDS symbol-pair codes with parameters $(n,d+2)_{q}$ where $n$ runs over all even integers $d+2<n\le N-3$.
\item For odd $d$ and odd $n:\,7\le d+2<n\le N-2$, in this case $k=N-2-d$ is even. Write $N-2=(k+1)s+r$ for some integers $s\ge 1$ and $0\le r\le k$.
Take the pre-evaluation set
\[
D_0=\{P_{1},P_{2},\cdots,P_{N-3},P_{N-2}\}
\]
and arrange it as following
\begin{equation*}
\begin{array}{rl}
D=&\{P_{1},\cdots,P_{k-1},P_{N-3},P_k,\cdots,P_{(s-1)k-1}, \\
& P_{N-1-s},P_{(s-1)k},\cdots,P_{sk-1},P_{N-2},P_{sk}, P_{sk+1},\cdots,P_{sk+r}\}.
\end{array}
\end{equation*}
If $r$ is even, then it is easy to see that by Proposition~\ref{ECSSP} there are no $k$ cyclically consecutive points whose sum is $O$.
If $r$ is odd, then similarly as the case when $d$ and $n$ are even, there may be some $k$ cyclically consecutive points whose sum is $O$ in the tail sequence. In this case, we just need process the same algorithm in the case $3$ to obtain a rearranged evaluation set $D$ such that there are no $k$ cyclically consecutive points whose sum is $O$.
And hence, the elliptic curve code $C_{\mathscr{L}}(D, G)$ is an MDS symbol-pair code with parameters $(N-2,d+2)_{q}$. So, similarly as above, by deleting pairs from the pre-evaluation set, we can obtain MDS symbol-pair codes with parameters $(n,d+2)_{q}$ where $n$ runs over all odd integers $7\le d+2<n\le N-2$.
\end{enumerate}
In conclusion, in the case that $N=N(\f{q})$ is odd, for any $7\le d+2\le n\le N(\f{q})-3$, no matter whether $d$ is odd or even, there exists an MDS
symbol-pair code with parameters $(n,d+2)_{q}$.
Case (II): $N=N(\f{q})$ is even. The proof is the same. Note that there are one or three non-zero elements of order $2$ in the group $E(\f{q})$.
Using these elements in the setting of the pre-evaluation set, the remainder of the argument is analogous. We omit the details here.
So, by the discussion above, we complete the proof of the theorem.
\end{proof}
\begin{remark}
From the proof, we see that in some cases, the length of the MDS symbol-pair code constructed from elliptic curve can attain $N(\f{q})-2$ or $N(\f{q})-1$. We omit the detail of the statements in the theorem to get a clear description of our result. Also, note that there are other works devoted to constructing almost MDS codes using curves {\rm \cite{BC}} besides elliptic curves. To construct MDS symbol-pair codes using these almost MDS codes, how to arrange the evaluation set becomes the difficult step.
\end{remark}
\section{Conclusion}\label{sec6}
In this paper, we first give a sufficient condition for the existence of linear MDS symbol-pair codes over $\mathbb{F}_{q}$.
On this basis, we show that a linear MDS $(n,5)_q$ symbol-pair code over $\mathbb{F}_{q}$ exists if and only if the length $n$ ranges from $5$ to $q^2+q+1$. Later, we introduce a special configuration in projective geometry called ovoid, which allows us to derive $q$-ary linear MDS symbol-pair codes with $d=6$ and length ranging from $6$ to $q^{2}+1$. This is an interesting method and deserves further investigation since it works well for both $d=5$ and $d=6$, and it may work for larger pair-distance. With the help of elliptic curves, we show that we can construct linear MDS $(n,d+2)_{q}$ symbol-pair codes for any $n,d$ satisfying $7\le d+2\le n\le q+\lfloor 2\sqrt{q}\rfloor+\delta(q)-3$. Compared with the results listed in Table \ref{tab}, our results provide a much larger range of parameters.
|
\section{Introduction}
In this paper, we study the existence and uniqueness of positive solution for the following singular fractional boundary value problem
$$
\left\{
\begin{array}{l}
D_{0+}^\alpha u(t)=f(t,u(t)),\,\,0<t<1\\
u(0)=u(1)=u'(0)=u'(1)=0,
\end{array}
\right.
\eqno (1)
$$
where $\alpha\in (3,4]$, and $D_{0+}^\alpha$ denotes the Riemann-Liouville fractional derivative. Moreover, $f: (0,1]\times [0,\infty)\rightarrow [0,\infty)$ with $\lim\limits_{t\rightarrow 0+} f(t,-)=\infty$ (i.e. $f$ is singular at $t=0$).
Similar problem was investigated in [1], in case when $\alpha\in (1,2]$ and with boundary conditions $u(0)=u(1)=0$. We note as well work [2], where
the following problem
$$
\left\{
\begin{array}{l}
D^\alpha u+f(t,u,u',D^\mu u)=0,\,\,0<t<1\\
u(0)==u'(0)=u'(1)=0,
\end{array}
\right.
$$
was under consideration. Here $\alpha\in (2,3), \mu\in (0,1)$ and function $f(t,x,y,z)$ is singular at the value of $0$ of its arguments $x,y,z$.
We would like notice some related recent works [3-5], which consider higher order fractional nonlinear equations for the subject of the existence of positive solutions.
\section{Preliminaries}
We need the following lemma, which appear in [6].
{\bf Lemma 1.} (Lemma 2.3 of [6]) Given $h\in C[0,1]$ and $3<\alpha\leq 4$, a unique solution of
$$
\left\{
\begin{array}{l}
D_{0+}^\alpha u(t)=h(t),\,\,0<t<1\\
u(0)=u(1)=u'(0)=u'(1)=0,
\end{array}
\right.\eqno (1)
$$
is
$$
u(t)=\int\limits_0^1G(t,s)h(s)ds,
$$
where
$$
G(t,s)=
\left\{
\begin{array}{l}
\frac{(t-s)^{\alpha-1}+(1-s)^{\alpha-2}t^{\alpha-2}\left[(s-t)+(\alpha-2)(1-t)s\right]}{\Gamma(\alpha)},\,\,0\leq s\leq 1\\
\frac{(1-s)^{\alpha-2}t^{\alpha-2}\left[(s-t)+(\alpha-2)(1-t)s\right]}{\Gamma(\alpha)},\,\,0\leq t\leq s\leq 1\\
\end{array}
\right.
$$
{\bf Lemma 2.} (Lemma 2.4 of [6]) The function $G(t,s)$ appearing in Lemma 1 satisfies:
(a) $G(t,s)>0$ for $t,s \in (0,1)$;
(b) $G(t,s)$ is continuous on $[0,1]\times [0,1]$.
For our study, we need a fixed point theorem. This theorem uses the following class of functions $\mathfrak{F}$.
By $\mathfrak{F}$ we denote the class of functions $\varphi: (0, \infty)\rightarrow \mathbb{R}$ satisfying the following conditions:
(a) $\varphi$ is strictly increasing;
(b) For each sequence $(t_n)\subset (0,\infty)$
$$
\lim\limits_{n\rightarrow \infty} t_n=0 \,\,\Leftrightarrow \lim\limits_{n\rightarrow \infty}\varphi(t_n)=-\infty;
$$
(c) There exists $k\in (0,1)$ such that $\lim\limits_{t\rightarrow 0^+}t^k\varphi(t)=0$.
Examples of functions belonging to $\mathfrak{F}$ are $\varphi(t)=-\frac{1}{\sqrt{t}}, \varphi(t)=\ln t, \varphi(t)=\ln t+t,\varphi(t)=\ln (t^2+t)$.
The result about fixed point which we use is the following and it appears in [7]:
\textbf{Theorem 3.} Let $(X,d)$ be a complete metric space and $T:X\rightarrow X$ a mapping such that there exist $\tau>0$ and $\varphi\in \mathfrak{F}$ satisfying for any $x,y\in X$ with $d(Tx,Ty)>0$,
$$
\tau+\varphi\left(d(Tx,Ty)\right)\leq \varphi(d(x,y)).
$$
Then $T$ has a unique fixed point.
\section{Main result}
Our starting point of this section is the following lemma.
\textbf{Lemma 4.} Let $0<\sigma<1$, $3<\alpha<4$ and $F:(0,1]\rightarrow \mathbb{R}$ is continuous function with $\lim\limits_{t\rightarrow 0^+} F(t)=\infty$. Suppose that $t^\sigma F(t)$ is a continuous function on $[0,1]$. Then the function defined by
$$
H(t)=\int\limits_0^1 G(t,s)F(s)ds
$$
is continuous on $[0,1]$, where $G(t,s)$ is the Green function appearing in Lemma 1.
\emph{Proof:} We consider three cases:
{\bf \texttt {1. Case No 1}}. $t_0=0$.
It is clear that $H(0)=0$. Since $t^\sigma F(t)$ is continuous on $[0,1]$, we can find a constant $M>0$ such that
$$
|t^\sigma F(t)|\leq M \,\,\,for \,\,any\,\,\, t\in [0,1].
$$
Moreover, we have
$$
\begin{array}{l}
|H(t)-H(0)|=|H(t)|=\left| \int\limits_0^1 G(t,s)F(s)ds\right|=\left| \int\limits_0^1 G(t,s)s^{-\sigma} s^\sigma F(s)ds\right|=\\
=\left| \int\limits_0^t \frac{(t-s)^{\alpha-1}+(1-s)^{\alpha-2}t^{\alpha-2}[(s-t)+(\alpha-2)(1-t)s]}{\Gamma(\alpha)}s^{-\sigma}s^\sigma F(s)ds\right.\\
\left.+
\int\limits_t^1 \frac{(1-s)^{\alpha-2}t^{\alpha-2}[(s-t)+(\alpha-2)(1-t)s]}{\Gamma(\alpha)}s^{-\sigma}s^\sigma F(s)ds\right|=\\
=\left|\int\limits_0^1 \frac{(1-s)^{\alpha-2}t^{\alpha-2}[(s-t)+(\alpha-2)(1-t)s]}{\Gamma(\alpha)}s^{-\sigma}s^\sigma F(s)ds+
\int\limits_0^t \frac{(t-s)^{\alpha-1}}{\Gamma(\alpha)}s^{-\sigma}s^\sigma F(s)ds\right|\leq\\
\leq \frac{Mt^{\alpha-2}}{\Gamma(\alpha)}\int\limits_0^1 (1-s)^{\alpha-2}|(s-t)+(\alpha-2)(1-t)s|s^{-\sigma}ds+
\frac{M}{\Gamma(\alpha)}\int\limits_0^t (t-s)^{\alpha-1}s^{-\sigma}ds\leq\\
\leq \frac{M(\alpha-1)t^{\alpha-2}}{\Gamma(\alpha)}\int\limits_0^1 (1-s)^{\alpha-2}s^{-\sigma}ds+
\frac{Mt^{\alpha-1}}{\Gamma(\alpha)}\int\limits_0^t \left(1-\frac{s}{t}\right)^{\alpha-1}s^{-\sigma}ds.\\
\end{array}
$$
Considering definition of Euler's beta-function, we derive
$$
|H(t)-H(0)|\leq \frac{M(\alpha-1)t^{\alpha-2}}{\Gamma(\alpha)} B(1-\sigma,\alpha-1)+\frac{Mt^{\alpha-\sigma}}{\Gamma(\alpha)} B(1-\sigma, \alpha).
$$
From this we deduce that $|H(t)-H(0)|\rightarrow 0$ when $t\rightarrow 0$.
This proves that $H$ is continuous at $t_0=0$.
{\bf \texttt {2. Case No 2}}. $t_0\in (0,1)$.
We take $t_n\rightarrow t_0$ and we have to prove that $H(t_n)\rightarrow H(t_0)$. Without loss of generality, we consider $t_n>t_0$. Then, we have
$$
\begin{array}{l}
\left|H(t_n)-H(t_0)\right|=\left|\int\limits_0^{t_n}\frac{(t_n-s)^{\alpha-1}+(1-s)^{\alpha-2}t_n^{\alpha-2}[(s-t_n)+(\alpha-2)(1-t_n)s]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds+\right.\\
\left.+\int\limits_{t_n}^1\frac{(1-s)^{\alpha-2}t_n^{\alpha-2}[(s-t_n)+(\alpha-2)(1-t_n)s]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds-\right.\\
-\int\limits_0^{t_0}\frac{(t_0-s)^{\alpha-1}+(1-s)^{\alpha-2}t_0^{\alpha-2}[(s-t_0)+(\alpha-2)(1-t_0)s]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds-\\
\left.-\int\limits_{t_0}^1\frac{(1-s)^{\alpha-2}t_0^{\alpha-2}[(s-t_0)+(\alpha-2)(1-t_0)s]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds\right|=\\
=\left|\int\limits_0^1\frac{(1-s)^{\alpha-2}t_n^{\alpha-2}[(s-t_n)+(\alpha-2)(1-t_n)s]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds+\int\limits_0^{t_n}\frac{(t_n-s)^{\alpha-1}}{\Gamma(\alpha)}s^{-\sigma} s^\sigma F(s)ds-\right.\\
\left.-\int\limits_0^1\frac{(1-s)^{\alpha-2}t_0^{\alpha-2}[(s-t_0)+(\alpha-2)(1-t_0)s]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds-\int\limits_0^{t_0}\frac{(t_0-s)^{\alpha-1}}{\Gamma(\alpha)}s^{-\sigma} s^\sigma F(s)ds\right|=\\
=\left|\int\limits_0^1\frac{(1-s)^{\alpha-2}\left(t_n^{\alpha-2}-t_0^{\alpha-2}\right)[(s-t_n)+(\alpha-2)(1-t_n)s]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds+\right.\\
\left.+\int\limits_0^1\frac{(1-s)^{\alpha-2}t_0^{\alpha-2}\left[(s-t_n)+(\alpha-2)(1-t_n)s-[(s-t_0)+(\alpha-2)(1-t_0)s]\right]}{\Gamma(\alpha)}
s^{-\sigma} s^\sigma F(s)ds+\right.\\
\left.+\int\limits_0^{t_0}\frac{\left[(t_n-s)^{\alpha-1}-(t_0-s)^{\alpha-1}\right]}{\Gamma(\alpha)}s^{-\sigma} s^\sigma F(s)ds+\int\limits_{t_0}^{t_n}\frac{(t_n-s)^{\alpha-1}}{\Gamma(\alpha)}s^{-\sigma} s^\sigma F(s)ds\right|\leq\\
\leq \frac{M\left|t_n^{\alpha-2}-t_0^{\alpha-2}\right|}{\Gamma(\alpha)}(\alpha-1)\int\limits_0^1(1-s)^{\alpha-2}s^{-\sigma}ds+\\
+\frac{M t_0^{\alpha-2}}{\Gamma(\alpha)}\int\limits_0^1(1-s)^{\alpha-2}|t_n-t_0|(\alpha-1)s^{-\sigma}ds+\\
+\frac{M}{\Gamma(\alpha)}\int\limits_0^{t_0}\left|(t_n-s)^{\alpha-1}-(t_0-s)^{\alpha-1}\right|s^{-\sigma}ds+
\frac{M}{\Gamma(\alpha)}\int\limits_{t_0}^{t_n}(t_n-s)^{\alpha-1}s^{-\sigma}ds\leq\\
\leq \frac{M}{\Gamma(\alpha)}\left(t_n^{\alpha-2}-t_0^{\alpha-2}\right)(\alpha-1)B(1-\sigma,\alpha-1)+
\frac{M(t_n-t_0)}{\Gamma(\alpha)}(\alpha-1)B(1-\sigma,\alpha-1)+\\
+\frac{M}{\Gamma(\alpha)}I_n^1+\frac{M}{\Gamma(\alpha)}I_n^2,\\
\end{array}
$$
where
$$
I_n^1=\int\limits_0^{t_0}\left[(t_n-s)^{\alpha-1}-(t_0-s)^{\alpha-1}\right]s^{-\sigma}ds,\,\,\,I_n^2=\int\limits_{t_0}^{t_n}(t_n-s)^{\alpha-1}s^{-\sigma}ds.
$$
In the sequel, we will prove that $I_n^1\rightarrow 0$ when $n\rightarrow \infty$. In fact, as
$$
\left[(t_n-s)^{\alpha-1}-(t_0-s)^{\alpha-1}\right]s^{-\sigma}\leq \left[|t_n-s|^{\alpha-1}-|t_0-s|^{\alpha-1}\right]s^{-\sigma}\leq 2s^{-\sigma}
$$
and $\int\limits_0^1 2s^{-\sigma}ds=\frac{2}{1-\sigma}<\infty$. By Lebesque's dominated convergence theorem $I_n^1\rightarrow 0$ when $n\rightarrow \infty$.
Now, we will prove that $I_n^2\rightarrow 0$ when $n\rightarrow \infty$. In fact, since
$$
I_n^2=\int\limits_{t_0}^{t_n}(t_n-s)^{\alpha-1}s^{-\sigma}ds\leq \int\limits_{t_0}^{t_n}s^{-\sigma}ds=\frac{1}{1-\sigma}\left(t_n^{1-\sigma}-t_0^{1-\sigma}\right)
$$
and as $t_n\rightarrow t_0$, we obtain the desired result.
Finally, taking into account above obtained estimates, we infer that $|H(t_n)-H(t_0)|\rightarrow 0$ when $n\rightarrow \infty$.
{\bf \texttt {3. Case No 3}}. $t_0=1$.
It is clear that $H(1)=0$ and following the same argument that in Case No 1, we can prove that continuity of $H$ at $t_0=1$.
\textbf{Lemma 5.} Suppose that $0<\sigma<1$. Then there exists
$$
N=\max\limits_{0\leq t\leq 1}\int\limits_0^1 G(t,s)s^{-\sigma}ds.
$$
\textbf{\emph{Proof:}} Considering representation of the function $G(t,s)$ and evaluations of Lemma 4, we derive
$$
\begin{array}{l}
\int\limits_0^1 G(t,s)s^{-\sigma}ds=\frac{1}{\Gamma(\alpha)}\left[t^{\alpha-\sigma}B(1-\sigma, \alpha)-t^{\alpha-1}
\left(B(1-\sigma, \alpha-1)+\right.\right.\\
\left.\left.+(\alpha-2)B(2-\sigma, \alpha-1)\right)+(\alpha-1)t^{\alpha-2}B(2-\sigma, \alpha-1)\right].
\end{array}
$$
Taking
$$
B(1-\sigma,\alpha)=\frac{\alpha-1}{\alpha-\sigma}B(1-\sigma, \alpha-1);\,\,B(2-\sigma, \alpha-1)=\frac{1-\sigma}{\alpha-\sigma}B(1-\sigma, \alpha-1),
$$
into account we infer
$$
\begin{array}{l}
\int\limits_0^1 G(t,s)s^{-\sigma}ds=\\
=\frac{B(1-\sigma, \alpha-1)}{\Gamma(\alpha)}\left[\frac{\alpha-1}{\alpha-\sigma}t^{\alpha-\sigma}-\left(1+\frac{(\alpha-2)(1-\sigma)}{\alpha-\sigma}\right)t^{\alpha-1}
+\frac{(\alpha-1)(1-\sigma)}{\alpha-\sigma}t^{\alpha-2}\right].
\end{array}
$$
Denoting $L(t)=\int\limits_0^1 G(t,s)s^{-\sigma}ds$, from the last equality one can easily derive that $L(0)=0,\, L(1)=0$. Since $G(t,s)\geq 0$, then $L(t)\geq 0$ and as $L(t)$ is continuous on $[0,1]$, it has a maximum. This proves Lemma 5.
\textbf{Theorem 6.} Let $0<\sigma<1, \, 3<\alpha\leq 4,\, f: \, (0,1]\times [0,\infty)$ be continuous and $\lim\limits_{t\rightarrow 0^+} f(t,\cdot)=\infty$, $t^\sigma f(t,y)$ be continuous function on $[0,1]\times [0,\infty)$. Assume that there exist constants $0<\lambda\leq \frac{1}{N}$, and $\tau>0$ such that for $x,y \in [0,\infty)$ and $t\in [0,1]$
$$
t^\sigma |f(t,x)-f(t,y)|\leq \frac{\lambda |x-y|}{\left(1+\tau\sqrt{|x-y|}\right)^2}.
$$
Then Problem (1) has a unique non-negative solution.
\textbf{\emph{Proof:}} Consider the cone $P=\left\{u\in C[0,1]:\, u\geq 0\right\}$. Notice that $P$ is a closed subset of $C[0,1]$ and therefore, $(P,d)$ is a complete metric space where
$$
d(x,y)=\sup\left\{|x(t)-y(t)|:\, t\in[0,1]\right\}\,\, for\,\, x,y\in P.
$$
Now, for $u\in P$ we define the operator $T$ by
$$
(Tu)(t)=\int\limits_0^1 G(t,s) f(s,u(s))ds=\int\limits_0^1 G(t,s) s^{-\sigma} s^\sigma f(s,u(s))ds.
$$
In virtue of Lemma 4, for $u\in P$, $Tu\in C[0,1]$ and, since $G(t,s)$ and $t^\sigma f(t,y)$ are non-negative functions, $Tu\geq 0$ for $u\in P$. Therefore, $T$ applies $P$ into itself.
Next, we check that assumptions of Theorem 3 are satisfied. In fact, for $u,v\in P$ with $d(Tu,Tv)>0$, we have
$$
\begin{array}{l}
d(Tu,Tv)=\max\limits_{t\in [0,1]} \left|(Tu)(t)-(Tv)(t)\right|=\\
=\max\limits_{t\in [0,1]}\left|\int\limits_0^1 G(t,s)s^{-\sigma}s^\sigma \left(f(s,u(s))-f(s,v(s))\right)ds\right|\leq\\
\leq \max\limits_{t\in [0,1]}\int\limits_0^1 G(t,s)s^{-\sigma}s^\sigma \left|f(s,u(s))-f(s,v(s))\right|ds\leq\\
\leq \max\limits_{t\in [0,1]}\int\limits_0^1 G(t,s)s^{-\sigma}\frac{\lambda|u(s)-v(s)|}{\left(1+\tau\sqrt{|u(s)-v(s)|}\right)^2}ds\leq
\max\limits_{t\in [0,1]}\int\limits_0^1 G(t,s)s^{-\sigma}\frac{\lambda d(u,v)}{\left(1+\tau\sqrt{d(u,v)}\right)^2}ds=\\
=\frac{\lambda d(u,v)}{\left(1+\tau\sqrt{d(u,v)}\right)^2}\max\limits_{t\in [0,1]}\int\limits_0^1 G(t,s)s^{-\sigma}ds=
\frac{\lambda d(u,v)}{\left(1+\tau\sqrt{d(u,v)}\right)^2} N\leq \frac{d(u,v)}{\left(1+\tau\sqrt{d(u,v)}\right)^2},\\
\end{array}
$$
where we have used that $\lambda\leq\frac{1}{N}$ and the non-decreasing character of the function $\beta(t)=\frac{t}{\left(1+\tau\sqrt{t}\right)^2}$.
Therefore,
$$
d(Tu,Tv)\leq \frac{d(u,v)}{\left(1+\tau\sqrt{d(u,v)}\right)^2}.
$$
This gives us
$$
\sqrt{d(Tu,Tv)}\leq \frac{\sqrt{d(u,v)}}{1+\tau\sqrt{d(u,v)}}
$$
or
$$
\tau-\frac{1}{\sqrt{d(Tu,Tv)}}\leq-\frac{1}{\sqrt{d(u,v)}}
$$
and the contractivity condition of the Theorem 3 is satisfied with the function $\varphi(t)=-\frac{1}{\sqrt{t}}$ which belongs to the class $\mathfrak{F}$.
Consequently, by Theorem 3, the operator $T$ has a unique fixed point in $P$. This means that Problem (1) has a unique non-negative solution in $C[0,1]$. This finishes the proof.
An interesting question from a practical point of view is that the solution of Problem (1) is positive. A sufficient condition for that solution is positive, appears in the following result:
\textbf{Theorem 7.} Let assumptions of Theorem 6 be valid. If the function $t^\sigma f(t,y)$ is non-decreasing respect to the variable $y$, then the solution of Problem (1) given by Theorem 6 is positive.
\textbf{\emph{Proof:}} In contrary case, we find $t^*\in (0,1)$ such that $u(t^*)$=0. Since $u(t)$ is a fixed point of the operator $T$ (see Theorem 6) this means that
$$
u(t)=\int\limits_0^1 G(t,s) f(s,u(s))ds\,\, for\,\, 0<t<1.
$$
Particularly,
$$
0=u(t^*)=\int\limits_0^1 G(T^*,s)f(s,u(s))ds.
$$
Since that $G$ and $f$ are non-negative functions, we infer that
$$
G(t^*,s)f(s,u(s))=0\,\,\,a.e.\,(s)\eqno (2)
$$
On the other hand, as $\lim\limits_{t\rightarrow0^+}f(t,0)=\infty$ for given $M>0$ there exists $\delta>0$ such that for $s\in (0,\delta)$ $f(s,0)>M$. Since $t^\sigma f (t,y)$ is increasing and $u(t)\geq 0$,
$$
s^\sigma f(s,u(s))\geq s^\sigma f(s,0)\geq s^\sigma M\,\,\,for\,\,\, s\in (0,\delta)
$$
and, therefore, $f(s,u(s))\geq M$ for $s\in(0,\delta)$ and $f(s,u(s))\neq 0 \,\,a.e.\,\,(s)$. But this is a contradiction since $G(t^*,s)$ is a function of rational type in the variable $s$ and, consequently, $G(t^*,s)\neq 0 \,\,a.e.\,\,(s)$. Therefore, $u(t)>0$ for $t\in (0,1)$. This finishes the proof.
\section{Acknowledgement}
The present work is partially supported by the project MTM-2013-44357-P.
|
\section*{Introduction}
\label{sec:introduction}
\noindent
The accurate delivery of various cargoes is of great importance for maintaining the correct function of cells and organisms. Particles, such as vesicles, proteins, organelles, have to be transported to their specific destinations. In order to enable this cargo transfer, cells are equipped with a complex filament network and specialized motor proteins. The cytoskeleton serves as tracks for molecular motors. They convert the energy provided by ATP (adenosine triphosphate) hydrolysis into active motion along the cytoskeletal filaments, while they simultaneously bind to cargo \cite{Vale2000,Schliwa2003}. In addition to intracellular transport, the dynamic cytoskeleton and its associated motors also stabilize the cell shape, adjust it to different environmental circumstances, and drive cell motility or division \cite{TheCell}.\\
\noindent
The two main constituents of the cytoskeleton involved in intracellular transport are the polarized microtubules and actin filaments. In cells with a centrosome, the rigid microtubules grow radially from the central MTOC (microtubule organizing center) towards the cell periphery. In conjunction with the associated motor proteins kinesin and dynein, microtubules manage fast long-range transport between the cell center and periphery. In contrast to microtubules, which spread through the whole cell, actin filaments are mostly accumulated in a random fashion underneath the plasma membrane and construct the so called actin cortex \cite{TheCell}. Myosin motors operate on actin filaments and are therefore specialized for lateral transport in the cell periphery. Consequently, the cytoskeletal structure is very inhomogeneous and characterized by a
thin actin cortical layer \cite{Salbreux2012,Eghiaian2015}.\\
\noindent
The `saltatory' transport \cite{Rebhun1967} by molecular motors is a cooperative mechanism. Several motors of diverse species are simultaneously attached to one cargo \cite{Welte2004,Vershinin2007,Balint2013,Hancock2014}. This enables a frequent exchange between actin and microtubule based transport, which is necessary for specific search problems. A prominent example of collaborative transport on actin and microtubule networks is the motion of pigment granules in fish and frog melanophores \cite{Rodionov2003, Gross2002}.
The activity level of the particular motor species, and thus the share in cooperation, is regulated by cell signaling \cite{Welte2004,Mallik2004}.
In fish and frog, pigment granules are accumulated near the nucleus by extracellular stimuli transduced via PKA (protein kinase A) \cite{Rodionov2003, Gross2002}. The motor activity is thereby controlled via the level of cAMP (cyclic adenosine monophosphate). While low concentrations promote the action of dyneins, intermediate values stimulate myosin motors and high amounts of cAMP activate kinesins \cite{Rodionov2003}. Moreover, the infection of cells by adenoviruses triggers signaling through PKA and MAP (mitogen activated protein kinase), which enhances transport to the nucleus \cite{Suomalainen2001}; and the net transport of lipid droplets in Drosophila is directed to the cell center (periphery) by absence (presence) of the transacting factor Halo \cite{Gross2003,Welte2004}.\\
\noindent
Another aspect of intracellular transport is its intermittent nature.
Molecular motors perform two phases of motility. Periods of directed active motion along cytoskeletal filaments interfere with effectively stationary states \cite{Bressloff2013}. Intersection nodes of the cytoskeleton cause the molecular motors to pause until they either manage to squeeze through the constriction and pass it on the same filament or switch to the crossing track and thus change the direction of the transported cargo \cite{Balint2013,Ali2007,Ross2008,Ross2008B}.
Motor proteins also detach of the filaments out of chemical reasons. In the cytoplasm the cargoes experience subdiffusive dynamics due to crowding effects \cite{Weiss2004}. The displacement is limited to the vicinity of the detachment site and negligible compared to the one of the active motion phase. Hence, detachment and reattachment processes effectively contribute to waiting times. However, cargo particles preferentially change their direction of motion at cytoskeletal intersections, which constitute motion barriers. The mean distance between two intersections, the mesh size of a network, is typically smaller than the processive run length of a single molecular motor \cite{Snider2004}.\\
\noindent
The spatial organization of the cytoskeleton as well as the activity of the different motor species and their behavior at network intersections establish a typical stochastic movement pattern of intracellular cargo, which suggests a random walk description \cite{Benichou2011,Bressloff2013}.
The narrow escape problem describes a common search process, where the target destination is represented by a specific, small region on the plasma membrane of a cell. Typical examples involve secretion processes in Cytotoxic T Lymphocytes (CTL) which play a key role in immune response and defeat tumorigenic or virus-infected cells. When in contact with an antigen-presenting cell, CTL form a connection with a diameter of the order of microns to it, which is called immunological synapse. Lytic granules are actively transported and released at the immunological synapse. They contain perforin and granzymes which induce apoptosis of the pathological cell. In order to prevent unintended damage of neighboring cells, the release of lytic granules is strictly confined to the immunological synapse. Thus, the transport of lytic granules towards the immunological synapse constitutes a narrow escape problem to be solved by CTL \cite{Grakoui1999,Qu2011,Angus2013,Ritter2013}. Moreover, the outgrowth of dendrites or axons from neurons \cite{Alberts2003,Chada2003} as well as repair mechanisms for corruptions of a cell's plasma membrane \cite{McNeil2005,Andrews2014} require the target-oriented transport of mitochondria and vesicles.\\
\noindent
In prior work on the narrow escape problem, Schuss et al. analytically investigated first passage properties of a purely diffusive searcher \cite{Schuss2007,Schuss2012}, while B\' enichou et al. identified benefits of search efficiency to small targets on the surface of spherical domains by intermittent phases of surface-mediated and bulk diffusion \cite{Benichou2010}.
Ballistic motion along the cytoskeleton and the different characteristics of microtubule- and actin-based transport was taken into account by Slepchenko et al. in order to model the spreading of pigment granules in melanophores \cite{Slepchenko2007}. They were able to determine the switching rate between microtubules and actin filaments by fitting their theoretical results to experimental data of aggregation and dispersion of pigment granules in fish melanophores, whose cytoskeleton appears to be rather homogeneous.
In order to study first passage properties of cargo transport from the nucleus to the complete plasma membrane, Ando et al. recently considered an inhomogeneous network distribution in which the cytoskeleton is confined to a delimited shell \cite{Ando2015}. Within a continuum model of increased bulk diffusion, they found that the transit time can be significantly reduced when the cytoskeletal shell is placed close to the nucleus. Moreover, they explicitly modeled cytoskeletal networks and investigated the impact of number, length and polarity of filaments as well as detachment and reattachment processes by considering intermittent phases of ballistic transport and cytoplasmic diffusion. Via extensive computer simulations, they showed, inter alia, that an outward-directed network polarity expectedly improves transit times while the actual distribution of filament orientations is less effective.
Nonetheless, transport in cells comprises an intricate interplay between motor performance and spatial organization of the cytoskeleton, which is generally inhomogeneous itself. The description of this interaction is a challenging theoretical task and we still lack precise knowledge about how cells adapt to various transport tasks and especially to narrow escape problems.\\
\noindent
In the following, we present a coarse grained model of intracellular transport by considering the effective movement between network nodes, while discarding the single steps of individual motors at the molecular level. We introduce a random velocity model with intermittent arrest states where the dynamic cytoskeleton is implicitly modeled by probability density functions for network orientation and mesh size.
The proposed model allows the study of diverse transport tasks. Here we focus on the narrow escape problem and address the effects of the interplay between inhomogeneous cytoskeletal architecture and motor performance on the search efficiency to small targets alongside the membrane of spherical cells.
\section*{Model}
\label{sec:model}
\subsection*{General Random Velocity Model}
\noindent
Cells establish specific search strategies for cargo transport by alterations of the cytoskeletal organization and regulation of the motor behavior at network intersections. In order to study the efficiency of spatially inhomogeneous search strategies we formulate a random velocity model in continuous space and time composed of two states of motility: (i) a ballistic motion state at constant speed $v{=}1$, which corresponds to active transport by molecular motors in between two successive intersections of the filamentous network and (ii) a waiting state, which is associated to pauses at intersection nodes of the cytoskeleton. The swaps from one state to another are arranged via constant but generally asymmetric transition rates $k_{m\rightarrow w}$ for a switch from motion to waiting and $k_{w\rightarrow m}$ for an inverse transition, see figure \ref{figure1} (b). These lead to exponentially distributed time periods $t_m$, $t_w$ spent in each state of motility
\begin{align}
p(t_m) &= k_{m\rightarrow w} e^{-k_{m\rightarrow w}t_m}, \\
p(t_w) &= k_{w\rightarrow m} e^{-k_{w\rightarrow m}t_w},
\end{align}
and mean residence times of $1/k_{m\rightarrow w}$, $1/k_{w\rightarrow m}$ for the motion and waiting state, respectively, which is biologically consistent as active lifetimes of cargo particles are exponentially distributed \cite{Arcizet2008}. The event rates are directly connected to biologically tractable properties of the cytoskeleton and the motor proteins. The mesh size $\ell$ of the underlying cytoskeletal network, which reflects the typical distance between two consecutive intersections, defines the rate $k_{m\rightarrow w} {=} v/\ell$ for a transition from the ballistic motion to the pausing state. Whereas the event rate $k_{w\rightarrow m}$ is determined by the characteristic waiting time at an intersection node.\\
\noindent
Whenever a particle has reached a network intersection and paused, it may either keep moving processively along the same filament with probability $p$ or it may change to a crossing track with probability $(1{-}p)$, as sketched in figure \ref{figure1} (d). This provides a typical timescale $[(1-p)k_{w\rightarrow m}]^{-1}$ ($[pk_{w\rightarrow m}]^{-1}$) of changing (remaining on) the filament subsequent to a waiting period. With regard to the rotational symmetry of a cell, the new direction $\theta {=} \phi +\alpha_\text{rot}$ is always chosen with respect to the radial direction $\tan(\phi){=}y/x$. The rotation angle $\alpha_\text{rot}$ is drawn from a distribution $f(\alpha_\text{rot})$ which is characteristic for the underlying cytoskeletal network. Due to the inhomogeneous structure of the cytoskeleton, $f(\alpha_\text{rot})$ depends on the location of the particle inside the cell.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figure1e-eps-converted-to.pdf}
\caption
{Random Velocity Model for Intracellular Search.
\textbf{a} The model geometry displays the confined and inhomogeneous architecture of a cell of radius $R_\text{membrane}$. While the interior is filled with radial microtubules, the periphery is dominated by random actin filaments. The width of the actin cortical layer is denoted by $\delta$ and the narrow escape hole has an angular diameter $\alpha_\text{target}$. \textbf{b} Pausing processes arise due to constricting intersection nodes of the cytoskeletal network. This is accounted in the model by two motility states with specific transition rates $k_{m\rightarrow w}$ and $k_{w\rightarrow m}$. \textbf{c} The rotation angle distributions $f_\text{K}$ and $f_\text{D}$ for kinesins and dyneins acting on microtubules are delta-peaked, while the ones of myosins may be uniformly distributed ($f^u_\text{M}$) or gaussian distributed ($f^g_\text{M}(\mu,\sigma)$) with various expectation values $\mu$ and standard deviations $\sigma$ regarding the orientation of actin filaments. \textbf{d} Sketch of the walk during two consecutive events and the corresponding transition rates, where $\theta$ denotes the new and $\Theta$ the previous direction.
}
\label{figure1}
\end{figure*}
\subsection*{Model Geometry}
\noindent
Our model system is designed according to the inhomogeneous internal organization of a cell, see figure \ref{figure1} (a). We assume a circular confined geometry of radius $R_\text{membrane}$, which displays the plasma membrane and will be fixed in the following ($R_\text{membrane}{=}1$). The cytoplasm is split into an interior region, where only microtubules are present, and a periphery, which is dominated by the actin cortex but may also be pervaded by microtubules. The width of the actin cortical layer is denoted by $\delta$, so that an internal margin of radius $R_\text{internal}{=}R_\text{membrane}-\delta$ emerges. Throughout this article, the target destination of the cargo is assumed to be a narrow escape hole in the plasma membrane with opening angle $\alpha_\text{target}$.
\subsection*{Random Velocity Model in the Interior}
\noindent
The interior of a cell is controlled by the radial network of microtubules with its associated kinesin and dynein motors. Since they manage fast long-range transport inside a cell, the rate to switch from the motion to the waiting state $k^i_{m\rightarrow w} {=} 0$ is fixed to zero for simplicity. This leads to uninterrupted radial movement along the internal microtubule network. However, the particle may be forced into the waiting state due to confinement events, as dyneins are assumed to stop at the central MTOC. Hence the transition rate to the motion state is set to a non-zero value $k^i_{w\rightarrow m}$.
\subsection*{Random Velocity Model in the Periphery}
\noindent
Due to the complex network structure of the periphery the searcher frequently encounters intersection nodes at rate $k^p_{m\rightarrow w}$ and switches to the waiting state. Subsequently, the particle may either keep moving along the previously used track at rate $p \, k^p_{w\rightarrow m}$ or it may change to a crossing filament at rate $(1-p) \, k^p_{w\rightarrow m}$. The rotation angle is thereby drawn of a distribution
\begin{align}
f(\alpha_\text{rot}) = q_\text{K} \, f_\text{K}(\alpha_\text{rot})+q_\text{D} \, f_\text{D}(\alpha_\text{rot})+q_\text{M} \, f_\text{M}(\alpha_\text{rot}),
\end{align}
which specifies the peripheral environment in terms of the filament orientation ($f_i(\alpha_\text{rot})$) and motor species activity ($q_i$). The probabilities $q_\text{K}$, $q_\text{D}$, $q_\text{M}$ correspond to directional changes induced by kinesin, dynein and myosin, respectively, with $q_\text{K}+q_\text{D}+q_\text{M}=1$. With regard to the radial orientation of microtubules and the directionality of the motors, the rotation angle distributions associated to kinesins and dyneins are delta peaked
\begin{align}
f_\text{K}(\alpha_\text{rot}) & = \delta(\alpha_\text{rot}),\quad~~\,\qquad\quad\text{ for } \alpha_\text{rot} \in ({-}\pi;\pi],\\
f_\text{D}(\alpha_\text{rot}) & = \delta(\alpha_\text{rot}-\pi),\qquad\quad\text{ for } \alpha_\text{rot} \in ({-}\pi;\pi].
\end{align}
In case of a directional change initiated by myosins, the rotation angle distribution is assumed to be either uniform or cut-off-gaussian, which takes into account the randomness of actin networks
\begin{align}
f_\text{M}^\text{u}(\alpha_\text{rot}) & = \frac{1}{2\pi},\qquad\qquad\qquad\quad\text{for } \alpha_\text{rot} \in ({-}\pi;\pi],\\
f_\text{M}^\text{g}(\alpha_\text{rot}) & = \frac{1}{2} f^+ + \frac{1}{2} f^-,
\label{eq:rotanglemyosin}
\end{align}
with
\begin{align}
f^+ & = \frac{\mathcal{N}}{\sigma\sqrt{2\pi}}\exp\left({-}\frac{(\alpha_\text{rot}{-}\mu)^2}{2\sigma^2}\right),\text{ for } \alpha_\text{rot} \in [0;\pi],\\
f^- & = \frac{\mathcal{N}}{\sigma\sqrt{2\pi}}\exp\left({-}\frac{(\alpha_\text{rot}{+}\mu)^2}{2\sigma^2}\right),\text{ for } \alpha_\text{rot} \in [{-}\pi;0],
\end{align}
where $\mu>0$ and $\mathcal{N}$ denotes the normalization constant, see figure \ref{figure1} (c). In conclusion, the peripheral network is characterized by a mean mesh size $1/k_{m\rightarrow w}$, its structure is reflected by the rotation angle distributions $f_i(\alpha_\text{rot})$ and the motor activity is defined via $q_i$.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figure2d-eps-converted-to.pdf}
\caption
{Homogeneous Search Strategy. \textbf{a} Sample trajectories (blue lines) during homogeneous search strategies on a circle of radius $R_\text{membrane}{=}1$ with a membranous narrow escape of opening angle $\alpha_\text{target}{=}0.1$ (red zone). The transition rate to the waiting state is fixed to $k_{m\rightarrow w}{=}0$. Consequently, the cargo particle does not encounter any network intersections and hence does not change its direction of motion inside the bulk. For $q_\text{M}{=}0$ the network is completely radial, whereas it is uniformly distributed in the case of $q_\text{M}{=}1$. Intermediate values of the probability $q_\text{M}$ correspond to homogeneous combinations of radial microtubules and random actin filaments, which is reflected by the trajectories. \textbf{b} Sample trajectories for $k_{m\rightarrow w}{=}1$. The searcher frequently changes its direction at the network intersections (blue crosses) according to the probabilities $q_\text{M}$, $q_\text{K}{=}q_\text{D}{=}(1-q_\text{M})/2$ and their associated rotation angle distributions until it reaches the membranous target zone $\alpha_\text{target}{=}0{.}1$. \textbf{c} MFPT in dependence of the target size $\alpha_\text{target}$ for different values of the transition rate $k_{m\rightarrow w}$ and various motor activities $q_\text{M}$, $q_\text{K}{=}q_\text{D}{=}(1-q_\text{M})/2$ for uniformly distributed ($f_\text{M}^\text{u}$) as well as outward-directed ($f_\text{M}^\text{g}(\mu{=}0,\sigma{=}1)$) actin structures.}
\label{figure2}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figure11c-eps-converted-to.pdf}
\caption
{Sample Trajectories for Homogeneous Search Strategies with $\alpha_\text{target}{=}0{.}1$. \textbf{a} Motion on a radial microtubule network ($q_\text{M}{=}0$), where $k_{m\rightarrow w}{=}0$. \textbf{b} Trajectory for a cargo moving along a uniformly random actin network ($q_\text{M}{=}1$, $f_\text{M}^\text{u}$) with $k_{m\rightarrow w}{=}10$. \textbf{c} Motion on a gaussian distributed actin network ($q_\text{M}{=}1$, $f_\text{M}^\text{g}(\mu,\sigma)$) with a mesh size defined by $k_{m\rightarrow w}{=}10$ for various $\mu$ and $\sigma$. For instance $\sigma{=}1$ demonstrates the diverse residence areas of the cargo with respect to $\mu$. While for $\mu<\pi/2$ the particle predominantly moves close to the boundary, for $\mu > \pi/2$ it is pushed away from the membrane. By increasing $\sigma$ the motion pattern gets randomized due to the broadening of the rotation angle distribution $f_\text{M}^\text{g}(\mu,\sigma)$.
}
\label{figure2b}
\end{figure*}
\subsection*{Confinement Events}
\noindent
The spatial geometry of the model system imposes various confinement events, which are further specified in the following.
At the onset, each searcher is assumed to start its walk in the center of the cell, where it is linked to a kinesin and runs in a uniformly distributed initial direction.
As soon as a particle encounters the outer membrane margin at radius $R_\text{membrane}$ it will switch into the pausing state, since it is assumed to detach of the filament, and check for the target zone $\alpha_\text{target}$. If it is found, the walk will be terminated, otherwise the particle will wait at rate $(1-p) \, k^p_{w\rightarrow m}$ and the rotation angle distribution will be restricted to allowed values.
The same holds in the case that a cargo transported by myosin hits the internal margin $R_\text{internal}$, which is created by the structural inhomogeneity of the cytoskeleton. Crossovers of the internal margin by kinesins or dyneins, happen uninterruptedly under a change of the characteristic event rates for interior and periphery.
Whenever a dynein coupled particle reaches the MTOC in the center of the cell, it will wait with rate $k^i_{w\rightarrow m}$ before it will change to kinesin motion in a uniformly distributed direction.
Consequently, the mean waiting time at confinement events is not necessarily the same as the one at network intersections in the bulk. In general it will be larger.
\section*{Results}
\noindent
In the following, we perform extensive Monte Carlo simulations in order to analyze the dependence of the search efficiency to narrow escapes alongside a cell's membrane on the spatial organization of the cytoskeleton as well as the motor performance at network intersections. For that purpose we define the mean first passage time (MFPT) to a target as the ensemble average over first passage events of $5\times 10^5$ independent realizations of the walk, in which all cargoes are initially located at the center of the cell and start moving in an uniformly distributed direction $\Theta\in[-\pi;\pi)$. At first, we will consider search strategies, where the motors operate on homogeneous cytoskeletal networks. Then, we will focus on the inhomogeneous architecture and elaborate the influence of the actin cortical width on the mean first passage properties of narrow escape problems.
\begin{figure*}[t]
\centering
\includegraphics[width=0.97\linewidth]{figure4-eps-converted-to.pdf}
\caption
{Inhomogeneous Search Strategy.
Interior parameters: $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}\infty$; periphery parameters: $k^p_{m\rightarrow w}\in \{0;1;10\}$, $k^p_{w\rightarrow m}{=}\infty$, $p{=}0$, $q_\text{K}{=}q_\text{D}{=}(1-q_\text{M})/2$, $q_\text{M}\in \{0;1/3;1\}$; target size: $\alpha_\text{target}\in\{0{.}1;0{.}01\}$. \textbf{a} MFPT to a narrow escape of opening angle $\alpha_\text{target}{=}0{.}1$ for an inhomogeneous cytoskeleton in dependence of the actin cortical width $\delta$. Since $k^p_{w\rightarrow m}{=}\infty$, the searcher changes its direction instantaneously at each cytoskeletal intersection in the periphery, which occur at various rates $k^p_{m\rightarrow w}$. The new direction is thereby chosen with probability $q_\text{M}$ according to a uniformly rotation angle distribution $f_\text{M}^\text{u}(\alpha_\text{rot})$, or with probabilities $q_\text{K}{=}q_\text{D}{=}(1-q_\text{M})/2$ according to a rotation angle drawn from $f_\text{K}(\alpha_\text{rot})$, $f_\text{D}(\alpha_\text{rot})$, respectively, thus the motors are not processive $p{=}0$ at network intersections. The dashed black line refers to a homogeneous search strategy on a pure and radial microtubule network. The insets show some sample trajectories (blue lines) for $\delta{=}0.3$ and different values of $q_\text{M}$. \textbf{b} MFPT to a narrow opening of angular diameter $\alpha_\text{target}{=}0{.}01$. The inset is a detail of $q_\text{M}{=}1$ for $\delta \in \left[0;0{.}1\right]$, which emphasized the shift of the optimal cortical width by a decrease of the target size.
}
\label{figure4}
\end{figure*}
\subsection*{Homogeneous Search Strategy}
\noindent
Here, we neglect the inhomogeneous structure of the cytoskeleton and assume that it spreads through the whole cell in a homogeneous manner. Since we aim to isolate the impact of the cytoskeletal architecture, we ignore the motor's processivity ($p{=}0$) and waiting processes ($k_{w\rightarrow m}{=}\infty$). Hence at each network intersection the particle immediately changes its direction according to
\begin{align}
f(\alpha_\text{rot}) = q_\text{K} \, f_\text{K}(\alpha_\text{rot})+q_\text{D} \, f_\text{D}(\alpha_\text{rot})+q_\text{M} \, f_\text{M}(\alpha_\text{rot}),
\end{align}
where $f_\text{M}\in\{f_\text{M}^\text{u};f_\text{M}^\text{g}(\mu,\sigma)\}$ and $q_\text{K}=q_\text{D}=(1-q_\text{M})/2$. In the case of $q_\text{M}{=}0$ myosin motors are deactivated and the transport is managed by kinesin and dynein on a radial network of microtubules, while $q_\text{M}{=}1$ leads to a pure actin mesh with myosins and either uniformly or gaussian distributed filaments. Intermediate values of $q_\text{M}$ correspond to cooperative transport on homogeneous combinations of microtubules and actin filaments with active kinesins, dyneins and myosins. Sample trajectories which implicitly reflect the network structure are given in figure \ref{figure2} for uniformly random actin orientations.\\
\noindent
In order to investigate the influence of the network composition and motor activity, we measure the MFPT in dependence of the target size $\alpha_\text{target}$ for different event rates $k_{m\rightarrow w}$ and various probabilities $q_\text{M}$. As expected, the MFPT increases monotonically with decreasing opening angle $\alpha_\text{target}$, see figure \ref{figure2} (c). Apparently, directional changes at intersections of a homogeneous cytoskeletal network do not improve the search efficiency nor does a cooperative behavior of different motor species ($q_\text{M}\in (0;1)$). The MFPT increases with greater transition rate $k_{m\rightarrow w}$ and is quite robust against alterations of the probability $q_\text{M}$. However, we notice that a random actin mesh ($q_\text{M}{=}1$) is preferable to a radial microtubule network ($q_\text{M}{=}0$).\\
\noindent
As expected, figure \ref{figure2} (c) shows that an outward-oriented actin network with $f_\text{M}^\text{g}(\mu,\sigma)$ and for example $\mu{=}0\!<\!\pi/2$, $\sigma{=}1$ is advantageous for large target sizes. This was also stated by Ando et al. for transport from the cell center to the whole membrane, i.e. $\alpha_\text{target}{=}2\pi$ \cite{Ando2015}. Remarkably, the search for small targets does also benefit from outward-directed actin filaments. While inward-directed ($\mu>\pi/2$) actin polarities drive the cargo towards the cell center, which lowers the probability to reach the membrane and detect the target, outward-directed ($\mu<\pi/2$) actin networks push the particle towards the plasma membrane, as indicated by the sample trajectories in figure \ref{figure2b}. This introduces a topologically induced technique to scan the membrane for small targets as the cargo is predominantly moving in vicinity of the cell boundary. This effect vanishes for increasing $\sigma$, since the resulting broadening of the distribution $f_\text{M}^\text{g}(\mu,\sigma)$ leads to approximately uniformly distributed filaments ($f_\text{M}^\text{u}$) and thus a randomization of the search. Consequently, as evident from figure \ref{figure2} (c) for small target sizes, a homogeneous outward-directed actin structure may be more efficient (even for $k_{m\rightarrow w}\neq 0$) than the best search strategy on a homogeneous microtubule network (i.e. $k_{m\rightarrow w}{=}0$). The reason is that large excursions to the cell center are inhibited by outward-directed actin polarities, while detours to the center are necessary for transport along microtubules in order to change radial direction and reach the target.
\subsection*{Inhomogeneous Search Strategy}
\noindent
In case of a homogeneous cytoskeleton, the most efficient search strategy is an uninterrupted motion on a pure actin network without any directional changes in the bulk. Such a motion scheme is sufficient for large membranous targets, but generally fails for narrow escape problems. May an inhomogeneous search strategy, like it is found in living cells, be the key to the efficient detection of small departure zones on the plasma membrane? Guided by this question, we check for the influence of the actin cortical width $\delta$ on the effectiveness of the narrow escape search problem.
\subsubsection*{Influence of the Actin Cortical Width}
\noindent
In order to investigate the pristine effect of an inhomogeneous cytoskeleton, we perform Monte Carlo simulations and assume $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}\infty$ in the interior and $k^p_{m\rightarrow w}\in \{0;1;10\}$, $k^p_{w\rightarrow m}{=}\infty$, $p{=}0$ in the periphery. Consequently, the cargo changes its direction of motion instantaneously at each intersection node. Figure \ref{figure4} displays the resulting MFPT to small membranous target sites ($\alpha_\text{target}{=}0{.}1$ and $\alpha_\text{target}{=}0{.}01$) in dependence of the actin cortical width $\delta$ for different peripheral mesh sizes defined by the node encountering rate $k^p_{m\rightarrow w}$ and motor activities $q_\text{M}$, $q_\text{K}{=}q_\text{D}{=}(1-q_\text{M})/2$ for $f_\text{M}^\text{u}$.\\
\noindent
Remarkably, we find that the MFPT responds sensitively to alterations of the motor activity $q_\text{M}$ and exhibits a nontrivial minimum for $q_\text{M}\neq 0$ and small targets. In the case of $q_\text{M}{=}0$, transport is solely managed by kinesins and dyneins on a radial but generally inhomogeneous microtubule network. The inhomogeneity is determined by $\delta$ and results in a change of the mesh size from the cell's interior (with $k^i_{m\rightarrow w}{=}0$) to the periphery (with various $k^p_{m\rightarrow w}$). Hence, for $k^p_{m\rightarrow w}{=}k^i_{m\rightarrow w}{=}0$ the network is homogeneous and the MFPT is constant in $\delta$, as shown in figure \ref{figure4}. Contrarily, the particle frequently changes its direction in the peripheral bulk according to $k^p_{m\rightarrow w}\neq 0$, which leads to a local back-and-forth motion of the cargo on the same filament for $q_\text{M}{=}0$. This back-and-forth pattern significantly hinders both the hitting with the membrane as well as the retraction to the central MTOC, which is necessary in order to change radial direction and detect the membranous target. Consequently, the MFPT monotonically increases with growing cortical width $\delta$, because a larger peripheral area increases the interruption by back-and-forth motion. In the case of $q_\text{M}\neq 0$, the minimum of the MFPT for $\delta \in \left(0;1\right)$ is most prominent for $q_\text{M}{=}1$, i.e. a peripheral motion dominated by transport along actin filaments, achieved by a high activity level of myosins, is most efficient. Contrarily, a high activity of dyneins would be favourable for virus trafficking or aggregation of pigment granules near the nucleus \cite{Slepchenko2007,Rodionov2003}. But for small membranous targets, inhomogeneous networks lead to a considerable gain of efficiency in comparison to the corresponding homogeneous limits $\delta\!\to\! 1$ and $\delta\!\to\! 0$. The case $\delta{=}1$ leads to a homogeneous search strategy which is defined by the features of the cell's periphery, i.e. the network possesses a mesh size defined by $k^p_{m\rightarrow w}$ and is composed of actin filaments and microtubules according to the probabilities $q_\text{M}$, $q_\text{K}{=}q_\text{D}{=}(1-q_\text{M})/2$. For instance $q_\text{M}{=}1$ and $\delta{=}1$ corresponds to intracellular transport on a homogeneous and uniformly random actin network. Contrarily, for $\delta{=}0$ the cargo's characteristics are purely determined by the cell's interior via $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}\infty$, which leads to motion along a homogeneous network of radial microtubules. \\
\noindent
While the MFPT is continuous in the limit $\delta\!\to\! 1$, it diverges for $\delta \!\to\! 0$ and $q_{M}\neq 0$. Consequently, a comparison to MFPT$(\delta{=}0)$ is not directly possible and we include this limit, which we studied in figure \ref{figure2}, by the black dashed line in figure \ref{figure4} and following figures. The divergence of the MFPT for $\delta\!\to\! 0$ originates from a motility restriction by a narrow cortex. Cargo particles which are transported by myosins get localized for narrow actin cortices, since the resulting prompt collisions with the two margins at radii $R_\text{internal}$ and $R_\text{membrane}$ inhibit substantially large displacements. Thus, the actin cortex can also act as a barrier for transport to a cell's plasma membrane.
The results shown in figure \ref{figure4} visualize that an increased chance of directional alterations by increase of $k^p_{m\rightarrow w}$, leads to a loss of search efficiency. This directly suggests a profit by a processive behavior of motor proteins at intersection nodes. Furthermore, a decrease in the opening angle $\alpha_\text{target}$ provokes a more prominent minimum of the MFPT at lower cortical widths $\delta$, compare figure \ref{figure4} (a) and (b).
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{figure5d-eps-converted-to.pdf}
\caption
{Influence of the Waiting Time Distribution on Inhomogeneous Search Strategies.
Interior parameters: $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}k_{w\rightarrow m}$; periphery parameters: $k^p_{m\rightarrow w}{=}k_{m\rightarrow w}$, $k^p_{w\rightarrow m}{=}k_{w\rightarrow m}$, $p{=}q_\text{K}{=}q_\text{D}{=}0$, $q_\text{M}{=}1$; target size: $\alpha_\text{target}{=}0{.}1$. \textbf{a} MFPT for $k_{m\rightarrow w}{=}10$ and various $k_{w\rightarrow m}$ in dependence of the actin cortex width $\delta$. The case of $k_{w\rightarrow m}{=}\infty$ corresponds to instantaneous directional changes. A systematic increase in the mean waiting time per arrest state, i.e. a decrease in $k_{w\rightarrow m}$, heightens the MFPT. Moreover, the optimal value of the actin cortical width $\delta_\text{opt}$, which minimizes the MFPT, is shifted for decreasing transition rates $k_{w\rightarrow m}$. This is further illustrated by the inset, where the MFPT is normalized according to $\text{MFPT}_\text{norm}=\text{MFPT}/\text{MFPT}(\delta{=}1)$. \textbf{b} The MMT depends on the rate $k_{m\rightarrow w}$ and exhibits a minimum for $\delta$. It does not depend on the event rate $k_{w\rightarrow m}$ and is of course equivalent to the MFPT without any waiting processes ($k_{w\rightarrow m}{=}\infty$). \textbf{c} The MWT for $k_{m\rightarrow w}{=}10$ displays a minimum and is shifted by decreasing transition rates $k_{w\rightarrow m}$ (the colors correspond to part a). This is further illustrated by the inset, a transformation to the mean number of waiting periods via $\text{MWT}\!\times\! k_{w\rightarrow m}$ is applied for various $k_{m\rightarrow w}$.
}
\label{figure5}
\end{figure*}
\subsubsection*{Influence of the Waiting Time Distribution}
\noindent
So far the waiting times, which are induced by intersection nodes and confinement events, have been neglected by fixing the rate to switch from waiting to motion to infinity, $k^i_{w\rightarrow m}{=}\infty$, $k^p_{w\rightarrow m}{=}\infty$. Here we address the impact of the mean waiting time per arrest state. For that purpose we assume radial microtubules in the cell interior, $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}k_{w\rightarrow m}$. The motion in the periphery is non-processive ($p{=}0$) and performed on a random actin network ($q_\text{M}{=}1$) of a given mesh size, defined by $k^p_{m\rightarrow w}{=}k_{m\rightarrow w}$. The mean waiting time is determined by $k^p_{w\rightarrow m}{=}k_{w\rightarrow m}$ and the target size is fixed to $\alpha_\text{target}{=}0{.}1$.\\
\noindent
As expected, a systematic decrease of the rate $k_{w\rightarrow m}$, i.e. increase of the mean waiting time per arrest state, extends the MFPT in comparison to instantaneous directional changes for $k_{w\rightarrow m}{=}\infty$, as shown in figure \ref{figure5} (a). Remarkably, the position of the minimum, which determines the optimal cortical width $\delta_\text{opt}$, is shifted by a reduction in $k_{w\rightarrow m}$. This is further illustrated in the inset of figure \ref{figure5} (a) by normalization of the MFPT to $\text{MFPT}_\text{norm}=\text{MFPT}/\text{MFPT}(\delta{=}1)$.\\
\noindent
The change in $\delta_\text{opt}$ is based on the enhanced impact of waiting times on the first passage properties. The MFPT is composed of the total mean motion time (MMT) and the total mean waiting time (MWT) which the cargo experiences in the course of the search. figure \ref{figure5} (b) shows that the MMT displays pronounced minima and is indeed independent of the rate $k_{w\rightarrow m}$ and thus fully determined by the mesh size of the network given via $k_{m\rightarrow w}$. In contrast to that, the MWT, given in figure \ref{figure5} (c), depends on both rates $k_{m\rightarrow w}$ and $k_{w\rightarrow m}$. Since the rate $k_{w\rightarrow m}$ determines the mean waiting time per arrest state, it influences the MWT only via a multiplicative factor, as obvious by the shift in figure \ref{figure5} (c) for $k_{m\rightarrow w}{=}10$. The impact of the transition rate $k_{w\rightarrow m}$ on the MWT can be excluded by the transformation to the mean number of waiting periods
\begin{align}
\#\,\text{waiting periods} = \text{MWT} \times k_{w\rightarrow m}.
\end{align}
The inset of figure \ref{figure5} (c) displays its dependence on $k_{m\rightarrow w}$ and $\delta$. Remarkably, the number of waiting periods may also exhibit a minimum for small cortical widths.\\
\noindent
Consequently, we can solve the question of the shift in $\delta_\text{opt}$ for $k_{m\rightarrow w}{=}10$ presented in figure \ref{figure5} (a). In the case of $k_{m\rightarrow w}{=}10$, the MMT exhibits a minimum at roughly $\delta_\text{opt}{=}0{.}1$ (figure \ref{figure5} (b)), while a minimum at $\delta_\text{opt}{=}0{.}2$ emerges for the number of waiting periods and thus also for the MWT (figure \ref{figure5} (c) and inset). Hence, the optimal cortical width $\delta_\text{opt}$ of the MFPT, which is the sum of MMT and MWT, undergoes a crossover from $\delta_\text{opt}{=}0{.}1$ to $\delta_\text{opt}{=}0{.}2$ as the impact of waiting times increases with decreasing rate $k_{w\rightarrow m}$.\\
\noindent
In conclusion, the MWT dominates the behavior of the first passage properties in the limit of $k_{w\rightarrow m}\rightarrow 0$, which is biologically relevant as the mean waiting time per arrest state is typically of the order of seconds \cite{Zaliapin2005,Slepchenko2007,Balint2013}. This results in a shift of the optimal width $\delta_\text{opt}$ in dependence of the mesh size of the underlying network defined by $k_{m\rightarrow w}$. For small rates $k_{m\rightarrow w}$ a raising impact of waiting times may lead to a shift of $\delta_\text{opt}$ from intermediate values to $\delta_\text{opt}{=}1$ (compare to the inset of figure \ref{figure5} (c)) and thus a favour of homogeneous cytoskeleton. Larger rates $k_{m\rightarrow w}$, and hence biologically relevant mesh sizes, conserve the gain of search efficiency by inhomogeneous cytoskeletal organizations, since the number of waiting periods exhibits a minimum for $\delta \in (0;1)$ (inset of figure \ref{figure5} (c)).
\begin{figure*}[t]
\centering
\includegraphics[width=0.97\linewidth]{figure6d-eps-converted-to.pdf}
\caption
{Influence of the Rotation Angle Distribution of the Actin Cortex on Inhomogeneous Search Strategies.
Interior parameters: $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}\infty$; periphery parameters: $k^p_{m\rightarrow w}{=}10$, $k^p_{w\rightarrow m}{=}\infty$, $p{=}q_\text{K}{=}q_\text{D}{=}0$, $q_\text{M}{=}1$; target size: $\alpha_\text{target}{=}0{.}1$. \textbf{a} Sample trajectories for $\delta{=}0{.}3$ and various values of the rotation angle distribution of actin filaments. \textbf{b} MFPT in dependence of the actin cortical width $\delta$ for various values of the rotation angle distribution of actin filaments $f_\text{M}$.
}
\label{figure6}
\end{figure*}
\subsubsection*{Influence of the Rotation Angle Distribution of the Actin Cortex}
\noindent
The polarity of the actin filaments in the cortex is typically random. Hence we have previously investigated uniform rotation angle distributions $f_\text{M}^\text{u}$. \\
However, actin filaments may align to the radial microtubule network \cite{Lopez2014} and for instance the protein complex Arp2/3 induces a formation of actin branches at a distinct angle ($\approx 70^\circ$) compared to the parent filament \cite{Mullins1998,Risca2012}. In general, such mechanisms influence the orientational distribution of the actin filaments. \\
\noindent
Here, we study the impact of the expectation value and width of a cut-off-gaussian rotation angle distribution $f_\text{M}^\text{g}$ (as given in equation \ref{eq:rotanglemyosin}) on the search efficiency to narrow membranous targets. A mean value $\mu\!\in\![0;\pi/2)$ ($\mu\!\in\!(\pi/2;\pi]$) leads to outwardly (inwardly) peaked actin filaments, whereas $\mu{=}\pi/2$ corresponds to a predominantly lateral orientation. In order to isolate the influence of the orientational distribution we assume $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}\infty$ and $k^p_{m\rightarrow w}{=}10$, $k^p_{w\rightarrow m}{=}\infty$. Furthermore, we fix $p{=}0$, $q_\text{M}{=}1$ and $\alpha_\text{target}{=}0{.}1$.\\
\noindent
Figure \ref{figure6} displays the MFPT in dependence of the cortex width $\delta$, which defines the inhomogeneity of the system, for various actin orientations defined by the uniform distribution $f^u_\text{M}$ or the gaussian distribution $f^g_\text{M}(\mu,\sigma)$. Please first focus on $\sigma{=}1$, which corresponds to strongly peaked gaussian distributions $f^g_\text{M}(\mu,\sigma)$ of the actin filaments into directions defined by $\mu$.\\
\noindent
Inward-directed actin polarities ($\mu\!>\!\pi/2$) drive the cargo towards the cell center, which lowers the probability to reach the membrane and detect the target. As evident from the sample trajectory in figure \ref{figure6}, decreasing the actin cortical width $\delta$ draws the particle nearer to the cell's boundary. Consequently, a thin actin cortex is essential in order to improve the search efficiency to membranous targets as manifested by the prominent minimum of the MFPT for $\mu\!>\!\pi/2$ in figure \ref{figure6}. Lateral actin orientations ($\mu{=}\pi/2$) result in an equal chance for outward- and inward-directed motion of the cargo. Hence the behavior of the MFPT is similar to the one of uniformly random actin networks defined by $f^u_\text{M}$ and an inhomogeneous cytoskeletal structure with a thin actin cortex $\delta$ generally advances the search efficiency.
To the contrary, for outward-pointing actin filaments ($\mu<\pi/2$) the particle is predominantly moving in close proximity $\Delta$ of the cell's boundary, as sketched for the sample trajectory in figure \ref{figure6} (a). For that reason the cortex width $\delta$ does not significantly influence the MFPT as long as it is out of range of $\Delta$ and the homogeneous limit $\delta{=}1$ is most efficient, as evident from figure \ref{figure6}. This limit is even more efficient than a homogeneous microtubule network due to the outward-directed network structure, as found in figure \ref{figure2}. In general, a larger standard deviation $\sigma$ broadens the distribution of actin filaments and randomizes the search. Consequently, the behavior of the MFPT converges to the uniform case $f^u_\text{M}$ for increasing $\sigma$ and an inhomogeneous cytoskeleton generally improves the search of small membranous targets again.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figure7b-eps-converted-to.pdf}
\caption
{Influence of the Processivity on Inhomogeneous Search Strategies without Waiting Times.
Interior parameters: $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}\infty$; periphery parameters: $k^p_{m\rightarrow w}{=}10$, $k^p_{w\rightarrow m}{=}\infty$, $q_\text{M}{=}1$; target size: $\alpha_\text{target}{=}0{.}1$. The crosses refer to intersection nodes. \textbf{a} Sample trajectories for $\delta{=}0{.}5$ and different values of the processivity $p$. \textbf{b} MFPT in dependence of the actin cortical width $\delta$ for diverse motor processivities $p$. An increase of the processive behavior at intersection nodes leads to a systematic decrease of the MFPT. Hence, purely processive motors are most efficient in the case of instantaneous directional changes. \textbf{c} The MFPT in dependence of the processivity $p$ exhibits a monotonic decrease for different values of the actin cortical width $\delta$. This emphasizes the benefit of highly processive motors.
}
\label{figure7}
\end{figure*}
\begin{figure*}[!htb]
\centering
\includegraphics[width=\linewidth]{figure8c-eps-converted-to.pdf}
\caption
{Influence of the Processivity on Inhomogeneous Search Strategies with Waiting Times.
Interior parameters: $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}k_{w\rightarrow m}$; periphery parameters: $k^p_{m\rightarrow w}{=}10$, $k^p_{w\rightarrow m}{=}k_{w\rightarrow m}$, $q_\text{M}{=}1$; target size: $\alpha_\text{target}{=}0{.}1$. \textbf{a} MFPT in dependence of the actin cortical width $\delta$ for different motor processivities $p$ and $k_{w\rightarrow m}{=}1$. There arises a threshold at which a further increase of the processivity is disadvantageous. \textbf{b} The number of waiting periods decreases monotonically with the processivity $p$ for diverse fixed cortical widths $\delta$. Contrarily, the MWT diverges in the limit of $p\rightarrow 1$. \textbf{c} The MFPT in dependence of the processivity $p$ displays a minimum at $p\neq 1$ for various event rates $k_{w\rightarrow m}$ and different values of the actin cortical width $\delta$.
}
\label{figure8}
\end{figure*}
\subsubsection*{Influence of the Motor Processivity}
\noindent
Molecular motors do not necessarily change their direction of motion at each intersection node they encounter. A main feature of molecular motors is their processivity. Motor driven cargoes may also overcome the barrier opposed by network intersections and keep moving ballistically along the same track, which has been neglected so far ($p{=}0$). Here, we would like to investigate the impact of the motor processivity $p$ on the search efficiency to narrow target zones. For that purpose, we assume $k^i_{m\rightarrow w}{=}0$, $k^i_{w\rightarrow m}{=}k_{w\rightarrow m}$ in the interior and a mesh size defined by $k^p_{m\rightarrow w} {=} 10$ in the periphery. After each waiting period at intersection nodes, the particle may either start moving processively at rate $p \, k_{w\rightarrow m}$ or it changes its direction at rate $(1-p) \, k_{w\rightarrow m}$. For simplicity, we fix $q_\text{M}{=}1$ and $\alpha_\text{target}{=}0{.}1$.
\noindent
The difference in the movement pattern by a change in processivity is visualized by sample trajectories presented in figure \ref{figure7} (a). figure \ref{figure7} (b) displays the MFPT in dependence of the actin cortical width $\delta$ for various values of the processivity $p$. By neglecting waiting processes at intersection nodes via $k_{w\rightarrow m}{=}\infty$, we find that a higher processivity systematically improves the search efficiency. This benefit is emphasized in figure \ref{figure7} (c). The MFPT for fixed actin cortical widths decreases monotonically in dependence on the processivity $p$ and is most efficient for $p{=}1$.\\
\noindent
Remarkably, the introduction of waiting processes results in the development of an optimal processivity $p\neq 1$, as presented in figure \ref{figure8} (a) for $k_{w\rightarrow m}{=}1$. For a fixed cortical width $\delta$, figure \ref{figure8} (b) shows that the mean number of waiting periods decreases with $p$ due to the overall gain in search efficiency by an enhanced processivity $p$. However the MWT drastically increases, since the mean waiting time at confinement events $[(1-p)\,k_{w\rightarrow m}]^{-1}$ diverges in the limit of $p\rightarrow 1$. figure \ref{figure8} (c) visualizes, that the MFPT is dominated by the MWT for small rates $k_{w\rightarrow m}$. Consequently, the MFPT exhibits a minimum for $p\neq 1$ in contrast to the MMT, which is minimal for $p{=}1$ (compare to figure \ref{figure7} (c)). Due to inevitable waiting processes, it may be more efficient to change directions with a specific probability $1{-}p$ rather than transport by completely processive motors ($p{=}1$). A specific processivity $p$ is also reported in biological systems \cite{Balint2013,Ali2007,Ross2008,Ross2008B}.
\FloatBarrier
\begin{figure*}[!htbp]
\includegraphics[width=0.88\linewidth]{figure10d-eps-converted-to.pdf}
\caption
{Comparison to Biological Dimensions. Interior Parameters: $k^i_{m\rightarrow w}{=}0/\text{s}$, $k^i_{w\rightarrow m}{=}k_{w\rightarrow m}$; periphery parameters: $k^p_{m\rightarrow w}{=}10/\text{s}$, $k^p_{w\rightarrow m}{=}k_{w\rightarrow m}$, $p{=}0{.}5$, $q_\text{M}{=}1$; velocity: $v{=}1\,\mu\text{m}/\text{s}$; search domain: $R_{m.}{=}5\,\mu\text{m}$; target size: $\alpha_\text{target}\in \{0{.}02,0{.}1,0{.}2\}$. The MMT (MFPT for $k_{w\rightarrow m}{=}\infty$) as well as the mean number of waiting periods display distinct minima at small cortical widths in dependence of the target size $\alpha_\text{target}$ for the 2D case (\textbf{a}) as well as the generalization to 3D (\textbf{b}). The optimal width of the actin cortex $\delta_\text{opt}$ varies according to $\alpha_\text{target}$ and the mean waiting time per arrest state $1/k_{w\rightarrow m}$ from approximately $\delta_\text{opt}{=}0{.}1$ $\mu$m to $\delta_\text{opt}{=}1{.}2$ $\mu$m in 2D and $\delta_\text{opt}{=}1$ $\mu$m in 3D, which is in qualitative agreement to biological measurements.
}
\label{figure9}
\end{figure*}
\vspace*{5cm}
\pagebreak
\vspace*{5cm}
\section*{Discussion}
\noindent
With the aid of a random velocity model with intermittent arrest states, we studied the first passage properties of intracellular narrow escape problems. Via extensive computer simulations we are able to systematically analyze the influence of the cytoskeletal structure as well as the motor performance on the search efficiency to small target zones on the plasma membrane of a cell.\\
\noindent
For a spatially homogeneous cytoskeleton, the MFPT diverges in the limit $\alpha_\text{target}\rightarrow 0$ and directional changes at intersection nodes ($k_{m\rightarrow w}\neq 0$, $p{=}0$) do not improve the search efficiency. Moreover, we find that a random actin mesh ($q_\text{M}{=}1$) is preferable to radial microtubule networks ($q_\text{M}{=}0$) as well as combinations of both ($q_\text{M} \in (0;1)$), which underlines the benefit of motor activity regulation by outer stimuli. Homogeneous motion pattern are sufficient for large target sizes, but actually fail in the biologically relevant case of narrow escape problems.\\
\noindent
By varying the width of the actin cortical layer, we elaborate the impact of the cytoskeletal inhomogeneity on the search efficiency to narrow escapes. Remarkably, we find that a cell can optimize the detection time by regulation of the motor performance and convenient alterations of the spatial organization of the cytoskeleton. An inhomogeneous architecture with a thin actin cortex constitutes an efficient intracellular search strategy for narrow targets and generally leads to a considerable gain of efficiency in comparison to the homogeneous pendant. A confinement of the search to a thin shell below the plasma membrane, where the target area is located, saves time as it prevents extensive excursions to the cell interior - but the shell must not be too thin because otherwise the motion of the searcher gets localized and a time loss due to many stops occurs. Consequently, the MFPT diverges in the limit of $\delta \rightarrow 0$, which outlines that the actin cortex can act as a transport barrier or functional gateway \cite{Papadopulos2013}.\\
\noindent
Molecular motors do not necessarily change their direction of motion at each filamentous intersection, they may also overcome the constriction and remain on the same track, a property referred to as processivity. We find that an increased motor processivity systematically improves the search efficiency in the case of instantaneous directional changes on networks of a mesh size defined by $k^p_{m\rightarrow w}$. Due to waiting processes an optimal value of the motor processivity $p\neq 1$ emerges, which minimizes the detection of membranous targets. Specific probabilities to overcome constricting filament crossings are also reported during intracellular transport \cite{Balint2013,Ali2007,Ross2008,Ross2008B}.\\
\noindent
So far we have investigated a model for intracellular search strategies in two-dimensional cells. However, a generalization of our approach to three-dimensional spherical cells is straightforward and the former principles stay valid. If we assume a cell radius of $5$ $\mu$m, which is consistent with the typical size of a CTL, the opening angle of $\alpha_\text{target}{=}0{.}02$ ($\alpha_\text{target}{=}0{.}2$) leads to an arc length of $0{.}02\times 5$ $\mu$m ${=}100$ nm ($1$ $\mu$m). For instance the diameter of an immunological synapse is of the order of microns \cite{Grakoui1999,Qu2011,Angus2013,Ritter2013}. We further assume that motors move processively at intersections with probability $p{=}0{.}5$ and their velocity typically is about $1$ $\mu$m$/$s. A transition rate to the waiting state of $10/$s thus leads to a mesh size of $100$ nm, which is biologically reasonable \cite{Eghiaian2015,Salbreux2012}. Under these conditions, figure \ref{figure9} reveals an optimal width of the actin cortex which varies from approximately $\delta_\text{opt}=0{.}1$ $\mu$m to $\delta_\text{opt}=1{.}2$ $\mu$m in 2D and $\delta_\text{opt}=1$ $\mu$m in 3D. This is in good qualitative agreement to biological data \cite{Eghiaian2015,Salbreux2012,Clark2013}.\\
\noindent
In summary, our model indicates that the spatial organization of the cytoskeleton of spherical cells with a centrosome minimizes the characteristic time necessary to detect small targets on the cell membrane by random intermittent search processes along the cytoskeletal filaments (see also \cite{Schwarz2016,Schwarz2016b} for similar findings in a model with intermittent diffusive search). The minimization is achieved by a small width of the actin cortical layer and by regulation of the motor activity and behavior at network intersections. Remarkably a thin actin cortex is also more economic than distributing cytoskeletal filaments in all directions over the cell body. Thus, our work outlines that a thin confinement of the actin cortex, besides its advantages concerning cell stability or motility, also serves as an efficient key to intracellular narrow escape problems.
\section*{Acknowledgements}
\noindent
This work was funded by the German Research Foundation (DFG) within the Collaborative Research Center SFB 1027.
|
\section{Introduction\label{S1}}
Particle creation from the vacuum by strong external electromagnetic fields
was studied already for a long time; see, for example, Refs. ~\cite%
{Sch51,Nikis70a,Nikis70b,Gitman77,Gitman77b,GMR85,FGS,BirDav82,Grib,GelTan15,
GelTan15b,GavGT06,GavGit15}. To
be observable, the effect needs very strong electric fields in magnitudes
compared with the Schwinger critical field. Nevertheless, recent progress in
laser physics allows one to hope that an experimental observation of the
effect can be possible in the near future, see Refs. \cite{Dun09,Dun09b,Dun09c,Dun09d,Dun09e} for the
review. Electron-hole pair creation from the vacuum becomes also an
observable effect in graphene and similar nanostructures in laboratory; see,
e.g., \cite{dassarma,dassarmab}. The particle creation from the vacuum by external
electric and gravitational backgrounds plays also an important role in
cosmology and astrophysics \cite{BirDav82,Grib,GelTan15,GelTan15b,AndMot14}.
It should be noted that the particle creation from the vacuum by external
fields is a nonperturbative effect and its calculation essentially depends
on the structure of the external fields. Sometimes calculations can be done
in the framework of the relativistic quantum mechanics, sometimes using
semiclassical and numerical methods (see Refs. \cite{BirDav82,Grib,GelTan15,GelTan15b} for
the review). In most interesting cases, when the semiclassical approximation
is not applicable,{\Large \ }the most convincing consideration of the effect
is formulated in the framework of quantum field theory, in particular, in
the framework of QED, see Ref. \cite{Gitman77,Gitman77b,FGS,GavGT06,GavGit15} and is
based on the existence of exact solutions of the Dirac equation with the
corresponding external field. Until now,\emph{\ }only few exactly solvable
cases are known for either time-dependent homogeneous or constant
inhomogeneous electric fields. One of them is related to the constant
uniform electric field \cite{Sch51}, another one to\ the\ so-called
adiabatic electric field $E\left( t\right) =E\cosh ^{-2}\left( t/T_{\mathrm{S%
}}\right) \,$\cite{NarNik70} (see also \cite{DunHal98,GavGit96}), the case
related to the so-called $T$-constant electric field \cite%
{BagGitShv75,GavGit96,GavGit08,GavGit08b}, the case related to a periodic alternating
electric field \cite{NarNIk74,NarNIk74b}, and several constant inhomogeneous electric
fields of the similar forms where the time $t$ is replaced by the spatial
coordinate $x$. The existence of exactly solvable cases of
particle creation is extremely important both for deep understanding of
quantum field theory in general and for studying quantum vacuum effects in
the corresponding external fields. In our recent work \cite{AdoGavGit14}, we
have presented a new exactly solvable case of particle creation in an
exponentially decreasing in time electric field.
In the present article, we consider for the first time particle creation in
the so-called peak electric field, which is a combination of two exponential
parts, one exponentially increasing and the other
exponentially decreasing. This is another new exactly solvable case. We
demonstrate that in the field under consideration, one can find exact
solutions with appropriate asymptotic conditions and perform nonperturbative
calculations of all the characteristics of particle creation process. In
some respects, the peak electric field shares similar features with the
Sauter-like electric field, while in other respects it{\Huge \ }can be
treated as a pulse created by laser beams. Switching the peak field on and
off, we can imitate electric fields that are specific to condensed matter
physics, in particular to graphene or Weyl semimetals as was reported, e.g.,
in Refs. \cite%
{lewkowicz10,lewkowicz10b,lewkowicz10c,Vandecasteele10,GavGitY12,Zub12,KliMos13,Fil+McL15,VajDorMoe15}.
In our calculations, we use the general theory of Ref. \cite{Gitman77,Gitman77b,FGS}
and follow in the main the consideration of particle creation effect in a
homogeneous electric field \cite{GavGit96}. To this end we find complete
sets of exact solutions of the Dirac and Klein-Gordon equations in the peak
electric field and use them to calculate differential mean numbers of
created particle, total number of created particles, and the probability for
a vacuum to remain a vacuum.\textrm{\ }Characteristic asymptotic regimes
(slowly varying peak field, short pulse field, and the most asymmetric case
related to exponentially decaying field) are discussed in detail and a
comparison with the pure asymptotically decaying field is considered.
\section{Peak electric field\label{S2}}
\subsection{General}
In this section we introduce the so-called peak electric field, that is a
time-dependent electric field directed along an unique direction\footnote{%
Greek indices refer to the Minkowski spacetime $\mu =0,...,D$ while Latin
indices refer to Euclidean space $i=1,...,D$. Here $d=D+1$ is the dimension
of the spacetime. Bold letters represent Euclidean vectors such as $\mathbf{%
r}=x^{1},x^{2},...,x^{D}$. The Minkowski metric tensor is diagonal $\eta
_{\mu \nu }=\mathrm{diag}\underset{d}{\underbrace{\left( +1,-1,...,-1\right)
}}$.}%
\begin{equation}
\mathbf{E}\left( t\right) =\left( E^{i}\left( t\right) =\delta
_{1}^{i}E\left( t\right) \,,\ \ i=1,...,D\right) \,, \label{s3.1}
\end{equation}%
switched on at $t=-\infty $, and\ off at $t=+\infty ,$ its maximum $E>0$
occurring at a very sharp time instant, say at $t=0$, such that the limit%
\begin{equation}
\lim_{t\rightarrow -0}\dot{E}\left( t\right) \neq \lim_{t\rightarrow +0}\dot{%
E}\left( t\right) \,, \label{s3.3}
\end{equation}%
is not defined. The latter property implies that a peak at $t=0$ is present.
Time-dependent electric fields of this form can, as usual in QED with
unstable vacuum \cite{GMR85,FGS,BirDav82,Grib} (see also \cite{AdoGavGit15}), be
described by $t$-electric potential steps,%
\begin{eqnarray}
&&A^{0}=0\,,\ \ \mathbf{A}\left( t\right) =\left( A^{i}\left( t\right)
=\delta _{1}^{i}A_{x}\left( t\right) \right) \,, \notag \\
&&\dot{A}_{x}\left( t\right) =\frac{dA_{x}\left( t\right) }{dt}\leq
0\rightarrow \left\{
\begin{array}{l}
E\left( t\right) =-\dot{A}_{x}\left( t\right) \geq 0 \\
A_{x}\left( -\infty \right) >A_{x}\left( +\infty \right)%
\end{array}%
\right. \,, \label{s3.2}
\end{eqnarray}%
where $A_{x}\left( -\infty \right) $, $A_{x}\left( +\infty \right) $ are
constants (for further discussion and details concerning the definition of $%
t $-electric potential steps; see Ref. \cite{AdoGavGit15}).
To study the peak electric field we consider an electric field that is
composed of independent parts, wherein for each one the Dirac equation is
exactly solvable. The field in consideration grows exponentially from the
infinitely remote past $t=-\infty $, reaches a maximal amplitude $E$ at $t=0$
and decreases exponentially to the infinitely remote future $t=+\infty $. We
label the exponentially increasing interval by $\mathrm{I}=\left( -\infty ,0%
\right] $ and the exponentially decreasing interval by $\mathrm{II}=\left(
0,+\infty \right) $, where the field and its $t$-electric potential step are%
\begin{equation}
E\left( t\right) =E\left\{
\begin{array}{l}
e^{k_{1}t}\,,\ \ t\in \mathrm{I}\,, \\
e^{-k_{2}t}\,,\ \ t\in \mathrm{II}%
\end{array}%
\right. \,,\ \ A_{x}\left( t\right) =E\left\{
\begin{array}{l}
k_{1}^{-1}\left( -e^{k_{1}t}+1\right) ,\ \ t\in \mathrm{I}\,, \\
k_{2}^{-1}\left( e^{-k_{2}t}-1\right) \,,\ \ t\in \mathrm{II}%
\end{array}%
\right. \,. \label{ns4.0}
\end{equation}%
Here $E,k_{1},k_{2}$ are positive constants. The field and its potential are
depicted below in Fig. \ref{Fig1}.
\begin{figure}[th]
\includegraphics[scale=0.42]{Peak-electric-potential.eps}
\caption{The peak electric field $E\left( t\right) $ and its vector
potential $A_{x}\left( t\right) $ (\protect\ref{ns4.0}). Each interval is
characterized by a distinct exponential constant which explains the
non-symmetrical form of the picture. Here $k_{1}>k_{2}$ has been chosen.}
\label{Fig1}
\end{figure}
\subsection{Dirac equation with peak electric field}
To describe the problem in the framework of QED with $t$-electric potential
steps it is necessary to solve the Dirac equation for each interval
discussed above. In any of them, the Dirac equation in a $d=D+1$ dimensional
Minkowski spacetime is, in its Hamiltonian form, represented by%
\begin{eqnarray}
&&i\partial _{t}\psi \left( x\right) =H\left( t\right) \psi \left( x\right)
\,,\ \ H\left( t\right) =\gamma ^{0}\left( \boldsymbol{\gamma }\mathbf{P}%
+m\right) \,, \notag \\
&&\,P_{x}=-i\partial _{x}-U\left( t\right) ,\ \ \mathbf{P}_{\bot }=-i%
\boldsymbol{\nabla }_{\perp },\ \ U\left( t\right) =-eA_{x}\left( t\right)
\,, \label{2.9}
\end{eqnarray}%
where the index $\perp $ stands for spacial components perpendicular to the
electric field, $\mathbf{x}_{\perp }=\left\{ x^{2},...,x^{D}\right\} $ and $%
\mathbf{P}_{\bot }=\left( P^{2},\ldots ,P^{D}\right) $. Here $\psi (x)$ is a
$2^{[d/2]}$-component spinor ($[d/2]$ stands for the integer part of the
ratio $d/2$), $m\neq 0$ is the electron mass, $\gamma ^{\mu }$ are the $%
\gamma ${\Huge \ }matrices in $d$ dimensions , $U\left( t\right) $ is the
potential energy of one electron, and the relativistic system of units is
used throughout in this paper ($\hslash =c=1$).
As customary for $t$-electric steps \cite{AdoGavGit15}, solutions of the
Dirac equation (\ref{2.9}) have the form%
\begin{eqnarray}
&&\psi _{n}\left( x\right) =\exp \left( i\mathbf{pr}\right) \psi _{n}\left(
t\right) ,\;\ n=(\mathbf{p},\sigma )\,, \notag \\
&&\psi _{n}\left( t\right) =\left\{ \gamma ^{0}i\partial _{t}-\gamma ^{1}
\left[ p_{x}-U\left( t\right) \right] -\boldsymbol{\gamma }\mathbf{p}_{\bot
}+m\right\} \phi _{n}(t)\,, \label{2.10}
\end{eqnarray}%
where $\psi _{n}\left( t\right) $ and $\phi _{n}(t)$ are spinors which
depend on $t$ alone. In fact, these are states with definite momenta $%
\mathbf{p}$. Substituting Eq. (\ref{2.10}) into Dirac equation (\ref{2.9}),
we obtain a second-order differential equation for the spinor $\phi _{n}(t)$,%
\begin{equation}
\left\{ \frac{d^{2}}{dt^{2}}+\left[ p_{x}-U\left( t\right) \right] ^{2}+\pi
_{\perp }^{2}-i\gamma ^{0}\gamma ^{1}\dot{U}\left( t\right) \right\} \phi
_{n}\left( t\right) =0\,,\;\pi _{\perp }=\sqrt{\mathbf{p}_{\perp }^{2}+m^{2}}%
\,. \label{2.11}
\end{equation}%
We separate the spinning variables by the substitution%
\begin{equation}
\phi _{n}(t)=\varphi _{n}\left( t\right) v_{\chi ,\sigma }\,, \label{2.12}
\end{equation}%
where $v_{\chi ,\sigma }$ for $\chi =\pm 1$ and $\sigma =(\sigma _{1},\sigma
_{2},\dots ,\sigma _{\lbrack d/2]-1})$ for $\sigma _{s}=\pm 1$ is a set of
constant orthonormalized spinors, satisfying the following equations:%
\begin{equation}
\gamma ^{0}\gamma ^{1}v_{\chi ,\sigma }=\chi v_{\chi ,\sigma }\,,\ \ v_{\chi
,\sigma }^{\dag }v_{\chi ^{\prime },\sigma ^{\prime }}=\delta _{\chi ,\chi
^{\prime }}\delta _{\sigma ,\sigma ^{\prime }\,}. \label{e2a}
\end{equation}%
The quantum numbers $\chi $ and $\sigma _{s}$ describe spin polarization and
provide a convenient parametrization of the solutions. Since in ($1+1$) and (%
$2+1$) dimensions ($d=2,3$) there are no spinning degrees of freedom, the
quantum numbers $\sigma $ are absent. In addition, in $d>3$, Eq. (\ref{2.11}) allows one to subject the constant spinors $v_{\chi ,\sigma }$ to some
supplementary conditions that, for example, can be chosen in the form%
\begin{eqnarray}
&&i\gamma ^{2s}\gamma ^{2s+1}v_{\chi ,\sigma }=\sigma _{s}v_{\chi ,\sigma }\
\mathrm{for}\ \mathrm{even\ }d\,, \notag \\
&&i\gamma ^{2s+1}\gamma ^{2s+2}v_{\chi ,\sigma }=\sigma _{s}v_{\chi ,\sigma
}\ \mathrm{for\ odd}\ d\,, \label{e2.5}
\end{eqnarray}%
Then the scalar functions $\varphi _{n}\left( t\right) $ have to obey the
second-order differential equation%
\begin{equation}
\left\{ \frac{d^{2}}{dt^{2}}+\left[ p_{x}-U\left( t\right) \right] ^{2}+\pi
_{\perp }^{2}-i\chi \dot{U}\left( t\right) \right\} \varphi _{n}\left(
t\right) =0\,. \label{s2}
\end{equation}
In $d$ dimensions, for any given set of quantum numbers $\mathbf{p}$, there
exist only $J_{(d)}=2^{[d/2]-1}$ different spin states. The projection
operator inside the curly brackets in Eq. (\ref{2.10}) does not commute with
the matrix $\gamma ^{0}\gamma ^{1}$ and, consequently, transforms $\phi
_{n}^{\left( \chi \right) }(x)$ with a given $\chi $ to a linear
superposition of functions $\phi _{n}^{\left( +1\right) }(x)$ and $\phi
_{n^{\prime }}^{\left( -1\right) }(x)$ with indices $n$ and $n^{\prime }$
corresponding to the same $\mathbf{p}$. For this reason, solutions of (\ref%
{2.10}) differing only by values of $\chi $ and $\sigma _{s}$ are linearly
dependent.\textrm{\ }That is why it is enough to select a particular value
of $\chi $ to perform some specific calculations, whose choice shall be
explicitly indicated when necessary.
Exact solutions of the Dirac equation with the exponentially decreasing
electric field have been obtained by us previously in \cite{AdoGavGit14}.
Thus, using some results of the latter work, below we summarize the
structure of solutions for each interval\textbf{\ }and unify it in a single
presentation. To this aim we introduce new variables $\eta _{j}$,%
\begin{eqnarray}
&&\eta _{1}\left( t\right) =ih_{1}e^{k_{1}t}\,,\ \ \eta _{2}\left( t\right)
=ih_{2}e^{-k_{2}t}\,, \notag \\
&&h_{j}=\frac{2eE}{k_{j}^{2}},\ \ j=1,2\,, \label{i.0}
\end{eqnarray}%
in place of $t$ and represent the scalar functions $\varphi _{n}\left(
t\right) $ as%
\begin{eqnarray}
&&\varphi _{n}^{j}\left( t\right) =e^{-\eta _{j}/2}\eta _{j}^{\nu _{j}}%
\tilde{\varphi}^{j}\left( \eta _{j}\right) , \notag \\
&&\nu _{j}=\frac{i\omega _{j}}{k_{j}}\,,\ \ \omega _{j}=\sqrt{\pi
_{j}^{2}+\pi _{\perp }^{2}}\,,\ \ \pi _{j}=p_{x}-\left( -1\right) ^{j}\frac{%
eE}{k_{j}}\,, \label{i.2}
\end{eqnarray}%
where the subscript $j$ distinguishes quantities associated to the intervals
$\mathrm{I}$ ($j=1$) and $\mathrm{II}$ ($j=2$), respectively. Then the
functions $\tilde{\varphi}^{j}\left( \eta _{j}\right) $ satisfy the
confluent hypergeometric equation \cite{BatE53},%
\begin{equation*}
\left[ \eta _{j}\frac{d^{2}}{d\eta _{j}^{2}}+\left( c_{j}-\eta _{j}\right)
\frac{d}{d\eta _{j}}-a_{j}\right] \tilde{\varphi}^{j}\left( \eta _{j}\right)
=0\,,
\end{equation*}%
whose parameters are%
\begin{equation}
c_{j}=1+2\nu _{j}\,,\ \ a_{j}=\frac{1}{2}\left( 1+\chi \right) +\left(
-1\right) ^{j}\frac{i\pi _{j}}{k_{j}}+\nu _{j}\,. \label{i.3}
\end{equation}%
A fundamental set of solutions for the equation is composed by two linearly
independent confluent hypergeometric functions:%
\begin{equation*}
\Phi \left( a_{j},c_{j};\eta _{j}\right) \ \ \mathrm{and}\ \ \eta
_{j}^{1-c_{j}}e^{\eta _{j}}\Phi \left( 1-a_{j},2-c_{j};-\eta _{j}\right) \,,
\end{equation*}%
where%
\begin{equation}
\Phi \left( a,c;\eta \right) =1+\frac{a}{c}\frac{\eta }{1!}+\frac{a\left(
a+1\right) }{c\left( c+1\right) }\frac{\eta ^{2}}{2!}+\ldots \,. \label{chf}
\end{equation}%
Thus the general solution of Eq.~(\ref{s2}) in the intervals $\mathrm{I}$
and $\mathrm{II}$ can be expressed as the following linear superposition:%
\begin{align}
& \varphi _{n}^{j}\left( t\right) =b_{2}^{j}y_{1}^{j}\left( \eta _{j}\right)
+b_{1}^{j}y_{2}^{j}\left( \eta _{j}\right) \,, \notag \\
& y_{1}^{j}\left( \eta _{j}\right) =e^{-\eta _{j}/2}\eta _{j}^{\nu _{j}}\Phi
\left( a_{j},c_{j};\eta _{j}\right) \,, \notag \\
& y_{2}^{j}\left( \eta _{j}\right) =e^{\eta _{j}/2}\eta _{j}^{-\nu _{j}}\Phi
\left( 1-a_{j},2-c_{j};-\eta _{j}\right) \,, \label{i.3.3}
\end{align}%
with constants $b_{1}^{j}$ and $b_{2}^{j}$ being fixed by the initial
conditions. The Wronskian of the $y$ functions is%
\begin{equation}
y_{1}^{j}\left( \eta _{j}\right) \frac{d}{d\eta _{j}}y_{2}^{j}\left( \eta
_{j}\right) -y_{2}^{j}\left( \eta _{j}\right) \frac{d}{d\eta _{j}}%
y_{1}^{j}\left( \eta _{j}\right) =\frac{1-c_{j}}{\eta _{j}}\,. \label{i.3.4}
\end{equation}
It is worth noting that the complete set of solutions for the Klein-Gordon
equation,%
\begin{equation}
\phi _{n}\left( x\right) =\exp \left( i\mathbf{pr}\right) \varphi _{n}\left(
t\right) \,. \label{KG1}
\end{equation}%
can be obtained from the solutions above by setting $\chi =0$ in all
formulas. In this case $n=\mathbf{p}$.
With the help of the exact solutions one may write Dirac spinors throughout
the time interval $t\in \left( -\infty ,+\infty \right) $. As can be seen
from (\ref{ns4.0}), the peak electric field is switched on at the infinitely
remote past $t\rightarrow -\infty $ and switched{\Huge \ }off at the
infinitely remote future $t\rightarrow +\infty $. At these regions, the
exact solutions represent free particles,%
\begin{equation}
\ _{\zeta }\varphi _{n}\left( t\right) =\ _{\zeta }\mathcal{N}e^{-i\zeta
\omega _{1}t}\,\mathrm{\ \ if\ }\ t\rightarrow -\infty ,\ \ ^{\zeta }\varphi
_{n}\left( t\right) =\ ^{\zeta }\mathcal{N}e^{-i\zeta \omega _{2}t}\mathrm{\
\ if}\ \ t\rightarrow +\infty \,, \label{i.4.0}
\end{equation}%
respectively, where $\omega _{1}$ denotes energy of initial particles at $%
t\rightarrow -\infty $, $\omega _{2}$ denotes energy of final particles at $%
t\rightarrow +\infty $ and $\zeta $ labels electron $\left( \zeta =+\right) $
and positron $\left( \zeta =-\right) $ states. Here$\;_{\zeta }\mathcal{N}$
and $\;^{\zeta }\mathcal{N}$ are normalization constants with respect to the
inner product\footnote{%
For a detailed explanation concerning the inner product for $t$-electric
potential steps see, e. g., Ref. \cite{AdoGavGit15}.}%
\begin{equation}
\left( \psi ,\psi ^{\prime }\right) =\int \psi ^{\dagger }\left( x\right)
\psi ^{\prime }\left( x\right) d\mathbf{r}\,,\ \ d\mathbf{r}%
=dx^{1}...dx^{D}\,, \label{t4.0}
\end{equation}%
These constants are%
\begin{eqnarray}
&&\ _{\zeta }\mathcal{N}=\ _{\zeta }CY\,,\ \ \ ^{\zeta }\mathcal{N}=\
^{\zeta }CY\,,\ \ Y=V_{\left( d-1\right) }^{-1/2}\,, \notag \\
&&_{\zeta }C=\left( 2\omega _{1}q_{1}^{\zeta }\right) ^{-1/2}\,,\ \ \
^{\zeta }C=\left( 2\omega _{2}q_{2}^{\zeta }\right) ^{-1/2}\,, \notag \\
&&q_{j}^{\zeta }=\omega _{j}-\chi \zeta \pi _{j}\,, \label{i.4.2}
\end{eqnarray}%
where $V_{\left( d-1\right) }$ is the spatial volume. By virtue of these
properties, electron (positron) states can be selected as follows:%
\begin{eqnarray}
\ _{+}\varphi _{n}\left( t\right) &=&\;_{+}\mathcal{N}\exp \left( i\pi \nu
_{1}/2\right) y_{2}^{1}\left( \eta _{1}\right) \,,\,\ _{-}\varphi _{n}\left(
t\right) =\;_{-}\mathcal{N}\exp \left( -i\pi \nu _{1}/2\right)
y_{1}^{1}\left( \eta _{1}\right) \,,\ \ t\in \mathrm{I}\,; \notag \\
\ ^{+}\varphi _{n}\left( t\right) &=&\;^{+}\mathcal{N}\exp \left( -i\pi \nu
_{2}/2\right) y_{1}^{2}\left( \eta _{2}\right) \,,\,\ ^{-}\varphi _{n}\left(
t\right) =\;^{-}\mathcal{N}\exp \left( i\pi \nu _{2}/2\right)
y_{2}^{2}\left( \eta _{2}\right) \,,\ \ t\in \mathrm{II}\,. \label{i.4.1}
\end{eqnarray}
\subsection{$g$ Coefficients and mean numbers of created particles\label%
{Ss2.3}}
Taking into account the complete set of exact solutions (\ref{i.3.3}), the
functions$\ \ _{-}\varphi _{n}\left( t\right) $ and $\ ^{+}\varphi
_{n}\left( t\right) $ can be presented in the form%
\begin{eqnarray}
\ \ ^{+}\varphi _{n}\left( t\right) &=&\left\{
\begin{array}{l}
g\left( _{+}|^{+}\right) \ _{+}\varphi _{n}\left( t\right) +\kappa g\left(
_{-}|^{+}\right) \ _{-}\varphi _{n}\left( t\right) \,,\ \ t\in \mathrm{I} \\
\;^{+}\mathcal{N}\exp \left( -i\pi \nu _{2}/2\right) y_{1}^{2}\left( \eta
_{2}\right) \,,\ \ \ \ \ \ \ \ \ \ \,t\in \mathrm{II}%
\end{array}%
\right. \,, \label{i.6.1} \\
\ _{-}\varphi _{n}\left( t\right) &=&\left\{
\begin{array}{l}
\;_{-}\mathcal{N}\exp \left( -i\pi \nu _{1}/2\right) y_{1}^{1}\left( \eta
_{1}\right) \,,\ \ \ \ \ \ \ \ \ \ \ \ \,t\in \mathrm{I} \\
g\left( ^{+}|_{-}\right) \ ^{+}\varphi _{n}\left( t\right) +\kappa g\left(
^{-}|_{-}\right) \ ^{-}\varphi _{n}\left( t\right) \,,\ \ t\in \mathrm{II}%
\end{array}%
\right. \,, \label{i.6.2}
\end{eqnarray}%
for the whole axis $t$, where the coefficients $g$\ are the diagonal matrix
elements,%
\begin{equation}
\left( \ _{\zeta ^{\prime }}\psi _{l},\ ^{\zeta }\psi _{n}\right) =\delta
_{l,n}g\left( _{\zeta ^{\prime }}|^{\zeta }\right) ,\ \ g\left( ^{\zeta
^{\prime }}|_{\zeta }\right) =g\left( _{\zeta ^{\prime }}|^{\zeta }\right)
^{\ast }. \label{t4.5}
\end{equation}%
These coefficients satisfy the unitary relations%
\begin{equation}
\sum_{\varkappa }g\left( ^{\zeta }|_{\varkappa }\right) g\left( _{\varkappa
}|^{\zeta ^{\prime }}\right) =\sum_{\varkappa }g\left( _{\zeta }|^{\varkappa
}\right) g\left( ^{\varkappa }|_{\zeta ^{\prime }}\right) =\delta _{\zeta
,\zeta ^{\prime }}\,. \label{3.16.1}
\end{equation}%
Here the constant $\kappa $ ($\kappa =+1$ above) allows us to cover the
Klein-Gordon case, whose details are discussed in Eqs. (\ref{KGC}) and (\ref%
{gpKG}) below.
The functions$\ _{-}\varphi _{n}\left( t\right) $ and $\ ^{+}\varphi
_{n}\left( t\right) $ and their derivatives satisfy the following continuity
conditions:
\begin{equation}
\left. \ _{-}^{+}\varphi _{n}(t)\right\vert _{t=-0}=\left. \ _{-}^{+}\varphi
_{n}(t)\right\vert _{t=+0}\,,\ \ \left. \partial _{t}\ _{-}^{+}\varphi
_{n}(t)\right\vert _{t=-0}=\left. \partial _{t}\ _{-}^{+}\varphi
_{n}(t)\right\vert _{t=+0}\,. \label{i.7}
\end{equation}%
Using Eq.~(\ref{i.7}) and the Wronskian (\ref{i.3.4}), one can find each
coefficient $g\left( _{\zeta }|^{\zeta ^{\prime }}\right) $ and $g\left(
^{\zeta }|_{\zeta ^{\prime }}\right) $ in Eqs.~(\ref{i.6.1}) and (\ref{i.6.2}%
). For example, applying these conditions to the set (\ref{i.6.1}), the
coefficient $g\left( _{-}|^{+}\right) $ takes the form%
\begin{eqnarray}
&&g\left( _{-}|^{+}\right) =\mathcal{C}\Delta \,,\ \ \mathcal{C}=-\frac{1}{2}%
\sqrt{\frac{q_{1}^{-}}{\omega _{1}q_{2}^{+}\omega _{2}}}\exp \left[ \frac{%
i\pi }{2}\left( \nu _{1}-\nu _{2}\right) \right] \,, \notag \\
&&\Delta =\left. \left[ k_{1}h_{1}y_{1}^{2}\left( \eta _{2}\right) \frac{d}{%
d\eta _{1}}y_{2}^{1}\left( \eta _{1}\right) +k_{2}h_{2}y_{2}^{1}\left( \eta
_{1}\right) \frac{d}{d\eta _{2}}y_{1}^{2}\left( \eta _{2}\right) \right]
\right\vert _{t=0}\,. \label{gp}
\end{eqnarray}%
Alternatively, we obtain from the set (\ref{i.6.1})%
\begin{eqnarray}
&&g\left( ^{+}|_{-}\right) =\mathcal{C}^{\prime }\Delta ^{\prime }\,,\ \
\mathcal{C}^{\prime }=-\frac{1}{2}\sqrt{\frac{q_{2}^{+}}{\omega
_{1}q_{1}^{-}\omega _{2}}}\exp \left[ \frac{i\pi }{2}\left( \nu _{2}-\nu
_{1}\right) \right] \,, \notag \\
&&\Delta ^{\prime }=\left\{ k_{2}h_{2}y_{1}^{1}\left( \eta _{1}\right) \frac{%
d}{d\eta _{2}}y_{2}^{2}\left( \eta _{2}\right) +k_{1}h_{1}y_{2}^{2}\left(
\eta _{2}\right) \frac{d}{d\eta _{1}}y_{1}^{1}\left( \eta _{1}\right)
\right\} _{t=0}\,. \label{nd5}
\end{eqnarray}%
Comparing Eqs.~(\ref{gp}) and (\ref{nd5}) one can easily verify that the
symmetry under a simultaneous change $k_{1}\leftrightarrows k_{2}$ and $\pi
_{1}\leftrightarrows -\pi _{2}$ holds,%
\begin{equation}
g\left( ^{+}|_{-}\right) \leftrightarrows g\left( _{-}|^{+}\right) \,.
\label{nd6}
\end{equation}
A formal transition to the Klein-Gordon case can be performed by setting $\chi =0$
and $\kappa =-1$\ in Eqs.~(\ref{i.6.1}) and (\ref{i.6.2}), and by replacing%
{\Huge \ }the normalization factors $_{\zeta }C,\ ^{\zeta }C$ written in (%
\ref{i.4.2}) by%
\begin{equation}
\ _{\zeta }C=\left( 2\omega _{1}\right) ^{-1/2}\,,\ \ \ ^{\zeta }C=\left(
2\omega _{2}\right) ^{-1/2}\,. \label{KGC}
\end{equation}%
After these substitutions, the coefficient $g\left( _{-}|^{+}\right) $ for
scalar particles reads%
\begin{equation}
g\left( _{-}|^{+}\right) =\mathcal{C}_{\mathrm{sc}}\left. \Delta \right\vert
_{\chi =0}\,,\ \ \mathcal{C}_{\mathrm{sc}}=\left( 4\omega _{1}\omega
_{2}\right) ^{-1/2}\exp \left[ i\pi \left( \nu _{1}-\nu _{2}\right) /2\right]
\,, \label{gpKG}
\end{equation}%
with $\Delta $ given by Eq.~(\ref{gp}). In this case it is worth noting that
the symmetry under the simultaneous change $k_{1}\leftrightarrows k_{2}$ and
$\pi _{1}\leftrightarrows -\pi _{2}$ holds as%
\begin{equation}
g\left( ^{+}|_{-}\right) \leftrightarrows -g\left( _{-}|^{+}\right) \,.
\label{nd6b}
\end{equation}%
Note that for scalar particles\ the coefficients $g$ satisfy the unitary
relations%
\begin{equation}
\sum_{\varkappa }\varkappa g\left( ^{\zeta }|_{\varkappa }\right) g\left(
_{\varkappa }|^{\zeta ^{\prime }}\right) =\sum_{\varkappa }\varkappa g\left(
_{\zeta }|^{\varkappa }\right) g\left( ^{\varkappa }|_{\zeta ^{\prime
}}\right) =\zeta \delta _{\zeta ,\zeta ^{\prime }}\,. \label{3.16.2}
\end{equation}
Using a unitary transformation $V$\ between the initial and final Fock spaces,
see \cite{FGS}, one finds that the differential mean number of
electron-positron pairs created from the vacuum can be expressed via the
coefficients $g$\ as%
\begin{equation}
N_{n}^{\mathrm{cr}}=\left\vert g\left( _{-}|^{+}\right) \right\vert ^{2}
\label{mN}
\end{equation}%
both for fermions and bosons. Then the total number of \ created pairs is
given by the sum%
\begin{equation}
N=\sum_{n}N_{n}^{\mathrm{cr}}=\sum_{n}\left\vert g\left( _{-}\left\vert
^{+}\right. \right) \right\vert ^{2} \label{TN}
\end{equation}%
and the vacuum-to-vacuum transition probability reads%
\begin{equation}
P_{v}=\exp \left\{ \kappa \sum_{n}\ln \left[ 1-\kappa N_{n}^{\mathrm{cr}}%
\right] \right\} \,. \label{vacprob}
\end{equation}
For Dirac particles, using $g\left( _{-}|^{+}\right) $ given by Eq.~(\ref{gp}%
), we find in the case under consideration%
\begin{equation}
N_{n}^{\mathrm{cr}}=\left\vert \mathcal{C}\Delta \right\vert ^{2}\,.
\label{4.0}
\end{equation}
For scalar particles, using $g\left( _{-}|^{+}\right) $ given by Eq.~(\ref%
{gpKG}) the same quantity has the form%
\begin{equation}
N_{n}^{\mathrm{cr}}=\left\vert \mathcal{C}_{\mathrm{sc}}\left. \Delta
\right\vert _{\chi =0}\right\vert ^{2}\,. \label{4b}
\end{equation}%
It is clear that $N_{n}^{\mathrm{cr}}$ is a function of modulus squared of
transversal momentum, $\mathbf{p}_{\perp }^{2}$. It follows from Eq.~(\ref%
{nd6}) and (\ref{nd6b}), respectively, that $N_{n}^{\mathrm{cr}}$ is
invariant under the simultaneous change $k_{1}\leftrightarrows k_{2}$ and $%
\pi _{1}\leftrightarrows -\pi _{2}$ for both fermions and bosons. Then if $%
k_{1}=k_{2}$, $N_{n}^{\mathrm{cr}}$ appears to be an even function of
longitudinal momentum $p_{x}$ too.
\section{Slowly varying field\label{S4}}
\subsection{Differential quantities\label{Ss4.1}}
We are primarily interested in a strong field, when $N_{n}^{\mathrm{cr}}$
are not necessarily small in some ranges of quantum numbers and
semiclassical calculations cannot{\Huge \ }be applied. The inverse
parameters $k_{1}^{-1}$, $k_{2}^{-1}$ represent scales of time duration for
increasing and decreasing phases of the electric field. In particular, we
have a slowly varying field at small values of both $k_{1},k_{2}\rightarrow
0 $. This case can be considered as a new two-parameter regularization for a
constant electric field [additional to the known one-parameter
regularizations by the Sauter-like electric field, $E\cosh ^{-2}\left( t/T_{%
\mathrm{S}}\right) $, and the $T$-constant electric field (an electric field
which effectively acts during a sufficiently large but finite time interval $%
T$)]. Let us consider only this case, supposing that $k_{1}$ and $k_{2}$ are
sufficiently small, obeying the conditions%
\begin{equation}
\min \left( h_{1},h_{2}\right) \gg \max \left( 1,m^{2}/eE\right) \,.
\label{4.1}
\end{equation}
Let us analyze how the numbers $N_{n}^{\mathrm{cr}}$ depend on the
parameters $p_{x}$ and $\pi _{\perp }$. It can be seen from semiclassical
analysis that $N_{n}^{\mathrm{cr}}$ is exponentially small in the range of
very large $\pi _{\perp }\gtrsim \min \left(
eEk_{1}^{-1},eEk_{2}^{-1}\right) $. Then the range of fixed $\pi _{\perp }$
is of interest, and in what follows, we assume that%
\begin{equation}
\sqrt{\lambda }<K_{\bot }\,,\ \ \lambda =\frac{\pi _{\perp }^{2}}{eE}\,,
\label{4.2}
\end{equation}%
where $K_{\bot }$ is any given number satisfying the condition%
\begin{equation}
\min \left( h_{1},h_{2}\right) \gg K_{\bot }^{2}\gg \max \left(
1,m^{2}/eE\right) \,. \label{4.3}
\end{equation}
By virtue of symmetry properties of $N_{n}^{\mathrm{cr}}$ discussed above,
one can only consider $p_{x}$ either{\Huge \ }positive or negative. Let us,
for example, consider the interval $-\infty <p_{x}\leq 0$. In this case $\pi
_{2}$ is negative and large, $-\pi _{2}\geq eE/k_{2}$, while $\pi _{1}$
varies from positive to negative values, $-\infty <\pi _{1}\leq eE/k_{1}$.
The case of large negative $\pi _{1}$, $-2\pi _{1}/k_{1}>K_{1}$, where $%
K_{1} $ is any given large number, $K_{1}\gg K_{\bot }$, is quite simple. In
this case, using the appropriate asymptotic expressions of the confluent
hypergeometric function one can see that $N_{n}^{\mathrm{cr}}$ is negligibly
small. To see this, Eq. (\ref{A10}) in Appendix \ref{Ap2} is useful in the
range $h_{1}\gtrsim -2\pi _{1}/k_{1}>K_{1}$ and the expression for large $%
c_{2}$ with fixed $a_{2}$ and $h_{2}$ and the expression for large $c_{1}$
with fixed $a_{1}-c_{1}$ and $h_{1}$, given in \cite{BatE53}, are useful in
the range $-2\pi _{1}/k_{1}\gg h_{1}$.
We expect a significant contribution in the range%
\begin{equation}
h_{1}\geq 2\pi _{1}/k_{1}>-K_{1}, \label{4.4}
\end{equation}%
that can be divided in four subranges%
\begin{eqnarray}
\mathrm{(a)} &&\;h_{1}\geq 2\pi _{1}/k_{1}>h_{1}\left[ 1-\left( \sqrt{h_{1}}%
g_{2}\right) ^{-1}\right] , \notag \\
\mathrm{(b)} &&\;h_{1}\left[ 1-\left( \sqrt{h_{1}}g_{2}\right) ^{-1}\right]
>2\pi _{1}/k_{1}>h_{1}\left( 1-\varepsilon \right) , \notag \\
\mathrm{(c)} &&\;h_{1}\left( 1-\varepsilon \right) >2\pi
_{1}/k_{1}>h_{1}/g_{1}, \notag \\
\mathrm{(d)} &&\;h_{1}/g_{1}>2\pi _{1}/k_{1}>-K_{1}, \label{4.5}
\end{eqnarray}%
where $g_{1}$, $g_{2}$, and $\varepsilon $ are any given numbers satisfying
the condition $g_{1}$ $\gg 1$, $g_{2}$ $\gg 1$, and $\varepsilon \ll 1$. $.$%
Note that $\tau _{1}=-ih_{1}/\left( 2-c_{1}\right) \approx \frac{h_{1}k_{1}}{%
2\left\vert \pi _{1}\right\vert }$ in the subranges (a), (b), and (c) and $%
\tau _{2}=ih_{2}/c_{2}\approx \frac{h_{2}k_{2}}{2\left\vert \pi
_{2}\right\vert }$ in the whole range (\ref{4.4}). In these subranges we
have for $\left\vert \tau _{2}\right\vert $%
\begin{eqnarray}
\mathrm{(a)} &&\;1\leq \tau _{2}^{-1}<\left[ 1+\left( \sqrt{h_{2}}%
g_{2}\right) ^{-1}\right] , \notag \\
\mathrm{(b)} &&\;\left[ 1+\left( \sqrt{h_{2}}g_{2}\right) ^{-1}\right] <\tau
_{2}^{-1}<\left( 1+\varepsilon k_{2}/k_{1}\right) , \notag \\
\mathrm{(c)} &&\;\left( 1+\varepsilon k_{2}/k_{1}\right) <\tau _{2}^{-1}<%
\left[ 1+k_{2}/k_{1}\left( 1-1/g_{1}\right) \right] , \notag \\
\mathrm{(d)} &&\;\left[ 1+k_{2}/k_{1}\left( 1-1/g_{1}\right) \right] <\tau
_{2}^{-1}\lesssim \left( 1+k_{2}/k_{1}\right) . \label{4.6}
\end{eqnarray}%
We see that $\tau _{1}-1\rightarrow 0$ and $\tau _{2}-1\rightarrow 0$ in the
range (a), while $\left\vert \tau _{1}-1\right\vert \sim 1$ in the range
(c), and $\left\vert \tau _{2}-1\right\vert \sim 1$ in the ranges (c) and
(d). In the range (b) these quantities vary from their values in the ranges
(a) and (c).
In the range (a) we can use the asymptotic expression of the confluent
hypergeometric function given by Eq.~(\ref{A1}) in Appendix \ref{Ap2}. Using
Eqs.~(\ref{A7}), (\ref{A8}), and (\ref{A9}) obtained in Appendix \ref{Ap2},
we finally find the leading term as%
\begin{equation}
N_{n}^{\mathrm{cr}}=e^{-\pi \lambda }\left[ 1+O\left( \left\vert \mathcal{Z}%
_{1}\right\vert \right) \right] , \label{4.7}
\end{equation}%
for fermions and bosons, where $\max \left\vert \mathcal{Z}_{1}\right\vert
\lesssim g_{2}^{-1}$ . In the range (c), we use the asymptotic expression of
the confluent hypergeometric function given by Eq.~(\ref{A10}) in Appendix %
\ref{Ap2}. Then we find that%
\begin{equation}
N_{n}^{\mathrm{cr}}=e^{-\pi \lambda }\left[ 1+O\left( \left\vert \mathcal{Z}%
_{1}\right\vert \right) ^{-1}+O\left( \left\vert \mathcal{Z}_{2}\right\vert
\right) ^{-1}\right] , \label{4.8}
\end{equation}%
where $\max \left\vert \mathcal{Z}_{1}\right\vert ^{-1}\lesssim \sqrt{%
g_{1}/h_{1}}$ and $\max \left\vert \mathcal{Z}_{2}\right\vert ^{-1}\lesssim
g_{2}^{-1}$. Using the asymptotic expression Eq.~(\ref{A1}) and taking into
account Eq.~(\ref{4.7}) and (\ref{4.8}), we can estimate that $N_{n}^{%
\mathrm{cr}}\sim e^{-\pi \lambda }$ in the range (b). In the range (d), the
confluent hypergeometric function $\Phi \left( a_{2},c_{2};ih_{2}\right) $
is approximated by Eq.~(\ref{A10a}) and the function $\Phi \left(
1-a_{1},2-c_{1};-ih_{1}\right) $ is approximated by Eq.~(\ref{A11}) given in
Appendix \ref{Ap2}. In this range the differential mean numbers in the
leading-order approximation are%
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \frac{\exp \left[ -\frac{\pi }{k_{1}}\left(
\omega _{1}-\pi _{1}\right) \right] }{\sinh \left( 2\pi \omega
_{1}/k_{1}\right) }\times \left\{
\begin{array}{l}
\sinh \left[ \pi \left( \omega _{1}+\pi _{1}\right) /k_{1}\right] \ \mathrm{%
for\ fermions} \\
\cosh \left[ \pi \left( \omega _{1}+\pi _{1}\right) /k_{1}\right] \ \mathrm{%
for\ bosons}%
\end{array}%
\right. \,. \label{4.9a}
\end{equation}%
It is clear that $N_{n}^{\mathrm{cr}}$ given by Eqs.~(\ref{4.9a}) tends to
Eq.~(\ref{4.8}), $N_{n}^{\mathrm{cr}}\rightarrow e^{-\pi \lambda }$, when $%
\pi _{1}\gg \pi _{\bot }$. Consequently, the forms (\ref{4.9a}) are valid in
the whole range (\ref{4.4}). Assuming $m/k_{1}\gg 1$, we see that \ values
of $N_{n}^{\mathrm{cr}}$ given by Eqs.~(\ref{4.9a}) are negligible in the
range $\pi _{1}\lesssim \pi _{\bot }$. Then we find for bosons and fermions
that significant value of $N_{n}^{\mathrm{cr}}$ is in the range $\pi _{\bot
}<\pi _{1}\leqslant eE/k_{1}$ and it has the form
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \exp \left[ -\frac{2\pi }{k_{1}}\left( \omega
_{1}-\pi _{1}\right) \right] . \label{4.10}
\end{equation}
Considering positive $p_{x}>0$, we can take into account that exact $N_{n}^{%
\mathrm{cr}}$ is invariant under the simultaneous exchange $%
k_{1}\leftrightarrows k_{2}$ and $\pi _{1}\leftrightarrows -\pi _{2}$. In
this case $\pi _{1}$ is positive and large, $\pi _{1}>eE/k_{1}$, while $\pi
_{2}$ varies from negative to positive values, $-eE/k_{2}<\pi _{2}<\infty $.
We find a significant contribution in the range
\begin{equation}
-h_{2}<2\pi _{2}/k_{2}<K_{2}, \label{4.11}
\end{equation}%
where $K_{2}$ is any given large number, $K_{2}\gg K_{\bot }$. In this
range, similarly to the case of the negative $p_{x}$, the differential mean
numbers in the leading-order approximation are%
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \frac{\exp \left[ -\frac{\pi }{k_{2}}\left(
\omega _{2}+\pi _{2}\right) \right] }{\sinh \left( 2\pi \omega
_{2}/k_{2}\right) }\times \left\{
\begin{array}{c}
\sinh \left[ \pi \left( \omega _{2}-\pi _{2}\right) /k_{2}\right] \ \mathrm{%
for\ fermions} \\
\cosh \left( \pi \left( \omega _{2}-\pi _{2}\right) /k_{2}\right) \ \mathrm{%
for\ bosons}%
\end{array}%
\right. \,. \label{4.12a}
\end{equation}%
Assuming $m/k_{2}\gg 1$, we find for bosons and fermions that significant
value of $N_{n}^{\mathrm{cr}}$ is in the range $-eE/k_{2}<\pi _{2}<-\pi
_{\bot }$ and it has a form%
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \exp \left[ -\frac{2\pi }{k_{2}}\left( \omega
_{2}+\pi _{2}\right) \right] . \label{4.13}
\end{equation}
Consequently, the quantity $N_{n}^{\mathrm{cr}}$ is almost constant over the
wide range of longitudinal momentum $p_{x}$ for any given $\lambda $
satisfying Eq.~(\ref{4.2}). When $h_{1},h_{2}\rightarrow \infty $, one
obtains the well-known result in a constant uniform electric field \cite%
{Nikis70a,Nikis70b},%
\begin{equation}
N_{n}^{\mathrm{cr}}\rightarrow N_{n}^{\mathrm{uni}}=e^{-\pi \lambda }.
\label{uni}
\end{equation}
\subsection{Total quantities\label{Ss4.2}}
In this subsection we estimate the total number $N^{\mathrm{cr}}$ of pairs
created by the peak electric field. To compute this number, one has to sum
the corresponding differential mean numbers $N_{n}^{\mathrm{cr}}$ over the
momenta $\mathbf{p}$ and, in the Fermi case, to{\Huge \ } sum over the spin
projections. Once $N_{n}^{\mathrm{cr}}$ does not depend on the spin
variables, the latter sum results in a multiplicative numerical factor $%
J_{(d)}=2^{\left[ d/2\right] -1}$ for fermions ($J_{\left( d\right) }=1$ for
bosons). Then replacing the sum over the momenta in Eq.~(\ref{TN}) by an
integral, the total number of pairs created from the vacuum takes the form%
\begin{equation}
N^{\mathrm{cr}}=\frac{V_{\left( d-1\right) }}{(2\pi )^{d-1}}J_{(d)}\int d%
\mathbf{p}N_{n}^{\mathrm{cr}}\,. \label{asy5}
\end{equation}
Due to the structure of the coefficients $g\left( _{-}|^{+}\right) $
presented in section \ref{Ss2.3}, it is clear that a direct integration of
combinations of{\Huge \ }hypergeometric functions involved in the absolute
value of (\ref{gp}) and (\ref{gpKG}) is overcomplicated. Nevertheless the
analysis presented in section \ref{Ss4.1}\ reveals that the dominant
contributions for particle creation by a slowly varying field occurs in the
ranges of large kinetic momenta, whose differential quantities have the
asymptotic forms (\ref{4.10}) for $p_{x}<0$ and (\ref{4.13}) for $p_{x}>0$.
Therefore, one may represent the total number (\ref{asy5}) as%
\begin{eqnarray}
&&N^{\mathrm{cr}}=V_{\left( d-1\right) }n^{\mathrm{cr}}\,,\ \ n^{\mathrm{cr}%
}=\frac{J_{(d)}}{(2\pi )^{d-1}}\int_{\sqrt{\lambda }<K_{\bot }}d\mathbf{p}%
_{\bot }I_{\mathbf{p}_{\bot }},\ \ I_{\mathbf{p}_{\bot }}=I_{\mathbf{p}%
_{\bot }}^{\left( 1\right) }+I_{\mathbf{p}_{\bot }}^{\left( 2\right) },
\notag \\
&&I_{\mathbf{p}_{\bot }}^{\left( 1\right) }=\int_{-\infty }^{0}dp_{x}N_{n}^{%
\mathrm{cr}}\approx \int_{\pi _{\perp }}^{eE/k_{1}}d\pi _{1}\exp \left[ -%
\frac{2\pi }{k_{1}}\left( \omega _{1}-\pi _{1}\right) \right] \,, \notag \\
&&I_{\mathbf{p}_{\bot }}^{\left( 2\right) }=\int_{0}^{\infty }dp_{x}N_{n}^{%
\mathrm{cr}}\approx \int_{\pi _{\perp }}^{eE/k_{2}}d\left\vert \pi
_{2}\right\vert \exp \left[ -\frac{2\pi }{k_{2}}\left( \omega
_{2}-\left\vert \pi _{2}\right\vert \right) \right] \,. \label{tot2}
\end{eqnarray}
Using the change of the variables%
\begin{equation*}
s=\frac{2}{k_{1}\lambda }\left( \omega _{1}-\pi _{1}\right) \,,
\end{equation*}%
and neglecting exponentially small contributions, we represent the quantity $%
I_{\mathbf{p}_{\bot }}^{\left( 1\right) }$ as%
\begin{equation}
I_{\mathbf{p}_{\bot }}^{\left( 1\right) }\approx \int_{1}^{\infty }\frac{ds}{%
s}\omega _{1}e^{-\pi \lambda s}\,. \label{4.15}
\end{equation}%
Similarly, using the change of variables%
\begin{equation*}
s=\frac{2}{k_{2}\lambda }\left( \omega _{2}-\left\vert \pi _{2}\right\vert
\right) \,,
\end{equation*}%
we represent the quantity $I_{\mathbf{p}_{\bot }}^{\left( 2\right) }$ as%
\begin{equation}
I_{\mathbf{p}_{\bot }}^{\left( 2\right) }\approx \int_{1}^{\infty }\frac{ds}{%
s}\omega _{2}e^{-\pi \lambda s}\,. \label{4.16}
\end{equation}%
The leading contributions for both integrals (\ref{4.15}) and (\ref{4.16})
come from the range near $s\rightarrow 1$, where $\omega _{1}$ and $\omega
_{2}$ are approximately given by,%
\begin{equation*}
\omega _{1}\approx \frac{eE}{sk_{1}}\,,\ \ \omega _{2}\approx \frac{eE}{%
sk_{2}}\,.
\end{equation*}%
Consequently the leading term in $I_{\mathbf{p}_{\bot }}$ (\ref{tot2}) takes
the following final form,%
\begin{equation}
I_{\mathbf{p}_{\bot }}\approx \left( \frac{eE}{k_{1}}+\frac{eE}{k_{2}}%
\right) \int_{1}^{\infty }\frac{ds}{s^{2}}e^{-\pi \lambda s}=eE\left( \frac{1%
}{k_{1}}+\frac{1}{k_{2}}\right) e^{-\pi \lambda }G\left( 1,\pi \lambda
\right) , \label{4.17}
\end{equation}%
where%
\begin{equation}
G\left( \alpha ,x\right) =\int_{1}^{\infty }\frac{ds}{s^{\alpha +1}}%
e^{-x\left( s-1\right) }=e^{x}x^{\alpha }\Gamma \left( -\alpha ,x\right) ,
\label{4.19a}
\end{equation}%
and $\Gamma \left( -\alpha ,x\right) $ is the incomplete gamma function.
Neglecting the exponentially small contribution, one can represent the integral
over $\mathbf{p}_{\bot }$ in Eq.~(\ref{tot2}) (where $I_{\mathbf{p}_{\bot }}$
is given by Eq.~(\ref{4.17})) as
\begin{equation*}
\int_{\sqrt{\lambda }<K_{\bot }}d\mathbf{p}_{\bot }I_{\mathbf{p}_{\bot
}}\approx \int_{\sqrt{\lambda }<\infty }d\mathbf{p}_{\bot }I_{\mathbf{p}%
_{\bot }}.
\end{equation*}%
Then calculating the Gaussian integral,%
\begin{equation}
\int d\mathbf{p}_{\bot }\exp \left( -\pi s\frac{\mathbf{p}_{\bot }^{2}}{eE}%
\right) =\left( \frac{eE}{s}\right) ^{d/2-1}, \label{4.18}
\end{equation}%
we find%
\begin{equation}
n^{\mathrm{cr}}=r^{\mathrm{cr}}\left( \frac{1}{k_{1}}+\frac{1}{k_{2}}\right)
G\left( \frac{d}{2},\pi \frac{m^{2}}{eE}\right) ,\;\;r^{\mathrm{cr}}=\frac{%
J_{(d)}\left( eE\right) ^{d/2}}{(2\pi )^{d-1}}\exp \left\{ -\pi \frac{m^{2}}{%
eE}\right\} . \label{4.19}
\end{equation}
Using the considerations presented above, one can perform the summation
(integration) in Eq.~(\ref{vacprob}) to obtain the vacuum-to-vacuum
probability $P_{v}$,%
\begin{eqnarray}
&&P_{v}=\exp \left( -\mu N^{\mathrm{cr}}\right) ,\;\;\mu =\sum_{l=0}^{\infty
}\frac{(-1)^{(1-\kappa )l/2}\epsilon _{l+1}}{(l+1)^{d/2}}\exp \left( -l\pi
\frac{m^{2}}{eE}\right) \;, \notag \\
&&\epsilon _{l}=G\left( \frac{d}{2},l\pi \frac{m^{2}}{eE}\right) \left[
G\left( \frac{d}{2},\pi \frac{m^{2}}{eE}\right) \right] ^{-1}. \label{4.20}
\end{eqnarray}
These results allow us to establish an immediate comparison with the
one-parameter regularizations of the constant field, namely the $T$-constant
and Sauter-like electric fields \cite{GavGit96}.\textsf{\ }We note that in
all these cases the quantity\ is quasiconstant over the wide range of the
longitudinal momentum $p_{x}$\ for any given\emph{\ }$\lambda ,$ i.e.,\emph{%
\ }$N_{n}^{\mathrm{cr}}\sim e^{-\pi \lambda }$\emph{. }Pair creation effects
in such fields are proportional to increments of longitudinal kinetic
momentum, $\Delta U=e\left\vert A_{x}\left( +\infty \right) -A_{x}\left(
-\infty \right) \right\vert $, which are%
\begin{eqnarray}
\Delta U_{\mathrm{p}} &=&eE\left( k_{1}^{-1}+k_{2}^{-1}\right) \;\mathrm{%
for\;peak\ field,} \notag \\
\Delta U_{\mathrm{T}} &=&eET\;\;\mathrm{for\;T}\text{\textrm{-}}\mathrm{%
const\ field,} \notag \\
\Delta U_{\mathrm{S}} &=&2eET_{\mathrm{S}}\;\;\mathrm{for\;Sauter}\text{%
\textrm{-}}\mathrm{like\ field}\text{.} \label{4.21}
\end{eqnarray}%
This fact allows one to compare pair creation effects in such fields. Using
the quantities introduced, we can represent the densities $n^{\mathrm{cr}}$
as follows:
\begin{eqnarray}
n^{\mathrm{cr}} &=&r^{\mathrm{cr}}\frac{\Delta U_{\mathrm{p}}}{eE}G\left(
\frac{d}{2},\pi \frac{m^{2}}{eE}\right) ,\ \mathrm{for\;peak\ field,} \notag
\\
n^{\mathrm{cr}} &=&r^{\mathrm{cr}}\frac{\Delta U_{\mathrm{T}}}{eE}\;\;%
\mathrm{for\;T}\text{\textrm{-}}\mathrm{constant\ field}, \notag \\
n^{\mathrm{cr}} &=&r^{\mathrm{cr}}\frac{\Delta U_{\mathrm{S}}}{2eE}\delta \;%
\mathrm{for\;Sauter}\text{\textrm{-}}\mathrm{like\ field}, \label{4.23}
\end{eqnarray}%
where%
\begin{equation*}
\delta =\int_{0}^{\infty }dtt^{-1/2}(t+1)^{-\left( d-2\right) /2}\exp \left(
-t\pi \frac{m^{2}}{eE}\right) =\sqrt{\pi }\Psi \left( \frac{1}{2},-\frac{d-2%
}{2};\pi \frac{m^{2}}{eE}\right) ,
\end{equation*}%
and $\Psi \left( a,b;x\right) $ is the confluent hypergeometric function
\cite{BatE53}.
Thus, for a given magnitude of the electric field $E$ one can compare the
pair creation effects in fields with equal increment of the longitudinal
kinetic momentum, or one can determine such increments of the longitudinal
kinetic momenta,\emph{\ }for which particle creation effects are the same.
In Eq.(\ref{4.21}) $T$ is the time duration of the $T$-constant field.
Equating the densities $n^{\mathrm{cr}}$ for Sauter-like field and
for the peak field to the density $n^{\mathrm{cr}}$ for the $T$-constant field, we find an effective duration time $T_{eff}$ in both cases,%
\begin{eqnarray}
T_{eff} &=&T_{\mathrm{S}}\delta \;\mathrm{for\;Sauter}\text{\textrm{-}}%
\mathrm{like\ field}, \notag \\
T_{eff} &=&\left( k_{1}^{-1}+k_{2}^{-1}\right) G\left( \frac{d}{2},\pi \frac{%
m^{2}}{eE}\right) \;\mathrm{for\;the\ peak\ field}. \label{4.24}
\end{eqnarray}%
By the definition $T_{eff}=T$ for the $T$-constant field. One can say that
the Sauter-like and the peak electric fields with the same $T_{eff}=T$ are
equivalent to the $T$-constant field in pair production.
If the electric field $E$ is weak, $m^{2}/eE\gg $ $1$, one can use
asymptotic expressions for the $\Psi $-function and the incomplete gamma
function. Thus, we obtain
\begin{equation}
G\left( \frac{d}{2},\pi \frac{m^{2}}{eE}\right) \approx \frac{eE}{\pi m^{2}}%
,\ \ \delta \approx \sqrt{eE}/m\ . \label{4.26}
\end{equation}%
If the electric field $E$ is strong enough, $m^{2}/eE\ll 1$, it follows from
a corresponding representation for the $\Psi $-function, see Ref.~\cite%
{BatE53}, that its leading term does not depend on the dimensionless
parameter $m^{2}/eE$ and reads%
\begin{equation}
\Psi \left( \frac{1}{2},-\frac{d-2}{2};\pi \frac{m^{2}}{eE}\right) \approx
\Gamma \left( d/2\right) /\Gamma \left( d/2+1/2\right) . \label{asy12}
\end{equation}%
Then, for example, $\delta \approx \pi /2$ if $d=3$ and $\delta \approx 4/3$
if $d=4$. The leading term of $G$-function, which is given by Eq. (\ref%
{4.19a}), does not depend on the parameter $m^{2}/eE$ either,%
\begin{equation}
G\left( \frac{d}{2},\pi \frac{m^{2}}{eE}\right) \approx \frac{2}{d}.
\label{4.27}
\end{equation}
It is clear that there is a time range where Sauter-like and the peak
electric fields coincide with a $T$-constant field. Out of this range both
these fields have an exponential behavior and can be compared. Assuming $%
k_{1}\sim k_{2}$, we have%
\begin{eqnarray}
\dot{U} &\approx &eEe^{-2\left\vert t\right\vert /T_{\mathrm{S}}}\;\mathrm{if%
}\;\left\vert t\right\vert /T_{\mathrm{S}}\gg 1\;\mathrm{for\;Sauter}\text{%
\textrm{-}}\mathrm{like\ field}, \notag \\
\dot{U} &\approx &eEe^{-k_{1}\left\vert t\right\vert }\;\mathrm{if}%
\;k_{1}\left\vert t\right\vert \gg 1\;\mathrm{for\;peak\ field}.
\label{4.28}
\end{eqnarray}%
If the field is weak, $m^{2}/eE\gg $ $1$, we see that
\begin{equation*}
k_{1}T_{\mathrm{S}}=\frac{2\sqrt{eE}}{\pi m}\ll 1,
\end{equation*}%
that is, the peak electric field switches on and off much more slowly than the
Sauter-like field. If the field is strong, $m^{2}/eE\ll 1$, this
dimensionless parameter turns to unity,%
\begin{equation*}
k_{1}T_{\mathrm{S}}=\left\{
\begin{array}{l}
8/\left( 3\pi \right) \;\;\mathrm{if}\;d=3 \\
3/4\;\;\mathrm{if}\;d=4%
\end{array}%
\right. .
\end{equation*}%
In this case, the peak electric field switches on and off not much slowly
than the Sauter-like field.
Another global{\Huge \ }quantity is the vacuum-to-vacuum transition
probability $P_{v}$. It is given by Eq.~(\ref{4.20}) for the peak field and
has the form similar to that for the $T$-constant and the Sauter-like fields
with the corresponding $N^{\mathrm{cr}}$, and%
\begin{eqnarray}
\epsilon _{l} &=&\epsilon _{l}^{\mathrm{T}}=1\;\mathrm{for\;T}\text{\textrm{-%
}}\mathrm{constant\ field}, \notag \\
\epsilon _{l} &=&\epsilon _{l}^{\mathrm{S}}=\delta ^{-1}\sqrt{\pi }\Psi
\left( \frac{1}{2},-\frac{d-2}{2};l\pi \frac{m^{2}}{eE}\right) \;\mathrm{%
for\;Sauter}\text{\textrm{-}}\mathrm{like\ field}. \label{4.29}
\end{eqnarray}%
If the field is weak, $m^{2}/eE\gg 1$, then $\;\epsilon _{l}^{\mathrm{S}%
}\approx l^{-1/2}$ for the Sauter-like field and $\epsilon _{l}\approx
l^{-1} $ for the peak field. Then $\mu \approx 1$ for both fields and we see
that the identification with $T_{eff}=T$, given by Eq.~(\ref{4.24}), is the
same as the one extracted from the comparison of total densities $n^{\mathrm{%
cr}}$. In the case of a strong field, $m^{2}/eE\ll 1$, all the terms with
different $\epsilon _{l}^{\mathrm{S}}$ and $\epsilon _{l}$ contribute
significantly to the sum in Eq.~(\ref{4.20}) if $l\pi m^{2}/eE\sim 1$, and
the $\mu $ quantities differ essentially from the case of the $T$-constant
field. However, for a very strong field, $l\pi m^{2}/eE\ll 1$, the leading
contribution for $\epsilon _{l}$ has a quite simple form $\epsilon _{l}^{%
\mathrm{S}}\approx \epsilon _{l}\approx 1$. In this case the quantities $\mu
$ are the same for all these fields, namely%
\begin{equation*}
\mu \approx \sum_{l=0}^{\infty }\frac{(-1)^{(1-\kappa )l/2}}{(l+1)^{d/2}},
\end{equation*}%
and the identification with $T_{eff}=T$ is the same as the one extracted
from the comparison of the total densities $n^{\mathrm{cr}}$.
It is clear that different total quantities, such as the total number of
created pairs and the vacuum-to-vacuum transition probability {\Huge \ }%
discussed above, in the general case lead to different identifications with $%
T_{eff}=T$. We believe that some of these quantities are more adequate for
such an identification. In this connection, it should be noted that in
small-gradient fields, the total vacuum mean values, such as mean electric
currents and the mean energy-momentum tensor, are usually of interest; see,
e.g. Refs. \cite{GavGit08,GavGit08b,GavGitY12}. These total quantities are represented
by corresponding sums of differential numbers of created particles.
Therefore, relations between the total numbers and parameters $\Delta U_{%
\mathrm{p}}$, $\Delta U_{\mathrm{T}}$, and $\Delta U_{\mathrm{S}}$ derived
above are also important. Such relations derived from the vacuum-to-vacuum
transition probability $P_{v}$\ are interesting in semiclassical approaches
based on Schwinger's technics \cite{Sch51}. We recall that the semiclassical%
{\LARGE \ }approaches work in the case of weak external fields $m^{2}/eE\gg $%
\emph{\ }$1$. It should be noted that in the case of a strong field when the
semiclassical approach is not applicable, the probability $P_{v}$\ has no
direct relation to vacuum mean values of the above discussed physical
quantities.
\section{Configurations with sharp fields\label{S5}}
\subsection{Short pulse field\label{Ss5.1}}
Choosing certain parameters of the peak field, one can obtain electric
fields that exist only for a short time in a vicinity of the time instant $%
t=0$. The latter fields switch on and/or switch off{\Huge \ }%
\textquotedblleft abruptly\textquotedblright\ near the time instant $t=0$.
Let us consider large parameters $k_{1}$, $k_{2}\rightarrow \infty $ with a
fixed ratio $k_{1}/k_{2}$. The corresponding asymptotic potentials, $U\left(
+\infty \right) =eEk_{2}^{-1}$ and $U\left( -\infty \right) =-eEk_{1}^{-1}$
define finite increments of the longitudinal kinetic momenta $\Delta U_{1}$
and $\Delta U_{2}$ for increasing and decreasing parts, respectively, \ \ \
\ \ ,%
\begin{equation}
\Delta U_{1}=U\left( 0\right) -U\left( -\infty \right)
=eEk_{1}^{-1},\;\;\Delta U_{2}=U\left( +\infty \right) -U\left( 0\right)
=eEk_{2}^{-1}. \label{5.1}
\end{equation}%
Such a case corresponds to a very short pulse of the electric field. At the
same time this configuration imitates well enough a $t$-electric rectangular
potential step (it is an analog of the Klein step, which is an $x$-electric
rectangular step; see Ref.~\cite{GavGit15}) and coincides with it as $k_{1}$%
, $k_{2}\rightarrow \infty $. Thus, these field configurations can be
considered as regularizations of rectangular step. We assume that
sufficiently large $k_{1}$ and $k_{2}$ satisfy the following inequalities:
\begin{equation}
\Delta U_{1}/k_{1}\ll 1,\;\;\Delta U_{2}/k_{2}\ll 1,\;\;\max \left( \omega
_{1}/k_{1},\omega _{2}/k_{2}\right) \ll 1 \label{5.2}
\end{equation}%
for any given $\pi _{\perp }$ and $\pi _{1,2}=p_{x}-U\left( \mp \infty
\right) $. In this case the confluent hypergeometric function can be
approximated by the first two terms in Eq. (\ref{chf}), which are $\Phi
\left( a,c;\eta \right) $, $c_{j}\approx 1$, and $\ \ a_{j}\approx \left(
1+\chi \right) /2$. Then for fermions, we obtain the result
\begin{equation}
N_{n}^{\mathrm{cr}}=\frac{\left( \omega _{1}+\pi _{1}\right) \left( \Delta
U_{2}+\Delta U_{1}+\omega _{2}-\omega _{1}\right) ^{2}}{4\omega _{1}\omega
_{2}\left( \omega _{2}-\pi _{2}\right) } \label{5.3}
\end{equation}%
which does not depend on $k_{1,2}$. For bosons, we obtain%
\begin{equation}
N_{n}^{\mathrm{cr}}=\frac{\left( \omega _{2}-\omega _{1}\right) ^{2}}{%
4\omega _{1}\omega _{2}}. \label{5.4}
\end{equation}
In contrast to the Fermi case, where $N_{n}^{\mathrm{cr}}\leq 1$, in the
Bose case, the differential numbers $N_{n}^{\mathrm{cr}}$ are unbounded in
two ranges of the longitudinal kinetic momenta, in the range where $\omega
_{1}/\omega _{2}\rightarrow \infty $ and in the range where $\omega
_{2}/\omega _{1}\rightarrow \infty $. In these ranges they are%
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \frac{1}{4}\max \left\{ \omega _{1}/\omega
_{2},\omega _{2}/\omega _{1}\right\} . \label{5.5}
\end{equation}
If $k_{1}=k_{2}$ (in this case $\Delta U_{2}=\Delta U_{1}=\Delta U/2$), we
can compare the above results with the results of the regularization of
rectangular steps by the Sauter-like potential \cite{AdoGavGit15},\ obtained
for a small $T_{\mathrm{S}}\rightarrow 0$\ and constant $\Delta U=2eET_{%
\mathrm{S}}$\ under the conditions\emph{\ }$\Delta UT_{\mathrm{S}}\ll 1$%
\emph{\ }and\emph{\ }$\max \left\{ T_{\mathrm{S}}\omega _{1},T_{\mathrm{S}%
}\omega _{2}\right\} \ll 1$. We see that both regularizations are in
agreement for fermions under the condition $\left\vert \omega _{2}-\omega
_{1}\right\vert \ll \Delta U$ , and for bosons under the condition $\left(
\omega _{2}-\omega _{1}\right) ^{2}\gg \left( \Delta U\right) ^{4}T_{\mathrm{%
S}}^{2}/4$, which is the general condition for applying the Sauter-like
potential for the regularization of rectangular step\emph{\ }for bosons.
\subsection{Exponentially decaying field\label{Ss5.2}}
In the examples, considered above, the pick field switches on and off
relatively smooth. Here we are going to consider a different essentially
asymmetric configuration of the peak field, when for example, the field
switches abruptly on at $t=0$, that is, $k_{1}$ is sufficiently large, while
the value of parameter $k_{2}>0$ remains arbitrary and includes the case of
a smooth switching off. Note that due to the invariance of the mean numbers $%
N_{n}^{\mathrm{cr}}$ under the simultaneous change $k_{1}\leftrightarrows
k_{2}$ and $\pi _{1}\leftrightarrows -\pi _{2}$, one can easily transform
this situation to the case with a large $k_{2}$ and arbitrary $k_{1}>0$.
Let us assume that a sufficiently large $k_{1}$ satisfies the inequalities
\begin{equation}
\Delta U_{1}/k_{1}\ll 1,\;\;\omega _{1}/k_{1}\ll 1. \label{5.6}
\end{equation}%
Then Eqs.~(\ref{4.0}) and (\ref{4b}) can be reduced to the following form%
\begin{equation}
\left\vert \Delta \right\vert ^{2}\approx \left\vert \Delta _{\mathrm{ap}%
}\right\vert ^{2}=e^{i\pi \nu _{2}}\left. \left\vert \left[ \chi \Delta
U_{1}+\omega _{2}-\omega _{1}+k_{2}h_{2}\left( -\frac{1}{2}+\frac{d}{d\eta
_{2}}\right) \right] \Phi \left( a_{2},c_{2};\eta _{2}\right) \right\vert
^{2}\right\vert _{t=0}\,. \label{5.7}
\end{equation}%
Under the condition%
\begin{equation}
-2p_{x}\Delta U_{1}\ll \pi _{\perp }^{2}+p_{x}^{2}, \label{5.8}
\end{equation}%
one can disregard the term $\chi \Delta U_{1}$ in Eq.~(\ref{5.7}) and write
approximately $\pi _{1}\approx p_{x}$. Thus, $\omega _{1}\approx \sqrt{%
p_{x}^{2}+\pi _{\perp }^{2}}$. In this approximation, leading terms do not
contain $\Delta U_{1}$, so that we obtain
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \left\{
\begin{array}{l}
\left\vert \mathcal{C}\Delta _{\mathrm{ap}}\right\vert ^{2}\ \mathrm{for\
fermions} \\
\left\vert \mathcal{C}_{\mathrm{sc}}\left. \Delta _{\mathrm{ap}}\right\vert
_{\chi =0}\right\vert ^{2}\ \mathrm{for\ bosons}%
\end{array}%
\right. \,. \label{5.9}
\end{equation}%
In fact, differential mean numbers obtained in these approximations are the
same as in the so-called exponentially decaying electric field, given by the
potential%
\begin{equation}
A_{x}^{\mathrm{ed}}\left( t\right) =E\left\{
\begin{array}{l}
0\,,\ \ t\in \mathrm{I}\, \\
k_{2}^{-1}\left( e^{-k_{2}t}-1\right) \,,\ \ t\in \mathrm{II}%
\end{array}%
\right. \,. \label{com1}
\end{equation}%
The effect of pair creation in the exponentially decaying electric field was
studied previously by us in Ref. \cite{AdoGavGit14}. Note that the pair
creation due to an exponentially decaying background has been studied in de
Sitter spacetime and for the constant electric field in two dimensional de
Sitter spacetime; see, e.g., \cite{AndMot14,StStHue16} and references
therein. Under condition (\ref{5.8}), the results presented by Eqs.~(\ref%
{5.9}) for arbitrary $k_{2}>0$ are in agreement with ones obtained in Ref.
\cite{AdoGavGit14}.
Let us consider the most asymmetric case when Eqs.~(\ref{5.9}) hold and when
the increment of the longitudinal kinetic momentum due to exponentially
decaying electric field is sufficiently large ($k_{2}$ are sufficiently
small),%
\begin{equation}
h_{2}=2\Delta U_{2}/k_{2}\gg \max \left( 1,m^{2}/eE\right) \,. \label{5.11}
\end{equation}
As it was noted in Section \ref{Ss4.1}, in this case only the range of fixed
$\pi _{\perp }$ is essential and we assume that the inequality (\ref{4.2})
holds. In the case under consideration $K_{\bot }$ is any given number
satisfying the condition%
\begin{equation}
h_{2}\gg K_{\bot }^{2}\gg \max \left( 1,m^{2}/eE\right) \,. \label{5.12}
\end{equation}
It should be noted that the distribution $N_{n}^{\mathrm{cr}}$, given by
Eqs.~(\ref{5.9}) for this most asymmetric case\emph{\ }coincides with the
one obtained in our recent work \cite{AdoGavGit14}, where the exponentially
decreasing field was considered. However, the detailed study of this
distribution was not performed there. In the following, we study how this
distribution depends on the parameters $p_{x}$ and $\pi _{\perp }$.
In the case of large negative $p_{x}$, $p_{x}<0$ and $\left\vert
p_{x}\right\vert /\sqrt{eE}>K_{\bot }$, using appropriate asymptotic
expressions of the confluent hypergeometric function, given in Appendix \ref%
{Ap2}, one can conclude that numbers $N_{n}^{\mathrm{cr}}$ are negligibly
small both for fermions and bosons. The same holds true for very large
positive $p_{x}$, such that $2\pi _{2}/k_{2}>K_{2}$, where $K_{2}$ is any
given large number, $K_{2}\gg K_{\bot }$. We see that $N_{n}^{\mathrm{cr}}$
are nonzero\ only in the range%
\begin{equation}
-K_{\bot }<p_{x}/\sqrt{eE},\;\;2\pi _{2}/k_{2}<K_{2}. \label{5.13}
\end{equation}%
This range can be divided in three subranges,%
\begin{eqnarray}
\mathrm{(a)} &&\;\left( 1-\varepsilon \right) h_{2}\leq -2\pi
_{2}/k_{2}<\left( 1+\varepsilon \right) h_{2}, \notag \\
\mathrm{(b)} &&\;h_{2}/g_{1}<-2\pi _{2}/k_{2}<\left( 1-\varepsilon \right)
h_{2}, \notag \\
\mathrm{(c)} &&\;-K_{2}<-2\pi _{2}/k_{2}<h_{2}/g_{1}, \label{5.14}
\end{eqnarray}%
where $g_{1}$ and $\varepsilon $ are any given numbers satisfying the
conditions $g_{1}$ $\gg 1$ and $\varepsilon \ll 1$. We assume that $%
\varepsilon \sqrt{h_{2}}\gg 1$. Note that%
\begin{equation*}
\tau _{2}=ih_{2}/c_{2}\approx \frac{h_{2}k_{2}}{2\left\vert \pi
_{2}\right\vert }
\end{equation*}%
in the ranges (a) and (b). Then in the ranges (a) and (b), $\tau _{2}$
varies from $1-\varepsilon $ to $g_{1}$. In the range (b), parameters $\eta
_{2}$ and $c_{2}$ are large with $a_{2}$ fixed and $\tau _{2}>1$. In this
case, using the asymptotic expression of the confluent hypergeometric
function given by Eq.~(\ref{A10}) in Appendix \ref{Ap2}, we find that%
\begin{equation}
N_{n}^{\mathrm{cr}}=\exp \left[ -\frac{2\pi }{k_{2}}\left( \omega _{2}+\pi
_{2}\right) \right] \left[ 1+O\left( \left\vert \mathcal{Z}_{2}\right\vert
^{-1}\right) \right] \label{5.15}
\end{equation}%
both for fermions and bosons, where $\mathcal{Z}_{2}$ is given by Eq.~(\ref%
{A5}) in the Appendix \ref{Ap2}. We note that\textrm{\ }modulus $\left\vert
\mathcal{Z}_{2}\right\vert ^{-1}$ varies from $\left\vert \mathcal{Z}%
_{2}\right\vert ^{-1}\sim \left( \varepsilon \sqrt{h_{2}}\right) ^{-1}$ to $%
\left\vert \mathcal{Z}_{2}\right\vert ^{-1}\sim \left[ \left( g_{1}-1\right)
\sqrt{h_{2}}\right] ^{-1}$. Approximately, expression (\ref{5.15}) can be
written as%
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \exp \left( -\frac{\pi \pi _{\perp }^{2}}{%
k_{2}\left\vert \pi _{2}\right\vert }\right) . \label{5.16}
\end{equation}%
Note that $eE/g_{1}<k_{2}\left\vert \pi _{2}\right\vert <\left(
1-\varepsilon \right) eE$ in the range (b).
It is clear that the distribution $N_{n}^{\mathrm{cr}}$ given by Eq.~(\ref%
{5.16}) has the following limiting form:%
\begin{equation*}
N_{n}^{\mathrm{cr}}\rightarrow e^{-\pi \lambda }\ \ \mathrm{as\ \ }%
k_{2}\left\vert \pi _{2}\right\vert \rightarrow \left( 1-\varepsilon \right)
eE\ .
\end{equation*}%
In the range (a), we can use the asymptotic expression for the confluent
hypergeometric function given by Eq.~(\ref{A1}) in Appendix \ref{Ap2} to
verify that $N_{n}^{\mathrm{cr}}$ is finite and restricted, $N_{n}^{\mathrm{%
cr}}\lesssim e^{-\pi \lambda },$ both for fermions and bosons. Thus, we see
that the well-known distribution obtained by Nikishov \cite{Nikis70a} in a
constant uniform electric field is reproduced in an exponentially decaying
electric field in the range of a large increment of the longitudinal kinetic
momentum, $-\pi _{2}\sim eE/k_{2}$.
In the range (c), we can use the asymptotic expression of the confluent
hypergeometric function for large $h_{2}$ with fixed $a_{2}$ and $c_{2}$
given by Eq.~(\ref{A11}) in Appendix \ref{Ap2} to get the following result:
\begin{equation}
N_{n}^{\mathrm{cr}}\approx \frac{\exp \left[ -\frac{\pi }{k_{2}}\left(
\omega _{2}+\pi _{2}\right) \right] }{\sinh \left( 2\pi \omega
_{2}/k_{2}\right) }\times \left\{
\begin{array}{c}
\sinh \left[ \pi \left( \omega _{2}-\pi _{2}\right) /k_{2}\right] \ \mathrm{%
for\ fermions} \\
\cosh \left( \pi \left( \omega _{2}-\pi _{2}\right) /k_{2}\right) \ \mathrm{%
for\ bosons}%
\end{array}%
\right. \, \label{5.17}
\end{equation}%
in the leading-order approximation. The same distribution was obtained for $%
p_{x}>0$ in a slowly varying field; see Eq.~(\ref{4.12a}).
For $m/k_{2}\gg 1$, distribution~(\ref{5.17}) has the form (\ref{5.15}),
which means that distribution (\ref{5.15}) holds in the range (c) as well.
Using the above considerations, we can estimate the total number $N^{\mathrm{%
cr}}$ (\ref{asy5}) of pairs created by an exponentially decaying electric
field. To this end, we represent the leading terms of integral (\ref{asy5})
as a sum of two contributions, one due to the range (a) and another due to
the ranges (b) and (c):
\begin{eqnarray}
&&N^{\mathrm{cr}}=V_{\left( d-1\right) }n^{\mathrm{cr}}\,,\ \ n^{\mathrm{cr}%
}=\frac{J_{(d)}}{(2\pi )^{d-1}}\int_{\sqrt{\lambda }<K_{\bot }}d\mathbf{p}%
_{\bot }I_{\mathbf{p}_{\bot }},\ \ I_{\mathbf{p}_{\bot }}=I_{\mathbf{p}%
_{\bot }}^{\left( 1\right) }+I_{\mathbf{p}_{\bot }}^{\left( 2\right) },
\notag \\
&&I_{\mathbf{p}_{\bot }}^{\left( 1\right) }=\int_{\pi _{2}\in \mathrm{(a)}%
}d\pi _{2}N_{n}^{\mathrm{cr}}\,,\ \ I_{\mathbf{p}_{\bot }}^{\left( 2\right)
}=\int_{\pi _{2}\in \mathrm{(b)\cup (c)}}d\pi _{2}N_{n}^{\mathrm{cr}}\,.
\label{5.18}
\end{eqnarray}%
Note that numbers $N_{n}^{\mathrm{cr}}$ given by Eqs.~(\ref{5.17}) are
negligibly small in the range $-\pi _{2}\lesssim \pi _{\bot }$. Then the
integral $I_{\mathbf{p}_{\bot }}^{\left( 2\right) }$ in Eq.~(\ref{5.18}) can
be taken from Eq.~(\ref{tot2}). Using the results of section \ref{Ss4.2}, we can
verify that the leading term in $I_{\mathbf{p}_{\bot }}^{\left( 2\right) }$
takes the following final form:%
\begin{equation}
I_{\mathbf{p}_{\bot }}^{\left( 2\right) }\approx \frac{eE}{k_{2}}%
\int_{1}^{\infty }\frac{ds}{s^{2}}e^{-\pi \lambda s}=\frac{eE}{k_{2}}e^{-\pi
\lambda }G\left( 1,\pi \lambda \right) , \label{5.19}
\end{equation}%
where $G\left( \alpha ,x\right) $ is given by Eq.~(\ref{4.19a}). The
integral $I_{\mathbf{p}_{\bot }}^{\left( 1\right) }$ in Eq.~(\ref{5.18}) is
of the order of $e^{-\pi \lambda }\varepsilon eE/k_{2}$, such that it is
relatively small comparing to integral (\ref{5.19}). Thus, the dominant
contribution is given by integral (\ref{5.19}), $I_{\mathbf{p}_{\bot
}}\approx I_{\mathbf{p}_{\bot }}^{\left( 2\right) }$. Then calculating the
Gaussian integral, we find%
\begin{equation}
n^{\mathrm{cr}}=\frac{r^{\mathrm{cr}}}{k_{2}}G\left( \frac{d}{2},\pi \frac{%
m^{2}}{eE}\right) , \label{5.20}
\end{equation}%
where $r^{\mathrm{cr}}$ is given by Eq.~(\ref{4.19}). We see that $n^{%
\mathrm{cr}}$ given by Eq.~(\ref{5.20}) is the $k_{2}$-dependent part of the
mean number density of pairs created in the slowly varying peak field~(\ref%
{4.19}). In the case of a strong field, $n^{\mathrm{cr}}$ given by Eq.~(\ref%
{5.20}) has the form obtained in Ref.\textrm{\cite{AdoGavGit14}.}
Finally, we can see that the vacuum-to-vacuum probability is%
\begin{equation}
P_{v}=\exp \left( -\mu N^{\mathrm{cr}}\right) , \label{5.21}
\end{equation}%
where $\mu $ is given by Eq.~(\ref{4.20}).
\section{Concluding remarks}
Using the strong-field QED, we consider for the first time particle creation
in the so-called peak electric field, which is a combination of two
exponential parts, one exponentially increasing and another
exponentially decreasing. This is an addition to the few previously known
exactly solvable cases, where one can perform nonperturbative calculations
of all the characteristics of particle creation process. For a
certain choice of parameters, the peak electric field produces a particle%
creation effect similar to that of the Sauter-like electric field,
or of{\Huge \ }the constant electric field, or of the electric pulse of
laser beams. Besides, by varying the peak field parameters we can change the
asymmetry rate so that the resulting field turns out to be effectively
equivalent to the exponentially decaying field. All these asymptotic regimes
are discussed in detail and a comparison with the pure asymptotically
decaying field is considered. Moreover, the results obtained allow one to
study how the effects of switching on and off together or separately affect the
particle creation processes. Changing the parameters of the peak electric field
we can adjust its form to a specific physical situation, in particular,
imitate field configurations characteristic for graphene, Weyl semimetals
and so on.
\section*{Acknowledgements}
The work of the authors{\large \ }was supported by a grant from the Russian
Science Foundation, Research Project No. 15-12-10009.
|
\section{Introduction}
\label{sec:intro}
Schr\"{o}inger equation is the most famous equation in non-relativistic quantum mechanics. With the aim of this equation we are able to investigate many different systems in presence of various situations \cite{1,2}. There is vacant place in non-relativistic quantum mechanics, that is, in this formalism we can’t consider spin of particles in interactions similar what we do in relativistic version of quantum mechanics \cite{3}. Recently Ajaib published an article in which a fundamental form Schrödinger equation has been derived as \cite{4}
\begin{equation}\label{1}
- i{\partial _z}\psi = \left( {i\eta {\partial _t} + {\eta ^\dag }m} \right)\psi.
\end{equation}
Actually the author iterated Eq. (1) to obtain Schr\"{o}dinger equation. It yields to have
\begin{equation}\label{2}
{\left( { - i{\partial _z}} \right)^2}\psi = {\left( {i\eta {\partial _t}} \right)^2}\psi + im\left\{ {\eta ,{\eta ^\dag }} \right\}{\partial _t}\psi + {\left( {{\eta ^\dag }} \right)^2}{m^2}\psi,
\end{equation}
then in order to reach Schrödinger equation, the parameters have to obey below conditions
\begin{equation}\label{3}
\begin{array}{l}
{\eta ^2} = 0,\\\\
{\left( {{\eta ^\dag }} \right)^2} = 0,\\\\
\left\{ {\eta ,{\eta ^\dag }} \right\} = 2{I}.
\end{array}
\end{equation}
In which $I$ is the unit matrix and
\begin{equation}\label{4}
\eta = \frac{1}{{\sqrt 2 }}\left( {\begin{array}{*{20}{c}}
0&{ - i}&0&{ - 1}\\
{ - i}&0&1&0\\
0&1&0&{ - i}\\
{ - 1}&0&{ - i}&0
\end{array}} \right) = \frac{i}{{\sqrt 2 }}\left( {\begin{array}{*{20}{c}}
{{\sigma _1}}&{{\sigma _2}}\\
{ - {\sigma _2}}&{{\sigma _1}}
\end{array}} \right) = \frac{i}{{\sqrt 2 }}\left( {\begin{array}{*{20}{c}}
0&{I + {\sigma _2}}\\
{I - {\sigma _2}}&0
\end{array}} \right),
\end{equation}
Where ${\sigma _i}(i = 1,2)$ are Pauli matrixes and $\eta $ is non-Hermitian $({\eta ^\dag } = {\eta ^*} \ne \eta )$ and symmetric matrix $({\eta ^T} = \eta )$.
Eq. (1) can be extended in 3 dimensional by writing
\begin{equation}\label{5}
\left( {i({\mu _j}{\partial _j} + \eta {\partial _t}) + {\eta ^\dag }m} \right)\psi = 0, \qquad (j = 1,2,3)
\end{equation}
in which the parameters are
\begin{equation}\label{6}
\begin{array}{l}
{\mu _1} = i{\gamma _1}{\gamma _2} = \left( {\begin{array}{*{20}{c}}
{{\sigma _3}}&0\\
0&{{\sigma _3}}
\end{array}} \right),\\ \\
{\mu _2} = {\gamma _0}{\gamma _2} = \left( {\begin{array}{*{20}{c}}
0&{{\sigma _2}}\\
{{\sigma _2}}&0
\end{array}} \right),\\ \\
{\mu _3} = {\gamma _2}{\gamma _5} = \left( {\begin{array}{*{20}{c}}
{{\sigma _2}}&0\\
0&{ - {\sigma _2}}
\end{array}} \right),
\end{array}
\end{equation}
with the famous gamma matrices. In this article, we derive the Lagrange density corresponding Eq. (\ref{1}) in section 2. Then in Sec. 3 we derive time evolution relation of Eq. (\ref{1}) to probe the time evolution of such system. Section 4 contains theory of Lewis-Riesenfeld dynamical invariant method and details of deriving wave function in presence of time-dependent interaction. In the last section using the Lie algebra method we also derive the time evolution operator.
\section{The Lagrange Density}
In the quantum field theory the start point of any problem or system is a simple concept, the
Lagrange density $\ell ({\psi _\sigma },\frac{{\partial {\psi _\sigma }}}{{\partial {x_\mu }}})$ that Lagrange function can be obtained by integrating over the
three space
\begin{equation}\label{7}
L = \int\limits_V^{} {\ell ({\psi _\sigma },\frac{{\partial {\psi _\sigma }}}{{\partial {x_\mu }}})} {d^3}x,
\end{equation}
where,$\ell ({\psi _\sigma },\frac{{\partial {\psi _\sigma }}}{{\partial {x_\mu }}})$ as is shown in our notation, is function of wave field and all derivatives
because of not lead nonlocal theories, the higher order of derivatives are not considered \cite{3}. Following the variational principle, it is possible to write
\begin{equation}\label{8}
\delta \int {Ldt = } \delta \int {\ell ({\psi _\sigma },\frac{{\partial {\psi _\sigma }}}{{\partial {x_\mu }}}){d^4}x} = 0,
\end{equation}
which results in
\begin{equation}\label{9}
\frac{\partial }{{\partial {x^\mu }}}\frac{{\partial \ell }}{{\partial (\partial {\psi _\nu }/\partial {x^\mu })}} - \frac{{\partial \ell }}{{\partial {\psi _\nu }}} = 0,
\end{equation}
to have equation of motion. Let us to bring a simpler form of Eq. (\ref{9}) as \cite{3}
\begin{equation}\label{10}
\frac{\partial }{{\partial t}}\frac{{\partial \ell }}{{\partial ({{\dot {\bar{ \psi }}}_\nu })}} + \nabla .\frac{{\partial \ell }}{{\partial (\nabla {{\bar \psi }_\nu })}} - \frac{{\partial \ell }}{{\partial {{\bar \psi }_\nu }}} = 0,
\end{equation}
where dot is time partial derivative. Comparing with Eq. (\ref{1}), we find out that
\begin{equation}\label{11}
i\eta \frac{{\partial \psi }}{{\partial t}} = \frac{\partial }{{\partial t}}(\frac{{\partial \ell }}{{\partial \dot {\bar{\psi }}}}),
\end{equation}
or
\begin{equation}\label{12}
\ell = i\eta \dot {\bar {\psi}} \psi + f(\nabla \bar \psi ,\bar \psi ),
\end{equation}
that $f(\nabla \bar \psi ,\bar \psi )$ should be determined. Using Eq. (\ref{12}) and another comparison Eqs. (\ref{10}) and (\ref{1}) results
\begin{equation}\label{13a}
\nabla .(\frac{{\partial \ell }}{{\partial (\nabla \bar \psi )}}) = i{\mu _j}\nabla \psi ,
\end{equation}
so that
\begin{equation}\label{13b}
\frac{{\partial f(\nabla \bar \psi ,\bar \psi )}}{{\partial (\nabla \bar \psi )}} = i{\mu _j}\psi ,
\end{equation}
which reads
\begin{equation}\label{14}
f(\nabla \bar \psi ,\bar \psi ) = i{\mu _j}\nabla \bar \psi \psi + g(\bar \psi ),
\end{equation}
Where $g(\bar{\psi} )$ will be ascertained. Substituting Eq. (\ref{14}) into (\ref{12}), caused to rewrite Eq. (\ref{12}) clearly
\begin{equation}\label{15}
\ell = i\eta \dot {\bar {\psi}} \psi + i{\mu _j}\nabla \bar \psi \psi + g(\bar \psi ).
\end{equation}
For the last time, we again compare Eqs. (\ref{10}) and (\ref{1}) to gain
\begin{equation}\label{16}
- \frac{{\partial \ell }}{{\partial \bar \psi }} = \frac{{\partial g(\bar \psi )}}{{\partial \bar \psi }} = {\eta ^\dag }m\psi ,
\end{equation}
that easily results
\begin{equation}\label{17}
g(\bar \psi ) = - {\eta ^\dag }m\bar \psi \psi ,
\end{equation}
Finally, inserting Eq. (\ref{17}) into Eq. (\ref{12}) leads to the Lagrange density
\begin{equation}\label{18}
\ell = i\eta \dot {\bar {\psi}} \psi + i{\mu _j}\nabla \bar \psi \psi - {\eta ^\dag }m\bar \psi \psi.
\end{equation}
\section{Time Evolution Relation}
Since Ajaib showed that Eq. (\ref{1}) can turn into Schrödinger equation, we thought that it has the
ability to have similar time evolution relation as Schrödinger equation does. To obtain such
relation first we should be multiplied Eq. (\ref{1}) by ${\eta ^\dag }$ from the left and using Eq. (\ref{4}), it reaches
\begin{equation}\label{19}
- i{\eta ^\dag }{\partial _z}\psi = i{\eta ^\dag }\eta {\partial _t}\psi ,
\end{equation}
as well as taking complex conjugate of Eq. (\ref{1}) then multiply it by $\eta$ from the left caused to write the final form as
\begin{equation}\label{20}
- i\eta {\partial _z}\psi = i\eta {\eta ^\dag }{\partial _t}\psi ,
\end{equation}
Summing Eq. (\ref{19}) and (\ref{20}) and with aim of Eq. (\ref{3}) we obtain
\begin{equation}\label{21}
i{\partial _t}\psi = \frac{{ - i}}{2}\left( {\eta + {\eta ^\dag }} \right){\partial _z}\psi,
\end{equation}
As is mentioned in Ref. \cite{4}, the parentheses is equal to $ - i\sqrt 2 {\gamma _2}$ , the $\gamma$ is the gamma matrices. Rewriting Eq. (\ref{21}) helps us to get to the time evolution relation as
\begin{equation}\label{22}
i{\partial _t}\psi = {H_{F - Sch}}\psi ,
\end{equation}
where
\begin{equation}\label{23}
{H_{F - Sch}} = \frac{{i{\gamma _2}}}{{\sqrt 2 }}{P_z} + V(z,t).
\end{equation}
So, Eq. (\ref{23}) by the potential as $V(z,t) = f(t)z$ in which $f(t)$ is an arbitrary function of time has the form of
\begin{equation}\label{24}
{H_{F - Sch}} = \frac{{i{\gamma _2}}}{{\sqrt 2 }}{P_z} + f(t)z.
\end{equation}
Now, we are in position to probe time evolution of Eq. (\ref{24}).
\section{Time Evolution and Lewis-Riesenfeld Approach}
In 1969, Lewis and Riesenfeld introduced an approach which is able to investigate time evolution of time dependent systems \cite{5}. In this manner, it is supposed that there exist a Hermitian and invariant operator such as $\hat{I}$ that has time evolution as
\begin{equation}\label{25}
\frac{{d\hat I(t)}}{{dt}} = \frac{{\partial \hat I(t)}}{{\partial t}} + \frac{1}{i}[\hat I(t),H(t)] = 0,
\end{equation}
Acting Eq. (\ref{25}) from the left on the $\left| \psi \right\rangle $ , after some calculations, it is possible to write
\begin{equation}\label{26}
i\hbar \frac{{\partial (\hat I\left| \Psi \right\rangle )}}{{\partial t}} = H(\hat I\left| \Psi \right\rangle ),
\end{equation}
Eq. (26) is showing that the action $\hat{I}(t)$ also satisfies the time evolution equation of the ket.
On the other hand, the ket can be expressed in terms of Eigen kets of $\hat{I}(t)$ as
\begin{equation}\label{27}
\left| \Psi \right\rangle = \sum\limits_{\lambda ,\kappa } {{c_{\lambda ,\kappa }}{e^{i{\alpha _{\lambda ,\kappa }}(t)}}} \left| {\lambda ,\kappa ;t} \right\rangle ,
\end{equation}
where $\lambda$ is the eigen value, $\kappa$ denotes the other quantum numbers, $c_{\lambda, \kappa}$ is constant and $\alpha_{\lambda, \kappa}(t)$ derived from the below equation
\begin{equation}\label{28}
\hbar \frac{{d{\alpha _{\lambda ,\kappa }}(t)}}{{dt}} = \left\langle {\lambda ,\kappa } \right|i\hbar \frac{\partial }{{\partial t}} - H(t)\left| {\lambda ,\kappa } \right\rangle.
\end{equation}
By this theory and following the methods in Refs \cite{5,6}, we derived the dynamical invariant for our considered system, Eq. (\ref{24}), as
\begin{equation}\label{29}
\hat I(t) = {\varepsilon _1}{P_z} + {\varepsilon _2}z + {\varepsilon _3}(t),
\end{equation}
with constants $\varepsilon_1$ and $\varepsilon_2$ and
\begin{equation}\label{30}
{\varepsilon _3}(t) = \frac{{ - i{\gamma _2}}}{{\sqrt 2 }}{\varepsilon _2}t + {\varepsilon _1}\int {f(t)dt + c', \qquad{\rm{(c' = constant)}}}.
\end{equation}
It is reasonable to consider that because $\hat{I}(t)$ produces a first order differential equation for its own eigen functions, we propose the general form of the eigen function as below
\begin{equation}\label{31}
{\phi _\lambda }(z,t) = \exp \left[{\mu _1}(t)z + {\mu _2}(t){z^2} \right],
\end{equation}
then we propose wave function according the eigen function in form of
\begin{equation}\label{32}
{\psi }(z,t) = K(t)\exp \left[{\mu _1}(t)z + {\mu _2}(t){z^2} \right].
\end{equation}
Using(\ref{32}) and (\ref{24}), it is easy to derived
\begin{equation}\label{33}
\begin{array}{l}
{\mu _1}(t) = - i \left( \sqrt 2 {\gamma _2}{\mu _2}t + \int {f(t)dt} \right)+ c'',\qquad(c'' = {\rm{constant}}), \\\\
{\mu _2} = {\rm{constant}},\\\\
K(t) = \exp \left[ \frac{{ - i{\gamma _2}}}{{\sqrt 2 }}\int {{\mu _1}(t)dt + c'''} \right],\qquad(c''' = {\rm{constant}}).
\end{array}
\end{equation}
It is instructive that as special cases we consider two cases as
\begin{itemize}
\item \textbf{Case A:}$f(t) = {d_1}{e^{i\omega t}},\,\,(\omega ,{d_1} = {\rm{constant}})$
This case represents periodic phenomena. So Eq.(\ref{33}) changes to
\begin{equation}\label{34}
\begin{array}{l}
{\mu _1}(t) = - i \left( \sqrt 2 {\gamma _2}{\mu _2}t + \frac{{{d_1}}}{{i\omega }}{e^{i\omega t}} \right) + {d_2},\,\,\,({d_2} = {\rm{constant}}), \\\\
K(t) = \exp \left[ {\frac{{ - i{\gamma _2}}}{{\sqrt 2 }}\left( { - i\frac{{{\gamma _2}{\mu _2}{t^2}}}{{\sqrt 2 }} - \frac{{{d_1}}}{{{\omega ^2}}}{e^{i\omega t}} + {d_2}t} \right) + {d_3}} \right],\,\,\,({d_3} = {\rm{constant}}).
\end{array}
\end{equation}
\item \textbf{Case B:}$f(t) = {g_1}{t^s},\,\,(s,{g_1} = {\rm{constant}})$
This case represents all linear or harmonics interactions. By the way Eq.(\ref{33}) turns into
\begin{equation}\label{35}
\begin{array}{l}
{\mu _1}(t) = - i \left( \sqrt 2 {\gamma _2}{\mu _2}t + \frac{{{g_1}}}{{s + 1}}{t^{s + 1}} \right) + {g_2},\,\,\,({g_2} = {\rm{constant}} ),\\\\
K(t) = \exp \left[ {\frac{{ - i{\gamma _2}}}{{\sqrt 2 }}\left( { - i\frac{{{\gamma _2}{\mu _2}{t^2}}}{{\sqrt 2 }} - \frac{{{g_1}}}{{(s + 1)(s + 2)}}{t^{s + 2}} + {g_2}t} \right) + {g_3}} \right],\,\,\,({g_3} = {\rm{constant}}).
\end{array}
\end{equation}
\end{itemize}
\section{Time Evolution Operator And Lie Algebra Method}
Other aspect of time evolution of a system is finding time evolution operator for the system. Having such an operator causes to have wave function in any desired time. According method of Ref. \cite{7} the Ansatz for this operator is as
\begin{equation}\label{36}
U(t) = \exp \left[ {{\xi _1}(t){P_z} + {\xi _2}(t)z + {\xi _3}(t)} \right],
\end{equation}
in which arbitrary functions ${\xi _i}(t),(i = 1,2,3)$ can be derived from
\begin{equation}\label{37}
i{\partial _t}U(t){U^{ - 1}}(t) = H.
\end{equation}
So, we have
\begin{equation}\label{38}
{\partial _t}U(t){U^{ - 1}}(t) = {\dot \xi _1}(t){P_z} + {\dot \xi _2}(t){e^{{\xi _1}(t){P_z}}}\left( z \right){e^{ - {\xi _1}(t){P_z}}} + {\dot \xi _1}(t).
\end{equation}
Following Baker–Hausdorff formula \cite{7}, we can write as
\begin{equation}\label{39}
\begin{array}{l l}
{e^{{\xi _1}(t){P_z}}}\left( z \right){e^{ - {\xi _1}(t){P_z}}} &= z + {\xi _1}(t)[{P_z},z],\\\\
&= z - {\xi _1}(t)i.
\end{array}
\end{equation}
Inserting (\ref{39}) into (\ref{38}) and some sorting result
\begin{equation}\label{40}
i\left( {{{\dot \xi }_1}(t){P_z} + {{\dot \xi }_2}(t)z + \frac{{{\xi _1}(t){{\dot \xi }_2}(t)}}{i} + {{\dot \xi }_3}(t)} \right) = \frac{{i{\gamma _2}}}{{\sqrt 2 }}{P_z} + f(t)z,
\end{equation}
equating each coefficient, we have
\begin{equation}\label{41}
\begin{array}{l}
{\xi _1}(t) = \frac{{{\gamma _2}}}{{\sqrt 2 }}t + {q_1},\\\\
{\xi _2}(t) = - i\int {f(t)dt} + {q_2},\\\\
{\xi _3}(t) = - \frac{{{\gamma _2}}}{{\sqrt 2 }}\int {f(t)dt} + {q_3},
\end{array}
\end{equation}
In which ${q_i},(i = 1,2,3)$ are constant and having the explicit form of $f(t)$ , the exact form of ${\xi _i},(i = 1,2,3)$ can be derived.
|
\section{Appendix}
\begin{lemma}[Eigenvector Estimation via Spectral Norm Matrix Approximation]\label{spectral_error_conversion}
Let $\bv{A}^\top\bv{A}$ have top eigenvector $1$, top eigenvector $v_1$ and eigenvalue gap $\mathrm{gap}$. Let $\bv{B}^\top \bv{B}$ be some matrix with $\norm{\bv{A}^\top\bv{A}-\bv{B}^\top\bv{B}}_2 \le O(\sqrt{\epsilon} \cdot \mathrm{gap})$. Let $x$ be the top eigenvector of $\bv{B}^\top \bv{B}$. Then:
\begin{align*}
|x^\top v_1 | \ge 1 - \epsilon ~.
\end{align*}
\end{lemma}
\begin{proof}
We can any unit vector $y$ as $y = c_1 v_1 + c_2 v_2$ where $v_2$ is the component of $x$ orthogonal to $v_1$ and $c_1^2 + c_2^2 = 1$. We know that
\begin{align*}
v_1^\top \bv{B}^\top \bv{B} v_1 &= v_1^\top \bv{A}^\top\bv{A}v_1-v_1^T(\bv{A}^\top\bv{A}-\bv{B}^\top\bv{B})v_1\\
1 -\sqrt{\epsilon} \mathrm{gap} &\le v_1^\top \bv{B}^\top \bv{B} v_1\le 1 +\sqrt{\epsilon} \mathrm{gap}
\end{align*}
Similarly we can compute:
\begin{align*}
v_2^\top \bv{B}^\top \bv{B} v_2 &= v_2^\top \bv{A}^\top\bv{A}v_2-v_2^T(\bv{A}^\top\bv{A}-\bv{B}^\top\bv{B})v_2\\
1 -\mathrm{gap} -\sqrt{\epsilon} \mathrm{gap} &\le v_2^\top \bv{B}^\top \bv{B} v_2 \le 1 -\mathrm{gap} +\sqrt{\epsilon} \mathrm{gap}.
\end{align*}
and
\begin{align*}
|v_1^\top \bv{B}^\top \bv{B} v_2| &= |v_1^\top \bv{A}^\top\bv{A}v_2-v_1^T(\bv{A}^\top\bv{A}-\bv{B}^\top\bv{B})v_2|\\
&\le \sqrt{\epsilon} \mathrm{gap}.
\end{align*}
We have $x^\top \bv{B}\bv{B}^\top x = c_1^2 (v_1^\top \bv{B}^\top \bv{B} v_1) + c_2^2 (v_2^\top \bv{B}^\top \bv{B} v_2) + 2c_1c_2 \cdot v_2^\top \bv{B}^\top \bv{B} v_1$.
We want to bound $c_1 \ge 1-\epsilon$ so $c_1^2 \ge 1- O(\epsilon)$. Since $x$ is the top eigenvector of $\bv{BB}^\top$ we have:
\begin{align*}
x^\top \bv{B}\bv{B}^\top x &\ge v_1^\top \bv{B}\bv{B}^\top v_1\\
c_2^2 (v_2^\top \bv{B}^\top \bv{B} v_2) + 2c_2 v_2^\top \bv{B}^\top \bv{B} v_1 &\ge (1-c_1^2)v_1^\top \bv{B}\bv{B}^\top v_1\\
2\sqrt{1-c_1^2} \sqrt{\epsilon}\mathrm{gap} &\ge (1-c_1^2)\left (v_1^\top \bv{B}\bv{B}^\top v_1 -v_2^\top \bv{B}\bv{B}^\top v_2 \right )\\
\frac{1}{\sqrt{1-c_1^2}} &\ge \frac{(1-2\sqrt{\epsilon})\mathrm{gap}}{2\sqrt{\epsilon}\mathrm{gap}}\\
\frac{1}{1-c_1^2} &\ge \frac{1-5\sqrt{\epsilon}}{4\epsilon}
\end{align*}
This means we need have $1-c_1^2 \le O(\epsilon)$ meaning $c_1^2 \ge 1-O(\epsilon)$ as desired.
\end{proof}
\begin{lemma}[Inverted Power Method progress in $\ell_2$ and $\bv{B}$ norms]\label{same_progress}
Let $x$ be a unit vector with $\inprod{x,v_1} \neq 0$ and let $\xtilde = \mvar{b}^{-1}w$, i.e. the power method update of $\mvar{b}^{-1}$ on $x$. Then, we have both:
\begin{align}\label{lb_progress}
\frac{\norm{\bv{P}_{v_1^{\perp}}\xtilde}_{\mvar{b}}}{\norm{\bv{P}_{v_1}\xtilde}_{\mvar{b}}} \le \frac{\lambda_2(\bv{B}^{-1})}{\lambda_1(\bv{B}^{-1})} \cdot \frac{\norm{\bv{P}_{v_1^{\perp}}x}_{\mvar{b}}}{\norm{\bv{P}_{v_1}x}_{\mvar{b}}}
\end{align}
and
\begin{align}\label{l2_progress}
\frac{\norm{\bv{P}_{v_1^{\perp}}\xtilde}_{2}}{\norm{\bv{P}_{v_1}\xtilde}_{2}} \le \frac{\lambda_2(\bv{B}^{-1})}{\lambda_1(\bv{B}^{-1})} \cdot \frac{\norm{\bv{P}_{v_1^{\perp}}x}_{2}}{\norm{\bv{P}_{v_1}x}_{2}}
\end{align}
\end{lemma}
\begin{proof}
\eqref{lb_progress} was already shown in Lemma \ref{thm:powermethod}. We show \eqref{l2_progress} similarly.
Writing $x$ in the eigenbasis of $\bv{B}^{-1}$, we have $x=\sum_i \alpha_i v_i$ and $\xtilde = \sum_i \alpha_i \lamiBinv{i}v_i$. Since $\inprod{x,v_1} \neq 0$, $\alpha_1 \neq 0$ and we have:
\begin{align*}
\frac{\norm{\bv{P}_{v_1^{\perp}}\xtilde}_{2}}{\norm{\bv{P}_{v_1}\xtilde}_{2}} = \frac{\sqrt{\sum_{i\geq 2} \alpha_i^2 \lambda^2_{i}(\bv{B}^{-1})}}{\sqrt{\alpha_1^2 \lambda^2_{1}(\bv{B}^{-1})}}
\leq \frac{\lamiBinv{2}}{\lamiBinv{1}} \cdot \frac{\sqrt{\sum_{i\geq 2} \alpha_i^2}}{\sqrt{\alpha_1^2}}
= \frac{\lamiBinv{2}}{\lamiBinv{1}} \cdot \frac{\norm{\bv{P}_{v_1^{\perp}}x}_{2}}{\norm{\bv{P}_{v_1}x}_{2}}.
\end{align*}
\end{proof}
\section{Algorithmic Framework}\label{framework}
Here we develop our robust shift-and-invert framework. In Section~\ref{sec:framework:basics} we provide a basic overview of the framework and in Section~\ref{sec:framework:potential} we introduce the potential function we use to measure progress of our algorithms. In Section~\ref{sec:framework:power-iteration} we show how to analyze the framework given access to an exact linear system solver and in Section~\ref{sec:framework:approximate-power} we strengthen this analysis to work with an inexact linear system solver. Finally, in Section~\ref{sec:framework:init} we discuss initializing the framework.
\subsection{Shifted-and-Inverted Power Method Basics}
\label{sec:framework:basics}
We let $\mvar{b}_{\lambda}\defeq\lambda\mI-\mvar{\Sigma}$ denote the shifted matrix that we will use in our implementation of the shifted-and-inverted power method. As discussed, in order for $\bv{B}_\lambda^{-1}$ to have a large eigenvalue gap, $\lambda$ should be set to $(1+c \cdot \mathrm{gap}) \lambda_1$ for some constant $c \geq 0$. Throughout this section we assume that we have a crude estimate of $\lambda_1$ and $\mathrm{gap}$ and fix $\lambda$ to be a value satisfying $\left (1 + \frac{\mathrm{gap}}{150}\right)\lambda_1 \le \lambda \le \left (1 + \frac{\mathrm{gap}}{100}\right)\lambda_1$. (See Section~\ref{parameter_free} for how we can compute such a $\lambda$). For the remainder of this section we work with such a fixed value of $\lambda$ and therefore for convenience denote $\bv{B}_\lambda$ as $\bv{B}$.
Note that $\lamiBinv{i} = \frac{1}{\lambda_i(\mvar{b})} = \frac{1}{\lambda-\lambda_i}$ and so $\frac{\lamiBinv{1}}{\lamiBinv{2}} = \frac{\lambda-\lambda_2}{\lambda-\lambda_1} \ge \frac{\mathrm{gap}}{\mathrm{gap}/100} = 100.$ This large gap will ensure that, assuming the ability to apply $\bv{B}^{-1}$, the power method will converge very quickly. In the remainder of this section we develop our error analysis for the shifted-and-inverted power method which demonstrates that approximate application of $\bv{B}^{-1}$ in each iteration in fact suffices.
\subsection{Potential Function}
\label{sec:framework:potential}
Our analysis of the power method focuses on the objective of maximizing
the Rayleigh quotient, $x^\top \mvar{\Sigma} x$ for a unit vector $x$. Note that as the following lemma shows, this has a direct correspondence to the error in maximizing $|v_1^\top x|$:
\begin{lemma}[Bounding Eigenvector Error by Rayleigh Quotient]
\label{lem:ray_to_evec}
For a unit vector $x$ let $\epsilon = \lambda_1 - x^\top \bv{\Sigma} x$. If $\epsilon \leq \lambda_1 \cdot \mathrm{gap}$ then
\[
\left | v_1^\top x \right |
\geq \sqrt{1 - \frac{\epsilon}{\lambda_1 \cdot \mathrm{gap}}}.
\]
\end{lemma}
\begin{proof}
Among all unit vectors $x$ such that $\epsilon = \lambda_1 - x^\top \mvar{\Sigma} x$, a minimizer of $\left | v_1^\top x\right |$ has the form $x = (\sqrt{1 - \delta^2}) v_1 + \delta v_2$ for some $\delta$. We have
\begin{align*}
\epsilon
&=
\lambda_1 - x^\top \bv \Sigma x
=
\lambda_1 - \lambda_1 (1 - \delta^2) - \lambda_2 \delta^2
=
(\lambda_1 - \lambda_2) \delta^2.
\end{align*}
Therefore by direct computation,
\begin{align*}
\left | v_1^\top x \right | = \sqrt{1 - \delta^2} = \sqrt{1 - \frac{\epsilon}{\lambda_1-\lambda_2}} = \sqrt{1 - \frac{\epsilon}{\lambda_1 \cdot \mathrm{gap}}} ~ .
\end{align*}
\end{proof}
\iffalse
\sk{I tweaked the below to make it more intuitive. Please check.}
write $x$ in the eigenbasis as $x=\sum_i \alpha_i v_i$,
and, in order to track how close $x$ is to $v_1$, let us
\fi
In order to track the progress of our algorithm we use a more
complex potential function than just the Rayleigh quotient error, $\lambda_1 - x^\top \bv{\Sigma}
x$. Our potential function $G$ is defined for $x \neq 0$ by
\begin{align*}
G(x) \defeq
\frac{\norm{\bv{P}_{v_1^{\perp}}x}_{\mvar{b}}}
{\norm{\bv{P}_{v_1}x}_{\mvar{b}}}
\end{align*}
where $\bv{P}_{v_1}$ and $\bv{P}_{v_1^{\perp}}$ are the
projections onto $v_1$ and the subspace
orthogonal to $v_1$ respectively. Equivalently, we have that:
\begin{align}\label{gequiv}
G(x) =
\frac{\sqrt{\norm{x}_{\mvar{b}}^{2}-\left(v_1^{\top}\mvar{b}^{1/2}x\right)^{2}}}{\abs{v_1^\top
\mvar{b}^{1/2}x}} = \frac{\sqrt{\sum_{i\geq 2} \frac{\alpha_i^2}{ \lambda_{i}(\bv{B}^{-1})}}}{\sqrt{\frac{\alpha_1^2}{ \lambda_{1}(\bv{B}^{-1})}}}.
\end{align}
where $\alpha_i = v_i^\top x$.
When the Rayleigh quotient error $\epsilon = \lambda_1 - x^\top\mvar{\Sigma}
x$ of $x$ is small, we can show a strong relation between $\epsilon$ and $G(x)$. We prove this in two parts. We first give a technical lemma, Lemma~\ref{lem:potfunc1}, that we will use several times for bounding the numerator of $G$. We then prove the connection in Lemma~\ref{lem:rayquot-potential}.
\begin{lemma} \label{lem:potfunc1} For a unit vector $x$ and $\epsilon = \lambda_1 - x^\top \mvar{\Sigma} x$ if $\epsilon \leq \lambda_1 \cdot \mathrm{gap}$ then
\[
\epsilon \leq
x^\top \mvar{b} x - (v_1^\top \mvar{b} x) (v_1^\top x)
\leq \epsilon \left (1 +
\frac{\lambda - \lambda_1}{\lambda_1 \cdot \mathrm{gap}}\right).
\]
\end{lemma}
\begin{proof}
Since $\mvar{b} = \lambda \mI - \mvar{\Sigma}$ and since $v_1$ is an eigenvector of
$\mvar{\Sigma}$ with eigenvalue $\lambda_1$ we have
\begin{align*}
x^\top \mvar{b} x - (v_1^\top \mvar{b} x) (v_1^\top x)
&=
\lambda \norm{x}_2^2 - x^\top \mvar{\Sigma} x
- (\lambda v_1^\top x - v_1^\top \mvar{\Sigma} x) (v_1^\top x)
\\
&=
\lambda - \lambda_1 + \epsilon -
(\lambda v_1^\top x - \lambda_1 v_1^\top x) (v_1^\top x)
\\
&=
(\lambda - \lambda_1) \left(1 - (v_1^\top x)^2\right) + \epsilon.
\end{align*}
Now by Lemma~\ref{lem:ray_to_evec} we know that $| v_1^\top x |
\geq \sqrt{1 - \frac{\epsilon }{\lambda_1 \cdot \mathrm{gap}}}$, giving us the upper bound. Furthermore, since trivially $\left |v_1^\top x\right | \leq 1$ and $\lambda - \lambda_1 > 0$, we have the lower bound.
\end{proof}
\begin{lemma}[Potential Function to Rayleigh Quotient Error Conversion]\label{lem:rayquot-potential}
For a unit vector $x$ and $\epsilon = \lambda_1 - x^\top \mvar{\Sigma} x$ if $\epsilon \leq \frac{1}{2}\lambda_1 \cdot \mathrm{gap}$, we have:
\begin{align*}
\frac{\epsilon}{\lambda-\lambda_1} \leq G(x)^2 \leq \left(1+\frac{\lambda-\lambda_1}{\lambda_1 \cdot \mathrm{gap}}\right) \left(1+\frac{2\epsilon}{\lambda_1 \cdot \mathrm{gap}}\right)\frac{\epsilon}{\lambda-\lambda_1}.
\end{align*}
\end{lemma}
\begin{proof}
Since $v_1$ is an eigenvector of $\mvar{b}$, we can write $G(x)^2 = \frac{x^\top \mvar{b} x - (v_1^\top \mvar{b} x) (v_1^\top x)}{(v_1^\top \mvar{b} x) (v_1^\top x)}$. Lemmas~\ref{lem:ray_to_evec} and~\ref{lem:potfunc1} then give us:
\begin{align*}
\frac{\epsilon}{\lambda-\lambda_1} \leq G(x)^2 \leq \left(1+\frac{\lambda-\lambda_1}{\lambda_1 \cdot \mathrm{gap}}\right)\frac{\epsilon}{\left(\lambda-\lambda_1\right) \left(1-\frac{\epsilon}{\lambda_1 \cdot \mathrm{gap}}\right)}.
\end{align*}
Since $\epsilon \leq \frac{1}{2}\lambda_1 \cdot \mathrm{gap}$, we have $\frac{1}{1-\frac{\epsilon}{\lambda_1 \cdot \mathrm{gap}}} \leq 1+\frac{2\epsilon}{\lambda_1 \cdot \mathrm{gap}}$. This proves the lemma.
\end{proof}
\subsection{Power Iteration}
\label{sec:framework:power-iteration}
Here we show that the shifted-and-inverted power iteration in fact makes progress with respect to our objective function given an exact linear system solver for $\mvar{b}$. Formally, we show that applying $\mvar{b}^{-1}$ to a vector $x$ decreases the potential function $G(x)$ geometrically.
\begin{theorem}\label{thm:powermethod}
Let $x$ be a unit vector with $\inprod{x,v_1} \neq 0$ and let $\xtilde = \mvar{b}^{-1}x$, i.e. the power method update of $\mvar{b}^{-1}$ on $x$. Then, under our assumption on $\lambda$, we have:
\begin{align*}
G(\xtilde) \leq \frac{\lamiBinv{2}}{\lamiBinv{1}} G(x)
\leq \frac{1}{100} G(x).
\end{align*}
\end{theorem}
Note that $\xtilde$ may no longer be a unit vector. However, $G(\xtilde, v_1) = G(c \xtilde, v_1)$ for any scaling parameter $c$, so the theorem also holds for $\xtilde$ scaled to have unit norm.
\begin{proof}
Writing $x$ in the eigenbasis, we have $x=\sum_i \alpha_i v_i$ and $\xtilde = \sum_i \alpha_i \lamiBinv{i}v_i$. Since $\inprod{x,v_1} \neq 0$, $\alpha_1 \neq 0$ and by the equivalent formulation of $G(x)$ given in \eqref{gequiv}:
\begin{align*}
G(\xtilde) = \frac{\sqrt{\sum_{i\geq 2} \alpha_i^2 \lambda_{i}(\bv{B}^{-1})}}{\sqrt{\alpha_1^2 \lambda_{1}(\bv{B}^{-1})}}
\leq \frac{\lamiBinv{2}}{\lamiBinv{1}} \cdot \frac{\sqrt{\sum_{i\geq 2} \frac{\alpha_i^2}{\lambda_{i}(\bv{B}^{-1})}}}{\sqrt{\frac{\alpha_1^2}{\lambda_1(\bv{B}^{-1})}}}
= \frac{\lamiBinv{2}}{\lamiBinv{1}} \cdot G(x) ~.
\end{align*}
Recalling that $\frac{\lamiBinv{1}}{\lamiBinv{2}} = \frac{\lambda-\lambda_2}{\lambda-\lambda_1} \ge \frac{\mathrm{gap}}{\mathrm{gap}/100} = 100$ yields the result.
\end{proof}
The challenge in using the above theorem, and any traditional analysis of the shifted-and-inverted power method, is that we don't actually have access to $\mvar{b}^{-1}$. In the next section we show that the shifted-and-inverted power method is robust -- we still make progress on our objective function even if we only approximate $\mvar{b}^{-1}x$ using a fast linear system solver.
\subsection{Approximate Power Iteration}
\label{sec:framework:approximate-power}
We are now ready to prove our main result. We show that each iteration of the shifted-and-inverted power method makes constant factor expected progress on our potential function assuming we:
\begin{enumerate}
\item Start with a sufficiently good $x$ and an approximation of $\lambda_1$
\item Can apply $\bv{B}^{-1}$ approximately using a system solver such that the function error (i.e. distance to $\bv{B}^{-1} x$ in the $\bv{B}$ norm) is sufficiently small \emph{in expectation}.
\item Can estimate Rayleigh quotients over $\bv{\Sigma}$ well enough to only accept updates that do not hurt progress on the objective function too much.
\end{enumerate}
This third assumption is necessary since the second assumption is quite weak. An expected progress bound on the linear system solver allows, for example, the solver to occasionally return a solution that is entirely orthogonal to $v_1$, causing us to make unbounded backwards progress on our potential function. The third assumption allows us to reject possibly harmful updates and ensure that we still make progress in expectation. In the offline setting, we can access $\bv{A}$ and are able to compute Rayleigh quotients exactly in time $\nnz(\bv{A})$ time. However, we only assume the ability to estimate quotients since in the online setting we only have access to $\mvar{\Sigma}$ through samples from $\mathcal{D}$.
Our general theorem for the approximate power iteration, Theorem~\ref{thm:powermethod-perturb}, assumes that we can solve linear systems to some absolute accuracy in expectation. This is not completely standard. Typically, system solver analysis assumes an initial approximation to $\bv{B}^{-1} x$ and then shows a relative progress bound -- that the quality
of the initial approximation is improved geometrically in each iteration of the algorithm. In Corollary~\ref{cor:constant_factor_corollary} we show how to find a coarse initial approximation to $\bv{B}^{-1} x$, in fact just approximating $\bv{B}^{-1}$ with $\frac{1}{x^\top \mvar{b} x} x$. Using this approximation, we show that Theorem~\ref{thm:powermethod-perturb} actually implies that traditional system solver relative progress bounds suffice.
Note that in both claims we measure error of the linear system solver using $\norm{\cdot}_\mvar{b}$. This is a natural norm in which geometric convergence is shown for many linear system solvers and directly corresponds to the function error of minimizing $f(w) = \frac{1}{2} w^\top \mvar{b} w - w^\top x$ to compute $\mvar{b}^{-1} x$.
\begin{theorem}[Approximate Shifted-and-Inverted Power Iteration -- Warm Start]\label{thm:powermethod-perturb}
Let $x=\sum_i \alpha_i v_i$ be a unit vector such that $G(x) \leq \frac{1}{\sqrt{10}}$. Suppose we know some shift parameter $\lambda$ with $\left (1 + \frac{\mathrm{gap}}{150} \right ) \lambda_1< \lambda \le \left (1 + \frac{\mathrm{gap}}{100} \right ) \lambda_1$ and an estimate $\lambdah_1$ of $\lambda_1$ such that
$
\frac{10}{11} \left(\lambda-\lambda_1\right) \leq \lambda-\lambdah_1 \leq \lambda-\lambda_1
$.
Furthermore, suppose we have a subroutine $\mathrm{solve}(\cdot)$ such that on any input $x$
\begin{align*}
\expec{\norm{\solve{x} - \mvar{b}^{-1}x}_{\mvar{b}}} \leq \frac{c_1}{1000}
\sqrt{\lambda_1(\bv{B}^{-1})},
\end{align*}
for some $c_1 < 1$,
and a subroutine $\rayquoth{\cdot}$ that on any input $x \neq 0$
\begin{align*}
\abs{\rayquoth{x} - \mathrm{quot}(x)}
\leq \frac{1}{30}\left(\lambda-\lambda_1\right) \; \text{ for all nonzero } \; x \in \mathbb{R}^d.
\end{align*}
where $\mathrm{quot}(x) \defeq \frac{x^\top \mvar{\Sigma} x}{x^\top x}$.
Then the following update procedure:
\begin{align*}
\text {Set }\xhat = \solve{x},
\end{align*}
\begin{align*}
\text {Set } \xtilde = \left\{
\begin{array}{cc}
\xhat& \mbox{ if } \left\{
\begin{array}{c}
\rayquoth{\xhat} \geq \lambdah_1 - \left(\lambda-\lambdah_1\right)/6 \mbox{ and } \\
\norm{\xhat}_2 \geq \frac{2}{3}\frac{1}{\lambda-\lambdah_1}
\end{array} \right. \\
x & \mbox{ otherwise,}
\end{array}
\right.
\end{align*}
satisfies the following:
\begin{itemize}
\item $G(\xtilde)\leq \frac{1}{\sqrt{10}}$ and
\item $\expec{G(\xtilde)}\leq \frac{3}{25} G(x) + \frac{c_1}{500}$.
\end{itemize}
\end{theorem}
That is, not only do we decrease our potential function by a constant factor in expectation, but we are guaranteed that the potential function will never increase beyond $1/\sqrt{10}$.
\begin{proof}
The first claim follows directly from our choice of $\xtilde$ from $x$ and $\xhat$.
If $\xtilde = x$, it holds trivially by our assumption that $G(x) \leq \frac{1}{\sqrt{10}}$. Otherwise, $\xtilde = \xhat$ and we know that
\begin{align*}
\lambda_1 - \mathrm{quot}\left(\xhat\right)
&\leq \lambdah_1 - \mathrm{quot}\left(\xhat\right) \leq \lambdah_1 - \rayquoth{\xhat} + \abs{\rayquoth{\xhat} - \mathrm{quot}\left(\xhat\right)}
\\
&\leq \frac{\lambda-\lambdah_1}{6} + \frac{\lambda-\lambda_1}{30}
\leq \frac{\lambda-\lambda_1}{5} \le \frac{\lambda_1 \cdot \mathrm{gap}}{500} ~.
\end{align*}
The claim then follows from Lemma~\ref{lem:rayquot-potential} as
\begin{align*}
G(\xhat)^2
&\leq
\left(1 + \frac{\lambda - \lambda_1}{\lambda_1 \cdot \mathrm{gap}}\right)
\left(1 + \frac{2\left(\lambda_1 - \mathrm{quot}\left(\xhat\right)\right)}{\lambda_1 \cdot \mathrm{gap}}\right)
\frac{\lambda_1 - \mathrm{quot}\left(\xhat\right)}{\lambda-\lambda_1} \\
&\leq \frac{101}{100}\cdot \frac{251}{250} \cdot
\frac{\left(\frac{\lambda_1 \cdot \mathrm{gap}}{500}\right)}
{\left(\frac{\lambda_1 \cdot \mathrm{gap}}{150}\right)}
\le \frac{1}{\sqrt{10}} ~.
\end{align*}
All that remains is to show the second claim, that $\expec{G(\xtilde)}\leq \frac{3}{25} G(x) + \frac{4c_1}{1000}$. Let $\mcalF$ denote the event that we accept our iteration and set $x= \xhat = \solve{x}$. That is:
\begin{align*}
\mcalF &\defeq \set{\rayquoth{\xhat} \geq \lambdah_1 -
\frac{\lambda - \lambdah_1}{6}} \cup \set{\norm{\xhat}_2 \geq \frac{2}{3}\frac{1}{\lambda-\lambdah_1}}.
\end{align*}
Using our bounds on $\lambdah_1$ and $\rayquoth{\cdot}$, we know that $\rayquoth{x} \leq \mathrm{quot}(x) + (\lambda - \lambda_1) / 30$ and $\lambda - \lambdah_1 \leq \lambda - \lambda_1$.
Therefore, since $-1/6 - 1/30 \geq -1/2$ we have
\begin{align*}
\mathcal{F} &\subseteq \set{\mathrm{quot}\left(\xhat\right) \geq \lambda_1 - \left(\lambda-\lambda_1\right)/2} \cup \set{\norm{\xhat}_2 \geq \frac{2}{3}\frac{1}{\lambda-\lambda_1}},
\end{align*}
We will complete the proof in two steps. First we let $\xi \defeq \xhat - \mvar{b}^{-1}x$ and show that assuming $\mcalF$ is true then $G(\xhat)$ and
$\norm{\xi}_{\mvar{b}}$ are linearly related, i.e. expected error bounds on $\norm{\xi}_{\mvar{b}}$ correspond to expected error bounds on $G(\xhat)$. Second, we bound the probability that $\mcalF$ does not occur and bound error incurred in this case. Combining these yields the result.
To show the linear relationship in the case where $\mcalF$ is true, first note Lemma~\ref{lem:ray_to_evec} shows that in this case $\abs{v_1^\top \frac{\xhat}{\norm{\xhat}_2}} \geq \sqrt{1-\frac{\lambda_1-\mathrm{quot}(\xhat)}{\lambda_1 \cdot \mathrm{gap}}} \geq \frac{3}{4}$. Consequently,
\[
\norm{\bv{P}_{v_1}\xhat}_{\mvar{b}}
= \abs{v_1^\top\xhat}\sqrt{\lambda-\lambda_1}
= \abs{v_1^\top \frac{\xhat}{\norm{\xhat}_2}}\cdot \norm{\xhat}\sqrt{\lambda-\lambda_1}
\geq \frac{3}{4}\cdot \frac{2}{3} \frac{1}{\sqrt{\lambda-\lambda_1}}
=
\frac{\sqrt{\lambda_1(\bv{B}^{-1})}}{2} ~.
\]
However,
\[
\norm{\bv{P}_{v_1^{\perp}}\xhat}_{\mvar{b}}
\leq \norm{\bv{P}_{v_1^{\perp}}\mvar{b}^{-1} x}_{\mvar{b}} + \norm{\bv{P}_{v_1^{\perp}}\xi}_\mvar{b}
\leq \norm{\bv{P}_{v_1^{\perp}}\mvar{b}^{-1} x}_{\mvar{b}} + \norm{\xi}_{\mvar{b}}
\]
and by Theorem~\ref{thm:powermethod} and the definition of $G$ we have
\[
\norm{\bv{P}_{v_1^{\perp}}\mvar{b}^{-1} x}_{\mvar{b}}
= \norm{\bv{P}_{v_1}\mvar{b}^{-1} x}_{\mvar{b}} \cdot G(\mvar{b}^{-1} x)
\leq \left(\abs{\inprod{x,v_1}} \sqrt{\lambda_1(\bv{B}^{-1})} \right)
\cdot \frac{G(x)}{100} ~.
\]
Taking expectations, using that $\abs{\inprod{x,v_1}} \leq 1$, and combining these three inequalities yields
\begin{equation}
\expec{G(\xhat) \middle\vert \mcalF}
=
\expec{\frac{\norm{\bv{P}_{v_1^{\perp}}\mvar{b}^{-1} x}_{\mvar{b}}}{\norm{\bv{P}_{v_1}\mvar{b}^{-1} x}_{\mvar{b}}} \middle\vert \mcalF}
\leq \frac{G(x)}{50}
+ 2 \frac{ \expec{\norm{\xi}_{\mvar{b}}
\middle\vert \mcalF}}{\sqrt{\lambda_1(\bv{B}^{-1})}}
\label{eqn:expec-potential}
\end{equation}
So, conditioning on making an update and changing $x$ (i.e. $\mathcal{F}$ occurring), we see that our potential function changes exactly as in the exact case (Theorem \ref{thm:powermethod}) with additional additive error due to our inexact linear system solve.
Next we upper bound $\prob{\mcalF}$ and use it to compute $\expec{\norm{\xi}_{\mvar{b}}\middle\vert \mcalF}$. We will show that
$$\mcalG \defeq \set{\norm{\xi}_{\mvar{b}} \leq \frac{1}{100}\cdot \sqrt{\lamiBinv{1}} } \subseteq \mcalF$$ which then implies by Markov inequality that
\begin{align}
\prob{\mcalF} &\geq \prob{\norm{\xi}_{\mvar{b}} \leq \frac{1}{100}\cdot \sqrt{\lamiBinv{1}} }
\geq 1- \frac{\expec{\norm{\xi}_{\mvar{b}}}}{\frac{1}{100}\cdot \sqrt{\lamiBinv{1}} } \geq \frac{9}{10},\label{eqn:PF}
\end{align}
where we used the fact that $\mathbb{E} [\norm{\xi}_{\mvar{b}}] \le \frac{c_1}{1000} \sqrt{\lambda_1(\bv{B}^{-1})}$ for some $c_1 < 1$.
Let us now show that $\mcalG \subseteq \mcalF$. Suppose $\mcalG$ is occurs. We can bound $\norm{\xhat}_2$ as follows:
\begin{align}
\norm{\xhat}_2
& \geq \norm{\mvar{b}^{-1}x}_2 - \norm{\xi}_2
\geq \norm{\mvar{b}^{-1}x}-\sqrt{\lamiBinv{1}}\norm{\xi}_{\mvar{b}} \nonumber \\
&\geq \abs{\alpha_1}\lamiBinv{1} - \frac{1}{100}\cdot \lamiBinv{1} \nonumber \\
&= \frac{1}{\lambda-\lambda_1} \left(\abs{\alpha_1} - \frac{1}{100}\right)
\geq \frac{3}{4}\frac{1}{\lambda-\lambda_1},\label{eqn:f2}
\end{align}
where we use Lemmas~\ref{lem:potfunc1} and~\ref{lem:rayquot-potential} to conclude that $\abs{\alpha_1} \geq \sqrt{1-\frac{1}{10}}$. We now turn to showing the Rayleigh quotient condition required by $\mcalF$. In order to do this, we first bound $\xhat^\top \mvar{b} \xhat - \left(v_1^\top \mvar{b} \xhat\right)\left(v_1^\top \xhat\right)$ and then use Lemma~\ref{lem:potfunc1}. We have:
\begin{align*}
\sqrt{\xhat^\top \mvar{b} \xhat - \left(v_1^\top \mvar{b} \xhat\right)\left(v_1^\top \xhat\right)}
&= \norm{\bv{P}_{v_1^{\perp}}\xhat}_\mvar{b}
\leq \norm{\bv{P}_{v_1^{\perp}}\mvar{b}^{-1}x}_\mvar{b} + \norm{\bv{P}_{v_1^{\perp}} \xi}_\mvar{b} \\
&\leq \sqrt{\sum_{i\geq 2} \alpha_i^2 \lamiBinv{i}} + \frac{1}{100}\cdot \sqrt{\lamiBinv{1}} \\
&\leq \sqrt{\lamiBinv{2}} + \frac{1}{100}\cdot \sqrt{\lamiBinv{1}}
\leq \frac{1}{9}\sqrt{\lambda-\lambda_1},
\end{align*}
where we used the fact that $\lamiBinv{2}\leq \frac{1}{100}\lamiBinv{1}$ since $\lambda \le \lambda_1 + \frac{\mathrm{gap}}{100}$ in the last step. Now, using Lemma~\ref{lem:potfunc1} and the bound on $\norm{\xhat}_2$, we conclude that
\begin{align}
\lambdah_1 - \rayquoth{\xhat}
&\leq \lambda_1 - \mathrm{quot}\left(\xhat\right) + \abs{\mathrm{quot}\left(\xhat\right)-\rayquoth{\xhat}} + \lambdah_1 - \lambda_1\nonumber \\
&\leq \frac{\xhat^\top \mvar{b} \xhat - \left(v_1^\top \mvar{b} \xhat\right)\left(v_1^\top \xhat\right)}{\norm{\xhat}^2_2} + \frac{\lambda-\lambda_1}{30} + \frac{\lambda-\lambda_1}{11} \nonumber \\
&\leq \frac{1}{81\left(\lambda-\lambda_1\right)} \cdot \frac{16}{9}\left(\lambda-\lambda_1\right)^2 + \frac{\lambda-\lambda_1}{8} \nonumber \\
&\leq \left(\lambda-\lambda_1\right)/6
\leq \left(\lambda-\lambdah_1\right)/4. \label{eqn:f1}
\end{align}
Combining~\eqref{eqn:f2} and~\eqref{eqn:f1} shows that $\mcalG \subseteq \mcalF$ there by proving~\eqref{eqn:PF}.
Using this and the fact that $\norm{\cdot}_{\mvar{b}} \geq 0$ we can upper bound $\expec{\norm{\xi}_{\mvar{b}}\middle\vert \mcalF}$ as follows:
\begin{align*}
\expec{\norm{\xi}_{\mvar{b}}\middle\vert \mcalF} \leq \frac{1}{\prob{\mcalF}} \cdot \expec{\norm{\xi}_{\mvar{b}}}
\leq \frac{c_1}{900}\cdot \sqrt{\lambda_1(\bv{B}^{-1})}
\end{align*}
Plugging this into~\eqref{eqn:expec-potential}, we obtain:
\[
\expec{G(\xhat)\middle\vert \mcalF}
\leq
\frac{1}{50} G(x) + \frac{2\expec{\norm{\xi}_{\mvar{b}}\middle\vert \mcalF}}{\sqrt{\lambda_1(\bv{B}^{-1})}}
\leq \frac{1}{50}\cdot G(x) + \frac{2c_1}{900}.
\]
We can now finally bound $\expec{G(\xtilde)}$ as follows:
\begin{align*}
\expec{G(\xtilde)} &= \prob{\mcalF} \cdot \expec{G(\xhat)\middle\vert \mcalF} + \left(1-\prob{\mcalF}\right) G(x) \\
&\leq \frac{9}{10} \left ( \frac{1}{50}\cdot G(x) + \frac{2c_1}{900}\right) + \frac{1}{10} G(x)
= \frac{3}{25} G(x) + \frac{2c_1}{1000}.
\end{align*}
This proves the theorem.
\end{proof}
\begin{corollary} [Relative Error Linear System Solvers]
\label{cor:constant_factor_corollary}
For any unit vector $x$, we have:
\begin{equation}
\label{eq:constant_factor_corollary}
\norm{\frac{1}{x^\top \mvar{b} x} x - \mvar{b}^{-1}x}_{\mvar{b}} \leq \alpha_1\sqrt{\lambda_1(\bv{B}^{-1})} \cdot G(x) = \lamiBinv{1} \sqrt{\sum_{i\geq 2} \frac{\alpha_i^2}{\lamiBinv{i}}},
\end{equation}
where $x=\sum_i \alpha_i v_i$ is the decomposition of $x$ along $v_i$.
Therefore, instantiating Theorem \ref{thm:powermethod-perturb} with $c_1 = \alpha_1 G(x)$ gives $\mathbb{E} [G(\xtilde)] \le \frac{4}{25} G(x)$ as long as:
\begin{align*}
\expec{\norm{\solve{x}-\mvar{b}^{-1}x}_{\mvar{b}}} \le \frac{1}{1000} \norm{\frac{1}{\lambda-x^\top \bv{\Sigma}x} x - \mvar{b}^{-1}x}_{\mvar{b}}.
\end{align*}
\end{corollary}
\begin{proof}
Since $\mvar{b}$ is PSD we see that if we let $f(w) = \frac{1}{2} w^\top \mvar{b} w - w^\top x$, then the minimizer is $\mvar{b}^{-1} x$. Furthermore note that $\frac{1}{x^\top \mvar{b} x} = \argmin_\beta
f(\beta x)$ and therefore
\begin{align*}
\norm{\frac{1}{x^\top \mvar{b} x} x - \mvar{b}^{-1}x}_{\mvar{b}}^2
&= x^\top \mvar{b}^{-1} x - \frac{1}{x^\top \mvar{b} x}
= 2 \left [f\left(\frac{x}{x^\top \mvar{b} x}\right) - f(\mvar{b}^{-1} x) \right] \\
= & 2 \left [\min_\beta f(\beta x) - f(\mvar{b}^{-1} x) \right]
\le 2 \left[f(\lamiBinv{1}x) - f(\mvar{b}^{-1} x) \right] \\
= &\lamiBinv{1}^2 x^\top \mvar{b} x - 2 \lamiBinv{1} x^\top x + x^\top \mvar{b}^{-1} x \\
= & \sum_{i=1}^d \abs{v_i^\top \mvar{b}^{\frac{1}{2}} x}^2(\lamiBinv{1}-\lamiBinv{i})^2
\le \lamiBinv{1}^2\sum_{i\ge 2} \abs{v_i^\top \mvar{b}^{\frac{1}{2}} x}^2 \\
= & \lamiBinv{1}^2\sum_{i\geq 2} \frac{\alpha_i^2}{\lamiBinv{i}},
\end{align*}
which proves \eqref{eq:constant_factor_corollary}.
Consequently
\begin{align*}
\frac{c_1}{1000}\sqrt{\lambda_1(\bv{B}^{-1})} = \frac{1}{1000} \alpha_1 G(x)\sqrt{\lambda_1(\bv{B}^{-1})} \ge \frac{1}{1000} \norm{\frac{1}{x^\top \mvar{b} x} x - \mvar{b}^{-1}x}_{\mvar{b}}
\end{align*}
which with Theorem~\ref{thm:powermethod-perturb} then completes the proof.
\end{proof}
\subsection{Initialization}
\label{sec:framework:init}
Theorem \ref{thm:powermethod-perturb} and Corollary \ref{cor:constant_factor_corollary} show that, given a good enough approximation to $v_1$, we can rapidly refine this approximation by applying the shifted-and-inverted power method. In this section, we cover initialization. That is, how to obtain a good enough approximation to apply these results.
We first give a simple bound on the quality of a randomly chosen start vector $x_0$.
\begin{lemma}[Random Initialization Quality]\label{lem:init-random}
Suppose $x \sim \mathcal{N}(0, \bv{I})$, and we initialize $x_0$ as $\frac{x}{\norm{x}_2}$, then with probability greater than $1- O\left (\frac{1}{d^{10}} \right )$, we have:
\begin{equation*}
G(x_0) \le \sqrt{\kappa(\mvar{b}^{-1})} d^{10.5} \le 15 \frac{1}{\sqrt{\mathrm{gap}}} \cdot d^{10.5}
\end{equation*}
where $\kappa(\mvar{b}^{-1}) = \lambda_1(\mvar{b}^{-1}) / \lambda_d(\mvar{b}^{-1)}$.
\end{lemma}
\begin{proof}
\begin{align*}
G(x_0) =& G(x) = \frac{\norm{\bv{P}_{v_1^{\perp}}x}_{\mvar{b}}} {\norm{\bv{P}_{v_1}x}_{\mvar{b}}} = \frac{\sqrt{\norm{x}_{\mvar{b}}^{2}-\left(v_1^\top\mvar{b}^{1/2}x\right)^{2}}}{\abs{v_1^\top \mvar{b}^{1/2}x}}
=\frac{\sqrt{\sum_{i\geq2} \frac{(v_i^\top x)^2}{\lamiBinv{i}}}}{\sqrt{\frac{(v_1^\top x)^2}{\lamiBinv{1}}}}, \\
\le& \sqrt{\kappa(\mvar{b}^{-1})} \cdot \frac{\sqrt{\sum_{i\geq2}(v_i^\top x)^2}}{\abs{v_1^\top x}}
\end{align*}
Since $\{v_i^\top x\}_i$ are independent standard normal Gaussian variables.
By standard concentration arguments, with probability greater than $1- e^{-\Omega(d)}$, we have $\sqrt{\sum_{i\geq2}(v_i^\top x)^2} = O(\sqrt{d})$. Meanwhile, $v_1^\top x$ is just a one-dimensional standard Gaussian. It is easy to show $\Pr \left(\abs{v_1^\top x} \le \frac{1}{d^{10}} \right) = O\left (\frac{1}{d^{10}} \right)$, which finishes the proof.
\end{proof}
We now show that we can rapidly decrease our initial error to obtain the required $G(x) \le \frac{1}{\sqrt{10}}$ bound for Theorem \ref{thm:powermethod-perturb}.
\begin{theorem}[Approximate Shifted-and-Inverted Power Method -- Burn-In]\label{thm:init-offline}
Suppose we initialize $x_0$ as in Lemma \ref{lem:init-random} and suppose we have access to a subroutine $\solve{\cdot}$ such that
\begin{align*}
\expec{\norm{\solve{x}-\mvar{b}^{-1}x}_{\mvar{b}}} \leq \frac{1}{3000 \kappa(\mvar{b}^{-1})d^{21}} \cdot \norm{\frac{1}{\lambda-x^\top \bv \Sigma x} x - \mvar{b}^{-1}x}_{\mvar{b}}
\end{align*}
where $\kappa(\mvar{b}^{-1}) = \lambda_1(\mvar{b}^{-1}) / \lambda_d(\mvar{b}^{-1)}$.
Then the following procedure,
\begin{align*}
x_{t} = \solve{x_{t-1}}/\norm{\solve{x_{t-1}}}_2
\end{align*}
after $T = O\left(\log d + \log \kappa(\mvar{b}^{-1}))\right)$ iterations satisfies:
\begin{align*}
G(x_T) \leq \frac{1}{\sqrt{10}},
\end{align*}
with probability greater than $1- O(\frac{1}{d^{10}})$.
\end{theorem}
\begin{proof}
As before, we first bound the numerator and denominator of $G(\xhat)$ more carefully as follows:
\begin{align*}
\begin{array}{lrl}
\textbf{Numerator:} &\norm{\bv{P}_{v_1^{\perp}}\xhat}_{\mvar{b}}
&\leq \norm{\bv{P}_{v_1^{\perp}}\mvar{b}^{-1} x}_{\mvar{b}} + \norm{\bv{P}_{v_1^{\perp}}\xi}_\mvar{b}
\leq \norm{\bv{P}_{v_1^{\perp}}\mvar{b}^{-1} x}_{\mvar{b}} + \norm{\xi}_{\mvar{b}} \\
& &=\sqrt{\sum_{i\geq2}\left(v_{i}^{T}B^{-1/2}x\right)^{2}}+\norm{\xi}_{\mvar{b}} = \sqrt{\sum_{i\geq2} \alpha_{i}^{2}\lamiBinv{i}} + \norm{\xi}_{\mvar{b}}, \\
\textbf{Denominator:} &\norm{\bv{P}_{v_1}\xhat}_{\mvar{b}}
&\geq \norm{\bv{P}_{v_1}\mvar{b}^{-1} x}_{\mvar{b}} - \norm{\bv{P}_{v_1}\xi}_\mvar{b}
\geq \norm{\bv{P}_{v_1}\mvar{b}^{-1} x }_{\mvar{b}} - \norm{\xi}_{\mvar{b}} \\
&&=\abs{v_{i}^{T}{\mvar{b}}^{-1/2}x}-\norm{\xi}_{\mvar{b}} = \alpha_1\sqrt{\lamiBinv{1}} - \norm{\xi}_{\mvar{b}}
\end{array}
\end{align*}
We now use the above estimates to bound $G(\xhat)$.
\begin{align*}
G(\xhat) &\leq \frac{\sqrt{\sum_{i\geq2} \alpha_{i}^{2}\lamiBinv{i}} + \norm{\xi}_{\mvar{b}}}{\alpha_1\sqrt{\lamiBinv{1}} - \norm{\xi}_{\mvar{b}}}
\leq \frac{\lamiBinv{2}\sqrt{\sum_{i\geq2} \frac{\alpha_{i}^{2}}{\lamiBinv{i}}} + \norm{\xi}_{\mvar{b}}}{\lamiBinv{1}\sqrt{\frac{\alpha_1^2}{\lamiBinv{1}}} - \norm{\xi}_{\mvar{b}}} \\
&= G(x) \frac{\lamiBinv{2} + \norm{\xi}_{\mvar{b}}/\sqrt{\sum_{i\geq2} \frac{\alpha_{i}^{2}}{\lamiBinv{i}}}}{\lamiBinv{1} - \norm{\xi}_{\mvar{b}}/\sqrt{\frac{\alpha_1^2}{\lamiBinv{1}}}}
\end{align*}
By Lemma \ref{lem:init-random}, we know with at least probability $1- O(\frac{1}{d^{10}})$,
we have $G(x_0) \le \sqrt{\kappa(\mvar{b}^{-1})} d^{10.5}$.
Conditioned on high probability result of $G(x_0)$, we now use induction to prove $G(x_t) \le G(x_0)$. It trivially holds for $t=0$.
Suppose we now have $G(x) \le G(x_0)$, then by the condition in Theorem \ref{thm:init-offline} and Markov inequality, we know with probability greater than $1-\frac{1}{100\sqrt{\kappa(\mvar{b}^{-1})} d^{10.5}}$ we have:
\begin{align*}
\norm{\xi}_\mvar{b} \le &\frac{1}{30\sqrt{\kappa(\mvar{b}^{-1})} d^{10.5}} \cdot \norm{\frac{1}{\lambda-x^\top \bv \Sigma x} x - \mvar{b}^{-1}x}_{\mvar{b}}\\
\le & \frac{1}{30} \cdot \norm{\frac{1}{\lambda-x^\top \bv \Sigma x} x - \mvar{b}^{-1}x}_{\mvar{b}} \min \left \{1, \frac{1}{G(x_0)} \right \} \\
\le &\frac{1}{30} \cdot \norm{\frac{1}{\lambda-x^\top \bv \Sigma x} x - \mvar{b}^{-1}x}_{\mvar{b}} \min \left \{1, \frac{1}{G(x)} \right \} \\
\le & \frac{\lamiBinv{1}-\lamiBinv{2}}{4}\min\left\{\sqrt{\sum_{i\geq2} \frac{\alpha_{i}^{2}}{\lamiBinv{i}}}, \sqrt{\frac{\alpha_1^2}{\lamiBinv{1}}}\right\}
\end{align*}
The last inequality uses Corollary~\ref{cor:constant_factor_corollary} with the fact that $\lamiBinv{2} \le \frac{1}{100}\lamiBinv{1}$. Therefore, we have:
We will have:
\begin{equation*}
G(\xhat) \le \frac{\lamiBinv{1} + 3\lamiBinv{2}}{3\lamiBinv{1}+\lamiBinv{2}} \times G(x)
\le \frac{1}{2}G(x)
\end{equation*}
This finishes the proof of induction.
Finally, by union bound, we know with probability greater than $1- O(\frac{1}{d^{10}})$ in $T = O(\log d + \log \kappa(\mvar{b}^{-1}))$ steps,
we have:
\begin{equation*}
G(x_T) \le \frac{1}{2^T} G(x_0) \le \frac{1}{\sqrt{10}}
\end{equation*}
\end{proof}
\section{Gap-Free Bounds}\label{sec:gapfree}
In this section we demonstrate that our techniques can easily be extended to obtain gap-free runtime bounds, for the regime when $\epsilon \ge \mathrm{gap}$. In many ways these bounds are actually much easier to achieve than the gap dependent bounds since they require less careful error analysis.
Let $\epsilon$ be our error parameter and $m$ be the number of eigenvalues of $\Sigma$ that are $\ge (1-\epsilon/2)\lambda_1$.
Choose $\lambda = \lambda_1 + \epsilon/100$. We have $\lambda_1(\mvar{b}^{-1}) = \frac{100}{\epsilon\lambda_1}$. For $i > m$ we have $\lambda_i(\mvar{b}^{-1}) < \frac{2}{\epsilon\lambda_1}$. $\kappa(\bv{B}^{-1}) \le \frac{100}{\epsilon}$.
Let $\bv{V}_b$ have columns equal to all \emph{bottom} eigenvectors with eigenvalues $\lambda_i < (1-\epsilon/2) \lambda_1$. Let $\bv{V}_t$ have columns equal to the $m$ remaining \emph{top} eigenvectors.
We define a simple modified potential:
\begin{align*}
\bar G(x) \mathbin{\stackrel{\rm def}{=}} \frac{\norm{\bv{P}_{\bv{V}_b}x}_\mvar{b}}{\norm{\bv{P}_{v_1}x}_\mvar{b}} = \frac{\sqrt{\sum_{i > m} \frac{\alpha_i^2}{\lambda_i(\mvar{b}^{-1})}}}{\sqrt{\frac{\alpha_1^2}{\lambda_1(\mvar{b}^{-1})}}}
\end{align*}
We have the following Lemma connecting this potential function to eigenvalue error:
\begin{lemma}
For unit $x$, if $\bar G(x) \le c\sqrt{\epsilon}$ for sufficiently small constant $c$ then $\lambda_1 - x^\top \bv{\Sigma}x \le \epsilon \lambda_1$.
\end{lemma}
\begin{proof}
\begin{align*}
\bar G(x) \ge \frac{\norm{\bv{P}_{\bv{V}_b}x}_2}{\norm{\bv{P}_{v_1}x}_2} \ge \frac{\norm{\bv{P}_{\bv{V}_b}x}_2}{\norm{\bv{P}_{\bv{V}_t}x}_2}
\end{align*}
So if $\bar G(x) \le c\sqrt{\epsilon}$ then $\norm{\bv{P}_{\bv{V}_t}x}^2_2 c^2\epsilon \ge \norm{\bv{P}_{\bv{V}_b}x}^2_2$ and since $\norm{\bv{P}_{\bv{V}_t}x}^2_2 + \norm{\bv{P}_{\bv{V}_b}x}^2_2 = 1$, this gives $\norm{\bv{P}_{\bv{V}_t}x}^2_2 \ge \frac{1}{1+c^2\epsilon}$. So we have $x^T \bv \Sigma x \ge \bv{P}_{\bv{V}_t}x^T \bv \Sigma x\bv{P}_{\bv{V}_t} \ge \frac{(1-\epsilon/2)\lambda_1}{1+c^2\epsilon} \ge 1-\epsilon$ for small enough $c$, giving the lemma.
\end{proof}
We now follow the proof of Lemma \ref{thm:init-offline}, which is actually much simpler in the gap-free case.
\begin{theorem}[Approximate Shifted-and-Inverted Power Method -- Gap-Free]\label{burnInGapFree}
Suppose we randomly initialize $x_0$ as in Lemma \ref{lem:init-random} and suppose we have access to a subroutine $\solve{\cdot}$ such that
\begin{align*}
\expec{\norm{\solve{x}-\mvar{b}^{-1}x}_{\mvar{b}}} \leq \frac{\epsilon^3}{3000 d^{21}} \sqrt{\lambda_d(\mvar{b}^{-1})}
\end{align*}
Then the following procedure,
\begin{align*}
x_{t} = \solve{x_{t-1}}/\norm{\solve{x_{t-1}}}
\end{align*}
after $T = O\left(\log d/\epsilon \right)$ iterations satisfies:
\begin{align*}
\bar G(x_T) \leq c\sqrt{\epsilon},
\end{align*}
with probability greater than $1- O(\frac{1}{d^{10}})$.
\end{theorem}
\begin{proof}
By Lemma \ref{lem:init-random}, we know with at least probability $1- O(\frac{1}{d^{10}})$,
we have $\bar G(x_0) \le G(x_0) \le \sqrt{\kappa(\mvar{b}^{-1})} d^{10.5} = \frac{100d^{10.5}}{\epsilon}$.
We want to show by induction that at iteration $i$ we have $\bar G(x_i) \le \frac{1}{2^i} \cdot \frac{100d^{10.5}}{\epsilon}$, which will give us the lemma if we set $T = \log_2 \left ( \frac{100d^{10.5}}{c\epsilon^{1.5}} \right) = O(\log(d/\epsilon))$.
Let $\xhat = \solve{x}$ and $\xi = \xhat - \mvar{b}^{-1}x$. Following Lemma \ref{thm:init-offline} we have:
\begin{align*}
\norm{\bv{P}_{\bv{V}_b}\left(\xhat \right)}_{\mvar{b}} &\leq \norm{\bv{P}_{\bv{V}_b}\left(\mvar{b}^{-1} x \right)}_{\mvar{b}} + \norm{\bv{P}_{\bv{V}_b}\left(\xi\right)}_\mvar{b} \leq \norm{\bv{P}_{\bv{V}_b}\left(\mvar{b}^{-1} x \right)}_{\mvar{b}} + \norm{\xi}_{\mvar{b}} \\
&=\sqrt{\sum_{i>m}\alpha_i^{2}\lambda_{i}(\mvar{b}^{-1})}+\norm{\xi}_{\mvar{b}}\\
&\le \lambda_{m+1}(\mvar{b}^{-1}) \left (\sqrt{\sum_{i>m} \frac{\alpha_i^2}{\lambda_i(\mvar{b}^{-1})}} + \frac{\epsilon^3}{3000d^{21}\sqrt{\lambda_{m+1}(\mvar{b}^{-1})}}\right )\\
&\le 2\lambda_{m+1}(\mvar{b}^{-1}) \max \left \{ \sqrt{\sum_{i>m} \frac{\alpha_i^2}{\lambda_i(\mvar{b}^{-1})}},\frac{\epsilon^3}{3000d^{21}\sqrt{\lambda_{m+1}(\mvar{b}^{-1})}}\right \}
\end{align*}
and
\begin{align*}
\norm{\bv{P}_{v_1}\left(\xhat \right)}_{\mvar{b}}
&\geq \norm{\bv{P}_{v_1}\left(\mvar{b}^{-1} x \right)}_{\mvar{b}} - \norm{\bv{P}_{v_1}\left(\xi\right)}_\mvar{b}
\geq \norm{\bv{P}_{v_1}\left(\mvar{b}^{-1} x \right)}_{\mvar{b}} - \norm{\xi}_{\mvar{b}} \\
&=\sqrt{\alpha_1^2 \lambda_1(\bv{B}^{-1})} - \norm{\xi}_{\mvar{b}}\\
& \ge \lambda_1(\bv{B}^{-1}) \sqrt{\frac{\alpha_1^2 - \frac{\epsilon^6}{(3000d^{21})^2}}{ \lambda_1(\bv{B}^{-1})}}.
\end{align*}
Initially, we have with high probability, by the argument in Lemma \ref{lem:init-random}, $\alpha_1 \ge \frac{1}{d^{10}}$ so we have $\norm{\bv{P}_{v_1}\left(\xhat \right)}_{\mvar{b}} \ge \frac{\lambda_1(\bv{B}^{-1})}{2} \sqrt{\frac{\alpha_1^2}{ \lambda_1(\bv{B}^{-1})}}$. This also holds by induction in each iteration.
Let $\hat \alpha_1=|v_1^\top \xhat|/\norm{\xhat}_2$. $\norm{\bv{P}_{v_1}\left(\xhat \right)}_{\mvar{b}}^2 = \frac{\hat \alpha_1^2\norm{\hat x}_2^2}{\lambda_1(\bv{B}^{-1})}$ so we have
\begin{align*}
\hat \alpha_1^2 &\ge \frac{\lambda_1(\bv{B}^{-1})^2}{\norm{\hat x}_2^2} \left (\alpha_1^2 - \frac{\epsilon^6}{(3000d^{21})^2}\right )
\end{align*}
and since $\norm{\hat x}_2^2 \le 2 \left (\norm{\bv{B}^{-1}x}_2^2 + 2\norm{\xi}_2^2 \right ) \le \lambda_1(\bv{B}^{-1})^2 + 2\frac{\epsilon^6}{(3000d^{21})^2} \le \lambda_1(\bv{B}^{-1})^2\left (2 + 2\frac{\epsilon^6}{(3000d^{21})^2} \right )$ we have:
\begin{align*}
\hat \alpha_1^2 &\ge \frac{1}{2.1} \left (\alpha_1^2 - \frac{\epsilon^6}{(3000d^{21})^2}\right ) \ge \frac{1}{3} \alpha_1^2.
\end{align*}
So over all $\log_2 \left ( \frac{100d^{10.5}}{c\epsilon^{1.5}} \right)$ iterations, we always have $\hat \alpha_1^2 \ge \frac{1}{d^{10}}\cdot \left ( \frac{c\epsilon^{1.5}}{100d^{10.5}}\right )^{\log_2 3}$ and so $\frac{\epsilon^6}{(3000d^{21})^2} << 1/2\alpha_1^2$.
Combining the above bounds:
\begin{align*}
\bar G(\xhat) &\le \frac{2\lamiBinv{m+1}}{\lamiBinv{1}/2} \cdot \frac{\max \left \{ \sqrt{\sum_{i>m} \frac{\alpha_i^2}{\lambda_i(\mvar{b}^{-1})}},\frac{\epsilon^3}{3000d^{21}\sqrt{\lambda_{m+1}(\mvar{b}^{-1})}}\right \}}{\sqrt{\frac{\alpha_1^2}{ \lambda_1(\bv{B}^{-1})}}}\\
&\le \frac{4}{50} \max \left \{\bar G(x), O(\sqrt{\epsilon}) \right \}.
\end{align*}
This is enough to give the Theorem.
\end{proof}
Finally, we combine Theorem \ref{burnInGapFree} with the SVRG based solvers of Theorem \ref{offline_solver} and \ref{accelerated_offline_solver} to obtain:
\begin{theorem}[Gap-Free Shifted-and-Inverted Power Method With SVRG]\label{main_gapfree_theorem}
Let $\bv{B} = \lambda \bv{I} - \bv{A}^\top \bv{A}$ for $\lambda = \left ( 1+\frac{\epsilon}{100}\right)$ and let $x_0 \sim \mathcal{N}(0,\bv I)$ be a random initial vector. Running the inverted power method on $\bv{B}$ initialized with $x_0$, using the SVRG solver from Theorem \ref{offline_solver} to approximately apply $\bv{B}^{-1}$ at each step, returns $x$ such that with probability $1-O\left (\frac{1}{d^{10}}\right)$, $x^\top \bv{\Sigma}x \ge (1-\epsilon) \lambda_1$ in time $$O \left (\left(\nnz(\bv A) + \frac{d \mathsf{r}(\bv A)}{\epsilon^2} \right )\cdot \log^2\left(\frac{d}{\epsilon}\right) \right ).$$
\end{theorem}
\begin{theorem}[Accelerated Gap-Free Shifted-and-Inverted Power Method With SVRG]\label{acell_gapfree_theorem}
Let $\bv{B} = \lambda \bv{I} - \bv{A}^\top \bv{A}$ for $\lambda = \left ( 1+\frac{\epsilon}{100}\right)$ and let $x_0 \sim \mathcal{N}(0,\bv I)$ be a random initial vector. Running the inverted power method on $\bv{B}$ initialized with $x_0$, using the SVRG solver from Theorem \ref{accelerated_offline_solver} to approximately apply $\bv{B}^{-1}$ at each step, returns $x$ such that with probability $1-O\left (\frac{1}{d^{10}}\right)$, $x^\top \bv{\Sigma}x \ge (1-\epsilon) \lambda_1$ in total time $$O \left (\frac{\nnz(\bv A)^{3/4} (d \mathsf{r}(\bv A))^{1/4}}{\sqrt{\epsilon}} \cdot \log^3\left(\frac{d}{\epsilon}\right) \right ).$$
\end{theorem}
\section{Introduction}
Given $\bv{A} \in \mathbb{R}^{n \times d}$, computing the top eigenvector
of $\bv{A}^{\top}\bv{A}$ is a fundamental problem in numerical linear algebra, applicable to principal component analysis \cite{jolliffe2002principal}, spectral clustering and learning \cite{ng2002spectral,vempala2004spectral}, pagerank computation, and many other graph computations \cite{page1999pagerank,koren2003spectral,spielman2007spectral}. For instance, a degree-$k$ principal component analysis is nothing more than performing $k$ leading eigenvector computations.
Given the ever-growing size of modern datasets, it is thus a key challenge to come up with more efficient algorithms for this basic computational primitive.
In this work we provide improved algorithms for computing the top eigenvector,
both in the \emph{offline} case, when $\bv{A}$ is given
explicitly and in the \emph{online} or \emph{statistical} case where we access samples from a distribution $\mathcal{D}$ over
$\mathbb{R}^{d}$ and wish to estimate the top eigenvector of the covariance matrix $\mathbb{E}_{a\sim\mathcal{D}} \left [aa^\top \right ]$.
In the offline case, our algorithms are the fastest to date in a wide and meaningful regime of parameters. Notably, while the running time of most popular methods for eigenvector computations is a product of the size of the dataset (i.e. number of non-zeros in $\bv{A}$) and certain spectral characteristics of $\bv{A}$, which both can be quite large in practice, we present running times that actually split the dependency between these two quantities, and as a result may yield significant speedups.
In the online case, our results yield improved sample complexity bounds and allow for very efficient \textit{streaming} implementations with memory and processing-time requirements that are proportional to the size of a single sample.
On a high-level, our algorithms are based on a robust analysis of the classic idea of \emph{shift-and-invert} preconditioning \cite{saad1992numerical}, which allows us to efficiently reduce eigenvector computation to \emph{approximately} solving a \emph{short} sequence of \emph{well-conditioned} linear systems in $\lambda \bv I - \bv{A}^\top \bv {A}$ for some shift parameter $\lambda \approx \lambda_1(\bv{A})$. We then apply state-of-the-art stochastic gradient methods to approximately solve these linear systems.
\subsection{Our Approach}
The well known power method for computing the top eigenvector of $\bv{A}^\top \bv{A}$ starts with an initial vector $x$ and repeatedly multiplies by $\bv{A}^\top \bv{A}$, eventually causing $x$ to converge to the top eigenvector. For a random start vector, the power method converges in $O (\log(d/\epsilon)/\mathrm{gap} )$ iterations, where $\mathrm{gap} = (\lambda_1-\lambda_2)/\lambda_1$, $\lambda_i$ denotes the $i^{th}$ largest eigenvalue of $\bv{A}^\top \bv{A}$, and we assume a high-accuracy regime where $\epsilon < \mathrm{gap}$.
The dependence on this gap ensures that the largest eigenvalue is significantly amplified in comparison to the remaining values.
If the eigenvalue gap is small, one approach is replace $\bv{A}^\top \bv{A}$ with a preconditioned matrix -- i.e. a matrix with the same top eigenvector but a much larger gap.
Specifically, let $\bv{B} = \lambda \bv{ I} - \bv{A}^\top \bv{A}$ for some shift parameter $\lambda$. If $\lambda > \lambda_1$, we can see that the smallest eigenvector of $\bv{B}$ (the largest eigenvector of $\bv{B}^{-1}$) is equal to the largest eigenvector of $\bv{A}^\top \bv{A}$. Additionally, if $\lambda$ is close to $\lambda_1$, there will be a constant gap between the largest and second largest values of $\bv{B}^{-1}$. For example, if $\lambda = (1+\mathrm{gap})\lambda_1$, then we will have $\lambda_1\left(\bv{B}^{-1} \right ) = \frac{1}{\lambda-\lambda_1} = \frac{1}{\mathrm{gap} \cdot \lambda_1}$ and $\lambda_2\left(\bv{B}^{-1} \right ) = \frac{1}{\lambda - \lambda_2} = \frac{1}{2\cdot\mathrm{gap} \cdot \lambda_1}$.
This constant factor gap ensures that the power method applied to $\bv{B}^{-1}$ converges to the top eigenvector of $\bv{A}^\top \bv{A}$ in just $O (\log (d/\epsilon) )$ iterations. Of course, there is a catch -- each iteration of this \emph{shifted-and-inverted power method} must solve a linear system in $\bv{B}$, whose condition number is proportional $\frac{1}{\mathrm{gap}}$. For small gap, solving this system via iterative methods is more expensive.
Fortunately, linear system solvers are incredibly well studied and there are many efficient iterative algorithms we can adapt to apply $\bv{B}^{-1}$ approximately. In particular, we show how to accelerate the iterations of the shifted-and-inverted power method using variants of Stochastic Variance Reduced Gradient (SVRG) \cite{johnson2013accelerating}.
Due to the condition number of $\bv{B}$, we will not entirely avoid a $\frac{1}{\mathrm{gap}}$ dependence, however, we can separate this dependence from the input size $\nnz(\bv A)$.
Typically, stochastic gradient methods are used to optimize convex functions that are given as the sum of many convex components. To solve a linear system $(\bv{M}^\top\bv{M}) x = b$ we minimize the convex function $f(x) = \frac{1}{2} x^\top (\bv{M}^\top\bv{M}) x - b^\top x$ with components $\psi_i(x) = \frac{1}{2} x^\top \left (m_i m_i^\top \right )x - \frac{1}{n}b^\top x$ where $m_i$ is the $i^{th}$ row of $\bv M$. Such an approach can be used to solve systems in $\bv{A}^\top \bv{A}$, however solving systems in $\bv{B} = \lambda \bv I - \bv A^\top \bv A$ requires more care. We require an analysis of SVRG that guarantees convergence even when some of our components are \emph{non-convex}. We give a simple analysis for this setting, generalizing recent work in the area \cite{shalev2015sdca,csiba2015primal}.
Given fast approximate solvers for $\bv{B}$, the second main piece of our algorithmic framework is a new error bound for the shifted-and-inverted power method, showing that it is robust to approximate linear system solvers, such as SVRG.
We give a general analysis, showing
exactly what accuracy each system must be solved to, allowing for faster implementations using linear solvers with weaker guarantees. Our proofs center around the potential function: $G(x) \defeq
\norm{\bv{P}_{v_1^{\perp}}x}_{\mvar{b}}/\norm{\bv{P}_{v_1}x}_{\mvar{b}}$,
where $\bv{P}_{v_1}$ and $\bv{P}_{v_1^{\perp}}$ are the projections onto the top eigenvector and its complement respectively. This function resembles tangent based potential functions used in previous work \cite{hardt2014noisy} except that we use the $\bv{B}$ norm rather than the $\ell_2$ norm. For the exact power method, this is irrelevant -- progress is identical in both norms (see Lemma \ref{same_progress} of the Appendix). However, $\norm{\cdot}_\mvar{b}$ is a natural norm for measuring the progress of linear system solvers for $\mvar{b}$, so our potential function makes it possible to extend analysis to the case when $\bv{B}^{-1}x$ is computed up to error $\xi$ with bounded $\norm{\xi}_\mvar{b}$.
\subsection{Our Results}
Our algorithmic framework described above offers several advantageous. We obtain improved running times for computing the top eigenvector in the offline model. In Theorem \ref{main_offline_theorem} we give an algorithm running in time $O\left(\left [\nnz(\bv{A}) + \frac{d \mathsf{r} \bv{A}}{\mathrm{gap}^2}\right ] \cdot \left [ \log \frac{1}{\epsilon} + \log^2 \frac{d}{\mathrm{gap}} \right ]\right)$, where $\mathsf{r}(\bv{A}) = \norm{\bv A}_F^2 / \norm{\bv A}_2^2 \le \rank(\bv{A})$ is the stable rank and $\nnz(\bv{A})$ is the number of non-zero entries. Up to log factors, our runtime is in many settings proportional to the input size $\nnz(\bv{A})$, and so is very efficient for large matrices. In the case when $\nnz(\bv{A}) \le \frac{d\mathsf{r}(\bv{A})}{\mathrm{gap}^2}$ we apply the results of \cite{frostig2015regularizing, lin2015catalyst} to provide an accelerated runtime of $O\left(\left [\frac{\nnz(\bv{A})^{3/4}(d \mathsf{r} (\bv{A}))^{1/4}}{\sqrt{\mathrm{gap}}}\right ] \cdot \left [\log \frac{d}{\mathrm{gap}} \log \frac{1}{\epsilon} + \log^3 \frac{d}{\mathrm{gap}} \right ]\right)$, shown in Theorem \ref{accelerated_offline_theorem}. Finally, in the case when $\epsilon > \mathrm{gap}$, our results easily extend to give gap-free bounds (Theorems \ref{main_gapfree_theorem} and \ref{acell_gapfree_theorem}), identical to those shown above but with $\mathrm{gap}$ replaced by $\epsilon$. Note that our offline results hold for any $\bv{A}$ and require no initial knowledge of the top eigenvector. In Section \ref{parameter_free} we discuss how to estimate the parameters $\lambda_1$, $\mathrm{gap}$, with modest additional runtime cost.
Our algorithms return an approximate top eigenvector $x$ with $x^\top\bv{A}^\top \bv{A} x \ge (1-\epsilon)\lambda_1$. By choosing error $\epsilon \cdot \mathrm{gap}$, we can ensure that $x$ is actually close to $v_1$ -- i.e. that $|x^\top v_1| \ge 1-\epsilon$. Further, we obtain the same asymptotic runtime since $O\left (\log \frac{1}{\epsilon \cdot \mathrm{gap}} + \log^2 \frac{d}{\mathrm{gap}} \right ) = O\left (\log \frac{1}{\epsilon} + \log^2 \frac{d}{\mathrm{gap}} \right )$. We compare our runtimes with previous work in Table \ref{offline_table}.
In the online case, in Theorem \ref{warmstart_online_theorem}, we show how to improve an $O(\mathrm{gap})$ approximation to the top eigenvector to an $\epsilon$ approximation with constant probability using $O\left (\frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap} \cdot \epsilon}\right)$ samples where $\mathsf{v}(\mathcal{D})$ is a natural variance measure. Our algorithm is based on the streaming SVRG algorithm of \cite{frostig2014competing}. It requires just $O(d)$ amortized time per sample, uses just $O(d)$ space, and is easily parallelized. We can apply our result in a variety of regimes, using existing algorithms to obtain the initial $O(\mathrm{gap})$ approximation and our algorithm to improve. As shown in Table \ref{online_table}, this gives improved runtimes and sample complexities over existing work. Notably, we give better asymptotic sample complexity than known matrix concentration results for general distributions, and give the first streaming algorithm that is asymptotically optimal in the popular Gaussian spike model.
Overall, our robust shifted-and-inverted power method analysis gives new understanding of this classical technique. It gives a means of obtaining provably accurate results when each iteration is implemented using fast linear system solvers with weak accuracy guarantees.
In practice, this reduction between approximate linear system solving and eigenvector computation shows that optimized regression libraries can be leveraged for faster eigenvector computation in many cases. Furthermore, in theory we believe that the reduction suggests computational limits inherent in eigenvector computation as seen by the often easier-to-analyze problem of linear system solving. Indeed, in Section \ref{sec:lower}, we provide evidence that in certain regimes our statistical results are optimal.
\subsection{Previous Work}
\subsubsection*{Offline Eigenvector Computation}\label{previous_work_offline}
Due to its universal applicability, eigenvector computation in the offline case is extremely well studied. Classical methods, such as the QR algorithm, take roughly $O(nd^2)$ time to compute a full eigendecomposition. This can be accelerated to $O(nd^{\omega - 1})$, where $\omega < 2.373$ is the matrix multiplication constant \cite{williams2012multiplying,le2014powers}, however this is still prohibitively expensive for large matrices. Hence, faster iterative methods are often employed, especially when only the top eigenvector (or a few of the top eigenvectors) is desired.
As discussed, the popular power method requires $O \left ( \frac{\log(d/\epsilon)}{\mathrm{gap}} \right )$ iterations to converge to an $\epsilon$ approximate top eigenvector. Using Chebyshev iteration, or more commonly, the Lanczos method, this bound can be improved to $O \left ( \frac{\log(d/\epsilon)}{\sqrt{\mathrm{gap}}} \right )$ \cite{saad1992numerical}, giving total runtime of $O \left (\nnz(\bv A) \cdot \frac{\log(d/\epsilon)}{\sqrt{\mathrm{gap}}} \right )$. When $\epsilon > \mathrm{gap}$, the $\mathrm{gap}$ terms in these runtimes can be replaced by $\epsilon$. While we focus on the high-precision regime when $\epsilon < \mathrm{gap}$, we also give gap-free bounds in Section \ref{sec:gapfree}.
Unfortunately, if $\nnz(\bv{A})$ is very large and $\mathrm{gap}$ is small, the above runtimes can still be quite expensive, and there is a natural desire to separate the $\frac{1}{\sqrt{\mathrm{gap}}}$ dependence from the $\nnz(\bv A)$ term. One approach is to use random subspace embedding matrices \cite{ailon2009fast,clarkson2013low} or fast row sampling algorithms \cite{cohen2015uniform}, which can be applied in $ O(\nnz(\bv A))$ time and yield a matrix $\bv{\tilde A}$ which is a good spectral approximation to the original. The number of rows in $\bv{\tilde A}$ depends only on the stable rank of $\bv A$ and the error of the embedding -- hence it can be significantly smaller than $n$. Applying such a subspace embedding and then computing the top eigenvector of $\bv{\tilde A}^\top \bv{\tilde A}$ requires runtime $O \left (\nnz(\bv A) + \mathop\mathrm{poly}(\mathsf{r}(\bv{A}),\epsilon,\mathrm{gap}) \right )$, achieving the goal of reducing runtime dependence on the input size $\nnz(\bv A)$. Unfortunately, the dependence on $\epsilon$ is significantly suboptimal -- such an approach cannot be used to obtain a linearly convergent algorithm. Further, the technique does not extend to the online setting, unless we are willing to store a full subspace embedding of our sampled rows.
Another approach, which we follow more closely, is to apply stochastic optimization techniques, which iteratively update an estimate to the top eigenvector, considering a random row of $\bv{A}$ with each update step. Such algorithms naturally extend to the online setting and have led to improved dependence on the input size for a variety of problems \cite{bottou2010large}. Using variance-reduced stochastic gradient techniques, \cite{shamir2015stochastic} achieves runtime $ O\left (\left (\nnz(\bv A) + \frac{dr^2n^2}{\mathrm{gap}^2\lambda_1^2} \right ) \cdot \log(1/\epsilon) \log\log(1/\epsilon) \right )$ for approximately computing the top eigenvector of a matrix with constant probability. Here $r$ is an upper bound on the squared row norms of $\bv{A}$. In the \emph{best case}, when row norms are uniform, this runtime can be simplified to $O\left (\left (\nnz(\bv A) + \frac{d\mathsf{r}(\bv A)^2}{\mathrm{gap}^2} \right ) \cdot \log(1/\epsilon) \log\log(1/\epsilon) \right )$.
The result in \cite{shamir2015stochastic} makes an important contribution in separating input size and gap dependencies using stochastic optimization techniques.
Unfortunately, the algorithm requires an approximation to the eigenvalue gap and a starting vector that has a constant dot product with the top eigenvector. In \cite{shamir2015fast} the analysis is extended to a random initialization, however loses polynomial factors in $d$. Furthermore, the dependencies on the stable rank and $\epsilon$ are suboptimal -- we improve them to $\mathsf{r}(\bv A)$ and $\log(1/\epsilon)$ respectively, obtaining true linear convergence.
\begin{table}[h]
\def1{1}
\begin{center}
\begin{tabular}{|>{\centering}m{6.5cm}|c|}
\hline
\textbf{Algorithm} & \textbf{Runtime}\\
\hline
Power Method & $O\left (\nnz(\bv{A})\frac{\log(d/\epsilon)}{\mathrm{gap}} \right )$ \\
\hline
Lanczos Method & $O\left (\nnz(\bv{A})\frac{\log(d/\epsilon)}{\sqrt{\mathrm{gap}}} \right )$ \\
\hline
Fast Subspace Embeddings \cite{clarkson2013low} Plus Lanczos & $O\left (\nnz(\bv{A}) + \frac{d\mathsf{r}(\bv{A})}{\max \left \{ \mathrm{gap}^{2.5} \epsilon, \epsilon^{2.5} \right \} } \right )$ \\
\hline
SVRG \cite{shamir2015stochastic} (assuming bounded row norms, warm-start)& $O\left (\left (\nnz(\bv A) + \frac{d\mathsf{r}(\bv A)^2}{\mathrm{gap}^2} \right ) \cdot \log(1/\epsilon) \log\log(1/\epsilon) \right )$\\
\hline
\textbf{Theorem \ref{main_offline_theorem}} & $O\left(\left [\nnz(\bv{A}) + \frac{d \mathsf{r}(\bv{A})}{\mathrm{gap}^2}\right ] \cdot \left [\log \frac{1}{\epsilon} + \log^2 \frac{d}{\mathrm{gap}} \right ]\right)$\\
\hline
\textbf{Theorem \ref{accelerated_offline_theorem}} & $O\left(\left [\frac{\nnz(\bv{A})^{3/4}(d \mathsf{r}(\bv{A}))^{1/4}}{\sqrt{\mathrm{gap}}}\right ] \cdot \left[\log \frac{d}{\mathrm{gap}} \log \frac{1}{\epsilon} + \log^3 \frac{d}{\mathrm{gap}} \right ]\right)$\\
\hline
\end{tabular}
\caption{Comparision to previous work on Offline Eigenvector Estimation. We give runtimes for computing a unit vector $x$ such that $x^\top \bv{A}^\top\bv{A} x \ge (1-\epsilon)\lambda_1$ in the regime $\epsilon = O(\mathrm{gap})$.}
\label{offline_table}
\end{center}
\vspace{-2.5em}
\end{table}
\subsubsection*{Online Eigenvector Computation}
While in the offline case the primary concern is computation time, in the online, or statistical setting, research also focuses on minimizing the number of samples that are drawn from $\mathcal{D}$ in order to achieve a given accuracy. Especially sought after are results that achieve asymptotically optimal accuracy as the sample size grows large.
While the result we give in Theorem \ref{warmstart_online_theorem} works for any distribution parameterized by a variance bound, in this section, in order to more easily compare to previous work, we normalize $\lambda_1 = 1$ and assume we have the row norm bound $\norm{a}_2^2 \le O(d)$ which then gives us the variance bound $\norm{\mathbb{E}_{a\sim\mathcal{D}}\left[(aa^\top)^2\right]}_2 = O(d)$. Additionally, we compare runtimes for computing some $x$ such that $|x^\top v_1| \ge 1-\epsilon$, as this is the most popular guarantee studied in the literature. Theorem \ref{warmstart_online_theorem} is easily extended to this setting as obtaining $x$ with $x^T \bv{AA}^\top x \ge (1-\epsilon\cdot \mathrm{gap}) \lambda_1$ ensures $|x^\top v_1| \ge 1-\epsilon$. Our algorithm requires $O\left ( \frac{d}{\mathrm{gap}^2 \epsilon} \right )$ samples to find such a vector under the assumptions given above.
The simplest algorithm in this setting is to take $n$ samples from $\mathcal{D}$ and compute the leading eigenvector of the empirical estimate $\widehat\mathbb{E}[a a^\top] = \frac{1}{n} \sum_{i=1}^n a_i a_i^\top$. By a matrix Bernstein bound, such as inequality of Theorem 6.6.1 of \cite{tropp2015introduction}, $O\left ( \frac{d\log d}{\mathrm{gap}^2 \epsilon} \right )$ samples is enough to insure $\norm{\widehat\mathbb{E}[a a^\top] - \mathbb{E}[a a^\top]}_2 \le \sqrt{\epsilon}\cdot \mathrm{gap}$.
By Lemma \ref{spectral_error_conversion} in the Appendix, this gives that, if $x$ is set to the top eigenvector of $\widehat\mathbb{E}[a a^\top] $ it will satisfy $|x^\top v_1| \ge 1-\epsilon$. $x$ can be approximated with any offline eigenvector algorithm.
A large body of work focuses on improving this simple algorithm, under a variety of assumptions on $\mathcal{D}$. A common focus is on obtaining \emph{streaming algorithms}, in which the storage space is just $O(d)$ - proportional to the size of a single sample.
In Table \ref{online_table} we give a sampling of results in this area. All listed results rely on distributional assumptions at least as strong as those given above.
Note that, in each setting, we can use the cited algorithm to first compute an $O(\mathrm{gap})$ approximate eigenvector, and then refine this approximation to an $\epsilon$ approximation using $O\left ( \frac{d}{\mathrm{gap}^2 \epsilon} \right )$ samples by applying our streaming SVRG based algorithm.
This allows us to obtain improved runtimes and sample complexities.
To save space, we do not include our improved runtime bounds in Table \ref{online_table}, however they are easy to derive by adding the runtime required by the given algorithm to achieve $O(\mathrm{gap})$ accuracy, to $O\left (\frac{d^2}{\mathrm{gap}^{2} \epsilon}\right)$ -- the runtime required by our streaming algorithm.
The bounds given for the simple matrix Bernstein based algorithm described above, Krasulina/Oja's Algorithm \cite{balsubramani2013fast}, and SGD \cite{shamir2015convergence} require no additional assumptions, aside from those given at the beginning of this section.
The streaming results cited for \cite{mitliagkas2013memory} and \cite{hardt2014noisy} assume $a$ is generated from a
Gaussian spike model, where $a_i = \sqrt{\lambda_1}\gamma_i{v_1} + Z_i$
and $\gamma_i \sim \mathcal{N}(0, 1), Z_i \sim \mathcal{N}(0, I_d)$. We note that under this model, the matrix Bernstein results improve by a $\log d$ factor and so match our results in achieving asymptotically optimal convergence rate. The results of \cite{mitliagkas2013memory} and \cite{hardt2014noisy} sacrifice this optimality in order to operate under the streaming model. Our work gives the best of both works -- a streaming algorithm giving asymptotically optimal results.
The streaming Alecton algorithm \cite{sa2015global} assumes $\mathbb{E}\norm{aa^\top \bv W a a^\top} \le O(1)\text{tr}(\bv W)$ for any symmetric $\bv W$ that commutes with $\mathbb{E} aa^\top$. This is strictly stronger than our assumption that
\\$\norm{\mathbb{E}_{a\sim\mathcal{D}} \left[(aa^\top)^2\right]}_2 = O(d)$.
\begin{table}[h]
\def1{1}
\small
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Algorithm} & \specialcell{\textbf{Sample}\\ \textbf{Size}} & \textbf{Runtime} & \textbf{Streaming?} & \specialcell{\textbf{Our Sample}\\ \textbf{Complexity}}\\
\hline
\specialcell{Matrix Bernstein plus \\ Lanczos (explicitly forming\\ sampled matrix)} & $O\left (\frac{d\log d}{gap^2 \epsilon}\right) $ & $O\left(\frac{d^3\log d}{gap^2\epsilon}\right)$ & $\times$ & $O\left (\frac{d\log d}{gap^3} + \frac{d}{gap^2 \epsilon}\right) $ \\
\hline
\specialcell{Matrix Bernstein plus \\Lanczos (iteratively applying\\ sampled matrix)} & $O\left (\frac{d\log d}{gap^2 \epsilon}\right) $ & $O\left (\frac{d^2\log d\cdot \log(d/\epsilon)}{gap^{2.5}\epsilon}\right)$ & $\times$ & $O\left (\frac{d\log d}{gap^3} + \frac{d}{gap^2 \epsilon}\right) $\\
\hline
\specialcell{Memory-efficient PCA \\\cite{mitliagkas2013memory, hardt2014noisy}} & $O\left(\frac{d\log (d/\epsilon)}{gap^3 \epsilon}\right )$ & $O\left (\frac{d^2\log (d/\epsilon)}{gap^3 \epsilon}\right)$ & $\surd$ & $O\left(\frac{d\log (d/\mathrm{gap})}{gap^4} + \frac{d}{gap^2 \epsilon}\right )$\\
\hline
Alecton \cite{sa2015global} & $O(\frac{d\log (d/\epsilon)}{gap^2 \epsilon})$ & $O(\frac{d^2\log (d/\epsilon)}{gap^2 \epsilon})$ & $\surd$ & $O(\frac{d\log (d/\mathrm{gap})}{gap^3} + \frac{d}{gap^2 \epsilon})$ \\
\hline
\specialcell{Krasulina / Oja's \\Algorithm \cite{balsubramani2013fast}} & $O(\frac{d^{c_1}}{gap^2 \epsilon^{c_2}})$ &$O(\frac{d^{c_1+1}}{gap^2 \epsilon^{c_2}})$ & $\surd$ & $O(\frac{d^{c_1}}{gap^{2+c_2}} + \frac{d}{gap^2 \epsilon})$ \\
\hline
SGD \cite{shamir2015convergence} & $O(\frac{d^3\log (d/\epsilon)}{\epsilon^2})$ & $O(\frac{d^4\log (d/\epsilon)}{\epsilon^2})$ & $\surd$ & $O\left (\frac{d^3\log (d/\mathrm{gap})}{\mathrm{gap}^2} +\frac{d}{\mathrm{gap}^2 \epsilon} \right ) $ \\
\hline
\end{tabular}
\caption{Summary of existing work on Online Eigenvector Estimation and improvements given by our results. Runtimes are for computing a unit vector $x$ such that $|x^\top v_1| \ge 1-\epsilon$. For each of these results we can obtain improved running times and sample complexities by running the algorithm to first compute an $O(\mathrm{gap})$ approximate eigenvector, and then running our algorithm to obtain an $\epsilon$ approximation using an additional $O\left ( \frac{d}{\mathrm{gap}^2 \epsilon} \right )$ samples, $O(d)$ space, and $O(d)$ work per sample.
}
\label{online_table}
\end{center}
\end{table}
\vspace{-2.5em}
\subsection{Paper Organization}
\begin{description}
\item[Section \ref{prelims}] Review problem definitions and parameters for our runtime and sample bounds.
\item[Section \ref{framework}] Describe the shifted-and-inverted power method and show how it can be implemented using approximate system solvers.
\item[Section \ref{sec:offline}] Show how to apply SVRG to solve systems in our shifted matrix, giving our main runtime results for offline eigenvector computation.
\item[Section \ref{sec:online}] Show how to use an online variant of SVRG to run the shifted-and-inverted power method, giving our main sampling complexity and runtime results in the statistical setting.
\item[Section \ref{parameter_free}] Show how to efficiently estimate the shift parameters required by our algorithms.
\item[Section \ref{sec:lower}] Give a lower bound in the statistical setting, showing that our results are asymptotically optimal for a wide parameter range.
\item[Section \ref{sec:gapfree}] Give gap-free runtime bounds, which apply when $\epsilon > \mathrm{gap}$.
\end{description}
\section{Lower Bounds}\label{sec:lower}
Here we show that our online eigenvector estimation algorithm (Theorem \ref{warmstart_online_theorem}) is asymptotically optimal - as sample size grows large it achieves optimal accuracy as a function of sample size.
We rely on the following lower bound for eigenvector estimation in the Gaussian spike model:
\begin{lemma}[Lower bound for Gaussian Spike Model \cite{birnbaum2013minimax}] \label{lowerbound_spike_lemma}
Suppose data is generated as
\begin{equation}\label{spike_model}
a_i = \sqrt{\lambda}\iota_i{v^\star} + Z_i
\end{equation}
where $\iota_i \sim \mathcal{N}(0, 1)$, and $Z_i \sim \mathcal{N}(0, I_d)$.
Let $\hat{v}$ be some estimator of the top eigenvector $v^\star$.
Then, there is some universal constant $c_0$, so that for $n$ sufficiently large, we have:
\begin{equation*}
\inf_{\hat{v}} \max_{v^\star \in \mathbb{S}^{d-1}} \mathbb{E} \norm{\hat{v} - v^\star}_2
\ge c_0 \frac{(1+\lambda)d}{\lambda^2 n}
\end{equation*}
\end{lemma}
\begin{theorem}\label{lowerbound_online_theorem}
Consider the problem of estimating the top eigenvector $v_1$ of $\mathbb{E}_{a\sim \mathcal{D}} aa^\top$, where we observe $n$ i.i.d samples from unknown distribution $\mathcal{D}$. If $\mathrm{gap} < 0.9$, then there exists some universal constant c, such that for any estimator $\hat{v}$ of top eigenvector, there always exists some hard distribution $\mathcal{D}$ so that for $n$ sufficiently large:
\begin{align*}
\mathbb{E}\norm{\hat{v}-v_1}^2_2 \ge c \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2n}
\end{align*}
\end{theorem}
\begin{proof}
Suppose the claim of theorem is not true, then there exist some estimator $\hat{v}$ so that
\begin{equation*}
\mathbb{E}\norm{\hat{v}-v_1}^2_2 < c' \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2n}
\end{equation*}
holds for all distribution $\mathcal{D}$, and for any fixed constant $c'$ when $n$ is sufficiently large.
Let distribution $\mathcal{D}$ be the Gaussian Spike Model specified by Eq.(\ref{spike_model}), then
by calculation, it's not hard to verify that:
\begin{equation*}
\mathsf{v}(\mathcal{D}) = \frac{\norm{\mathbb{E}_{a \sim \mathcal{D}}\left [ \left (a a^\top \right )^2 \right ]}_2}{\norm{\mathbb{E}_{a \sim \mathcal{D}} (aa^\top)}_2^2} =\frac{d+2+3\lambda}{1+\lambda}
\end{equation*}Since we know $\mathrm{gap} = \frac{\lambda}{1+\lambda} <0.9$, this implies $\lambda <9$, which gives
$\mathsf{v}(\mathcal{D}) < \frac{d+29}{1+\lambda} <\frac{30d}{1+\lambda}$.
Therefore, we have that:
\begin{equation*}
\mathbb{E}\norm{\hat{v}-v^\star}^2_2 < c' \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2n} <30c' \frac{(1+\lambda)d}{\lambda^2 n}
\end{equation*}
holds for all $v^\star \in \mathbb{S}^{d-1}$. Choose $c' = \frac{c_0}{30}$ in Lemma \ref{lowerbound_spike_lemma} we have a contradiction.
\end{proof}
$\norm{\hat{v}-v_1}^2_2 = 2 - 2 \hat{v}^\top v_1$, so this bound implies that- to obtain $|\hat{v}^\top v_1| \ge 1- \epsilon$, we need $\frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2 n} = O(\epsilon)$ so $n = \Theta \left (\frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2 \epsilon} \right )$. This exactly matches the sample complexity given by Theorem \ref{warmstart_online_theorem}.
\section{Acknowledgements}
Sham Kakade acknowledges funding from the Washington Research
Foundation for innovation in Data-intensive Discovery.
\bibliographystyle{alpha}
\section{Offline Eigenvector Computation}\label{sec:offline}
In this section we show how to instantiate the framework of Section \ref{framework} in order to compute an approximate top eigenvector in the offline setting. As discussed, in the offline setting we can trivially compute the Rayleigh quotient of a vector in $\nnz(\mvar{a})$ time as we have explicit access to $\mvar{a}^\top \mvar{a}$. Consequently the bulk of our work in this section is to show how we can solve linear systems in $\mvar{b}$ efficiently in expectation, allowing us to apply Corollary \ref{cor:constant_factor_corollary} of Theorem \ref{thm:powermethod-perturb}.
In Section~\ref{sec:offline:svrg} we first show how Stochastic Variance Reduced Gradient (SVRG) \cite {johnson2013accelerating} can be adapted to solve linear systems of the form $\mvar{b} x = b$.
If we wanted, for example, to solve a linear system in a positive definite matrix like $\bv{A}^\top \bv A$, we would optimize the objective function $f(x) = \frac{1}{2}x^\top \bv{A}^\top \bv A x - b^\top x$. This function can be written as the sum of $n$ \emph{convex components}, $\psi_i(x) = \frac{1}{2} x^\top \left (a_i a_i^\top \right )x - \frac{1}{n}b^\top x$. In each iteration of traditional gradient descent, one computes the full gradient of $f(x_i)$ and takes a step in that direction. In stochastic gradient methods, at each iteration, a single component is sampled, and the step direction is based only on the gradient of the sampled component. Hence, we avoid a full gradient computation at each iteration, leading to runtime gains.
Unfortunately, while we have access to the rows of $\bv{A}$ and so can solve systems in $\bv A^\top \bv A$, it is less clear how to solve systems in $\bv{B} = \lambda \bv I - \bv A^\top \bv A$. To do this, we will split our function into components of the form $\psi_i(x) = \frac{1}{2} x^\top \left (w_i \bv I - a_i a_i^\top \right )x - \frac{1}{n}b^\top x$ for some set of weights $w_i$ with $\sum_{i \in [n]} w_i = \lambda$.
Importantly, $(w_i \bv I - a_i a_i^\top)$ may not be positive semidefinite. That is, we are minimizing a sum of functions which is convex, but consists of non-convex components. While recent results for minimizing such functions could be applied directly \cite{shalev2015sdca,csiba2015primal} here we show how to obtain stronger results by using a more general form of SVRG and analyzing the specific properties of our function (i.e. the variance).
Our analysis shows that we can make constant factor progress in solving linear systems in $\bv{B}$ in time $O\left (\nnz(\bv A) + \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2} \right )$. If $\frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2} \le \nnz(\bv A)$ this gives a runtime proportional to the input size -- the best we could hope for.
If not, we show in Section~\ref{sec:offline:acceleration} that it is possible to \emph{accelerate} our system solver, achieving runtime $\tilde O \left (\frac{\nnz(\bv A)^{3/4}(d\mathsf{r}(\bv A))^{1/4}}{\sqrt{\mathrm{gap}}} \right )$. This result uses the work of \cite{frostig2015regularizing, lin2015catalyst} on accelerated approximate proximal point algorithms.
With our solvers in place, in
Section~\ref{sec:offline:results} we pull our results together, showing how to use these solvers in the framework of Section \ref{framework} to give faster running times for offline eigenvector computation.
\subsection{SVRG Based Solver}
\label{sec:offline:svrg}
Here we provide a sampling based algorithm for solving linear systems in $\mvar{b}$. In particular we provide an algorithm for solving the more general problem where we are given a strongly convex function that is a sum of possibly non-convex functions that obey smoothness properties. We provide a general result on bounding the progress of an algorithm that solves such a problem by non-uniform sampling in Theorem~\ref{lem:svrg-nonconv} and then in the remainder of this section we show how to bound the requisite quantities for solving linear systems in $\mvar{b}$.
\newcommand{\overline{S}}{\overline{S}}
\begin{theorem}[SVRG for Sums of Non-Convex Functions]
\label{lem:svrg-nonconv}
Consider a set of functions, $\{\psi_{1}, \psi_2,...\psi_n \}$, each mapping $\mathbb{R}^{d}\rightarrow\mathbb{R}$.
Let $f(x) = \sum_{i}\psi_{i}(x)$ and let $\opt{x} \mathbin{\stackrel{\rm def}{=}} \argmin_{x \in \mathbb{R}^{d}} f(x)$. Suppose we have a probability distribution $p$ on $[n]$,
and that starting from some initial point $x_{0} \in \mathbb{R}^d$ in each iteration $k$ we pick $i_{k} \in [n]$ independently with probability $p_{i_k}$ and let
\[
x_{k+1}
:=
x_{k}
-\frac{\eta}{p_{i}}
\left(\gradient\psi_{i}(x_k)-\gradient\psi_{i}(x_0 )\right) + \eta \gradient f(x_0)
\]
for some $\eta$. If $f$ is $\mu$-strongly convex and if for all $x \in \mathbb{R}^d$
we have
\begin{equation}
\label{eq:svrg-nonconvex-avgsmooth}
\sum_{i \in [n]} \frac{1}{p_{i}} \norm{\gradient \psi_{i}(x) - \gradient\psi_{i}(\opt{x})}_{2}^{2}
\leq
2 \overline{S} \left[f(x) - f(\opt{x}) \right],
\end{equation}
where $\overline{S}$ is a variance parameter,
then for all $m \geq 1$ we have
\[
\mathbb{E} \left[\frac{1}{m} \sum_{k\in[m]} f(x_{k}) - f(\opt{x})\right] \leq \frac{1}{1-2\eta\bar{S}}
\left[\frac{1}{\mu\eta m}+2\eta\overline{S} \right] \cdot
\left[f(x_{0})-f(\opt{x})\right]
\]
Consequently, if we pick $\eta$ to be a sufficiently small multiple of $1/\bar{S}$
then when $m = O(\overline{S} / \mu)$ we can decrease the error by
a constant multiplicative factor in expectation.
\end{theorem}
\begin{proof}
We first note that $\mathbb{E}_{i_{k}}[x_{k+1} - x_{k}] = \eta \gradient f(x_{k})$. This is, in each iteration, in expectation, we make a step in the direction of the gradient. Using this fact we have:
\begin{align*}
\mathbb{E}_{i_{k}} \norm{x_{k+1} - \opt{x}}_{2}^{2} &= \mathbb{E}_{i_{k}} \norm{(x_{k+1} -x_k) + (x_k - \opt{x}) }_{2}^{2} \\
&= \norm{x_k - \opt{x} }_{2}^{2} - 2 \mathbb{E}_{i_{k}}(x_{k+1} -x_k)^\top(x_k - \opt{x}) + \mathbb{E}_{i_{k}}\norm{x_{k+1} - x_k }_{2}^{2} \\
&=\norm{x_k - \opt{x} }_{2}^{2} - 2 \eta \gradient f(x_k)^{\top} \left(x_{k} - \opt{x} \right)\\
&+ \sum_{i \in [n]} \eta^{2} p_{i} \normFull{
\frac{1}{p_{i}}
\left(\gradient \psi_{i}(x_{k}) - \gradient\psi_{i}(x_{0}) \right)
+ \gradient f(x_{0})}_{2}^{2}
\end{align*}
We now apply the fact that $\norm{x + y}_2^2 \leq 2\norm{x}_2^2 + 2\norm{y}_2^2$ to give:
\begin{align*}
\sum_{i \in [n]}
& p_i \normFull{\frac{1}{p_i}
\left(\gradient \psi_i (x_k) - \gradient \psi_i (x_0)\right) + \gradient f(x_0)
}_2^2
\\
&
\leq
\sum_{i \in [n]}
2 p_i \normFull{\frac{1}{p_i}
\left(\gradient \psi_i (x_k) - \gradient \psi_i (\opt{x}) \right)
}_2^2
+
\sum_{i \in [n]}
2 p_i \normFull{\frac{1}{p_i}
\left(\gradient \psi_i (x_0) - \gradient \psi_i(\opt{x}) \right) - \gradient f(x_0)
}_2^2.
\end{align*}
Then, using that $\gradient f(\opt{x}) = 0$ by optimality, that $\mathbb{E}\norm{x-\mathbb{E} x}_{2}^{2} \leq \mathbb{E}\norm{x}_{2}^{2}$, and \eqref{eq:svrg-nonconvex-avgsmooth} we have:
\begin{align*}
\sum_{i \in [n]}
& p_i \normFull{\frac{1}{p_i}
\left(\gradient \psi_i (x_k) - \gradient \psi_i (x_0)\right) + \gradient f(x_0)
}_2^2
\\
&
\le
\sum_{i \in [n]}
\frac{2}{p_i}
\normFull{\gradient \psi_i (x_k) - \gradient \psi_i (\opt{x})
}_2^2
+
\sum_{i \in [n]}
2 p_i \normFull{\frac{1}{p_i}
\left(\gradient \psi_i (x_0) - \gradient \psi_i(\opt{x})) - (\gradient f(x_0) - \gradient f(\opt{x})\right)
}_2^2
\\
&
\leq
\sum_{i \in [n]}
\frac{2}{p_i}
\normFull{\gradient \psi_i (x_k) - \gradient \psi_i (\opt{x})
}_2^2
+
\sum_{i \in [n]}
2 p_i \normFull{\frac{1}{p_i}
\gradient \psi_i (x_0) - \gradient \psi_i(\opt{x}))
}_2^2
\\
&
\leq 4 \overline{S}
\left[ f(x_{k})-f(\opt{x}) + f(x_{0}) - f(\opt{x}) \right]
\end{align*}
Since $f(\opt{x}) - f(x_k) \geq \gradient f(x_k)^\top (\opt{x} - x_k)$ by the convexity of $f$, these inequalities imply
\begin{align*}
\mathbb{E}_{i_{k}}\norm{x_{k+1} - \opt{x}}_{2}^{2}
&
\leq\norm{x_{k} - \opt{x}}_{2}^{2}
- 2 \eta \left[f(x_{k}) - f(\opt{x})\right]
+ 4 \eta^{2} \overline{S} \left[f(x_{k}) - f(\opt{x}) + f(x_{0}) - f(\opt{x})\right]
\\
&
= \norm{x_{k} - \opt{x}}_{2}^{2}-2\eta(1-2\eta S)\left(f(x_{k})-f(\opt{x})\right)+4\eta^{2}\bar{S}\left(f(x_{0})-f(\opt{x})\right)
\end{align*}
Rearranging, we have:
\begin{align*}
2\eta(1-2\eta S)\left(f(x_{k})-f(\opt{x})\right) \le \norm{x_{k} - \opt{x}}_{2}^{2} - \mathbb{E}_{i_{k}}\norm{x_{k+1} - \opt{x}}_{2}^{2} + 4\eta^{2}\bar{S}\left(f(x_{0})-f(\opt{x})\right).
\end{align*}
And summing over all iterations and taking expectations we have:
\begin{align*}
\mathbb{E} \left [2\eta(1-2\eta \bar{S})\sum_{k\in[m]}f(x_{k})-f(\opt{x}) \right ] \le \norm{x_0 - \opt{x}}_2^2 + 4m\eta^{2}\bar{S}\left[f(x_{0})-f(\opt{x})\right].
\end{align*}
Finally, we use that
by strong convexity, $\norm{x_{0}-\opt{x}}_{2}^{2}\leq\frac{2}{\mu}\left(f(x_{0})-f(\opt{x})\right)$ to obtain:
\[
\mathbb{E} \left [2\eta(1-2\eta \bar{S})\sum_{k\in[m]}f(x_{k})-f(\opt{x})\right ]\leq \frac{2}{\mu}\left[f(x_{0})-f(\opt{x})\right]+4m\eta^{2}\bar{S}\left[f(x_{0})-f(\opt{x})\right]
\]
and thus
\[
\mathbb{E} \left [\frac{1}{m}\sum_{k\in[m]}f(x_{k})-f(\opt{x}) \right ]\leq\frac{1}{1-2\eta\bar{S}}\left[\frac{1}{\mu\eta m}+2\eta\bar{S}\right]\cdot\left[f(x_{0})-f(\opt{x})\right]
\]
\end{proof}
Theorem \ref{lem:svrg-nonconv} immediately yields a solver for $\bv{B}x = b$. Finding the minimum norm solution to this system is equivalent to minimizing $f(x) = \frac{1}{2} x^\top \bv{B} x - b^\top x$. If we take the common approach of applying a smoothness bound for each $\psi_i$ along with a strong convexity bound on $f(x)$ we obtain:
\begin{lemma}[Simple Variance Bound for SVRG]\label{simple_variance_bound}
Let
\begin{align*}
\psi_i(x) \defeq \frac{1}{2} x^\top \left(\frac{\lambda\norm{a_i}_2^2}{\norm{\bv A}_F^2} \mI - a_i a_i^\top \right) x
- \frac{1}{n}b^\top x
\end{align*}
so we have $\sum_{i\in[n]} \psi_i(x) = f(x) = \frac{1}{2} x^\top \bv{B} x - b^\top x$.
Setting $p_i = \frac{\norm{a_i}_2^2}{\norm{\bv A}_F^2}$ for all $i$, we have
\begin{align*}
\sum_{i \in [n]} \frac{1}{p_i} \norm{\gradient \psi_i(x)-\gradient \psi_i(\opt{x})}_2^2 = O \left (
\frac{\norm{\bv A}_F^4}{\lambda - \lambda_1}
\left[f(x) - f(\opt{x})\right] \right )
\end{align*}
\end{lemma}
\begin{proof}
We first compute, for all $i \in [n]$
\begin{align}\label{basic_gradient_computation}
\gradient \psi_i(x) =
\left(\frac{\lambda \norm{a_i}_2^2}{\normFro{\mvar{a}}^2} \mI
- a_i a_i^\top \right) x
- \frac{1}{n}b.
\end{align}
We have that each $\psi_i$ is $\frac{\lambda \norm{a_i}_2^2}{\normFro{\mvar{a}}^2} + \norm{a_i}^2$ smooth with respect to $\norm{\cdot}_2$. Specifically,
\begin{align*}
\norm{\gradient \psi_i(x)-\gradient \psi_i(\opt{x})}_2 &= \norm {\left (\frac{\lambda \norm{a_i}_2^2}{\normFro{\mvar{a}}^2} \mI
- a_i a_i^\top \right) (x-\opt{x}) }_2 \\
&\le \left (\frac{\lambda \norm{a_i}_2^2}{\normFro{\mvar{a}}^2} + \norm{a_i}^2 \right ) \norm{x-\opt{x}}_2.
\end{align*}
Additionally,
$f(x)$ is $\lambda_d(\bv B) = \lambda - \lambda_1$ strongly convex so we have $\norm{x - \opt{x}}_2^2 \leq \frac{2}{\lambda - \lambda_1} \left[f(x) - f(\opt{x})\right]$ and putting all this together we have
\begin{align*}
\sum_{i \in [n]} \frac{1}{p_i} \norm{\gradient \psi_i(x) - \gradient \psi_i (\opt{x})}_2^2
&\leq
\sum_{i \in [n]} \frac{\norm{\bv A}_F^2}{\norm{a_i}_2^2}
\cdot
\norm{a_i}_2^4 \left(\frac{\lambda}{\norm{\bv A}_F^2} + 1\right)^2
\cdot \frac{2}{\lambda - \lambda_1}
\left[f(x) - f(\opt{x})\right]
\\
&= O \left (
\frac{\norm{\bv A}_F^4}{\lambda - \lambda_1}
\left[f(x) - f(\opt{x})\right] \right )
\\
\end{align*}
where the last step uses that $\lambda \le 2\lambda_1 \le 2\norm{\bv A}_F^2$ so $\frac{\lambda}{\norm{\bv A}_F^2} \le 2$.
\end{proof}
Assuming that $\lambda = (1+c \cdot \mathrm{gap})\lambda_1$ for some constant $c$, the above bound means that we can make constant progress on our linear system by setting $m = O(\overline{S}/\mu) = O\left(\frac{\norm{\bv A}^4_F}{(\lambda-\lambda_1)^2}\right)= O\left ( \frac{\mathsf{r}(\bv{A})^2}{\mathrm{gap}^2} \right )$. This dependence on stable rank matches the dependence given in \cite{shamir2015stochastic} (see discussion in Section \ref{previous_work_offline}), however we can show that it is suboptimal.
We show to improve the bound to $O\left ( \frac{\mathsf{r}(\bv{A})}{\mathrm{gap}^2} \right )$ by using a better variance analysis. Instead of bounding each $\norm{\gradient \psi_i(x) - \gradient \psi_i (\opt{x})}_2^2$ term using the smoothness of $\psi_i$, we more carefully bound the sum of these terms.
\begin{lemma}(Improved Variance Bound for SVRG)\label{variance_lemma} For $i \in [n]$ let
\begin{align*}
\psi_i(x) \defeq \frac{1}{2} x^\top \left(\frac{\lambda\norm{a_i}_2^2}{\norm{\bv A}_F^2} \mI - a_i a_i^\top \right) x
- \frac{1}{n}b^\top x
\end{align*}
so we have $\sum_{i\in[n]} \psi_i(x) = f(x) = \frac{1}{2} x^\top \bv{B} x - b^\top x$.
Setting $p_i = \frac{\norm{a_i}_2^2}{\norm{\bv A}_F^2}$ for all $i$, we have for all $x$
\[
\sum_{i \in [n]}
\frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x})}_2^2
\leq
\frac{4\lambda_1 \normFro{\mvar{a}}^2}{\lambda - \lambda_1} \cdot \left[f(x) - f(\opt{x}) \right].
\]
\end{lemma}
\begin{proof}
Using the gradient computation in \eqref{basic_gradient_computation} we have
\begin{align}
\sum_{i \in [n]}
\frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x})}_2^2 &= \sum_{i \in [n]}
\frac{\normFro{\mvar{a}}^2}{\norm{a_i}_2^2}
\normFull{\left( \frac{\lambda\norm{a_i}_2^2}{\normFro{\mvar{a}}^2} \mI
- a_i a_i^\top \right) (x - \opt{x})}_2^2\nonumber\\
&=
\sum_{i \in [n]}
\frac{\lambda^2 \norm{a_i}_2^2}{\normFro{\mvar{a}}^2}
\normFull{x - \opt{x}}^2_2
- 2 \sum_{i \in [n]} \lambda \normFull{x - \opt{x}}_{a_i a_i^\top}^2
\nonumber\\&+ \sum_{i \in [n]} \frac{\normFro{\mvar{a}}^2}{\norm{a_i}^2} \normFull{x - \opt{x}}_{\norm{a_i}_2^2 a_i a_i^\top}^2
\nonumber\\
&=
\lambda^2 \normFull{x - \opt{x}}_2^2
- 2 \lambda \normFull{x - \opt{x}}_{\mvar{\Sigma}}^2
+ \normFro{\mvar{a}}^2 \normFull{x - \opt{x}}_{\mvar{\Sigma}}^2.\nonumber\\
&\le
\lambda \normFull{x - \opt{x}}_{\bv B}^2
+ \normFro{\mvar{a}}^2 \normFull{x - \opt{x}}_{\mvar{\Sigma}}^2.\label{broken_up_norms}
\end{align}
Now since
\[
\mvar{\Sigma} \preceq \lambda_1 \mI \preceq \frac{\lambda_1}{\lambda - \lambda_1} \mvar{b}
\]
we have
\begin{align*}
\sum_{i \in [n]}
\frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x})}_2^2 &\le \left (\frac{ \lambda(\lambda-\lambda_1) +\normFro{\mvar{a}}^2\cdot \lambda_1}{\lambda - \lambda_1} \right )\normFull{x - \opt{x}}_{\mvar{b}}^2\\
&\le \left (\frac{2\normFro{\mvar{a}}^2 \lambda_1}{\lambda - \lambda_1} \right )\normFull{x - \opt{x}}_{\mvar{b}}^2
\end{align*}
where in the last inequality we just coarsely bound $\lambda(\lambda-\lambda_1) \le \lambda_1\norm{\bv A}_F^2$.
Now since $\bv{B}$ is full rank, $\bv{B}\opt{x} = b$, we can compute:
\begin{align}\label{norm_to_function_error_conversion}
\norm{x - \opt{x}}_{\mvar{b}}^2 = x^\top \mvar{b} x - 2b^\top x + b^\top \opt{x} = 2[f(x) - f(\opt{x})].
\end{align}
The result follows.
\end{proof}
Plugging the bound in Lemma \ref{variance_lemma} into Theorem \ref{lem:svrg-nonconv} we have:
\begin{theorem}(Offline SVRG-Based Solver)\label{offline_solver}
Let $\overline{S} =\frac{2\lambda_1 \normFro{\mvar{\Sigma}}^2}{\lambda - \lambda_1}$, $\mu =\lambda-\lambda_1$. The iterative procedure described in Theorem \ref{lem:svrg-nonconv} with $f(x) = \frac{1}{2}x^\top \bv{B} x - b^\top x$, $\psi_i(x) = \frac{1}{2} x^\top \left(\frac{\lambda\norm{a_i}_2^2}{\norm{\bv \Sigma}_F^2} \mI - a_i a_i^\top \right) x - b^\top x$, $p_i = \frac{\norm{a_i}_2^2}{\norm{\bv \Sigma}_F^2}$, $\eta = 1/(8\overline{S})$ and $m$ chosen uniformly at random from $[64 \overline{S}/\mu]$ returns a vector $x_m$ such that
\begin{align*}
\mathbb{E} \norm{x_m - \opt{x}}_\mvar{b}^2 \le \frac{1}{2} \norm{x_0-\opt{x}}_\mvar{b}^2.
\end{align*}
Further, assuming $\left (1 + \frac{\mathrm{gap}}{150} \right ) \lambda_1< \lambda \le \left (1 + \frac{\mathrm{gap}}{100} \right ) \lambda_1$, this procedure runs in time $O\left (\nnz(\bv A) + \frac{d\cdot \mathsf{r}(\bv{A})}{\mathrm{gap}^2} \right )$.
\end{theorem}
\begin{proof}
Lemma \ref{variance_lemma} tells us that \[
\sum_{i \in [n]}
\frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x})}_2^2
\leq
2 \overline{S} \left[f(x) - f(\opt{x}) \right].
\]
Further $f(x) = \frac{1}{2}x^\top \bv{B} x - b^\top x$ is $\lambda_d(\bv B)$-strongly convex and $\lambda_d(\bv B) = \lambda-\lambda_1 = \mu$.
Plugging this into Theorem \ref{lem:svrg-nonconv} and using \eqref{norm_to_function_error_conversion} which shows $\norm{x-\opt{x}}_\mvar{b}^2 = 2[f(x)-f(\opt{x})]$ we have, for $m$ chosen uniformly from $[64 \overline{S}/ \mu]$:
\begin{align*}
\mathbb{E} \left[\frac{1}{64 \overline{S}/ \mu} \sum_{k\in[64 \overline{S}/ \mu]} f(x_{k}) - f(\opt{x})\right] &\leq 4/3 \cdot
\left[1/8+1/8 \right] \cdot
\left[f(x_{0})-f(\opt{x})\right]\\
\mathbb{E} \left [f(x_m) - f(\opt{x}) \right ] &\le \frac{1}{2}\left[f(x_{0})-f(\opt{x})\right] \\
\mathbb{E} \norm{x_m - \opt{x}}_\mvar{b}^2 &\le \frac{1}{2} \norm{x_0-x^{opt}}_\mvar{b}^2.
\end{align*}
The procedure requires $O\left(\nnz(\bv A)\right)$ time to initially compute $\gradient f(x_0)$, along with each $p_i$ and the step size $\eta$ which depend on $\norm{\bv A}_F^2$ and the row norms of $\bv{A}$. Each iteration then just requires $O(d)$ time to compute $\gradient \psi_i(\cdot)$ and perform the necessary vector operations. Since there are at most $[64 \overline{S}/ \mu] = O\left ( \frac{\lambda_1\norm{\bv A}_F^2}{(\lambda-\lambda_1)^2} \right )$ iterations, our total runtime is
\begin{align*}
O \left (\nnz(\bv{A}) + d \cdot \frac{\lambda_1\norm{\bv A}_F^2}{(\lambda-\lambda_1)^2} \right ) = O\left (\nnz(\bv A) + \frac{d\cdot \mathsf{r}(\bv{A})}{\mathrm{gap}^2} \right ).
\end{align*}
Note that if our matrix is uniformly sparse - i.e. all rows have sparsity at most $d_s$, then the runtime is actually at most $O\left (\nnz(\bv A) + \frac{d_s \cdot \mathsf{r}(\bv{A})}{\mathrm{gap}^2} \right )$.
\end{proof}
\subsection{Accelerated Solver}
\label{sec:offline:acceleration}
Theorem \ref{offline_solver} gives a linear solver for $\bv{B}$ that makes progress in expectation and which we can plug into Theorems \ref{thm:powermethod-perturb} and \ref{thm:init-offline}. However, we first show that the runtime in Theorem \ref{offline_solver} can be accelerated in some cases. We apply a result of \cite{frostig2015regularizing}, which shows that, given a solver for a regularized version of a convex function $f(x)$, we can produce a fast solver for $f(x)$ itself. Specifically:
\begin{lemma}[Theorem 1.1 of \cite{frostig2015regularizing}]\label{acceleration_primitive}
Let $f(x)$ be a $\mu$-strongly convex function and let $\opt{x} \mathbin{\stackrel{\rm def}{=}} \argmin_{x\in\mathbb{R}^d} f(x)$. For any $\gamma > 0$ and any $x_0 \in \mathbb{R}^d$, let $f_{\gamma,x_0}(x) \mathbin{\stackrel{\rm def}{=}} f(x) + \frac{\gamma}{2} \norm{x-x_0}_2^2$. Let $\opt{x}_{\gamma,x_0} \mathbin{\stackrel{\rm def}{=}} \argmin_{x\in\mathbb{R}^d} f_{\gamma,x_0} (x)$.
Suppose that, for all $x_0 \in \mathbb{R}^d$, $c > 0$, $\gamma > 0$, we can compute a point $x_c$ such that
\begin{align*}
\mathbb{E} f_{\gamma,x_0}(x_c) - f_{\gamma,x_0}(\opt{x}_{\gamma,x_0} ) \le \frac{1}{c} \left [f_{\gamma,x_0} - f_{\gamma,x_0}(\opt{x}_{\gamma,x_0} ) \right ]
\end{align*}
in time $\mathcal{T}_c$. Then given any $x_0$, $c > 0$, $\gamma > 2 \mu$, we can compute $x_1$ such that
\begin{align*}
\mathbb{E} f(x_1) - f(\opt{x}) \le \frac{1}{c} \left [f(x_0) - f(\opt{x}) \right ]
\end{align*}
in time $O\left(\mathcal{T}_{4\left (\frac{2\gamma + \mu}{\mu} \right )^{3/2}} \sqrt{\lceil \gamma/\mu \rceil} \log c \right ).$
\end{lemma}
We first give a new variance bound on solving systems in $\bv{B}$ when a regularizer is used. The proof of this bound is very close to the proof given for the unregularized problem in Lemma \ref{variance_lemma}.
\begin{lemma}\label{regularized_variance_lemma} For $i \in [n]$ let
\begin{align*}
\psi_i(x) \defeq \frac{1}{2} x^\top \left(\frac{\lambda\norm{a_i}_2^2}{\norm{\bv A}_F^2} \mI - a_i a_i^\top \right) x
- \frac{1}{n}b^\top x + \frac{\gamma\norm{a_i}_2^2}{2\norm{\bv A}_F^2} \norm{x - x_0}_2^2
\end{align*}
so we have $\sum_{i\in[n]} \psi_i(x) = f_{\gamma,x_0} (x) = \frac{1}{2} x^\top \bv{B} x - b^\top x + \frac{\gamma}{2}\norm{x - x_0}_2^2$.
Setting $p_i = \frac{\norm{a_i}_2^2}{\norm{\bv A}_F^2}$ for all $i$, we have for all $x$
\[
\sum_{i \in [n]}
\frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x}_{\gamma,x_0} )}_2^2
\leq
\left ( \frac{\gamma^2 + 12\lambda_1\norm{\bv A}_F^2}{\lambda-\lambda_1+\gamma}\right ) \left [f_{\gamma,x_0}(x)-f_{\gamma,x_0}(\opt{x}_{\gamma,x_0}) \right ]
\]
\end{lemma}
\begin{proof}
We have for all $i \in [n]$
\begin{align}\label{regularized_gradient_computation}
\gradient \psi_i(x) =
\left(\frac{\lambda \norm{a_i}_2^2}{\normFro{\mvar{a}}^2} \mI
- a_i a_i^\top \right) x
- \frac{1}{n}b + \frac{\gamma\norm{a_i}_2^2}{2\norm{\bv A}_F^2} (x-2x_0)
\end{align}
Plugging this in we have:
\begin{align*}
\sum_{i \in [n]}
\frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x}_{\gamma,x_0} )}_2^2 &= \sum_{i \in [n]}
\frac{\normFro{\mvar{a}}^2}{\norm{a_i}_2^2}
\normFull{\left( \frac{\lambda\norm{a_i}_2^2}{\normFro{\mvar{a}}^2} \mI
- a_i a_i^\top \right) (x - \opt{x}_{\gamma,x_0} ) + \frac{\gamma\norm{a_i}_2^2}{2\norm{\bv A}_F^2} (x - \opt{x}_{\gamma,x_0} )}_2^2
\end{align*}
For simplicity we now just use the fact that $\norm{x+y}_2^2 \le 2\norm{x}_2^2 + 2\norm{y}_2^2$ and apply our bound from equation \eqref{broken_up_norms} to obtain:
\begin{align*}
\sum_{i \in [n]}
\frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x}_{\gamma,x_0} )}_2^2 &\le
2\lambda^2 \normFull{x - \opt{x}_{\gamma,x_0}}_2^2
- 4 \lambda \normFull{x - \opt{x}_{\gamma,x_0}}_{\mvar{\Sigma}}^2
+ 2\normFro{\mvar{\Sigma}}^2 \normFull{x - \opt{x}_{\gamma,x_0}}_{\mvar{\Sigma}}^2 \\&+ 2\sum_{i \in [n]}\frac{\norm{a_i}_2^2}{\normFro{\mvar{a}}^2} \frac{\gamma^2}{4} \norm{x-\opt{x}_{\gamma,x_0} }_2^2\\
&\le
\left (2\lambda^2 +\gamma^2/2 + 2\lambda_1\norm{\bv A}_F^2 - 4\lambda_1\lambda \right ) \normFull{x - \opt{x}_{\gamma,x_0} }_{2}^2\\
&\le
\left (\gamma^2/2 + 6\lambda_1\norm{\bv A}_F^2 \right ) \normFull{x - \opt{x}_{\gamma,x_0} }_{2}^2
\end{align*}
Now, $f_{\gamma,x_0}(\cdot)$ is $\lambda-\lambda_1+\gamma$ strongly convex, so
\begin{align*}
\normFull{x - \opt{x}_{\gamma,x_0} }_{2}^2 \le \frac{2}{\lambda-\lambda_1+\gamma} \left [f_{\gamma,x_0}(x)-f_{\gamma,x_0}(\opt{x}_{\gamma,x_0}) \right ].
\end{align*}
So overall we have:
\begin{align*}
\sum_{i \in [n]} \frac{1}{p_i} \normFull{\gradient \psi_i(x) - \gradient \psi_i(\opt{x}_{\gamma,x_0} )}_2^2 &\le \left ( \frac{\gamma^2 + 12\lambda_1\norm{\bv A}_F^2}{\lambda-\lambda_1+\gamma}\right ) \left [f_{\gamma,x_0}(x)-f_{\gamma,x_0}(\opt{x}_{\gamma,x_0}) \right ]
\end{align*}
\end{proof}
We can now use this variance bound to obtain an accelerated solver for $\bv{B}$. We assume $\nnz(\bv{A}) \le \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2}$, as otherwise, the unaccelerated solver in Theorem \ref{offline_solver} runs in $O(\nnz(\bv{A}))$ time and cannot be accelerated further.
\begin{theorem}[Accelerated SVRG-Based Solver]\label{accelerated_offline_solver}
Assuming $\left (1 + \frac{\mathrm{gap}}{150} \right ) \lambda_1< \lambda \le \left (1 + \frac{\mathrm{gap}}{100} \right ) \lambda_1$ and $nnz(\bv{A}) \le \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2}$, applying the iterative procedure described in Theorem \ref{lem:svrg-nonconv} along with the acceleration given by Lemma \ref{acceleration_primitive} gives a solver that returns $x$ with
\begin{align*}
\mathbb{E} \norm{x - \opt{x}}_\mvar{b}^2 \le \frac{1}{2} \norm{x_0-\opt{x}}_\mvar{b}^2.
\end{align*}
in time $O\left ( \frac{\nnz(\bv A)^{3/4} (d \mathsf{r}(\bv A))^{1/4}}{\sqrt{\mathrm{gap}}} \cdot \log \left (\frac{d}{\mathrm{gap}}\right)\right)$.
\end{theorem}
\begin{proof}
Following Theorem \ref{offline_solver}, the variance bound of Lemma \ref{regularized_variance_lemma} means that we can make constant progress in minimizing $f_{\gamma,x_0}(x)$ in $O\left (\nnz(\bv{A}) + d m \right)$ time where $m = O \left ( \frac{\gamma^2 + 12\lambda_1\norm{\bv\Sigma}_F^2}{(\lambda-\lambda_1+\gamma)^2}\right )$. So, for $\gamma \ge 2(\lambda-\lambda_1)$ we can make $4\left (\frac{2\gamma + (\lambda-\lambda_1)}{\lambda-\lambda_1} \right )^{3/2}$ progress, as required by Lemma \ref{acceleration_primitive} in time $O\left (\left (\nnz(\bv A) + dm \right) \cdot \log \left (\frac{\gamma}{\lambda-\lambda_1}\right) \right )$ time. Hence by Lemma \ref{acceleration_primitive} we can make constant factor expected progress in minimizing $f(x)$ in time:
\begin{align*}
O\left ( \left (\nnz(\bv A) + d\frac{\gamma^2 + 12\lambda_1\norm{\bv A}_F^2}{(\lambda-\lambda_1+\gamma)^2} \right ) \log \left (\frac{\gamma}{\lambda-\lambda_1}\right) \sqrt {\frac{\gamma}{\lambda-\lambda_1}}\right)
\end{align*}
By our assumption, we have $\nnz(\bv{A}) \le \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2} = \frac{d \lambda_1 \norm{\bv A}_F^2}{(\lambda-\lambda_1)^2}$. So, if we let $\gamma = \Theta \left (\sqrt{\frac{d\lambda_1 \norm{\mvar{a}}_F^2}{\nnz(\bv A)}}\right)$ then using a sufficiently large constant, we have $\gamma \ge 2(\lambda-\lambda_1)$. We have $\frac{\gamma}{\lambda-\lambda_1} = \Theta \left (\sqrt{\frac{d\lambda_1 \norm{\bv A}_F^2}{\nnz(\bv A)\lambda_1^2 \mathrm{gap}^2}}\right) = \Theta \left (\sqrt{\frac{d\mathsf{r}(\bv A)}{\nnz(\bv A)\mathrm{gap}^2}}\right) $ and can make constant expected progress in minimizing $f(x)$ in time:
\begin{align*}
O\left ( \frac{\nnz(\bv A)^{3/4} (d \mathsf{r}(\bv A))^{1/4}}{\sqrt{\mathrm{gap}}} \cdot \log \left (\frac{d}{\mathrm{gap}}\right)\right).
\end{align*}
\end{proof}
\subsection{Shifted-and-Inverted Power Method}
\label{sec:offline:results}
Finally, we are able to combine the solvers from Sections \ref{sec:offline:svrg} and \ref{sec:offline:acceleration} with the framework of Section \ref{framework} to obtain faster algorithms for top eigenvector computation.
\begin{theorem}[Shifted-and-Inverted Power Method With SVRG]\label{main_offline_theorem}
Let $\bv{B} = \lambda \bv{I} - \bv{A}^\top \bv{A}$ for $\left ( 1+\frac{\mathrm{gap}}{150}\right) \lambda_1 \le \lambda \le \left(1+ \frac{\mathrm{gap}}{100} \right ) \lambda_1$ and let $x_0 \sim \mathcal{N}(0,\bv I)$ be a random initial vector. Running the inverted power method on $\bv{B}$ initialized with $x_0$, using the SVRG solver from Theorem \ref{offline_solver} to approximately apply $\bv{B}^{-1}$ at each step, returns $x$ such that with probability $1-O\left (\frac{1}{d^{10}}\right)$, $x^\top \bv{\Sigma}x \ge (1-\epsilon) \lambda_1$ in total time $$O \left (\left(\nnz(\bv A) + \frac{d \mathsf{r}(\bv A)}{\mathrm{gap}^2} \right )\cdot \left (\log^2\left(\frac{d}{\mathrm{gap}}\right) + \log\left(\frac{1}{\epsilon}\right) \right ) \right ).$$
\end{theorem}
Note that by instantiating the above theorem with $\epsilon' = \epsilon\cdot \mathrm{gap}$, and applying Lemma \ref{lem:ray_to_evec} we can find a unit vector $x$ such that $| v_1^\top x | \ge 1-\epsilon$ in the same asymptotic running time (an extra $\log(1/\mathrm{gap})$ term is absorbed into the $\log^2(d/\mathrm{gap})$ term).
\begin{proof}
By Theorem \ref{thm:init-offline}, if we start with $x_0 \sim \mathcal{N}(0,\bv I)$ we can run $O\left (\log \left (\frac{d}{\mathrm{gap}} \right ) \right )$ iterations of the inverted power method, to obtain $x_1$ with $G(x_1) \le \frac{1}{\sqrt{10}}$ with probability $1-O\left (\frac{1}{d^{10}}\right)$. Each iteration requires applying an linear solver that decreases initial error in expectation by a factor of $\frac{1}{\mathop\mathrm{poly}(d,1/\mathrm{gap})}$. Such a solver is given by applying the solver in Theorem \ref{offline_solver} $O\left (\log \left (\frac{d}{\mathrm{gap}} \right ) \right )$ times, decreasing error by a constant factor in expectation each time. So overall in order to find $x_1$ with $G(x_1) \le \frac{1}{\sqrt{10}}$, we require time $O \left (\left(\nnz(\bv A) + \frac{d \mathsf{r}(\bv A)}{\mathrm{gap}^2} \right )\cdot\log^2\left(\frac{d}{\mathrm{gap}}\right)\right )$.
After this initial `burn-in' period we can apply Corollary \ref{cor:constant_factor_corollary} of Theorem \ref{thm:powermethod-perturb}, which shows that running a single iteration of the inverted power method will decrease $G(x)$ by a constant factor in expectation. In such an iteration, we only need to use a solver that decreases initial error by a constant factor in expectation. So we can perform each inverted power iteration in this stage in time $O \left (\nnz(\bv A) + \frac{d \mathsf{r}(\bv A)}{\mathrm{gap}^2} \right ).$
With $O\left (\log \left (\frac{d}{\epsilon}\right)\right)$ iterations, we can obtain $x$ with $\mathbb{E} \left [G(x)^2 \right ] = O\left (\frac{\epsilon}{d^{10}}\right)$ So by Markov's inequality, we have $G(x)^2 = O(\epsilon)$, giving us $x^T \bv{\Sigma} x \ge (1-O(\epsilon))\lambda_1$ by Lemma \ref{lem:rayquot-potential}. Union bounding over both stages gives us failure probability $O\left ( \frac{1}{d^{10}}\right )$, and adding the runtimes from the two stages gives us the final result. Note that the second stage requires $O\left (\log \left (\frac{d}{\epsilon}\right)\right) = O(\log d + \log (1/\epsilon))$ iterations to achieve the high probability bound. However, the $O(\log d)$ term is smaller than the $O\left (\log^2 \left (\frac{d}{\mathrm{gap}} \right ) \right )$ term, so is absorbed into the asymptotic notation.
\end{proof}
We can apply an identical analysis using the accelerated solver from Theorem \ref{accelerated_offline_solver}, obtaining the following runtime which beats Theorem \ref{main_offline_theorem} whenever $\nnz(\bv A) \le \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2}$:
\begin{theorem}[Shifted-and-Inverted Power Method Using Accelerated SVRG]\label{accelerated_offline_theorem}
Let $\bv{B} = \lambda \bv{I} - \bv{A}^\top \bv{A}$ for $\left ( 1+\frac{\mathrm{gap}}{150}\right) \lambda_1 \le \lambda \le \left(1+ \frac{\mathrm{gap}}{100} \right ) \lambda_1$ and let $x_0 \sim \mathcal{N}(0,\bv I)$ be a random initial vector. Assume that $\nnz(\bv A) \le \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2}$. Running the inverted power method on $\bv{B}$ initialized with $x_0$, using the accelerated SVRG solver from Theorem \ref{accelerated_offline_solver} to approximately apply $\bv{B}^{-1}$ at each step, returns $x$ such that with probability $1-O\left (\frac{1}{d^{10}}\right)$, $| v_1^\top x | \ge 1-\epsilon$ in total time $$O \left (\left(\frac{\nnz(\bv A)^{3/4}(d \mathsf{r}(\bv A))^{1/4}}{\sqrt{\mathrm{gap}}} \right )\cdot \left (\log^3\left(\frac{d}{\mathrm{gap}}\right) + \log\left(\frac{d}{\mathrm{gap}}\right)\log\left(\frac{1}{\epsilon}\right) \right ) \right ).$$
\end{theorem}
\section{Online Eigenvector Computation}\label{sec:online}
Here we show how to apply the shifted-and-inverted power method framework of Section \ref{framework} to the online setting. This setting is more difficult than the offline case. As there is no canonical matrix $\bv{A}$, and we only have access to the distribution $\mathcal{D}$ through samples, in order to apply Theorem \ref{thm:powermethod-perturb}
we must show
how to both estimate the Rayleigh quotient (Section~\ref{sec:online:rayleigh_quotient}) as well as solve the requisite linear systems in expectation (Section~\ref{sec:online:system-solver}).
After laying this ground work, our main result is given in Section \ref{sec:online:result}.
Ultimately, the results in this section allow us to achieve more efficient algorithms for computing the top eigenvector in the statistical setting as well as improve upon the previous best known sample complexity for top eigenvector computation. As we show in Section~\ref{sec:lower} the bounds we provide in this section are in fact tight for general distributions.
\subsection{Estimating the Rayleigh Quotient}
\label{sec:online:rayleigh_quotient}
Here we show how to estimate the Rayleigh quotient of a vector with respect to $\mvar{\Sigma}$. Our analysis is standard -- we first approximate the Rayleigh quotient by its empirical value on a batch of $k$ samples and prove using Chebyshev's inequality that the error on this sample is small with constant probability. We then repeat this procedure $O(\log (1/p))$ times and output the median. By a Chernoff bound this yields a good estimate with probability $1 - p$. The formal statement of this result and its proof comprise the remainder of this subsection.
\begin{theorem}[Online Rayleigh Quotient Estimation]\label{online_rayleigh_estimation}
Given $\epsilon \in (0,1]$, $p \in [0, 1]$, and unit vector $x$ set $k = \lceil 4\mathsf{v}(\mathcal{D}) \epsilon^{-2} \rceil$ and $m = O(\log(1/p))$.
For all $i \in [k]$ and $j \in [m]$ let $a_{i}^{(j)}$ be drawn independently from $\mathcal{D}$ and set $R_{i,j} = x^\top a_i^{(j)} (a_i^{(j)})^\top x$ and $R_{j} = \frac{1}{k} \sum_{i \in [k]} R_{i,j}$. If we let $z$ be median value of the $R_j$ then with probability $1 - p$ we have
$
\left|z - x^\top \mvar{\Sigma} x\right| \leq \epsilon \lambda_1
$.
\end{theorem}
\begin{proof}
\begin{align*}
\mathrm{Var}_{a \sim \mathcal{D}} (x^\top a a^\top x)
&=
\mathbb{E}_{a \sim \mathcal{D}} (x^\top a a^\top x)^2 - (\mathbb{E}_{a \sim \mathcal{D}} x^\top a a^\top x)^2
\\
&\leq
\mathbb{E}_{a \sim \mathcal{D}} \norm{a}_2^2 x^\top a a^\top x - (x^\top \mvar{\Sigma} x)^2
\\
&\leq \normFull{\mathbb{E}_{a \sim \mathcal{D}} \norm{a}_2^2 a a^\top}_2 =\mathsf{v}(\mathcal{D}) \lambda_1^2
\end{align*}
Consequently, $\mathrm{Var}(R_{i,j}) \leq \mathsf{v}(\mathcal{D}) \lambda_1^2$, and since each of the $a_{i}^{(j)}$ were drawn independently this implies that we have that $\mathrm{Var}(R_j) \leq \mathsf{v}(\mathcal{D}) \lambda_1^2 / k$. Therefore, by Chebyshev's inequality
\[
\Pr\left[\left|R_j - \mathbb{E}[R_j]\right| \geq 2 \sqrt{\frac{\mathsf{v}(\mathcal{D}) \lambda_1^2}{k}}\right]
\leq \frac{1}{4}.
\]
Since $\mathbb{E}[R_j] = x^\top \mvar{\Sigma} x$ and since we defined $k$ appropriately this implies that
\begin{equation}
\label{eq:rayleighlemma:1}
\Pr\left[\left|R_j - x^\top \mvar{\Sigma} x \right| \geq \epsilon \lambda_1 \right]
\leq \frac{1}{4}.
\end{equation}
The median $z$ satisfies $|z - x^\top \mvar{\Sigma} x| \leq \epsilon$ as more than half of the $R_j$ satisfy $|R_j - x^\top \mvar{\Sigma} x| \leq \epsilon$. This happens with probability $1 - p$ by Chernoff bound, our choice of $m$ and \eqref{eq:rayleighlemma:1}.
\end{proof}
\subsection{Solving the Linear system}
\label{sec:online:system-solver}
Here we show how to solve linear systems in $\mvar{b} = \lambda \mI - \mvar{\Sigma}$ in the streaming setting. We follow the general strategy of the offline algorithms in Section~\ref{sec:offline}, replacing traditional SVRG with the streaming SVRG algorithm of \cite{frostig2014competing}. Similarly to the offline case we minimize $f(x) = \frac{1}{2} x^\top \bv B x - b^\top x$ and define for all $a \in \mathrm{supp}(\mathcal{D})$,
\begin{align}\label{distribution_obj_function}
\psi_{a}(x)
\defeq
\frac{1}{2} x^{\top} (\lambda \mI - aa^{\top})x - b^{\top}x.
\end{align}
insuring that
$
f(x) = \mathbb{E}_{a \sim \mathcal{D}} \psi_{a}(x).
$.
The performance of streaming SVRG \cite{frostig2014competing} is governed by three regularity parameters. As in the offline case, we use the fact that $f(\cdot)$ is $\mu$-strongly convexity for $\mu = \lambda - \lambda_1$ and we require a smoothness parameter, denoted $\overline{S}$, that satisfies:
\begin{equation}
\label{eq:smoothness}
\forall x \in \mathbb{R}^d \enspace : \enspace
\mathbb{E}_{a \sim \mathcal{D}}\norm{\gradient\psi_{a}(x)-\gradient\psi_{a}(\opt{x})}_2^2
\leq 2 \overline{S} \left [f(x) - f(\opt{x})\right] ~.
\end{equation}
Furthermore, we require an upper bound the variance, denoted $\sigma^2$, that satisfies:
\begin{equation}
\label{eq:variance}
\mathbb{E}_{a \sim \mathcal{D}} \frac{1}{2}
\normFull{\gradient \psi_{a} (\opt{x})}_{
\left(\hessian f (\opt{x})\right)^{-1}}^2
\leq \sigma^2
~.
\end{equation}
With the following two lemmas we bound these parameters.
\begin{lemma} [Streaming Smoothness]
\label{lem:online:smooth}
The smoothness parameter
$
\overline{S} \defeq \lambda + \frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{\lambda - \lambda_1}
$
satisfies \eqref{eq:smoothness}.
\end{lemma}
\begin{proof} Our proof is similar to the one for Lemma~\ref{simple_variance_bound}.
\begin{align*}
\mathbb{E}_{a \sim \mathcal{D}}
\normFull{\gradient\psi_{a}(x) - \gradient \psi_{a} (\opt{x})}_2^2 &= \mathbb{E}_{a \sim \mathcal{D}} \normFull{(\lambda \mI - aa^\top) (x - \opt{x})}_2^2\\
&=
\lambda^2 \norm{x - \opt{x}}_2^2 - 2\lambda \mathbb{E}_{a \sim \mathcal{D}} \norm{x - \opt{x}}_{aa^\top}^2 + \mathbb{E}_{a \sim \mathcal{D}} \norm{aa^\top (x - \opt{x})}_2^2\\
&\le
\lambda^2 \norm{x - \opt{x}}_2^2 - 2\lambda \norm{x - \opt{x}}_{\bv \Sigma}^2 + \normFull{\mathbb{E}_{a \sim \mathcal{D}} \norm{a}_2^2 a a^\top}_2 \cdot \norm{x - \opt{x}}_2^2\\
&\leq \lambda \normFull{x - \opt{x}}_\mvar{b}^2 + \mathsf{v}(\mathcal{D})\lambda_1^2\norm{x - \opt{x}}_2^2.
\end{align*}
Since $f$ is $\lambda - \lambda_1$-strongly convex, $\norm{x - \opt{x}}_2^2 \leq \frac{2}{\lambda - \lambda_1} [f(x) - f(\opt{x})]$. Furthermore, since direct calculation reveals, $2[f(x) - f(\opt{x})] = \norm{x - \opt{x}}_{\mvar{b}}^2$, the result follows.
\end{proof}
\begin{lemma}[Streaming Variance]
\label{lem:online:streamvar}
The variance parameter $\sigma^2 \defeq
\frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{\lambda - \lambda_1}
\normFull{\opt{x}}_2^2$
satisfies \eqref{eq:variance}.
\end{lemma}
\begin{proof}
We have
\begin{align*}
\mathbb{E}_{a \sim \mathcal{D}} \frac{1}{2}
\normFull{\gradient \psi_{a} (\opt{x})}_{
\left(\hessian f (\opt{x})\right)^{-1}}^2
&=
\mathbb{E}_{a \sim \mathcal{D}} \frac{1}{2} \normFull{\left(\lambda \mI
- aa^{\top}\right) \opt{x} - b}_{\mvar{b}^{-1}}^{2}
\\
&=
\mathbb{E}_{a \sim \mathcal{D}} \frac{1}{2} \normFull{\left(\lambda \mI
- aa^{\top}\right) \opt{x} - \mvar{b} \opt{x}}_{\mvar{b}^{-1}}^{2}
\\
&=
\mathbb{E}_{a \sim \mathcal{D}} \frac{1}{2} \normFull{\left(\mvar{\Sigma}
- aa^{\top}\right) \opt{x}}_{\mvar{b}^{-1}}^{2}.
\end{align*}
Applying $\mathbb{E} \norm{a -\mathbb{E} a}_{2}^{2} = \mathbb{E}\norm a_{2}^{2}-\norm{\mathbb{E} a}_{2}^{2}$ gives:
\[
\mathbb{E}_{a \sim \mathcal{D}} \normFull{\left(\mvar{\Sigma}
- aa^{\top}\right) \opt{x}}_{\mvar{b}^{-1}}^{2}
= \mathbb{E}_{a \sim \mathcal{D}} \normFull{\opt{x}}_{a a^\top \mvar{b}^{-1} a a^\top}^2
- \normFull{\opt{x}}_{\mvar{\Sigma} \mvar{b}^{-1} \mvar{\Sigma}}^2 \le \mathbb{E}_{a \sim \mathcal{D}} \normFull{\opt{x}}_{a a^\top \mvar{b}^{-1} a a^\top}^2.
\]
Furthermore, since $\mvar{b}^{-1} \preceq \frac{1}{\lambda - \lambda_1} \mI$ we have
\[
\mathbb{E}_{a \sim \mathcal{D}} a a^{\top} \mvar{b}^{-1} a a^\top
\preceq
\frac{1}{\lambda - \lambda_1} \mathbb{E}_{a \sim \mathcal{D}} (a a^\top)^2
\preceq
\left(
\frac{\normFull{\mathbb{E}_{a \sim \mathcal{D}} (a a ^\top)^2}_2}{\lambda - \lambda_1}\right) \mI
=
\left(\frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{\lambda - \lambda_1}\right) \mI ~.
\]
Combining these three equations yields the result.
\end{proof}
With the regularity parameters bounded we can apply the streaming SVRG algorithm of \cite{frostig2014competing} to solve systems in $\bv{B}$. We encapsulate the core iterative step of Algorithm $1$ of \cite{frostig2014competing} as follows:
\begin{definition}[Streaming SVRG Step]\label{svrg_step_def}
Given $x_0 \in \mathbb{R}^d$ and $\eta, k, m > 0$ we define a \emph{streaming SVRG step}, $x = \mathrm{ssvrg\_iter}(x_0, \eta, k, m)$ as follows. First we take $k$ samples $a_1, ..., a_k$ from $\mathcal{D}$ and set $g = \frac{1}{k} \sum_{i \in [k]} \psi_{a_i}$ where $\psi_{a_i}$ is as defined in \eqref{distribution_obj_function}. Then for $\widetilde{m}$ chosen uniformly at random from $\{1, ..., m\}$ we draw $\tilde m$ additional samples $\widetilde{a}_1, ..., \widetilde{a}_{\widetilde{m}}$ from $\mathcal{D}$. For $t = 0, ... , \widetilde{m} - 1$ we let
\[
x_{t+1}
:=
x_t -
\frac{\eta}{L}
\left(
\gradient \psi_{\widetilde{a}_{t}}(x_t)
- \gradient \psi_{\widetilde{a}_{t}}(x_0)
+ \gradient g(x_0)
\right)
\]
and return $x_{\widetilde{m}}$ as the output.
\end{definition}
The accuracy of the above iterative step is proven in Theorem 4.1 of \cite{frostig2014competing}, which we include, using our notation below:
\begin{theorem} [Theorem 4.1 of \cite{frostig2014competing} \footnote{Note that
Theorem~4.1 in \cite{frostig2014competing} has an additional parameter of $\alpha$, which bounds the Hessian of $f(\opt{x})$ in comparison to the Hessian everywhere else. In our setting this parameter is $1$ as $\hessian f(y) = \hessian f(z)$ for all $y$ and $z$.}]\label{streaming_svrg_perf_bound}
Let $f(x) = \mathbb{E}_{a\sim \mathcal{D}} \psi_a(x)$ and let $\mu$, $\overline{S}$, $\sigma^2$ be the strong convexity, smoothness, and variance bounds for $f(x)$. Then for any distribution over $x_0$ we have that
$x := \mathrm{ssvrg\_iter}(x_0, \eta, k, m)$ has
$\mathbb{E}[f(x) - f(\opt{x})]$ upper bounded by
\[
\frac{1}{1 - 4\eta}
\left[
\left(\frac{\overline{S}}{\mu m\eta} + 4 \eta \right)
\left [\mathbb{E} f(x_0) - f(\opt{x}) \right ]
+ \frac{1 + 2\eta}{k}
\left(
\sqrt{\frac{\overline{S}}{\mu} \cdot \left [\mathbb{E} f(x_0) - f(\opt{x})\right ]}
+ \sigma
\right)^2
\right].
\]
\end{theorem}
Using Theorem~\ref{streaming_svrg_perf_bound} we can immediately obtain the following guarantee for solve system in $\bv{B}$:
\begin{corollary}[Streaming SVRG Solver - With Initial Point]\label{streaming_solverOLD}
Let $\mu =\lambda-\lambda_1$, $\overline{S} =\lambda + \frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{\lambda-\lambda_1}$, and $\sigma^2 = \frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{\lambda-\lambda_1} \norm{\opt{x}}_2^2$. Let $c_2,c_3 \in (0,1)$ be any constants and set $\eta = \frac{c_2}{8}$, $m = \left [\frac{\overline{S}}{\mu c_2^2}\right]$, and $k = \max \left \{ \left [\frac{\overline{S}}{\mu c_2}\right], \left [\frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{(\lambda-\lambda_1)^2c_3} \right] \right \}$. If to solve $\bv{B}x = b$ for unit vector $b$ with initial point $x_0$, we use the iterative procedure described in Definition \ref{svrg_step_def} to compute $x = \mathrm{ssvrg\_iter}(x_0, \eta, k, m)$ then:
\[
\mathbb{E} \norm{x - \opt{x}}_{\mvar{b}}^2
\leq 22c_2 \cdot \norm{x_0-\opt{x}}_\mvar{b}^2 + 10 c_3\lambda_1(\bv{B}^{-1}).
\]
Further, the procedure requires $O \left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2} \left [\frac{1}{c_2^2} + \frac{1}{c_3} \right ] \right )$ samples from $\mathcal{D}$.
\end{corollary}
\begin{proof}
Using the inequality $(x + y)^2 \leq 2x^2 + 2y^2$ we have that
\[
\left(\sqrt{\frac{\overline{S}}{\mu} \cdot \mathbb{E}[f(x_0) - f(\opt{x})]} + \sigma \right)^2
\leq \frac{2 \overline{S}}{\mu} \cdot \mathbb{E}[f(x_0) - f(\opt{x})]
+ 2\sigma^2
\]
Additionally, since $b$ is a unit vector, we know that $\norm{\opt{x}}_2^2 = \norm{\mvar{b}^{-1} b}_2^2 \le \frac{1}{(\lambda-\lambda_1)^2}$.
Using equation \eqref{norm_to_function_error_conversion}, i.e. that $\norm{x-\opt{x}}_\bv{B}^2 = 2 [f(x)-f(\opt{x})]$ for all $x$, we have by Theorem \ref{streaming_svrg_perf_bound}:
\begin{align*}
\mathbb{E} \norm{x-\opt{x}}_\bv{B}^2
&\leq
\frac{1}{1 - c_2/2} \left[
\left(8c_2 + \frac{c_2}{2} + \frac{4+c_2}{2}\cdot c_2 \right)\cdot \norm{x_0-\opt{x}}_\mvar{b}^2
+ \frac{4+c_2}{4k}\cdot \frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{(\lambda-\lambda_1)^3}
\right]
\\
&\le 22c_2 \cdot \norm{x_0-\opt{x}}_\mvar{b}^2 + \frac{10 c_3}{\lambda-\lambda_1}
= 22c_2 \cdot \norm{x_0-\opt{x}}_\mvar{b}^2 + 10 c_3 \lambda_1(\bv{B}^{-1}).
\end{align*}
Since $1/(\lambda - \lambda_1) = \lambda_1(\bv{B}^{-1})$ we see that $\mathbb{E} \norm{x-\opt{x}}_\bv{B}^2$ is as desired. All that remains is to bound the number of samples we used.
Now the number of samples used to compute $x$ is clearly at most $m + k$ Now
\[
m = \frac{\overline{S}}{\mu c_2^2} = O \left(\frac{\lambda}{c_2^2(\lambda-\lambda_1)} + \frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{c_2^2(\lambda-\lambda_1)^2}\right ) = O\left (\frac{1}{c_2^2\mathrm{gap}} + \frac{\mathsf{v}(\mathcal{D})}{c_2^2\mathrm{gap}^2} \right ) ~.
\]. However since $\mathrm{gap} < 1$ and $\mathsf{v}(\mathcal{D}) \ge 1$ this simplifies to $m = O\left ( \frac{\mathsf{v}}{c_2^2\mathrm{gap}^2}\right )$. Next to bound $k$ we can ignore the $\left [\frac{\overline{S}}{\mu c_2}\right]$ term since this was already included in our bound of $m$ and just bound $\frac{\mathsf{v}(\mathcal{D})\lambda_1^2}{c_3(\lambda-\lambda_1)^2} = O\left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2c_3}\right )$ yielding our desired sample complexity.
\end{proof}
Whereas in the offline case, we could ensure that our initial error $\norm{x_0 - \opt{x}}_\mvar{b}^2$ is small by simply scaling by the Rayleigh quotient (Corollary~\ref{cor:constant_factor_corollary}) in the online case estimating the Rayleigh quotient to sufficient accuracy would require too many samples. Instead, here simply show how to simply apply Corollary~\ref{streaming_solverOLD} iteratively to solve the desired linear systems to absolute accuracy without an initial point. Ultimately, due to the different error dependences in the online case this guarantee suffices and the lack of an initial point is not a bottleneck.
\begin{corollary}[Streaming SVRG Solver] \label{streaming_solver}
There is a streaming algorithm that iteratively applies the solver of Corollary~\ref{streaming_solverOLD} to solve $\bv{B}x = b$ for unit vector $b$ and returns a vector $x$ that satisfies
$
\mathbb{E} \norm{x - \opt{x}}_{\mvar{b}}^2
\leq 10 c\lambda_1(\bv{B}^{-1})
$
using $O \left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2 \cdot c}\right )$ samples from $\mathcal{D}$.
\end{corollary}
\begin{proof}
Let $x_0 = 0$. Then $\norm{x_0-\opt{x}}_\mvar{b}^2 = \norm{\bv{B}^{-1} b}_\mvar{b}^2 \le \lambda_1(\bv{B}^{-1})$ since $b$ is a unit vector. If we apply Corollary \ref{streaming_solverOLD} with $c_2 = \frac{1}{44}$ and $c_3 = \frac{1}{20}$, then we will obtain $x_1$ with $\mathbb{E} \norm{x_1-\opt{x}}_\mvar{b}^2 \le \frac{1}{2}\lambda_1(\mvar{b}^{-1})$. If we then double $c_3$ and apply the solver again we obtain $x_2$ with $\mathbb{E} \norm{x_1-\opt{x}}_\mvar{b}^2 \le \frac{1}{4}\lambda_1(\mvar{b}^{-1})$. Iterating in this way, after $\log(1/c)$ iterations we will have the desired guarantee: $\mathbb{E} \norm{x - \opt{x}}_{\mvar{b}}^2 \leq 10 c\lambda_1(\bv{B}^{-1}).$ Our total sample cost in each iteration is, by Corollary \ref{streaming_solverOLD}, $O \left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2} \left [\frac{1}{44^2} + \frac{1}{c_3} \right ] \right ).$ Since we double $c_3$ each time, the cost corresponding to the $\frac{1}{c_3}$ terms is dominated by the last iteration when we have $c_3 = O(c)$. So our overall sample cost is just:
\begin{align*}
O \left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2}\left [\frac{1}{c} + \log(1/c) \right ] \right ) = O \left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2 \cdot c}\right ).
\end{align*}
\end{proof}
\subsection{Online Shifted-and-Inverted Power Method}
\label{sec:online:result}
We now apply the results in Section~\ref{sec:online:rayleigh_quotient} and Section~\ref{sec:online:system-solver} to the shifted-and-inverted power method framework of Section~\ref{framework} to give our main result in the online setting, an algorithm that quickly refines a coarse approximation to $v_1$ into a finer approximation.
\begin{theorem}[Online Shifted-and-Inverted Power Method -- Warm Start]\label{warmstart_online_theorem}
Let $\bv{B} = \lambda \bv{I} - \bv{A}^\top \bv{A}$ for $\left ( 1+\frac{\mathrm{gap}}{150}\right) \lambda_1 \le \lambda \le \left(1+ \frac{\mathrm{gap}}{100} \right ) \lambda_1$ and let $x_0$ be some vector with $G(x_0)\leq \frac{1}{\sqrt{10}}$. Running the shifted-and-inverted power method on $\bv{B}$ initialized with $x_0$, using the streaming SVRG solver of Corollary \ref{streaming_solver} to approximately apply $\bv{B}^{-1}$ at each step, returns $x$ such that $x^\top \bv{\Sigma} x \ge (1-\epsilon) \lambda_1$ with constant probability for any target $\epsilon < \mathrm{gap}$. The algorithm uses
$
O (\frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap} \cdot \epsilon})
$ samples and amortized $O(d)$ time per sample.
\end{theorem}
We note that by instantiating Theorem \ref{warmstart_online_theorem}, with $\epsilon' = \epsilon\cdot \mathrm{gap}$, and applying Lemma \ref{lem:ray_to_evec} we can find $x$ such that $| v_1^\top x | \ge 1-\epsilon$ with constant probability in time $O \left (\frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2 \cdot \epsilon } \right ).
$
\begin{proof}
By Lemma \ref{lem:rayquot-potential} it suffices to have $G^2(x) = O(\frac{\epsilon}{\mathrm{gap}})$ or equivalently $G(x) = O(\sqrt{\epsilon/\mathrm{gap}})$. In order to succed with constant probability it suffices to have $\expec{ G(x)} = O( \sqrt{\epsilon/\mathrm{gap}})$ with constant probability. Since we start with $G(x_0)\leq \frac{1}{\sqrt{10}}$, we can achieve this
using $\log(\mathrm{gap}/\epsilon)$ iterations of the approximate shifted-and-inverted power method of Theorem \ref{thm:powermethod-perturb}. In each iteration $i$ we choose the error parameter for Theorem \ref{thm:powermethod-perturb} to be $c_1(i) = \frac{1}{\sqrt{10}}\cdot \left( \frac{1}{5}\right)^i$. Consequently,
\begin{align*}
\expec{G(x_i)} \le \frac{3}{25} G(x_{i-1}) + \frac{4}{1000}\frac{1}{\sqrt{10}}\cdot \left( \frac{1}{5}\right)^i
\end{align*}
and by induction $\expec{G(x_i)} \le \frac{1}{5^i} \frac{1}{\sqrt{10}}$. We halt when $(\frac{1}{5})^i = O(\sqrt{\epsilon/\mathrm{gap}})$ and hence $c_1(i) = O(\sqrt{\epsilon/\mathrm{gap}})$.
In order to apply Theorem~\ref{thm:powermethod-perturb} we need a subroutine $\rayquoth{x}$ that lets us approximate $\mathrm{quot}(x)$ to within an additive error $\frac{1}{30} (\lambda-\lambda_1) = O(\mathrm{gap} \cdot\lambda_1)$. Theorem~\ref{online_rayleigh_estimation} gives us such a routine, requiring $O \left (\frac{\mathsf{v}(\mathcal{D})\log \log(\mathrm{gap}/\epsilon)}{\mathrm{gap}^2} \right) = O(\frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap} \cdot \epsilon})$ samples to succeed with probability $1- O\left(\frac{1}{\log(\mathrm{gap}/\epsilon)}\right)$ (since $\epsilon < \mathrm{gap}$). Union bounding, the estimation succeeds in all rounds with constant probability.
By Corollary \ref{streaming_solver} with $c = \Theta(c_1(i)^2)$ the cost for solving each linear system solve is
$
O \left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap}^2c_1(i)^2} \right )
$. Since $c_1(i)$ multiplies by a constant factor with each iteration the cost over all $O(\log(\mathrm{gap}/d\epsilon)$ iterations is just a truncated geometric series and is proportional to cost in the last iteration, when $c = \Theta\left(\frac{ \epsilon}{\mathrm{gap}}\right)$. So the total cost for solving the linear systems is
$
O \left ( \frac{\mathsf{v}(\mathcal{D})}{\mathrm{gap} \cdot \epsilon} \right )
$. Adding this to the number of samples for the Rayleigh quotient estimation yields the result.
\end{proof}
\section{Parameter Estimation for Offline Eigenvector Computation}\label{parameter_free}
In Section \ref{sec:offline}, in order to invoke Theorems \ref{thm:powermethod-perturb} and \ref{thm:init-offline} we assumed knowledge of some $\lambda$ with $(1 + c_1 \cdot \mathrm{gap})\lambda_1 \le \lambda \le (1 + c_2 \cdot \mathrm{gap})\lambda_1$ for some small constant $c_1$ and $c_2$. Here we show how to estimate this parameter using Algorithm \ref{algo:eigestimate}, incurring a modest additional runtime cost.
In this section, for simplicity we initially assume that we have oracle access to compute $\mvar{b}_{\lambda}^{-1}x$ for any given $x$, and any $\lambda > \lambda_1$. We will then show how to achieve the same results when we can only compute $\mvar{b}_{\lambda}^{-1}x$ approximately.
We use a result of~\cite{Musco2015} that gives gap free bounds for computing eigenvalues using the power method. The following is a specialization of Theorem 1 from~\cite{Musco2015}:
\begin{theorem}\label{thm:musco}
For any $\epsilon > 0$, any matrix $\mM \in \mathbb{R}^{d\times d}$ with eigenvalues $\lambda_1,...,\lambda_d$, and $k \leq d$, let $\bv{W} \in \mathbb{R}^{d \times k}$ be a matrix with entries drawn independently from $\mathcal{N}(0,1)$. Let $eigEstimate (\bv{Y})$ be a function returning for each $i$, $\tilde \lambda_i = \tilde v_i^\top \bv{M} \tilde v_i$ where $\tilde v_i$ is the $i^{th}$ largest left singular vector of $\bv{Y}$. Then setting $[\tilde \lambda_1,...,\tilde \lambda_k ] = eigEstimate\left (\bv{M}^t \bv{W} \right )$, for some fixed constant $c$ and $t = c\alpha \log d$ for any $\alpha > 1$, with probability $1-\frac{1}{d^{10}}$, we have for all $i$:
\begin{align*}
|\tilde \lambda_i - \lambda_i| \le \frac{1}{\alpha}\lambda_{k+1}
\end{align*}
\end{theorem}
\begin{algorithm}[t]
\caption{Estimating the eigenvalue and the eigengap}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input: }}
\renewcommand{\algorithmicensure}{\textbf{Output: }}
\REQUIRE $\bv{A}\in\mathbb{R}^{n\times d},\;\alpha$
\STATE $w = \left[w_{1},w_{2}\right]\leftarrow\mathcal{N}\left(0,1\right)^{d\times 2}$
\STATE $t\leftarrow O\left(\alpha\log d\right)$
\STATE $\left[\lamtilij 01,\lamtilij 02\right]\leftarrow eigEstimate\left(\left(\bv{A}^{T}\bv{A}\right)^{t}w\right)$
\STATE $\lambari 0\leftarrow(1+\frac{1}{2})\lamtilij 01$
\STATE $i\leftarrow0$
\WHILE{$\lambari i-\lamtilij i1<\frac{1}{10}\left(\lambari i-\lamtilij i2\right)$}
\STATE $i\leftarrow i+1$
\STATE $w = \left[w_{1},w_{2}\right]\leftarrow\mathcal{N}\left(0,1\right)^{d\times 2}$
\STATE $\left[\lamhatij i1,\lamhatij i2\right]\leftarrow eigEstimate\left(\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-t}w\right)$
\STATE $\left[\lamtilij i1,\lamtilij i2\right]\leftarrow\left[\lambari{i-1}-\frac{1}{\lamhatij i1},\lambari{i-1}-\frac{1}{\lamhatij i2}\right]$
\STATE $\lambari i\leftarrow\frac{1}{2}\left(\lamtilij i1+\lambari{i-1}\right)$
\ENDWHILE
\ENSURE $\lambda$
\end{algorithmic}
\label{algo:eigestimate}
\end{algorithm}
Throughout the proof, we assume $\alpha$ is picked to be some large constant - e.g. $\alpha>100$. Theorem \ref{thm:musco} implies:
\begin{lemma}
\label{lem:topeig1} Conditioning on the event that Theorem \ref{thm:musco} holds for all iterates $i$, then the iterates
of Algorithm \ref{algo:eigestimate} satisfy:
\begin{align*}
0\leq\lamj 1-\lamtilij 01\leq\frac{1}{\alpha}\lamj 1\;\;\mbox{and}\;\; & \frac{1}{2}\left(1-\frac{3}{\alpha}\right)\lamj 1\leq\lambari 0-\lamj 1\leq\frac{1}{2}\lamj 1,\;\;\mbox{and,}
\end{align*}
\begin{align*}
0\leq\lamj 1-\lamtilij i1\leq\frac{1}{\alpha-1}\left(\lambari{i-1}-\lamj 1\right)\;\;\mbox{and}\;\; & \frac{1}{2}\left(1-\frac{1}{\alpha-1}\right)\left(\lambari{i-1}-\lamj 1\right)\leq\lambari i-\lamj 1\leq\frac{1}{2}\left(\lambari{i-1}-\lamj 1\right).
\end{align*}
\end{lemma}
\begin{proof}
The proof can be decomposed into two parts:
\textbf{Part I (Lines 3-4):} Theorem \ref{thm:musco} tells us that $\lamtilij 01\geq\left(1-\frac{1}{\alpha}\right)\lamj 1$.
This means that we have
\begin{align*}
0\leq\lamj 1-\lamtilij 01\leq\frac{1}{\alpha}\lamj 1\;\;\mbox{and}\;\; & \frac{1}{2}\left(1-\frac{3}{\alpha}\right)\lamj 1\leq\lambari 0-\lamj 1\leq\frac{1}{2}\lamj 1.
\end{align*}
\textbf{Part II (Lines 5-6):} Consider now iteration $i$. We now
apply Theorem \ref{thm:musco} to the matrix $\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-1}$.
The top eigenvalue of this matrix is $\left(\lambari{i-1}-\lamj 1\right)^{-1}$.
This means that we have $\left(1-\frac{1}{\alpha}\right)\left(\lambari{i-1}-\lamj 1\right)^{-1}\leq\lamhatij i1\leq\left(\lambari{i-1}-\lamj 1\right)^{-1}$,
and hence we have,
\begin{align*}
0\leq\lamj 1-\lamtilij i1\leq\frac{1}{\alpha-1}\left(\lambari{i-1}-\lamj 1\right)\;\;\mbox{and}\;\; & \frac{1}{2}\left(1-\frac{1}{\alpha-1}\right)\left(\lambari{i-1}-\lamj 1\right)\leq\lambari i-\lamj 1\leq\frac{1}{2}\left(\lambari{i-1}-\lamj 1\right).
\end{align*}
This proves the lemma.\end{proof}
\begin{lemma}
\label{lem:secondeig}Recall we denote $\lamj 2\defeq\lamj 2\left(\bv{A}^{T}\bv{A}\right)$
and $\mathrm{gap}\defeq\frac{\lamj 1-\lamj 2}{\lamj 1}$. Then conditioning on the event that Theorem \ref{thm:musco} holds for all iterates $i$, the iterates
of Algorithm \ref{algo:eigestimate} satisfy $\abs{\lamj 2-\lamtilij i2}\leq\frac{1}{\alpha-1}\left(\lambari{i-1}-\lamj 2\right)$,
and $\lambari i-\lamtilij i2\geq\frac{\mathrm{gap} \lamj 1}{4}$.\end{lemma}
\begin{proof}
Since $\left(\lambari{i-1}-\lamj 2\right)^{-1}$ is the second eigenvalue
of the matrix $\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-1}$, Theorem \ref{thm:musco}
tells us that
\[
\left(1-\frac{1}{\alpha}\right)\left(\lambari{i-1}-\lamj 2\right)^{-1}\leq\lamhatij i2\leq\left(1+\frac{1}{\alpha}\right)\left(\lambari{i-1}-\lamj 2\right)^{-1}.
\]
This immediately yields the first claim. For the second claim, we
notice that
\begin{align*}
\lambari i-\lamtilij i2
&=\lambari i-\lamj 2+\lamj 2-\lamtilij i2\\
& \stackrel{(\zeta_{1})}{\geq}\lambari i-\lamj 2-\frac{1}{\alpha-1}\left(\lambari{i-1}-\lamj 2\right) \\
& =\lambari i-\lamj 1-\frac{1}{\alpha-1}\left(\lambari{i-1}-\lamj 1\right)+\left(1-\frac{1}{\alpha-1}\right)\left(\lamj 1-\lamj 2\right) \\
& \stackrel{\left(\zeta_{2}\right)}{\geq}\frac{1}{2}\left(1-\frac{3}{\alpha-1}\right)\left(\lambari{i-1}-\lamj 1\right)+\left(1-\frac{1}{\alpha-1}\right)\left(\lamj 1-\lamj 2\right)\geq\frac{\mathrm{gap}\lamj 1}{4},
\end{align*}
where $\left(\zeta_{1}\right)$ follows from the first claim of this lemma,
and $\left(\zeta_{2}\right)$ follows from Lemma \ref{lem:topeig1}.
\end{proof}
We now state and prove the main result in this section:
\begin{theorem}\label{thm:parafree}
Suppose $\alpha>100$, and after $T$ iterations, Algorithm \ref{algo:eigestimate} exits. Then
with probability $1- \frac{\ceil{\log\frac{10}{\mathrm{gap}}}+1}{d^{10}}$, we have
$T \le \ceil{\log\frac{10}{\mathrm{gap}}}+1$, and:
\begin{equation*}
\left(1+\frac{\mathrm{gap}}{120}\right)\lambda_1
\le \lambari T
\le \left(1+\frac{\mathrm{gap}}{8}\right)\lambda_1
\end{equation*}
\end{theorem}
\begin{proof}
By union bound, we know with probability $1- \frac{\ceil{\log\frac{10}{\mathrm{gap}}}+1}{d^{10}}$, Theorem \ref{thm:musco} will hold for all iterates where $i\le\ceil{\log\frac{10}{\mathrm{gap}}}+1$.
Let $\overline{i}=\ceil{\log\frac{10}{\mathrm{gap}}}$, suppose the algorithm has not exited yet after $\overline{i}$
iterations, then since $\lambari i-\lamj 1$ decays geometrically, we have $\lambari{\overline{i}}-\lamj 1\leq\frac{\mathrm{gap}\lamj 1}{10}$.
Therefore, Lemmas \ref{lem:topeig1} and \ref{lem:secondeig} imply
that $\lambari{\overline{i}+1}-\lamtilij{\overline{i}+1}1\leq\left(\frac{1}{2}+\frac{1}{\alpha-1}\right)\left(\lambari{\overline{i}}-\lamj 1\right)\leq\frac{\mathrm{gap}\lamj 1}{15}$,
and
\begin{align*}
\lambari{\overline{i}+1}-\lamtilij{\overline{i}+1}2
&\geq\lambari{\overline{i}+1}-\lamj 2-\abs{\lamj 2-\lamtilij{\overline{i}+1}2}\geq\lamj 1-\lamj 2-\frac{1}{\alpha-1}\left(\lambari{\overline{i}}-\lamj 2\right)\\
&=\mathrm{gap}\lamj 1-\frac{1}{\alpha-1}\left(\lambari{\overline{i}}-\lamj 1+\lamj 1-\lamj 2\right)\geq\frac{3}{4}\mathrm{gap}\lamj 1
\end{align*}
This means that the exit condition on Line $6$ must be triggered in $\overline{i}+1$ iteration, proving
the first part of the lemma.
For upper bound, by Lemmas \ref{lem:topeig1}, \ref{lem:secondeig} and exit condition we know:
\begin{align*}
\lambari T- \lambda_1 &\le \lambari T-\lamtilij T1
\le \frac{1}{10} (\lambari T-\lamtilij T2)
\le \frac{1}{10} \left(\lambari T - \lambda_2 + \abs{\lamj 2-\lamtilij T2}\right)\\
&\leq \frac{1}{10} \left(\lambari T - \lambda_2 + \frac{1}{\alpha-1}(\lambari{T-1}-\lamj 2)\right)\\
&= \frac{1}{10} \left( \frac{\alpha}{\alpha-1} \mathrm{gap} \lambda_1 + (\lambari T - \lambda_1) + \frac{1}{\alpha-1}\left(\lambari{T-1}-\lamj 1\right) \right)\\
&\le \frac{1}{10} \left( \frac{\alpha}{\alpha-1} \mathrm{gap} \lambda_1 + \frac{\alpha}{\alpha-2}\left(\lambari{T}-\lamj 1\right) \right)
\end{align*}
Since $\alpha > 100$, this directly implies $\lambari{T}-\lamj 1 \le \frac{\mathrm{gap}}{8}\lambda_1$.
For lower bound, since as long as the Algorithm \ref{algo:eigestimate} does not exists, by Lemmas \ref{lem:secondeig}, we have
$\lambari {T-1}-\lamtilij {T-1}1\geq\frac{1}{10}\left(\lambari {T-1}-\lamtilij {T-1}2\right)\geq\frac{\mathrm{gap}\lamj 1}{40}$, and thus:
\begin{align*}
\lambari {T-1}- \lambda_1 &= \lambari {T-1}-\lamtilij {T-1}1 - (\lambda_1 - \lamtilij {T-1}1)
\ge \frac{\mathrm{gap}\lamj 1}{40} - \frac{1}{\alpha-1}\left(\lambari{T-1}-\lamj 1\right) \\
& \ge \frac{\mathrm{gap}\lamj 1}{40} - \frac{2}{\alpha-2}\left(\lambari{T}-\lamj 1\right)
\ge \frac{\mathrm{gap}\lamj 1}{50}
\end{align*}
By Lemma \ref{lem:topeig1}, we know $\lambari {T}- \lambda_1\ge \frac{1}{2}(1-\frac{1}{\alpha-1}(\lambari {T-1}- \lambda_1)) >\frac{\mathrm{gap}}{120}\lamj 1$
\end{proof}
Note that, although we proved the upper bound and lower bound in Theorem \ref{thm:parafree} with specific constants coefficient $\frac{1}{8}$ and $\frac{1}{120}$, this analysis can easily be extended to any smaller constants by modifying the constant in the exit condition, and choosing $\alpha$ larger.
Also in the failure probability
$$1- \frac{\ceil{\log\frac{10}{\mathrm{gap}}}+1}{d^{10}},$$ the term $d^{10}$ can be replaced by any $\mathop\mathrm{poly}(d)$ by adjusting the constant in setting $t \leftarrow O(\alpha \log d)$ in Algorithm \ref{algo:eigestimate}. Assuming $\log \frac{1}{\mathrm{gap}} < \mathop\mathrm{poly}(d)$, thus gives that Theorem \ref{thm:parafree} returns a correct result with high probability.
Finally, we can also bound the runtime of algorithm \ref{algo:eigestimate}, when we use SVRG based approximate linear system solvers for $\bv{B}_\lambda$.
\begin{theorem}\label{thm:parafree_runtime}
With probability $1- O(\frac{1}{d^{10}}\log\frac{1}{\mathrm{gap}})$, Algorithm \ref{algo:eigestimate} runs in time $$O\left (\left [\nnz(\bv A) + \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2} \right ] \cdot \log^3\left (\frac{d}{\mathrm{gap}} \right ) \right)$$.
\end{theorem}
\begin{proof}
By Theorem \ref{thm:parafree}, we know only $O(\log 1/\mathrm{gap})$ iterations of the algorithm are needed.
In each iteration, the runtime is dominated by running $eigEstimate\left(\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-t}w\right)$, which is dominated by computing $\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-t}w$.
Since $t=O(\log d)$, it's easy to verify that: to make Theorem \ref{thm:musco} hold, we only need to approximate $\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-1}w$ up to accuracy $\mathop\mathrm{poly}(\mathrm{gap}/d)$. By Theorem \ref{offline_solver}, we know this approximation can be calculated in time $$O\left (\left [ \nnz(\bv A) + \frac{d\mathsf{r}(\bv A)\lambda_1^2}{(\lambari{i-1} - \lambda_1)^2} \right ] \cdot \log\left (\frac{d}{\mathrm{gap}}\right ) \right)$$. Combining Theorem \ref{thm:parafree} with Lemma \ref{lem:topeig1}, we know $\lambari{i-1} - \lambda_1 \ge
\lambari{T} - \lambda_1 \ge \frac{\mathrm{gap}}{120}$, thus approximately solving $\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-1}w$ can be done in time $\tilde{O}\left (\nnz(\bv A) + \frac{d\mathsf{r}(\bv A)}{\mathrm{gap}^2} \right)$. Finally, since the runtime of Algorithm \ref{algo:eigestimate} is dominated by repeating this subroutine $t\times T = O(\log d \cdot \log (1/\mathrm{gap}))$ times, we finish the proof.
\end{proof}
Note that we can accelerate the runtime of Algorithm \ref{algo:eigestimate} to $\tilde O \left (\frac{\nnz(\bv A)^{3/4}(d\mathsf{r}(\bv A))^{1/4}}{\sqrt{\mathrm{gap}}} \right )$, by simply
replacing the base solver for $\left(\lambari{i-1}\bv{I}-\bv{A}^{T}\bv{A}\right)^{-1}w$ with the accelerated solver in Theorem \ref{accelerated_offline_solver}.
\section{Preliminaries}\label{prelims}
We bold all matrix variables.
We use $[n] \defeq \{1,...,n\}$. For a symmetric positive semidefinite (PSD) matrix $\bv{M}$ we let $\norm x_{\bv{M}} \defeq \sqrt{x^{\top}\bv{M} x}$ and $\lambda_1(\bv M), ..., \lambda_d(\bv M)$ denote its eigenvalues in decreasing order.
We use $\bv{M} \preceq \bv{N}$ to denote the condition that $x^{\top}\bv{M} x\leq x^{\top}\bv{N} x$
for all $x$.
\subsection{The Offline Problem }
We are given a matrix $\mvar{a}\in\mathbb{R}^{n\times d}$ with rows $a^{(1)},...,a^{(n)}$
and wish to compute an approximation to the top eigenvector of $\mvar{\Sigma}\defeq\mvar{a}^{\top}\mvar{a}$. Specifically, for error parameter $\epsilon$ we want a unit vector $x$ such that $x^\top \bv{\Sigma} x \ge (1-\epsilon) \lambda_1(\bv{\Sigma})$.
\subsection{The Statistical Problem}
We have access to an oracle returning independent samples from a distribution $\mathcal{D}$
on $\mathbb{R}^{d}$ and wish to compute the top eigenvector of $\mvar{\Sigma}\defeq\mathbb{E}_{a\sim\mathcal{D}} \left [aa^\top \right ]$. Again, for error parameter $\epsilon$ we want to return a unit vector $x$ such that $x^\top \bv{\Sigma} x \ge (1-\epsilon) \lambda_1(\bv{\Sigma})$.
\subsection{Problem Parameters}
We parameterize the running times and sample complexities of our algorithms in terms of several natural properties of $\mvar{a}$, $\mathcal{D}$, and $\mvar{\Sigma}$. Let $\lambda_{1},...,\lambda_{d}$ denote the eigenvalues of $\mvar{\Sigma}$
in decreasing order and $v_1,... ,v_d$ denote their corresponding
eigenvectors. We define the \emph{eigenvalue gap} by $\mathrm{gap}\defeq\frac{\lambda_{1}-\lambda_{2}}{\lambda_{1}}$.
We use the following additional parameters for the offline and statistical problems respectively:
\begin{itemize}
\item \textbf{Offline Problem}: Let $\mathsf{r}(\mvar{a}) \defeq\sum_{i}\frac{\lambda_{i}}{\lambda_{1}} = \frac{\norm{\mvar{a}}_F^2}{\norm{\mvar{a}}_2^2}$ denote the stable rank of $\mvar{a}$. Note that we always have $\mathsf{r}(\mvar{a}) \le \rank(\mvar{a})$. Let $\nnz(\mvar{a})$ denote the number of non-zero entries in $\mvar{a}$.
\item \textbf{Online Problem}: Let $\mathsf{v}(\mathcal{D}) \defeq \frac{\norm{\mathbb{E}_{a \sim \mathcal{D}}\left [ \left (a a^\top \right )^2 \right ]}_2}{\norm{\mathbb{E}_{a \sim \mathcal{D}} (aa^\top)}_2^2} = \frac{\norm{\mathbb{E}_{a \sim \mathcal{D}}\left [ \left (a a^\top \right )^2 \right ]}_2}{\lambda_1^2}$ denote a natural upper bound on the variance of $\mathcal{D}$ in various settings. Note that $\mathsf{v}(\mathcal{D}) \ge 1$.
\end{itemize}
|
\section{Introduction}
\label{sect-intro}
If $\s{x} = \s{uv}$ for some \s{u} and nonempty \s{v},
then \s{vu} is said to be the $|\s{u}|^{\mbox{th}}$ \itbf{rotation} of \s{x},
written $\s{vu} = R_{|\s{u}|}(x)$.
If there exists a string \s{u} and an integer $e > 1$
such that $\s{x} = \s{u}^e$,
then \s{x} is said to be a \itbf{repetition};
otherwise \s{x} is \itbf{primitive}.
A primitive string \s{x} that is
lexicographically least among all its rotations
$R_k(\s{x}), k= 0,1,\ldots,|\s{x}|\- 1$,
is said to be a \itbf{Lyndon word}.
The \itbf{Lyndon array} $\s{\lambda} = \s{\lambda}_{\s{x}}[1..n]$
(equivalently, $\mathcal{L} = \mathcal{L}_{\s{x}}[1..n]$)
of a given nonempty string $\s{x} = \s{x}[1..n]$
gives at each position $i$ the length
(equivalently, the end position)
of the longest Lyndon word starting at $i$:
\begin{equation}
\label{ex1}
\begin{array}{rccccccccc}
\scriptstyle 1 & \scriptstyle 2 & \scriptstyle 3 & \scriptstyle 4 & \scriptstyle 5 & \scriptstyle 6 & \scriptstyle 7 & \scriptstyle 8 & \scriptstyle 9 & \scriptstyle 10 \\
\s{x} = a & b & a & a & b & a & b & a & a & b \\
\s{\lambda} = 2 & 1 & 5 & 2 & 1 & 2 & 1 & 3 & 2 & 1 \\
\mathcal{L} = 2 & 2 & 7 & 5 & 5 & 7 & 7 & 10 & 10 & 10
\end{array}
\end{equation}
The Lyndon array has recently become of interest since
Bannai {\it et al.} \cite{BIINTT14} showed that it could
be used to efficiently compute all the maximal periodicities (``runs'')
in a string.
In this paper we describe four algorithms to compute $\s{\lambda}_{\s{x}}$,
three of them shown experimentally to be running in $\Theta(n)$ time in practice.
Section~\ref{sect-prelim} makes various observations that
apply generally to the Lyndon array and its computation.
In Section~\ref{sect-folklore} we describe three
algorithms, two that require $\mathcal{O}(n^2)$ time in the worst case,
of which one is very fast and apparently linear in practice,
the other supralinear in practice and $\mathcal{O}(n\log n)$ in the average case on binary strings.
The third algorithm is simple and worst-case linear-time,
but requires suffix array construction and so is a little slower.
Section~\ref{sect-newalgs} describes two variants of a new algorithm
that uses only elementary data structures (no suffix arrays).
One variant is $\mathcal{O}(n^2)$ in the worst case,
the other guarantees $\mathcal{O}(n\log n)$ time,
but with no clear advantage in processing time.
Section~\ref{sect-exp} describes the results of
preliminary experiments on the algorithms;
Section~\ref{sect-future} outlines future work.
\section{Preliminaries}
\label{sect-prelim}
Here we make various observations that apply to
the algorithms described below.
\begin{obs}
\label{obs-decomp}
Let $\s{x} = \s{w_1w_2\cdots w_k}$ be the Lyndon decompostion \cite{CFL58,D83} of \s{x},
with Lyndon words $\s{w_1} \ge \s{w_2} \ge \cdots \ge \s{w_k}$.
Then every Lyndon word $\s{x}[i..\mathcal{L}[i]]$ of length $\s{\lambda}[i]$
is a substring of some $\s{w_h},\ h \in 1..k$.
\end{obs}
\begin{proof}
For some $h \in 1..k\- 1$, consider \s{w_h} with nonempty proper suffix \s{v_h},
and for some $t \in 1..k\- h$, consider \s{w_{h+t}} with nonempty prefix \s{u_{h+t}}.
Since \s{w_h} is a Lyndon word, $\s{w_h} < \s{v_h}$,
and by lexorder, $\s{u_{h+t}} \le \s{w_{h+t}}$. Thus
$\s{v_h} > \s{w_h} \ge \s{w_{h+t}} \ge \s{u_{h+t}}$,
and so $\s{v_hw_{h+1}}\cdots\s{w_{h+t-1}u_{h+t}}$
cannot be a Lyndon word for any choice of $h$ or $t$. \hfill \quad\qedbox46\newline\smallbreak
\end{proof}
Therefore to compute $\mathcal{L}_{\s{x}}$ it suffices to consider separately
each distinct element \s{w_h} in the Lyndon decomposition of \s{x}.
Hence, without loss of generality suppose \s{x} is a Lyndon word
and write it in the form $\s{x_1x_2}\cdots\s{x_m}$,
where for each $r \in 1..m$, $|\s{x_r}| = \ell_r$ and
\begin{equation}
\label{range}
\s{x_r}[1] \le \s{x_r}[2] \le \cdots \le \s{x_r}[\ell_r],
\end{equation}
while for $1 \le r < m$,
\begin{equation}
\label{drop}
\s{x_r}[\ell_r] > \s{x_{r+1}}[1].
\end{equation}
We call \s{x_r} a \itbf{range} in \s{x} and the boundary between
\s{x_r} and \s{x_{r+1}} a \itbf{drop}.
We identify a position $j$ in range \s{x_r}, $1 \le j \le \ell_r$,
with its equivalent position $i$ in \s{x} by writing
$i = S_{r,j} = \sum_{r'=1}^{r-1} \ell_{r'}\+ j$.
\begin{obs}
\label{obs-endrange}
Let $i = S_{r,j}$ be a position in \s{x} that corresponds to position $j$
in range \s{x_r}.
\begin{itemize}
\item[$(a)$]
If $\s{x_r}[j] = \s{x_r}[\ell_r]$, then $\mathcal{L}[i] = i$.
\item[$(b)$]
Otherwise, $\mathcal{L}[i] = i'$, where $i'$ is the final position in some range $\s{x_{r'}},\ r' \ge r$;
that is, $i' = \sum_{s=1}^{r'} \ell_s$.
\end{itemize}
\end{obs}
\begin{proof}
(a) is an immediate consequence of (\ref{range}) and (\ref{drop}).
To prove (b), suppose that $\s{x}[i..\mathcal{L}[i]]$ is a maximum-length Lyndon word,
where $\mathcal{L}[i]$ falls within range $r'$ but $\mathcal{L}[i] < i'$.
Since by (\ref{range}) $\s{x}[\mathcal{L}(i)] \le \s{x}[\mathcal{L}[i]\+ 1]$,
there are two consecutive Lyndon words $\s{x}[i..\mathcal{L}[i]],\s{x}[\mathcal{L}[i]\+ 1]$
that by the Lyndon decomposition theorem \cite{CFL58}
can be merged into a single Lyndon word $\s{x}[i..\mathcal{L}[i]\+ 1]$.
Thus $\s{x}[i..\mathcal{L}[i]]$ is not maximum-length, a contradiction. \hfill \quad\qedbox46\newline\smallbreak
\end{proof}
We see then that if $\s{x_r}[j] < \s{x_r}[\ell_r]$,
then $\s{x_r}[j..\ell_r]$ is a (not necessarily maximum-length) Lyndon word,
and for $i = S_{r,j}$, $\mathcal{L}[i] \ge S_{r,\ell_r}$:
\begin{equation}
\label{ex2}
\begin{array}{rcccccccccccc}
\scriptstyle 1 & \scriptstyle 2 & \scriptstyle 3 & \scriptstyle 4 & \scriptstyle 5 & \scriptstyle 6 & \scriptstyle 7 & \scriptstyle 8 & \scriptstyle 9 & \scriptstyle 10 & \scriptstyle 11 & \scriptstyle 12 & \scriptstyle 13 \\
\s{x} = a & a & a & b\,| & a & a & b\,| & a & b\,| & a & a & b & b \\
\mathcal{L} = 13 & 13 & 4 & 4 & 9 & 7 & 7 & 9 & 9 & 13 & 13 & 12 & 13
\end{array}
\end{equation}
More generally, the vectors $(i,\mathcal{L}[i])$ satisfy a ``Monge'' property
that is exploited by Algorithm $\mbox{NSV}^*$ (Section~\ref{sect-newalgs}):
\begin{obs}
\label{obs-monge}
Suppose positions $i,j$ in $\s{x}[1..n]$ satisfy $1 \le i < j \le n$.
Then either $\mathcal{L}[i] \le j$ or $\mathcal{L}[i] \ge \mathcal{L}[j]$:
the vectors $(i,\mathcal{L}[i])$ and $(j,\mathcal{L}[j])$ are nonintersecting.
\end{obs}
\begin{proof}
Suppose two such vectors do intersect.
Then the maximum-length Lyndon words
$\s{w_1} = \s{x}[i..\mathcal{L}[i]]$ and $\s{w_2} = \s{x}[j..\mathcal{L}[j]]$
have a nonempty overlap,
so that we can write
$\s{w_1} = \s{uv},\ \s{w_2} = \s{vv'}$ for some nonempty \s{v}.
But then, by well-known properties of Lyndon words,
$\s{w_1} < \s{v} < \s{w_2} < \s{v'}$,
implying that $\s{w_1v'}$ is a Lyndon word,
contradicting the assumption that \s{w_1} is maximum-length. \hfill \quad\qedbox46\newline\smallbreak
\end{proof}
Expressing a string in terms of its ranges has the same useful
lexorder property that writing it in terms of its letters does:
\begin{obs}
\label{obs-compare}
Suppose strings \s{x} and \s{y} are expressed in terms of their ranges:
$\s{x} = \s{x_1x_2}\cdots\s{x_m},\ \s{y} = \s{y_1y_2}\cdots\s{y_n}$.
Suppose further that for some least integer $r \in 1..\min(m,n)$,
$\s{x_r} \ne \s{y_r}$.
Then $\s{x} < \s{y}$ (respectively, $\s{x} > \s{y}$)
according as $\s{x_r} < \s{y_r}$ (respectively, $\s{x_r} > \s{y_r}$).
\end{obs}
\begin{proof}
If $\s{x_r} < \s{y_r}$, then either
\begin{itemize}
\item[$(a)$]
\s{x_r} is a nonempty proper prefix of \s{y_r}; or
\item[$(b)$]
there is some least position $j$ such that $\s{x_r}[j] < \s{y_r}[j]$.
\end{itemize}
In case (a), if $r = m$, then \s{x} is actually a prefix of \s{y},
so that $\s{x} < \s{y}$,
while if $r < m$, then by (\ref{drop}),
$\s{x_{r+1}}[1] < \s{y_r}[|\s{x_r}|\+ 1]$, and again $\s{x} < \s{y}$.
In case (b) the result is immediate.
The proof for $\s{x_r} > \s{y_r}$ is similar. \hfill \quad\qedbox46\newline\smallbreak
\end{proof}
\section{Basic Algorithms}
\label{sect-folklore}
Here we outline three algorithms
for which no clear exposition is available in the literature.
We remark that the Lyndon array computation is equivalent to ``Lyndon bracketing'',
for which an $\mathcal{O}(n^2)$ algorithm has been described \cite{SR03}.
\subsection{Folklore --- Iterated MaxLyn}
\label{subsect-maxlyn}
For a string \s{x} of length $n$,
recall that the \itbf{prefix table} $\pi[1..n]$
is an integer array in which for every $i \in 1\mathinner{\ldotp\ldotp} n$,
$\pi[i]$ is the length of the longest substring beginning at position $i$
of \s{x} that matches a prefix of \s{x}.
Given a nonempty string \s{x} on alphabet $\Sigma$, let us define $\s{x'} = \s{x}\$$,
where the sentinel $\$ < \mu$ for every letter $\mu \in \Sigma$.
\begin{obs}
\label{obs2}
\s{x} is a Lyndon word if and only if for every $i \in 2\mathinner{\ldotp\ldotp} n$,
$\s{x'}[1+k] < \s{x'}[i+k]$, where $k = \pi[i]$.
\end{obs}
This result forms the basis of the algorithm
given in Figure~\ref{fig-maxlyn} that computes
the length $\max \in 1\mathinner{\ldotp\ldotp} n-j +1$
of the longest Lyndon factor at a given position $j$ in $\s{x}[1..n]$.
Its efficiency is a consequence of the instruction
$i \leftarrow i+k+1$ that skips over positions in the range $i+1\mathinner{\ldotp\ldotp} i+k-1$,
effectively assuming that for every position $i^*$
in that range, $i^*+\pi[i^*] \le i\+ k$.
Lemma~\ref{lemm-maxlyn}, given in Appendix 1, justifies this assumption.
Simply repeating MaxLyn at every position $j$ of \s{x}
gives a simple, fast $\mathcal{O}(n^2)$ time and $\mathcal{O}(1)$ additional space
algorithm to compute $\s{\lambda}_{\s{x}}$.
\begin{figure}[ht]
{\leftskip=2.5cm\obeylines\sfcode`;=3000
{\bf procedure\ } MaxLyn$(\s{x}[1\mathinner{\ldotp\ldotp} n],j,\Sigma,\prec)\ :\ integer$
$i \leftarrow j + 1;\ max\leftarrow 1$
{\bf while\ } $i \le n$ {\bf do\ }
\quad $k \leftarrow 0$
\quad {\bf while\ } $\s{x'}[j+k] = \s{x'}[i+k]$ {\bf do\ }
\qquad $k \leftarrow k+1$
\quad {\bf if\ } $\s{x'}[j+k] \prec \s{x'}[i+k]$ {\bf then\ }
\qquad $i \leftarrow i+k+1;\ max \leftarrow i-1$
\quad {\bf else\ }
\qquad {\bf return\ } max
}
\caption{Algorithm MaxLyn}
\label{fig-maxlyn}
\end{figure}
Recent work on the prefix table
\cite{BKS13,CRSW15} has confirmed its importance as a data structure
for string algorithms.
In this context it is interesting to find that Lyndon words \s{x}
can be characterized in terms of $\pi_{\s{x}}$:
\begin{obs}
\label{obs-prefix}
Suppose $\s{x} = \s{x}[1\mathinner{\ldotp\ldotp} n]$ is a string on alphabet $\Sigma$
such that $\s{x}[1]$ is the least letter in \s{x}.
Then \s{x} is a Lyndon word over $\Sigma$ if and only if
for every $i \in 2\mathinner{\ldotp\ldotp} n$,
\begin{itemize}
\item[$(a)$]
$i + \pi_{\s{x}}[i] < n+1$; and
\item[$(b)$]
for every $j \in i+1\mathinner{\ldotp\ldotp} i+\pi_{\s{x}}[i]-1$,
$j + \pi_{\s{x}}[j] \le i + \pi_{\s{x}}[i]$.
\end{itemize}
\end{obs}
\subsection{Recursive Duval Factorization: Algorithm RDuval}
Rather than independently computing the maximum-length Lyndon factor
at each position $i$, as MaxLyn does, Algorithm RDuval
recursively computes the Lyndon decomposition into
maximum factors, at each step taking advantage of the fact that
$\mathcal{L}[i]$ is known for the first position $i$ in each factor,
then recomputing with the first letters removed.
By Observation~\ref{obs-decomp},
whenever $\s{x} = \s{x}[1..n]$ is a Lyndon word,
we know that $\mathcal{L}[1] = n$.
Thus computing the Lyndon decomposition $\s{x} = \s{w_1w_2}\cdots\s{w_k}$,
$\s{w_1} \ge \s{w_2} \ge \cdots \ge \s{w_k}$,
allows us to assign $\s{\lambda}[i_j] = |\s{w_j}|$,
where $i_j$ is the first position of \s{w_j}, $j = 1,2,\ldots,k$.
Algorithm RDuval applies this strategy recursively,
by assigning $\s{\lambda}[i_j] \leftarrow |\s{w_j}|$,
then removing the first letter $i_j$ from each $\s{w_j}$ to form $\s{w'_j}$,
to which the Lyndon decomposition is applied in the next recursive step.
This process continues until each Lyndon word is reduced to a single letter.
The asymptotic time required for RDuval is bounded above by
$n$ times the maximum depth of the recursion, thus $O(n^2)$ in the worst case ---
consider, for example, the string $\s{x} = a^{n-1}b$.
However, to estimate expected behaviour, we can make use of a result
of Bassino {\it et al.} \cite{BassinoCN05}.
Given a Lyndon word \s{w}, they call $\s{w} = \s{uv}$
the \itbf{standard factorization} of \s{w} if \s{u} and \s{v}
are both Lyndon words and \s{v} is of maximum size.
They then show that if \s{w} is a binary string ($\Sigma = \{a,b\}$),
the average length of \s{v} is asymptotically $3|\s{w}|/4$.
Thus each recursive application of RDuval yields a left Lyndon factor
of expected length $|\s{w}|/4$
and a remainder of length $3|\s{w}|/4$ to be factored further.
It follows that the expected number of recursive calls
of RDuval is $\mathcal{O}(\log_{4/3} n)$. Hence
\begin{lemm}
\label{lemm-RD}
On binary strings
RDuval executes in $O(n\log_{4/3} n)$ time on average.
\end{lemm}
\begin{examp}
For
\begin{equation*}
\label{ex1}
\begin{array}{cccccccccccccc}
& & \scriptstyle 1 & \scriptstyle 2 & \scriptstyle 3 & \scriptstyle 4 & \scriptstyle 5 & \scriptstyle 6 & \scriptstyle 7 & \scriptstyle 8 & \scriptstyle 9 & \scriptstyle 10 & \scriptstyle 11 & \scriptstyle 12 \\
\s{x} & = & a & a & b & a & a & b & b & a & b & b & a & b \\
\s{\lambda} & = & 12 & 2 & 1 & 9 & 3 & 1 & 1 & 3 & 1 & 1 & 2 & 1
\end{array}
\end{equation*}
the factors considered are first 1--12, then
\begin{itemize}
\item[$\bullet$]
2--3 and 4--12 in the first level of recursion;
\item[$\bullet$]
3, 5--7, 8--10 and 11--12 in the second level;
\item[$\bullet$]
$6,7,9,10,12$ in the third level.
\end{itemize}
Positions are assigned as follows:
$\s{\lambda}[1] \leftarrow 12; \s{\lambda}[2] \leftarrow 2, \s{\lambda}[4] \leftarrow 9;
\s{\lambda}[3] \leftarrow 1, \s{\lambda}[5] \leftarrow 3, \s{\lambda}[8] \leftarrow 3, \s{\lambda}[11] \leftarrow 2;
\s{\lambda}[6] \leftarrow 1, \s{\lambda}[7] \leftarrow 1, \s{\lambda}[9] \leftarrow 1,
\s{\lambda}[10] \leftarrow 1, \s{\lambda}[12] \leftarrow 1$.
\end{examp}
\subsection{NSV Applied to the Inverse Suffix Array}
\label{subsect-nsvisa}
The idea of the ``next smaller value'' (NSV) array for a given array (string) \s{x}
has been proposed in various forms and under various names
\cite{AGKR02,FMN08,OG11,GB13}.
\begin{defn}[Next Smaller Value]
Given an array $\s{x}[1..n]$ of ordered values,
$\mbox{NSV} = \mbox{NSV}_{\s{x}}[1..n]$ is the \itbf{next smaller value array} of \s{x}
if and only if for every $i \in 1..n$, $\mbox{NSV}[i] = j$, where
\begin{itemize}
\item[$(a)$]
for every $h \in 1..j\- 1,\ \s{x}[i] \le \s{x}[i\+ h]$; and
\item[$(b)$]
either $i\+ j = n\+ 1$ or $\s{x}[i] > \s{x}[i\+ j]$.
\end{itemize}
\end{defn}
\begin{examp}
\begin{equation*}
\label{ex1}
\begin{array}{rccccccccc}
\scriptstyle 1 & \scriptstyle 2 & \scriptstyle 3 & \scriptstyle 4 & \scriptstyle 5 & \scriptstyle 6 & \scriptstyle 7 & \scriptstyle 8 & \scriptstyle 9 & \scriptstyle 10 \\
\s{x} = 3 & 8 & 7 & 10 & 2 & 1 & 4 & 9 & 6 & 5 \\
\mbox{NSV}_{\s{x}} = 4 & 1 & 2 & 1 & 1 & 5 & 4 & 1 & 1 & 1
\end{array}
\end{equation*}
\end{examp}
As shown in various contexts in \cite{GB13},
$\mbox{NSV}_{\s{x}}$ can be computed in $\Theta(n)$ time using a stack.
Our main observation here, touched upon in \cite{HR03},
is that $\s{\lambda}_{\s{x}}$ can be computed merely by applying NSV
to the inverse suffix array $\mbox{ISA}_{\s{x}}$.
Proof of this claim can be found in Appendix 2;
here we present the very simple $\Theta(n)$-time, $\Theta(n)$-space algorithm for this calculation:
\begin{figure}[ht]
{\leftskip=3.4cm\obeylines\sfcode`;=3000
{\bf procedure\ } NSVISA$(\s{x}[1\mathinner{\ldotp\ldotp} n])\ :\ \s{\lambda}_{\s{x}}[1\mathinner{\ldotp\ldotp} n]$
Compute $SA_{\s{x}}$\ \ \ \ (see \cite{NZC09,PST07})
Compute $\mbox{ISA}_{\s{x}}$ from $\mbox{SA}_{\s{x}}$ in place\ \ \ \ (see \cite{PST07})
$\s{\lambda}_{\s{x}} \leftarrow$ NSV$(\mbox{ISA}_{\s{x}})$ (in place)
}
\caption{Apply NSV to $\mbox{ISA}_{\s{x}}$}
\label{fig-3line}
\end{figure}
\section{Elementary Computation of $\s{\lambda}_{\s{x}}$ Using Ranges}
\label{sect-newalgs}
In this section we describe an approach to the computation of $\s{\lambda}_{\s{x}}$
that applies a variant of the NSV idea to the ranges of \s{x}.
Figure~\ref{nsv*} gives pseudocode for Algorithm $\mbox{NSV}^*$
that uses the NSV stack ACTIVE to compute $\s{\lambda}$.
The processing identifies ranges in a single left-to-right scan
of \s{x}, making use of two range comparison routines, COMP and MATCH.
COMP compares adjacent individual ranges \s{x_r} and \s{x_{r+1}}, returning $\delta_1 = -1,0,+1$
according as $\s{x_r} < \s{x_{r+1}}$, $\s{x_r} = \s{x_{r+1}}$, $\s{x_r} > \s{x_{r+1}}$.
MATCH similarly returns $\delta_2$ for adjacent {\it sequences} of ranges; that is,
\begin{eqnarray*}
\s{X_r} &=& \s{x_rx_{r+1}\cdots x_{r+s}}, \mbox{ for some } s \ge 1; \\
\s{X_{r+s+1}} &=& \s{x_{r+s+1}x_{r+s+2}\cdots x_{r+s+t}}, \mbox{ for some } t \ge 1.
\end{eqnarray*}
Algorithm $\mbox{NSV}^*$ is based on the idea encapsulated in Lemma~\ref{maxlyn-suf}
of Appendix 2,
the main basis of the correctness of Algorithm NSVISA.
We process \s{x} from left to right, using a stack ACTIVE initialized with index 1.
At each iteration, the top of the stack (say, $j$)
is compared with the current index (say, $i$).
In particular, we need to compare $\s{s}_{\s{x}}(i)$ with $\s{s}_{\s{x}}(j)$,
where $\s{s}_{\s{x}}(i) \equiv \s{x}[i..n]$.
As long as $\s{s}_{\s{x}}(i) \succeq \s{s}_{\s{x}}(j)$,
$\mbox{NSV}^*$ pushes the current index and continues to the next.
When $\s{s}_{\s{x}}(i) \prec \s{s}_{\s{x}}(j)$,
it pops the stack and puts
appropriate values in the corresponding indices of $\s{\lambda}_{\s{x}}$.
As noted above, especially Observations~\ref{obs-decomp}--\ref{obs-monge},
ranges are employed to expedite these suffix comparisons.
\begin{figure}[ht]
{\leftskip=1.0cm\obeylines\sfcode`;=3000
{\bf procedure\ } NSV* $(\s{x},\s{\lambda})$
$\mbox{nextequal} \leftarrow 0^n;\ \mbox{period} \leftarrow 0^n$
{\bf push}$(\rm{ACTIVE}) \leftarrow 1$
\com{$\s{x}[n\+ 1] = \$$, a letter smaller than any in $\Sigma$.}
{\bf for\ } $i \leftarrow 2$ {\bf to\ } $n\+ 1$ {\bf do\ }
\quad $\mbox{prev} \leftarrow 0;\ j \leftarrow$ {\bf peek}$(\rm{ACTIVE})$
\com{COMP compares suffixes specified by $i,j$ of two ranges.}
\quad $\delta_1 \leftarrow \rm{COMP}(\s{x}[j],\s{x}[i]);\ \delta_2 \leftarrow 1$
\qquad {\bf while\ } ($\delta_1 \ge 0$ {\bf and\ } $\delta_2 > 0$) {\bf do\ }
\qquad\quad {\bf if\ } $\delta_1 = 0$ {\bf then\ } $\delta_2 \leftarrow \rm{MATCH}(\s{x}[j],\s{x}[i])$
\qquad\quad {\bf if\ } $\delta_2 > 0$ {\bf then\ }
\qquad\quad\q {\bf if\ } $\mbox{prev} = 0$ {\bf or\ } $\mbox{nextequal}[j] \neq \mbox{prev}$ {\bf then\ } $\s{\lambda}[j] \leftarrow i\!-\! j$
\qquad\quad\q {\bf else\ }
\qquad\qq\quad $\s{\lambda}[j] \leftarrow \mbox{offset} \leftarrow \mbox{prev}\!-\! j$
\qquad\qq\quad {\bf if\ } $\mbox{period}[\mbox{prev}]=0$ {\bf then\ }
\qquad\qq\quad\q {\bf if\ } $\s{\lambda}[\mbox{prev}] > \mbox{offset}$ {\bf then\ }
\qquad\qq\qquad\quad $\s{\lambda}[j] \leftarrow \s{\lambda}[j]\+ \s{\lambda}[\mbox{prev}]$
\qquad\qq\quad {\bf else\ }
\qquad\qq\quad\q {\bf if\ } $\mbox{nextequal}[j]=\mbox{prev}$ {\bf and\ } $\mbox{offset} \neq \s{\lambda}[\mbox{prev}]$ {\bf then\ }
\qquad\qq\qquad\quad $\s{\lambda}[j] \leftarrow \s{\lambda}[j]\+ \mbox{period}[\mbox{prev}]$
\qquad\qq\quad {\bf if\ } $\s{\lambda}[\mbox{prev}] = \mbox{offset}$ {\bf then\ }
\qquad\qq \com{Current position is a part of periodic substring}
\qquad\qq\qquad {\bf if\ } $\mbox{period}[\mbox{prev}] = 0$ {\bf then\ }
\qquad\qq\qquad\quad $\mbox{period}[j] \leftarrow \mbox{period}[\mbox{prev}]+2\times \mbox{offset}$
\qquad\qq\quad\q {\bf else\ }
\qquad\qq\qquad\quad $\mbox{period}[j] \leftarrow \mbox{period}[\mbox{prev}]\+ \mbox{offset}$
\qquad\quad\q {\bf pop}$(\rm{ACTIVE})$
\qquad\quad\q $\mbox{prev} \leftarrow j;\ j \leftarrow \bf{peek}(\rm{ACTIVE})$
\qquad\qq \com{Empty stack implies termination.}
\qquad\quad\q {\bf if\ } $j = 0$ {\bf then\ } \rm{EXIT}
\qquad\quad\q $\delta_1 \leftarrow \rm{COMP}(\s{x}[j],\s{x}[i])$
\com{Finished processing $i$ --- it goes to stack.}
\quad {\bf if\ } $\delta_2 = 0$ {\bf then\ } $\mbox{nextequal}[j] \leftarrow i$
\quad {\bf push}$(\rm{ACTIVE}) \leftarrow i$
}
\caption{Computing $\s{\lambda}_x$ using modified NSV}
\label{nsv*}
\end{figure}
Two auxiliary arrays,
\textbf{nextequal}\ and \textbf{period},
are required to handle situations in which MATCH
finds that a suffix of a previous range at position $j$
equals the current range at position $i$.
Thus, when $\delta_2 = 0$, the algorithm assigns
$\mbox{nextequal}[j] \leftarrow i$ before $i$ is pushed onto ACTIVE.
Then when a later MATCH yields $\delta_2 = 0$,
the value of \textbf{period}\ --- that is, the extent of the following periodicity --- may need to be set or adjusted,
as shown in the following example:
\begin{equation*}
\label{ex1}
\begin{array}{ccccccccccccccccc}
& & \scriptstyle 1 & \scriptstyle 2 & \scriptstyle 3 & \scriptstyle 4 & \scriptstyle 5 & \scriptstyle 6 & \scriptstyle 7 & \scriptstyle 8 & \scriptstyle 9 & \scriptstyle 10 & \scriptstyle 11 & \scriptstyle 12 & \scriptstyle 13 & \scriptstyle 14 & \scriptstyle 15 \\
\s{x} & = & a & a & a & b & a & a & b & a & a & b & a & a & b & a & b \\
\mbox{nextequal} & = & 0 & 5 & 0 & 0 & 8 & 0 & 0 & 11 & 0 & 0 & 0 & 14 & 0 & 0 & 0 \\
\mbox{period} & = & 0 & 12 & 0 & 0 & 9 & 0 & 0 & 6 & 0 & 0 & 0 & 4 & 0 & 0 & 0
\end{array}
\end{equation*}
A straightforward implementation of COMP and MATCH could require
a number of letter comparisons equal to the length of the shorter
of the two sequences of ranges being matched.
However, by performing $\Theta(n)$-time preprocessing,
we can compare two ranges in $\mathcal{O}(\sigma)$ time, where $\sigma = |\Sigma|$
is the alphabet size.
Given $\Sigma = \{\mu_1,\mu_2,\ldots,\mu_{\sigma}\}$,
we define Parikh vectors $P_r[1..\sigma]$,
where $P_r[j]$ is the number of occurrences of $\mu_j$ in range $\s{x_r}$.
Since ranges are monotone nondecreasing in the letters of the alphabet,
it is easy to compute all the $P_r, r = 1,2,\ldots,m$, in linear time
in a single scan of \s{x}.
Similarly, during the processing of each range \s{x_r},
any value $P_{r,j}$, the Parikh vector of the suffix $\s{x_r}[j..\ell_r]$,
can be computed in constant time for each position considered.
Thus we can determine the lexicographical order of any two ranges (or part ranges)
\s{x_r} and \s{x_{r'}} in $\mathcal{O}(\sigma)$ time
rather than time $\mathcal{O}(\max(\ell_r,\ell_{r'}))$.
The variant of $\mbox{NSV}^*$ that uses Parikh vectors is called P$\mbox{NSV}^*$;
otherwise NP$\mbox{NSV}^*$ for Not Parikh.
In Appendix 3 we describe briefly another approach to
this suffix comparison problem, which also
achieves run time $\mathcal{O}(n \log n)$
by maintaining a simple data structure requiring $\mathcal{O}(n \log n)$ space.
Now consider the worst case behaviour of Algorithm $\mbox{NSV}^*$.
Given the initial string $\s{x_0} = a^hba^hc_0,\ h \ge 1$,
$c_0 > b > a$,
let $\s{x^{(h)}_k} = \s{x_k} = \s{x_{k-1}x^*_{k-1}},\ k = 1,2,\ldots,$
with \s{x^*_{k-1}} identical to \s{x_{k-1}} except in the last position,
where the letter $c_k > c_{k-1}$ replaces $c_{k-1}$.
Then \s{x_k} has length $n = (h\+ 1)m$,
where $m = 2^{k+1}$ is the number of ranges in \s{x_k}.
In Appendix 4 it is shown in Lemma~\ref{lemm-rm}
that $\s{x_k}$
is a worst-case input for Algorithm $\mbox{NSV}^*$,
which requires $\mathcal{O}(n\log n)$ range matches in such cases.
Since $\mbox{PNSV}^*$ compares two ranges in $\mathcal{O}(\sigma)$ time,
it therefore requires $\mathcal{O}(\sigma n\log n)$ time
in the worst case, thus $\mathcal{O}(n\log n)$ for constant $\sigma$.
In Appendix 4 we argue that $\mbox{NPNSV}^*$ is also $\mathcal{O}(n\log n)$
in the worst case.
\section{Experimental Results}
\label{sect-exp}
We have done preliminary tests on the algorithms described above,
including the two variants of $\mbox{NSV}^*$.
The equipment used was an Intel(R) Core i3 at 1.8GHz and 4GB main memory
under a 64-bit Windows 7 operating system.
Figure~\ref{fig-binary} shows the results of exhaustive tests of the algorithms on all
binary strings of lengths 11--22,
with all but RDuval displaying linear-time behaviour.
MaxLyn and NP$\mbox{NSV}^*$ are roughly equivalent in time requirement,
with NSVISA several times slower, P$\mbox{NSV}^*$ perhaps 10 times slower.
We have also tested the linear average-case algorithms on much longer binary strings,
several megabytes in length,
both random and highly periodic \cite{FSS03}.
On random strings, P$\mbox{NSV}^*$ and NP$\mbox{NSV}^*$ are comparable in speed
and fastest by a factor of 2 or 3,
while on the periodic strings, MaxLyn has an advantage by approximately
the same margin.
More testing needs to be done,
especially on strings defined on larger alphabets,
but of the current collection,
it appears that the two new $\mathcal{O}(n\log n)$-time algorithms
are the algorithms of choice.
\begin{figure}[htbp]
\centering
\includegraphics*[scale = 0.6]{28janplot.pdf}
\caption{Five algorithms compared on all binary strings of lengths $n \in 11..22$: the average processing time for each $n$ is given in $10^{-4}$ seconds.}
\label{fig-binary}
\end{figure}
\section{Future Work}
\label{sect-future}
There is reason to believe \cite{K14}
that the Lyndon array computation is less hard than
suffix array construction.
Thus the authors conjecture that there is a linear-time elementary
algorithm (no suffix arrays) to compute the Lyndon array.
\defInternational Conference on Algorithmic Aspects in Information \& Management{International Conference on Algorithmic Aspects in Information \& Management}
\defAustralasian J.\ Combinatorics{Australasian J.\ Combinatorics}
\defAustralasian Workshop on Combinatorial Algs.{Australasian Workshop on Combinatorial Algs.}
\defAnnual Symp.\ Combinatorial Pattern Matching{Annual Symp.\ Combinatorial Pattern Matching}
\defAnnual International Computing \& Combinatorics Conference{Annual International Computing \& Combinatorics Conference}
\defIEEE Symp.\ Found.\ Computer Science{IEEE Symp.\ Found.\ Computer Science}
\defAnnual European Symp.\ on Algs.{Annual European Symp.\ on Algs.}
\defInternat.\ Conf.\ on Language \& Automata Theory \& Applications{Internat.\ Conf.\ on Language \& Automata Theory \& Applications}
\defInternat.\ Workshop on Combinatorial Algs.{Internat.\ Workshop on Combinatorial Algs.}
\defAustralasian Workshop on Combinatorial Algs.{Australasian Workshop on Combinatorial Algs.}
\defSymp.\ Theoretical Aspects of Computer Science{Symp.\ Theoretical Aspects of Computer Science}
\defInternat.\ Colloq.\ Automata, Languages \& Programming{Internat.\ Colloq.\ Automata, Languages \& Programming}
\defInternat.\ J.\ Foundations of Computer Science{Internat.\ J.\ Foundations of Computer Science}
\defInternat.\ Symp.\ Algs.\ \& Computation{Internat.\ Symp.\ Algs.\ \& Computation}
\defScandinavian Workshop on Alg.\ Theory{Scandinavian Workshop on Alg.\ Theory}
\defPrague Stringology Conf.{Prague Stringology Conf.}
\defAlgorithmica {Algorithmica }
\defACM Computing Surveys{ACM Computing Surveys}
\defDiscrete Applied Math.{Discrete Applied Math.}
\defFundamenta Informaticae{Fundamenta Informaticae}
\defInform.\ Process.\ Lett.\ {Inform.\ Process.\ Lett.\ }
\defInform.\ \& Computation{Inform.\ \& Computation}
\defInform.\ Sciences{Inform.\ Sciences}
\defJ.\ Assoc.\ Comput.\ Mach.\ {J.\ Assoc.\ Comput.\ Mach.\ }
\defJ.\ Combinatorial Theory, Series A{J.\ Combinatorial Theory, Series A}
\defCommun.\ Assoc.\ Comput.\ Mach.\ {Commun.\ Assoc.\ Comput.\ Mach.\ }
\defMath.\ in Computer Science{Math.\ in Computer Science}
\defSIAM J.\ Computing{SIAM J.\ Computing}
\defSIAM J.\ Discrete Math.\ {SIAM J.\ Discrete Math.\ }
\defJ.\ Computational Biology\ {J.\ Computational Biology\ }
\defJ.\ Algorithms {J.\ Algorithms }
\defJ.\ Discrete Algorithms {J.\ Discrete Algorithms }
\defJ.\ Automata, Languages \& Combinatorics\ {J.\ Automata, Languages \& Combinatorics\ }
\defACM-SIAM Symp.\ Discrete Algs.\ {ACM-SIAM Symp.\ Discrete Algs.\ }
\defSoftware, Practice \& Experience\ {Software, Practice \& Experience\ }
\defThe Computer Journal{The Computer Journal}
\defTheoret.\ Comput.\ Sci.\ {Theoret.\ Comput.\ Sci.\ }
\bibliographystyle{plain}
|
\section{Introduction}
Compared to heat transfer by phonons in crystalline materials, heat transfer in amorphous materials is complicated by the existence of three regimes of vibrational modes.\cite{allen99} Low-frequency propagons are delocalized and have a well-defined wave vector and group velocity,\cite{feldman93} similar to phonons in crystals, while high-frequency locons are localized and contribute negligibly to thermal conduction.\cite{leitner01} Diffusons have intermediate frequencies and are delocalized, but do not have well-defined wave vectors or group velocities. The contribution of diffusons to thermal conduction can be notable, however, as they occupy the majority of the vibrational spectrum.\cite{feldman93}
From kinetic theory,\cite{ziman} the contribution of an individual phonon or propagon mode to thermal conductivity is proportional to its mean free path (MFP). Because diffusons do not have a well-defined group velocity, it is not clear if they have a MFP or how it can be defined. Their contribution to thermal conductivity can be predicted using their diffusivity, which is well-defined and can be calculated from Allen-Feldman theory.\cite{allen93,feldman93} Nevertheless, it would be insightful to identify a frequency-dependent length scale for propagons and diffusons describing the decay of the heat flux at each vibrational frequency. Such a definition would lift the (fundamentally) arbitrary distinction between propagons and diffusons and enable a unified description of heat transfer at all vibrational frequencies.
In this paper, we apply the spectrally-decomposed MFP method\cite{saaskilahti15} to probe the non-equilibrium MFPs of vibrational heat carriers in amorphous silicon (a-Si). This method is based on calculating the spectrally-decomposed heat current (SDHC)\cite{saaskilahti14b} in systems of different lengths using non-equilibrium molecular dynamics (NEMD) simulations. The MFPs are determined from the variation of the SDHC as a function of system length at each vibrational frequency.\cite{saaskilahti15b}
We previously used the spectrally-decomposed MFP method to calculate the non-equilibrium MFPs in low-dimensional systems such as carbon nanotubes\cite{saaskilahti15} and anharmonic chains.\cite{saaskilahti15b} We demonstrated that the non-equilibrium MFPs transparently describe the ballistic-to-diffusive transition in the length-dependence of thermal conductivity. Compared to previous calculations for a-Si, the spectrally-decomposed MFP method has several advantages. Unlike in modal life-time calculations,\cite{he11} we do not need to estimate the group velocities of individual modes to calculate their MFPs. We also do not need to distinguish between propagons and diffusons\cite{larkin14} nor resort to the harmonic approximation.\cite{feldman93} In contrast to recent calculations studying the spectral conductivity of a-Si in fixed-size systems, \cite{lv15,zhou15} we focus on the MFPs of heat carriers and the system-size dependence of thermal conductivity.
The rest of the paper is organized as follows. The calculation methods are presented in Sec. \ref{sec:setup} and the numerical results are discussed in Sec. \ref{sec:results}. We also introduce a simple method for the quantum correction of thermal conductivity from classical MD simulations, based on weighting the SDHC by the quantum occupation function. Because this quantum correction method operates at the frequency level, it is more reasonable than quantum-correction methods that operate at the system level (see Ref. \citenum{turney09a} and references therein) and allows us to compare our predictions to experimental measurements.
\section{Simulation setup and methods} \label{sec:setup}
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.6cm]{fig1.ps}
\caption{Schematic illustration of the a-Si system for $L=10$ nm. The spectral heat flux $q(\omega)$ is calculated at the cross-section in the middle of the structure (dashed line). The length $L$ between the Langevin heat baths is varied to extract the vibrational mode MFPs based on the decrease of $q(\omega)$ as a function of $L$.}
\label{fig:geom}
\end{center}
\end{figure}
All simulations are carried out using the LAMMPS package\cite{plimpton95} with a time step of 2.5 fs. The Si-Si interactions are modeled by the Stillinger-Weber potential\cite{stillinger85} with the parameters of Ref. \citenum{vink01}. The NEMD simulation geometry is shown in Fig. \ref{fig:geom}. The atomic coordinates for a-Si are generated by following the melt-quench procedure of Ref. \citenum{france-lanord14} and the final density is 2,291 kg/m$^3$. After the equilibration of the quenched system, atoms located within a distance $L_{\textrm{bath}}=5$ nm from the left and right edges of the structure are coupled to Langevin heat baths at temperatures $T_H=T+\Delta T/2$ and $T_C=T-\Delta T/2$ with bath relaxation times of 1 ps. To prevent sublimation, atoms at the far left and right edges are fixed to their equilibrium positions. Periodic boundary conditions are applied at the boundaries transverse to the current flow. The width of the system cross-section is $7$ nm. System lengths $L$ (i.e., the region between the baths) between 1 and 10 nm at intervals of 1 nm are considered.
The SDHC is calculated through the plane of decomposition located halfway between the hot and cold baths (dashed line in Fig. \ref{fig:geom}).\cite{saaskilahti15,saaskilahti15b} The SDHC $q_{i \to j}(\omega)$ between particles $i$ and $j$ located on opposite sides of this plane is given by the pair-wise SDHC equation \cite{saaskilahti14b}
\begin{equation}
q_{i \to j}(\omega) = -\frac{2}{t_{\textrm{simu}} \omega} \sum_{\alpha,\beta\in\{x,y,z\}} \textrm{Im}\left.\left\langle \hat{v}_i^{\alpha}(\omega)^* K_{ij}^{\alpha\beta} \hat{v}_j^{\beta}(\omega) \right\rangle \right. \label{eq:qomega_expr},
\end{equation}
where $t_{\textrm{simu}}$ is the simulation time, $\omega$ is the angular frequency, and the interatomic force constant $K_{ij}^{\alpha\beta}$ is defined as
\begin{equation}
K_{ij}^{\alpha\beta} = - \left. \frac{\partial ^2 \mathcal{V}}{\partial u_i^{\alpha} u_j^{\beta}} \right|_{\mathbf{u}=\mathbf{0}}. \label{eq:K}
\end{equation}
The velocities $\hat v_i^{\alpha}(\omega)$ and $\hat v_j^{\beta}(\omega)$ are the discrete Fourier transforms of the atomic velocities $v_i^{\alpha}(t)=\dot{u}_i^{\alpha}(t)$ and $v_j^{\beta}(t)=\dot{u}_j^{\beta}(t)$ (the exact definitions are in Ref. \citenum{saaskilahti15}), where $u_i^{\alpha}$ and $u_j^{\beta}$ are the displacements of atoms $i$ and $j$ from their equilibrium positions in directions $\alpha$,$\beta \in \{x,y,z\}$. In Eq.~\eqref{eq:K}, $\ca{V}$ is the interatomic potential energy function. The spectral flux through the plane of decomposition is obtained from Eq.~\eqref{eq:qomega_expr} by summing over all pairs of atoms (one of the left side, denoted by $\tilde{L}$, and one on the right side, denoted by $\tilde{R}$) within the potential cut-off distance of each other and dividing by the interface area $A$:
\begin{equation}
q(\omega) = \frac{1}{A} \sum_{i\in \tilde L} \sum_{j\in \tilde R} q_{i \to j}(\omega). \label{eq:qomega_sum}
\end{equation}
While Eq.~\eqref{eq:qomega_expr} is the first-order approximation to the inter-particle SDHC,\cite{saaskilahti14b} we have confirmed that the contribution of higher-order terms is negligible for a-Si by comparing the integral of Eq. \eqref{eq:qomega_sum}, which we denote as $Q$, to the total flux determined from the work done by the heat baths. The two results agree within 4\%. We attribute this good agreement to the stiffness of the interatomic bonds in a-Si, which ensures that the first-order term in the current (proportional to the harmonic force constants $K_{ij}^{\alpha \beta}$) dominates the higher-order terms that are related to anharmonic force constants. This restriction to the first-order current term at the plane of decomposition does not, however, mean that anharmonic scattering in the bulk is neglected, because all anharmonic effects are included in the NEMD simulations.\cite{saaskilahti14b,saaskilahti15}
Frequency-dependent MFPs $\Lambda(\omega)$ are calculated by determining $q(\omega,L)$ for different system lengths $L$ and fitting the length-dependent $q(\omega,L)$ to the equation \cite{saaskilahti15,saaskilahti15b}
\begin{equation}
q(\omega,L) = \frac{q^0(\omega)}{1+L/[2\Lambda(\omega)]}, \label{eq:mfp}
\end{equation}
where $q^0(\omega)$ is the spectral flux when the baths are in contact ($L\to 0^+$). Both $q^0(\omega)$ and $\Lambda(\omega)$ are determined from the fitting procedure. While $q^0(\omega)$ depends on the details of the heat baths, the MFPs extracted from the length-dependence are not expected to depend on the bath details.\cite{saaskilahti15b} The frequency-dependent MFPs determined from Eq. \eqref{eq:mfp} are mode-averaged and projected along the direction of heat transfer.\cite{saaskilahti15} The MFPs are independent of the system length and therefore correspond to the bulk values.\cite{saaskilahti15} We note that the MFPs determined from Eq. \eqref{eq:mfp} correspond to the decay length of the heat flux. This definition does not necessarily coincide with the conventional definition of the MFP as the decay length of a wave packet.\cite{ziman} This distinction is important for diffusons, which do not have a well-defined wave-vector so that the traditional definition cannot be applied.
Once the spectral MFPs are determined, the thermal conductivity $\kappa$ for length $L$ can be determined from\cite{saaskilahti15}
\begin{equation}
\kappa = \frac{QL}{\Delta T}= \frac{L}{\Delta T} \int_0^{\infty} \frac{d\omega}{2\pi} \frac{q^0(\omega)}{1+L/[2\Lambda(\omega)]}. \label{eq:kappa_est}
\end{equation}
The length-dependence of the thermal conductivity can be intuitively understood by writing Eq. \eqref{eq:kappa_est} in the equivalent form
\begin{equation}
\kappa = \frac{2}{\Delta T} \int_0^{\infty} \frac{d\omega}{2\pi} q^0(\omega) \underbrace{\frac{1}{(L/2)^{-1}+\Lambda(\omega)^{-1}}}_{\Lambda_{\textrm{eff}}(\omega)^{-1}},
\end{equation}
where the ``effective'' MFP $\Lambda_{\textrm{eff}}(\omega)$ has been introduced, which is similar to the well-known Matthiessen rule.\cite{chen} The effective MFP accounts for boundary scattering through the additional $L/2$ term and is limited to below this value.
Finally, Eq.~\eqref{eq:kappa_est} allows for a simple quantum correction to the thermal conductivity prediction, because the contributions of different frequencies can be weighted by the vibrational mode energy and occupation,
as in the Landauer-B\"uttiker formalism \cite{rego98,angelescu98,saaskilahti13}. We define the quantum corrected thermal conductivity as
\begin{equation}
\kappa = \frac{L}{\Delta T} \int_0^{\infty} \frac{d\omega}{2\pi} \frac{q^0(\omega)}{1+L/[2\Lambda(\omega)]} \frac{\hbar \omega}{k_\mathrm{B}} \frac{\partial f_{\mathrm {BE}}(\omega,T)}{\partial T}, \label{eq:kappa_quantum}
\end{equation}
where $f_\mathrm{BE}(\omega,T)=\left[ \exp(\hbar \omega/k_\mathrm{B} T)-1\right]^{-1}$ is the Bose-Einstein distribution function, $k_\mathrm{B}$ is the Boltzmann constant, and $\hbar$ is the Planck constant divided by $2\pi$. By defining the dimensionless, length-dependent bath-to-bath transmission function as
\begin{equation}
\ca{T}(\omega,L) = \frac{q^0(\omega)A}{k_\mathrm{B} \Delta T} \frac{1}{1+L/[2\Lambda(\omega)]},
\end{equation}
Eq. \eqref{eq:kappa_quantum} can be written in the familiar Landauer-B\"uttiker form as \cite{rego98}
\begin{equation}
\kappa = \frac{L}{A} \int_0^{\infty} \frac{d\omega}{2\pi} \hbar \omega \ca{T}(\omega,L) \frac{\partial f_B(\omega,T)}{\partial T}. \label{eq:kappa_quantum2}
\end{equation}
The proposed quantum correction accounts for the quantum specific heat of the modes at each frequency, but does not account for quantum effects in the dynamics. The method is thus similar to the one recently introduced by Lv and Henry,\cite{lv15} who weight the modal contributions to the equilibrium Green-Kubo thermal conductivity by the quantum population function.
All our NEMD simulations are performed at a mean temperature of $T=300$ K with temperature bias $\Delta T=100$ K. Choosing a relatively large temperature bias allows for very good signal-to-noise ratio in the spectral heat flux, suppressing the statistical noise. We checked that halving the bias to $\Delta T=50$ K does not change the spectral MFPs. In addition, we checked that the heat flux is not sensitive to the exact arrangement of atoms arising from the melt-and-quench procedure, which we attribute to the large cross-section of the system giving rise to spatial averaging in the spectral currents. Therefore, we performed a single melt-quench for each system length.
\section{Results}\label{sec:results}
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.6cm]{fig2.eps}
\caption{Spectral heat flux $q(\omega)$ for different system lengths $L$. Increasing the system length reduces the heat flux, especially at high frequencies, where the MFPs are shorter. At frequencies below 2 THz, the heat current is independent of system length, suggesting ballistic transport.}
\label{fig:flux}
\end{center}
\end{figure}
The spectral heat flux $q(\omega)$ for selected system lengths $L$ as a function of frequency $f=\omega/(2\pi)$ is plotted in Fig. \ref{fig:flux}. The spectral distribution of the heat flux for $L=20$ nm was recently analyzed in detail by Zhou and Hu,\cite{zhou15} so we focus here on its length-dependence. As expected, increasing the system length reduces the heat current throughout the whole frequency range because of increased phonon-phonon scattering. The reduction is strongest at high frequencies, where the MFPs are shorter compared to low frequencies. At frequencies less than 2 THz, the spectral current is nearly independent of system length. Such nearly ballistic conduction suggest that the low frequency MFPs are notably longer than the system sizes considered here.
Equation \eqref{eq:mfp} suggests that the inverse of the spectral flux will be linearly proportional to the system length, with the slope given by $[2\Lambda(\omega)]^{-1}$. To determine the MFPs, we calculated the spectral flux for system sizes $L \in \{1,2,\dots,10\}$ nm, plotted $q(\omega)^{-1}$ versus $L$ and fitted a linear function at each frequency using least squares fitting.\cite{saaskilahti15,saaskilahti15b} A linear function accurately reproduces the length-dependence of $q(\omega)^{-1}$ (not shown), as previously also observed for other systems.\cite{saaskilahti15,saaskilahti15b}
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.6cm]{fig3.eps}
\caption{Log-log plot of the spectral MFPs determined by fitting to Eq. \eqref{eq:mfp}. The shaded regions correspond to the 95 \% confidence interval.}
\label{fig:mfps}
\end{center}
\end{figure}
Because of the high computational cost associated with calculating the spectral heat fluxes for large systems, we limited our study to systems at most 10 nm long. This restriction precludes extracting MFPs longer than 10 nm accurately, limiting the current analysis to frequencies greater than 2 THz. The MFPs $\Lambda(\omega)$ extracted from the linear fitting procedure are shown in a log-log plot in Fig.~\ref{fig:mfps}. At frequencies below 5 THz, the MFPs obey a power-law scaling $\Lambda(\omega) \propto \omega^{-2}$. This scaling agrees with modal life-time calculations on a-Si.\cite{larkin14} At frequencies greater than 5 THz, the power-law scaling breaks down and the MFPs increase with increasing frequency, giving rise to a local maximum around 8 THz. A similar maximum for a-Si has been observed in effective MFPs \cite{he11} and in lifetimes.\cite{larkin14} At higher frequencies, the MFPs decrease again and fall below 1 nm, which is on the order of the silicon-silicon bond length. At such high frequencies, the uncertainty is large because of the sensitivity of the spectral flux to the system size.
Larkin and McGaughey reported a propagon-diffuson transition frequency of 1.8 THz,\cite{larkin14} such that the frequency range considered in Fig. \ref{fig:mfps} mostly corresponds to diffuson-like modes. In the analysis below, we assume that the scaling $\Lambda(\omega) \propto \omega^{-2}$ (solid line in Fig. \ref{fig:mfps}) remains valid at frequencies below 2 THz. With such scaling, the MFPs exceed 100 nm below 530 GHz and 1 $\mu$m below 170 GHz. While the $\omega^{-2}$ scaling may break down in real situations because of defects, boundary scattering, or even the onset of a Rayleigh-like $\omega^{-4}$ scaling at very low frequencies, \cite{feldman93} we assume it to hold for simplicity.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.6cm]{fig4.eps}
\caption{Thermal conductivity versus system length. The thermal conductivities calculated from direct NEMD simulation are marked by circles and the estimated thermal conductivity from Eq.~\eqref{eq:kappa_est} using classical statistics is indicated by the solid line. The error bars in the NEMD thermal conductivities correspond to the 95\% confidence interval.}
\label{fig:kappa}
\end{center}
\end{figure}
We now investigate the length-dependence of the thermal conductivity using Eq.~\eqref{eq:kappa_est}. The calculated thermal conductivity (continuous line) is compared with that determined directly from NEMD simulations (data points) for lengths up to $100$ nm in Fig.~\ref{fig:kappa}. In the evaluation of the integral in Eq.~\eqref{eq:kappa_est}, the MFPs have been assumed to scale as $\Lambda(\omega)\propto \omega^{-2}$ at frequencies below 2 THz.
These calculations were carried out without the quantum correction as the NEMD simulations are classical. Equation \eqref{eq:kappa_est} combined with the MFP data of Fig. \ref{fig:mfps} reproduces the length-dependence of thermal conductivity up to lengths $L=100$ nm to within 2\%. This close agreement (i) supports the assumption of $\Lambda(\omega)\propto \omega^{-2}$ scaling at low frequencies, (ii) shows that the MFP data in Fig. \ref{fig:mfps}, which were determined from simulations of systems shorter than $10$ nm, can be reliably used to estimate the relative contributions of different vibrational frequencies to thermal transport in much larger systems, and (iii) provides support for the accuracy of Eq. \eqref{eq:kappa_est} in describing the length-dependence.
Because the Debye temperature of a-Si is $530$ K \cite{mertig84}, which is well above room temperature, we need to apply the quantum correction to compare the predicted temperature-dependence of thermal conductivity to experimental data. To do so, we use Eq.~\eqref{eq:kappa_quantum} and evaluate the integral as a function of temperature using the MFPs from Fig. \ref{fig:mfps}, again assuming the scaling $\Lambda(\omega)\propto \omega^{-2}$ at frequencies below 2 THz. A full quantum-corrected analysis would require determining the MFPs at each temperature. For simplicity and based on the recent results of Lv and Henry \cite{lv15}, we assume that the MFPs calculated at a temperature of 300 K remain valid at other temperatures.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.6cm]{fig5.eps}
\caption{Quantum-corrected thermal conductivity versus temperature for system lengths of 50, 250, and 1000 nm. The MFPs are assumed to scale as $\Lambda(\omega)\propto \omega^{-2}$ below frequencies of 2 THz and to be independent of temperature. Also plotted is the thermal conductivity of a 520 nm thick hydrogenated a-Si thin film measured by Cahill \textit{et al.} \cite{cahill94}.}
\label{fig:kappa_qm}
\end{center}
\end{figure}
The quantum-corrected thermal conductivity [Eq. \eqref{eq:kappa_quantum}] is plotted as a function of temperature for system lengths of 50, 250, and 1000 nm in Fig.~\ref{fig:kappa_qm}. As noted above, assuming finite $L$ in Eq. \eqref{eq:kappa_quantum} limits the MFPs to $L/2$. Experimental data from Cahill {\em et al.} for a 520 nm thick film of hydrogenated a-Si with one atomic percent hydrogen content are also plotted.\cite{cahill94} Because available thermal conductivity measurements for a-Si contain significant scatter,\cite{larkin14} we use the data of Cahill \textit{et al.} to check the trend of our predictions, but do not expect agreement. Differences may also exist due to the use of the Stillinger-Weber potential and the classical nature of the NEMD simulations. The increase of thermal conductivity with increasing temperature is well described by the quantum-corrected thermal conductivity. At temperatures higher than 300 K, the experimentally measured thermal conductivity increases slightly slower as a function of temperature than our prediction, but this disagreement may be related to our approximation that the MFPs are independent of temperature. At such high temperatures, anharmonic scattering will reduce MFPs and therefore decrease the thermal conductivity. Without the quantum-correction, the predicted thermal conductivity would depend only very weakly on temperature (due to the weak temperature-dependence of the MFPs), precluding reasonable agreement with the trend of the experimental data. We caution that this quantum correction has only been examined for a-Si here and that its application to other systems warrants further investigation.
\section{Conclusion}
\label{sec:conclusion}
We investigated vibrational heat transfer in a-Si by determining the SDHC and MFPs from NEMD simulations. The calculated MFPs directly reflect the decay of the heat flux at each vibrational frequency and do not rely on the existence of a well-defined modal wave vector, thereby avoiding the separate treatment of diffusons and propagons. As shown in Fig. \ref{fig:mfps}, the MFPs exhibit $\omega^{-2}$ scaling at frequencies above 2 THz and below 5 THz. At frequencies higher than 10 THz, the MFPs fall below 1 nm, corresponding to strongly localized vibrations. The length-independent MFPs can be used to accurately predict the thermal conductivity in systems as long as 100 nm (Fig. \ref{fig:kappa}). Weighting the SDHC by the frequency-dependent quantum occupation function provides a simple method for a quantum-correction of thermal conductivity and is able to reproduce the experimentally measured temperature-dependence of thermal conductivity, as shown in Fig. \ref{fig:kappa_qm}.
In the future, it would be useful to calculate the SDHC for systems longer than those considered here, enabling direct extraction of MFPs at frequencies below 2 THz. Such an analysis could inform the ongoing discussion\cite{larkin14} of the low-frequency scaling of MFPs in a-Si. It will also be important to compare the non-equilibrium MFPs to those calculated from equilibrium molecular dynamics simulations.
\section{Acknowledgements}
This work was initiated during the research visit of K.S. to Carnegie Mellon University. The computational resources were provided by the Finnish IT Center for Science and Aalto Science-IT project. The work was partially funded by the Aalto Energy Efficiency Research Programme (AEF) and the Academy of Finland.
|
\section{Introduction}
One way to search for physics beyond the Standard Model is to search
for decays that are forbidden or predicted to occur at a negligible level.
Observing such decays would constitute evidence for new physics, and
measuring their branching fractions would provide insight into how to
modify our theoretical understanding.
For example, the absence of flavor-changing neutral currents (FCNCs) in
kaon decays led to the prediction of the charm quark \cite{ref::cleo1},
and the observation of $B^0-\bar B^0$ mixing, a FCNC process, indicated
a very large top-quark mass \cite{ref::cleo2}. Till now, the rare and
forbidden charm decays have been less informative and less extensively
studied.
FCNC processes in charm decays are highly suppressed by the
Glashow-Iliopoulos-Maiani (GIM) mechanism \cite{ref::gim}, and can only
occur via higher-order diagrams within SM , but the estimated branching
fractions are $10^{-8}$ to $10^{-6}$ \cite{ref::fcnc}.
Such a small branching fractions are touching the sensitivity of current
experiments. However, if additional new particles or mechanism exist, they
could contribute additional amplitudes that would make these modes observable.
Thus, the hints of $D^+$ FCNC decays might provide indication of non-SM physics
or of unexpectedly large rates, such as at $10^{-5}$ or $10^{-6}$ level \cite{ref::long-distance},
for long-distance SM processes $D^+\to\pi^+ V$, $V\to e^+e^-$, with a real or
virtual vector meson $V$ (can be a $\rho$, $\omega$, or $\phi$).
The LNV decays $D^+\to K^-e^+e^+$ and $D^+\to\pi^-e^+e^+$ are strictly
forbidden in the SM. They could be induced by a Majorana neutrino, but with
a branching fraction only of order $10^{-23}$ \cite{ref::majorana}. So any
observation at experimentally accessible levels would be clear evidence of new physics.
The searches for these decay modes have been carried out in several experiments
(see table \ref{tab::historical}).
In this document, we present latest results of searching for the FCNC
decays of $D^0\to\gamma\gamma$, $D^+\to K^+e^+e^-$ and $D^+\to\pi^+e^+e^-$
together with the lepton-number violating (LNV) decays of $D^+\to K^-e^+e^+$
and $D^+\to\pi^-e^+e^+$ at the BESIII experiment.
\begin{table*}[htbp]
\begin{center}
\caption{Comparisons of the upper limits ($10^{-6}$) on the branching fractions for $D^+\to h^{\pm}e^{\mp}e^{+}$ at a 90\% C.L..}\label{tab::historical}
\begin{tabular}{lccccccc}
\hline
Experiments & $D^+\to K^+e^+e^-$ & $D^+\to K^-e^+e^+$ & $D^+\to\pi^+e^+e^-$ & $D^+\to \pi^-e^+e^+$ \\
\hline
CLEO \cite{ref::cleo-1988} & - & - & 2600 & - \\
MARK2 \cite{ref::mark-1990} & 4800 & 9100 & 2500 & 4800 \\
E687 \cite{ref::e687-1997} & 200 & 120 & 110 & 110 \\
E791 \cite{ref::e791-1999} & 200 & - & 52 & 96 \\
CLEO \cite{ref::cleo-2010} & 3.0 & 3.5 & 5.9 & 1.1 \\
Babar \cite{ref::babar-2011}& 1.0 & 0.9 & 1.1 & 1.9 \\
PDG \cite{ref::pdg2014} & 1.0 & 0.9 & 1.1 & 1.1 \\
\hline
This work & 1.2 & 0.6 & 0.3 & 1.2 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\section{Technique}
For the $e^+e^-$ annihilation experiment around the $\psi(3770)$ peak,
the $D$ mesons are produced in pairs, i.e., if a $D$ meson is reconstructed
in an event, which is called a {\it singly tagged $D$ event}, there must exist
a $\bar D$ meson in the recoiling side. If the pair $D\bar D$is fully
reconstructed in an event, the event is called a {\it doubly tagged $D\bar D$ event}.
Traditionally, there are two methods to perform the searching for the rare/forbidden
decays. One is based on singly tagged events which will provide large statics with
high backgrounds, another is using doubly tagged events presenting extremly low
backgrounds while bad statics (see table \ref{tab::technique}).
On which technique a searching will employ depends on both background contaminations
and the statistics.
\begin{table*}[htbp]
\begin{center}
\caption{Two techniques on rare/forbidden searching.}\label{tab::technique}
\begin{tabular}{lccccccc}
\hline
Method & Statistics (charged/neutral) & Background & Sensitivity \\
\hline
Single Tag Method & $1.7\times10^7$/$2.1\times10^7$ & not good & Bkg. vs Stat. \\
Double Tag Method & $1.6\times10^6$/$2.8\times10^6$ & clean & Bkg. vs Stat. \\
\hline
\end{tabular}
\end{center}
\end{table*}
\section{$D^+\to h^{\pm}e^+e^{\mp}$}
To searching the decays of $D^+\to h^{\pm}e^+e^{\mp}$, where $h$ means $K$/$\pi$, we
check the accepted events in the signal boxes in the scatter plots of $M_{\rm BC}$
versus $\Delta E$ which are shown in figure \ref{fig::box-data}.
The signal box, which are kept blind before cuts optimizing and backgrounds studies
based on MC and sideband data, is deined with the mean value and the resolution of $M_{\rm BC}$
and $\Delta E$ determined from MC simulations, i.e.
$|\Delta E -\Delta E_{\rm mean}|<3\sigma_{\Delta E}$
and
$|M_{\rm BC}- M_{D^+}| <3\sigma_{M_{D^+}}$,
where $\Delta E_{\rm mean}$ is the mean value of the $\Delta E$ distribution,
$M_{D^+}$ is the $D^+$ nominal mass,
$\sigma_{\Delta E}$ and $\sigma_{M_{D^+}}$ are the corresponding resolutions.
Events falling into the signal box, shown as blue rectangle in figure \ref{fig::box-data},
are taken as candidate signal events.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[height=9cm,width=14cm]{box-data.eps}
\put(-220,-10){\bf \large $\Delta E$ [GeV]}
\put(-400,90){\rotatebox{90}{\bf \large $M_{\rm BC}$ [GeV/$c^2$]} }
\put(-325,235){\bf $D^+\to K^+e^+e^-$}
\put(-330,225){\bf BESIII preliminary}
\put(-125,235){\bf $D^+\to \pi^+e^+e^-$}
\put(-130,225){\bf BESIII preliminary}
\put(-325,107){\bf $D^+\to K^-e^+e^+$}
\put(-330,97){\bf BESIII preliminary}
\put(-125,107){\bf $D^+\to \pi^-e^+e^+$}
\put(-130,97){\bf BESIII preliminary}
\caption{The scatter plots of $M_{\rm BC}$ versus $\Delta E$ of the accepted events in data,
where the blue rectangle denotes the signal box.
The contour plot is determined by MC simulations. The scale of the MC is arbitrary.} \label{fig::box-data}
\end{center}
\end{figure*}
Table \ref{tab::numbers} summarizes the numbers of events inside
($N^{\rm data}_{\rm inside}$) and outside ($N^{\rm data}_{\rm outside}$)
the signal boxes, the scale factors ($f_{\rm scale}$), detection efficiencies ($\epsilon$),
systematic uncertainties ($\Delta_{\rm sys}$), calculated upper limits for
the observed events ($\mathcal{s}_{90}$) and branching fractions ($\mathcal{B}$).
The upper limits on the numbers of these decays is determined by utilizing a
frequentist method \cite{ref::TROLKE} with unbounded profile likelihood treatment
of systematic uncertainties, where the number of the observed events is assumed
to follow a Poisson distribution, the number of background events and the efficiency
are assumed to follow Gaussian distributions, and the systematic uncertainty is
considered as the standard deviation of the efficiency.
Our results for $D^+\to\pi^+e^+e^-$ and $D^+\to K^-e^+e^+$ are significantly
improved than the previous restrictions and the other two limits are comparable
with the world best results.
\begin{table}[htbp]
\begin{center}
\caption{Summary of the numbers.}
\label{tab::numbers}
\begin{tabular}{ccccccccccc}
\hline
& $N^{\rm data}_{\rm inside}$ & $N^{\rm data}_{\rm outside}$ & $f_{\rm scale}$
& $\epsilon$ [\%] & $\Delta_{\rm sys}$ [\%] & $\mathcal{s}_{90}$ & $\mathcal{B}$ [$\times10^{-6}$] \\
\hline
$D^+\to K^+e^+e^-$ & 5 & 69 & $0.08\pm0.01$ & 22.53 & 5.4 & 19.4 & $<1.2$ \\
$D^+\to K^-e^+e^+$ & 3 & 55 & $0.08\pm0.01$ & 24.08 & 6.1 & 10.2 & $<0.6$ \\
$D^+\to \pi^+e^+e^-$ & 3 & 65 & $0.09\pm0.02$ & 25.72 & 5.9 & 4.2 & $<0.3$ \\
$D^+\to \pi^-e^+e^+$ & 5 & 68 & $0.06\pm0.02$ & 28.08 & 6.8 & 20.5 & $<1.2$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{$D^0\to\gamma\gamma$\cite{ref::BESIII-2gamma}}
To suppress the backgrounds from QED continuum processes, potential
$\psi(3770)\to {\rm non}-D\bar D$ decays, as well as $D^+D^−$ decays,
we perform a double tag technique in the analysis.
In this work, singly tagged events are selected as the first step, then
$\gamma\gamma$ final states will be investigated in the system recoiling side.
Single tag candidates are selected by reconstructing a $\bar D^0$ in one of the following
five hadronic final states: $\bar D^0\to K^+\pi^−$, $K^+\pi^−\pi^0$, $K^+\pi^−\pi^+\pi^−$,
$K^+\pi^−\pi^+\pi^−\pi^0$, and $K^+\pi^−\pi^0\pi^0$, constituting approximately 37\% of
all $D^0$ decays \cite{ref::pdg2012}.
The absolute branching fraction for the signal mode is determined as,
\begin{equation}
\mathcal{B}=\frac{N_{\rm tag, \gamma\gamma}}{\Sigma_{i}N^{i}_{\rm tag}\cdot(\epsilon^{i}_{\rm tag, \gamma\gamma}/\epsilon^{i}_{\rm tag})},
\end{equation}
where $i$ runs over each of the five tag modes, $N_{\rm tag}$ and $\epsilon_{\rm tag}$ are
the single tag yield and reconstruction efficiency, and $N_{\rm tag, \gamma\gamma}$ and
$\epsilon^{i}_{\rm tag, \gamma\gamma}$ are the yield and efficiency for the double tag
combination of a hadronic tag and a $D^0\to\gamma\gamma$ decay.
We extract the single tag yield for each tag mode and the combined yields
of all five modes from fits to $M^{\rm tag}_{\rm BC}$ distributions.
The signal shape is derived from the MC simulation which includes the effects of
beam-energy smearing, initial-state radiation, the $\psi(3770)$ line shape, and detector
resolution. We then convolute the line shape with a Gaussian to compensate for a difference
in resolution between data and our MC simulation. Mean and width of the convoluted Gaussian,
along with the overall normalization, are left free in our nominal fitting procedure.
The background is described by an ARGUS function \cite{ref::argus}, which models combinatorial
contributions. In the fit, all parameters of the background function are left free, except
its endpoint which is fixed at 1.8865 GeV$/c^2$.
Figure \ref{fig::tags} shows the fits to our tag-candidate samples.
\begin{figure}[htb]
\centering
\includegraphics[height=9cm,width=12cm]{tags.eps}
\caption{Fits (solid line) to the $M^{\rm tag}_{\rm BC}$ distributions
in data (points) for the five $D^0$ tag modes: (a) $K^+\pi^−$, (b) $K^+\pi^−\pi^0$,
(c) $K^+\pi^−\pi^+\pi^−$, (d) $K^+\pi^−\pi^+\pi^−\pi^0$, and (e) $K^+\pi^−\pi^0\pi^0$.
The gray shaded histograms are arbitrarily scaled generic MC backgrounds.}
\label{fig::tags}
\end{figure}
Although we can suppress most of the background with the double tag method,
there remain residual contributions from continuum processes, primarily
doubly-radiative Bhabha events for $Kπ$ tags and $e^+e^−\to q\bar q$ for
other modes. In order to correctly estimate their sizes, we take a data-driven
approach by performing an unbinned maximum likelihood fit to the two-dimensional
distribution of $\Delta E^{\gamma\gamma}$ versus $\Delta E^{\rm tag}$, as shown
in figure \ref{fig::sig-tag}.
We use $\Delta E^{\gamma\gamma}$ distributions rather than $M^{\gamma\gamma}_{\rm BC}$
distributions as the background from non-$D\bar D$ decays is more easily addressed
in the fit. Also, the background from $D^0\to\pi^0\pi^0$ peaks in
$M^{\gamma\gamma}_{\rm BC}$ at the same place as the signal does, whereas it is
shifted in $\Delta E^{\gamma\gamma}$.
The fitting ranges are $|\Delta E^{\gamma\gamma}| < 0.25$ GeV and
$|\Delta E^{\rm tag}| < 0.1$ GeV. These wide ranges are chosen to have adequate
statistics of the continuum backgrounds in our fit.
For the signal and the $D\to\pi^0\pi^0$ background, we extract probability
density functions (PDFs) from MC simulations. For the background from continuum
processes, we include a flat component in two dimensions, allowing the
normalization to float. The contribution from $D^+D^−$ decays is completely
negligible. We model the background from other $D^0\bar D^0$ decays with a pair
of functions. In the $\Delta E^{\rm tag}$ dimension we use a Crystal Ball
function (CB) \cite{ref::Crystal-Ball} plus a Gaussian, and in the
$\Delta E^{\gamma\gamma}$ dimension, we use a second-order exponential polynomial
function.
In our nominal fitting procedure, we fix the following parameters based on MC:
the power-law tail parameters of the CB, the coefficients of the polynomial, and
the mean and the width of the Gaussian function.
The normalization for the background from all other $D^0\bar D^0$ decays is left
free in the fit, as are the mean and width of the CB and the ratio of the areas
of the CB and Gaussian functions.
Figure \ref{fig::sig-tag} shows projections of the fit to the DT data sample
onto $\Delta E^{\gamma\gamma}$ (top) and $\Delta E^{\rm tag}$ (bottom).
\begin{figure}[htb]
\centering
\includegraphics[height=9cm,width=7cm]{sig_tag.eps}
\caption{Fit to the DT sample in data (points), projected onto $\Delta E^{\gamma\gamma}$ (a)
and $\Delta E^{\rm tag}$ (b). The dashed lines show the overall fits, while the dotted
histograms represent the estimated background contribution from $D^0\to\pi^0\pi^0$.
The solid line superimposed on the ∆Eγγ projection indicates the expected signal for
$\mathcal{B}(D^0\to\gamma\gamma)=10\times10^{−6}$. Also overlaid are the overall MC-estimated
backgrounds (gray shaded histograms) and the background component from non-$D\bar D$ processes
(diagonally hatched histograms).}
\label{fig::sig-tag}
\end{figure}
The fit yields $N_{\rm tag, \gamma\gamma} = (−1.0^{+3.7}_{-2.3})$, demonstrating that
there is no signal for $D^0\to\gamma\gamma$ in our data. This corresponds to
$\mathcal{B}(D^0\to\gamma\gamma) = 3.8× 10^{−6}$ including the systematic uncertainties.
If the systematic uncertainty were ignored in setting this limit it would be reduced
by $0.1\times10^{−6}$.
\section{Summary}
In summary, by analyzing 2.92 fb$^{-1}$ data collected at $\sqrt{s}=3.773$ GeV
with the BESIII detector at the BEPCII collider, we search for the
FCNC decays $D^0\to\gamma\gamma$, $D^+\to h^+e^+e^-$ and the LNV decays $D^+\to h^-e^+e^+$.
No signal excess is observed. As a result, we set the upper limits on the branching fractions
for these decays at a 90\% CL.
The results for $D^+\to\pi^+e^+e^-$ and $D^+\to K^-e^+e^+$ are significantly
improved than the previous restrictions, while those for $D^+\to\pi^-e^+e^+$ and $D^+\to K^+e^+e^-$ are comparable
with the world best results.
The result for $D^0\to\gamma\gamma$ is consistent with the upper limit previously set
by the BABAR Collaboration \cite{ref::babar-2gamma}.
Our result is the first experimental study of this decay using data at open-charm threshold,
where the backgrounds from non-$D\bar D$ decays can be effectively suppressed.
The resulting upper limits are still above the Standard Model predictions, no hints for
New Physics have been found yet.
\bigskip \bigskip \begin{center} \begin{large
I would like to thank the committee of CHARM 2015 for the invitation and their excellent
organizing.
|
\section{How to identify a Martingale among the many Local Martingales}
\section{True or Strict Local Martingale}
\subsection{Introduction}
The question whether a non-negative local martingale is a strict local or a true martingale is of fundamental probabilistic nature. It relates to existence of solutions to stochastic differential equations and martingale problems, as well as to
absolute continuity
of laws.
The question has been tackled in different settings most prominently by integrability conditions of Novikov- and Kazamaki-type, c.f. \cite{J79,KS(2002b),kazamaki77,LM,Novikov73,protter,RufNK}. A series of papers further explored the one-dimensional diffusion setting that thanks to the work \cite{MU(2012)} is particularly well understood.
Their main result relates martingality, i.e. the martingale property, of a generalized stochastic exponential driven by a solution \((X, W)\) to an SDE of the type
\begin{align}\label{SDE1}
\operatorname{d}\hspace{-0.003cm} X_t = b(X_t)\operatorname{d}\hspace{-0.003cm} t + a(X_t)\operatorname{d}\hspace{-0.003cm} W_t
\end{align} to exit times of a solution to a modified SDE
\begin{align}\label{SDE2}
\operatorname{d}\hspace{-0.003cm} Y_t = \widetilde{b}(Y_t) \operatorname{d}\hspace{-0.003cm} t + a(Y_t)\operatorname{d}\hspace{-0.003cm} B_t.
\end{align}
In particular, the result distinguishes at which end of the state space \((l, r)\) the modified solution exits.
A remarkable observation is that there are examples where the generalized stochastic exponential is a martingale while both the solution of the driving and the modified SDE exit their state space. In the case where it is the real line, this means that both SDEs may be explosive.
Already the topology of multidimensional state spaces does not allow a direct generalization of this subtle observation
to higher dimensions.
In finite-dimensional diffusion settings \cite{HR,Sin} connect the martingality of a non-explosive stochastic exponential driven by a solution to an SDE to existence properties of a related modified SDE.
\cite{RufSDE} starts in a possibly explosive diffusion setting and gives a condition for martingality, which is in the spirit of absolute continuity conditions for laws of semimartingales as given by \cite{JS}.
The proof of his condition is based on the Dambis-Dubins-Schwarz theorem and hence restricted to a continuous path setting.
A condition for martingality in a jump-diffusion setting can be deduced from the main result of \cite{CFY}.
\cite{KMK,Mayerhofer2011568} provide conditions for affine processes.
Based on an extension of probability measures \cite{kardaras2015} provide techniques to construct strict local martingales.
\cite{BR16} study martingality from a perspective of weak convergence.
In this article we raise the question how martingality of non-negative local martingales driven by Hilbert-space valued semimartingales on stochastic intervals relates to path properties of their driving processes.
Building our framework on semimartingale problems (SMPs) in the sense of \cite{J79} allows us to connect martingality with several interesting properties of semimartingales.
To briefly describe the concept let us be given a filtered space, a process and a triplet, which serves as a candidate for semimartingale characteristics. A probability measure under which the process is a semimartingale with the candidate triplet as characteristics is called a solution to the SMP.
We particularly allow for infinite-dimensions and stochastic lifetimes.
Let \(Z\) be a non-negative local martingale (usually) defined on a filtered probability space equipped with a solution to a given SMP.
In the diffusion setting this boils down to the assumption that \(Z\) is driven by ~\eqref{SDE1}.
We now summarize our main contributions.
Firstly, we prove under mild topological assumptions that if \(Z\) is a martingale then a modified SMP, comparable to the modified SDE \eqref{SDE2}, has a solution, c.f. Theorem \ref{main theorem new 2}. This is not only an existence result for solutions to SMPs, but also provides a condition for strict local martingality via contradiction.
Secondly, we replace the topological assumption by a uniqueness condition on the modified SMP and assume existence of a solution.
Our uniqueness condition is related to the definition of local uniqueness used by
\cite{JS} in their study of absolute continuity of laws. In Markovian settings it is implied by well-posedness of the SMP. If the SMP corresponds to a not necessarily Markovian SDE the uniqueness is implied by pathwise-uniqueness.
In this setting we formulate abstract convergence conditions in terms of a sequence of stopping times, c.f. Proposition \ref{main prop uni 1} and \ref{main prop uni 2}.
We use these results to formulate a boundedness condition of local type that also can be verified in cases where the conditions of Novikov- and Kazamaki-type are too strict, c.f. Section \ref{MSE}.
In Section \ref{Martingality of Generalized Stochastic Exponentials} we investigate
stochastic exponentials as explicit examples.
To illustrate these claims in a non-Markovian and infinite-dimensional setting let
\[W^* := \sup_{s < \cdot} W_s\ \textup{ and }\ N := W^* \cdot W,
\]
where \(W\) is an (infinite-dimensional) Brownian motion.
In the one-dimensinal case, we have \(\mathrm{I\kern-.2em E}[\exp\{(W^*)^2/2 \cdot I_t\}] \geq \mathrm{I\kern-.2em E}[\exp\{W^2/2 \cdot I_t\}] = \infty\), for \(t\) large enough, where \(I_t = t\) denotes the identity process. This implies that Novikov's condition is violated.
We will see in Section \ref{A Case Study - Martingality in Terms of SDEs driven by Hilbert-Space-Valued Brownian Motion} that \(Z\) is in fact a true martingale.
Thirdly, we show that under appropriate mild assumptions existence of a solution to the modified SMP implies martingality even if the original SMP has no global solution, c.f. Corollary \ref{explosion cond coro}.
This can be interpreted as a type of explosion condition as given by \cite{MU(2012)}, but now in
an infinite-dimensional and discontinuous setting.
Fourthly, we impose additional topological assumptions on the underlying filtered space.
This setting traces back to ideas employed by \cite{follmer72,perkowski2015} in the context of the F\"ollmer measure.
Thanks to an extension result for probability measures these conditions guarantee uniqueness and existence properties of the modified SMP, which lead to an explosion-type condition for martingality, c.f. Section \ref{Martingality on Standard Systems}.
Martingale criteria for non-negative local martingales driven by semimartingales are also interesting from an applications point of view.
The class of semimartingales lies at the heart of financial modeling as it plays a fundamental role for stochastic integration, c.f. \cite{bichteler1981stochastic,DellacherieMeyer78,Pro} and \cite{DS}, Chapter 7.3.
In a semimartingale setting the seminal work \cite{DelbaenSchachermayer94,Delbaen96thefundamental} relates absense of arbitrage to existence of an equivalent martingale measure (EMM). The existence of an EMM or more generally of changes of probability measures is deeply connected to the question when a candidate density process is a true martingale.
Typically candidate processes are already non-negative local martingales.
On a more practical side, changes of measures frequently lead to a substantial reduction of
computational complexity. In mathematical finance for instance they arise as changes of numeraire and simplify the computation of option prices as
highlighted in the influential work
\cite{GEK95}.
In the context of utility maximization, \cite{RePEc:wsi:ijtafx:v:13:y:2010:i:03:p:459-477} connects the martingality of a local martingale with optimal terminal wealth.
In all these cases, criteria for martingality of non-negative local martingales are substantial.
Recently also strict local martingales have attracted more attention and are used to model financial bubbles, c.f. \cite{Cox2005,Jarrow2007,MAFI:MAFI394,kardaras2015,Pal20101424,Pro(2013)}.
Examples of strict local martingales are given by \cite{Chybiryakov2007,Delbaen1995,ELY,KellerRessel2015,MU(2012),protter2015strict}.
Martingality of local martingales is also interesting from a purely probabilistic point of view, as
it determines absolute continuity or singularity of laws. This observation was exploited in the seminal work \cite{doi:10.1137/1105027}, c.f. also \cite{JM76,KLS-LACOM1,KLS-LACOM2} for a semimartingale perspective.
Such absolute continuity results are valuable tools to study the statistics of stochastic processes, c.f. \cite{10.2307/4616544}, and existence properties of SDEs and martingale problems, c.f. \cite{deprato,KaraShre,RY}.
Let us shortly summarize the structure of the article.
In Section \ref{First important Observations} we present simple characterizations of martingality of a non-negative local martingale.
These observations prove useful for establishing our main results.
In Section ~\ref{Hilbert-Vlaued Semimartingales and Semimartingale Problems on Stochastic Intervals} we introduce our mathematical setting.
The main results are given in Section \ref{section: Martingality of local martingales}.
Finally, in Section \ref{A Case Study - Martingality in Terms of SDEs driven by Hilbert-Space-Valued Brownian Motion}, we deduce explicit martingale conditions for stochastic exponentials driven by infinite-dimensional Brownian motion.
\textbf{A Short Remark Concerning Notations:}
All non-explained notations can be found in the monograph of \cite{JS}.
Moreover, we usually skip the terminology \emph{up to indistinguishability}. \emph{Equality of processes} as well as \emph{uniqueness of processes} should be read up to indistinguishability.
\subsection{A Simple Description of Martingality}\label{First important Observations}
We start with a characterization of the martingality of non-negative local matingales, which is as elementary as useful.
Fatou's lemma implies that a non-negative local martingale is a supermartingale, which itself is a martingale if and only if it has constant expectation.
Therefore, the martingality of non-negative martingales is a property that only depends on the expectation.
The following two consequences take advantage of this important observation.
We fix a filtered probability space \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\), where \(\ensuremath{\mathcal{F}} = (\mathcal{F}_t)_{t \geq 0}\) is a right-continuous filtration,
and a one-dimensional non-negative local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale \(Z\) with \({\mathbf{P}}\)-a.s. \(Z_0=1\).
The assumption that \(Z_0\) is \({\mathbf{P}}\)-a.s. constant can be relaxed to the case where \(Z_0\) is a positive \(\mathcal{F}_0\)-measurable and \({\mathbf{P}}\)-integrable random variable by replacing \(Z\) with \(Z/Z_0\).
The following lemma is in the spirit of Theorem 1.3.5 in \cite{SV}, Lemma III.3.3 in \cite{JS}, and Corollary 2.1 in \cite{BR16}.
\begin{lemma}\label{ruf mimic}
The following is equivalent:
\begin{enumerate}
\item[\textup{(i)}]
\(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\item[\textup{(ii)}] There exists an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-localization sequence \((\rho_n)_{n \in \mathbb{N}}\) of \(Z\) such that
\begin{align}\label{difficult to check}
\lim_{n \to \infty} {\mathbf{Q}}_n (\rho_n> t) = 1\ \textit{ for all } t \geq 0,
\end{align}
where \({\mathbf{Q}}_n := Z_{\rho_n} \cdot {\mathbf{P}}\), i.e. \({\mathbf{Q}}_n(\operatorname{d}\hspace{-0.003cm} \omega) = Z_{\rho_n}(\omega){\mathbf{P}}(\operatorname{d}\hspace{-0.003cm} \omega)\), for \(n \in \mathbb{N}\).
\item[\textup{(iii)}]
For all \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-localizing sequences \((\rho_n)_{n \in \mathbb{N}}\) of \(Z\) the convergence \eqref{difficult to check} holds.
\end{enumerate}
\end{lemma}
\begin{proof}
The implication (iii) \(\Longrightarrow\) (ii) is trivial. We show that (ii) \(\Longrightarrow\) (i) \(\Longrightarrow\) (iii).
\(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale if and only if for all \(t \geq 0\) we have \(\mathrm{I\kern-.2em E}[Z_t] = \mathrm{I\kern-.2em E}[Z_0] = 1\).
Let \((\rho_n)_{n \in \mathbb{N}}\) be an arbitrary \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-localizing sequence. Due to monotone convergence, Doob's stopping theorem and \(\{\rho_n > t\} \in \mathcal{F}_{t \wedge \rho_n}\), we have
\begin{align*}
\mathrm{I\kern-.2em E}[Z_t] &= \mathrm{I\kern-.2em E}\big[Z_t \lim_{n \to \infty} \ensuremath{1}_{\{\rho_n > t\}}\big] = \lim_{n \to \infty} \mathrm{I\kern-.2em E}\big[Z_{t \wedge \rho_n} \ensuremath{1}_{\{\rho_n> t\}}\big] = \lim_{n \to \infty} \mathrm{I\kern-.2em E}\big[\mathrm{I\kern-.2em E}\big[Z_{\rho_n}|\mathcal{F}_{t \wedge \rho_n}\big]\ensuremath{1}_{\{\rho_n> t\}}\big]
\\&= \lim_{n \to \infty} \mathrm{I\kern-.2em E}\big[\mathrm{I\kern-.2em E}\big[Z_{\rho_n} \ensuremath{1}_{\{\rho_n > t\}}|\mathcal{F}_{t \wedge \rho_n}\big]\big]
= \lim_{n \to \infty} {\mathbf{Q}}_n(\rho_n> t).
\end{align*}
This finishes the proof.
\end{proof}
If we relinquish on an equivalence statement we can relax the assumption in Lemma \ref{ruf mimic} that \((\rho_n)_{n \in \mathbb{N}}\) is a localizing sequence.
\begin{lemma}\label{neues lemma}
Assume that \((\rho_n)_{n \in \mathbb{N}}\) is a increasing sequence of \(\ensuremath{\mathcal{F}}\)-stopping times such that for all \(n \in \mathbb{N}\) the process \(Z^{\rho_n}\) is an uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\begin{enumerate}
\item[\textup{(i)}]
If \eqref{difficult to check} holds for all \(t \geq 0\), then \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\item[\textup{(ii)}]
If
\begin{align*}
\lim_{n \to \infty} {\mathbf{Q}}_n(\rho_n = \infty) = 1,\textit{ where } {\mathbf{Q}}_n = Z_{\rho_n} \cdot {\mathbf{P}},
\end{align*}
then \(Z\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\end{enumerate}
\end{lemma}
\begin{proof}
We first prove (i) and obtain that
\begin{align*}
1 \geq \mathrm{I\kern-.2em E}[Z_t] \geq \mathrm{I\kern-.2em E}[Z_t \lim_{n \to \infty} \ensuremath{1}_{\{\rho_n> t\}}] = \lim_{n \to \infty} \mathrm{I\kern-.2em E}[Z_t \ensuremath{1}_{\{\rho_n > t\}}] = \lim_{n \to \infty} {\mathbf{Q}}_n(\rho_n > t) = 1,
\end{align*}
by the same argumentation as in the proof of Lemma \ref{ruf mimic}. Therefore, as \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-supermartingale, \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
We now turn to (ii).
\(Z\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale if \(\mathrm{I\kern-.2em E}[Z_\infty] \geq 1\), where \(Z_\infty := \lim_{t \to \infty} Z_t\) exists due to the supermartingale convergences theorem. We obtain that
\begin{align*}
\mathrm{I\kern-.2em E}[Z_\infty] &\geq \lim_{n \to \infty} \mathrm{I\kern-.2em E}[Z_\infty \ensuremath{1}_{\{\rho_n = \infty\}}] = \lim_{n \to \infty} \mathrm{I\kern-.2em E}[ \mathrm{I\kern-.2em E}[Z_\infty |\mathcal{F}_{\rho_n}] \ensuremath{1}_{\{\rho_n = \infty\}}]
= \lim_{n \to \infty} {\mathbf{Q}}_n(\rho_n = \infty) = 1,
\end{align*}
thanks to \(\{\rho_n = \infty\} \in \mathcal{F}_{\rho_n}\) and Doob's stopping theorem.
This finishes the proof.
\end{proof}
Our main results introduce more structure for these conditions and thereby increase their applicability.
In order to identify a promising structure, let us observe the following:
If there exists a probability measure \({\mathbf{Q}}\) such that
\begin{align*}
{\mathbf{Q}} = {\mathbf{Q}}_n = Z_{\rho_n}\cdot {\mathbf{P}} \textup{ on } \mathcal{F}_{\rho_n},\textup{ for all } n \in \mathbb{N},
\end{align*}
then \eqref{difficult to check} boils down to
\begin{align}\label{cond q}
\lim_{n \to \infty} {\mathbf{Q}}(\rho_n > t) = 1,\textup{ for all } t \geq 0.
\end{align}
Moreover, if \eqref{cond q} holds and \({\mathbf{P}}\)-a.s. \(\rho_n \uparrow \infty\) as in Lemma \ref{ruf mimic}, then Lemma III.3.3 in \cite{JS} yields that \({\mathbf{Q}} \ll_\textup{loc} {\mathbf{P}}\) with density process \(Z\).
To use this observation and identify a structure of \({\mathbf{Q}}\), let us introduce an additional player, an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-semimartingale \(X\).
It is well-known that \({\mathbf{Q}} \ll_\textup{loc}{\mathbf{P}}\) implies that \(X\) is also an \((\ensuremath{\mathcal{F}}, {\mathbf{Q}})\)-semimartingale and its characteristics are related to \(X\) and \(Z\) via Girsanov's theorem.
Therefore \({\mathbf{Q}}\) is a probability measure under which \(X\) is a semimartingale with known characteristics.
This observation leads us to the concept of semimartingale problems.
\section{Hilbert-space-valued Semimartingales and \\ Semimartingale Problems on Stochastic Intervals}
\label{Hilbert-Vlaued Semimartingales and Semimartingale Problems on Stochastic Intervals}
Throughout this section we fix a filtered probability space \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\), where \(\ensuremath{\mathcal{F}} = (\mathcal{F}_t)_{t \geq 0}\) is a right-continuous filtration.
For the first part of Section
\ref{Processes on Stochastic Intervals} the probability measure \({\mathbf{P}}\) does not play a ~role.
\subsection{Processes on Stochastic Intervals}
\label{Processes on Stochastic Intervals}
In this first section
we introduce the concept of stochastic processes on stochastic intervals, which build the mathematical background to study possibly explosive processes.
We partially follow \cite{HWY,J79, JS}.
Related topics were also studied by \cite{emery1982, Schwartz1981,sharpe}.
We fix a
Polish space \(E\), which serves as the \emph{state space}.
Recall that for two \(\ensuremath{\mathcal{F}}\)-stopping times \(\tau\) and \(\rho\) we set
\begin{align*}
[\hspace{-0.06cm}[ \tau, \rho]\hspace{-0.06cm}] := \big\{(\omega, t) \in \Omega \times \mathbb{R}^+ : \tau(\omega) \leq t \leq \rho(\omega)\big\},
\end{align*}
and \([\hspace{-0.06cm}[ \tau, \rho[\hspace{-0.06cm}[, ]\hspace{-0.06cm}] \tau, \rho]\hspace{-0.06cm}], ]\hspace{-0.06cm}] \tau, \rho[\hspace{-0.06cm}[\) analogeously. Moreover, we denote \([\hspace{-0.06cm}[ \tau, \tau]\hspace{-0.06cm}] =: [\hspace{-0.06cm}[ \tau]\hspace{-0.06cm}]\).
\begin{definition}
We call a random set \(A\) a \emph{set of interval type}, if there exists an increasing sequence of \(\ensuremath{\mathcal{F}}\)-stopping times \((\tau_n)_{n \in \mathbb{N}}\) such that
\begin{align*}
A = \bigcup_{n \in \mathbb{N}} [\hspace{-0.06cm}[ 0, \tau_n]\hspace{-0.06cm}].
\end{align*}
\end{definition}
Sets of interval type are \(\ensuremath{\mathcal{F}}\)-predictable.
With a little abuse of terminology, we call the sequence \((\tau_n)_{n \in \mathbb{N}}\) \emph{announcing sequence for \(A\)}.
The set of all random sets of interval type is denoted by \(\mathcal{I}(\ensuremath{\mathcal{F}})\).
Let \(\mathcal{C}\) be a class of \(E\)-valued processes.
For \(A \in \mathcal{I}(\ensuremath{\mathcal{F}})\) we set
\begin{align*}
\mathcal{C}^{A} &:= \big\{ X : A \to\ E : \exists \textup{ announcing sequence } (\tau_n)_{n \in \mathbb{N}} \textup{ for } A \textup{ s.th. }
X^{\tau_n} \in \mathcal{C}\ \forall n \in \mathbb{N}\big\}.
\end{align*}
Since \((\omega, t \wedge \tau_n (\omega)) \in A\) for all \((\omega, t) \in [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[\), we have \(X^{\tau_n} : [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[\ \to E\) for all \(X \in \mathcal{C}^A\).
\begin{proposition}\label{subinterval semimartingale}
Let \(\mathcal{C}\) be stable under stopping and localization, \(X \in \mathcal{C}^{A}\), and \(\rho\) an \(\ensuremath{\mathcal{F}}\)-stopping time such that
\([\hspace{-0.06cm}[ 0, \rho]\hspace{-0.06cm}]\subseteq A\). Then \(X^{\rho}\in \mathcal{C}\).
\end{proposition}
\begin{proof}
The claim follows identically to \cite{HWY}, Theorem 8.20.
\end{proof}
\vspace{-0.3cm}
Examples of classes which are stable under stopping and localization are the class of \(E\)-valued local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingales, denoted by \(\mathcal{M}_\textup{loc}(E, \ensuremath{\mathcal{F}}, {\mathbf{P}})\), and the class of \(E\)-valued \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-semimartingales, denoted by \(\mathcal{S}(E, \ensuremath{\mathcal{F}}, {\mathbf{P}})\), where \(E\) is a real separable Hilbert space.
More generally, if \(\mathcal{C}\) is stable under stopping, then its localized class \(\mathcal{C}_\textup{loc}\) is stable under stopping and localization, c.f. \cite{JS}, Lemma I.1.35.
We obtain two obvious consequences of Proposition \ref{subinterval semimartingale}:
\begin{corollary}\label{ordinary semimartingale}
Let \(\mathcal{C}\) be stable under stopping and localization.
\begin{enumerate}
\item[\textup{(i)}]
\(Y\in\mathcal{C}^{A}\) if and only if for all \(\ensuremath{\mathcal{F}}\)-stopping times \(\rho\) such that \([\hspace{-0.06cm}[ 0, \rho]\hspace{-0.06cm}] \subseteq A\), \(Y^{\rho}\in \mathcal{C}\).
\item[\textup{(ii)}]
\(\mathcal{C}^{[\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[} = \mathcal{C}\).
\end{enumerate}
\end{corollary}
To construct non-negative local martingales from semimartingale on sets of interval type the following extension due to \cite{J79} proves itself useful in the Sections \ref{MSE} and \ref{Martingality of Generalized Stochastic Exponentials}.
For \(X\in \mathcal{C}^A\), where \(\mathcal{C}\) is a class of \([- \infty, \infty]\)-valued processes, we define the extension \(\widetilde{X}\) by
\begin{align}\label{extension}
\widetilde{X}_t := \begin{cases}
X_t&\textup{ on } \{t < \tau\},\\
X_{\tau},&\textup{ on } \{t \geq \tau\} \cap G^c,\\
\liminf_{s \uparrow \tau, s \in A \cap \mathbb{Q} \times \Omega} X_s,&\textup{ on } \{t \geq \tau\} \cap G,
\end{cases}
\end{align}
where
\(
\tau := \lim_{n \to \infty} \tau_n\) and \(G := \bigcap_{n \in \mathbb{N}} \{\tau_n < \tau < \infty\}
\)
for an arbitrary announcing sequence \((\tau_n)_{n \in \mathbb{N}}\) for ~\(A\).
\begin{remark}
If \([\hspace{-0.06cm}[ 0, \rho]\hspace{-0.06cm}] \subseteq A\) for an \(\ensuremath{\mathcal{F}}\)-stopping time \(\rho\), then \(\widetilde{X}^\rho = X^{\rho}\).
\end{remark}
Clearly, in general \(\widetilde{X} \not \in \mathcal{C}\), however, if \(X \in \mathcal{M}^A_\textup{loc}(\mathbb{R}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\), non-negativity implies
\(\widetilde{X} \in \mathcal{M}_\textup{loc}(\mathbb{R}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\), c.f. Proposition \ref{lemma tilde Z local martingale} in Appendix \ref{Extensions of non-negative local martingales on sets of interval type}.
\subsection{Characteristics of Hilbert-space-valued Semimartingales on Sets of Interval Type}
In what follows it is of little additional cost to proceed directly with Hilbert-space-valued processes. We are following closely the monographs \cite{MP80,metivier}.
For the definition Hilbert-space-valued semimartingales, which can be done parallel to the finite-dimensional case, we refer to Appendix \ref{A HVS}.
Let \(\mathbb{H}\) be a real separable Hilbert space and
denote by \(\mathcal{N}(\mathbb{H}, \mathbb{H})\) the space of nuclear operators from \(\mathbb{H}\) into \(\mathbb{H}\).\footnote{\(\mathcal{N}(\mathbb{H}, \mathbb{H})\) is a separable Banach space, c.f. \cite{gohberg2013classes}, p.106}
Let \(A \in \mathcal{I}(\ensuremath{\mathcal{F}})\).
\newpage
\begin{definition}\label{def A-c}
Let \(X\in \mathcal{S}^A(\mathbb{H}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\) and assume that \((B(h), C, \nu)\) is a triplet consisting of the following:
\begin{enumerate}
\item[\textup{(i)}] an \(\ensuremath{\mathcal{F}}\)-predictable \(\mathbb{H}\)-valued process \(B(h)\) on \(A\),
\item[\textup{(ii)}] an \(\ensuremath{\mathcal{F}}\)-predictable \(\mathcal{N}(\mathbb{H}, \mathbb{H})\)-valued process \(C\) on \(A\),
\item[\textup{(iii)}] an \(\ensuremath{\mathcal{F}}\)-predictable random measure \(\nu\) on \(\mathbb{R}^+ \times \mathbb{H}\) such that
\begin{align}\label{nu dec}
\nu(\omega, \operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x) = \ensuremath{1}_{A \times \mathbb{H}} (\omega, t, x) \nu(\omega, \operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x).
\end{align}
\end{enumerate}
If there exists an announcing sequence \((\tau_n)_{n \in \mathbb{N}}\) for \(A\) such that \((B(h)^{\tau_n}, C^{\tau_n}, \nu^{\tau_n})\), where
\begin{align*}
\nu^{\tau_n} (\omega, \operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x) := \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_n ]\hspace{-0.06cm}] \times \mathbb{H}}(\omega, t, x) \nu(\omega, \operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x),
\end{align*}
is \({\mathbf{P}}\)-indistinguishable of the \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-characteristics of \(X^{\tau_n} \in \mathcal{S}(\mathbb{H}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\) for all \(n \in \mathbb{N}\
, then we call the triplet \((B(h), C, \nu)\) the \emph{\((\ensuremath{\mathcal{F}}, {\mathbf{P}}, A)\)-characteristics} of \(X\).
\end{definition}
The most important consistency observations concerning the \(A\)-characteristics, such as \emph{existence} and \emph{uniqueness}, and some additional notations, such as the \emph{continuous local martingale part on \(A\)}, are given in Appendix \ref{facts HSVSM SI}.
\subsection{Semimartingale Problems on Stochastic Invervals}\label{Semimartingale Problems on Predictable Stochastic Invervals}
We denote the Borel \(\sigma\)-field of \(\mathbb{H}\) by \(\mathcal{H}\).
In order to deal with possibly explosive processes
we introduce the notion of a \emph{grave}. For a Polish space \(E\), i.e. in particular for \(\mathbb{H}\) and for \(\mathcal{N}(\mathbb{H}, \mathbb{H})\),\footnote{since separable Banach spaces are Polish spaces} let \(\Delta\) be a point outside of \(E\) and denote \(E_\Delta = E\cup \{\Delta\}\).
We define \(\mathbb{D}^{E_\Delta}\) as the space of all functions \(\alpha : \mathbb{R}^+ \to E_\Delta\), such that \(\alpha(s) = \Delta,\) for all \(s \in [\xi(\alpha), \infty)\), where \(\xi(\alpha) := \inf(t \geq 0 : \alpha(t) = \Delta)\), which are c\`adl\`ag on \((0, \xi(\alpha))\).
We assume that we are given the following:
\begin{enumerate}
\item[\textup{(i)}]
an \(\ensuremath{\mathcal{F}}\)-adapted
process \(X\) with paths in \(\mathbb{D}^{\mathbb{H}_\Delta}\),
which we call \textit{candidate process},
\item[\textup{(ii)}]
a triplet \((B(h), C, \nu)\), called a \textit{candidate triplet},
consisting of
\\[-1.3ex]
\begin{enumerate}
\item[--] an \(\mathfrak{F}\)-predictable process \(B(h)\) with paths in \(\mathbb{D}^{\mathbb{H}_\Delta}\),
\\[-1.5ex]
\item[--] an \(\mathfrak{F}\)-predictable process \(C\) with paths in \(\mathbb{D}^{\hspace{0.04cm}\mathcal{N}(\mathbb{H}, \mathbb{H})_\Delta}\),
\\[-1.5ex]
\item[--] an \(\mathfrak{F}\)-predictable random measure \(\nu\) on \(\mathbb{R}^+ \times \mathbb{H}\),
\\[-1.5ex]
\end{enumerate}
\item[\textup{(iii)}]
a probability measure \(\eta\) on \((\mathbb{H}, \mathcal{H})\), which is called \textit{initial law}.
\end{enumerate}
The process \(B(h)\) may depend on the truncation function \(h\).
\begin{definition}\label{Definition Semimartingale Problem}
We call a probability measure \({\mathbf{P}}\) on \((\Omega, \mathcal{F})\) a \emph{solution to the SMP} associated with \((\mathbb{H};\rho;\eta; X; B(h), C, \nu)\), where \(\rho\) is an \(\ensuremath{\mathcal{F}}\)-stopping time, \(\eta\) is an initial law, \(X\) a candidate process and \((B(h), C, \nu)\) a candidate triplet, if
\\[-1.3ex]
\begin{enumerate}
\item[\textup{(i)}] \(X^{\rho}\) is \({\mathbf{P}}\)-indistinguishable from an \(\mathbb{H}\)-valued \((\mathfrak{F}, {\mathbf{P}})\)-semimartingale,
\\[-1.3ex]
\item[\textup{(ii)}] the \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-characteristics of \(X^{\rho}\)
are \({\mathbf{P}}\)-indistinguishable from \((B^{\rho}(h), C^{\rho}, \nu^{\rho})\),
\\[-1.3ex]
\item[\textup{(iii)}] \({\mathbf{P}} \circ X^{-1}_0 = \eta\).\\[-1.3ex]
\end{enumerate}
We call \({\mathbf{P}}\) a solution to the SMP associated with \((\mathbb{H}; A; \eta; X; B(h), C, \nu)\), where \(A \in \mathcal{I}(\ensuremath{\mathcal{F}})\), if there exists an announcing sequence \((\tau_n)_{n \in \mathbb{N}}\) for \(A\), such that for all \(n \in \mathbb{N}\), \({\mathbf{P}}\) is a solution to the SMP associated with \((\mathbb{H};\tau_n; \eta; X; B(h), C, \nu)\).
\end{definition}
\begin{remark}
\({\mathbf{P}}\) is a solution to the SMP associated with \((\mathbb{H}; A; X; \eta; B(h), C, \nu)\) if and only ~if
\begin{enumerate}
\item[\textup{(i)}]
\(X|_A\) is \({\mathbf{P}}\)-indistinguishable from a process in \(\mathcal{S}^A(\mathbb{H},\ensuremath{\mathcal{F}}, {\mathbf{P}})\),
\item[\textup{(ii)}]
\((B(h), C, \nu)|_A\) is \({\mathbf{P}}\)-indistinguishable from the \((\ensuremath{\mathcal{F}}, {\mathbf{P}},A)\)-characteristics of \(X|_A\),
\item[\textup{(iii)}]
\({\mathbf{P}}\circ X^{-1}_0 = \eta\).
\end{enumerate}
\end{remark}
\begin{proposition}\label{SMP subinterval}
\({\mathbf{P}}\) is a solution to the SMP associated with \((\mathbb{H}; A; \eta; X; B(h), C, \nu)\) if and only if for all \(\ensuremath{\mathcal{F}}\)-stopping times \(\rho\) such that \([\hspace{-0.06cm}[ 0, \rho]\hspace{-0.06cm}]\subseteq A\), \({\mathbf{P}}\) is a solution to the SMP associated with \((\mathbb{H}; \rho; \eta; X; B(h), C, \nu)\).
\end{proposition}
\begin{proof}
The implication \(\Longleftarrow\) is trivial. The implication \(\Longrightarrow\) follows from
Corollary \ref{ordinary semimartingale} and Proposition \ref{exist A c} in Appendix \ref{facts HSVSM SI}.
\end{proof}
As an immediate consequence we obtain the following corollary.
\begin{corollary}\label{coro subinterval}
If \({\mathbf{P}}\) is a solution to the SMP associated with \((\mathbb{H}; A; \eta; X; B(h), C, \nu)\), and \(\mathcal{I}(\ensuremath{\mathcal{F}})\ni A^* \subseteq A\), then \({\mathbf{P}}\) is also a solution to the SMP associated with \((\mathbb{H}; A^*; \eta; X;\) \(B^*(h),\) \(C^*, \nu^*)\) for any candidate triplet \((B^*(h), C^*, \nu^*)\) which coincides on \(A^*\) with \((B(h), C, \nu)\).
\end{corollary}
Our next goal is to introduce a concept of uniqueness which is in the spirit of the \emph{local uniqueness} concept introduced by \cite{J79}, Section 12.3.e).
\begin{definition}
\begin{enumerate}
\item[\textup{(i)}]
We say that the SMP associated with \((\mathbb{H}; \rho; \eta; X; B(h), C, \nu)\) satisfies \emph{uniqueness}, if all solutions coincide on \(\mathcal{F}_{\rho-}\).
\item[\textup{(ii)}]
Let \(\Lambda\) be a set of \(\ensuremath{\mathcal{F}}\)-stopping times such that for all \(\rho\in \Lambda\) we have \([\hspace{-0.06cm}[ 0, \rho]\hspace{-0.06cm}] \subseteq A\).
We say that the SMP associated with \((\mathbb{H}; A; \eta; X; B(h), C, \nu)\) satisfies \emph{\(\Lambda\)-uniqueness} if for all \(\rho \in \Lambda\) the SMP associated with \((\mathbb{H}; \rho; \eta; X; B(h), C, \nu)\) satisfies uniqueness.
\end{enumerate}
\end{definition}
\begin{remark}
To hope for uniqueness we usually have to assume that \((\mathcal{F}, \ensuremath{\mathcal{F}})\) is \emph{generated by the candidate process \(X\)}, i.e. that
\begin{align*}
\mathcal{F} = \sigma(X_t, t \geq 0),\ \mathcal{F}^{\hspace{0.03cm}0}_t = \sigma(X_s, s\leq t)\textup{ and } \mathcal{F}_t = \bigcap_{s > t} \mathcal{F}^{\hspace{0.03cm}0}_s.
\end{align*}
The choice of the \(\sigma\)-field \(\mathcal{F}_{\rho-}\) in the definition of uniqueness has its origin in classical Markovian settings.
More precisely, standard techniques show that all solutions coincide on \(\mathcal{F}^{\hspace{0.03cm}0}_{\rho}\) instead of \(\mathcal{F}_\rho\). However, for a large class of stopping times \(\rho\) it holds that \(\mathcal{F}_{\rho-} \subseteq \mathcal{F}^{\hspace{0.03cm}0}_\rho\).
\end{remark}
In a finite dimensional Markovian setting a related concept of uniqueness is well-studied, c.f. \cite{J79}, Section 13.3. In a non-Markovian setting we derive explicit conditions implying uniqueness in Appendix \ref{SMPs and SDEs}.
The following remark is an immediate consequence of the definition of \emph{uniqueness}.
\begin{remark}\label{uni prop}
Let \(\Lambda\) be a set of \(\ensuremath{\mathcal{F}}\)-stopping times such that \([\hspace{-0.06cm}[ 0, \rho]\hspace{-0.06cm}] \subseteq A^*\in \mathcal{I}(\ensuremath{\mathcal{F}})\) for all \(\rho \in \Lambda\), let \(A\in \mathcal{I}(\ensuremath{\mathcal{F}})\) such that \(A^* \subseteq A\), and let \((B(h),C,\nu)\), \((B^*(h), C^*, \nu^*)\) be two candidate triplets which coincide on \(A^*\). Then the following is equivalent:
\begin{enumerate}
\item[\textup{(i)}] The SMP associated with \((\mathbb{H}; A; \eta; X; B(h), C, \nu)\) satisfies \(\Lambda\)-uniqueness.
\item[\textup{(ii)}] The SMP associated with \((\mathbb{H}; A^*; \eta; X; B^*(h), C^*, \nu^*)\) satisfies \(\Lambda\)-uniqueness.
\end{enumerate}
\end{remark}
\section{Main Results}\label{section: Martingality of local martingales}
Let \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\) be a given filtered probability space with right-continuous filtration \(\ensuremath{\mathcal{F}} = (\mathcal{F}_t)_{t \geq 0}\).
We are interested in the martingality of an \(\mathbb{R}^+\)-valued local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale \(Z\) with \({\mathbf{P}}\)-a.s. \(Z_0 = 1\).
Later we will assume that \(Z\) is driven by a Hilbert-space-valued semimartingale.
We describe the martingality of \(Z\) by convergence properties of sequences of stopping times satisfying the following convention which we impose from now on.
\begin{C}\label{conv}
Let \((\rho_n)_{n \in \mathbb{N}}\) be an increasing sequence of \(\ensuremath{\mathcal{F}}\)-stopping times such that \(Z^{\rho_n}\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale for all \(n \in \mathbb{N}\).
\end{C}
A sequence as in Convention \ref{conv} always exists, namely any localizing sequence of \(Z\).
It is important to note that the sequence in Convention \ref{conv} need not to increase \({\mathbf{P}}\)-a.s. to infinity.
In Appendix \ref{An Integrability Condition to Indentify} we present an integrability condition to identify such sequences, c.f. \cite{CFY} for a related approach based on a Novikov-type condition.
In Section \ref{Martingality in Terms of SMPs}, we relate martingality to existence and uniqueness properties of SMPs.
In Section \ref{Martingality on Standard Systems}, under topological assumptions on the underlying filtered space, we derive sufficient conditions for the martingality of \(Z\) in terms of an almost sure convergence of ~\((\rho_n)_{n \in \mathbb{N}}\).
\subsection{Martingality in Terms of SMPs}\label{Martingality in Terms of SMPs}
We consider five situations: Firstly, we derive a condition for the strict local martingality of \(Z\) in terms of existence properties of SMPs. Secondly, we impose the assumption of \(\Lambda\)-uniqueness for a modified SMP.
This additional structure enables us to investigate the martingality of \(Z\) by examining the limit behavior of the sequence of stopping times \((\rho_n)_{n \in \mathbb{N}}\) under a solution to an SMP.
Thirdly, we construct an explicit sequence \((\rho_n)_{n \in \mathbb{N}}\) driven solely by \(Z\). This results in martingality conditions only depending on the path of \(Z\).
Fourthly, we consider the situation where \(Z\) is an generalized stochastic exponential. In this case the previously established integrability condition boils down to a localized Novikov-type condition, which is suitable for applications also when classical Novikov-type conditions fail to hold.
Fifthly, we present another application of our abstract results which starts with the explicit choice of \((\rho_n)_{n \in \mathbb{N}}\) to be exit times of compacta. Thanks to this assumption, we derive an explosion-type condition in the spirit of \cite{MU(2012)}.
We impose the following additional assumption.
\begin{SA}\label{SA}
We assume that the probability measure \({\mathbf{P}}\) is a solution to the SMP
associated with \((\mathbb{H}; A; \eta; X; B(h), C, \nu)\).
We denote the unique
decomposition of \(X(h)\)
by \(X(h) = X_0 + M(h) + B(h)\), c.f. \eqref{B(h) def} in Appendix \ref{Semimartingale Characterstics}.
\end{SA}
All following standing assumptions are only assumed to hold in the particular subsection.
\subsubsection{A Condition for Strict Local Martingality}
We now consider the situation where the underlying filtered space allows an extension of every consistent family.
This topological property is called \emph{fullness}, c.f. Definition \ref{def full} in Appendix \ref{Classical Path-Spaces}, and is for instance possessed by all standard path spaces.
It enables us to extend the consistent family \(({\mathbf{Q}}_t, \mathcal{F}_t)_{t \geq 0}\) where \({\mathbf{Q}}_t := Z_t \cdot {\mathbf{P}}\) to a measure \({\mathbf{Q}}\) which is locally absolutely continuous w.r.t. \({\mathbf{P}}\) with density process
\(Z\).
An application of Girsanov's theorem then yields the existence of a solution to a modified SMP.
In view of this observation a contradiction argument yields a sufficient condition for strict local martingality.
In the following we formalize this idea.
\begin{theorem}\label{main theorem new 2}
Assume that \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}})\) is full.
If \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale, there exists a solution \({\mathbf{Q}}\) to the SMP
associated with \((\mathbb{H}; A; \eta; X; \dot{B}(h), C, \dot{\nu})\), where on \(A\)
\begin{equation}\label{candidtate triplet main theorem}
\begin{split}
\dot{B}(h) = B(h) + &1/Z_- \ensuremath{1}_{\{Z_- > 0\}}\cdot \langle\hspace{-0.085cm}\langle Z, M(h)\rangle\hspace{-0.085cm}\rangle^{{\mathbf{P}}}\textit{ and } \dot{\nu} = Y \cdot \nu,\\ \textit{with } Y &= 1/Z_-\ensuremath{1}_{\{Z_- > 0\}} M^{\mathbf{P}}_{\mu^X} (Z |\widetilde{\mathcal{P}}(\mathfrak{F}))
\end{split}
\end{equation}
for the notation c.f. Appendix \ref{Girsanov's Theorem for Hilbert-Space-valued Semimartingales}.
Additionally we have \({\mathbf{Q}} \ll_\textup{loc} {\mathbf{P}}\) with density process \(Z\).
\end{theorem}
\begin{proof}
Thanks to the assumption that \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}})\) is full, there exists a probability measure \({\mathbf{Q}}\) such that for all \(t \geq 0\) we have \({\mathbf{Q}} = {\mathbf{Q}}_t = Z_t \cdot {\mathbf{P}}\) on \(\mathcal{F}_t\).
Since for all \(t \geq 0\) it holds that \({\mathbf{Q}}_t \ll {\mathbf{P}}\) on \(\mathcal{F}_t\), we conclude that \({\mathbf{Q}} \ll_{\textup{loc}} {\mathbf{P}}\) with density process \(Z\).
Denote by \((\tau_n)_{n \in \mathbb{N}}\) an announcing sequence for \(A\). Then \(X^{\tau_n}\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-semimartingale with characteristics \((B(h)^{\tau_n}, C^{\tau_n}, \nu^{\tau_n})\), since \({\mathbf{P}}\) solves the SMP associated with \((\mathbb{H}; A; \eta; X; B(h), C, \nu)\).
Proposition \ref{pred finie var 0} in Appendix \ref{A HVS} yields that the unique decomposition \eqref{B(h) def} of \(X(h)^{\tau_n}\) is given by \(X(h)^{\tau_n} =X_0+ M(h)^{\tau_n} + B(h)^{\tau_n}\).
Moreover, the identity \([\hspace{-0.085cm}[ Z, M(h)^{\tau_n}]\hspace{-0.085cm}]^{\mathbf{P}} = [\hspace{-0.085cm}[ Z, M(h)]\hspace{-0.085cm}]^{\mathbf{P}}_{\cdot \wedge \tau_n}\) and again Proposition \ref{pred finie var 0} in Appendix \ref{A HVS} imply that
\begin{align*}
\langle\hspace{-0.085cm}\langle Z, M(h)^{\tau_n}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}} = \langle\hspace{-0.085cm}\langle Z, M(h)\rangle\hspace{-0.085cm}\rangle^{{\mathbf{P}}}_{\cdot \wedge \tau_n},
\end{align*}
where \(\langle\hspace{-0.085cm}\langle Z, M(h)^{\tau_n}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}\) is the unique c\`adl\`ag
\(\ensuremath{\mathcal{F}}\)-predictable process of finite variation such that the process \([\hspace{-0.085cm}[ Z, M(h)^{\tau_n}]\hspace{-0.085cm}]^{\mathbf{P}} - \langle\hspace{-0.085cm}\langle Z, M(h)^{\tau_n}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}\) is a local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale, c.f. Theorem \ref{GT} in Appendix ~\ref{Girsanov's Theorem for Hilbert-Space-valued Semimartingales}.
Furthermore, Proposition II.1.30 in \cite{JS} yields that
\begin{align*}
M^{\mathbf{P}}_{\mu^{X^{\tau_n}}}(Z|\widetilde{\mathcal{P}}(\ensuremath{\mathcal{F}})) = \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_n]\hspace{-0.06cm}]} M^{\mathbf{P}}_{\mu^X}(Z|\widetilde{\mathcal{P}}(\ensuremath{\mathcal{F}})).
\end{align*}
Therefore, the Girsanov-type theorem given by Theorem \ref{GT} in Appendix \ref{Girsanov's Theorem for Hilbert-Space-valued Semimartingales} yields that \(X^{\tau_n}\) is also an \((\ensuremath{\mathcal{F}}, {\mathbf{Q}})\)-semimartingale with \((\ensuremath{\mathcal{F}}, {\mathbf{Q}})\)-characteristics \((\dot{B}(h)^{\tau_n}, C^{\tau_n}, \dot{\nu}^{\tau_n})\).
Now the identity \({\mathbf{Q}} \circ X_0^{-1} = {\mathbf{Q}}_0 \circ X_0^{-1} = {\mathbf{P}} \circ X_0^{-1} = \eta\), as \({\mathbf{P}}\)-a.s. \(Z_0 = 1\), yields that \({\mathbf{Q}}\) is a solution to the SMP associated with \((\mathbb{H}; \tau_n; \eta; X; \dot{B}(h), C, \dot{\nu})\).
Since this holds for all \(n \in \mathbb{N}\), the claim is proven.
\end{proof}
As a corollary we obtain the following sufficient condition for strict local ~martingality.
\begin{corollary}\label{coro strict explosion}
Let \((\Omega,\mathcal{F}, \ensuremath{\mathcal{F}})\) be full. If the SMP
associated with \((\mathbb{H}; A; \eta; X; \dot{B}(h), C, \dot{\nu})\), where \(\dot{B}(h)\) and \(\dot{\nu}\) are given by \eqref{candidtate triplet main theorem} on \(A\), has no solution, then the process \(Z\) is a strict local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\end{corollary}
\begin{remark}
In the one-dimensional diffusion setting the \textit{Feller test for explosion}, c.f. \cite{Feller52} or \cite{KaraShre}, Section 5.5.C, provides a technique to apply
Corollary \ref{coro strict explosion}.
A comprehensive discussion is given by \cite{MU(2012)}.
\end{remark}
\subsubsection{The Martingale Property under Uniqueness Assumptions}\label{main section}
In this section we impose existence and uniqueness assumptions on the modified SMP
and derive sufficient conditions for the martingality of \(Z\).
We denote by \((\tau_n)_{n \in \mathbb{N}}\) an arbitrary announcing sequence for \(A\) and set \(\bar{A} := \bigcup_{n \in \mathbb{N}} [\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}] \in \mathcal{I}(\ensuremath{\mathcal{F}})\) and \(\tau := \lim_{n \to \infty} \tau_n\).
\begin{condition}\label{cond main uni}
We define the following conditions:
\begin{enumerate}
\item[\textup{(I)}]
The SMP associated with \((\mathbb{H}; \bar{A}; \eta; X;\dot{B}(h), C, \dot{\nu})\), where \(\dot{B}(h)\) and \(\dot{\nu}\) are given by \eqref{candidtate triplet main theorem} on \(\bar{A}\),
has a solution \({\mathbf{Q}}\) and satisfies \(\{\tau_n \wedge \rho_m, n,m \in \mathbb{N}\}\)-uniqueness.
\item[\textup{(II)}] For all \(n \in \mathbb{N}\) and \(t \geq 0\) we have \(\{\rho_n > t\}\in \mathcal{F}_{(\tau \wedge \rho_n)-}\).
\item[\textup{(III)}] For all \(n \in \mathbb{N}\) we have \(\{\rho_n = \infty\} \in \mathcal{F}_{(\tau \wedge \rho_n)-}\).
\end{enumerate}
\end{condition}
Note that (II) and (III) are structural assumptions which relate the sequences \((\rho_n)_{n \in \mathbb{N}}\) and \((\tau_n)_{n \in \mathbb{N}}\).
Now we are in the position to give the main results of this section.
\begin{proposition}\label{main prop uni 1}
Assume \textup{(I)} and \textup{(II)} in Conditions \ref{cond main uni}.
\begin{enumerate}
\item[\textup{(i)}]
If we have
\begin{align}\label{Main condition}
\lim_{n \to \infty}{\mathbf{Q}}(\rho_n > t) = 1,\textit{ for all } t \geq 0,
\end{align}
then \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\item[\textup{(ii)}]
If \({\mathbf{P}}\)-a.s. \(\rho_n \uparrow_{n \to \infty} \infty\) and \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale, then \eqref{Main condition} holds.
\end{enumerate}
\end{proposition}
\begin{proposition}\label{main prop uni 2}
Assume \textup{(I)} and \textup{(III)} in Conditions \ref{cond main uni}.
\begin{enumerate}
\item[\textup{(i)}]
If we have
\begin{align}\label{Main condition ui}
\lim_{n \to \infty} {\mathbf{Q}}(\rho_n = \infty) = 1,
\end{align}
then \(Z\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\item[\textup{(ii)}]
If \(\bigcup_{n \in \mathbb{N}} \{\rho_n = \infty\}\) is a \({\mathbf{P}}\)-a.s. event and \(Z\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale, then \eqref{Main condition ui} holds.
\end{enumerate}
\end{proposition}
The proofs of these two results are based on the following two lemmata.
\begin{lemma}\label{lemma stopped}
For all \(m, n \in \mathbb{N}\), the probability measure \({\mathbf{Q}}_{m, n} := Z_{\tau_m \wedge \rho_n} \cdot {\mathbf{P}}\) solves the SMP associated with \((\mathbb{H}; \tau_m \wedge \rho_n; \eta; X; \dot{B}(h), C, \dot{\nu})\), where
\(\dot{B}(h)\) and \(\dot{\nu}\) are given by \eqref{candidtate triplet main theorem} on \(\bar{A}\).
\end{lemma}
\begin{proof}
Since \({\mathbf{P}}\)-a.s. \(Z_0 = 1\), it holds that
\begin{align}\label{qn initial}
{\mathbf{Q}}_{m,n} \circ X^{-1}_0 = \eta.
\end{align}
Moreover, since \(X^{\tau_m \wedge \rho_n}\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-semimartingale
and \({\mathbf{Q}}_{m,n} \ll {\mathbf{P}}\) with density process \(Z^{\tau_m \wedge \rho_n}\), the Girsanov-type theorem as given in Theorem \ref{GT} in Appendix \ref{Girsanov's Theorem for Hilbert-Space-valued Semimartingales} implies that \(X^{\tau_m \wedge \rho_n}\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{Q}})\)-semimartingale with \((\ensuremath{\mathcal{F}}, {\mathbf{Q}})\)-characteristics \((\bar{B}(h), C^{\tau_m \wedge \rho_n}, \bar{\nu})\), where
\begin{align*}
\bar{B}(h) &= B(h)^{\tau_m \wedge \rho_n} + \ensuremath{1}_{\{Z^{\tau_m \wedge \rho_n}_- > 0\}}/Z^{\tau_m \wedge \rho_n}_-
\cdot \langle\hspace{-0.085cm}\langle Z^{\tau_m \wedge \rho_n}, M(h)^{\tau_m \wedge \rho_n}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}} \\
&= B(h)^{\tau_m \wedge \rho_n} + \ensuremath{1}_{\{Z_- > 0\}}/Z_- \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_m \wedge \rho_n]\hspace{-0.06cm}]} \cdot \langle\hspace{-0.085cm}\langle Z, M(h)\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}
\\&= \dot{B}(h)^{\tau_m \wedge \rho_n},\\
\bar{\nu}(\operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x) &= \frac{\ensuremath{1}_{\{Z^{\tau_m \wedge \rho_n}_{t-} > 0\}}}{Z^{\tau_m \wedge \rho_n}_{t-}} M^{\mathbf{P}}_{\mu^{X^{\tau_m \wedge \rho_n}}}(Z^{\tau_m \wedge \rho_n} | \widetilde{\mathcal{P}}(\ensuremath{\mathcal{F}}))(t, x) \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_m \wedge \rho_n]\hspace{-0.06cm}] \times \mathbb{H}}(\cdot, t, x) \nu(\operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x)
\\&= \frac{\ensuremath{1}_{\{Z_{t-} > 0\}}}{Z_{t-}} M^{\mathbf{P}}_{\mu^X}(Z|\widetilde{\mathcal{P}}(\ensuremath{\mathcal{F}}))(t,x) \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_m \wedge \rho_n]\hspace{-0.06cm}] \times \mathbb{H}}(\cdot, t, x) \nu(\operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x)
\\&= \dot{\nu}^{\tau_m \wedge \rho_n} (\operatorname{d}\hspace{-0.003cm} t, \operatorname{d}\hspace{-0.003cm} x).
\end{align*}
This and \eqref{qn initial} implies that
\({\mathbf{Q}}_{m,n}\) is a solution to the SMP
\((\mathbb{H};\tau_m\wedge \rho_n; \eta; X; \dot{B}(h), C, \dot{\nu})\).
\end{proof}
\begin{lemma}\label{lemma 2}
Assume part \textup{(I)} of Conditions \ref{cond main uni}, then
\begin{align*}
{\mathbf{Q}} = Z_{\tau \wedge \rho_n} \cdot {\mathbf{P}} \textit{ on }
\mathcal{F}_{(\tau \wedge \rho_n)-}.
\end{align*}
\end{lemma}
\begin{proof}
Due to the assumption of \(\{\tau_m\wedge \rho_n, n, m \in \mathbb{N}\}\)-uniqueness and Lemma \ref{lemma stopped} we have
\begin{align*}
{\mathbf{Q}} =Z_{\tau_m \wedge \rho_n} \cdot {\mathbf{P}} \textup{ on } \mathcal{F}_{(\tau_m \wedge \rho_n)-}
\end{align*}
Let \(F \in \mathcal{F}_{(\tau_m \wedge \rho_n)-}\), then
\begin{align*}
{\mathbf{Q}}(F) = \mathrm{I\kern-.2em E}[Z_{\tau_m\wedge \rho_n} \ensuremath{1}_F] = \mathrm{I\kern-.2em E}[\mathrm{I\kern-.2em E}[Z_{\tau \wedge \rho_n}| \mathcal{F}_{\tau_m \wedge \rho_n}] \ensuremath{1}_F] = \mathrm{I\kern-.2em E}[Z_{\tau \wedge \rho_n} \ensuremath{1}_F],
\end{align*}
due to Doob's stopping theorem.
This proves that \({\mathbf{Q}} = Z_{\tau \wedge \rho_n} \cdot {\mathbf{P}}\) on \(\bigcup_{m \in \mathbb{N}} \mathcal{F}_{(\tau_m \wedge \rho_n)-}\). Then a monotone class argument
yields
\begin{align*}
{\mathbf{Q}} = Z_{\tau \wedge \rho_n} \cdot {\mathbf{P}}\ \textup{ on } \bigvee_{m \in \mathbb{N}} \mathcal{F}_{(\tau_m \wedge \rho_n)-}.
\end{align*}
Finally, thanks to \cite{DellacherieMeyer78}, Theorem IV.56, \(\bigvee_{m \in \mathbb{N}} \mathcal{F}_{(\tau_m \wedge \rho_n)-} = \mathcal{F}_{(\tau \wedge \rho_n)-}\),
which finishes the proof.
\end{proof}
\hspace{-0.45cm}\textit{Proof of Proposition \ref{main prop uni 1}:}
(i).
Due to part (II) of Conditions \ref{cond main uni}, Lemma \ref{lemma 2} and Doob's stopping theorem we obtain
\begin{align}\label{q comp}
{\mathbf{Q}}(\rho_n > t) &= \mathrm{I\kern-.2em E}\big[Z_{\tau \wedge \rho_n} \ensuremath{1}_{\{\rho_n > t\}}\big] = \mathrm{I\kern-.2em E}\big[ \mathrm{I\kern-.2em E}\big[Z_{\rho_n}|\mathcal{F}_{\tau \wedge \rho_n}\big]\ensuremath{1}_{\{\rho_n > t\}}\big] = \mathrm{I\kern-.2em E}\big[Z_{\rho_n} \ensuremath{1}_{\{\rho_n > t\}}\big].
\end{align}
Therefore, Lemma \ref{neues lemma} (i) yields the claim.
(ii). Since \({\mathbf{P}}\)-a.s. \(\rho_n \uparrow_{n \to \infty} \infty\), Lemma \ref{ruf mimic} and \eqref{q comp} yield the claim.
\qed
\\\\
\hspace{-0.45cm}\textit{Proof of Proposition \ref{main prop uni 2}:}
(i).
The claim follows identically to the proof of Proposition \ref{main prop uni 1} ~(i).
(ii).
Part (III) of Conditions \ref{cond main uni}, Lemma \ref{lemma 2}, Doob's optional stopping theorem, dominated convergence and the assumption that \(\bigcup_{n \in \mathbb{N}} \{\rho_n = \infty\}\) is a \({\mathbf{P}}\)-a.s. event yield
\begin{align*}
\lim_{n \to \infty} {\mathbf{Q}}(\rho_n = \infty) = \lim_{n \to \infty} \mathrm{I\kern-.2em E}[Z_{\rho_n} \ensuremath{1}_{\{\rho_n = \infty\}}] =
\lim_{n \to \infty} \mathrm{I\kern-.2em E}[Z_\infty \ensuremath{1}_{\{\rho_n = \infty\}}] = \mathrm{I\kern-.2em E}[Z_\infty] = 1.
\end{align*}
This finishes the proof.
\qed
\begin{remark}
For explicit conditions implying existence of a solution to the modified SMP we refer to Theorem \ref{theorem extension2} below.
\end{remark}
\subsubsection{A first explicit definition of \((\rho_n)_{n \in \mathbb{N}}\)}\label{MSE}
In this section we benefit from a boundedness condition implying uniformly integrable martingality due to \cite{J79} in order to construct an explicit sequence \((\rho_n)_{n \in \mathbb{N}}\).
Let \((\sigma_n)_{n \in \mathbb{N}}\) be a sequence of \(\ensuremath{\mathcal{F}}\)-stopping times, denote \(A' = \bigcup_{n \in \mathbb{N}} [\hspace{-0.06cm}[ 0, \sigma_n]\hspace{-0.06cm}]\), and let \(Z \in \mathcal{M}^{A'}_\textup{loc}(\mathbb{R}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\) be non-negative.
We extend \(Z\) as in \eqref{extension} and denote the extension by \(\widetilde{Z}\). Due to Proposition \ref{lemma tilde Z local martingale} in Appendix \ref{Extensions of non-negative local martingales on sets of interval type}, \(\widetilde{Z}\) is a non-negative local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
We set
\begin{align}\label{C(Z)}
C(\widetilde{Z}) := \langle\hspace{-0.085cm}\langle \widetilde{Z}^c, \widetilde{Z}^c\rangle\hspace{-0.085cm}\rangle^{{\mathbf{P}}} + \sum_{s \leq \cdot} \bigg(\widetilde{Z}_{s-}- \sqrt{\widetilde{Z}_s \widetilde{Z}_{s-}}\hspace{0.05cm}\bigg)^2.
\end{align}
Thanks to Lemma 8.5 in \cite{J79} it holds that \(C(\widetilde{Z}) \in \mathcal{A}^+_\textup{loc}(\ensuremath{\mathcal{F}}, {\mathbf{P}})\), implying that \(C(\widetilde{Z})^p\), the \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-compensator of \(C(\widetilde{Z})\), is well-defined.
Moreover, we define the sequence \((\rho_n)_{n \in \mathbb{N}}\) of \(\ensuremath{\mathcal{F}}\)-stopping times by
\begin{align}\label{gamma}
\rho_n &:= \inf\big(t \geq 0 : \ensuremath{1}_{\{\widetilde{Z}_- > 0\}}/\widetilde{Z}_-^2 \cdot C(\widetilde{Z})^p_{t} \geq n\big).
\end{align}
\begin{condition}\label{gen bound cond}
We define the following conditions:
\begin{enumerate}
\item[\textup{(I)}] We have \(A' \subseteq A\).
\item[\textup{(II)}] The random time \(\sigma := \lim_{n \to \infty} \sigma_n\) is an \(\ensuremath{\mathcal{F}}\)-predictable time.
\item[\textup{(III)}]
For each \(n \in \mathbb{N}\) there exists a constant \(c_n > 0\) such that
\begin{align}\label{jump condition}
\frac{\Delta C(\widetilde{Z})^p_{\rho_n}}{\widetilde{Z}_{\rho_n-}^2} \ensuremath{1}_{\{\widetilde{Z}_{\rho_n-} > 0\}} \leq c_n.
\end{align}
\end{enumerate}
\end{condition}
The main result of this section is the following:
\begin{proposition}\label{gen coro extended stoch exp}
Assume that the
Part \textup{(I)} of Conditions \ref{cond main uni} and \ref{gen bound cond} hold.
\begin{enumerate}
\item[\textup{(i)}]
If for all \(t \geq 0\) we have
\begin{align}\label{Q cont assup}
{\mathbf{Q}}\left(\ \frac{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}}{\widetilde{Z}_-^2} \cdot C(\widetilde{Z})^p_t < \infty\right) = 1,
\end{align}
then \(\widetilde{Z}\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\item[\textup{(ii)}]
If we have
\begin{align*}
{\mathbf{Q}}\left(\ \frac{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}}{\widetilde{Z}_-^2} \cdot C(\widetilde{Z})^p_\infty< \infty\right) = 1,
\end{align*}
then \(\widetilde{Z}\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\end{enumerate}
\end{proposition}
Before the give a proof we shortly discuss an example.
\\\\
\textbf{Example.}
We now consider a finite-dimensional continuous and Markovian setting to illustrate an application of Proposition \ref{gen coro extended stoch exp}.
Let \(\beta= (\beta^i)_{i \leq d}: \mathbb{R}^+ \times \mathbb{R}^d \to \mathbb{R}^d\) and \(b = (b^i)_{i \leq d} : \mathbb{R}^+ \times \mathbb{R}^d \to \mathbb{R}^d\) be Borel functions which are locally bounded, i.e. for each \(n \in \mathbb{N}\) bounded on \([0, n] \times \{x \in \mathbb{R}^d : |x| \leq n\}\), and \(a=(a^{ij})_{i, j \leq d} : \mathbb{R}^+ \times \mathbb{R}^d \to \mathbb{S}^d_+\) be continuous, where \(\mathbb{S}^d_+\) denotes the space of positive definit symmetric \(d \times d\) matrices.
Next we specify the SMP of Standing Assumption \ref{SA}.
We assume that \(A = [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[\), that the underlying filtered space \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}})\) is the path space \((\mathbb{C}^d, \mathcal{C}^d, \mathfrak{C}^d)\) and that \(X\) is the coordinate process \(X_t(\omega) = \omega\hspace{0.01cm}_t\). Moreover, for \(\omega \in \mathbb{C}^d\) we set
\begin{align*}
B (\omega) = b(\omega)\cdot I,\quad C (\omega) = a(\omega)\cdot I, \quad \nu = 0,
\end{align*}
where \(b(\omega)_t := b(t, \omega\hspace{0.01cm}_t)\) and \(a(\omega)_t := a(t, \omega\hspace{0.01cm}_t)\).
Due to the local boundedness assumptions on \(a\) and \(b\), the SMP associated with \((\mathbb{R}^d; A; \varepsilon_{x_0};X; B, C, 0)\) is equivalent to the \emph{classical martingale problem} of \cite{SV} associated to the operator
\begin{align*}
L_t f (x) = \sum_{i \leq d} b^i(t, x) \partial_i f (x)+ \frac{1}{2} \sum_{i, j \leq d} a^{ij}(t, x) \partial^2_{ij} f (x),
\end{align*}
and the initial law \(\varepsilon_{x_0}\) for \(x_0 \in \mathbb{R}^d\),
c.f. Theorem 13.55 in \cite{J79} and Proposition 5.4.11 in \cite{KaraShre}.
Sufficient conditions for the existence of a solution can for instance be found in \cite{IW89,J79,SV}.
Now we are interested in the martingality of the local \((\mathfrak{C}^d, {\mathbf{P}})\)-martingale
\begin{align*}
Z = \mathcal{E}\left(\beta(X) \cdot X^c\right),
\end{align*}
where \(\beta(X)_t := \beta(t, X_{t})\). The local martingality follows immediately from the local boundedness assumption of \(\beta\).
\begin{corollary}\label{coro markov}
\(Z\) is a \((\mathfrak{C}^d, {\mathbf{P}})\)-martingale if and only if the SMP \((\mathbb{R}^d; A; \varepsilon_{x_0};X; \dot{B}, \dot{C}, 0)\), where
\begin{align}\label{triplet example}
\dot{B}(\omega) = b (\omega) \cdot I + (\beta a)(\omega) \cdot I \textit{ and } (\beta a)(\omega)_t := \beta(t,\omega\hspace{0.01cm}_t) a(t,\omega\hspace{0.01cm}_t),
\end{align}
has a solution.
\end{corollary}
\begin{proof}
Since \((\mathbb{C}^d, \mathcal{C}^d, \mathbf{C}^d)\) is full, c.f. Remark \ref{bichteler remark full} in Appendix \ref{Classical Path-Spaces}, the implication \(\Longrightarrow\) follows from Theorem \ref{main theorem new 2}.
We now show the converse implication.
Note that the Conditions ~\ref{gen bound cond} are trivially satisfied and that
\begin{align*}
\frac{\ensuremath{1}_{\{Z_- > 0\}}}{Z^2_-} \cdot C(Z)^p_t = (\beta c \beta^*)(X) \cdot I_t,\ \textup{ where } \ (\beta c \beta^*)(X) := \sum_{i, j \leq d} \beta^i(X) c^{ij}(X) \beta^j(X),
\end{align*}
which is clearly finite for each \(t \geq 0\) if \(X\) is a continuous process. Hence \eqref{Q cont assup} is satisfied
and the claim follows from Proposition \ref{gen coro extended stoch exp} if the SMP associated with \((\mathbb{R}^d; A; \varepsilon_{x_0}; \dot{B}, C, 0)\) satisfied \(\Upsilon := \{n \wedge \rho_m, n, m \in \mathbb{N}\}\)-uniqueness,
where \(\rho_m = \inf(t \geq 0 : (\beta c \beta^*)(X) \cdot I_t \geq m)\).
Due to Exercise 13.12 in \cite{J79} the SMP associated with \((\mathbb{R}^d; A; \varepsilon_{x_0};X; \dot{B}, C, 0)\) satisfies \(\mathfrak{P}\)-uniqueness, where \(\mathfrak{P}\) denotes the set of \(\mathfrak{C}^d\)-predictable times.
Thanks to Proposition I.2.13 in \cite{JS} it holds that \(\Upsilon \subseteq \mathfrak{P}\) and the proof is finished.
\end{proof}
In a one-dimensional setting this corollary is contained in Corollary 2.2 in \cite{MU(2012)}.
We shortly give a simple example where Corollary \ref{coro markov} implies martingality and Novikov's and Kazamaki's conditions fail.
Let \(d = 1\) and set \(x_0 = 0, \beta(x) = x, b = 0\) and \(a = 1\), i.e. let \(X\) be a one-dimensional standard Brownian motion and \(Z = \mathcal{E}(X\cdot X)\).
Then the SMP associated to \((\mathbb{R}; A; \varepsilon_0; X;\beta (X)\cdot I, I, 0)\) has a solution due to classical Lipschitz conditions.
Hence \(Z\) is a \((\mathfrak{C}^d, {\mathbf{P}})\)-martingale. However, for large enough \(t \geq 0\) it holds that \(\mathrm{I\kern-.2em E}[\exp(X/2 \cdot X_t)] = \infty\) which implies that Novikov's and Kazamaki's conditions fail.
The invertibility assumptions on \(a\) may be exchanged by a global uniqueness assumption. More precisely, it follows from Theorem 12.73 in \cite{J79} and Exercise 6.7.4 in \cite{SV} that well-posedness, i.e. the existence of a unique global solution, and locally bounded coefficients imply \(\mathfrak{P}\)-uniqueness.
We state this observation in the following corollary.
\begin{corollary}
Assume that \(\beta\) and \(b\) are as above and \(a : \mathbb{R}^+ \times \mathbb{R}^d \to \mathbb{S}^d\) is a locally bounded Borel function, where \(\mathbb{S}^d\) denotes the set of non-negative definite symmetric \(d \times d\)-matrices. If the SMP associated with \((\mathbb{R}^d; A; \varepsilon_{x_0}; X;\dot{B}, C, 0)\), where \(\dot{B}\) is given as in \eqref{triplet example}, is well-posed, then \(Z\) is a \((\mathfrak{C}^d, {\mathbf{P}})\)-martingale.
\end{corollary}
For an extension beyond Markovianity and finite dimensions we refer to Section \ref{A Case Study - Martingality in Terms of SDEs driven by Hilbert-Space-Valued Brownian Motion} below.
We now turn to the proof of Proposition \ref{gen coro extended stoch exp} which is based on the following lemma.
\begin{lemma}\label{con extended loc mart}
If Conditions \ref{gen bound cond} \textup{(II), (III)} hold, then for all \(n \in \mathbb{N}\) the process \(\widetilde{Z}^{\rho_n}\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
Moreover, we have for all \(t \geq 0\),
\begin{center}
\(
\{\rho_n > t\} \in \mathcal{F}_{(\sigma \wedge \rho_n)-}
\)
and \(\{\rho_n = \infty\} \in \mathcal{F}_{(\sigma \wedge \rho_n)-}\).
\end{center}
\end{lemma}
\begin{proof}
Note that
\begin{align*}
\frac{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}}{\widetilde{Z}_-^2} \cdot C(\widetilde{Z})^p_{\rho_n} \leq n + c_n,
\end{align*}
c.f. part (III) of Conditions \ref{gen bound cond}. Hence it follows from Lemma \ref{besser als novi lemma} in Appendix \ref{An Integrability Condition to Indentify} that \(\widetilde{Z}^{\rho_n}\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
Due to construction we have \(\widetilde{Z} = \widetilde{Z}^\sigma\).
This yields
\begin{align*}
\frac{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}}{\widetilde{Z}_-^2} \cdot C(\widetilde{Z})^p = \frac{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}}{\widetilde{Z}_-^2} \cdot C (\widetilde{Z})^p_{\cdot \wedge \sigma} =: U.
\end{align*}
Since \(U\) is \(\ensuremath{\mathcal{F}}\)-predictable, Proposition I.2.4 in \cite{JS} yields that for all \(t \geq 0\), the random variable \(U_t\) is \(\mathcal{F}_{\sigma-}\)-measurable.
Each \(\rho_n\) is \(\ensuremath{\mathcal{F}}\)-predictable thanks to part (III) of Conditions \ref{gen bound cond} and Proposition I.2.13 in \cite{JS}. Therefore, since \(\sigma\) is also \(\ensuremath{\mathcal{F}}\)-predictable, c.f. part (II) of Conditions \ref{gen bound cond}, we have
\begin{align}\label{pred cap F}
\mathcal{F}_{(\sigma \wedge \rho_n)-} = \mathcal{F}_{\sigma-} \cap \mathcal{F}_{\rho_n-},\end{align}
c.f. \cite{DellacherieMeyer78}, p. 119.
Therefore, since \(U\) is an increasing process, we obtain \(\{\rho_n > t\} = \{U_t < n\} \in \mathcal{F}_{\sigma -}\).
This, together with \eqref{pred cap F} and \(\{\rho_n > t\} \in \mathcal{F}_{\rho_n -}\), yields that \(\{\rho_n > t\} \in \mathcal{F}_{(\sigma \wedge \rho_n)-}\).
Since
\(U_{\sigma} \ensuremath{1}_{\{\sigma < \infty\}}\) is \(\mathcal{F}_{\sigma-}\)-measurable, c.f. Proposition I.2.4 in \cite{JS}, and \(\{\sigma < \infty\} \in \mathcal{F}_{\sigma -}\), we also obtain that
\begin{align*}
\{\sigma< \infty\} &\cap \{\rho_n = \infty\}
= \{\sigma < \infty\}\cap \{U_\sigma \ensuremath{1}_{\{\sigma < \infty\}} < n\} \in \mathcal{F}_{\sigma -}
\end{align*}
This, together with
\begin{align*}
\{\sigma = \infty\} \cap \{\rho_n = \infty\} = \{\sigma \wedge \rho_n = \infty\} \in \mathcal{F}_{(\sigma \wedge \rho_n) -} \subseteq \mathcal{F}_{\sigma-}
\end{align*}
yields that \(\{\rho_n = \infty\} \in \mathcal{F}_{\sigma-}\). Employing again \eqref{pred cap F} and \(\{\rho_n = \infty\} \in \mathcal{F}_{\rho_n -}\), we conclude that \(\{\rho_n = \infty\} \in \mathcal{F}_{(\sigma \wedge \rho_n)-}\).
This finishes the proof.
\end{proof}
\hspace{-0.6cm}
\textit{Proof of Proposition \ref{gen coro extended stoch exp}:}
Part (I) of Conditions \ref{gen bound cond} implies \(\sigma \leq \lim_{n \to \infty} \tau_n = \tau\).
Therefore \(\mathcal{F}_{(\sigma\wedge \rho_n)-} \subseteq \mathcal{F}_{(\tau \wedge \rho_n)-}\).
Note that
\begin{align*}
\{\rho_n > t\} &= \big\{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}/\widetilde{Z}_-^2 \cdot C(\widetilde{Z})^p_t < n\big\},\\
\{\rho_n = \infty\} &= \big\{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}/\widetilde{Z}_-^2 \cdot C(\widetilde{Z})^p_\infty < n,\ \forall t \geq 0\big\}.
\end{align*}
Now the claim of (i) follows from
Lemma \ref{con extended loc mart} and Proposition \ref{main prop uni 1}, and the claim of (ii) follows from Lemma \ref{con extended loc mart},
Proposition \ref{main prop uni 2}, and the identity
\begin{align*}
\bigcup_{n \in \mathbb{N}} \{\rho_n = \infty\} &= \big\{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}/\widetilde{Z}_-^2 \cdot C(\widetilde{Z})^p_\infty < \infty \big\}.
\end{align*}
This finishes the proof. \qed
\subsubsection{Martingality of Generalized Stochastic Exponentials}\label{Martingality of Generalized Stochastic Exponentials}
Let us now expand the setting considered by \cite{KLS-LACOM2} in their study of local absolute continuity of laws of semimartingales by allowing for stochastic lifetimes.
The continuous case was studied by \cite{RufSDE}.
Here we derive condition for a general semimartingale setting.
We assume
the following:
\begin{enumerate}
\item[\textup{(I)}]
\(\mathbb{H} = \mathbb{R}^d\) for \(d \in \mathbb{N}\).
\item[\textup{(II)}]
For any announcing sequence \((\tau_n)_{n \in \mathbb{N}}\) for \(A\), the random time \(\tau := \lim_{n \to \infty}\tau_n\) is an \(\ensuremath{\mathcal{F}}\)-predictable time.
\item[\textup{(III)}]
We choose a \emph{good version} of \(C\), i.e. we assume that \(C = c \cdot \bar{c}\), where \(c\) is \(\ensuremath{\mathcal{F}}\)-predictable and \(\bar{c}\) is real-valued, continuous and increasing.
\item[\textup{(IV)}]
We have identically \(a_t := \nu(\{t\} \times \mathbb{R}^d) \leq 1\) for all \(t \geq 0\).
\item[\textup{(V)}]
We are given a positive \(\mathcal{P}(\ensuremath{\mathcal{F}}) \otimes \mathcal{B}^d\)-measurable function \(U\) such that
\begin{align*}
\widehat{U}_t := \int_{\mathbb{R}^d} U(t, x)\nu(\{t\} \times \operatorname{d}\hspace{-0.003cm} x) \leq 1,
\end{align*}
for all \(t \geq 0\), and \(\{a = 1\} \subseteq \{\widehat{U} = 1\}\).
\item[\textup{(VI)}]
We are given an \(\ensuremath{\mathcal{F}}\)-predictable \(\mathbb{R}^d\)-valued process \(K\). The transposed of \(K\) is denoted by \(K^*\).
\end{enumerate}
We define by \(R\) a process on \(A\) such that
\begin{align}\label{R}
&R^{\tau_n} =KcK^* \cdot \bar{c}_{\cdot \wedge \tau_n} + \big(1 - \sqrt{U}\hspace{0.02cm}\big)^2* \nu_{\cdot \wedge \tau_n} + \sum_{s \leq \cdot \wedge \tau_n} \bigg( \sqrt{1 - a_s} - \sqrt{1 - \widehat{U}}\ \bigg)^2,
\end{align}
and denote its extension \eqref{extension} by \(\widetilde{R}\).
We furthermore set
\begin{align*}
U' := U - 1 + \frac{\widehat{U} - a}{1 - a}\textup{ and } \rho_n := \inf(t \geq 0 : \widetilde{R}_{t} \geq n).
\end{align*}
\begin{SA}\label{SA2}
Assume that \(\widetilde{R}\) \emph{does not jump to infinity}, i.e. \(\widetilde{R}_{\rho-} = \infty\) on \(\{\rho < \infty\}\), where \(\rho := \inf(t \geq 0 : \widetilde{R}_{t} = \infty)\).
\end{SA}
\begin{remark}\label{pred remark}
Standing Assumption \ref{SA2} implies that \(\widetilde{R}\) is right-continuous and hence that \((\rho_n)_{n \in \mathbb{N}}\) is a sequence of \(\ensuremath{\mathcal{F}}\)-stopping times.
Moreover, for all \(n \in \mathbb{N}\) we have \(\rho_n < \rho\) on \(\{\rho < \infty\}\). Therefore \(\rho\) is an \(\ensuremath{\mathcal{F}}\)-predictable time as it can be fortelled by the sequence \((\rho_n \wedge n)_{n \in \mathbb{N}}\) and \(\{\rho = 0\} = \emptyset\), c.f. \cite{DellacherieMeyer78}, Theorem IV.71.
\end{remark}
Next we
define the process
\begin{align*}
N^{\tau_n \wedge \rho_n} := K \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]} \cdot X^c + U' \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]\times \mathbb{R}^d} * (\mu^X - \nu),
\end{align*}
whose well-definedness is established in the following lemma.
\begin{lemma}\label{Lemma N}
For all \(n \in \mathbb{N}\) the process \(N^{\tau_n \wedge \rho_n}\) is a local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale. Moreover, it holds that \(\Delta N^{\tau_n \wedge \rho_n} > -1\).
\end{lemma}
\begin{proof}
Since \(\rho_n < \rho\) on \(\{\rho < \infty\}\) we obtain
\begin{align}\label{Delta R bound}
\Delta \widetilde{R}_{\tau_n \wedge \rho_n} = 2\bigg(1 - \widehat{\sqrt{U_{\tau_n \wedge \rho_n}}} - \sqrt{(1 - a_{\tau_n \wedge \rho_n})(1 - \widehat{U}_{\tau_n \wedge \rho_n})}\ \bigg) \leq 2,
\end{align}
where we denote
\begin{align*}
\widehat{\sqrt{U_{\tau_n \wedge \rho_n}}} := \int_{\mathbb{R}^d} \sqrt{U(\tau_n \wedge \rho_n, x)} \nu(\{\tau_n \wedge \rho_n\} \times \operatorname{d}\hspace{-0.003cm} x).
\end{align*}
Hence it holds that
\begin{align}\label{R bound}
\widetilde{R}_{\tau_n \wedge \rho_n} \leq n + 2.
\end{align}
This yields
\(K\ensuremath{1}_{ [\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]} \in L^2(X_{\cdot \wedge \tau_n}^c, \ensuremath{\mathcal{F}}, {\mathbf{P}})\).
Note that for \((\omega, t) \in [\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]\) we have
\begin{equation}\label{delta N}
\begin{split}
\widetilde{U}'_t (\omega):\hspace{-0.13cm}&= U'(\omega, t, \Delta X_t(\omega)) \ensuremath{1}_{\{\Delta X_t(\omega) \not= 0 \}} - \int_{\mathbb{R}^d} U'(\omega, t, x) \nu(\omega, \{t\} \times \operatorname{d}\hspace{-0.003cm} x)
\\&= (U(\omega, t, \Delta X_t(\omega)) - 1) \ensuremath{1}_{\{\Delta X_t(\omega) \not = 0\}} - \frac{\widehat{U}_t (\omega)- a_t(\omega)}{1 - a_t(\omega)}\ensuremath{1}_{\{\Delta X_t(\omega) = 0\}},
\end{split}
\end{equation}
with the convention that \(0/0 = 0\).
Since \(U\) is positive, we have \(\widetilde{U}' > -1\) on \([\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}] \cap \{\Delta X \not = 0\}\).
The assumption that identically \(\widehat{U} \leq 1\) implies \(\widehat{U} - a \leq 1 - a\), where equality holds on \(\{\widehat{U} = 1\}\). Thanks to the assumption that \(\{a = 1\}\subseteq \{\widehat{U} = 1\}\) and the convention that \(0/0 = 0\),
we obtain that \(\widetilde{U}' > -1\) on \([\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]\cap\{\Delta X= 0\}\), from which
we deduce that \(\widetilde{U}' > -1\) on \([\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]\).
Due to Equation 12.41 in \cite{J79} we have on \([\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]\)
\begin{equation}\label{12.41}
\begin{split}
\bigg(1 - \sqrt{1 + U' - \widehat{U}'}\ &\bigg)^2 * \nu + \sum_{s \leq \cdot}(1 - a_s)\bigg(1 - \sqrt{1 - \widehat{U}'}\ \bigg)^2
\\&= \big(1 - \sqrt{U}\big)^2 * \nu + \sum_{s \leq \cdot} \bigg(\sqrt{1 - a_s} - \sqrt{1 - \widehat{U}}\ \bigg)^2.
\end{split}
\end{equation}
Thanks to Theorem II.1.33 in \cite{JS} and the bound \eqref{R bound} we conclude that \(U' \ensuremath{1}_{[\hspace{-0.06cm}[ 0,\tau_n \wedge \rho_n]\hspace{-0.06cm}]\times \mathbb{R}^d} \in G_\textup{loc}(\mu^{X^{\tau_n}}, \ensuremath{\mathcal{F}}, {\mathbf{P}}).\)
This implies that \(N^{\tau_n \wedge \rho_n}\) is a well-defined local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
Due to the definition of the integral process \(U' \ensuremath{1}_{[\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}] \times \mathbb{R}^d} * (\mu^X - \nu)\), we have \(\Delta N^{\tau_n \wedge \rho_n} = \widetilde{U}'_{\cdot \wedge \tau_n \wedge \rho_n} >-1\).
This finishes the proof.
\end{proof}
In view of this lemma we can define a non-negative local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale on \(\bar{A} := \bigcup_{n \in \mathbb{N}} [\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]\) by \(Z := \mathcal{E}(N)\).
Now we extend \(Z\) to \(\widetilde{Z}\) in the same manner as in \eqref{extension}. The extension \(\widetilde{Z}\) is a non-negative local \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale, c.f. Proposition \ref{lemma tilde Z local martingale} in Appendix ~\ref{Extensions of non-negative local martingales on sets of interval type}.
\begin{remark}\label{pred remark}
In the notation of the previous section we have \(\sigma_n = \tau_n \wedge \rho_n\).
Since \(\tau = \lim_{n \to \infty} \tau_n\) is assumed to be \(\ensuremath{\mathcal{F}}\)-predictable, and \(\rho\) is \(\ensuremath{\mathcal{F}}\)-predictable due to Standing Assumption \ref{SA2}, c.f. Remark \ref{pred remark}, the random time \(\lim_{n \to \infty} \sigma_n
= \tau \wedge \rho\) is \(\ensuremath{\mathcal{F}}\)-predictable.
\end{remark}
The following corollary generalizes one direction of Theorem 3.3 in \cite{RufSDE}.
For another generalization we refer to Section \ref{Martingality on Standard Systems} below.
\begin{corollary}\label{coro j 1}
Assume that the SMP associated with \((\mathbb{R}^d; \bar{A}; \eta; X; \dot{B}(h), C, \dot{\nu})\), where on \(\bar{A}\)
\begin{equation}\label{applied triplet}
\begin{split}
\dot{B}(h) &= B(h) + KcK^* \cdot \bar{c} + h(x)(U - 1) * \nu\textit{ and } \dot{\nu} = U\cdot \nu,
\end{split}
\end{equation}
has a solution \({\mathbf{Q}}\) and satisfies \(\{\tau_n \wedge \rho_m, n, m \in \mathbb{N}\}\)-uniqueness.
\begin{enumerate}
\item[\textup{(i)}] If for all \(t \geq 0\) we have
\begin{align*}
{\mathbf{Q}}\big(\widetilde{R}_{t \wedge \rho} < \infty\big) = 1
\end{align*}
then \(\widetilde{Z}\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\item[\textup{(ii)}]
If we have
\begin{align*}
{\mathbf{Q}}\big(\widetilde{R}_{\rho} < \infty\big) = 1
\end{align*}
then \(\widetilde{Z}\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\end{enumerate}
\end{corollary}
The proof of this corollary is based on the following lemma and Proposition \ref{gen coro extended stoch exp}.
\begin{lemma}\label{iden R lemma}
We have
\begin{align}\label{R = C stuff}
\widetilde{R}^{\rho} = \frac{\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}}{\widetilde{Z}_-^2} \cdot C(\widetilde{Z})^p.
\end{align}
\end{lemma}
\begin{proof}
For each \(n \in \mathbb{N}\) we obtain that on \([\hspace{-0.06cm}[ 0, \tau_n \wedge \rho_n]\hspace{-0.06cm}]\)
\begin{align*}
\ensuremath{1}_{\{\widetilde{Z}_- > 0\}}/\widetilde{Z}_-^2 \cdot C(\widetilde{Z})^p
&= \ensuremath{1}_{\{\mathcal{E}(N^{\tau_n \wedge \rho_n})_- > 0\}}/\mathcal{E}(N^{\tau_n \wedge \rho_n})^2_- \cdot C(\mathcal{E}(N^{\tau_n \wedge \rho_n}))^p
= \widetilde{R},
\end{align*}
where we used that identically \(\mathcal{E}(N^{\tau_n \wedge \rho_n})_- > 0\), c.f. Theorem I.4.61 in \cite{JS}, Lemma 8.8 in \cite{J79}, and \eqref{12.41}.
Therefore the identity \eqref{R = C stuff} holds on \(\bar{A}\).
Due to construction, we have
\begin{align*}
\widetilde{R}^\rho = \widetilde{R}^{\tau \wedge \rho}\textup{ and } \ensuremath{1}_{\{\widetilde{Z}_- > 0\}}/\widetilde{Z}_-^2 \cdot C(\widetilde{Z})^p = \ensuremath{1}_{\{\widetilde{Z}_- > 0\}}/\widetilde{Z}_-^2 \cdot C(\widetilde{Z})^p_{\cdot \wedge \tau \wedge \rho}.
\end{align*}
Therefore it remains to study \eqref{R = C stuff} on
\begin{equation}
\label{remains}
\begin{split}
\big(\bar{A} &\ \cup\ ]\hspace{-0.06cm}] \tau \wedge \rho, \infty[\hspace{-0.06cm}[\big)^c
= \begin{cases}
[\hspace{-0.06cm}[ \tau]\hspace{-0.06cm}]& \textup{ on } \{\tau < \rho\} \cap \big(\bigcap_{n \in \mathbb{N}} \{\tau_n < \tau < \infty\}\big) \times \mathbb{R}^+,
\\
[\hspace{-0.06cm}[ \rho ]\hspace{-0.06cm}]&\textup{ on }
\{\rho \leq \tau\} \cap \big(\{ \rho < \infty\}
\times \mathbb{R}^+\big),
\\
\ \emptyset&\textup{ otherwise}.
\end{cases}
\end{split}
\end{equation}
In the first two cases the definition of the extension \eqref{extension} and the c\`adl\`ag paths of \(\widetilde{Z}\) imply \(\widetilde{Z}_{\tau \wedge \rho} = \widetilde{Z}_{(\tau \wedge \rho)-}\). Hence
\(\Delta C(\widetilde{Z})^p_{\tau \wedge \rho} = 0\), using I.3.21 and Theorem I.2.28 (ii) in \cite{JS}.
In the first case of \eqref{remains}, this implies that the identity \eqref{R = C stuff} holds due to the definition of \(\widetilde{R}\), which yields that \(\Delta \widetilde{R}_{\tau \wedge \rho} = 0\).
In the second case of \eqref{remains},
Standing Assumption \ref{SA2} yields
that \(\Delta \widetilde{R}_{\tau \wedge \rho} = 0\).
This finishes the proof.
\end{proof}
\hspace{-0.57cm}
\textit{Proof of Corollary \ref{coro j 1}:}
We set \(\sigma_n = \tau_n \wedge \rho_n\), then part (I) of Conditions \ref{gen bound cond} holds obviously, and part (II) of Conditions \ref{gen bound cond} holds due to Remark \ref{pred remark}.
It follows by the same arguments as in the proof of Theorem III.3.24 and Lemma III.5.27 in \cite{JS}, that on \(\bar{A}\) we have
\begin{equation}\label{ausgerechnet}
\begin{split}
1/ \widetilde{Z}_- \ensuremath{1}_{\{\widetilde{Z}_- > 0\}} M^{{\mathbf{P}}}_{\mu^X}(\widetilde{Z} |\widetilde{\mathcal{P}}(\ensuremath{\mathcal{F}})) &= U,\\
1/\widetilde{Z}_-\ensuremath{1}_{\{\widetilde{Z}_- > 0\}} \cdot \langle\hspace{-0.085cm}\langle \widetilde{Z}, M(h)\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}} &=KcK^* \cdot \bar{c}+ h(x) (U- 1) * \nu.
\end{split}
\end{equation}
Moreover, it follows similarly to \eqref{Delta R bound}, that for all \(n, m \in \mathbb{N}\), \(\Delta \widetilde{R}_{\tau_m \wedge \rho_n} \leq 2\). Therefore,
\begin{align*}
\Delta \widetilde{R}_{\rho_n}
&= \sum_{m \in \mathbb{N}} \Delta \widetilde{R}_{\tau_m \wedge \rho_n} \ensuremath{1}_{]\hspace{-0.06cm}] \tau_{m-1}, \tau_m]\hspace{-0.06cm}]}(\cdot, \rho_n) \leq 2,
\end{align*}
where \(\tau_0 := 0\). Hence, in view of Lemma \ref{iden R lemma} and noting that
\begin{align*}
\rho_n = \inf(t \geq 0 : \widetilde{R}^{\rho} \geq n),
\end{align*}
we conclude that part (III) of Conditions \ref{gen bound cond} holds.
Now the claim follows from the identities \eqref{ausgerechnet}, Lemma \ref{iden R lemma} and Proposition ~\ref{gen coro extended stoch exp}.
\qed
\subsubsection{An Explosion Condition}\label{Explosion criteria}
In this section we derive an explosion-type condition.
Let \((m_n)_{n \in \mathbb{N}}\) be an increasing sequence of non-negative real numbers.
We define
\begin{align}\label{explosion sequence}
\rho_n := \inf(t \geq 0 : \|X_t\| \geq m_n \textup{ or } \|X_{t-}\|\geq m_n) \wedge n,
\end{align}
and impose the following condition:
\begin{SA}\label{SA3}
Assume that for all \(n \in \mathbb{N}\) we have \(\rho_n \leq \lim_{m \to \infty} \tau_m\), where \((\tau_n)_{n \in \mathbb{N}}\) is an arbitrary announcing sequence for \(A\), and
that \((\rho_n)_{n \in \mathbb{N}}\) satisfies Convention \ref{conv}.
\end{SA}
\begin{remark}\label{local character remark}
If for all \(n \in \mathbb{N}\) there exists a positive constant \(c_n\) such that
\begin{align*}
\frac{\ensuremath{1}_{\{Z_- > 0\}}}{Z^2_-} \cdot C(Z)^p_{\rho_n} \leq c_n,
\end{align*}
where \(C(Z)\) is defined as in \eqref{C(Z)}, then the sequence \((\rho_n)_{n \in \mathbb{N}}\) satisfies Convention \ref{conv}, c.f. Lemma \ref{besser als novi lemma} in Appendix \ref{An Integrability Condition to Indentify}.
\end{remark}
Standing Assumption \ref{SA3} has a \emph{local character}, c.f. Remark \ref{local character remark}, and therefore a broad scope.
The following result gives a sufficient condition for the martingality of \(Z\). In the one-dimensional diffusion setting a similar result may be deduced from Theorem 2.1 in \cite{MU(2012)}.
\begin{corollary}\label{explosion cond coro}
\begin{enumerate}
\item[\textup{(i)}]
Assume part \textup{(I)} of Conditions \ref{cond main uni}.
If we have for all \(t \geq 0\)
\begin{align}\label{explosion cond}
\lim_{n \to \infty} {\mathbf{Q}}\bigg(\sup_{s \leq t} \|X_s\| < m_n\bigg) = 1,
\end{align}
then \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale. If \({\mathbf{P}}\)-a.s. \(\rho_n \uparrow_{n \to \infty} \infty\) and \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale, then \eqref{explosion cond} holds.
\item[\textup{(ii)}]
Let \((\bar{B}(h), C, \bar{\nu})\) be a candidate triplet which coincides with \((\dot{B}(h), C, \dot{\nu})\) from \eqref{candidtate triplet main theorem} on \(\bar{A}\).
If the SMP associated with \((\mathbb{H}; [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[; \eta; X; \bar{B}(h), C, \bar{\nu})\) has a solution \({\mathbf{Q}}\) and satisfies \(\{\rho_n \wedge \tau_m, n, m \in \mathbb{N}\}\)-uniqueness and if \(m_n \uparrow_{n \to \infty} \infty\) then
\(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\end{enumerate}
\end{corollary}
\begin{proof}
(i). Due to Standing Assumption \ref{SA3}, part (II) of Conditions \ref{cond main uni} is satisfied.
Moreover, note that for all \(t \geq 0\) we have
\(
\{\rho_n > t\} = \{ \sup_{s \leq t} \|X_s\| < m_n\} \cap \{n > t\}.
\)
Therefore, Proposition \ref{main prop uni 1} implies that \eqref{explosion cond} is sufficient for \(Z\) to be an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
If \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale and \({\mathbf{P}}\)-a.s. \(\rho_n \uparrow_{n \to \infty} \infty\), then Lemma \ref{ruf mimic}, Lemma \ref{lemma 2} and Standing Assumption \ref{SA3} imply that \eqref{explosion cond} is satisfied.
(ii). Corollary \ref{coro subinterval} and Remark \ref{uni prop} yield that part (I) of Conditions \ref{cond main uni} is satisfied. Moreover, since \(m_n\uparrow_{n \to \infty} \infty\) and \({\mathbf{Q}}\)-a.e. path of \(X\) is in \(\mathbb{D}^\mathbb{H}\) we obtain that
\begin{align*}
\lim_{n \to \infty} {\mathbf{Q}}\left(\sup_{s \leq t} \|X_s\| < m_n\right) = {\mathbf{Q}}\left(\sup_{s \leq t} \|X_s\| < \infty\right) = 1.
\end{align*}
Therefore part (i) yields the claim.
\end{proof}
Note that the condition \eqref{explosion cond} only depends on the process \(X\) under \({\mathbf{Q}}\). The condition may even imply that \(Z\) is a true martingale, while \(X\) explodes under \({\mathbf{P}}\), c.f. Example 3.1 in \cite{MU(2012)}.
It is fascinating and surprising that such a condition
not only holds in the well-studied one-dimensional diffusion setting, but also in our general infinite dimensional setting which allows for discontinuities.
In the case where \(A = [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[\) and \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}})\) is full we obtain a classical explosion condition which we state in the following corollary.
\begin{corollary}
Assume that \(m_n \uparrow_{n \to \infty} \infty\), \(A = [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[\), that \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}})\) is full, and that the SMP \((\mathbb{H}; A; \eta; X; \dot{B}(h), C, \dot{\nu})\), where \(\dot{B}(h)\) and \(\dot{\nu}\) are given by \eqref{candidtate triplet main theorem} on \(A\), satisfies \(\{\rho_n \wedge \tau_m, n, m \in \mathbb{N}\}\)-uniqueness.
Then the following is equivalent:
\begin{enumerate}
\item[\textup{(i)}] \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\item[\textup{(ii)}] The SMP associated with \((\mathbb{H}; A; \eta; X; \dot{B}(h), C, \dot{\nu})\) has a solution.
\end{enumerate}
\end{corollary}
\begin{proof} The claim follows immediately from Theorem \ref{main theorem new 2} and Corollary \ref{explosion cond coro} (ii).
\end{proof}
\subsection{Martingality on Standard Systems}\label{Martingality on Standard Systems}
The following approach depends on topological properties of the underlying filtered space and is related to the setting of the F\"ollmer measure, c.f. \cite{follmer72,perkowski2015}.
In particular, the result applies if the underlying filtered space is given by \((\mathbb{D}^{E_\Delta}, \mathcal{D}^{E_\Delta}, \mathfrak{D}^{E_\Delta})\) where \(E\) is an arbitrary Polish space. Hence we even allow \(Z\) to be driven by \(E_\Delta\)-valued processes.
Connecting the approach to the SMP setting of the previous sections allows us to drop the uniqueness assumptions of Conditions ~\ref{cond main uni}.
For the terminology used in this section we refer to the Appendix \ref{Extension of Probability Measures}.
We recall the notation \(\rho = \lim_{n \to \infty} \rho_n\), where \((\rho_n)_{n \in \mathbb{N}}\) is given as in Convention ~\ref{conv}.
\begin{proposition}\label{ext prop}
Assume that \((\Omega, \mathcal{F}_{\rho_n-})_{n \in \mathbb{N}}\) is a standard system. Then there exists a unique probability measure \({\mathbf{Q}}\) on \((\Omega, \mathcal{F}_{\rho-})\) such that for all \(n \in \mathbb{N}\), \({\mathbf{Q}}(F) = \mathrm{I\kern-.2em E}[Z_{\rho_n} \ensuremath{1}_F]\) for all \(F \in \mathcal{F}_{\rho_n-}\).
If in addition \((\Omega, \mathcal{F}_{\infty-})\) is a standard Borel space and \(\mathcal{F}_{\rho-}\) is countably generated, then \({\mathbf{Q}}\) has an extension to \((\Omega, \mathcal{F}_{\infty-})\).
\end{proposition}
\begin{proof}
Denote \({\mathbf{Q}}_n := Z_{\rho_n} \cdot {\mathbf{P}}\) and note that for all \(A \in \mathcal{F}_{\rho_{n}-}\) we have
\begin{align*}
{\mathbf{Q}}_{n+1} (A) = \mathrm{I\kern-.2em E}[\mathrm{I\kern-.2em E}[Z_{\rho_{n+1}}\ensuremath{1}_A|\mathcal{F}_{\rho_n}]] = \mathrm{I\kern-.2em E}[Z_{\rho_n}\ensuremath{1}_{A}] = {\mathbf{Q}}_n(A),
\end{align*}
where we use Doob's optional stopping theorem.
Hence the first claim of the proposition is an immediate consequence of Parthasarathy's extension theorem, c.f. Theorem \ref{extension P} in Appendix \ref{Extension of Probability Measures}, and the identity \(\bigvee_{n \in \mathbb{N}} \mathcal{F}_{\rho_n-} = \mathcal{F}_{\rho-}\).
The second claim follows from Theorem \ref{extension P2} in Appendix \ref{Extension of Probability Measures}.
\end{proof}
As an immediate consequence of Proposition \ref{ext prop}, Lemma \ref{ruf mimic} and Lemma \ref{neues lemma} we obtain the following condition for the martingality of \(Z\).
\begin{proposition}\label{theorem extension1}
Assume that \((\Omega, \mathcal{F}_{\rho_n-})_{n \in \mathbb{N}}\) is a standard system and denote by \({\mathbf{Q}}\) the probability measure as given in Proposition \ref{ext prop}.
\begin{enumerate}
\item[\textup{(i)}]
If \eqref{Main condition} holds,
then \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale. If \({\mathbf{P}}\)-a.s. \(\rho_n \uparrow_{n \to \infty} \infty\) and \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale, then \eqref{Main condition} holds.
\item[\textup{(ii)}]
If \eqref{Main condition ui} holds,
then \(Z\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\end{enumerate}
\end{proposition}
In general, due to the lack of information on \({\mathbf{Q}}\), these conditions are difficult to check.
However, \eqref{Main condition} holds if \(\rho_n(\omega) \uparrow_{n \to \infty} \infty\) for all \(\omega \in \Omega\). This path-wise condition is stated in the following corollary, which generalizes Corollary 3.4 in \cite{RufSDE}.
\begin{corollary}\label{path assumption}
Assume that \((\Omega, \mathcal{F}_{\rho_n-})_{n \in \mathbb{N}}\) is a standard system.
If \(\{\rho_n > t\} \uparrow_{n \to \infty} \Omega\) for all \(t \geq 0\), then \(Z\) is an \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale. If even \(\{\rho_n =\infty\} \uparrow_{n \to \infty} \Omega\), then \(Z\) is a uniformly integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale.
\end{corollary}
\begin{remark}
On the path-space \((\mathbb{C}^E, \mathcal{C}^E, \mathfrak{C}^E)\), the pathwise condition \(\{\rho_n > t\} \uparrow_{n \to \infty} \mathbb{C}^E\) from Corollary \ref{path assumption} is not sufficient for \(Z\) to be a \((\mathfrak{C}^E, {\mathbf{P}})\)-martingale, c.f. \cite{RufSDE} for a counterexample.
\end{remark}
We now assume a setting comparable to the previous section, which enables us to derive structural properties of \({\mathbf{Q}}\).
These allow us to express \({\mathbf{Q}}\) in Proposition \ref{theorem extension1} as a solution to a SMP:
\begin{condition}\label{cond topo}
\begin{enumerate}
\item[\textup{(I)}]
Assume that \((\Omega, \mathcal{F}_{\rho_n-})_{n \in \mathbb{N}}\) is a standard system, that \(\mathcal{F}_{\rho -}\) is countably generated and that \((\Omega, \mathcal{F}_{\infty-})\) is a standard Borel space.
\item[\textup{(II)}]
Assume that \(\mathcal{F} = \mathcal{F}_{\infty - }\) and that for each \(n \in \mathbb{N}\) there exists an \(m_n > n\) such that
\(\rho_n \wedge \tau_n < \rho_{m_n} \wedge \tau_{m_n}\) on \(\{\rho_n \wedge \tau_n< \infty\}\).
\item[\textup{(III)}]
Assume that \((\tau_n \wedge \rho_n)_{n \in \mathbb{N}}\) is a sequence of \(\ensuremath{\mathcal{F}}\)-predictable times, and that \(\ensuremath{\mathcal{F}}\) is quasi-left continuous.
\end{enumerate}
\end{condition}
\begin{theorem}\label{theorem extension2}
Assume Standing Assumption \ref{SA}, part \textup{(I)} of Conditions \ref{cond topo}, and additionally part \textup{(II)} or \textup{(III)} of Conditions \ref{cond topo}. Moreover, denote by \({\mathbf{Q}}\) the extension of the probability measure as given in Proposition ~\ref{ext prop}.
Then
\({\mathbf{Q}}\) is a solution to the SMP associated with \((\mathbb{H}; \bar{A}; \eta; X; \dot{B}(h), C, \dot{\nu})\), where \(\bar{A} = A \cap (\bigcup_{n \in \mathbb{N}} [\hspace{-0.06cm}[ 0, \rho_n]\hspace{-0.06cm}])\), and \(\dot{B}(h)\) and \(\dot{\nu}\) are given by \eqref{candidtate triplet main theorem}.
\end{theorem}
\begin{proof}
Lemma \ref{lemma stopped} yields that the probability measure \({\mathbf{Q}}_{n, n} = Z_{\tau_n \wedge \rho_n} \cdot {\mathbf{P}}\) is a solution to the SMP associated with \((\mathbb{H}; \tau_n \wedge \rho_n; \eta; X; \dot{B}(h), C, \dot{\nu})\).
Due to the construction of \({\mathbf{Q}}\) we have \({\mathbf{Q}}_{n, n} = {\mathbf{Q}}\) on \(\mathcal{F}_{(\tau_n \wedge \rho_n)-}\). If we assume part (III) of Conditions \ref{cond topo}, then \({\mathbf{Q}}_{n, n} = {\mathbf{Q}}\) on \(\mathcal{F}_{\tau_n \wedge \rho_n}\) which yields the claim.
Now assume part (II) of Conditions \ref{cond topo}. Following the proof of Proposition 3.1 in \cite{RufSDE}, we obtain for all \(F \in \mathcal{F}_{\tau_n \wedge \rho_n}\)
\begin{align*}
{\mathbf{Q}}(F) = \mathrm{I\kern-.2em E}[Z_{\tau_{m_n} \wedge \rho_{m_n}} \ensuremath{1}_{F \cap \{\tau_n \wedge \rho_{n} < \infty\}}] + \mathrm{I\kern-.2em E}[Z_{\tau_n \wedge \rho_n} \ensuremath{1}_{F \cap \{\tau_n \wedge \rho_n = \infty\}}] = \mathrm{I\kern-.2em E}[Z_{\tau_n \wedge \rho_n} \ensuremath{1}_F],
\end{align*}
where we used \cite{DellacherieMeyer78}, Theorem V.56, and Doob's optional stopping theorem.
This finishes the proof.
\end{proof}
Combining Proposition \ref{theorem extension1} and Theorem \ref{theorem extension2} leads to a martingale condition similarly to the Propositions \ref{main prop uni 1} and \ref{main prop uni 2}.
\begin{remark}
If the underlying filtered space is given by
\((\mathbb{D}^{\mathbb{H}_\Delta}, \mathcal{D}^{\mathbb{H}_\Delta}, \mathfrak{D}^{\mathbb{H}_\Delta}),
\)
the topological assumptions of Theorem \ref{theorem extension2}, part (I) of Conditions \ref{cond topo}, are satisfied, c.f. Appendix \ref{Extension of Probability Measures}.
\end{remark}
\section{A Case Study - Martingality in Terms of Infinite-Dimensional SDEs}
\label{A Case Study - Martingality in Terms of SDEs driven by Hilbert-Space-Valued Brownian Motion}
The generality of our setting allows us to derive sufficient conditions for the martingality of stochastic exponentials driven by Hilbert-space-valued Brownian motion.
The main result of this section is in the spirit of Theorem 3.1.1 in \cite{BR16}.
Let us first introduce some additional terminology.
As in the previous section let \(\mathbb{H}\) be a real separable Hilbert space. Moreover, we fix a filtered probability space \((\Omega, \mathcal{F}, \ensuremath{\mathcal{F}}, {\mathbf{P}})\) with right-continuous filtration \(\ensuremath{\mathcal{F}}\) and a non-negative and self-adjoint nuclear operator \(Q\in \mathcal{N}(\mathbb{H}, \mathbb{H})\).
\begin{definition}
An \(\mathbb{H}\)-valued process \(W\) is called an \((\mathbb{H}, Q, \ensuremath{\mathcal{F}}, {\mathbf{P}})\)-\emph{Brownian motion}, if the following holds:
\begin{enumerate}
\item[\textup{(i)}]
\({\mathbf{P}}\)-a.s. \(W_0 = 0\).
\item[\textup{(ii)}]
\(W\) has \({\mathbf{P}}\)-a.s. continuous paths.
\item[\textup{(iii)}]
For all \(s < t\) the random variable \(W_t - W_s\) is \({\mathbf{P}}\)-independent of \(\mathcal{F}_s\);
\item[\textup{(iv)}]
For all \(s < t\) and all \(h \in \mathbb{H}\) the (real-valued) random variable \((W_t - W_s) \bullet h\) is centered Gaussian with variance \((t-s) (Qh \bullet h)\).
\end{enumerate}
\end{definition}
The operator \(Q\) is usually called \emph{covariance} of \(W\).
Moreover, as in the finite-dimensional case we have a L\'evy characterization theorem: An \(\mathbb{H}\)-valued continuous \(\ensuremath{\mathcal{F}}\)-adapted process \(W\) with \({\mathbf{P}}\)-a.s. \(W_0 = 0\) is an \((\mathbb{H},Q, \ensuremath{\mathcal{F}}, {\mathbf{P}})\)-Brownian motion if and only if it is a square-integrable \((\ensuremath{\mathcal{F}}, {\mathbf{P}})\)-martingale with \(\langle\hspace{-0.085cm}\langle W, W\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}= I Q\), where \(I_t = t\) denotes the identity process, c.f. Proposition 4.11 in \cite{MP80} or
Theorem 4.6 in \cite{deprato}.
It is also well-known that \(Q\) has a decomposition of the form \(Q = Q^{1/2} Q^{1/2}\), where \(Q^{1/2}\) is a non-negative and self-adjoint Hilbert-Schmidt operator.
We denote by \(\mathbb{C}^\mathbb{H}_0\) the set of continuous functions \(\mathbb{R}^+ \to \mathbb{H}\) which start at \(0\).
The coordinate process is denoted by \(\widehat{X}\).
We equip \(\mathbb{C}^\mathbb{H}_0\) with the \(\sigma\)-field \(\mathcal{C}^\mathbb{H}_0 := \sigma(\widehat{X}_t, t \geq 0)\) and define the filtrations \(\mathfrak{C}^{\mathbb{H}, 0}_0 := (\mathcal{C}^{\mathbb{H}, 0}_{0, t})_{t \geq 0} := (\sigma(\widehat{X}_s, s \leq t))_{t \geq 0}\) and \(\mathfrak{C}^\mathbb{H}_0 := (\mathcal{C}^{\mathbb{H}}_{0, t})_{t \geq 0} := (\bigcap_{s > t} \mathcal{C}^{\mathbb{H}, 0}_{0, s})_{t \geq 0}\).
Let \({\mathbf{P}}\) be a solution to the SMP on \((\mathbb{C}^\mathbb{H}_0, \mathcal{C}^\mathbb{H}_0, \mathfrak{C}^\mathbb{H}_0)\) associated with \((\mathbb{H}; [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[; \epsilon_0; \widehat{X}; 0, IQ, 0)\).
In other words, we assume that the coordinate process is an \((\mathbb{H}, Q, \mathfrak{C}^\mathbb{H}_0, {\mathbf{P}})\)-Brownian motion.
Moreover, let \(\varphi : \mathbb{R}^+ \times \mathbb{C}^\mathbb{H}_0 \to \mathbb{H}\) be \(\mathfrak{C}^\mathbb{H}_0\)-predictable such that \(\varphi(\cdot, 0)\) is constant.
We use the notation \(\varphi(\cdot, \omega) =: \varphi_\cdot(\omega)\).
Next we formulate the main result of this section, which states that under a local Lipschitz and a linear growth condition
the stochastic exponential
\begin{align*}
Z: = \mathcal{E}\left(\varphi(\widehat{X}) \cdot \widehat{X}\right)
\end{align*}
is a true martingale.
For details concerning stochastic integration w.r.t. Hilbert-space-valued processes we refer to the monographs of \cite{metivier,MP80}.
To clarify the notation of \(\varphi(\widehat{X})\cdot\widehat{X}\), we note that we identify \(\varphi\) with \(v \rightsquigarrow \varphi \bullet v\).
\begin{theorem}\label{existence uniqueness concrete}
Assume the following:
\begin{enumerate}
\item[\textup{(i)}]
For all \(\alpha \in (0, \infty)\) there exists a positive c\`adl\`ag increasing function \(L^\alpha\) such that
\begin{align*}
\qquad\quad\|Q\varphi(t, \omega) - Q\varphi(t, \omega^*)\|_{\mathbb{H}} \leq L^{\alpha}_t \sup_{s < t} \|\omega(s) - \omega^*(s)\|_\mathbb{H}
\end{align*}
for all \(t \geq 0\), and all \(\omega, \omega^* \in \{\bar{\omega} \in \mathbb{C}^{\mathbb{H}}_0 : \sup_{s < t} \|\bar{\omega}(s)\|_\mathbb{H} \leq \alpha\}\).
\item[\textup{(ii)}] There exists a constant \(\lambda > 0\) such that
\begin{align*}
\|Q^{1/2} \varphi(t, \omega)\|^2_\mathbb{H} \leq \lambda\bigg(1 + \sup_{s < t} \|\omega(s)\|^2_\mathbb{H}\bigg),
\end{align*}
for all \((t, \omega) \in \mathbb{R}^+ \times \mathbb{C}^{\mathbb{H}}_0\).
\end{enumerate}
Then the process
\(
Z = \mathcal{E}(\varphi(\widehat{X}) \cdot \widehat{X})
\)
is an \((\mathfrak{C}^\mathbb{H}_0, {\mathbf{P}})\)-martingale.
\end{theorem}
\subsection{Proof of Theorem \ref{existence uniqueness concrete}}\label{Proof of Theorem}
We assume the notation of Section \ref{MSE}. Then \(A' = A = [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[\) and we may set \(\sigma_n = \tau_n = n\), and define
\begin{align}\label{rho neu}
\rho_n := \inf\big(t \geq 0 : \|\widehat{X}_t\| \geq n\big) \wedge n.
\end{align}
Let us first show that \(Z\) is a local \((\mathfrak{C}^{\mathbb{H}}_0, {\mathbf{P}})\)-martingale with \((\mathfrak{C}^\mathbb{H}_0, {\mathbf{P}})\)-localizing sequence \((\rho_n)_{n \in \mathbb{N}}\).
Due to Assumption (ii) in Theorem \ref{existence uniqueness concrete} we have identically
\begin{align*}
\left(\varphi(\widehat{X}) \bullet Q\varphi(\widehat{X})\right) \cdot I_{\rho_n} \leq \lambda n \left(1 + \sup_{s < \rho_n} \|\widehat{X}_s\|^2_\mathbb{H}\right) \leq \lambda n(1 + n^2).
\end{align*}
Hence Theorem 2.3 in \cite{gawarecki2010stochastic} yields that \(\varphi(\widehat{X}) \cdot \widehat{X}_{\cdot \wedge \rho_n}\) is a square-integrable \((\mathfrak{C}^\mathbb{H}_0, {\mathbf{P}})\)-martingale, implying that \(Z\) is a local \((\mathfrak{C}^\mathbb{H}_0, {\mathbf{P}})\)-martingale.
This furthermore implies that \((\rho_n)_{n \in \mathbb{N}}\) satisfies Convention \ref{conv}.
Now the claim of Theorem \ref{existence uniqueness concrete} is an immediate consequence of Corollary \ref{explosion cond coro} (ii) and
Lemma \ref{key1} below.
\begin{lemma}\label{key1}
Under the assumptions of Theorem \ref{existence uniqueness concrete},
the SMP on \((\mathbb{C}^\mathbb{H}_0, \mathcal{C}^\mathbb{H}_0, \mathfrak{C}^\mathbb{H}_0)\) associated with \((\mathbb{H}; A; \varepsilon_0; \widehat{X}; \dot{B}, IQ, 0)\), where
\begin{align}\label{BC}
\dot{B} = \frac{\ensuremath{1}_{\{Z_- > 0\}}}{Z_-} \cdot \langle\hspace{-0.085cm}\langle Z, M(h)\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}},
\end{align}
has a solution and satisfies \(\{\rho_n \wedge \tau_m, n, m \in \mathbb{N}\}\)-uniqueness.
\end{lemma}
\begin{proof}
First we compute \(\langle\hspace{-0.085cm}\langle Z, M(h)\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}} = \langle\hspace{-0.085cm}\langle Z, \widehat{X}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}\). In view of Equation 4.1.4 in \cite{MP80}, \(\langle\hspace{-0.085cm}\langle Z, \widehat{X}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}\) is the unique
\(\mathfrak{C}^\mathbb{H}_0\)-predictable process of finite variation such that \(Z \widehat{X} - \langle\hspace{-0.085cm}\langle Z, \widehat{X}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}\) is a local \((\mathfrak{C}^\mathbb{H}_0, {\mathbf{P}})\)-martingale.
Since \(\mathbb{H}\) is assumed to be separable, weak and strong measurability are equivalent. Therefore \(Z \widehat{X} - \langle\hspace{-0.085cm}\langle Z, \widehat{X}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}}\) is an \(\mathbb{H}\)-valued local \((\mathfrak{C}_0^\mathbb{H}, {\mathbf{P}})\)-martingale if and only if for all \(v \in \mathbb{H}\) the process \(Z \widehat{X} \bullet v - \langle\hspace{-0.085cm}\langle Z, \widehat{X}\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}} \bullet v\) is an \(\mathbb{R}\)-valued local \((\mathfrak{C}^\mathbb{H}_0, {\mathbf{P}})\)-martingale, c.f. \cite{gawarecki2010stochastic}, p. 21.
Define \(L^{\pm v} : \mathbb{H} \to \mathbb{R}\) by \(L^{\pm v} w = (\varphi(\widehat{X}) \pm v) \bullet w\), and note that the adjoint of \(L^{\pm v}\) is given by \(w \rightsquigarrow (\varphi(\widehat{X}) \pm v) w\).
Employing the polarization identity, the formula for the quadratic variation of Brownian integrals, c.f. Theorem 4.27 in \cite{deprato}, and Proposition A.2.2 in \cite{liu2015stochastic},
we obtain
\begin{align*}
\langle Z, \widehat{X} \bullet v\rangle^{\mathbf{P}}
&= Z_-/4 \cdot \big(\big\langle L^{+v} \cdot \widehat{X}, L^{+v}\cdot \widehat{X}\big\rangle^{\mathbf{P}} + \big\langle L^{-v} \cdot \widehat{X}, L^{-v} \cdot \widehat{X}\big\rangle^{\mathbf{P}}\big)
\\&=Z_-/4 \cdot \big( L^{+v} Q (\varphi(\widehat{X}) + v) \cdot I + L^{-v} Q (\varphi(\widehat{X}) - v)\cdot I\big)
\\&= \big(Z_- Q \varphi(\widehat{X}) \cdot I \big) \bullet v.
\end{align*}
Therefore we conclude \(\langle\hspace{-0.085cm}\langle Z, M(h)\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}} = Z_- Q \varphi(\widehat{X}) \cdot I\), which implies that
\begin{align*}
\frac{\ensuremath{1}_{\{Z_- > 0\}}}{Z_-} \cdot \langle\hspace{-0.085cm}\langle Z, M(h)\rangle\hspace{-0.085cm}\rangle^{\mathbf{P}} =
Q \varphi(\widehat{X}) \cdot I.
\end{align*}
Now we note that
\begin{align*}
\|Q\varphi(t, \omega)\|_\mathbb{H}^2 \leq \|Q^{1/2}\|_{\textup{HS}}^2 \|Q^{1/2}\varphi(t, \omega)\|_\mathbb{H}^2 \leq \lambda \|Q^{1/2}\|^2_{\textup{HS}} \left(1 + \sup_{s < t}\|\omega(s)\|^2_\mathbb{H}\right),
\end{align*}
for all \((t, \omega)\in \mathbb{R}^+ \times \mathbb{C}^\mathbb{H}_0\), where \(\|\cdot\|_\textup{HS}\) denotes the Hilbert-Schmidt norm.
Hence we may apply Proposition \ref{theorem existence uniqueness SMP} in Appendix \ref{SMPs and SDEs} with \(F(t, \omega) = Q\varphi(t, \omega)\) and \(G(t, \omega) = \operatorname{Id}\), which yields
that the SMP on \((\mathbb{C}^\mathbb{H}_0, \mathcal{C}^\mathbb{H}_0, \mathfrak{C}^\mathbb{H}_0)\) associated with \((\mathbb{H}; [\hspace{-0.06cm}[ 0, \infty[\hspace{-0.06cm}[; \eta; \widehat{X}; \dot{B}, C, 0)\)
has a solution and satisfies \(\mathcal{T}^*\)-uniqueness.
The continuity of \(\widehat{X}\) and
Lemma 74.2 in \cite{RW} imply that for all \(n, m \in \mathbb{N}\), \(\rho_n \wedge \tau_m = \rho_n \wedge m\) are positive \(\mathfrak{C}^{\mathbb{H}, 0}_0\)-stopping times, i.e. \(\rho_n \wedge \tau_m \in \mathcal{T}^*\).
This implies the claim.
\end{proof}
|
\section{Introduction}
The ridge regression~\cite{tikhonov:regression, tihonov_ridge_1963, hoerl_ridge_1970} has been introduced as a regularization method to estimate the unknown matrix $B$ in the linear regression model
$
\mathbf{y}= B ^\top \mathbf{x} + \boldsymbol{\varepsilon},
$
when the covariance matrix of the explanatory variables $\Sigma _{\bf x} := {\rm Cov} ( {\bf x} , {\bf x} ) \in \Bbb S _p $ is singular or poorly conditioned and, therefore, the standard least squares estimator is ill-defined. The ridge regularization method produces an estimate $ \widehat{B} _\lambda $ ($\lambda$ is a regularization strength that we introduce later on) whose properties are always formulated in terms of the coefficient $B$ that is assumed to solve the least squares problem. Nevertheless, there are many applications in biology, medicine, physics, engineering, or machine learning (see, for instance,~\cite{Geman1992}) in which an underlying linear model with a $B$ that solves the least squares problem does not exist and only a regularized version of it with parameter $B _\lambda $ is available. This is always the case in high-dimensional problems in which the number of covariates $p$ exceeds the sample size $N$ or whenever there is quasi-collinearity among some of the explanatory variables.
In this paper we place ourselves in that fully singular situation and nevertheless, we manage to establish the conditional distribution properties of the finite sample ridge regression estimator $ \widehat{B} _\lambda $ directly in terms of $B _\lambda $ without using the least squares coefficient $B$, whose existence is not needed. These results require that certain hypotheses on the conditional homoscedasticity of the regularized regression residuals hold.
The second main contribution of this paper consists of using the properties of this estimator to write down explicit expressions that evaluate the regression (also called training) and the generalization (also called testing) errors committed by a regularized regression model. We recall that the regression or training error is the one committed by the regularized regression model when calculated with the finite sample that has been used to obtain the estimate $\widehat{B} _\lambda $ of the regression coefficient ${B} _\lambda $; for the generalization or testing error we keep $\widehat{B} _\lambda $ and we compute the error committed by the corresponding regression model using another sample that may have different size or even different statistical properties. In both cases our error formulas incorporate the error committed at the time of parameter estimation using the finite training sample.
The paper is structured in two main sections. The first one is Section~\ref{Multivariate linear ridge regression and ridge estimator} that contains a description of our setup and the hypotheses that we invoke. It also presents the properties of the ridge estimator in the singular framework that we have chosen to work on. Section~\ref{Evaluation of the ridge regression errors: training and testing} contains the results on the evaluation of the training and testing errors. W e illustrate our developments with various classical examples that give an idea of the scope of our results. All the proofs of the results and the technical details are included in the appendices in Section~\ref{Appendices}.
\medskip
\noindent {\bf Notation and conventions:}
Column vectors are denoted by bold lower or upper case symbol like $\mathbf{v}$ or $\mathbf{V}$. We write $\mathbf{v} ^\top $ to indicate the transpose of $\mathbf{v} $. Given a vector $\mathbf{v} \in \mathbb{R} ^n $, we denote its entries by $v_i$, with $i \in \left\{ 1, \dots, n
\right\} $; we also write $\mathbf{v}=(v _i)_{i \in \left\{ 1, \dots, n\right\} }$.
The symbols $\mathbf{i} _n$ and $ \mathbf{0} _n $ stand for the vectors of length $n$ consisting of ones and zeros, respectively. We denote by $\mathbb{M}_{n , m }$ the space of real $n\times m$ matrices with $m, n \in \mathbb{N} $. When $n=m$, we use the symbol $\mathbb{M} _n $ to refer to the space of square matrices of order
$n$. Given a matrix $A \in \mathbb{M} _{n , m} $, we denote its components by $A _{ij} $ and we write $A=(A_{ij})$, with $i \in \left\{ 1, \dots, n\right\} $, $j \in \left\{ 1, \dots m\right\} $. If $A$ and $B$ are two matrices with the same number of rows, we denote by $(A|| B) $ the matrix resulting from their horizontal concatenation. We write $\mathbb{I} _n $ to denote the identity matrix of dimension $n$.
We use $\mathbb{S}_n $ to indicate the subspace $\mathbb{S} _n \subset \mathbb{M} _n $ of symmetric matrices, that is, $\mathbb{S} _n = \left\{ A \in \mathbb{M} _n \mid A ^\top = A\right\}$.
The symbol $||A||_{\rm Frob}$ denotes the Frobenius norm of $A \in \mathbb{M} _{m,n}$ defined as $\|A\|_{{\rm Frob}}^2:={\rm trace} \left(A^T A\right) $~\cite{Meyer:book:matrix}. The symbols ${\rm E}[\cdot]$ and ${\rm Cov}(\cdot,\cdot)$ denote the expectation and the covariance of random variables, respectively. Given a random variable $X$ the symbols ${\rm E}_X[\cdot]$ and ${\rm Cov}_X(\cdot,\cdot)$ denote the conditional expectation and the conditional covariance with respect to $X$, respectively. Given a random vector $\mathbf{x} $ and a random matrix $X$, we will use the symbols $\boldsymbol{\mu}_{\mathbf{x}} $ and $M_X$ to denote the mean of $\mathbf{x} $ and $X$, respectively. Let $Z$ be a $m$ by $n$ matrix random variable; the symbol $Z\sim {\rm MN} _{m,n}(M_Z, U_Z, V_Z)$ indicates that $Z$ is distributed according to the matrix valued normal distribution with mean matrix $M_Z\in \Bbb M _{m,n}$ and scale matrices $U_Z\in \Bbb S _{m}$ and $V_Z\in \Bbb S _{n}$ (see \cite{bookMatrixDistributions2000} for additional definitions and properties).
\medskip
\noindent {\bf Glossary of symbols.} $p$ is the number of explanatory variables (dimension of $\mathbf{x}$), $q$ is the number of dependent variables (dimension of $\mathbf{y}$), $N$ is the dimension of the (random) sample size, $\lambda$ is the regularization strength.
\noindent $\boldsymbol{\mu} _{\mathbf{x}}:= {\rm E}\left[\mathbf{x}\right]$, $\boldsymbol{\mu} _{\mathbf{y}}:= {\rm E}\left[\mathbf{y}\right]$, $\Sigma _{\bf x} := {\rm Cov} ( {\bf x} , {\bf x} ) \in \Bbb S _p $, $\Sigma _{\bf y} := {\rm Cov} ( {\bf y} , {\bf y} ) \in \Bbb S _q$, $\Sigma _{{\bf x}{\bf y}} := {\rm Cov} ( {\bf x} , {\bf y} ) \in \Bbb M _{p, q}$.
\noindent$B _\lambda := \left(\Sigma _{\bf x} + \lambda \mathbb{I}_p \right) ^{-1} \Sigma _{\bf xy}$ is the ridge regression matrix.
\noindent $\boldsymbol{\varepsilon} _\lambda := {\bf y} - B _\lambda^\top {\bf x} $ are the ridge regression residuals.
\noindent $\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda}: ={\rm E}\left[ \boldsymbol{\varepsilon} _\lambda \right] = \boldsymbol{\mu} _{\mathbf{y}} - B _\lambda ^\top \boldsymbol{\mu} _{\mathbf{x}} $, $\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} }: ={\rm E} _\mathbf{x}\left[ \boldsymbol{\varepsilon} _\lambda \right] = {\rm E}_\mathbf{x}\left[ \mathbf{y} \right] - B _\lambda ^\top \mathbf{x} $.
\noindent $\Sigma _{ {\boldsymbol{\varepsilon}} _\lambda} :={\rm Cov} \left( \boldsymbol{\varepsilon} _\lambda, \boldsymbol{\varepsilon} _\lambda\right)$, $\Sigma _{ {\boldsymbol{\varepsilon}} _\lambda|{\bf x}} :={\rm Cov} _{\bf x}\left( \boldsymbol{\varepsilon} _\lambda, \boldsymbol{\varepsilon} _\lambda\right).$
\noindent $X:=\left( \mathbf{x} _1 ||\mathbf{x} _2 || \dots ||\mathbf{x} _N \right) \in \mathbb{M} _{p, N}$, $Y:=\left( \mathbf{y} _1 ||\mathbf{y} _2 || \dots ||\mathbf{y} _N \right) \in \mathbb{M} _{q, N}$, and $E _\lambda:=\left( \boldsymbol{\varepsilon} _{\lambda ,1}|| \dots||\boldsymbol{\varepsilon} _{\lambda ,N} \right) \in \mathbb{M} _{q, N}$.
\noindent $A_N:= \mathbb{I} _N - \dfrac{1}{N} \mathbf{i} _N \mathbf{i} _N ^\top $.
\noindent$ \widehat{B} _\lambda :=(XA_NX^\top + \lambda N \mathbb{I}_p ) ^{-1} XA_NY^\top$ is the ridge regression matrix estimator.
\noindent $Z _\lambda := R _{\lambda} XA_N X ^\top=\mathbb{I}_p - \lambda N R _{\lambda}$.
\noindent $
R _{\lambda}:= (XA _N X ^\top + \lambda N \mathbb{I}_p ) ^{-1}.
$
\noindent For $XA _NX^\top$ invertible:
\noindent$ \widehat{B} :=\widehat{B}_0=(XA_NX^\top) ^{-1} XA_NY^\top$ is the ordinary least squares matrix estimator.
\noindent $\widehat{B} _\lambda = Z _\lambda \widehat{B}$.
\noindent$Z _\lambda := R _{\lambda} XA_N X ^\top=\mathbb{I}_p - \lambda N R _{\lambda}= \left( \mathbb{I} _p + \lambda N(XA _N X ^\top \right) ^{-1} ) ^{-1}$.
\medskip
\noindent {\bf Acknowledgments:} We acknowledge partial financial support of the French ANR ``BIPHOPROC" project (ANR-14-OHRI-0002-02).
\section{Multivariate ridge regressions and the ridge estimator}
\label{Multivariate linear ridge regression and ridge estimator}
\subsection{The setup}
\label{The covariates and the explained variables}
\medskip
Let $ {\bf x} $ and $ {\bf y} $ be respectively $\mathbb{R} ^p $ and $\mathbb{R} ^q $-valued random variables defined on a given probability space $( \Omega , \mathcal{F}, \mathbb{P} )$. All along this paper we work under the following assumption:
\medskip
\noindent {\bf (A1)} {\it The random variables $ {\bf x}, {\bf y}$ belong to $ {L}^2 ( \Omega , \mathcal{F}, \mathbb{P} )$, that is their first and second order moments exist and are finite, that is, ${\rm E} \left[ {\bf x} \right] < \infty$, ${\rm E} \left[ {\bf y} \right] < \infty$, ${\rm E} \left[ {\bf x}{\bf x}^\top \right] < \infty$, ${\rm E} \left[ {\bf y} {\bf y}^\top \right] < \infty$. Additionally, the Cauchy-Schwarz inequality implies the finiteness of the covariance between $ {\bf x}$ and $ {\bf y}$, that is, ${\rm Cov} \left( {\bf x},{\bf y} \right) < \infty$.}
\medskip
\noindent In the sequel we denote $\boldsymbol{\mu} _{\bf x}:={\rm E} \left[ {\bf x} \right] \in \mathbb{R} ^p $, $\boldsymbol{\mu} _{\bf y}:={\rm E} \left[ {\bf y} \right] \in\mathbb{R} ^q $ and use the following notation for the second order central moments: $\Sigma _{\bf x} := {\rm Cov} ( {\bf x} , {\bf x} ) \in \Bbb S _p $, $\Sigma _{\bf y} := {\rm Cov} ( {\bf y} , {\bf y} ) \in \Bbb S _q$, and $\Sigma _{{\bf x}{\bf y}} := {\rm Cov} ( {\bf x} , {\bf y} ) \in \Bbb M _{p, q}$.
\medskip
\noindent {\bf The multivariate ridge regression optimization problem.} Consider the ridge penalized least squares problem and define:
\begin{align}
\label{ridge optimization problem statement}
{ B}_{\lambda }&:=\mathop{\rm arg\, min}_{{ B} \in \mathbb{M}_{p,q} } \left({\rm trace} \left({\rm E} \left[ ({ B} ^{ \top} {\bf x} - {\bf y} ) ({ B} ^{ \top} {\bf x} -{\bf y} )^\top \right] \right) + \lambda \|{ B} \|^2_{\rm Frob}\right),
\end{align}
where $ \lambda \in \mathbb{R} ^+$ is regularization strength and $||\cdot||^2_{\rm Frob}$ denotes the squared Frobenius norm of the coefficient matrix $B$. The optimization problem \eqref{ridge optimization problem statement} does not contain an intercept which does not imply any loss of generality since its presence can be accommodated by replacing $B \in \mathbb{M} _{p,q}$ and $\mathbf{x} \in \mathbb{R} ^p $ in~\eqref{ridge optimization problem statement} by $\tilde{B} \in \mathbb{M} _{p + 1,q}$ and $\tilde{\mathbf{x}} \in \mathbb{R} ^{p+1} $, respectively, that are constructed by vertical concatenation of the intercept vector ${\bf b}_0^\top \in \mathbb{R} ^q $ with $B$ and of $1$ with $\mathbf{x} \in \mathbb{R} ^p $, respectively. In this case the penalization summand in \eqref{ridge optimization problem statement} has to be replaced by $\lambda \|{ \tilde{B}} \|^{2 ^\ast }_{\rm Frob}$ where $|| \tilde{B}||^{2 ^\ast }_{\rm Frob}$ would denote the squared Frobenius norm of the $p \times q$ lower submatrix of $\tilde{B}$.
In the presence of the assumption {\bf (A1)} the multivariate ridge regression problem \eqref{ridge optimization problem statement} has a unique solution $B_\lambda \in \mathbb{M} _{p,q}$ given by
\begin{equation}
\label{B lambda}
B _\lambda := \left(\Sigma _{\bf x} + \lambda \mathbb{I}_p \right) ^{-1} \Sigma _{\bf xy}.
\end{equation}
We will refer to $B _\lambda $ as the ridge regression matrix. We emphasize that $B _\lambda$ always exists even if the covariance matrix $\Sigma _{\bf x}$ is singular or ill-conditioned, provided that $\lambda > 0 $.
\medskip
\noindent {\bf The regression residuals.} Given the random variables $ {\bf x}, {\bf y}$ that satisfy the hypothesis {\bf (A1)}, we can define, for each $\lambda \in \mathbb{R}^+ $, the regression residuals using the corresponding and uniquely determined ridge regression matrix $B _\lambda $ in~\eqref{B lambda} as:
\begin{equation}
\label{regression residuals}
\boldsymbol{\varepsilon} _\lambda := {\bf y} - B _\lambda^\top {\bf x}.
\end{equation}
The residuals $\boldsymbol{\varepsilon} _\lambda \in \mathbb{R} ^q $ in \eqref{regression residuals} are a $\mathbb{R}^q$-valued distributed random variable with mean $\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda } \in \mathbb{R} ^q $ and covariance matrix $\Sigma _{ {\boldsymbol{\varepsilon}} _\lambda} \in \Bbb S ^q $, that is,
\begin{equation}
\label{distr eps}
\boldsymbol{\varepsilon} _\lambda \sim D( \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda }, \Sigma _{ {\boldsymbol{\varepsilon}} _\lambda}),\ \mbox{with} \ \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda } :=\boldsymbol{\mu} _{\bf y} - B _\lambda ^\top \boldsymbol{\mu} _{{\bf x}},\ \Sigma _{ {\boldsymbol{\varepsilon}} _\lambda}:=\Sigma_{\bf y} - B _\lambda ^\top \Sigma _{\bf xy} - \Sigma _{\bf xy} B _\lambda + B _\lambda ^\top (\Sigma _{\bf x}+ \boldsymbol{\mu} _{{\bf x}} \boldsymbol{\mu} _{{\bf x}}^\top ) B _\lambda.
\end{equation}
In the following sections we will also use the conditional expectation and variance of the residuals $\boldsymbol{\varepsilon} _\lambda \in \mathbb{R} ^q $ with respect to $\mathbf{x}$, that we denote by
\begin{equation}
\label{mu eps lambda cond x}
\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} }: ={\rm E} _\mathbf{x}\left[ \boldsymbol{\varepsilon} _\lambda \right] = {\rm E}_\mathbf{x}\left[ \mathbf{y} \right] - B _\lambda ^\top \mathbf{x} \quad \mbox{and} \quad \Sigma _{ {\boldsymbol{\varepsilon}} _\lambda|{\bf x}} :={\rm Cov} _{\bf x}\left( \boldsymbol{\varepsilon} _\lambda, \boldsymbol{\varepsilon} _\lambda\right),
\end{equation}
respectively. Recall that by the laws of total expectation and variance, these moments and conditional moments are related by
\begin{equation*}
\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda } = {\rm E} \left[\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} }\right] \quad \mbox{and} \quad
\Sigma _{ {\boldsymbol{\varepsilon}} _\lambda}={\rm E}\left[\Sigma _{ {\boldsymbol{\varepsilon}} _\lambda|{\bf x}} \right]+
{\rm Cov} \left(\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} },\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} }\right).
\end{equation*}
We emphasize that in our treatment of the ridge regression problem we do not assume, as it is customary in the literature, that there is an underlying stochastically perturbed linear functional relation between $\mathbf{x} $ and $ {\bf y} $. The aim of the error formulas that we present later on in Section~\ref{Evaluation of the ridge regression errors: training and testing} is indeed the quantitative evaluation of the inaccuracy committed when the actual functional link between those two random variables is approximated by a linear model obtained via the solution of the regularized regression problem~\eqref{ridge optimization problem statement}. More explicitly, we proceed by first presenting the two random variables $\mathbf{x} $ and $ {\bf y} $ that satisfy hypothesis {\bf (A1)} and by solving, for a fixed penalization strength $\lambda $, the ridge penalized least squares problem~\eqref{ridge optimization problem statement}; as we see later on, this strategy allows for the definition of the regression residuals ${\boldsymbol{\varepsilon}} _\lambda$ and, when they satisfy certain homoscedasticity conditions, it can be used to characterize the statistical properties of the ridge estimator in fully singular situations ($\Sigma _{\bf x} $ is not invertible) and to evaluate the regression (also called training) and the generalization (also called testing) errors.
\subsection{The finite sample ridge estimator, the residuals, and the homoscedasticity hypothesis}
\label{The ridge estimator and its properties}
We now study the finite sample properties of the ridge estimator. We start by considering two random samples of size $N$, that is, $\left\{ \mathbf{x} _i\right\}_{i \in \left\{ 1, \dots, N \right\}} $ and $\left\{ \mathbf{y} _i\right\} _{i \in \left\{ 1, \dots, N \right\}}$. These samples are constituted by $N$ different $ \mathbb{R} ^p $-valued (respectively, $ \mathbb{R} ^q $-valued), mutually independent, and identically distributed random variables $\mathbf{x} _i \in {L}^2( \Omega , \mathcal{F}, \mathbb{P} )$ (respectively, $\mathbf{y} _i \in {L}^2( \Omega , \mathcal{F}, \mathbb{P} )$), $i \in \left\{ 1, \dots, N \right\} $. Notice that we require that all these random variables satisfy the hypothesis {\bf(A1)}. We now horizontally concatenate the elements of each of these two random samples and construct the corresponding random matrices $X \in \mathbb{M} _{p, N}$ and $Y \in \mathbb{M} _{q, N}$, respectively, that is $X:=\left( \mathbf{x} _1 ||\mathbf{x} _2 || \dots ||\mathbf{x} _N \right) $ and $Y:=\left( \mathbf{y} _1 ||\mathbf{y} _2 || \dots ||\mathbf{y} _N \right) $.
The finite sample estimator $\widehat{B}_ \lambda $ of the ridge regression matrix $B _\lambda $ is constructed by plugging in \eqref{B lambda} the natural estimators for the second order moments $\Sigma _{\bf x}$ and $\Sigma _{\bf xy}$, that is,
\begin{align}
\label{Sigma x}
\widehat{\Sigma }_{\bf x} &:= \dfrac{1}{N} XA_NX^\top,\\
\widehat{\Sigma }_{\bf xy} &:= \dfrac{1}{N} XA_NY^\top, \label{Sigma xy}
\end{align}
with \begin{equation}
\label{A N}
A_N:= \mathbb{I} _N - \dfrac{1}{N} \mathbf{i} _N \mathbf{i} _N ^\top.
\end{equation}
The expressions \eqref{Sigma x}-\eqref{Sigma xy} substituted in \eqref{B lambda} yield
\begin{equation}
\label{B lambda random sample est}
\widehat{B} _\lambda :=(XA_NX^\top + \lambda N \mathbb{I}_p ) ^{-1} XA_NY^\top,
\end{equation}
which can be subsequently evaluated for any given realization $\mathcal{X}$ and $\mathcal{Y}$ of the random matrices $X$ and $Y$.
Consider now the $\mathbb{R} ^q $-valued random residuals $\left\{ \boldsymbol{\varepsilon} _{\lambda ,i} \right\} _{i \in \left\{ 1, \dots, N \right\} } $ obtained using the random samples $\left\{ \mathbf{x} _{i} \right\} _{i \in \left\{ 1, \dots, N \right\} } $ and $\left\{ \mathbf{y} _i \right\} _{i \in \left\{ 1, \dots, N \right\} } $, for each $i \in \left\{ 1, \dots, N \right\} $ via the assignment $\boldsymbol{\varepsilon} _{ \lambda , i} := \mathbf{y} _i - B _\lambda \mathbf{x}_i $. Similarly to the procedure used in the construction of the random matrices $X$ and $Y$ we horizontally concatenate the entries $\boldsymbol{\varepsilon} _{ \lambda ,i} \in\mathbb{R} ^q $, $i \in \left\{ 1, \dots, N \right\} $, and we obtain the random matrix $E _\lambda \in \mathbb{M} _{q, N}$ given by $E _\lambda:=\left( \boldsymbol{\varepsilon} _{\lambda ,1}|| \dots||\boldsymbol{\varepsilon} _{\lambda ,N} \right) $ and that satisfies the relation $E _\lambda = Y - B_\lambda ^\top X$. We denote the unconditional first moment of $E _\lambda$ by $M_{ E _\lambda}\in \mathbb{M} _{q,N}$ and note that
\begin{equation}
\label{mu E lambda}
M_{ E _\lambda }:={\rm E} \left[ E _\lambda \right] = \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_{\lambda} } \mathbf{i} _N^\top,
\end{equation}
where $ \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_{\lambda} }$ is given by \eqref{distr eps}. We analogously specify the conditional expectation of $E _\lambda$ given a random matrix $X$ as
\begin{equation}
\label{mu E lambda cond X}
M_{ E _\lambda | X}:={\rm E}_X \left[ E _\lambda \right] = (\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_{\lambda, 1} | { \mathbf{x} _1 } }|| \dots|| \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_{\lambda, N} | { \mathbf{x} _N } }),
\end{equation}
where each $\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_{\lambda, i} | { \mathbf{x} _i } }$, $i \in \left\{ 1, \dots, N \right\} $ is given by \eqref{mu eps lambda cond x}.
The relations that we just provided for the conditional and unconditional first order moments of the residuals random matrix $E _\lambda $ can be used to establish the expressions of their corresponding second-order counterparts. In general, one needs to use \eqref{distr eps} and \eqref{mu eps lambda cond x} in order to determine the second order moments of the residuals $\left\{ \boldsymbol{\varepsilon} _{\lambda ,i} \right\} _{i \in \left\{ 1, \dots, N \right\} } $ and this implies that, in principle, there exist $N$ different conditional covariance matrices $ \Sigma _{ \boldsymbol{\varepsilon} _{ \lambda, i} | \mathbf{x} _i }$, $i \in \left\{ 1, \dots, N \right\} $, defined for each $\boldsymbol{\varepsilon} _{ \lambda , i}$ and given some corresponding random sample element $ \mathbf{x}_i $. We consider in what follows a particular situation in which all these second order conditional moments are constant and equal for all $i \in \left\{ 1, \dots, N \right\} $. As we discuss later on in the text, this conditional homoscedasticity property is natural and is satisfied in many standard situations. In particular, we will show that linear and nonlinear models with additive noise (examples A and B) have conditionally homoscedastic residuals unlike the case with multiplicative noise (Example C) that in general does not satisfy this property.
\begin{lemma}
\label{Lemma 2}
In the conditions that were just introduced consider the ridge residuals $\left\{ \boldsymbol{\varepsilon} _{\lambda ,i} \right\} _{i \in \left\{ 1, \dots, N \right\} } \sim D( \boldsymbol{\mu}_{ \boldsymbol{\varepsilon} _\lambda}, \Sigma _{ \boldsymbol{\varepsilon} _ \lambda})$ and suppose that they are homoscedastic, that is, we assume that the conditional covariance matrices $ \Sigma _{ \boldsymbol{\varepsilon} _{ \lambda, i} | \mathbf{x} _i }$, $i \in \left\{ 1, \dots, N \right\} $, in \eqref{sigma lambda x} are constant and equal to some matrix $\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda \in \Bbb S^+ _q $.
Then the following relations hold true:
\begin{description}
\item [{\bf (i)}] For the conditional second-order raw moments:
\begin{align}
\label{E Elambda tr Elambda cond X}
{\rm E}_X \left[ E _\lambda ^\top E _\lambda \right] - M_{ E _\lambda | X} ^\top M_{ E _\lambda | X} &= {\rm trace}(\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda) \mathbb{I} _N,\\
{\rm E} _X\left[ E _\lambda E _\lambda^\top \right] -M_{ E _\lambda | X} M_{ E _\lambda | X}^\top&=N \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda, \label{E Elambda Elambda tr cond X}
\end{align}
\item [{\bf (ii)}] For the unconditional second-order raw moments:
\begin{align}
\label{E Elambda tr Elambda}
{\rm E} \left[ E _\lambda ^\top E _\lambda \right] - M_{E_\lambda}^\top M_{E_\lambda} &= {\rm trace}(\Sigma _{ \boldsymbol{\varepsilon} _\lambda } ) \mathbb{I} _N,\\
{\rm E} \left[ E _\lambda E _\lambda^\top \right] - M_{E_\lambda} M_{E_\lambda} ^\top&=N \Sigma _{ \boldsymbol{\varepsilon} _\lambda } .\label{E Elambda Elambda}
\end{align}
\end{description}
In these relations, the mean $M_{ E _\lambda} $ and conditional mean $M_{ E _\lambda | X} $ matrices are given by~\eqref{mu E lambda} and~\eqref{mu E lambda cond X}, respectively.
\end{lemma}
The proof of this lemma is not provided since it is straightforward. Lemma \ref{Lemma 2} has an important implication under an additional normality assumption for the ridge residuals that we state in the following corollary.
\begin{corollary} \label{Corollary Lemma 2}
Suppose that the hypotheses of Lemma \ref{Lemma 2} are satisfied and that, additionally, the conditional residuals are independent and normally distributed, that is, $\boldsymbol{\varepsilon} _{\lambda, i} | \mathbf{x}_i \sim {\rm IN} ( \boldsymbol{\mu} _{ \boldsymbol{\varepsilon} _{\lambda, i}| \mathbf{x} _i }, \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda)$, for each $i \in \left\{ 1, \dots, N\right\} $. Then, the random matrix of residuals $E _{\lambda}$ conditional on $X$ are distributed according to a matrix normal distribution with conditional mean $M_{ E _\lambda | X}$, and $\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda$ and $\mathbb{I}_N$ as scale matrices, that is,
\begin{equation}
\label{E lambda cond X distribution}
E _{\lambda} |X \sim {\rm MN} ( M_{ E _\lambda | X}, \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda, \mathbb{I}_N ).
\end{equation}
\end{corollary}
\subsection{The properties of the ridge regression estimator}
\label{The properties of the ridge regression estimator}
The results presented in Lemma \ref{Lemma 2} and its Corollary \ref{Corollary Lemma 2} allow us to take an important step. Indeed, we now see how, in the presence of the conditional normality and homoscedasticity hypotheses on the ridge residuals that we introduced above, we can spell out the properties of the ridge regression matrix estimator $\widehat{B} _\lambda$ introduced in~\eqref{B lambda random sample est}, conditional on the covariates sample $X$. As we already pointed out in the introduction, our results generalize the classical ones in \cite{hoerl_ridge_1970} by formulating the properties of $\widehat{B} _\lambda$ directly in terms of the solution $B _\lambda $ of the ridge regularized regression problem and without assuming the existence of a least squares matrix coefficient $B$ or, equivalently, the regularity of the covariance matrix ${\Sigma }_{ \mathbf{x} }$ of the covariates, that in many applications is just not available.
\begin{theorem}
\label{Theorem 1}
Consider the random samples $\left\{ \mathbf{x} _i \right\} _{i \in \left\{ 1, \dots, N \right\} } $ and $\left\{ \mathbf{y} _i \right\} _{i \in \left\{ 1 ,\dots, N \right\} } $ of length $N$ that consist of the random variables $\mathbf{x} _i \in \mathbb{R} ^p $, $\mathbf{y} _i \in \mathbb{R} ^q $, $i \in \left\{ 1, \dots, N \right\} $, respectively, that satisfy the hypothesis {\bf (A1)}. Let $X\in \mathbb{R} _{p, N}$ and $Y \in \mathbb{M} _{q, N}$ be the random matrices constructed by horizontal concatenation of all the corresponding entries of $\left\{ \mathbf{x} _i \right\} _{i \in \left\{ 1, \dots, N \right\} } $ and $\left\{ \mathbf{y} _i \right\} _{i \in \left\{ 1 ,\dots, N \right\} } $, respectively. Additionally, let $\left\{ \boldsymbol{\varepsilon} _{ \lambda , i} \right\} _{i \in \left\{ 1, \dots, N \right\} }$ be the associated residuals of the ridge regression with regularization strength $ \lambda \in \mathbb{R} ^+$ defined by $\boldsymbol{\varepsilon} _{ \lambda , i} := \mathbf{y} _i - B _\lambda \mathbf{x}_i $ and that we assume to be conditionally homoscedastic. Let $E _\lambda $ be the random matrix obtained by horizontal concatenation of the ridge residuals conditioned on $X$ and suppose that it is normally distributed as in \eqref{E lambda cond X distribution} with conditional mean $M_{ E _\lambda | X} $ and some constant scale matrix $\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda$. Finally, let $\widehat{B} _\lambda $ be the ridge estimator of the ridge regression matrix $B _\lambda $ based on the random sample matrices $X$ and $Y$. Then
\begin{equation}
\label{B lambda hat minus B lambda}
(\widehat{B} _\lambda - B_\lambda )|X \sim {\rm MN} (-\lambda N R _\lambda B_\lambda + R_\lambda X A _N M_{ E _\lambda | X} ^\top, Z _\lambda R_\lambda ,\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda),
\end{equation}
where the conditional mean $M_{ E _\lambda | X}$ is given in \eqref{mu E lambda cond X}, the conditional covariance matrix $\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda$ is provided in point {\bf (i)} of Lemma \ref{Lemma 2} and, additionally,
\begin{align}
\label{R lambda repeat}
R _\lambda &:=(XA _N X ^\top +\lambda N \mathbb{I}_p )^{-1},\\
Z _\lambda &:=R _\lambda X A_N X ^\top = \mathbb{I} _p - \lambda N R _\lambda \label{Z lambda B lambda}
\end{align}
with $A_N$ as in \eqref{A N}.
\end{theorem}
\subsection{Examples}
We now illustrate the different elements and developments in the paper, as well as the reach of the hypotheses that are invoked, with three examples. Example A contains the standard linear multivariate regression model under the Gauss-Markov hypotheses; we will show in the context of this example that the classical results in~\cite{hoerl_ridge_1970} can be obtained as a particular case of Theorem~\ref{Theorem 1}. Examples~B and C consider fully nonlinear functional relations in the generation of $ {\bf y} $ that are subjected to additive and multiplicative stochastic disturbances, respectively. A particular case of Example~B will be worked out in Section~\ref{Numerical illustration} in order to illustrate some of the consequences of the error formulas in Section~\ref{Evaluation of the ridge regression errors: training and testing}.
\medskip
\noindent $\blacktriangleright$ {\bf Example~A. Linear multivariate regression model}. Consider a multivariate linear regression model of the form
\begin{equation}
\label{regression model}
\mathbf{y} = B^\top \mathbf{x} + \boldsymbol{\varepsilon},
\end{equation}
where $\mathbf{x} $ and $\mathbf{y} $ are $ \mathbb{R} ^p $ and $\mathbb{R} ^q $-valued random variables, respectively, subjected to the hypothesis {\bf (A1)}. The term $\boldsymbol{\varepsilon}$ is a $\mathbb{R} ^q $-valued random variable with zero mean and covariance matrix $\Sigma _{\boldsymbol{\varepsilon} }$, that is, $\boldsymbol{\varepsilon} \sim {\rm N}( {\bf 0} _q , \Sigma _{\boldsymbol{\varepsilon} })$. We place ourselves in a Gauss-Markov setup by assuming that $\boldsymbol{\varepsilon}$ and $\mathbf{x} $ are independent random variables.
In these conditions it is easy to show that for any $ \lambda \in \mathbb{R} ^+$, the ridge regression matrix $B _\lambda $ in \eqref{B lambda} and the linear regression coefficient matrix $B$ in \eqref{regression model} are related by
\begin{equation}
\label{B lambda on B}
B_\lambda = ( \Sigma _{\bf x} +\lambda \mathbb{I}_p )^{-1} \Sigma _{\bf x} B, \enspace {\rm or,} \enspace {\rm equivalently,} \enspace B_\lambda - B = - \lambda ( \Sigma _{\bf x} +\lambda \mathbb{I}_p )^{-1}B,
\end{equation}
where we used in \eqref{B lambda} that $\Sigma _{\bf xy} = \Sigma _{\bf x} B$. Since the ridge residuals are determined by the equality $\mathbf{y}= B _\lambda ^\top \mathbf{x} + \boldsymbol{\varepsilon} _\lambda $ and, at the same time, $\mathbf{x} $ and $\mathbf{y} $ are related by~\eqref{regression model}, we have that
\begin{equation*}
\boldsymbol{\varepsilon} _\lambda := \mathbf{y} - B _\lambda ^\top \mathbf{x} = (B - B _\lambda )^\top \mathbf{x} + \boldsymbol{\varepsilon}.
\end{equation*}
The unconditional first and the second moments of the ridge residuals $\boldsymbol{\varepsilon} _\lambda$ can be obtained in a straightforward way out of the relations in \eqref{distr eps} as follows:
\begin{align}
\label{mu eps lambda example}
\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda }&= (B - B _\lambda ) ^\top \boldsymbol{\mu}_{\bf x},\\
\label{sigma lambda example}
\Sigma _{ \boldsymbol{\varepsilon} _\lambda} &= \Sigma _ {\boldsymbol{\varepsilon} } + (B-B_\lambda )^\top \Sigma _{\bf x} (B - B _\lambda ),
\end{align}
where we used that $\Sigma _{\bf y} = \Sigma _ { \boldsymbol{\varepsilon} } + B ^\top \Sigma _{\bf x} B$ and that $\Sigma _{\bf yx} =B^\top \Sigma _{\rm x} $. The conditional counterparts of \eqref{mu eps lambda example}-\eqref{sigma lambda example} can be obtained with the help of \eqref{mu eps lambda cond x}:
\begin{align}
\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x}} &= (B-B _\lambda ) ^\top \mathbf{x},\label{mu eps lambda cond x example}\\
\label{sigma lambda x}
\Sigma _{ \boldsymbol{\varepsilon} _\lambda|{\bf x}}&={\rm E} _{ \mathbf{x} }\left[ \boldsymbol{\varepsilon} _\lambda \boldsymbol{\varepsilon} _{\lambda}^ \top \right] - \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x}} \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x}} ^\top = \Sigma _ {\boldsymbol{\varepsilon}}.
\end{align}
Notice that this expression shows that the ridge regression residuals of the linear regression model are conditionally homoscedastic.
We now derive the particular expression of the ridge regression matrix estimator in the conditions implied by~\eqref{regression model}. We first recall that the ordinary least squares estimator $\widehat{B}$ of $B$ is given by \eqref{B lambda random sample est} with zero regularization strength, that is $\widehat{B}=\widehat{B}_0$. More explicitly,
\begin{equation}
\label{B lambda0 random sample est}
\widehat{B} := \widehat{B} _0 =\widehat{\Sigma} _{\bf x} ^{-1} \widehat{\Sigma} _{\bf xy}=(XA_NX^\top) ^{-1} XA_NY^\top.
\end{equation}
It is important to underline that the ordinary least squares estimator \eqref{B lambda0 random sample est} is available only when the sample covariance matrix $\widehat{\Sigma} _{\bf x} $ in \eqref{Sigma x} is invertible which, as we already pointed out in the introduction, is one of its main deficiencies.
The following lemma, whose proof is in Appendix \ref{Proof of Lemma 1}, provides the relation between the ridge regression $\widehat{B} _\lambda$ and the ordinary least squares estimator $\widehat{B}$. We obviously restrict ourselves to the case in which the evaluation of \eqref{B lambda0 random sample est} is feasible, that is, when $\widehat{\Sigma} _{\bf x} $ is invertible. The relations that we now state are not new and are well established in the literature~\cite{hoerl_ridge_1970}. We provide them for future reference and in order to complete Example A.
\begin{lemma}
\label{Lemma 1}
In the conditions of {Example~A} the following statements hold true:
\begin{description}
\item[{\bf (i)}] The relation between the finite sample ridge estimator $\widehat{B} _\lambda $ of $B _\lambda $ and the finite sample ordinary least squares estimator $\widehat{B}$ of $B$ is given by
\begin{equation}
\label{B lambda ZB}
\widehat{B} _\lambda = Z _\lambda \widehat{B}
\end{equation}
where\begin{align}
\label{Z lambda}
Z _\lambda &:= R _{\lambda} XA_N X ^\top,\\
\label{R lambda}
R _{\lambda}&:= (XA _N X ^\top + \lambda N \mathbb{I}_p ) ^{-1}
\end{align}
with $A_N$ as in \eqref{A N}.
Expression \eqref{B lambda ZB} is the finite sample based analogue of the first relation in \eqref{B lambda on B}.
\item[{\bf (ii)}] The multiplier $Z _\lambda $ in \eqref{Z lambda} can be equivalently expressed as
\begin{equation}
\label{Z1}
Z _\lambda = \mathbb{I}_p - \lambda N R _{\lambda},
\end{equation}
or
\begin{equation}
Z _\lambda = \left( \mathbb{I} _p + \lambda N(XA _N X ^\top \right) ^{-1} ) ^{-1}, \label{Z2}
\end{equation}
with $R _\lambda $ in \eqref{Z1} defined as in \eqref{R lambda}.
\end{description}
\end{lemma}
\begin{remark}
\normalfont
The relations $Z _\lambda = \mathbb{I}_p - \lambda N R _{\lambda}$ in~\eqref{Z1} and $Z _\lambda = R _{\lambda} XA_N X ^\top$ in~\eqref{Z lambda} hold in a general context beyond the specific conditions of Example A and even if $XX^\top$ is singular or ill-conditioned. They are used later on in the paper.
\end{remark}
We now show how the classical results on the ridge estimator properties in Section 4a of \cite{hoerl_ridge_1970} that assume the existence of a least squares estimate follow as a corollary of Theorem \ref{Theorem 1}. The proof of the following result is provided in Appendix \ref{Proof of Corollary of Theorem 1}.
\begin{corollary}[Section 4a in \cite{hoerl_ridge_1970}]
\label{Corollary Theorem 1}
Consider the $\mathbb{R}^q$-valued linear regression $Y = B ^\top X + E $ obtained out of random samples of length $N$ that satisfy the relation~\eqref{regression model}. Given that, by hypothesis, the error term satisfies $\boldsymbol{\varepsilon} \sim {\rm N}( {\bf 0} _q , \Sigma _{\boldsymbol{\varepsilon} })$, it is easy to see that the residual random matrix $E\in \mathbb{M} _{q, N}$ obtained by horizontal concatenation, is matrix normal distributed as $E \sim {\rm MN} \left( \mathbb{O} _q , \Sigma _{\boldsymbol{\varepsilon}}, \mathbb{I} _N \right) $. The following statements hold:
\begin{description}
\item[{\bf (i)}] The ridge regression matrix estimator is conditionally distributed as
\begin{equation}
\label{B lambda minus B our notation}
(\widehat{B} _\lambda - B)|X \sim {\rm MN} ( - \lambda N R _\lambda B, Z_\lambda R_\lambda, \Sigma _{ \boldsymbol{\varepsilon} }),
\end{equation}
where $R _\lambda $ and $Z _\lambda $ are defined in \eqref{R lambda repeat} and in \eqref{Z lambda B lambda}, respectively.
\item[{\bf (ii)}] If the matrix $\widehat{\Sigma}_{\mathbf{x} }$ defined in \eqref{Sigma x} is invertible, then \eqref{B lambda minus B our notation} can be rewritten as
\begin{equation}
\label{B lambda minus B}
(\widehat{B} _\lambda - B)|X \sim {\rm MN} ( - \lambda N R _\lambda B, Z_\lambda (X A _N X ^\top) ^{-1} Z_\lambda , \Sigma _{ \boldsymbol {\varepsilon} }),
\end{equation}
with $Z _\lambda $ defined as in
\eqref{Z lambda B lambda} or, equivalently, as in \eqref{Z2}. $\blacktriangleleft $
\end{description}
\end{corollary}
\medskip
\noindent $\blacktriangleright$ {\bf Example~B. Nonlinear one-dimensional model with additive stochastic disturbance term}. Consider the following data generating process for the variable ${y} $
\begin{equation}
\label{exampleB model}
{y} = f(\xi ) + {\varepsilon},
\end{equation}
where $f : \mathbb{R} \rightarrow \mathbb{R} $ is a continuous function, ${\varepsilon} \sim D_ \varepsilon ( 0 , \sigma_{ \varepsilon }^2 )$ represents a disturbance term and $\xi $ is distributed with zero mean, variance $ \sigma ^2 _\xi$, and probability density function $g _{ \xi }$. The random variables $\varepsilon $ and $\xi $ are assumed to be independent and subjected to {\bf (A1)}.
We now approximate the model \eqref{exampleB model} using a univariate ridge regression in which the $p$ explanatory variables are the first $p$ elements $\left\{ \phi _i\right\}_{i \in \left\{ 1, \dots, p \right\}}$ of a given ordered generating set of the function space to which the function $f$ in~\eqref{exampleB model} belongs, all of them evaluated at $\xi $. More explicitly, consider an approximate model of the following form:
\begin{equation}
\label{approximated model example B}
y = {\bf B} _\lambda ^\top \mathbf{x} + \varepsilon _\lambda, \enspace \varepsilon_\lambda \sim D_{ \varepsilon _\lambda }( \mu _{ \varepsilon_\lambda }, \sigma _{ \varepsilon_\lambda }^2),
\end{equation}
where the vector of regressors $\mathbf{x} \in \mathbb{R} ^p $ is constructed as
\begin{equation}
\label{x example B}
\mathbf{x} := \left( \widetilde{\xi_1}, \dots, \widetilde{ \xi _p} \right) ^\top, \enspace {\rm with }\enspace x _i = \widetilde{ \xi _i} := \dfrac{ \xi _i - {\rm E} \left[ \xi _i\right] }{ \sigma (\xi _i)}, \enspace \xi _i = \phi _i ( \xi ), \enspace i \in \left\{ 1, \dots, p\right\}.
\end{equation}
We emphasize that for each $i \in \left\{ 1, \dots, p\right\} $, the explanatory variable $x _i $ is constructed out of the standardized version of the evaluation $\xi _i = \phi _i ( \xi ) $ of the generating function associated $\phi _i$ using the standard deviation:
\begin{equation}
\label{std}
\sigma (\xi _i):= ( {\rm E} \left[ \xi _i^ {2}\right] - {\rm E} \left[ \xi _ {i}\right]^2)^{1/2}, \enspace {\rm for} \enspace {\rm each} \enspace i\in \left\{ 1, \dots, p \right\}.
\end{equation}
We now provide all the elements associated to the approximate model~\eqref{approximated model example B}. First,~\eqref{B lambda} determines, for any regularization strength $\lambda \in \mathbb{R} ^+$, the ridge regression matrix $B _\lambda $ once the central moments $\Sigma _{\bf x} \in \Bbb S _p $ and $ {\mathbf{\Sigma}} _{{\bf x}y} \in \mathbb{R}^p $ have been computed. It is easy to see that
\begin{align}
\label{sigma x example B}
(\Sigma _{\bf x})_{i,j}=&{\rm E}\left[\widetilde{ \xi _i}\widetilde{ \xi _j} \right] = \dfrac{1}{\sigma (\xi _i)\sigma (\xi _j)}\int_{-\infty}^{\infty} \left( \phi _i (z) - {\rm E} \left[ \xi _i\right]\right) \left( \phi _j (z) - {\rm E} \left[ \xi _j\right]\right) g_{\xi }(z) dz, \enspace i, j\in \left\{ 1, \dots, p \right\}, \\
(\mathbf{\Sigma} _{{\bf x}y})_{i}=&{\rm E}\left[\widetilde{ \xi _i}f(\xi) \right] = \dfrac{1}{\sigma (\xi _i)}\int_{-\infty}^{\infty} \left( \phi _i (z) - {\rm E} \left[ \xi _i\right]\right) f(z) g_{\xi }(z) dz, \enspace i\in \left\{ 1, \dots, p \right\}, \label{sigma xy example B}
\end{align}
with $\sigma ( \xi_i)$ as in \eqref{std} and where $g_{ \xi }$ is the probability density function of $ \xi $.
We now study the defining features of the ridge regression residuals $ \varepsilon _\lambda $ and their moments. First, by \eqref{approximated model example B} and \eqref{exampleB model}, the residuals $ \varepsilon _\lambda$ are given by
\begin{equation*}
\varepsilon _\lambda = f( \xi ) - {\bf B}_\lambda ^\top \mathbf{x} + \varepsilon.
\end{equation*}
This expression can be used to explicitly compute the unconditional and conditional first and second moments of $ \varepsilon _\lambda $; indeed, by~\eqref{distr eps}:
\begin{align}
\label{mu eps lambda example B}
\mu _{\varepsilon _\lambda } &= {\rm E} \left[ f( \xi )\right] = \int _{-\infty}^{\infty} f(z) g_{ \xi } (z) dz,
\end{align}
where we used that ${\rm E}\left[ \mathbf{x} \right] =0$ by construction in \eqref{x example B}. Analogously, \eqref{distr eps} yields
\begin{align}
\sigma ^2 _{\varepsilon_\lambda } &= \sigma ^2 _{y} - {\bf B}_\lambda ^\top \mathbf{ \Sigma }_{\mathbf{x} y} - \mathbf{ \Sigma }_{\mathbf{x} y} ^\top {\bf B}_\lambda + {\bf B}_\lambda ^\top \Sigma _{ \mathbf{x} } {\bf B}_\lambda - \mu _{\varepsilon _\lambda }^2, \label{sigma eps lambda example B}
\end{align}
where
\begin{equation*}
\sigma ^2 _{y} = {\rm E} \left[ f (\xi) ^2 \right] + \sigma _{ \varepsilon } ^2
\end{equation*}
with
\begin{equation}
\label{Efx2}
{\rm E} \left[ f (\xi) ^2 \right] = \int_{-\infty}^{\infty} f(z)^2 g_{ \xi }(z) dz.
\end{equation}
Regarding the conditional moments, by \eqref{mu eps lambda cond x} we have
\begin{align}
\mu _{\varepsilon _\lambda | \mathbf{x} } &= f ( \xi ) - {\bf B} _\lambda ^\top\mathbf{x} ,\nonumber\\
\sigma ^2 _{\varepsilon_\lambda | \mathbf{x} } &= {\rm E}_ \mathbf{x}\left[ \varepsilon _\lambda ^2 \right] - \mu _{\varepsilon _\lambda | \mathbf{x} }^2\nonumber\\
&= f( \xi ) ^2 + \sigma _\varepsilon ^2 + {\bf B}_\lambda ^\top \mathbf{x} \mathbf{x} ^\top {\bf B}_\lambda - 2 f ( \xi ) {\bf B}_\lambda ^\top \mathbf{x} - f( \xi ) ^2 - {\bf B}_\lambda ^\top \mathbf{x} \mathbf{x} ^\top {\bf B}_\lambda + 2 f ( \xi ) {\bf B}_\lambda ^\top \mathbf{x} = \sigma _\varepsilon ^2. \label{sigma eps lambda cond x example B}
\end{align}
Notice that the expression \eqref{sigma eps lambda cond x example B} proves the conditional homoscedasticity of the ridge residuals in this setup. This observation has important implications and we hence frame it in the following proposition.
\begin{proposition}
\label{Exampe B homoscdasticity}
Let $X$ and $Y$ be random sample matrices generated by the regression model~\eqref{approximated model example B} in Example B that assumes that $ y$ is an additive stochastic perturbation of some continuous function of some random variable $\xi $. If, additionally, the stochastic perturbation and the random variable $\xi $ are independent, then the ridge regression residuals of~\eqref{approximated model example B} are always conditionally homoscedastic. $\blacktriangleleft$
\end{proposition}
\medskip
\noindent$\blacktriangleright$ {\bf Example C. Non-linear one-dimensional model with mutiplicative stochastic disturbance term}. Consider the following data generating process for the variable ${y} $
\begin{equation}
\label{exampleC model}
{y} = f(\xi ) {\varepsilon},
\end{equation}
where ${\varepsilon} \sim D_ \varepsilon ( \mu _\varepsilon , \sigma_{ \varepsilon }^2 )$ is a disturbance term and $\xi $ is distributed with zero mean, variance $ \sigma ^2 _\xi$, and probability density function $g _{ \xi }$. As in Examples~A and B, the random variables $\varepsilon $ and $\xi $ are assumed to be independent. In the conditions of this example we choose to approximate the model \eqref{exampleC model} by the same linear regression model \eqref{approximated model example B} with the covariates in \eqref{x example B} that are used in Example B. Notice that in this case, the relations \eqref{std}-\eqref{sigma xy example B} in Example B hold true. As for the ridge regression residuals $ \varepsilon _\lambda $, the choice \eqref{exampleC model} for the data generating process implies that
\begin{equation*}
\varepsilon _\lambda = f( \xi )\varepsilon - {\bf B}_\lambda ^\top \mathbf{x}.
\end{equation*}
It is easy to verify, using the independence of $\xi $ and $\varepsilon $, that the unconditional first moment of the ridge regression residuals is
\begin{equation*}
\mu _{\varepsilon _\lambda } = {\rm E} \left[ f( \xi ) \right] \mu _\varepsilon =\mu _\varepsilon \int _{-\infty}^{\infty} f(z) g_{ \xi } (z) dz,
\end{equation*}
As to the unconditional second moment, the relation \eqref{sigma eps lambda example B} holds true with $\sigma ^2 _{y} $ determined by
\begin{equation}
\label{sigma y Example C}
\sigma ^2 _{y} = {\rm E} \left[ f (\xi) ^2 \varepsilon ^2 \right] - \mu _{\varepsilon _\lambda }^2 ={\rm E} \left[ f (\xi) ^2 \right] (\sigma ^2 _{ \varepsilon } + \mu _\varepsilon^2 )- \mu _{\varepsilon _\lambda }^2
\end{equation}
and with ${\rm E} \left[ f (\xi) ^2 \right]$ as in \eqref{Efx2}.
The conditional first and second moments are determined by
\begin{align}
\mu _{ \varepsilon _\lambda | \mathbf{x} } = &f( \xi ) \mu _\varepsilon - {\bf B}_\lambda ^\top \mathbf{x},\nonumber \\
\label{sigma eps lambda cond x example C}
\sigma ^2 _{ \varepsilon _\lambda | \mathbf{x} }= &{\rm E}_ \mathbf{x}\left[ \varepsilon _\lambda ^2 \right] - \mu _{\varepsilon _\lambda | \mathbf{x} }^2= f( \xi ) ^2 (\sigma ^2 _{ \varepsilon } + \mu _\varepsilon^2 ) - 2 \mu _\varepsilon f (\xi) {\bf B}_\lambda ^\top \mathbf{x} + {\bf B}_\lambda ^\top \mathbf{x} \mathbf{x} ^\top {\bf B}_\lambda\nonumber\\
& - f( \xi ) ^2\mu _\varepsilon^2 +2 \mu _\varepsilon f (\xi) {\bf B}_\lambda ^\top \mathbf{x} - {\bf B}_\lambda ^\top \mathbf{x} \mathbf{x} ^\top {\bf B}_\lambda= f( \xi ) ^2 \sigma _\varepsilon ^2,
\end{align}
respectively.
Relation \eqref{sigma eps lambda cond x example C} shows that, in this case, the ridge residuals are not conditionally homoscedastic.
$\blacktriangleleft$
\section{Evaluation of the ridge regression and generalization errors}
\label{Evaluation of the ridge regression errors: training and testing}
In this section we use the properties of the ridge estimator that we spelled out in Theorem~\ref{Theorem 1} in order to write down explicit expressions that allow for the evaluation of the regression (also called training) and the generalization (also called testing) errors committed by a regularized regression model whose coefficient $\widehat{B} _\lambda $ has been estimated using a finite sample of a given size. The regression or training error is the one committed by the regression model when evaluated with the sample that has been used to obtain $\widehat{B} _\lambda $; for the generalization or testing error, we keep $\widehat{B} _\lambda $ and we evaluate the error committed by the corresponding regression model using another sample that may have different size or even different statistical properties. As we will see, in both cases our error formulas incorporate the error committed at the time of parameter estimation.
These two errors are of much importance in specific applications involving a finite sample framework since their relative values give indications about the pertinence of various modeling choices. Indeed, a typical behavior in a regression model is that an increase in the number of covariates makes smaller the training error. The testing error follows the same trend up to a point in which it starts increasing; that turning point in the testing error has to do with the appearance of overfitting in the training, that is, the overabundance of regressors is fitting not only the underlying deterministic model but also the noise that is perturbing it. As we will see in the example that we work out in Section~\ref{Numerical illustration}, the formulas presented in this section can be used to avoid this phenomenon at the time of deciding on the model architecture.
All along this section we use the same hypotheses and notation as in Theorem \ref{Theorem 1} unless explicitly stated otherwise.
\subsection{The characteristic errors of the regression model}
\label{Characteristic errors of the regression model}
We define the characteristic error as the one committed by the regression model without taking into account estimation errors, that is, the characteristic error is evaluated assuming that the model parameters are known. We separately address two instances of the characteristic error. First, we consider the standard unconditional characteristic error corresponding to the ridge model with a given regularization strength $\lambda \in \mathbb{R} ^+$. This error corresponds to the expected regression error for arbitrary realizations of the explanatory and dependent variables $\mathbf{x} $ and $\mathbf{y} $, respectively; it is sometimes referred to in the literature as the irreducible error (see for example \cite{Elements:learning:book}). Second, we assume that the dependent variables $\mathbf{x} $ are fixed and we compute the characteristic error conditional on those values.
\medskip
\noindent {\bf Characteristic (irreducible) error.} It is given by:
\begin{equation}
\label{MSE}
{\rm MSE}_{\rm char} ^\lambda :={\rm E} \left[ (\mathbf{y} - B _\lambda \mathbf{x} )^\top (\mathbf{y} - B _\lambda \mathbf{x} )\right] = {\rm E} \left[ \boldsymbol{\varepsilon} _{ \lambda }^\top \boldsymbol{\varepsilon} _{ \lambda }\right] = {\rm trace} ( \Sigma _{ {\boldsymbol{\varepsilon}} _\lambda} + \boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda }\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda }^\top),
\end{equation}
where $\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda }$ and $ \Sigma _{ {\boldsymbol{\varepsilon}} _\lambda}$ are given in \eqref{distr eps}.
\medskip
\noindent {\bf Conditional characteristic error.} It is the characteristic error associated to the ridge regression model conditional on the covariates value $\mathbf{x} $:
\begin{equation}
\label{MSE cond}
{\rm MSE}_{{\rm char}| \mathbf{x} } ^\lambda :={\rm E}_{\mathbf{x}} \left[ (\mathbf{y} - B _\lambda \mathbf{x} )^\top (\mathbf{y} - B _\lambda \mathbf{x} ) \right] = {\rm E}_{ \mathbf{x} } \left[ \boldsymbol{\varepsilon}_\lambda ^\top \boldsymbol{\varepsilon}_\lambda\right] = {\rm trace} ( \Sigma _{\boldsymbol{\varepsilon} _\lambda |\mathbf{x} }+\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} }\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} }^\top),
\end{equation}
with $\boldsymbol{\mu} _{\boldsymbol{\varepsilon}_\lambda | {\bf x} }$ and $ \Sigma _{\boldsymbol{\varepsilon} _\lambda |\mathbf{x} }$ as in \eqref{mu eps lambda cond x}.
\medskip
\noindent We now evaluate these errors for the examples that we consider along the paper.
\medskip
\noindent$\blacktriangleright$ {\bf Example~A (continued)}.
\normalfont
The expressions for the characteristic and the conditional characteristic errors in \eqref{MSE} and in \eqref{MSE cond} for the standard regression model in \eqref{regression model} can be made explicit by using the particular form of the covariance matrix $\Sigma _{ {\boldsymbol{\varepsilon}} _\lambda} $ provided in \eqref{sigma lambda example} and of the corresponding conditional covariance $\Sigma _{\boldsymbol{\varepsilon} _\lambda |\mathbf{x} }$ given in \eqref{sigma lambda x}. Indeed:
\begin{align}
\label{MSE char}
{\rm MSE}_{\rm char} ^\lambda &={\rm trace}( \Sigma _{\boldsymbol{\varepsilon}} + (B - B _\lambda ) ^\top(\Sigma _ {\mathbf{x} } + \boldsymbol{\mu}_{\bf x}\boldsymbol{\mu}_{\bf x}^\top)(B - B _\lambda ))=
{\rm trace}( \Sigma _{\boldsymbol{\varepsilon}} + (B - B _\lambda ) ^\top{\rm E}\left[\mathbf{x} \mathbf{x}^\top\right](B - B _\lambda )),
\end{align}
and for the conditional characteristic error:
\begin{align}
{\rm MSE}_{{\rm char}| \mathbf{x} } ^\lambda &={\rm trace}( \Sigma _{\boldsymbol{\varepsilon}} + (B - B _\lambda ) ^\top\mathbf{x} \mathbf{x}^\top(B - B _\lambda )).\label{MSE char x}
\end{align}
From these expressions it is obvious that the relative magnitudes of these characteristic and conditional characteristic errors depend on the specific value of the random variable $\mathbf{x} \in \mathbb{R} ^p $ used in the conditioning. $\blacktriangleleft$
\medskip
\noindent$\blacktriangleright$ {\bf Example~B (continued)}. In this particular instance the characteristic error is:
\begin{align}
\label{MSE char example B}
{\rm MSE}_{\rm char} ^\lambda &= \sigma _{ \varepsilon } ^2 + {\rm E} \left[ f (\xi) ^2 \right] + {\bf B}_\lambda ^\top \Sigma _{ \mathbf{x} } {\bf B}_\lambda -2 {\bf B}_\lambda ^\top \mathbf{ \Sigma }_{\mathbf{x} y},
\end{align}
where ${\rm E} \left[ f (\xi) ^2 \right]$, $\mathbf{ \Sigma }_{\mathbf{x} y}$, and $\mathbf{ \Sigma }_{\mathbf{x} }$ are given in \eqref{Efx2}, \eqref{sigma xy example B}, and \eqref{sigma x example B}, respectively.
Additionally, the expression of the conditional characteristic error is
\begin{align}
{\rm MSE}_{{\rm char}| \mathbf{x} } ^\lambda &= \sigma _\varepsilon ^2 + f (\xi) ^2 + {\bf B}_\lambda^\top \mathbf{x} \mathbf{x} ^\top {\bf B}_\lambda - 2 f ( \xi ) {\bf B}_\lambda ^\top \mathbf{x}.\label{MSE char x example B}
\end{align}
It can be easily seen from \eqref{MSE char example B} and \eqref{MSE char x example B} that as in Example A, any ordering between these two errors is possible depending on the value of the random variable $\mathbf{x} \in \mathbb{R} ^p $.
$\blacktriangleleft$
\medskip
\noindent$\blacktriangleright$ {\bf Example~C
(continued)}. The characteristic error is:
\begin{align}
\label{MSE char example C}
{\rm MSE}_{\rm char} ^\lambda &= \sigma ^2 _{ \varepsilon } {\rm E} \left[ f (\xi) ^2 \right] + {\bf B}_\lambda ^\top \Sigma _{ \mathbf{x} } {\bf B}_\lambda - 2 {\bf B}_\lambda ^\top \mathbf{ \Sigma }_{\mathbf{x} y} ,
\end{align}
where ${\rm E} \left[ f (\xi) ^2 \right]$, $\mathbf{ \Sigma }_{\mathbf{x} y}$, $\mathbf{ \Sigma }_{\mathbf{x} }$, and $\mu _{\varepsilon _\lambda }$ are given in \eqref{Efx2}, \eqref{sigma xy example B}, \eqref{sigma x example B}, and \eqref{mu eps lambda example B}, respectively. The expression of the conditional characteristic error is provided by:
\begin{align}
{\rm MSE}_{{\rm char}| \mathbf{x} } ^\lambda &= \sigma _\varepsilon ^2 f( \xi ) ^2 + f( \xi ) ^2\mu _\varepsilon^2 -2 \mu _\varepsilon f (\xi) {\bf B}_\lambda ^\top \mathbf{x} + {\bf B}_\lambda ^\top \mathbf{x} \mathbf{x} ^\top {\bf B}_\lambda .\label{MSE char x example C}
\end{align}
As in Examples A and B, any ordering between these two errors is possible depending on the value of the random variable $\mathbf{x} \in \mathbb{R} ^p $.
$\blacktriangleleft$
\subsection{The conditional total training error}
\label{Conditional total training error}
In this section we provide an explicit expression for the total error committed when using a multivariate ridge regularized regression model with a coefficient matrix that has been estimated using a finite sample. In that situation there are two sources of error: first, the characteristic error associated to the regression model and second, the estimation error on the ridge regression matrix that appears due to the finiteness of the estimation sample. We first focus on the total regression or training error in which the error is evaluated using the same sample that has been earlier exploited to obtain an estimate of the the ridge regression matrix. The expression for the error that we provide is conditional on the covariates sample used at the time of estimation.
Consider the random matrices $X \in \mathbb{M} _{p, N}$, $Y\in \mathbb{M} _{q, N}$ constructed according to the prescriptions in Section~\ref{The ridge estimator and its properties} and let $\widehat{B} _\lambda \in \mathbb{M} _{p,q}$ be the estimator of $B_\lambda $ based on $X$ and $Y$ that we introduced in~\eqref{B lambda random sample est}. The conditional total training error is defined as
\begin{equation}
\label{MSE training}
{\rm MSE}_{{\rm training}|X}^\lambda := \dfrac{1}{N} {\rm trace} \left( {\rm E} _X\left[ (Y - \widehat{B} _\lambda ^\top X) ^\top (Y - \widehat{B} _\lambda ^\top X)\right] \right) .
\end{equation}
The following theorem, whose proof is given in Appendix~\ref{Proof of Theorem 2}, provides an explicit expression for the conditional total training error.
\begin{theorem}
\label{Theorem 2}
Suppose that we are in the hypotheses of Theorem \ref{Theorem 1}, that is, we consider a ridge regression between square summable explanatory and dependent variables that exhibits conditionally normal and homoscedastic residuals. Then, the total training error conditional on a random sample $X$ of length $N$ of the covariates is:
\begin{align}
\label{MSE longer}
{\rm MSE}^\lambda&_{{\rm training}|X} = \nonumber\\
=&{\rm trace} (\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda ) + \dfrac{1}{N} \Big\{{\rm trace}(\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda) {\rm trace}\left( Z _\lambda (R_\lambda X X ^\top - 2 \mathbb{I}_p )\right) + {\rm trace} ( M _{(\widehat{B} _\lambda -{B} _\lambda) }M _{(\widehat{B} _\lambda -{B} _\lambda) }^\top X X ^\top ) \nonumber\\
& -2 \ {\rm trace}\Big( (X ^\top R _\lambda X A_N - \dfrac{1}{2} \mathbb{I}_N )M _{E_\lambda |X} ^\top M _{E_\lambda |X}+ (X ^\top R _\lambda X A _N -\mathbb{I}_N ) X^\top B _\lambda \boldsymbol{\mu} _{E_\lambda |X} \Big) \Big\} ,
\end{align}
or, equivalently,
\begin{align}
\label{MSE shorter}
{\rm MSE}^\lambda&_{{\rm training}|X} = \nonumber\\=&{\rm trace} (\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda ) + \dfrac{1}{N} \Big\{{\rm trace}(\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda) {\rm trace}\left( Z _\lambda (R_\lambda X X ^\top - 2 \mathbb{I}_p )\right) + {\rm trace} \left( M _{(\widehat{B} _\lambda -{B} _\lambda) }M _{(\widehat{B} _\lambda -{B} _\lambda) }^\top X X ^\top \right)\nonumber\\
&-2 \ {\rm trace}((X ^\top R _\lambda X A_N - \dfrac{1}{2} \mathbb{I}_N ) M_{E_\lambda |X} ^\top M _{E_\lambda |X})\Big\} + {2}\lambda{\rm trace}( X ^\top R _\lambda B _\lambda M_{E_\lambda |X} ),
\end{align}
where
\begin{equation}
\label{mu M}
M _{(\widehat{B} _\lambda -{B} _\lambda) }:=-\lambda N R_\lambda B_\lambda + R_\lambda X A_N M _{E_\lambda |X} ^\top
\end{equation}
is the bias of the estimator $\widehat{B} _\lambda $ of $B_\lambda $ in \eqref{B lambda hat minus B lambda} and where \begin{align}
\label{Z lambda theorem}
Z _\lambda &= R _\lambda X A _N X^\top= \mathbb{I} _p - \lambda N R _\lambda,\\
R _\lambda&=\left(X A_{N} X ^\top + \lambda N \mathbb{I}_p \right) ^{-1}.\label{R train}
\end{align}
\end{theorem}
\medskip
\noindent$\blacktriangleright$ {\bf Example~A (continued).} In the conditions of the linear regression model~\eqref{regression model} of Example~A, the conditional total training error in \eqref{MSE training} can be made more explicit by using the expressions for the conditional covariance matrix $\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda $ in~\eqref{sigma lambda x} that shows that the homoscedasticity hypothesis is satisfied in this setup. Moreover, as we saw in Corollary~\ref{Corollary Theorem 1}, the properties of the estimator $\widehat{B} _\lambda $ can be formulated in this case in terms of the matrix $B$. The proof of the following result is provided in Appendix \ref{Proof of Corollary error Example A}.
\begin{corollary}
\label{Corollary error Example A}
In the conditions of Example~A, the conditional total training error~\eqref{MSE training} is given by the expression:
\begin{align}
\label{MSE training example A}
{\rm MSE}_{{\rm training}|X}^\lambda = &{\rm trace} ( \Sigma _ { \boldsymbol{\varepsilon} } ) + \dfrac{1}{N} {\rm trace}(\Sigma _ { \boldsymbol{\varepsilon} }) {\rm trace}\left( Z _\lambda (R_\lambda X X ^\top - 2 \mathbb{I}_p )\right) + \lambda ^2 N {\rm trace}( R_\lambda B B ^\top R_\lambda X X ^\top),
\end{align}
where $Z _\lambda$ is defined as in Corollary~\ref{Corollary Theorem 1}.
\end{corollary}
Expression~\eqref{MSE training example A} has been formulated for the first time in Proposition 3 of \cite{RC3} in a machine learning context.$\blacktriangleleft$
\subsection{The conditional total testing or generalization error}
\label{Conditional total testing (generalization) error}
As we already pointed out in the introduction to this section, the minimization of the training error in the construction of a model does not suffice to guarantee its good out-of-sample performance due to the possible appearance of overfitting phenomena. This makes necessary the evaluation of the so called testing or generalization error that measures the performance of a given model that has been trained on a finite sample that is independent and may even have different statistical properties from the one that is used for the testing. In specific machine learning and forecasting applications it is mainly the testing error that needs to be used in order to decide on questions related to model architecture.
We proceed in the following way in order to carry this out: we assume that the training and testing are carried out on two different sets of samples; those labeled with a one (respectively, two) are the training (respectively, testing) samples. More specifically, consider two pairs of random samples of sizes $N _1 $ and $N _2 $. The first pair $ \left\{\mathbf{x} ^{(1)} _i \right\} $, $\left\{ \mathbf{y} _i ^{(1)} \right\} $, with elements $\mathbf{x} ^{(1)} _i \in \mathbb{R} ^p $, $\mathbf{y} ^{(1)} _i \in \mathbb{R} ^q $, $i \in \left\{ 1, \dots, N _1 \right\} $, will be used for training and the second one, denoted by $ \left\{\mathbf{x} ^{(2)} _j \right\} $, $\left\{ \mathbf{y} _j ^{(2)} \right\} $ with $\mathbf{x} ^{(2)} _j \in \mathbb{R} ^p $, $\mathbf{y} ^{(2)} _j \in \mathbb{R} ^q $, $j\in \left\{ 1, \dots, N _2 \right\} $, will be used for testing. Suppose that all the elements in these random samples satisfy hypothesis {\bf (A1)} and hence have finite first and second order moments.
Additionally, assume that ridge residuals $ \left\{\boldsymbol{\varepsilon} ^{(1)} _{ \lambda , i} \right\} _{i \in \left\{ 1, \dots, N _1 \right\} }$ associated to the training pair are conditionally normal and homoscedastic in the sense of Lemma \ref{Lemma 2}, that is, their corresponding conditional covariance matrices are constant and equal to some matrix $\Sigma _{ \boldsymbol{\varepsilon} | {\mathbf{x} } } ^ {\lambda, (1)}\in \Bbb S _q $. This hypothesis can be stated as $\boldsymbol{\varepsilon} ^{(1)} _{ \lambda , i} | \mathbf{x} ^{(1)} _i \sim {\rm N} (\boldsymbol{\mu} _{\boldsymbol{\varepsilon}^{(1)} _{\lambda , i}| \mathbf{x} ^{(1)} _i} , \Sigma _{ \boldsymbol{\varepsilon} | {\mathbf{x} } } ^ {\lambda, (1)})$ for each $i \in \left\{ 1, \dots, N _1 \right\} $.
Consider now the random matrices $X _1 \in \mathbb{M} _{p, N _1 }$, $Y _1 \in \mathbb{M} _{q, N _1 }$, $X _2 \in \mathbb{M} _{p, N_2 }$, $Y_2 \in \mathbb{M} _{q, N_2 }$ obtained by horizontal concatenation of the elements in the random samples $ \left\{\mathbf{x} ^{(1)} _i \right\} _{i \in \left\{ 1, \dots, N _1 \right\} }$, $\left\{ \mathbf{y} _i ^{(1)} \right\} _{i \in \left\{ 1, \dots, N _1 \right\} }$, $ \left\{\mathbf{x} ^{(2)} _i \right\} _{i \in \left\{ 1, \dots, N _2 \right\} }$, and $\left\{ \mathbf{y} _i ^{(2)} \right\} _{i \in \left\{ 1, \dots, N _2 \right\} }$ respectively. We call $( X _1 , Y _1 )$ and $( X _2 , Y _2 )$ the training and the testing samples, respectively.
Let now $\widehat {B} _\lambda $ be the ridge regression matrix estimator of ${B} _\lambda $ constructed using a training sample $( X _1 , Y _1 )$. In these conditions we define the conditional total testing ridge error as:
\begin{equation}
\label{MSE testing}
{\rm MSE} ^\lambda _{{\rm testing}|X_1 } = \dfrac{1}{N _2 } {\rm trace} \left( {\rm E}_{X_1} \left[\left(Y _2 - \widehat{B} _\lambda ^\top X _2 \right) ^\top \left(Y _2 - \widehat{B} _\lambda ^\top X _2 \right) \right] \right) .
\end{equation}
This expression measures the mean square error committed when training is carried out with a fixed explanatory variables training sample $X _1$ of length $N _1 $ and testing is implemented with any other testing sample $\left(X _2 , Y _2 \right) $ of length $N _2 $. On other words, we fix $X _1$, and we average the square errors committed when varying $Y _1 $ and the testing samples $\left(X _2 , Y _2 \right) $. The proof of the following result is provided in Appendix~\ref{Proof of Theorem testing error}.
\begin{theorem}
\label{Theorem testing error}
Consider two pairs of random samples $ \left\{\mathbf{x} ^{(1)} _i \right\} $, $\left\{ \mathbf{y} _i ^{(1)} \right\} $ and $ \left\{\mathbf{x} ^{(2)} _j \right\} $, $\left\{ \mathbf{y} _j ^{(2)} \right\} $, with $\mathbf{x} ^{(1)} _i, \mathbf{x} ^{(2)} _i \in \mathbb{R} ^p $, $\mathbf{y} ^{(1)} _i, \mathbf{y} ^{(2)} _i \in \mathbb{R} ^q $, of sizes $N _1 $ and $N _2 $, respectively. The first pair will be used for training and the second one for testing. Suppose that all the elements in these random samples satisfy hypothesis {\bf (A1)} and that the ridge residuals $ \left\{\boldsymbol{\varepsilon} ^{(1)} _{ \lambda , i} \right\} _{i \in \left\{ 1, \dots, N _1 \right\} }$ associated to the training pair are conditionally normal and homoscedastic in the sense of Lemma \ref{Lemma 2}, that is, their corresponding conditional covariance matrices are constant and equal to some matrix $\Sigma _{ \boldsymbol{\varepsilon} | {\mathbf{x} } } ^ {\lambda, (1)}\in \Bbb S _q $.
Consider now the random matrices $X _1 \in \mathbb{M} _{p, N _1 }$ and $Y _1 \in \mathbb{M} _{q, N _1 }$ obtained by horizontal concatenation of the elements in the training samples. Under these hypotheses, the conditional total testing error introduced in~\eqref{MSE testing} can be written as:
\begin{align}
\label{MSE test}
{\rm MSE} ^\lambda _{{\rm testing}|X_1 } &= {\rm trace} \left( \Sigma ^{(2)} _{\mathbf{y} } + \boldsymbol{\mu} ^{(2)} _{\mathbf{y} } \boldsymbol{\mu} _{\mathbf{y} }^ {(2)\top} \right) - 2\, {\rm trace}\left( \left(\Sigma ^{(2)}_{\mathbf{x} \mathbf{y} } + \boldsymbol{\mu} _{ \mathbf{x} }^{(2)}\boldsymbol{\mu} _{\mathbf{y} } ^{(2)\top} \right) M_{ \widehat{B} _\lambda}^\top \right) \nonumber\\
&+{\rm trace} \left[ \left( \Sigma^{(2)} _{\mathbf{x} } + \boldsymbol{\mu}^{(2)} _{\mathbf{x} } \boldsymbol{\mu} _{\mathbf{x} }^ {(2)\top} \right) \left({\rm trace} \left(\Sigma ^{\lambda, (1)}_{ \boldsymbol{\varepsilon} | \mathbf{x} }\right) Z _\lambda R _\lambda + M_{ \widehat{B} _\lambda}M ^{ \top}_{ \widehat{B} _\lambda} \right)\right],
\end{align}
where
\begin{align}
Z _\lambda &:=R _\lambda X _1 A_{N_1} X _1 ^\top ,\label{Z test}\\
R _\lambda &:=\left(X _1 A_{N_1} X _1 ^\top + \lambda N _1 \mathbb{I}_p \right) ^{-1},\label{R test} \\
M _{ \widehat{B} _\lambda}&:= B _\lambda - \lambda N _1 R _\lambda B _\lambda + R _\lambda X _1 A _{N_1} M _{ E_\lambda | X} ^{(1)\top}, \label{M test}\\
\Sigma ^{(2)} _{\mathbf{y} } &= {\rm Cov} ( \mathbf{y} ^{(2)} ,\mathbf{y} ^{(2)} ),\label{Sigma y test}\\
\Sigma ^{(2)} _{\mathbf{x} } &= {\rm Cov} (\mathbf{x} ^{(2)} ,\mathbf{x} ^{(2)} ),\label{Sigma x test}\\
\Sigma ^{(2)} _{\mathbf{x} \mathbf{y} } &= {\rm Cov} ( \mathbf{x} ^{(2)} ,\mathbf{y} ^{(2)} ).\label{Sigma xy test}
\end{align}
with
\begin{align}
\label{}
A _{N_1} &:= \mathbb{I} _{N_1} - \dfrac{1}{N _1 } \mathbf{i} _{N _1 }\mathbf{i} _{N _1 } ^\top.
\end{align}
\end{theorem}
\noindent$\blacktriangleright$ {\bf Example~A (continued).} As we already pointed out when working with the training error, the homoscedasticity hypothesis is satisfied in the setup of Example~A with $\Sigma ^{\lambda, (1)}_{ \boldsymbol{\varepsilon} | \mathbf{x} } = \Sigma ^{(1)}_{ \boldsymbol{\varepsilon} }$ and the properties of the estimator $\widehat{B} _\lambda $ can be formulated in this case in terms of the matrix $B$. These observations lead us to the following corollary of Theorem \ref{Theorem testing error} that provides an explicit formula for the generalization error in this setup.
\begin{corollary}
\label{Corollary testing error Example A}
In the conditions of Example~A the conditional total testing error is determined by the following expression:
\begin{align}
\label{MSE test example A}
{\rm MSE} ^\lambda _{{\rm testing}|X_1 } &= {\rm trace} \left( \Sigma ^{(2)} _{\mathbf{y} } + \boldsymbol{\mu} ^{(2)} _{\mathbf{y} } \boldsymbol{\mu} _{\mathbf{y} }^ {(2)\top} \right) - 2\, {\rm trace}\left( \left(\Sigma ^{(2)}_{\mathbf{x} \mathbf{y} } + \boldsymbol{\mu} _{ \mathbf{x} }^{(2)}\boldsymbol{\mu} _{\mathbf{y} } ^{(2)\top} \right) B^\top Z _\lambda ^\top \right) \nonumber\\
&+{\rm trace} \left[ \left( \Sigma^{(2)} _{\mathbf{x} } + \boldsymbol{\mu}^{(2)} _{\mathbf{x} } \boldsymbol{\mu} _{\mathbf{x} }^ {(2)\top} \right) \left({\rm trace} (\Sigma ^{ (1)}_{ \boldsymbol{\varepsilon} }) Z _\lambda R _\lambda + Z _\lambda B B ^\top Z_\lambda ^\top \right)\right],
\end{align}
with $R _\lambda$, $\Sigma ^{(2)} _{\mathbf{y} } $, $\Sigma ^{(2)} _{\mathbf{x} }$, and $\Sigma ^{(2)} _{\mathbf{x} \mathbf{y} }$ as in \eqref{R test}, \eqref{Sigma y test}, \eqref{Sigma x test}, and \eqref{Sigma xy test}, respectively. $\blacktriangleleft$,
\end{corollary}
\subsection{Numerical illustration}
\label{Numerical illustration}
The goal of this section is showing how to compute in an explicit example the total training and testing error formulas that we provided in Theorems~\ref{Theorem 2} and~\ref{Theorem testing error} and to show how they can be used at the time of deciding on a modeling architecture that minimizes overfitting and generalization errors. On other words, we will see how the theoretical results that we just provided can help in finding the optimal trade-off between modeling complexity and out-of-sample performance.
\medskip
\noindent {\bf The underlying model.} Consider a particular instance of Example B where the dependent variable ${y} $ is generated by the additive stochastic perturbation of a deterministic model of the form
\begin{equation}
\label{exampleB model2}
{y} = f(\xi ) + {\varepsilon},
\end{equation}
with
\begin{equation}
\label{f xi}
f(\xi ) = e^ \xi - k, \enspace {\rm with} \enspace
k = \dfrac{1}{2}( e - e^{-1}).
\end{equation}
The term ${\varepsilon} \sim {\rm N} ( 0 , \sigma_{ \varepsilon }^2 )$ in the nonlinear model \eqref{exampleB model2} represents the disturbance term. Additionally, we assume that $\xi $ is uniformly distributed in the interval $[-1,1] $, that is, ${ \xi }\sim \mathcal{U} ( -1 , 1 )$ and hence has probability density function $g_\xi(z) = \frac{1}{2} I_{[-1,1]}(z) $. As it was assumed in the general conditions of Example B, we suppose that $\varepsilon $ and $\xi $ are independent random variables and we require that ${\rm E} \left[ f( \xi )\right] = 0$, which is automatically guaranteed by the choice of $k$ in \eqref{f xi}.
\medskip
\noindent {\bf A polynomial approximation and the corresponding linear regression model.}
We now construct a polynomial approximation of \eqref{exampleB model2} and, as in Example B, we formulate this procedure as a linear univariate regularized regression model of the form
\begin{equation}
\label{approximated model example B1}
y = {{\bf B}} _\lambda ^\top \mathbf{x} + \varepsilon _\lambda,
\end{equation}
where the covariates $\mathbf{x} $ are adequately standardized powers of increasing order of the random variable $\xi$. More specifically, we set:
\begin{equation}
\label{x example B2}
\mathbf{x} := \left( \widetilde{\xi} , \widetilde{\xi^2}, \dots, \widetilde{ \xi^p } \right) ^\top, \enspace {\rm with }\enspace x _i = \widetilde{ \xi ^i} := \dfrac{ \xi ^i - {\rm E} \left[ \xi ^i\right] }{ \sigma (\xi ^i)}, \enspace i \in \left\{ 1, \dots, p\right\},
\end{equation}
Notice that the polynomials that we are using in the construction of the explanatory variables $\mathbf{x} $ are not orthogonal and, moreover, produce covariance matrices $\Sigma _{\bf x} $ whose condition numbers increase rapidly with the degree $p$. This feature makes pertinent in this context the use of the ridge regularization.
Given that ${ \xi }\sim \mathcal{U} ( -1 , 1 )$, we have:
\begin{equation}
\label{exp and std}
{\rm E} \left[ \xi ^i\right] = \dfrac{(-1)^i + 1}{2(i+1)}, \enspace {\rm and} \enspace \sigma \left( \xi ^i\right) = \left(\dfrac{1}{2i+1} - \dfrac{(-1)^i + 1}{2(i+1)^2}\right)^{1/2}.
\end{equation}
Once a regularization strength $\lambda \in \mathbb{R} ^+$ has been fixed, the corresponding regularized regression vector ${\bf B} _\lambda $ is given by \eqref{B lambda}, that is, ${\bf B} _\lambda := \left(\Sigma _{\bf x} + \lambda \mathbb{I}_p \right) ^{-1} {\mathbf{\Sigma}} _{{\bf x}y} $. In this relation $\Sigma _{\bf x} \in \Bbb S _p $ and $ {\mathbf{\Sigma}} _{{\bf x}y} \in \mathbb{R}^p $ can be explicitly computed using the fact that ${ \xi }\sim \mathcal{U} ( -1 , 1 )$. Indeed, for $i, j\in \left\{ 1, \dots, p \right\}$ we have:
\begin{align}
\label{sigma x example B2}
(\Sigma _{\bf x})_{i,j}=&\dfrac{1}{\sigma (\xi ^i)\sigma (\xi ^j)}\left( {\rm E} \left[ \xi ^{i+j}\right] - {\rm E} \left[ \xi ^i\right]{\rm E} \left[ \xi ^j\right]\right) ,\\
(\mathbf{\Sigma} _{{\bf x}y})_{i}=& \dfrac{1}{2\sigma (\xi ^i)} \Big( (-1)^i \sum^{i}_{s = 0} \dfrac{i!}{s!} ((-1)^s e - e ^{-1}) - 2 k {\rm E} \left[ \xi ^i\right]\Big), \label{sigma xy example B2}
\end{align}
with ${\rm E} \left[ \xi ^i\right] $ and $\sigma ( \xi^i)$ as in \eqref{exp and std}. In order to complete the model specification in \eqref{approximated model example B1} we determine the regression residuals $ \varepsilon _\lambda $ that are given by
\begin{equation}
\label{residuals lambda example B2}
\varepsilon _\lambda = e^ \xi -k - {\bf B}_\lambda ^\top \mathbf{x} + \varepsilon.
\end{equation}
The general results in Example B provide the following expressions for the first and the second unconditional moments of $\varepsilon _\lambda $
\begin{align}
\label{mu eps lambda example B2}
\mu _{\varepsilon _\lambda } &= 0,\\
\sigma ^2 _{\varepsilon_\lambda } &= \sigma ^2 _{y} - {\bf B}_\lambda ^\top \mathbf{ \Sigma }_{\mathbf{x} y} - \mathbf{ \Sigma }_{\mathbf{x} y} ^\top {\bf B}_\lambda + {\bf B}_\lambda ^\top \Sigma _{ \mathbf{x} } {\bf B}_\lambda, \label{sigma eps lambda example B2}
\end{align}
where
\begin{equation}
\label{sigma y ExampleB2}
\sigma ^2 _{y} = \dfrac{1}{2}(1-e^{-2}) + \sigma _\varepsilon ^2.
\end{equation}
\noindent {\bf The conditional training and testing errors.} We now use the formulas in Theorems~\ref{Theorem 2} and~\ref{Theorem testing error} in order to asses the performance of the
approximating linear regression model \eqref{approximated model example B1} as a function of its complexity or, more specifically, its ability to reproduce the underlying nonlinear relation between $\xi $ and $y$. We will carry this out in two different situations that put the accent in the overfitting and in the lack of generalization power that the model may incur when the selected number of covariates is too high. These features will be detected by studying the evolution of the relative values of the conditional training and testing errors as a function of the model parsimony.
Regarding the conditional total training error, as we explained earlier, the training sample is generated using the underlying model in \eqref{exampleB model2}-\eqref{f xi} with $\xi \sim \mathcal{U}(-1,1)$. The sample length is denoted by $N _1 $. In order to evaluate the conditional total training error in Theorem~\ref{Theorem 2}, the following expressions for
the conditional first and the second moments of the ridge residuals are required
\begin{align}
\label{mu eps lambda cond x example B1}
\mu _{\varepsilon _\lambda | \mathbf{x} } &= e^ \xi -k - {\bf B} _\lambda ^\top\mathbf{x} ,\\
\sigma ^2 _{\varepsilon_\lambda | \mathbf{x} } &= \sigma _\varepsilon ^2. \label{sigma eps lambda cond x example B1}
\end{align}
As to the conditional total testing error, we consider a general case in which the random testing sample $({y^{(2)}}, {\bf x}^{(2)})$ of length $N _2 $ is generated using the underlying model \eqref{exampleB model2}-\eqref{f xi} but, this time around, we allow the support of the random variable used to generate the dependent variable $y^{(2)} $ to be an arbitrary closed interval, that is,
\begin{equation}
\label{true model testing}
y^{(2)}=e^{\zeta } - k + \varepsilon,
\end{equation}
where $k$ is given in \eqref{f xi}, $\varepsilon \sim {\rm N} (0, \sigma _\varepsilon ^2 )$, and $\zeta \sim \mathcal{U}(a, b)$ with $a,b \in \mathbb{R} $ such that $a<b $. At the same time, the covariates sample ${\bf x}^{(2)}$ is constructed as in \eqref{x example B2}, that is,
\begin{equation}
\label{x example B2 test}
\mathbf{x}^{(2)} := \left( \widetilde{\zeta } , \widetilde{\zeta^2}, \dots, \widetilde{ \zeta^p } \right) ^\top, \enspace {\rm with }\enspace {(\mathbf{x}^{(2)})} _i = \widetilde{ \zeta ^i} := \dfrac{ \zeta ^i - {\rm E} \left[ \xi ^i\right] }{ \sigma (\xi ^i)}, \enspace i \in \left\{ 1, \dots, p\right\},
\end{equation}
where $ {\rm E} \left[ \xi ^i\right]$ and $\sigma \left( \xi ^i\right)$ for ${ \xi }\sim \mathcal{U} ( -1 , 1 )$ are defined in \eqref{exp and std}.
The sample matrix $X^{(2)} \in \mathbb{M} _{p, N_2}$ and vector ${\bf y}^{(2)\top} \in \mathbb{R} ^{N_2}$ are obtained by horizontal concatenation of the $N_2$ realizations of $({y^{(2)}}, {\bf x}^{(2)})$.
In order to evaluate the conditional total testing error in \eqref{MSE test}, we need the first and second order moments of the random testing sample. First, we notice that since $\xi^{(2)} \sim \mathcal{U}(a, b)$, we can conclude that:
\begin{align}
\label{}
(\boldsymbol{\mu} _{\mathbf{x} }^{(2)})_i = \dfrac{1}{\sigma \left( \xi^i\right)} \left( {\rm E} \left[ \zeta ^i\right] - {\rm E} \left[ \xi ^i\right]\right) , \enspace i \in \left\{ 1, \dots, p \right\},
\end{align}
where
\begin{equation}
\label{exp zeta}
{\rm E} \left[ \zeta ^i\right] = \dfrac{b^{i+1}-a^{i+1}}{(i+1)(b-a)}
\end{equation}
and where for each $i \in \left\{ 1, \dots, p \right\}$, $ {\rm E} \left[ \xi ^i\right]$ and $\sigma \left( \xi ^i\right)$ are given by \eqref{exp and std}.
Additionally, the first order moment of $y^{(2)}$ is given by
\begin{equation}
\label{mu y 2}
\mu _{y}^{(2)} = \dfrac{1}{b-a} (e^b-e^a)-k.
\end{equation}
The second order moments of the testing sample are:
\begin{align}
\label{sigma x example B test}
(\Sigma _{\bf x}^{(2)})_{i,j}=& \dfrac{1}{(b-a)\sigma (\xi ^i)\sigma (\xi ^j)}\int_{a}^{b} \left( z^i - {\rm E} \left[ \xi ^i\right]\right) \left( z^j - {\rm E} \left[ \xi ^j\right]\right) dz - (\boldsymbol{\mu} _{\mathbf{x} }^{(2)})_i(\boldsymbol{\mu} _{\mathbf{x} }^{(2)})_j, \enspace i, j\in \left\{ 1, \dots, p \right\}, \\
(\mathbf{\Sigma} _{{\bf x}y}^{(2)})_{i}=& \dfrac{1}{(b-a)\sigma (\xi ^i)}\int_{a}^{b} \left( z^i - {\rm E} \left[ \xi ^i\right]\right)(e^z -k) dz - (\boldsymbol{\mu} _{\mathbf{x} }^{(2)})_i \mu _y^{(2)}, \enspace i\in \left\{ 1, \dots, p \right\}, \label{sigma xy example B test} \\
(\sigma ^2_y)^{(2)} =& \dfrac{1}{2 (b-a)} (e^{2b} - e^{2a}) - \dfrac{1}{(b-a)^2} (e^b - e^a)^2 + \sigma_{ \varepsilon }^2.
\end{align}
These relations allow for the explicit evaluation of the conditional total testing error that we now numerically study for two particular choices for the values of $a$ and $b$ that determine the support of the distribution of $ \zeta $.
\medskip
\noindent {\bf Numerical illustration. Case I. Overfitting.} Consider first the case in which $[a,b]=[-1,1]$ and hence $ \zeta = \xi$. This means that the testing sample is generated in the same way as the training one. In this setup, an increase in the number of covariates $p$ amounts to an increase in the degree of the polynomial that we are using to fit the sample. When this degree is too high, the fit may lead to a match with the noise in the sample rather than with the underlying deterministic signal that we are trying to model. Figure~\ref{Figure 1} shows the evolution of the conditional training and total errors for this case (for a given training covariates sample $X _1$ and a fixed regularization strength reported in the caption). As we anticipated, the conditional total training error is a decreasing function with $p$ but, in turn, the generalization error decreases only up to a point and when overfitting to the training sample starts to occur, we observe a monotonous increase in the testing error. In this case, this turning point takes place for $p=9$, a value that seems to offer the optimal tradeoff between the quality of training and out-of-sample performance. Another way to state this observation is that, with this noise level, the best polynomial approximation (in the mean square sense) of the the exponential function in the interval $[-1,1]$ is obtained when using polynomials of degree $9$.
\begin{centering}
\begin{figure}[!h]
\includegraphics[width=1.1\textwidth]{./paper_F1_1interval_overfit_training_v2.eps}
\caption{Conditional total training and testing errors committed by the regularized ($ \lambda =0.9$) polynomial approximation of the nonlinear underlying model \eqref{exampleB model2}-\eqref{f xi} as a function of the number of covariates $p$ or, equivalently, the degree of the polynomial approximation. The lengths of the training and testing samples are $N _1 =20$ and $N _2 = 20$, respectively, and $\zeta, \xi \sim \mathcal{U}(-1,1)$. }
\label{Figure 1}
\end{figure}
\end{centering}
\medskip
\noindent {\bf Numerical illustration. Case II. Generalization performance.} In this case we allow for the support of distribution for $\zeta $ used to generate the testing sample to be different from the one that defines the construction of the training sample. Indeed, we use $\xi \sim \mathcal{U}(-1,1)$ and $\zeta \sim \mathcal{U}(1,2)$ and hence the training and testing are carried out in the different intervals where $\xi $ and $\zeta $ are defined. Equivalently, in this situation the testing will be measuring the ability of the approximate model to reckon the values of the underlying model in an interval different from the one in which the training has taken place; that is why we explicitly talk in this case of generalization performance. Figure~\ref{Figure 2} shows the evolution of the conditional training and total errors for this case (for a given training covariates sample $X _1$ and a fixed regularization strength reported in the caption). Again, the conditional total training error is a decreasing function of $p$. As to the testing error in this case, its absolute value is much higher than in Case I and its minimum is already attained for $p=4$, that is, in this case a fourth order polynomial provides the best tradeoff in the modeling of the exponential function.
\begin{figure}[!h]
\includegraphics[width=1\textwidth]{./Figure2.png}\\
\caption{Conditional total training and testing errors committed by the regularized ($ \lambda =1$) polynomial approximation of the nonlinear underlying model \eqref{exampleB model2}-\eqref{f xi} as a function of the number of covariates $p$ or, equivalently, the degree of the polynomial approximation.
The lengths of the training and testing samples are $N _1 =30$ and $N _2 = 40$, respectively. The random variables $\xi \sim \mathcal{U}(-1,1)$ and $\zeta \sim \mathcal{U}(1,2)$ have been used in the dependent variables generating model.}
\label{Figure 2}
\end{figure}
\section{Appendices}
\label{Appendices}
\subsection{Proof of Theorem \ref{Theorem 1}}
\label{Proof of Theorem 1}
We first notice that by definition of $R _\lambda $ in \eqref{R lambda repeat}, namely $R _\lambda =(XA _N X ^\top +\lambda N \mathbb{I}_p )^{-1}$, the relation \eqref{B lambda random sample est} can be rewritten as $\widehat{B }_\lambda= R _\lambda X A_N Y ^\top$. In this expression we use that $Y=B _\lambda ^\top X + E_\lambda $ and that, by the same definition of $R _\lambda $ in \eqref{R lambda repeat}, $XAX ^\top = R _\lambda ^{-1} -\lambda N \mathbb{I}_p $. Taking both observations into account, we obtain that $\widehat{B }_\lambda= R _\lambda X A_N Y ^\top = R _\lambda X A_N ( X ^\top B_\lambda + E _\lambda ^\top) = R _\lambda X A_N X ^\top B_\lambda + R _\lambda X A_N E _\lambda ^\top = B _\lambda - \lambda N R _\lambda B _\lambda + R _\lambda X A _N E _\lambda ^\top$ or, equivalently,
\begin{equation}
\label{}
\widehat{B} _\lambda - B _\lambda + \lambda N R _\lambda B_\lambda = R _\lambda X A _N E_\lambda ^\top.
\end{equation}
We now point out that the conditional homoscedasticity hypothesis on the residuals implies by Corollary~\ref{Corollary Lemma 2} that $E _{\lambda} |X \sim {\rm MN} ( M_{ E _\lambda | X}, \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda, \mathbb{I}_N )$ and hence by Theorem 2.3.10 in \cite{bookMatrixDistributions2000} we have that
\begin{equation}
\label{}
(\widehat{B} _\lambda - B _\lambda + \lambda N R _\lambda B_\lambda)|X \sim {\rm MN} ( R _\lambda X A _N M_{ E _\lambda | X} ^\top, R_\lambda X A _N A _N ^\top X ^\top R_\lambda , \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda).
\end{equation}
Since $A _N A _N ^\top = A_N $, the statement in \eqref{B lambda hat minus B lambda} follows. $\blacksquare$
\subsection{Proof of Lemma \ref{Lemma 1}}
\label{Proof of Lemma 1}
\noindent {\bf Part (i)}. By \eqref{B lambda random sample est} we have that $\widehat{B} _\lambda :=(XA_NX^\top + \lambda N \mathbb{I}_p ) ^{-1} XA_NY^\top$. At the same time the relation \eqref{B lambda0 random sample est} yields $XA_NY^\top = (XA_N X ^\top ) \widehat{B}$, which in \eqref{B lambda random sample est} gives $\widehat{B} _\lambda :=(XA_NX^\top + \lambda N \mathbb{I}_p ) ^{-1} (XA_N X ^\top ) \widehat{B}$. The last expression is equivalent to ~\eqref{B lambda ZB}, as required.
\noindent{\bf Part (ii)}. First, in order to establish \eqref{Z1}, we notice that from the definition of $R _\lambda $ in \eqref{R lambda}, it follows that $XA_N X ^\top = R _{\lambda} ^{-1} - \lambda N \mathbb{I} _p$ and substituting this relation into \eqref{Z lambda} in {\bf (i)} yields
\begin{align*}
Z _\lambda = R _{\lambda} XA_N X ^\top = R _{\lambda} \left( R _{\lambda} ^{-1} - \lambda N \mathbb{I} _p \right) = \mathbb{I}_p - \lambda N R _\lambda,
\end{align*}
as required. Second, we show that \eqref{Z2} holds. Rewriting the relation \eqref{Z lambda} using the definition of $R _\lambda $ in \eqref{R lambda}, we obtain that
\begin{align*}
Z _\lambda =\left((XA_N X ^\top) ^{-1} XA_N X ^\top + \lambda N (XA_N X ^\top) ^{-1} \right) ^{-1} = \left( \mathbb{I} _p + \lambda N (XA_N X ^\top) ^{-1} \right) ^{-1},
\end{align*}
we establish that it is equivalent to \eqref{Z2}, as required. $\blacksquare$
\subsection{Proof of Corollary \ref{Corollary Theorem 1}}
\label{Proof of Corollary of Theorem 1}
As $\widehat{B} _\lambda - B = (\widehat{B}_\lambda - B_\lambda ) + ({B}_\lambda - B )$, then from \eqref{B lambda hat minus B lambda} in Theorem \ref{Theorem 1} follows that
\begin{align}
\label{B lambda minus B proof}
(\widehat{B} _\lambda - B )|X \sim {\rm MN} (B _\lambda - B-\lambda N R _\lambda B_\lambda + R_\lambda X A _N M_{ E _\lambda | X} ^\top, R_\lambda X A _N X ^\top R_\lambda ,\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda).
\end{align}
First, notice that by \eqref{mu eps lambda cond x example} and by \eqref{mu E lambda cond X} we can conclude that $M_{ E _\lambda | X} ^\top = X ^\top (B - B _\lambda )$. Consequently,
\begin{align}
\label{mean proof B lambda hat minus B}
B _\lambda - B -\lambda N R _\lambda B_\lambda + R _\lambda X A_N X ^\top \left(B - B _\lambda \right)&= B _\lambda - B - \lambda N R _\lambda B _\lambda + R _\lambda (R _\lambda ^{-1} -\lambda N \mathbb{I}_p ) \left( B - B _\lambda \right) \nonumber \\
&= B _\lambda - B - \lambda N R _\lambda B_\lambda + B - B _\lambda -\lambda N R _\lambda B +\lambda N R _\lambda B _\lambda \nonumber \\
&= - \lambda N R_\lambda B,
\end{align}
where we used the definition of $R_\lambda $ in \eqref{R lambda repeat}. Second, we rewrite the first scale matrix in \eqref{B lambda minus B proof} as follows
\begin{align}
\label{RXAR proof}
R _\lambda XA_N X ^\top R _\lambda &= R_\lambda (R _\lambda^{-1} - \lambda N \mathbb{I}_p ) R _\lambda = \left( \mathbb{I} _p - \lambda N R _\lambda \right) R_\lambda = Z _\lambda R_\lambda.
\end{align}
At the same time the second scale matrix in \eqref{B lambda minus B proof} in the conditions of the statement is equal to $\Sigma _{\boldsymbol{\varepsilon} }$ and hence together with \eqref{mean proof B lambda hat minus B} and \eqref{RXAR proof} this yields \eqref{B lambda minus B our notation}.
Now, in order to show \eqref{B lambda minus B}, we first use the definition of $Z _\lambda $ in \eqref{Z1} and rewrite \eqref{RXAR proof} as
\begin{align}
R _\lambda XA_N X ^\top R _\lambda&= Z _\lambda R_\lambda = \dfrac{1}{ \lambda N} Z_\lambda \left( \mathbb{I} _p - Z _\lambda \right) = \dfrac{1}{ \lambda N} \left( Z _\lambda - Z _\lambda ^2 \right)=\dfrac{1}{\lambda N } Z_\lambda\left( Z _\lambda ^{-1} - \mathbb{I} _p\right) Z _\lambda.
\end{align}
As by hypothesis $X A _N X ^\top$ is invertible, then by \eqref{Z2} it holds that $(X A _N X ^\top )^{-1} = \dfrac{1}{ \lambda N}\left( Z _\lambda ^{-1} - \mathbb{I}_ p \right) $ and the expression \eqref{RXAR proof} becomes
\begin{align}
\label{RXAR proof2}
R _\lambda XA_N X ^\top R _\lambda = Z _\lambda(X A _N X ^\top )^{-1} Z _\lambda.
\end{align}
Consequently, the relations \eqref{B lambda minus B proof} and \eqref{mean proof B lambda hat minus B} together with \eqref{RXAR proof2} guarantee \eqref{B lambda minus B}, as required. $\blacksquare$
\subsection{Proof of Theorem \ref{Theorem 2}}
\label{Proof of Theorem 2} We start by using the definition of the conditional total training error provided in \eqref{MSE training}. First, we define $D_\lambda := \widehat{B} _\lambda - B _\lambda $ and subtract and add $ B _\lambda $ to $\widehat{B} _\lambda $ in \eqref{MSE training} which results in
\begin{align}
\label{proof mse training}
{\rm MSE}_{{\rm training}|X}^\lambda &= \dfrac{1}{N} {\rm trace}\left( {\rm E} _X\left[ (Y - (\widehat{B} _\lambda - B _\lambda + B _\lambda ) ^\top X) ^\top (Y - (\widehat{B} _\lambda - B _\lambda + B _\lambda )X)\right] \right) \nonumber \\
&= \dfrac{1}{N} {\rm trace}\left( {\rm E} _X\left[ (E _\lambda - (\widehat{B} _\lambda - B _\lambda ) ^\top X) ^\top (E _\lambda - (\widehat{B} _\lambda - B _\lambda ) ^\top X)\right] \right)\nonumber \\
&= \dfrac{1}{N} {\rm trace}\left( {\rm E} _X\left[ (E _\lambda - D _\lambda ^\top X) ^\top (E _\lambda - D _\lambda ^\top X)\right] \right)= \dfrac{1}{N} {\rm trace} ({\rm E} _X\left[ E _\lambda^\top E_\lambda \right] )\nonumber \\
& + \dfrac{1}{N} {\rm trace} (X ^\top {\rm E}_X \left[ D_\lambda D _\lambda ^\top\right] X ) - \dfrac{2}{N} {\rm trace}\left( X^\top {\rm E} _X\left[ D _\lambda E _\lambda \right] \right).
\end{align}
We now study separately the three summands in \eqref{proof mse training}. The first one can be computed in a straightforward way. Indeed, by {\bf (i)} in Lemma \ref{Lemma 2} we can immediately write
\begin{equation}
\label{proof trace ee x}
\dfrac{1}{N} {\rm trace} ({\rm E} _X\left[ E _\lambda ^\top E _\lambda \right] ) = {\rm trace} (\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda ) + \dfrac{1}{N} {\rm trace} (M _{E_\lambda |X} ^\top M _{E_\lambda |X}).
\end{equation}
As for the second summand, we notice first that, in the hypotheses of the theorem, the properties of the estimator $\widehat{B} _\lambda$ are as provided in Theorem~\ref{Theorem 1}, that is,
\begin{equation}
\label{B lambda hat minus B lambda repeat}
D_\lambda|X \sim {\rm MN} (M _{D _\lambda }, R_\lambda X A _N X ^\top R_\lambda ,\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda),
\end{equation}
with
\begin{equation}
\label{mu M proof}
M _{D _\lambda }:=-\lambda N R _\lambda B_\lambda + R_\lambda X A _N M_{ E _\lambda | X} ^\top
\end{equation}
and
\begin{equation}
\label{R lambda repeat proof}
R _\lambda :=(XA _N X ^\top +\lambda N \mathbb{I}_p )^{-1}.
\end{equation}
At the same time, by Theorem 2.3.10 in \cite{bookMatrixDistributions2000} (see also Lemma 6.3 in \cite{RC3}) we have that $D _\lambda$ satisfies the following relation
\begin{align}
\label{covRowproperty}
{\rm E}_X \left[ D _\lambda D_\lambda ^\top\right] & = {\rm trace}(\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda)R_\lambda X A _N X ^\top R_\lambda + M _{D _\lambda }M _{D _\lambda }^\top.
\end{align}
Using this expression, we see that the second summand in \eqref{proof mse training} can be written as
\begin{align}
\label{proof trace mlml x}
\dfrac{1}{N} {\rm trace}\left( X ^\top {\rm E}_X \left[ D _\lambda D_\lambda ^\top\right] X \right) &= \dfrac{1}{N} {\rm trace}\left( {\rm E}_X \left[ D _\lambda D_\lambda ^\top\right] XX ^\top \right) \nonumber\\
&= \dfrac{1}{N} {\rm trace}\Bigg[ \left( {\rm trace}( \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda) R _\lambda X A _N X ^\top R_\lambda + M _{D _\lambda }M _{D _\lambda }^\top \right) X X ^\top\Bigg]\nonumber\\
&=\dfrac{1}{N} {\rm trace}( \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda)Z _\lambda R _\lambda XX^\top +\dfrac{1}{N}{\rm trace}(M _{D _\lambda }M _{D _\lambda }^\top X X ^\top),
\end{align}
where \begin{equation}
\label{Z lambda repeat}
Z _\lambda := R _\lambda X A _N X^\top.
\end{equation}
Finally, we can rewrite the third summand in \eqref{proof mse training} by using the definition of the ridge regression matrix estimator $\widehat{B} _\lambda $ in \eqref{B lambda random sample est} and using $R_\lambda $ in \eqref{R lambda repeat proof}:
\begin{align*}
X ^\top {\rm E}_X \left[ D _\lambda E _\lambda \right] = & X ^\top {\rm E}_X \left[ \left(\widehat{B} _\lambda -B _\lambda \right) E_\lambda \right] = X ^\top {\rm E} _X\left[ \widehat {B} _\lambda E_\lambda \right] - X ^\top B_\lambda {\rm E} _X\left[ E _\lambda \right]\\
=&X^\top R _\lambda X A _N {\rm E}_X \left[ Y^\top E _\lambda \right] - X^\top B _\lambda {\rm E} _X\left[ E_\lambda \right] =X ^\top R _\lambda X_N A _N {\rm E} _X\left[ X ^\top B _\lambda E _\lambda + E_\lambda ^\top E _\lambda \right] \\
&- X^\top B _\lambda {\rm E} _X\left[ E_\lambda \right] =X ^\top R _\lambda X A_N (X ^\top B _\lambda {\rm E} _X\left[ E_\lambda \right] + {\rm E} _X\left[ E_\lambda^\top E_\lambda \right])- X^\top B _\lambda {\rm E} _X\left[ E_\lambda \right].
\end{align*}
In this expression we now use the part {\bf (i)} in Lemma \ref{Lemma 2} and hence conclude that
\begin{align}
\label{proof E xmle x}
-\dfrac{2}{N}{\rm trace} \left({\rm E}_X \left[ X ^\top D _\lambda E _\lambda \right] \right)=&-\dfrac{2}{N}{\rm trace}\Big( X ^\top R _\lambda X A_N (X ^\top B _\lambda M _{E_\lambda |X} + {\rm trace}( \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda ) \mathbb{I} _N + M _{E_\lambda |X} ^\top M _{E_\lambda |X} ) \nonumber\\
&- X^\top B _\lambda M _{E_\lambda |X} \Big) = -\dfrac{2}{N}\Big({\rm trace}( \Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda) {\rm trace} (Z _\lambda ) \nonumber\\
&+ {\rm trace} ( X ^\top R _\lambda X A_N M _{E_\lambda |X} ^\top M _{E_\lambda |X} ) + {\rm trace}( \left(X ^\top R_\lambda X A _N - \mathbb{I} _N \right) X ^\top
B _\lambda M _{E_\lambda |X} )\Big).
\end{align}
Using \eqref{proof trace ee x}, \eqref{proof trace mlml x}, and \eqref{proof E xmle x} in \eqref{proof mse training} yields the expression \eqref{MSE longer}. At the same time, in \eqref{MSE longer} the last term can be rewritten as
\begin{align}
\label{third term shorter}
-\dfrac{2}{N} \ {\rm trace}\left( (X ^\top R _\lambda X A _N -\mathbb{I}_N ) X^\top B _\lambda M _{E_\lambda |X} \right) &= -\dfrac{2}{N} \ {\rm trace}\left( X ^\top ( Z _\lambda -\mathbb{I}_p ) B _\lambda M _{E_\lambda |X} \right) \nonumber\\
&= {2}\lambda{\rm trace}( X ^\top R _\lambda B _\lambda M _{E_\lambda |X} ) ,
\end{align}
by using \eqref{Z1}. The substitution of this expression in
\eqref{MSE longer} provides the relation \eqref{MSE shorter}.
$\blacksquare$
\subsection{Proof of Corollary \ref{Corollary error Example A}}
\label{Proof of Corollary error Example A}
In order to prove \eqref{MSE training example A}, we just determine all the ingredients required for the evaluation of \eqref{MSE training} in the conditions of Example~A and we then show that the result follows. First, notice that by \eqref{sigma lambda x}
\begin{align}
\label{sigma lambda x repeat}
\Sigma ^\lambda_{\boldsymbol{\varepsilon} | \mathbf{x} }&= \Sigma _ { \boldsymbol{\varepsilon} }
\end{align}
and that by \eqref{mu eps lambda cond x example}
\begin{equation}
\label{mu eps lambda cond x example repeat}
M _{E_\lambda |X} = (B-B _\lambda ) ^\top X.
\end{equation}
Additionally,
\begin{align}
\label{}
M _{E_\lambda |X} ^\top M _{E_\lambda |X} = X ^\top (B-B _\lambda ) (B-B _\lambda ) ^\top X = X ^\top B B^\top X - X ^\top B B _\lambda ^\top X - X^\top B _\lambda B ^\top X + X ^\top B_\lambda B_\lambda ^\top X.
\end{align}
We now proceed by pointing out that $M _{(\widehat{B}_\lambda - B _\lambda ) } = M _{\widehat{B}_\lambda } + B _\lambda$ and by using that in the conditions of Example A, the mean matrix $M _{\widehat{B}_\lambda }$ is given by $M _{\widehat{B}_\lambda } = Z _\lambda B$, with $Z _\lambda $ as in \eqref{Z lambda B lambda}. We hence have that
\begin{align}M _{(\widehat{B}_\lambda - B _\lambda ) }&= B _\lambda + Z _\lambda B = B _\lambda + (\mathbb{I}_p - \lambda N R _\lambda ) B = B+B _\lambda - \lambda N R _\lambda B. \label{mu M example}
\end{align}
Additionally, we compute
\begin{align}
\label{mu M mu M t}
M _{(\widehat{B}_\lambda - B _\lambda ) }M _{(\widehat{B}_\lambda - B _\lambda ) } ^\top &= (B+B _\lambda - \lambda N R _\lambda B)(B+B _\lambda - \lambda N R _\lambda B)^\top =B B^\top - B B _\lambda^\top - B_\lambda B ^\top + B _\lambda B _\lambda^\top \nonumber\\
&- \lambda N B B ^\top R _\lambda - \lambda N R_\lambda B B ^\top +\lambda N R_\lambda B B _\lambda ^\top + \lambda N B _\lambda B ^\top R_\lambda + \lambda ^2 N ^2 R _\lambda B B ^\top R _\lambda.
\end{align}
We now substitute these expressions in the five summands that make formula \eqref{MSE shorter} and that we analyze one. First,
\begin{align}
\label{term1 example}
&{\rm trace} (\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda ) = {\rm trace} ( \Sigma _ { \boldsymbol{\varepsilon} } ),
\end{align}
\begin{align}
\label{term2 example}
\dfrac{1}{N} {\rm trace}(\Sigma _{ \boldsymbol{\varepsilon} | \mathbf{x} } ^ \lambda) {\rm trace}\left( Z _\lambda (R_\lambda X X ^\top - 2 \mathbb{I}_p )\right)= \dfrac{1}{N} {\rm trace}(\Sigma _ { \boldsymbol{\varepsilon} }) {\rm trace}\left( Z _\lambda (R_\lambda X X ^\top - 2 \mathbb{I}_p )\right),\end{align}
\begin{align}
\label{term3 example}
\dfrac{1}{N} {\rm trace} \left( M _{(\widehat{B}_\lambda - B _\lambda ) }M _{(\widehat{B}_\lambda - B _\lambda ) }^\top X X ^\top \right) =& \dfrac{1}{N} {\rm trace} ( (B B^\top - B B _\lambda^\top - B_\lambda B ^\top + B _\lambda B _\lambda^\top - \lambda N B B ^\top R _\lambda \nonumber\\
&- \lambda N R _\lambda B B ^\top+ \lambda N R_\lambda B B _\lambda ^\top + \lambda N B _\lambda B ^\top R_\lambda \nonumber\\
& + \lambda ^2 N ^2 R _\lambda B B ^\top _\lambda R _\lambda) X X ^\top ),\end{align}
\begin{align}
\label{term4 example}
- \dfrac{2}{N}{\rm trace}((X ^\top R _\lambda X A_N &- \dfrac{1}{2} \mathbb{I}_N ) M _{E_\lambda |X} ^\top M _{E_\lambda |X}) = - \dfrac{2}{N}{\rm trace}((X ^\top R _\lambda X A_N - \dfrac{1}{2} \mathbb{I}_N ) (X ^\top B B^\top X - X ^\top B B _\lambda ^\top X\nonumber\\
& - X^\top B _\lambda B ^\top X + X ^\top B_\lambda B_\lambda ^\top X)) = \dfrac{1}{N}{\rm trace} (( - B B ^\top + B B _\lambda ^\top + B _\lambda B ^\top - B _\lambda B _\lambda ^\top \nonumber\\
&+ 2 \lambda N R_\lambda B B ^\top - 2 \lambda N R _\lambda B B _\lambda ^\top - 2\lambda N R _\lambda B _\lambda B ^\top + 2 \lambda N R _\lambda B _\lambda B _\lambda ^\top )X X^\top ),\end{align}
and finally,
\begin{align}\label{term5 example}
{2}\lambda{\rm trace}\left( X ^\top R _\lambda B _\lambda M _{E_\lambda |X} \right) = {2}\lambda{\rm trace}\left( X ^\top R _\lambda B _\lambda (B-B _\lambda ) ^\top X \right) = {2}\lambda{\rm trace}\left( (R _\lambda B _\lambda B ^\top -R _\lambda B _\lambda B _\lambda^\top ) X X ^\top\right),
\end{align}
respectively.
Gathering the terms \eqref{term1 example}-\eqref{term5 example}, we obtain that
\begin{align*}
{\rm MSE}_{{\rm training}|X}^\lambda = &{\rm trace} ( \Sigma _ { \boldsymbol{\varepsilon} } ) + \dfrac{1}{N} {\rm trace}(\Sigma _ { \boldsymbol{\varepsilon} }) {\rm trace}\left( Z _\lambda (R_\lambda X X ^\top - 2 \mathbb{I}_p )\right) \nonumber\\
&+ \dfrac{1}{N} {\rm trace}( (B B^\top - B B _\lambda^\top - B_\lambda B ^\top + B _\lambda B _\lambda^\top - \lambda N B B ^\top R _\lambda - \lambda N R_\lambda B B ^\top \nonumber\\
&+ \lambda N B _\lambda B ^\top R_\lambda + \lambda N R_\lambda B B _\lambda^\top + \lambda ^2 N ^2 R _\lambda B B ^\top R_\lambda - B B ^\top + B B _\lambda ^\top + B _\lambda B ^\top - B _\lambda B _\lambda ^\top\nonumber\\
&+ 2 \lambda N R _\lambda B B ^\top - 2 \lambda N R _\lambda B B _\lambda ^\top - 2 \lambda N R _\lambda B _\lambda B ^\top + 2 \lambda N R _\lambda B _\lambda B _\lambda ^\top + 2 \lambda N R _\lambda B _\lambda B ^\top \nonumber\\&- 2\lambda N R _\lambda B _\lambda B_ \lambda ^\top ) X X ^\top)\nonumber\\
&= {\rm trace} ( \Sigma _ { \boldsymbol{\varepsilon} } ) + \dfrac{1}{N} {\rm trace}(\Sigma _ { \boldsymbol{\varepsilon} }) {\rm trace}\left( Z _\lambda (R_\lambda X X ^\top - 2 \mathbb{I}_p )\right) + \lambda ^2 N {\rm trace}( R_\lambda B B ^\top R_\lambda X X ^\top),
\end{align*}
which coincides with \eqref{MSE training example A}, as required. $\blacksquare$
\subsection{Proof of Theorem \ref{Theorem testing error}}
\label{Proof of Theorem testing error}
In order to prove \eqref{MSE test}, we first rewrite \eqref{MSE testing} as
\begin{align}
\label{MSE test proof}
{\rm MSE} ^\lambda _{{\rm testing}|X _1 } &= \dfrac{1}{N _2 } {\rm E}_{X _1} \left[ \left(Y _2 - \widehat {B} _\lambda ^\top X _2 \right) ^\top \left(Y _2 - \widehat {B} _\lambda ^\top X _2 \right) \right] \nonumber\\
&=\dfrac{1}{N _2 } \left[ {\rm trace} \left( {\rm E}_{X _1} \left[ Y _2 ^\top Y _2 \right] \right) - 2\, {\rm trace} \left( {\rm E}_{X _1} \left[ Y_2^\top \widehat{B} _\lambda ^\top X _2 \right] \right) + {\rm trace} \left( {\rm E} _{X _1}\left[ X _2 ^\top \widehat{B} _\lambda \widehat{B} _\lambda^\top X _2 \right] \right) \right].
\end{align}
In order to evaluate this expression, we study separately all its terms. We start with the first summand and use that the pairs of random samples $(X_1, Y_1)$ and $(X_2, Y_2)$ used for training and testing, respectively, are independent from each other. We hence rewrite the first term in \eqref{MSE test proof} as
\begin{align}
\label{term1}
\dfrac{1}{N _2 } {\rm trace} \left( {\rm E} _{X _1}\left[ Y _2 ^\top Y _2 \right] \right) = \dfrac{1}{N _2 } {\rm trace} \left( {\rm E} \left[ Y _2 ^\top Y _2 \right] \right) = {\rm trace} \left[ \Sigma ^{(2)} _{\mathbf{y} } +\boldsymbol{\mu} ^{(2)} _{\mathbf{y} } \boldsymbol{\mu} _{\mathbf{y} }^ {(2)\top} \right]
\end{align}
with \begin{align}
\label{}
\Sigma ^{(2)} _{\mathbf{y} } &= {\rm Cov} ( \mathbf{y} ^{(2)} ,\mathbf{y} ^{(2)} ),\\
\boldsymbol{\mu} ^{(2)} _{\mathbf{y} } &= {\rm E} \left[ \mathbf{y} ^{(2)}\right].
\end{align}
For the second summand we obtain
\begin{align}
\label{term2}
- \dfrac{2}{N_2} {\rm trace} \left( {\rm E}_{X _1} \left[ Y_2^\top \widehat{B} _\lambda ^\top X _2 \right] \right) &= - \dfrac{2}{N_2} {\rm trace} \left( {\rm E}_{X _1} \left[ X _2 Y_2^\top \widehat{B} _\lambda ^\top \right] \right) = - \dfrac{2}{N_2} {\rm trace} \left( {\rm E} \left[ X _2 Y_2^\top\right] {\rm E} _{X _1}\left[ \widehat{B} _\lambda ^\top \right] \right)\nonumber\\
&=- {2}{} {\rm trace} \left( \left( \Sigma ^{(2)}_{\mathbf{x} \mathbf{y} } + \boldsymbol{\mu} _{ \mathbf{x} }^{(2)}\boldsymbol{\mu} _{\mathbf{y} } ^{(2)\top}\right) M ^{ \top}_{ \widehat{B} _\lambda}\right),
\end{align}
with
\begin{align}
\label{mu B hat}
M _{ \widehat{B} _\lambda}&:=B _\lambda -\lambda N _1 R _\lambda B _\lambda + R _\lambda X _1 A _{N_1} M _{ E_\lambda | X} ^{(1) \top},\\
\Sigma ^{(2)} _{\mathbf{x}\mathbf{y} } &= {\rm Cov} ( \mathbf{x} ^{(2)} ,\mathbf{y} ^{(2)} ),\\
\boldsymbol{\mu} ^{(2)} _{\mathbf{x} } &= {\rm E} \left[ \mathbf{x} ^{(2)}\right].
\end{align}
In \eqref{term2} we used the properties of the ridge regression matrix estimator $\widehat{B} _\lambda$ provided in Theorem~\ref{Theorem 1}. Finally, we rewrite the third term as
\begin{align}
\label{term3}
\dfrac{1}{N_2}{\rm trace} \left( {\rm E}_{X _1} \left[ X _2 ^\top \widehat{B} _\lambda \widehat{B} _\lambda^\top X _2 \right] \right)&= \dfrac{1}{N_2}{\rm trace} \left( {\rm E} _{X _1}\left[ X _2 X _2 ^\top \widehat{B} _\lambda \widehat{B} _\lambda^\top \right] \right) =\dfrac{1}{N_2}{\rm trace} \left( {\rm E} \left[ X _2 X _2 ^\top \right] {\rm E} _{X _1}\left[ \widehat{B} _\lambda \widehat{B} _\lambda^\top \right] \right) \nonumber\\
&={\rm trace} \left[ \left( \Sigma^{(2)} _{\mathbf{x} } + \boldsymbol{\mu}^{(2)} _{\mathbf{x} } \boldsymbol{\mu} _{\mathbf{x} }^ {(2)\top} \right) \left({\rm trace} \left(\Sigma ^{\lambda, (1)}_{ \boldsymbol{\varepsilon} | \mathbf{x} }\right) Z _\lambda R _\lambda + M_{ \widehat{B} _\lambda}M ^{ \top}_{ \widehat{B} _\lambda} \right)\right],
\end{align}
with
\begin{align}
\label{}
\Sigma ^{(2)} _{\mathbf{x} } &= {\rm Cov} ( \mathbf{x} ^{(2)} ,\mathbf{x} ^{(2)} ),
\end{align}
and where $M _{ \widehat{B} _\lambda}$ is given in \eqref{mu B hat}. In this expression we again used Theorem~\ref{Theorem 1} and the property \eqref{covRowproperty}.
Substituting expressions \eqref{term1}-\eqref{term3} into \eqref{MSE test proof} immediately yields \eqref{MSE test}, as required. $\blacksquare$
\noindent
\addcontentsline{toc}{section}{Bibliography}
\bibliographystyle{wmaainf}
|
\section{Supplementary Material}
\setcounter{page}{1}
A. Feigel and A. Rozen, "Thermodynamic Emergence of a Brownian Motor".
\subsection{Calculation of velocity moments $<V^{n}>$}
Following\cite{Meurs2004}, here are the steps to calculate average moments of velocity $<V^{n}>$ in a 1D stochastic system. First, one derives transition rate probability $W(V,V')$ between velocities $V'\rightarrow V$. Second, probability $P(V,t)$ for velocity $V$ at time $t$ fits Boltzmann Master equation:
\begin{eqnarray}
\label{eq:34}
\frac{\partial P(V,t)}{\partial t} = &&\int W(V-\Delta V,\Delta V)P(V-\Delta V,t)d\Delta V - \\\nonumber
&&P(V,t)\int W(V,-\Delta V)d\Delta V,
\end{eqnarray}
where $\Delta V= V-V'$. Third, to calculate the moments:
\begin{eqnarray}
\label{eq:45}
<x^{n}> &=& \int x^{n}P(x,t)dx, x=\sqrt{\frac{M}{T^{eff}_{0}}}V,
\end{eqnarray}
eq. (\ref{eq:45}) is transformed using Kramers Moyal expansion:
\begin{eqnarray}
\label{eq:34}
\frac{\partial P(V,t)}{\partial t} = \sum_{n=1}^{\infty}\frac{(-1)^{n}}{n!}\frac{d^{n}}{dV^{n}}\left [a_{n}(V)P(V,t) \right ],
\end{eqnarray}
into a system of equations:
\begin{eqnarray}
\label{eq:35}
\partial <x>&=&<A_{1}(x)>,\\\nonumber
\partial <x^{2}>&=&2<xA_{1}(x)>+<A_{2}(x)>,\\\nonumber
\partial <x^{3}>&=&...,
\end{eqnarray}
where:
\begin{eqnarray}
\label{eq:47}
A_{n}&=&\left (\sqrt{\frac{M}{T^{eff}_{0}}} \right )^{n}a_{n},\\\nonumber
a_{n}(V) &=& \int \Delta V^{n}W(V,\Delta V)d\Delta V,
\end{eqnarray}
Forth, the system (\ref{eq:35}) is solved for $<x^{n}>$ under assumption of steady state $\partial<x^{n}>=0$. At this stage additional expansion by some small parameter is required to linearize equations (\ref{eq:35}) in $<x^{n}>$.
\subsection{Calculation of transition rate probability $W$}
\subsubsection{Single degree of freedom - translation}
Here is calculation of transition rate probability $W(V,V')$ from velocity $V'$ to any other velocity $V$ for a macroscopic body of general 2D convex shape in an ideal gas\cite{Meurs2004}. The body possesses mass $M$ and can move along axis $x$ only. The gas consists of identical particles of mass $m$ at density $\rho$ and temperature $T$. The velocities of gas particles $(v_{x},v_{y})$ follow Maxwell distribution. In the case body is coupled to the several thermal bathes the final $W$ is the sum over the thermal bathes.
This work highlights universality of $W$ for different systems. It allows further extension of the method to the systems with multiple degrees of freedom.
Consider collision of the body with a particle of the gas. In the frame of the reference of the body ($V'=0$), conservation of energy before and after the collision is:
\begin{eqnarray}
\label{eq:9}
&& -MV^2+\\\nonumber
&& m(v'_x-v_x)(v'_x+v_x)+m(v'_y-v_y)(v'_y+v_y)=0,
\end{eqnarray}
where $'$ indicates velocities before collision. Momentum conservation along x axis is:
\begin{eqnarray}
\label{eq:mom2}
-m\Delta v_x=MV,
\end{eqnarray}
where:
\begin{eqnarray}
\label{eq:mom1}
\Delta v_x=v_{x}-v'_{x},
\end{eqnarray}
In addition, in the case of instantaneous collision, momentum of the gas particle conserves along surface of body:
\begin{eqnarray}
\label{eq:momsurf}
v'_x\cos\theta+v'_y\sin\theta = v_x\cos\theta+v_y\sin\theta,
\end{eqnarray}
where $\theta$ is the angle between surface of the body and axis $x$. In addition, it is convenient to express $\Delta v_{y}=v_{y}-v'_{y}$ as a function of $V$:
\begin{eqnarray}
\label{eq:21}
\Delta v_{y}=\frac{M}{m}\frac{V}{\tan{\theta}}.
\end{eqnarray}
It follows from (\ref{eq:9}) and (\ref{eq:momsurf}) is:
An important consequence of the conservation laws (\ref{eq:mom2}) and (\ref{eq:21}) is:
\begin{eqnarray}
\label{eq:28}
\Delta v_{x}^{2}+\Delta v_{y}^{2}=\frac{M^{2}}{m^{2}}\frac{V^{2}}{\sin^{2}{\theta}},
\end{eqnarray}
In this expression we will see that geometric factor in the right part of this expression universally defines linear viscous coefficient for the corresponding degree of freedom.
To find velocity of the body $V$ after collision as a function of a gas particle velocity prior the collision $(v'_{x},v'_{y})$ let us rewrite energy conservation (\ref{eq:9}) as:
\begin{eqnarray}
\label{eq:kin2}
&&-MV^2- m\Delta v_{x}^{2}-m\Delta v_{y}^{2}-\\\nonumber
&&2mv'_{x}\Delta v_{x}-2mv'_{y}\Delta v_{y}=0,
\end{eqnarray}
It can be rewritten further using (\ref{eq:mom2}) and expressing $\Delta v_{x}$ and $\Delta v_{y}$ as a functions of $V$:
\begin{eqnarray}
\label{eq:kinlin1}
-V-\frac{M}{m}\frac{V}{\Gamma^{2}_{V}(c)}+2v'_x-2v'_y\frac{1}{\tan\theta}=0,
\end{eqnarray}
where $\Gamma_{V}=\sin{\theta}$. The final formula is:
\begin{eqnarray}
\label{eq:Vlinfin}
V = \frac{2m/M\Gamma^{2}_{V}(c)}{1+m/M\Gamma^{2}_{V}(c)}\left [v'_x-v'_y\frac{1}{\tan\theta} \right ],
\end{eqnarray}
Transition from the frame of reference of the body to the laboratory frame of reference:
\begin{eqnarray}
\label{eq:Vlinfin2}
V = V'+\frac{2m/M\Gamma^{2}_{V}(c)}{1+m/M\Gamma^{2}_{V}(c)}\left [v'_x-V'-v'_y\frac{1}{\tan\theta} \right ],
\end{eqnarray}
requires addition of $-V'$ to the velocities along $x$ coordinate:
Taking into account (\ref{eq:Vlinfin2}), transition probability $V'\rightarrow V$ is:
\begin{eqnarray}
\label{eq:tr1}
&&dW(VV')=SF(\theta)d\theta\int_{-\infty}^{\infty} dv'_x\int_{-\infty}^{\infty}dv'_y\\\nonumber
&&H[(V'-v')e_\perp]\times (V'-v')e_\perp \rho \phi(v'_x,v'_y)\times\\\nonumber
&&\times\delta\left [ V -V'-\frac{2m/M\Gamma^{2}_{V}(c)}{1+m/M\Gamma^{2}_{V}(c)}\left [v'_x-V'-v'_y\frac{1}{\tan\theta} \right ] \right ],
\end{eqnarray}
where $H$ is Heaviside step function and $\phi$ is Maxwell distribution:
\begin{eqnarray}
\label{eq:max1}
\phi(v'_{x},v'_{y})=\frac{m}{2\pi T}\exp\left (\frac{-m(v_{x}^{\prime 2}+v_{y}^{\prime 2})}{2T} \right ),
\end{eqnarray}
The average is calculated using Hadamard Stratonovich transformation.
The result is:
\begin{eqnarray}
\label{eq:30}
&&W(V,V')=\frac{1}{4}\sum_{i}S_{i}\rho_{i}\sqrt{\frac{m}{2\pi T_{i}}}\\\nonumber
&&\left | \Delta V\right |H(\Delta V \Gamma(c))\int dcF(c)\Gamma_{bath}^{2}(c)\left (\frac{M}{m\Gamma_{body}^{2}(c)}+1\right )^{2}\\\nonumber
&&\exp\left [-\frac{m\Gamma_{bath}^{2}(c)\left ( V'+\frac{1}{2}\left[ \Delta V\left (\frac{M}{m\Gamma_{body}^{2}(c)}+1\right )\right ]\right )^{2}}{2T_{i}}\right ],
\end{eqnarray}
where $\Gamma_{body}\equiv\Gamma_{V}$ and $\Gamma_{bath}=\Gamma_{body}$ to fit detailed balance in equilibrium.
\subsubsection{$\Gamma_{bath}=\Gamma_{body}$ equality from detailed balance at thermal equilibrium}
Detailed balance at thermal equilibrium requires:
\begin{eqnarray}
\label{eq:8}
P^{eq}(V')W(V,V')=P^{eq}(-V)W(-V',-V),
\end{eqnarray}
where $P^{eq}(V)$ is distribution of velocities at thermal equilibrium. In the case of single translation degree of freedom:
\begin{eqnarray}
\label{eq:29}
P^{eq}(V)\propto\exp\left (-\frac{MV^{2}}{2T}\right ),
\end{eqnarray}
Detailed balance (\ref{eq:8}), taking into account (\ref{eq:30}) with $T_{i}=T_{j}$, reduces to:
\begin{eqnarray}
\nonumber
&&m\Gamma_{bath}^{2}\left (V'^{2}+V'(V-V')K+\frac{1}{4}(V-V')^{2}K^{2}\right )-\nonumber\\
&&m\Gamma_{bath}^{2}\left (V^{2}+V(V'-V)K+\frac{1}{4}(V-V')^{2}K^{2}\right )=\nonumber\\
&&MV'^{2}-MV^{2},
\label{eq:4}
\end{eqnarray}
where:
\begin{eqnarray}
\label{eq:5}
1-K=-\frac{M}{m\Gamma_{body}^{2}},
\end{eqnarray}
Finally, equality of geometric factors:
\begin{eqnarray}
\label{eq:6}
\Gamma^{2}_{bath}=\Gamma^{2}_{body},
\end{eqnarray}
follows from (\ref{eq:4}) and (\ref{eq:5}).
\subsubsection{Single degree of freedom - rotation case}
Conservation of energy, conservation for angular momentum are:
\begin{eqnarray}
\label{eq:rotsys1}
-I\Omega^2+m(v'_x-v_x)(v'_x+v_x)+m(v'_y-v_y)(v'_y+v_y),
\end{eqnarray}
\begin{eqnarray}
\label{eq:43}
I\Omega+m\overrightarrow{r}\times \overrightarrow{v}=m\overrightarrow{r}\times
\overrightarrow{v'},
\end{eqnarray}
Taking into account (\ref{eq:momsurf}) one gets:
\begin{eqnarray}
\label{eq:38}
\Delta v_{x}^{2}+\Delta v_{y}^{2}=\frac{I^{2}\Omega^{2}}{m^{2}(r_{x}\cos\theta+r_{y}\sin\theta)^{2}},
\end{eqnarray}
\begin{eqnarray}
\label{eq:42} -\Omega-\frac{I}{m}\Omega\frac{1}{\Gamma^{2}_{\Omega}(c)}+\frac{2v'_{x}}{r_{y}+r_{x}/\tan\theta}-\frac{2v'_{y}}{r_{y}\tan\theta+r_{x}}=0,
\end{eqnarray}
where:
\begin{eqnarray}
\label{eq:36}
&&\Gamma_{\Omega}=\overrightarrow{r}\hat{e}_\parallel=\nonumber\\
&&r_{x}\cos\theta+r_{y}\sin\theta,
\end{eqnarray}
and:
\begin{eqnarray}
\label{eq:ev1755}
&&\hat{e}_\perp=(\sin\theta,-\cos\theta),\\\nonumber
&&\hat{e}_\parallel=(\cos\theta,\sin\theta),
\end{eqnarray}
Equation (\ref{eq:42}) is analogous to (\ref{eq:42}). Consequently $W(\Omega',\Omega)$ follows from (\ref{eq:30}) by substitution $V\rightarrow\Omega$, $M\rightarrow I$ and $\Gamma_{V}\rightarrow\Gamma_{\Omega}$.
\subsubsection{Two degrees of freedom}
In the case of both rotation and translation degrees of freedom,
conservation laws become:
\begin{eqnarray}
\label{eq:rotsys12}
&&-MV^2-I\Omega^2+m(v'_x-v_x)(v'_x+v_x)+\\\nonumber
&&m(v'_y-v_y)(v'_y+v_y)=0,\\\nonumber
&&-v'_x\sin\phi+v'_y\cos\phi=-v_x\sin\phi+v_y\cos\phi,\\\nonumber
&&-m(v'_x-v_x)=MV,\\\nonumber
&&I\Omega+r\times v=r\times v',
\end{eqnarray}
In this case, the expressions (\ref{eq:28}) and (\ref{eq:38}) hold separately and imply:
\begin{eqnarray}
\label{eq:4076}
M^{2}V^{2}\frac{1}{\Gamma_{V}^{2}(c)}=I^{2}\Omega^{2}\frac{1}{\Gamma_{\Omega}^{2}(c)},
\end{eqnarray}
Using (\ref{eq:4076}), energy conservation law:
\begin{eqnarray}
\label{eq:41}
-MV^{2}-I\Omega^{2}-m\sum_{i}\Delta v_{i}^{2}+2m\sum_{i}v_{i}\Delta v_{i}=0
\end{eqnarray}
may be rewritten as:
\begin{eqnarray}
\label{eq:42} -V-\frac{M}{I}V\frac{\Gamma^{2}_{\Omega}(c)}{\Gamma^{2}_{V}(c)}-\frac{M}{m}V\frac{1}{\Gamma^{2}_{V}(c)}+\sum_{i}v_{i}f_{i}(c)=0,
\end{eqnarray}
and further reduced to:
\begin{eqnarray}
\label{eq:42} -V-\frac{M}{m}V\frac{\mathfrak{G}^{2}_{V}(c)}{\Gamma^{2}_{V}(c)}+\sum_{i}v'_{i}f_{i}(c)=0,
\end{eqnarray}
The same can be done for rotation coordinate:
\begin{eqnarray}
\label{eq:42} -\Omega-\frac{I}{m}\Omega\frac{\mathfrak{G}^{2}_{\Omega}}{\Gamma^{2}_{\Omega}(c)}+\sum_{i}v'_{i}g_{i}(c)=0,
\end{eqnarray}
\begin{eqnarray}
\label{eq:53879}
\mathfrak{G}^{2}_{V}=1+\frac{m}{I}\Gamma^{2}_{\Omega},
\end{eqnarray}
\begin{eqnarray}
\label{eq:5365}
\mathfrak{G}^{2}_{\Omega}=1+\frac{m}{M}\Gamma^{2}_{V},
\end{eqnarray}
Each degree of freedom possesses dynamics analogous to the single degree cases (\ref{eq:kinlin1}) and (\ref{eq:42}) with rescaled mass:
\begin{eqnarray}
\label{eq:46}
M(c) \rightarrow M\mathfrak{G}_{V}^{2},
\end{eqnarray}
and momentum:
\begin{eqnarray}
\label{eq:49}
I(c)\rightarrow I\mathfrak{G}_{\Omega}^{2},
\end{eqnarray}
correspondingly. The rescaling of $M$ and $I$ remains unaffected during derivation of transition rate probability (\ref{eq:30}) using independently (\ref{eq:42}) and (\ref{eq:42}). The requirement for detailed balance at thermal equilibrium also preserves. Consequently, final expression of $W$ for each degree of freedom take its single degree of freedom form (\ref{eq:30}) with rescaled $M$ (\ref{eq:46}) and $I$ (\ref{eq:49}), see (\ref{eq:30969}).
Transition to original frame of reference in the case of two degrees of freedom changes in (\ref{eq:Vlinfin2}) velocities $(v'_{x},v'_{y})$ to $(v'_{x}+S_{x},v'_{y}+S_{y})$, where $(S_{x},S_{y})$ are the velocity of the second degree of freedom. Transition rate probability (\ref{eq:tr1}) should be averaged over $(S_{x},S_{y})$. Distribution of $(S_{x},S_{y})$ is much more narrow than Maxwell distribution of the gas particles and absolute values $|(S_{x},S_{y})|<<|(v'_{x},v'_{y})|$ under assumption $m<<M$. In this case $(S_{x},S_{y})$ can be neglected in derivation of transition rate probability ~(\ref{eq:30}). The same is valid in the case of arbitrary degrees of freedom.
\subsubsection{General N degree of freedom case}
Consider macroscopic body with arbitrary degrees of freedom that is immersed in an a gas of particles with mass $m$. As a consequence of the conservation laws, for each degree of freedom $k$ of the body holds:
\begin{eqnarray}
\label{eq:40}
\sum_{i}\Delta v_{i}^{2}=\frac{M_{k}^{2}}{m^{2}}V_{k}^{2}\frac{1}{\Gamma_{k}^{2}(c)},
\end{eqnarray}
where the sum goes over all degrees of freedom of the gas particles, e.g see expressions (\ref{eq:28}) and (\ref{eq:38}). The mass $M_{k}$ and velocity $v_{k}$ define energy of the corresponding degree of freedom $\propto M_{k}V_{k}^{2}$. For instance in the case of rotation mass $M$ together with velocity $V$ are replaced by momentum $I$ and frequency $\Omega$ correspondingly. Parameter $c$ in a geometric model indicates point of interaction on contour of the body but in general indicates channel of interaction with thermal bath.
As a consequence of (\ref{eq:40}), for each two degrees of freedom $k$ and $l$ holds:
\begin{eqnarray}
\label{eq:407876}
M_{k}^{2}V_{k}^{2}\frac{1}{\Gamma_{k}^{2}(c)}=M_{l}^{2}V_{l}^{2}\frac{1}{\Gamma_{l}^{2}(c)},
\end{eqnarray}
This expression connects different degrees of freedom.
The main task is to show that geometric factor $\Gamma (c)$ indeed defines linear viscous coefficient and therefore universally linked with macroscopic measurable parameter. In the limit of thermal equilibrium expression (\ref{eq:40}) should converge to diffusion coefficient of the body in velocity space. This diffusion coefficient is:
\begin{eqnarray}
\label{eq:44}
D_{V}=\left <\frac{V^{2}}{\Delta t}\right >=\frac{T\gamma}{M}
\end{eqnarray}
as a consequence of fluctuation dissipation theorem\cite{Kubo1966}. Thus $\oint \Gamma^{2}\propto \gamma$ under assumption that $\sum \Delta v_{i}^{2}\propto T$ at thermal equilibrium. Indeed, $\oint \Gamma^{2}\propto \gamma$ in the cases of the Brownian motors with either translation or rotation degrees of freedom.
These expressions (\ref{eq:40}) and (\ref{eq:407876}) are very general because correct number of constraint always allow reduction of conservation laws to (\ref{eq:40}). For instance, if the surface of the body is not convex then particle of the gas can make several collisions during the impact and expression (\ref{eq:momsurf}) should be different. The expression of the type (\ref{eq:40}) still can be derived and and it connection with viscosity is preserved. The same is correct for arbitrary number of degrees of freedom.
\begin{eqnarray}
\label{eq:309837}
&&W(V_{\xi},\Delta V_{\xi})=\frac{1}{4}\sum_{i}S_{i}\rho_{i}\sqrt{\frac{m}{2\pi T_{i}}}\int dc F_{i}(c)\\\nonumber
&&\left | \Delta V_{\xi}\right |H\left (\Delta V_{\xi} \frac{\Gamma_{\xi,i}(c)}{ \mathfrak{G}_{\xi,i}(c)}\right )\Gamma_{\xi,i}^{2}(c)\left (\frac{M_{\xi} \mathfrak{G}^{2}_{\xi,i}(c)}{m\Gamma_{\xi,i}^{2}(c)}+1\right )^{2}\\\nonumber
&&\exp\left [-\frac{m\Gamma_{\xi,i}^{2}(c)\left ( V_{\xi}+\frac{1}{2}\left[ \Delta V_{\xi}\left (\frac{M_{\xi} \mathfrak{G}^{2}_{\xi,i}(c)}{m\Gamma_{\xi,i}^{2}(c)}+1\right )\right ]\right )^{2}}{2T_{i}}\right ]
\end{eqnarray}
where rescaling parameter:
\begin{eqnarray}
\label{eq:597965}
\mathfrak{G}^{2}_{\xi,i}(c)=1+\sum_{\xi'\neq \xi}\frac{m}{M_{\xi',i}}\Gamma^{2}_{\xi',i}(c),
\end{eqnarray}
and index $i$ goes over all thermal bathes.
\subsection{Rectification and Stability}
To analyze rectification of motion and its stability, one should calculate the average velocities $<V_{\xi}>$ for the relevant degrees of freedom $\xi$. For instance, in the case of a motor with two degrees of freedom, see Fig. \ref{fig1}, rectified motion $<V(\phi)>$ depends on orientation of the motor $\phi$. Consequently rotation $<\Omega(\phi)>$ defines stability of the orientations. Stable points are characterized by in flux of $<\Omega(\phi)>$, see Figs. \ref{fig2} and \ref{fig4}.
To calculate the moments $<V_{\xi}^{n}>$, Boltzmann equation (\ref{eq:34}) with transition rate probability (\ref{eq:309837}) are transformed by Kramers Moyal expansion (\ref{eq:35}). In addition equations are expanded using small parameter $m/M_{\xi}$. The first three moments then are:
\begin{eqnarray}
\label{eq:52}
&&\frac{\partial <V>}{\partial t} = \sum_{i}\rho_{i}S_{i}\sqrt{\frac{T_{i}}{m}}\\\nonumber
&&\left [ -\sqrt{\frac{T_{i}}{M}}\oint\frac{\Gamma_{V,i}(c)}{\mathfrak{G}^{2}_{V,i}(c)}\epsilon_{V}^{1}\right .\\\nonumber
&&-2\sqrt{\frac{2}{\pi}}\oint \frac{\Gamma_{V,i}^{2}(c)}{\mathfrak{G}_{V,i}^{2}(c)}<V>\epsilon_{V}^{2}\\\nonumber
&&\left . +\left ( \oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{4}}\sqrt{\frac{T_{i}}{M}}-\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{2}}\sqrt{\frac{M}{T_{i}}}<V^{2}>\right ) \epsilon_{V}^{3} \right ],
\end{eqnarray}
\begin{eqnarray}
\label{eq:524534}
&&\frac{\partial <V^{2}>}{\partial t} = \\\nonumber
&&\sum_{i} \rho_{i}S_{i}\sqrt{\frac{T_{i}}{m}} \left [ -2\sqrt{\frac{T_{i}}{M}}\oint\frac{\Gamma_{V,i}}{\mathfrak{G}^{2}_{V,i}}<V>\epsilon_{V}^{1}\right .\\\nonumber
&&\left .+4\sqrt{\frac{2}{\pi}}\left( \oint\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}_{V,i}^{4}}\frac{T_{i}}{M}-\oint\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}^{2}_{V,i}}<V^{2}>\right )\epsilon_{V}^{2} \right .\\\nonumber
&&\left . -2\left(-4 \oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{4}}\sqrt{\frac{T_{i}}{M}}<V>+\sqrt{\frac{M}{T_{i}}}\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}^{2}_{V,i}}<V^{3}>\right )\epsilon_{V}^{3} \right ],
\end{eqnarray}
\begin{eqnarray}
\label{eq:52479}
&&\frac{\partial <V^{3}>}{\partial t} = \\\nonumber
&&\sum_{i} \rho_{i}S_{i}\sqrt{\frac{T_{i}}{m}} \left [ -3\sqrt{\frac{T_{i}}{M}}\oint\frac{\Gamma_{V,i}}{\mathfrak{G}^{2}_{V,i}}<V^{2}>\epsilon_{V}^{1}\right .\\\nonumber
&&\left .+6\sqrt{\frac{2}{\pi}}\left( 2\oint\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}_{V,i}^{4}}\frac{T_{i}}{\sqrt{M}}<V>-\sqrt{M}\oint\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}^{2}_{V,i}}<V^{3}>\right )\epsilon_{V}^{2} \right ],
\end{eqnarray}
where $\epsilon_{V}=m/M$, index $i$ goes over the thermal bathes and the index $\xi$ is omitted.
Eqs. (\ref{eq:52}),~(\ref{eq:524534}) and (\ref{eq:52479}) converge to the corresponding results in the case of a system with single degree of freedom, if there is no affect of the other degrees of freedom $\mathfrak{G}=1$. In this case the terms with the first degree of $\Gamma$ vanish because $\oint\Gamma=0$.
In the case of two degrees of freedom $V$ and $\Omega$:
\begin{eqnarray}
\label{eq:53}
\mathfrak{G}^{2}_{V,i}=1+\frac{m}{I}\Gamma^{2}_{\Omega,i},
\end{eqnarray}
\begin{eqnarray}
\label{eq:536567}
\mathfrak{G}^{2}_{\Omega,i}=1+\frac{m}{M}\Gamma^{2}_{V,i},
\end{eqnarray}
In the limit $m<<M,m<<I$:
\begin{eqnarray}
\label{eq:59292}
\frac{1}{\mathfrak{G}^{2}_{V,i}}\approx 1-\frac{m}{I}\Gamma^{2}_{\Omega,i},
\end{eqnarray}
\begin{eqnarray}
\label{eq:511390}
\frac{1}{\mathfrak{G}^{4}_{V,i}}\approx 1-2\frac{m}{I}\Gamma^{2}_{\Omega,i},
\end{eqnarray}
Equations for $\Omega$ follow by change $V\rightarrow\Omega$, $M\rightarrow I$, $\Gamma_{V}\rightarrow\Gamma_{\Omega}$ and $\epsilon_{V}\rightarrow\epsilon_{\Omega}$, where $\epsilon_{\Omega}=m/I$.
Neglecting the time derivatives in (\ref{eq:52}),~(\ref{eq:524534}) and (\ref{eq:52479}) one gets:
\begin{eqnarray}
\label{eq:10} &&<V^{2}>=\frac{1}{M}\frac{\sum_{i}\rho_{i}S_{i}T_{i}^{\frac{3}{2}}\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}_{V,i}^{4}}}{\sum_{i}\rho_{i}S_{i}T^{\frac{1}{2}}\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}_{V,i}^{2}}}\\\nonumber
&&-\frac{1}{2\sqrt{M}}\sqrt{\frac{\pi}{2}}\frac{\sum_{i}\rho_{i}S_{i}T_{i}\oint\frac{\Gamma_{V,i}(c)}{\mathfrak{G}^{2}_{i}(c)}}{\sum_{i}\rho_{i}S_{i}T^{\frac{1}{2}}_{i}\oint \frac{\Gamma_{V,i}^{2}(c)}{\mathfrak{G}_{V,i}^{2}(c)}}\frac{<V>}{\epsilon_{V}}+\frac{1}{2}\sqrt{\frac{\pi}{2}}\times\\\nonumber
&&\frac{ \sum_{i}\rho_{i}S_{i}\left (4 T_{i}<V>\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{4}}-M<V^{3}>\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{2}}\right )}{ \sum_{i}\rho_{i}S_{i}T^{\frac{1}{2}}_{i}\oint \frac{\Gamma_{V,i}^{2}(c)}{\mathfrak{G}_{V,i}^{2}(c)}
}\frac{\epsilon_{V}}{\sqrt{M}},
\end{eqnarray}
For the further calculations it is convenient to define the leading term as:
\begin{eqnarray}
\label{eq:16} V^{2}_{0}=\frac{1}{M}\frac{\sum_{i}\rho_{i}S_{i}T_{i}^{\frac{3}{2}}\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}_{V,i}^{4}}}{\sum_{i}\rho_{i}S_{i}T^{\frac{1}{2}}\frac{\Gamma_{V,i}^{2}}{\mathfrak{G}_{V,i}^{2}}},
\end{eqnarray}
The third moment is:
\begin{eqnarray}
\label{eq:10760} <V^{3}>=<V>\frac{2\sum_{i}\rho_{i}S_{i}T^{3/2}_{i}\oint\frac{\Gamma^{2}_{V,i}}{\mathfrak{G}_{V,i}^{4}}}{M\sum_{i}\rho_{i}S_{i}\sqrt{T_{i}}\oint\frac{\Gamma^{2}_{V,i}}{\mathfrak{G}_{V,i}^{2}}},
\end{eqnarray}
It can be rewritten as:
\begin{eqnarray}
\label{eq:17}
<V^{3}>=2<V>V^{2}_{0}.
\end{eqnarray}
using (\ref{eq:16}).
The leading term of velocity is:
\begin{eqnarray}
\label{eq:1}
<V>_{\Gamma} = -\frac{1}{2\sqrt{M}}\sqrt{\frac{\pi}{2}}\frac{\sum_{i}\rho_{i}S_{i}T_{i}\oint\frac{\Gamma_{V,i}(c)}{\mathfrak{G}^{2}_{i}(c)}}{\sum_{i}\rho_{i}S_{i}T^{\frac{1}{2}}_{i}\oint \frac{\Gamma_{V,i}^{2}(c)}{\mathfrak{G}_{V,i}^{2}(c)}}\epsilon_{V}^{-1},
\end{eqnarray}
It is of the order $\epsilon_{\Omega}^{2}/\epsilon_{V}$ taking into account (\ref{eq:59292}) and $\oint\Gamma=0$.
The corresponding second moment is:
\begin{eqnarray}
\label{eq:19}
<V^{2}>_{\Gamma}=V^{2}_{0}+<V>_{\Gamma}^{2},
\end{eqnarray}
In the case $\int\Gamma/G^{2}=0$:
\begin{eqnarray}
\label{eq:2}
&&<V>_{\Gamma^{3}} =\\\nonumber
&&\frac{1}{2}\sqrt{\frac{\pi}{2}}\frac{ \sum_{i}\rho_{i}S_{i}\left ( T_{i}\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{4}}-MV_{0}^{2}\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{2}}\right )}{ \sum_{i}\rho_{i}S_{i}T^{\frac{1}{2}}_{i}\oint \frac{\Gamma_{V,i}^{2}(c)}{\mathfrak{G}_{V,i}^{2}(c)}
}\frac{\epsilon_{V}}{\sqrt{M}},
\end{eqnarray}
and:
\begin{eqnarray}
\label{eq:20}
&&<V^{2}>_{\Gamma^{3}} =V_{0}^{2}+\\\nonumber
&&\sqrt{\frac{\pi}{2}}\frac{ \sum_{i}\rho_{i}S_{i}\left (2 T_{i}\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{4}}-MV_{0}^{2}\oint\frac{\Gamma_{V,i}^{3}}{\mathfrak{G}_{V,i}^{2}}\right )}{ \sum_{i}\rho_{i}S_{i}T^{\frac{1}{2}}_{i}\oint \frac{\Gamma_{V,i}^{2}(c)}{\mathfrak{G}_{V,i}^{2}(c)}
}\frac{<V>_{\Gamma^{3}}\epsilon_{V}}{\sqrt{M}},
\end{eqnarray}
Expressions (\ref{eq:1}),~(\ref{eq:19}),~(\ref{eq:2}) and (\ref{eq:20}) are the main results of this work that describe mutual influence of degrees of freedom on each other.
In the case of two degrees of freedom and two thermal bathes, see Fig. \ref{fig1}:
\begin{eqnarray}
\label{eq:22}
<\Omega>_{\Gamma}=\frac{1}{2}\frac{\sqrt{m}}{M}\frac{S_{\bigtriangleup}\rho_{\bigtriangleup} T_{\bigtriangleup}^{1/2}\oint\Gamma_{\Omega,\bigtriangleup}\Gamma^{2}_{V,\bigtriangleup}}{\sum_{i=\bigtriangleup,\bigcirc}S_{i}\rho_{i} T_{i}^{1/2}\oint\Gamma^{2}_{\Omega,i}},
\end{eqnarray}
and:
\begin{eqnarray}
\label{eq:23} <\Omega>_{\Gamma^{3}}=-\sqrt{\frac{\pi}{8}}\frac{m^{3/2}}{MI}\frac{S_{\bigtriangleup}\rho_{\bigtriangleup}(2T_{\bigtriangleup}-T^{eff}_{\Omega})\oint\Gamma^{3}_{\Omega,\bigtriangleup}\Gamma^{2}_{V,\bigtriangleup}}{\sum_{i=\bigtriangleup,\bigcirc}S_{i}\rho_{i} T_{i}^{1/2}\oint\Gamma^{2}_{\Omega,i}}.
\end{eqnarray}
Both (\ref{eq:22}) and (\ref{eq:23}) depend on $\phi$ because contour integrals with $\Gamma$ depend on $\phi$.
Stability of specific orientation $\phi_{0}$ is defined by linear expansion of (\ref{eq:22}) or (\ref{eq:23}) in $\phi$ near $\phi_{0}$, see Figs. \ref{fig2} and \ref{fig4}. Stability in the case $\int\Gamma/G^{2}\neq 0$ depends only on the contour integral $\oint\Gamma_{\Omega,\bigtriangleup}\Gamma^{2}_{V,\bigtriangleup}$, because the temperature term in (\ref{eq:22}) is always positive. Orientation of the asymmetric part in this case is independent of the temperatures. If $\int\Gamma/G^{2}=0$ then:
\begin{eqnarray}
\label{eq:26}
<\Omega>_{\Gamma^{3}}\propto -(2T_{\bigtriangleup}-T^{eff}_{\Omega})\oint\Gamma^{3}_{\Omega,\bigtriangleup}\Gamma^{2}_{V,\bigtriangleup},
\end{eqnarray}
and, therefore, stability depends on the temperatures of the thermal bathes.
Fluctuations of the velocities (\ref{eq:20}) are:
\begin{eqnarray}
\label{eq:24}
<\Omega^{2}>_{\Gamma}=\Omega_{0}^{2}+<\Omega>^{2}_{\Gamma},
\end{eqnarray}
and
\begin{eqnarray}
\label{eq:25}
&&<\Omega^{2}>_{\Gamma^{3}}=\Omega_{0}^{2}-\\\nonumber
&&\sqrt{\frac{\pi}{2}}\sqrt{\frac{m}{MI}}\frac{S_{\bigtriangleup}\rho_{\bigtriangleup}(4T_{\bigtriangleup}-T^{eff}_{\Omega})<\Omega>_{\Gamma^{3}}\oint\Gamma^{3}_{\Omega,\bigtriangleup}\Gamma^{2}_{V,\bigtriangleup}}{\sum_{i=\bigtriangleup,\bigcirc}S_{i}\rho_{i}T_{i}^{1/2}\oint\Gamma^{2}_{\Omega,i}},
\end{eqnarray}
correspondingly. In this case (\ref{eq:24}), any stable point with $<\Omega>_{\Gamma}=0$ correspond to minimum fluctuations. In the case of (\ref{eq:25}):
\begin{eqnarray}
\label{eq:27}
&&<\Omega^{2}>_{\Gamma^{3}}=\Omega_{0}^{2}+\\\nonumber
&&A_{+}(4T_{\bigtriangleup}-T^{eff}_{\Omega})(2T_{\bigtriangleup}-T^{eff}_{\Omega})\left (\oint\Gamma^{3}_{\Omega,\bigtriangleup}\Gamma^{2}_{V,\bigtriangleup}\right )^{2},
\end{eqnarray}
where $A_{+}$ is a positive coefficient. According to (\ref{eq:27}), a stable point might correspond to minimum or maximum fluctuations as a function of the temperature of the thermal bathes.
Stability conditions $<\Omega>$ together with properties of $<\Omega^{2}>$ change at the temperature boundaries $(2T_{\bigtriangleup}-T^{eff}_{\Omega})=$ and $(4T_{\bigtriangleup}-T^{eff}_{\Omega})=0$, see Fig. \ref{fig3}. These equations reduce to:
\begin{eqnarray}
\label{eq:7}
c+2\frac{T_{\bigcirc}}{T_{\bigtriangleup}}-\left (\frac{T_{\bigcirc}}{T_{\bigtriangleup}}\right )^{3}=0,
\end{eqnarray}
and
\begin{eqnarray}
\label{eq:12}
3c+4\frac{T_{\bigcirc}}{T_{\bigtriangleup}}-\left (\frac{T_{\bigcirc}}{T_{\bigtriangleup}}\right )^{3}=0,
\end{eqnarray}
correspondingly. Coefficient $c$ is:
\begin{eqnarray}
\label{eq:15}
c = \frac{\rho_{\bigtriangleup} S_{\bigtriangleup} }{\rho_{\bigcirc} S_{\bigcirc} },
\end{eqnarray}
\onecolumngrid
\begin{widetext}
\begin{figure}
\begin{center}
\begin{tabular}{c}
\multicolumn{1}{l}{{\bf\sf A}}\\
\resizebox{1.0\textwidth}{!}{\includegraphics{fig4A.png}}\\
\multicolumn{1}{l}{{\bf\sf B}}\\
\resizebox{1.0\textwidth}{!}{\includegraphics{fig4B.png}}\\
\end{tabular}
\caption{Stability as a function of rotation axis position. (A) Axis of rotation is located near the base of the triangle. $\Gamma_{V}$ is constant along an edge of the triangle. Integral over the base is vanishes because of equal contribution of positive and negative $\Gamma_{\Omega}$. $\Gamma_{\Omega}$ changes sign at the intersection point of an edge and line perpendicular to the edge that crosses axis of rotation. Indeed collision with particle on the different sides relative to this point will rotate triangle to opposite directions. The size of $\Gamma$ indicates absolute contribution to the contour integral either as a consequence of $\Gamma$ value or integration span along the surface. The sign of the contribution is provided inside the parenthesis. Contour integrals $\oint\Gamma_{\Omega}\Gamma^{2}_{V}$ and $\oint\Gamma^{3}_{\Omega}\Gamma^{2}_{V}$ possess negative derivative due to $\phi$ at the point $\phi=-\pi/2$. This point therefore is stable because flux $<\Omega>$ directed towards it. No motion can emerge. (B) Axis of rotation is located near the sharp corner. In this case, contour integrals $\oint\Gamma_{\Omega}\Gamma^{2}_{V}$ and $\oint\Gamma^{3}_{\Omega}\Gamma^{2}_{V}$ possess positive derivative due to $\phi$. The point is unstable and emergence of motion is possible.}
\label{fig4}
\end{center}
\end{figure}
\end{widetext}
\end{document} |
\section{Introduction}
Thermoelectric materials enable the direct conversion from heat to electricity, which may make essential contributions
to the crisis of energy\cite{s1,s2}. A good thermoelectric material should have high dimensionless figure of merit $ZT=S^2\sigma T/(\kappa_e+\kappa_L)$, where S, $\sigma$, T, $\kappa_e$ and $\kappa_L$ are the Seebeck coefficient, electrical conductivity, absolute temperature, the electronic and lattice thermal conductivities, respectively. A high $ZT$ material requires high power factor ($S^2\sigma$) and low thermal conductivity ($\kappa=\kappa_e+\kappa_L$). They generally influence each other, and a counteracted relationship between electrical conductivity and thermal conductivity or Seebeck coefficient is often found.
The excellent classic thermoelectric materials include bismuth-tellurium systems\cite{s3,s4}, silicon-germanium alloys\cite{s5,s6}, lead chalcogenides\cite{s7,s8} and skutterudites\cite{s9,s10}. For thermoelectric research, searching for potential high $ZT$ materials with classic common lattice structure, like perovskite structure, is interesting and challenging.
The cubic perovskite structure $\mathrm{SrTiO_3}$ has attracted growing attention for thermoelectric power generation\cite{s17}. The $ZT$ value of undoped $\mathrm{SrTiO_3}$ is less than 0.5 due to its high thermal conductivity\cite{s18}, which can be reduced by introducing point defects\cite{s19,s20}. Searching other perovskites for more efficient thermoelectric applications is imperative and amusing.
Hybrid $\mathrm{AMX_3}$ perovskites (A=Cs, $\mathrm{CH_3NH_3}$; M=Sn, Pb; X=halide) have recently attracted a great
deal of attention for solar cell designs, which can realize up to 15\% energy conversion efficiencies\cite{s11,s12,s13,s14}.
A fair amount of theoretical works have been performed to investigate their electronic structures, phonons and optical properties\cite{t1,t2,t3,t4,t5,t6}.
The $\mathrm{CsMI_3}$ (M=Sn and Pb) of them under reasonable hydrostatic pressure can turn into three-dimensional topological insulators, which has been predicted by first-principles calculations\cite{s15,s16}.
Here, we report on the thermoelectric properties of cubic $\mathrm{CsMI_3}$ (M=Sn and Pb) in perovskite structures from a combination of first-principles calculations and semiclassical Boltzmann transport theory.
The SOC can produce obvious effects on power factor of many thermoelectric materials\cite{so1,so2,so3,so4,so5,gsd3,gsd4,so6}, so the SOC is included in our calculations of electronic part. Calculated results show SOC has a noteworthy reduced influence for n-type power factor. It is noteworthy that ultralow lattice thermal conductivities are attained for $\mathrm{CsMI_3}$ (M=Sn and Pb), and the corresponding lattice thermal conductivity at 300 K is 0.54 $\mathrm{W m^{-1} K^{-1}}$ and 0.25 $\mathrm{W m^{-1} K^{-1}}$, which is compared with lattice thermal conductivity of 0.23 $\mathrm{W m^{-1} K^{-1}}$ in SnSe with an unprecedented $ZT$ of 2.6 at 923 K\cite{zhao}. Finally, the dimensionless thermoelectric figure of merit $ZT$ is calculated with $\tau$=$10^{-14}$ s or $\tau$=$10^{-15}$ s , and can be up
to about 0.6 or 0.4 at high temperature by the optimized doping.
\begin{figure}
\includegraphics[width=5cm]{Fig1.eps}
\caption{(Color online) The crystal structure of $\mathrm{CsMI_3}$ (M=Sn and Pb). The largest blue ball represent Cs atom, the medium green balls X, and the smallest red balls I.}\label{st}
\end{figure}
The rest of the paper is organized as follows. In the next section, we shall
describe computational details. In the third section, we shall present the electronic structures and thermoelectric properties of $\mathrm{CsMI_3}$ (M=Sn and Pb). Finally, we shall give our discussions and conclusion in the fourth
section.
\begin{figure}
\includegraphics[width=8.0cm]{Fig2.eps}
\caption{The energy band structures of $\mathrm{CsMI_3}$ (M=Sn and Pb) using mBJ (Left) and mBJ+SOC (Right).}\label{band}
\end{figure}
\begin{figure*}
\includegraphics[width=15cm]{Fig3.eps}
\caption{(Color online) At room temperature, transport coefficients of $\mathrm{CsSnI_3}$ (Top panel) and $\mathrm{CsPbI_3}$ (Bottom panel) as a function of doping level (N): Seebeck coefficient S, electrical conductivity with respect to scattering time $\mathrm{\sigma/\tau}$ and power factor with respect to scattering time $\mathrm{S^2\sigma/\tau}$ calculated with mBJ (Black solid lines) and mBJ+SOC (Red dash lines). The N means electrons (minus value) or holes (positive value) per unit cell.}\label{t1}
\end{figure*}
\begin{figure*}
\includegraphics[width=14cm]{Fig4.eps}
\caption{(Color online) The power factor with respect to scattering time $\mathrm{S^2\sigma/\tau}$ and electronic thermal conductivity with respect to scattering time $\mathrm{\kappa_e/\tau}$ of $\mathrm{CsSnI_3}$ (Top panel) and $\mathrm{CsPbI_3}$ (Bottom panel) as a function of doping level with temperature being 200, 400, 600 , 800 and 1000 (unit: K) using mBJ+SOC. }\label{t2}
\end{figure*}
\section{Computational detail}
The electronic structures of $\mathrm{CsMI_3}$ (M=Sn and Pb) are performed using
a full-potential linearized augmented-plane-waves method
within the density functional theory (DFT) \cite{1}, as implemented in
the WIEN2k package\cite{2}. We employ Tran and Blaha's mBJ
exchange potential plus local-density approximation (LDA)
correlation potential for the
exchange-correlation potential \cite{4}, which has been known to produce
more accurate band gaps than LDA and GGA.
The SOC was included self-consistently \cite{10,11,12,so} due to containing heavy elements, which leads to band splitting, and produces important effects on power factor. We use 5000 k-points in the
first Brillouin zone for the self-consistent calculation, make harmonic expansion up to $\mathrm{l_{max} =10}$ in each of the atomic spheres, and set $\mathrm{R_{mt}*k_{max} = 8}$. The self-consistent calculations are
considered to be converged when the integration of the absolute
charge-density difference between the input and output electron
density is less than $0.0001|e|$ per formula unit, where $e$ is
the electron charge. Transport calculations, including Seebeck coefficient, electrical conductivity and electronic
thermal conductivity,
are performed through solving Boltzmann
transport equations within the constant
scattering time approximation (CSTA) as implemented in
BoltzTrap\cite{b}, and reliable results have
been obtained for several materials\cite{b1,b2,b3}. To
obtain accurate transport coefficients, we use 140000 k-points in the
first Brillouin zone for the energy band calculation. The lattice thermal conductivities are calculated within the linearized phonon Boltzmann equation
by using Phono3py+VASP codes\cite{pv1,pv2,pv3,pv4}. For the third-order force constants, 2$\times$2$\times$2 supercells
are built, and reciprocal
spaces of the supercells are sampled by 3$\times$3$\times$3 meshes. To compute lattice thermal conductivities, the
reciprocal spaces of the primitive cells are sampled using the 20$\times$20$\times$20 meshes.
\section{MAIN CALCULATED RESULTS AND ANALYSIS}
$\mathrm{CsMI_3}$ (M=Sn and Pb) belong to perovskite semiconductors, which consists of a network
of corner-sharing $\mathrm{MI_6}$ octahedra, and the schematic crystal structure is shown in \autoref{st}.
Based on the experimental structures, the electronic structures of $\mathrm{CsMI_3}$ (M=Sn and Pb) are investigated using mBJ and mBJ+SOC, and the energy band structures are plotted in \autoref{band}. Both mBJ and mBJ+SOC show $\mathrm{CsMI_3}$ (M=Sn and Pb) are direct-gap semiconductor, with the conduction band minimum (CBM) and valence band maximum (VBM) at the R point. The mBJ and mBJ+SOC energy band gap values are 0.52 eV (1.69 eV) and 0.17 eV (0.50 eV) for $\mathrm{CsSnI_3}$ ($\mathrm{CsPbI_3}$), respectively. The conduction bands are dominated by M-6p states, and the CBM is threefold degenerate at the absence of SOC. The VBM is a mixture of I-p and M-s states, which is nondegenerate. The SOC can remove band degeneracy, and leads to a spin-orbital splitting value of 0.43 eV and 1.49 eV for $\mathrm{CsSnI_3}$ and $\mathrm{CsPbI_3}$ at CBM. It is clearly seen that, near the Fermi level, the SOC has more obvious influence on the conduction bands than on the valence bands.
The related data are shown in \autoref{tab0}, and the mBJ gaps are larger than GGA or LDA ones, but are less than GW or HSE ones\cite{t1,t3,t4,t5,t6}.
\begin{table}[!htb]
\centering \caption{ The experimental lattice constant $a$ ($\mathrm{{\AA}}$); the calculated energy band gap values with mBJ $E_1$ (eV) and mBJ+SOC $E_2$ (eV); $E_1-E_2$ (eV); spin-orbit splitting $\Delta$ (eV) at the CBM.}\label{tab0}
\begin{tabular*}{0.48\textwidth}{@{\extracolsep{\fill}}cccccc}
\hline\hline
Name & $a$ & $E_1$ & $E_2$&$E_1-E_2$ &$\Delta$\\\hline\hline
$\mathrm{CsSnI_3}$&6.22 & 0.52 &0.17&0.35&0.43\\\hline
$\mathrm{CsPbI_3}$&6.29 &1.69 & 0.50& 1.19&1.49\\\hline\hline
\end{tabular*}
\end{table}
\begin{table*}[!htb]
\centering \caption{ Peak $ZT$ for both n- and p-type at 1000 K with $\tau$=$10^{-14}$ s and $\tau$=$10^{-15}$ s, and the corresponding doping concentrations. The doping concentration equals $\mathrm{4.16\times10^{21}cm^{-3}}$($\mathrm{4.02\times10^{21}cm^{-3}}$)$\times$ doping level for $\mathrm{CsSnI_3}$ ($\mathrm{CsPbI_3}$). }\label{tab}
\begin{tabular*}{0.96\textwidth}{@{\extracolsep{\fill}}ccccccccc}
\hline\hline
& & $\tau$=$10^{-14}$ s & & & & $\tau$=$10^{-15}$ s& &\\\hline
&n&&p&&n&&p&\\
Name&($\mathrm{\times10^{19}cm^{-3}}$)&$ZT$&($\mathrm{\times10^{19}cm^{-3}}$)&$ZT$&($\mathrm{\times10^{19}cm^{-3}}$)&$ZT$&($\mathrm{\times10^{19}cm^{-3}}$)&$ZT$\\\hline\hline
$\mathrm{CsSnI_3}$&4.16&0.63&1.08&0.36&9.78&0.49&1.88&0.19\\
$\mathrm{CsPbI_3}$&0.53&0.64&0.60&0.65&1.14&0.38&1.53&0.41\\\hline\hline
\end{tabular*}
\end{table*}
Next, we calculate semi-classic transport coefficients using CSTA Boltzmann theory.
The rigid band approach is used to mimic the doping effects by shifting the Fermi level, which is reasonable, when the doping
level is low\cite{tt9,tt10,tt11}.
The semi-classic transport coefficients, such as Seebeck coefficient S, electrical conductivity with respect to scattering time $\mathrm{\sigma/\tau}$ and power factor with respect to scattering time $\mathrm{S^2\sigma/\tau}$, as a function of doping level at room temperature using mBJ and mBJ+SOC are plotted in \autoref{t1}. Due to electrical thermal conductivity $\kappa_e$=$L\sigma T$ (Lorenz number $L$=$\pi^2k_B^2/3e^2$, where $K_B$ is the Boltzmann constant, e is the charge of an electron.), the electrical thermal conductivity has similar outlines with electrical conductivity.
The Fermi level moves into conduction bands, which means n-type doping (negative doping levels) with the negative Seebeck coefficient. The p-type doping (positive doping levels) with the positive Seebeck coefficient can be achieved by shifting Fermi level into the valence bands.
Although the Seebeck
coefficient is very large, when the Fermi level is in the middle of band gap, the low electrical conductivity leads to very small power factor. Shifting the Fermi level into conduction bands or valence bands, the Seebeck coefficient (absolute value)
decreases, while the electrical conductivity increases, which leads to a maximum of power factor at certain doping level.
\begin{figure}
\includegraphics[width=8cm]{Fig5.eps}
\caption{(Color online) The lattice thermal conductivities $\kappa_L$ of $\mathrm{CsMI_3}$ (M=Sn and Pb) as a function of temperature using GGA.}\label{t3}
\end{figure}
It has been proved that SOC has very important effects on power factor in many thermoelectric materials containing heavy elements\cite{so1,so2,so3,so4,so5,gsd3,gsd4,so6}. Calculated results show that SOC has obvious reduced influences on S and $\mathrm{\sigma/\tau}$ in n-type doping for both $\mathrm{CsSnI_3}$ and $\mathrm{CsPbI_3}$, but weak effects for p-type.
The large slope of density of states (DOS) near the energy band gap can induce a large Seebeck coefficient in narrow-gap semiconductors. This can be understood by the following formula: $S=\frac{\pi^2}{3}(\frac{k_B^2T}{e})[\frac{1}{n}\frac{dn(E)}{dE}+\frac{1}{\mu}\frac{d\mu(E)}{dE}]_{E=E_f}$\cite{z1},
where $n(E)$ and $\mu(E)$ are energy dependent carrier density and mobility, respectively.
It is found that the slope of DOS using mBJ near the energy band gap is larger than that using mBJ+SOC for conduction bands,
while they are nearly the same for valence bands. Band degeneracy, namely band convergence, can induce lager slope of DOS. The SOC can reduce the slope of DOS by removing band degeneracy. The SOC effects on energy bands can explain SOC influences on S.
The SOC-induced reduced S and $\mathrm{\sigma/\tau}$ for n-type lead to a very remarkable detrimental influence on power factor, especially for $\mathrm{CsPbI_3}$. However, the SOC has weak effects on p-type power factor.
For $\mathrm{CsPbI_3}$, at the absence of SOC, the best n-type power factor is larger than that in p-type doping. However, including SOC, it is opposite in considered doping range. Similar SOC-induced switch of best power factor between n-type and p-type can be achieved in $\mathrm{Mg_2Sn}$ \cite{gsd3}.
Therefore, including SOC is very important in the theoretical prediction of thermoelectric properties of $\mathrm{CsMI_3}$ (M=Sn and Pb).
\begin{figure*}
\includegraphics[width=14cm]{Fig6.eps}
\caption{(Color online) The $ZT$ of $\mathrm{CsSnI_3}$ (Top panel) and $\mathrm{CsPbI_3}$ (Bottom panel) as a function of doping level with temperature being 200, 400, 600 , 800 and 1000 (unit: K), and the scattering time $\mathrm{\tau}$ is 1 $\times$ $10^{-14}$ s (Left) and 1 $\times$ $10^{-15}$ s (Right). }\label{t4}
\end{figure*}
The power factor and electronic thermal conductivity with respect to scattering time ($\mathrm{S^2\sigma/\tau}$ and $\mathrm{\kappa_e/\tau}$) of $\mathrm{CsSnI_3}$ and $\mathrm{CsPbI_3}$ as a function of doping level with temperature from 200 K to 1000 K using mBJ+SOC are shown in \autoref{t2}. For $\mathrm{CsSnI_3}$, n-type doping has larger power factor than p-type doping, while p-type power factor is larger than n-type one for $\mathrm{CsPbI_3}$. if we assume scattering time is constant, in considered doping and temperature range, the best power factor of $\mathrm{CsSnI_3}$ is nearly four times larger than one of $\mathrm{CsPbI_3}$, and about two times larger for electronic thermal conductivity. The lattice thermal conductivity is important factor, which significantly affects thermoelectric performance. The lattice thermal conductivities $\kappa_L$ of $\mathrm{CsMI_3}$ (M=Sn and Pb) as a function of temperature are shown in \autoref{t3}. Calculated results show ultralow lattice thermal conductivities in $\mathrm{CsMI_3}$ (M=Sn and Pb). The lattice thermal conductivity can be assumed to have weak dependence on doping level, and typically goes as 1/T.
The corresponding room-temperature lattice thermal conductivity is 0.54 $\mathrm{W m^{-1} K^{-1}}$ and 0.25 $\mathrm{W m^{-1} K^{-1}}$ for $\mathrm{CsSnI_3}$ and $\mathrm{CsPbI_3}$. Theoretically, ultralow lattice thermal conductivities in many compounds have been predicted, such as $\mathrm{PbRbI_3}$ (0.10 $\mathrm{W m^{-1} K^{-1}}$), PbIBr (0.13 $\mathrm{W m^{-1} K^{-1}}$) and $\mathrm{K_2CdPb}$ (0.45 $\mathrm{W m^{-1} K^{-1}}$)\cite{ltc1}.
It is found that lattice thermal conductivity of $\mathrm{CsSnI_3}$
is nearly two times larger than one of $\mathrm{CsPbI_3}$.
Due to the complexity of various carrier scattering mechanisms, it is difficult to calculate scattering time $\tau$ from the first principles. To estimate thermoelectric conversion efficiency, the
thermoelectric figure of merit $ZT$ is calculated with hypothetical $\tau$=$10^{-14}$ and $\tau$=$10^{-15}$ s, and are plotted
in \autoref{t4}. The peak $ZT$ and corresponding doping concentrations for both n- and p-type at 1000 K are summarized in \autoref{tab}. For $\mathrm{CsSnI_3}$, the n-type doping has larger $ZT$ than p-type doping, which is mainly due to the larger n-type Seebeck coefficient. However, the $ZT$ between n- and p-type are nearly the same for $\mathrm{CsPbI_3}$ due to almost the same Seebeck coefficient. According to \autoref{t2} and \autoref{t3}, the total thermal conductivity $\mathrm{\kappa}$ is dominated by the lattice thermal conductivity $\mathrm{\kappa_L}$ in the very low doping level, but the electronic thermal conductivity $\mathrm{\kappa_e}$ becomes very larger than lattice thermal conductivity $\mathrm{\kappa_L}$ in slightly high doping region. These leads to very low doping concentration for peak $ZT$. Therefore, electronic thermal conductivity of $\mathrm{CsMI_3}$ (M=Sn and Pb) is a fatal disadvantage to gain more higher $ZT$ value.
\section{Discussions and Conclusion}
The CBM of $\mathrm{CsMI_3}$ (M=Sn and Pb) is dominated by a giant spin-orbit coupling (SOC), especially for $\mathrm{CsPbI_3}$. The SOC removes the band degeneracy of CBM by spin-orbit splitting, which leads to obvious effects on n-type Seebeck coefficient, and further affects the power factor. The larger spin-orbit splitting $\Delta$ leads to the more obvious detrimental influence on n-type power factor, which can be observed from \autoref{t2}. The similar SOC-induced detrimental influence on power factor has been observed in $\mathrm{Mg_2Sn}$ and half-Heusler $\mathrm{ANiB}$ (A=Ti, Hf, Sc, Y; B=Sn, Sb, Bi)\cite{so2,gsd3}. Therefore, it is very important for electronic part of thermoelectric properties of $\mathrm{CsMI_3}$ (M=Sn and Pb)
to include SOC.
$\mathrm{CsMI_3}$ (M=Sn and Pb) have been predicted to be three-dimensional topological insulators under reasonable hydrostatic pressure using a tight-binding analysis and first-principles calculations\cite{s15,s16}, which means that the electronic structures of $\mathrm{CsMI_3}$ (M=Sn and Pb) are easily tuned by pressure. Pressure-induced enhanced power factor has been predicted in $\mathrm{Mg_2Sn}$\cite{gsd3} and BiTeI\cite{gsd7} by the first-principles calculations. Experimentally, it is possible to tune the thermoelectric properties of $\mathrm{CsMI_3}$ (M=Sn and Pb) by pressure.
In summary, mBJ and mBJ+SOC are chosen to investigate electronic structures and electronic part of thermoelectric properties of halide perovskites $\mathrm{CsMI_3}$ (M=Sn and Pb). The strength of SOC influences on CBM is very large, especially for $\mathrm{CsPbI_3}$, which gives rise to obvious detrimental influence on n-type power factor.
The lattice thermal conductivities of $\mathrm{CsMI_3}$ (M=Sn and Pb) are performed with GGA, and ultralow lattice thermal conductivities are predicted, which is very key for providing high thermoelectric performance. At 1000 K, in low doping level, the figure of merit $ZT$ is up to about 0.6 with $\tau$=$10^{-14}$, and about 0.4 with $\tau$=$10^{-15}$.
The present work provides a platform to search potential thermoelectric materials from perovskite compounds.
\begin{acknowledgments}
This work is supported by the National Natural Science Foundation of China (Grant No. 11404391). We are grateful to the Advanced Analysis and Computation Center of CUMT for the award of CPU hours to accomplish this work.
\end{acknowledgments}
|
\section{Introduction}
Thermoresponsive polymers are an important class of materials which exhibit temperature dependent structural changes and find application in drug delivery,\cite{Hoare2009,Okano1995a} surface modification,\cite{Hoare2008} and self-assembled structures.\cite{Sun2013} Our interest lies in the family of thermoresponsive polymers which exhibit a lower critical solution temperature (LCST) in aqueous solutions. Along with the LCST, these polymers also exhibit a coil-to-globule transition at the single chain level. A well known example of such thermoresponsive polymers is Poly(N-isopropylacrylamide) (PNiPAM) which exhibits a LCST in water and a upper critical solution temperature (UCST) in solvents such as methanol, ethanol, dimethyl sulfoxide, acetone.\cite{Costa2002a, Bischofberger2014a} There have been several experimental \cite{Wang1998, Wu1998, Ono2006} and simulation \cite{Deshmukh2009a, Deshmukh2011, Alaghemandi2012a, Deshmukh2012a, Tucker2012a, Chiessi2010a} studies on the mechanism of the LCST behavior of PNiPAM in water. The tunability of the LCST for different applications has also been explored by studying its variation with additives such as salt,\cite{Zhang2005a, Du2010a} surfactant,\cite{Schild1991a, Shinde2001a, Mohan2007a} co-solvents,\cite{Zhang2002, Scherzinger2014, Mukherji2013, Mukherji2014, Mukherji2015, Walter2010a} and by the change in macromolecular architecture such as branching and tacticity. \cite{Ray2005a, Katsumoto2008a}
The origin of LCST in thermoresponsive polymer solutions is an important question in the field of polymer science. To understand this phase transition, there have been several attempts, ranging from the mean-field theory to atomistic molecular dynamic simulations,\cite{Deshmukh2009a, Du2010a, Deshmukh2011, Alaghemandi2012a, Deshmukh2012a} focusing on the LCST of PNiPAM-water system. A mean-field model was proposed by Okada and Tanka \cite{Okada2005} who hypothesized that preferential interaction among bound water molecules (cooperative hydration) controls the transition. The simulation results of Deshmukh \textit{et al.}\cite{Deshmukh2012a} indicate that the stability of bound water structure is the driving force for the transition. Though insightful, one should keep in mind that the results of these simulations are obtained for a particular polymer-solvent system, and these approaches require very specific interactions, extensive chemical details, or forcefield parameters. Schild \textit{et al.}\cite{Schild1991} have studied the LCST behavior of PNiPAM in water-alcohol mixtures using a combination of experiments and the three-component Flory-Huggins model. In their model, the interaction parameters have been partially taken from experimental data, and it does not give us a clear idea about the generic mechanisms that lead to this phenomena.\\
\indent In the experimental studies, a rich diversity of material systems based on PNiPAM have been explored to obtain physical insights related to the effect of substituents, copolymers, solvents, and additives. From the view point of the mechanism, Ono and Shikata\cite{Ono2006} have calculated the number of water molecules per monomer using high frequency dielectric relaxation measurements. Their results showed that the LCST is driven by the complete dehydration of the PNiPAM chains, showing the importance of bound water near the polymer. Bischofberger and coworkers \cite{Bischofberger2014, Bischofberger2014a} have performed turbidity and dynamic light scattering on the ternary system of PNiPAM, water, and alcohol. Their results indicate that the thermodynamic description of the solvent is more important than the specific description of local solvent structure.\\
\indent Given the multiplicity of systems that can exhibit LCST, it is highly pertinent to come up with a model that can exibit LCST broad physical principles. An approach aimed at identifying the minimal model that exhibits LCST will help to understand the relative importance of different contributions. Generic polymer models\cite{Anderson2006a, Marrink2007a, Mella2010a, Polson2005, Hatakeyama2007a} are suitable candidates for this kind of approach. The coarse-grained nature of these models allows us to qualitatively study the importance of the competition between entropy and internal energy without invoking to a specific polymer or solvent. In this paper, we develop generic polymer models with spherically symmetric solvent and monomeric beads for simulation and theoretical studies of a coil-to-globule transition. Our results indicate that the LCST depends on the competition between the mean interaction energy difference between the bulk and bound solvent, and entropy of bound solvent. We show that a coarse-grained representation of the solvent is sufficient to exhibit a LCST behavior. This indicates that the mean interaction energy difference between the bound solvent and bulk solvent is more important in comparison to the structural arrangement of the bound solvent. An important point to note is that this work is aimed at a generic understanding of the LCST behavior in thermoresponsive polymers, that does not refer to any particular polymer.\\
\indent The rest of the paper is organized as follows: in Sec.~\ref{sec:model}, we propose a polymer-solvent model for molecular dynamic simulation studies and introduce the solvophobic potential\cite{Kolomeisky1999} used in the theoretical approach. Section ~\ref{sec:results} presents the simulation results and numerical calculations of the theoretical model. In Sec.~\ref{sec:sum}, our findings will be summarized.
\section{Models}
\label{sec:model}
In this section, we discuss our models used for the simulation and theoretical studies. While simulations are carried out using a bead-spring model for polymers in an homogenous single component solvent, a phenomenological model is used for the theoretical analysis. Below we describe these models in detail.
\subsection{Generic polymer model with explicit solvent}
For the simulation studies, we model the polymer as a linear chain consisting of alternating solvophobic and amphiphilic beads ($N$ total beads, $N/2$ solvophobic, and $N/2$ amphiphilic beads). The motivation behind the presence of two different kinds of beads is to capture the behavior of the acrylamide family of thermoresponsive polymers in a generic manner (see Fig.~\ref{fig:mapping}). The methylene units along the backbone are analogous to hydrophobic beads. The substituted methylene units, most generally will have both hydrophilic and hydrophobic groups, and therefore analogous to amphiphilic beads. We emphasize that the intention is to use a generic model, without relating to any specific polymer system.
\begin{figure*}[h!]
\begin{center}
\includegraphics[scale=0.22]{mapping.eps}
\end{center}
\caption{Ad-hoc mapping of the acrylamide family of thermoresponsive polymers. $R_{\rm 1}$ and $R_{\rm 2}$ can be any arbitrary groups. The part of the monomer within the red box may have both hydrophilic and hydrophobic groups due to which it is modeled as an amphiphilic bead.}
\label{fig:mapping}
\end{figure*}
The amphiphilic bead has attractive interactions with both the solvent and the solvophobic beads. The interaction between the solvophobic bead and the solvent is purely repulsive. The solvent can be introduced either by including it explicitly or by incorporating its effects implicitly within the interaction potentials. Since coarse-grained potentials are obtained by integrating out the internal degrees of freedom of the unit, it is temperature dependent in general. When the scale of coarse graining is small, such a dependence is weak and can be neglected. However, when the solvent is implicit, the interaction potentials have a stronger dependence on the temperature, and the nature of this dependence has to be assumed a priori. To avoid a specifically assumed temperature dependence of the interaction potentials, we explicitly
incorporate the solvent. The potential energy for the system is given by the following expression
\begin{widetext}
\begin{eqnarray}\label{eq:energy_simu}
E=\sum_{i=1}^{N-1} k_{\rm b}(b_{i}-b_{i0})^{2} + \sum_{i=1}^{N_{\rm t}}\sum_{j>i} 4\epsilon_{ij}\Bigg[\left(\frac{\sigma}{r_{ij}}\right)^{12}-\left(\frac{\sigma}{r_{ij}}\right)^{6} - \left(\frac{\sigma}{r_{{\rm c},ij}}\right)^{12}+\left(\frac{\sigma}{r_{{\rm c}, ij}}\right)^{6}\Bigg],
\end{eqnarray}
\end{widetext}
where $N$ is the number of beads in the polymer chain, as mentioned before. $k_{\rm b}$ the force constant for the bonded interaction, $b_{i}$ the bond length between neighboring beads, $N_{\rm t}$ the total number of beads in the system (polymer + solvent), $b_{i0}$ the equilibrium bond length and $r_{ij}$ the distance between two non-bonded beads. The second term is the Shifted Lennard Jones (SLJ) potential with $r_{{\rm c},ij}$ being the cutoff distance at which the potential is truncated and shifted to zero. The above form of SLJ potential ensures that all the beads are spherically symmetric and have size $\sigma$. All the interaction parameters are kept independent of the temperature. We define dimensionless quantities as $\overline{r}_{ij}=r_{ij}/\sigma$, $\overline{\epsilon}_{ij}=\epsilon_{ij}/\epsilon_{\rm ss}$, $\overline{k}_{\rm b}=\sigma^{2}k_{\rm b}/\epsilon_{\rm ss}$, $\overline{b}_{i0}=b_{i0}/\sigma$, $\overline{T}=k_{\rm B}T/\epsilon_{\rm ss}$, $\overline{P}=\sigma^{3}P/\epsilon_{\rm ss}$ and $\overline{t}= t\sqrt{\epsilon_{\rm ss}/(m\sigma^2)}$, where $\epsilon_{\rm ss}$ is the potential energy of interaction between two solvent beads. We fix the values to $\overline{b}_{i0}=1$ and $\overline{k}_{\rm b}=200$ for all the simulations. The values of the other interaction parameters are listed in Table~\ref{table:1}.
\begin{table}[h]
\caption{Interaction parameters of the SLJ potential. Amphiphilic, solvophobic and solvent are represented by A, H and S, respectively.}
\begin{tabular}{c c c c c c c c }
\hline
$ij$ \ \ \ \ & AA\ \ \ \ &HH\ \ \ \ &SS\ \ \ \ &AH\ \ \ \ &HS\ \ \ \ &AS\\
\hline
$\overline{\epsilon}_{ij}\ \ \ \ \ $&1\ \ \ \ &1\ \ \ \ &1\ \ \ \ &1\ \ \ \ &1\ \ \ \ & 1.4, 1.7, 1.8, 2.0\\
$\overline{r}_{{\rm c},ij}\ \ \ \ \ $&2.5\ \ \ \ &2.5\ \ \ \ &2.5\ \ \ \ &2.5\ \ \ \ &$2^{1/6}$\ \ \ \ &2.5\\
\hline
\end{tabular}
\label{table:1}
\end{table}
Molecular dynamic simulations were performed in an NPT ensemble using the Nose-Hoover thermostat for different temperatures at a constant pressure $\overline{P}=0.002$. The trajectories were generated using the Velocity-Verlet algorithm with a time-step $\Delta \overline{t}= 0.004$. The ratio of the number of polymer beads to the solvent beads was maintained at 0.04 for all simulations. Simulations of $N=200$ chain were performed at four different interactions; $\overline{\epsilon}_{\rm AS}=1.4, 1.7, 1.8, 2.0$. For each of these values, the temperature was varied from $\overline{T}=0.5$ to 0.7 with an interval of 0.05. Simulations were also performed for a $N=400$ chain for $\overline{\epsilon}_{\rm AS}=1.7$ at the temperatures ranging from $\overline{T}=0.5$ to 0.8 with an interval of 0.05. The 200 and 400 bead systems were equilibrated for $1 \times10^{8}$ steps, and the data was sampled after every $4 \times 10^{6}$ and $2 \times 10^{7}$ steps, respectively. Four different initial configurations were used for averaging. All simulations were performed using open source molecular dynamics code LAMMPS.\cite{Plimpton1995}
The simulation data were used for the calculation of different structural quantities. We calculated the radius of gyration, $R_{\rm g}$, of the polymer to monitor the swelling of the polymer chain. We define a dimensionless radius of gyration $\overline{R}_{\rm g}=R_{\rm g}/\sigma$ given by the following expression
\begin{equation}\label{eq:rg}
\overline{R}_{\rm g}=\sqrt{\frac{1}{N}\sum_{i=1}^{N}(\overline{r}_{ i}-\overline{r}_{\rm cm})^{2}},
\end{equation}
where $N=200$ and $400$, $\overline{r}_{\rm cm}$ and $\overline{r}_{i}$ are the dimensionless coordinates of the centre of mass of the polymer chain and the $i$-th bead, respectively.
The number of solvent beads, $N_{\rm s}$, in the first solvation shell of the polymer was calculated to determine the bound solvent content.
The solvent beads which were within a distance of $\overline{r}=r/\sigma=1.5$ from any of the polymeric beads were regarded to be part of the first solvation shell.
The effective interaction between the polymer beads was calculated using the potential of mean force $U_{\rm AH}$, which was calculated from the radial distribution function of the amphiphilic and solvophobic bead pairs using the following expression:\cite{Chandler1987}
\begin{equation}\label{eq:pmf}
\frac{\overline{U}_{\rm AH}(\overline{r})}{\overline{T}}=-\ln{g_{\rm AH}(\overline{r})}.
\end{equation}
where $\overline{U}_{\rm AH}=U_{\rm AH}/\epsilon_{\rm ss}$ is the dimensionless potential of mean force.
\subsection{Solvophobic Potential by Kolomeisky and Widom}
\label{sec:widom}
Our simulations indicate that the LCST is dependent on the entropy loss of the bound solvent, and the mean interaction energy difference between the bulk and the bound solvent (see Sec.~\ref{subsec:simulation}). To obtain further insights related to the nature of the transition, scaling behavior, and the effect of chain flexibility, a theoretical approach is adopted where we consider the hydrophobic potential proposed by Kolomeisky and Widom (KW).\cite{Kolomeisky1999} The KW model is spherically symmetric in nature and does not relate to a specific solvent or solute, which is in correspondence with the modeling framework employed in our simulations. The KW model belongs to a class of implicit solvent models \cite{Lee1996,Moelbert2003a} which incorporate different interaction energies depending on the proximity of the solvent to a solute molecule. These models have been used for studying the solubility of small solutes in water. The KW model is one of the simplest among these models as it has only two solvent interaction energies; one for the bound state and another for the bulk state. These interaction energies are analogous to the monomer-solvent interaction energy ($\epsilon_{\rm AS}$), and solvent-solvent interaction energy ($\epsilon_{\rm SS}$) in our simulation model, respectively. Hence it can be seen that the model contains those contributions which have been emphasized in our simulation results.
\begin{figure*}
\centering
\includegraphics[scale=0.3]{schematic}
\caption{Schematic representation for the theoretical model. Black and blue beads represent monomer and solvent molecules, respectively. (a) Single polymer chain in explicit solvent, (b) single polymer chain in implicit solvent where the effect of solvent is incorporated in the monomer-monomer interaction potential $\phi(r)$, and (c) $\phi(r)$ modeled by the solvophobic potential given by KW.\cite{Kolomeisky1999} The potential assumes a one-dimensional solvent lattice where each solvent molecule has $q$ states. Neighboring solvent molecules exist in a bounded state (BS) when both are in the state \enquote{1} and in unbounded state (US) state otherwise. Solute can occupy interstitial sites between bounded (BS) solvent molecules. }
\label{fig:lattice}
\end{figure*}
In the KW-model, solvent molecules form a one-dimensional lattice with a nearest neighbor interaction, and each solvent molecule can exist in $q$ different states denoted by $1, 2, \cdots q$
as shown in Fig.~\ref{fig:lattice}. The interaction energy between the neighboring solvent molecules is $w$ when both of them are in the state \enquote{1}, and $u$ otherwise with $u>w$. Here the former and the latter cases are termed as the bound state (BS) state and the unbound state (US) state, respectively. A solvent molecule in the BS state can exist only in one state, whereas that in the US state in $q-1$ states. Hence the entropy of a solvent molecule in the BS state vanishes, whereas that in the US state is given by $k_{\rm B}\ln{(q-1)}$. In other words, the BS state is energetically favorable ($w<u$), while the US state is entropically favorable. The energetic ($\Delta U$) and entropic ($\Delta S$) differences between the BS and US states are $w-u$ and $-k_{\rm B}\ln{(q-1)}$, respectively. The competition between the BS and the US states can be conveniently described by a dimensionless parameter $x$ defined as
\begin{equation}\label{eq:8}
x=e^{(\Delta U-T\Delta S)/k_{\rm B}T}=\frac{q-1}{c},
\end{equation}
where
\begin{equation}\label{eq:9}
c=e^{(u-w)/k_{\rm B}T}.
\end{equation}
Since the number of states $q$ is constant in the KW model, $x$ is a monotonically increasing function of the temperature. Solute molecules are allowed to occupy only the interstitial sites between the solvent molecules in the BS state. Based on these assumptions, KW obtained the solvent mediated attraction potential $\phi(r)$ between two solute molecules (implicit solvent) for $r>\sigma,$\cite{Kolomeisky1999}
\begin{equation}\label{eq:widom_potential}
\phi(r) = -k_{\rm B}T\ln{\left[1+\left(\frac{1+Q}{1-Q}\right)
\left(\frac{1-S}{1+S}\right)^{(r-\sigma)/\sigma}
\right]},
\end{equation}
where
\begin{equation}\label{eq:6}
S=\left[1-\frac{4 x}{(1+x)^2}\left(1-\frac{1}{c}\right)\right]^{1/2},
\end{equation}
\begin{equation}\label{eq:7}
Q=\frac{\rm \sgn{({\it x}-1)}}{\left[1+4 x/(x-1)^2 c\right]^{1/2}},
\end{equation}
and $\sgn(z)$ is the sign function. Fig.~\ref{fig:wr} shows the variation of $\phi/k_{\rm B}T$ as a function of $\tilde{r}=r/\sigma$ for two different temperatures. The range of the solvent mediated interaction becomes shorter when the temperature is increased. From the inset of Fig.~\ref{fig:wr}, it can be seen that the attraction becomes stronger for higher temperature when $\tilde{r}$ is small. Such a behavior indicates that the monomers tend to aggregate as the temperature is increased.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.35]{wr}
\end{center}
\caption{Variation of $\phi/k_{\rm B}T$ as function of $\tilde{r}=r/\sigma$ for different
$\tilde{T}$ when $q=5 \times 10^5$.
The inset shows the variation at low $\tilde{r}$ values.}
\label{fig:wr}
\end{figure}
The large value of $q$ used by KW was justified in order to match the temperature dependence of the
solubility of non-polar solutes in water.\cite{Kolomeisky1999} From Eq.~(\ref{eq:widom_potential}) it can be seen that the monomer-monomer potential is temperature dependent due to the implicit nature of solvent. An important point to note is that this temperature dependence is not ad-hoc but a consequence of the underlying solvent model summarized in Fig.~\ref{fig:lattice}.
\section{Results and Discussions}
\label{sec:results}
\subsection{Simulation}
\label{subsec:simulation}
To examine the structural change of the polymer chain with the temperature, we plot in Fig.~\ref{fig:a} the variation of $\overline{R}_{\rm g}$ with $\overline{T}$ for different $\overline{\epsilon}_{\rm AS}$ for the $N=200$ chain. The measured $\overline{R}_{\rm g}$ is distinctly larger when $\overline{\epsilon}_{\rm AS}$ is increased, implying swelling as a result of stronger association between the amphiphilic bead and the solvent. We observe that there are three different behaviors according to the value of $\overline{\epsilon}_{\rm AS}$; (i) when $\overline{\epsilon}_{\rm AS}=1.4$, the polymer chain remains in a globular state and its $\overline{R}_{\rm g}$ is almost independent of $\overline{T}$, (ii) when $\overline{\epsilon}_{\rm AS}=2.0$, the polymer chain is in the coiled state at all the temperatures, and (iii) when $\overline{\epsilon}_{\rm AS}=1.7$ and $1.8$, $\overline{R}_{\rm g}$ decreases with the temperature, which is similar to the LCST behavior in thermoresponsive polymers.
\begin{figure}[h]
\begin{center}
\subfigure{\label{fig:a}\includegraphics[scale=0.35]{rg_200}}
\subfigure{\label{fig:b}\includegraphics[scale=0.35]{rg_400}}
\end{center}
\caption{Variation of $\overline{R}_{\rm g}$ of polymer chain with $\overline{T}$ for (a) 200 bead polymer at different values of $\overline{\epsilon}_{\rm AS}$, (b) $N=400$ chain at $\overline{\epsilon}_{\rm AS}=1.7$.}
\label{fig:rg}
\end{figure}
To further probe the decrease in $\overline{R}_{\rm g}$ with the temperature, we focus on the $N=400$ chain for $\overline{\epsilon}_{\rm AS}=1.7$. From Fig.~\ref{fig:b}, we see that the size change with the temperature for the $N=400$ chain is more prominent compared to that of the $N=200$ chain, which is due to difference in scaling of $\overline{R}_{\rm g}$ with $N$ for the coiled and globular states.\cite{Grosberg1997} In Fig.~\ref{fig:b}, $\overline{R}_{\rm g}$ decreases by 23\% as the temperature is increased from $\overline{T}=0.55$ to 0.75. Such trends have been observed in atomistic simulations of PNiPAM-water system.\cite{Du2010a,Alaghemandi2012a, Deshmukh2012b} Unlike the first-order LCST transition behavior observed in experimental studies of thermoresponsive polymer solutions,\cite{Wu1998, Wang1999, Cao2005} the transition observed in Fig.~\ref{fig:b} shows a gradual change in $\overline{R}_{\rm g}$ with $\overline{T}$. It is known that the coil-to-globule transition is a first-order transition only in the case of rigid and semirigid chains.\cite{Grosberg1992, Graziano2000, Baysal2003} Hence the continuous transition in our simulation is not surprising because the polymer chain used in the simulations, is fully flexible due to the absence of the angular and dihedral interactions. This is further supported by the results of our theoretical model, where we observe that the coil-to-globule transition deviates from a first order behavior with increase in chain flexibility (see Sec~\ref{sec:khok}). Based on the above observations, we consider that the behavior observed for $\overline{\epsilon}_{\rm AS}=1.7$ is akin to the LCST phenomenon in thermoresponsive polymers. Figure ~\ref{fig:snap} shows representative snapshots of $N=400$ chain at $\epsilon_{\rm AS}=1.7$ for different temperatures, we can see that the polymer chain is in the coil-like state below $\overline{T}=0.6$, while it is is the globule-like
state above $\overline{T}=0.7$.
\begin{figure*}[h]
\centering
\includegraphics[scale=0.48]{snapshots}
\caption{Representative snapshots of equilibrated $N=400$ chain at $\epsilon_{\rm AS}=1.7$ for different temperatures. Green and red beads represent solvophobic and amphiphilic units, respectively. (a) $\tilde{T}=0.55$, (b) $\tilde{T}=0.6$, (c) $\tilde{T}=0.7$ and (d) $\tilde{T}=0.8$. Solvent beads are not included for clarity.}
\label{fig:snap}
\end{figure*}
Hereafter we will be referring to $\overline{\epsilon}_{\rm AS}=1.4$, $\overline{\epsilon}_{\rm AS}=1.7$ and $1.8$, and $\overline{\epsilon}_{\rm AS}=2.0$ as low, intermediate and high values, respectively. To further understand the behavior in these three states, we examine the variation of the number of bound solvent beads, $N_{\rm s}$, with the temperature. In Fig.~\ref{fig:wa}, the number of bound solvent is found to be almost independent of $\overline{T}$ for low and high values of $\overline{\epsilon}_{\rm AS}$, whereas it decreases by increasing the temperature for intermediate values. In Fig.~\ref{fig:wb}, we plot $N_{\rm s}$ for the $N=400$ chain as a function of $\overline{T}$, when $\epsilon_{\rm AS}=1.7$. The decrease of $N_{\rm s}$ with the temperature is more prominent as compared with that of the $N=200$ chain case. Moreover, $N_{\rm s}$ markedly decreases around $\overline{T}=0.65$ which coincides with the temperature around which $\overline{R}_{\rm g}$ also decreases (see Fig.~\ref{fig:b}). It should be stressed that the change in $\overline{R}_{\rm g}$ and $N_{\rm s}$ with the temperature is observed in our model with minimal interactions as in Eq.~(\ref{eq:energy_simu}). It is important to note that even when the interaction parameters are independent of the temperature, the temperature dependence of $\overline{R}_{\rm g}$ is induced by the bound solvent number $N_{\rm s}$ for a range of values of $\overline{\epsilon}_{\rm AS}$.
\begin{figure}
\centering
\subfigure{\label{fig:wa}\includegraphics[scale=0.35]{solvent_stats_200}}
\subfigure{\label{fig:wb}\includegraphics[scale=0.35]{solvent_400}}
\caption{(a) Variation of $N_{\rm s}$ with $\overline{T}$ at different values for (a) $N=200$ chain at different values of $\overline{\epsilon}_{\rm AS}$, (b) $N=400$ chain at $\overline{\epsilon}_{\rm AS}$=1.7.}
\label{fig:water}
\end{figure}
The above trends in the three different states can be rationalized by considering the different energy, and entropic contributions in the model system.
The attraction between the amphiphilic monomer and the solvent is energetically favorable since
it is stronger than the solvent-solvent interaction ($\epsilon_{\rm AS} > \epsilon_{\rm SS}$), leading to a coil-like polymer conformation with
many bound solvent beads. Whereas from the viewpoint of entropy, the solvent prefers to be in the bulk state rather than the bound state. Another important contribution is the solvation of solvophobic beads, which is unfavorable as the interaction between the solvophobic bead and the solvent is repulsive.
The interplay of these contributions determines the different states of the polymer chain. \\
\indent Concerning intermediate $\overline{\epsilon}_{\rm AS}$ values, the large values of $\overline{R}_{\rm g}$ in the coil-like state
(see Fig.~\ref{fig:b}) and large $N_{\rm s}$ at low temperatures (see Fig.~\ref{fig:wb}) indicate
that the attraction between the amphiphilic monomer and the solvent dominates the entropy
loss of the solvent.
Hence it is favorable for the solvent to solvate the solvophobic beads despite the repulsive
interaction between them. With increase in the temperature, the gain in the entropy due to unbinding of the solvent leads to a reduction in the bound solvent content, $N_{\rm s}$. In the transition region ($0.6 \leq T \leq 0.7$), there is marked decrease of the bound solvent
because the entropy gain of the solvent dominates over the amphiphilic-solvent attraction.
This change in the dominant contribution makes the solvation of solvophobic beads unfavorable, leading to an attraction between the polymeric beads.
Such an attraction becomes stronger with the temperature, and drives the transition from
the coil-like state to the globule-like state.
In Fig~\ref{fig:pmf_lcst}, we plot the potential of mean force, $\overline{U}_{\rm AH}$,
between the amphiphilic and the solvophobic beads as a function of the distance $\overline{r}$.
Here we see that the attraction between the polymeric beads increases with the temperature. At low temperature, the attraction between the polymeric beads is screened due to the presence of solvent beads. At high temperatures, the bound solvent beads unbind due to entropy gain which reduces the screening effect, leading to the collapse of the polymer. As shown in Fig.~\ref{fig:pmf_collapsed}, the increase in the attraction between polymeric beads with temperature is not observed for low $\overline{\epsilon}_{\rm AS}$ values.\\
\indent For low $\overline{\epsilon}_{\rm AS}$ values, both $\overline{R}_{\rm g}$ and $N_{\rm s}$ are small.
This means that at these interaction strengths the entropic gain of the free solvent beads dominates over the interaction between
the amphiphilic monomer and the solvent for all the temperatures.
Hence the solvation of solvophobic beads is unfavorable in the
entire temperature range. On the other hand, for high $\overline{\epsilon}_{\rm AS}$ values, both $\overline{R}_{\rm g}$ and $N_{\rm s}$ are large.
This indicates that the attractive interaction is stronger than the entropy loss which leads to the binding of the solvent to the polymer. Hence, the solvation of the solvophobic beads is favored at all temperatures despite the repulsive interaction between the solvophobic monomer and the solvent.
\begin{figure}
\centering
\subfigure{\label{fig:pmf_lcst}\includegraphics[scale=0.35]{pmfah_lcst}}
\subfigure{\label{fig:pmf_collapsed}\includegraphics[scale=0.35]{pmfah_collapsed}}
\caption{Variation of $\overline{U}_{\rm AH}/k_{\rm B}\overline{T}$ with $\overline{r}$ at different temperatures for (a) $N=400$ chain at $\overline{\epsilon}_{\rm AS}=1.7$ (b) $N=200$ chain at $\overline{\epsilon}_{\rm AS}=1.4$.}
\end{figure}
We find from our simulations that the transition from the coil-like state to the globule-like state
depends on how much the solvophobic beads can be solvated.
The solvation depends on the interplay between the entropy loss of the bound solvent, and the energetic
difference between the bound and the bulk solvent.
The bound solvent is energetically favored as the interaction between the amphiphilic monomer
and the solvent is stronger than the solvent-solvent interaction.
Therefore two kinds of beads having opposite interactions with the solvent are necessary to exhibit
the LCST as long as the interaction parameters are independent of the temperature as in our case.
Additionally in our simulation, we did not incorporate any specific chemical or structural details pertaining to
the solvent or the polymer chain.
Therefore we have demonstrated that the LCST behavior can be exhibited by a coarse-grained description of
the polymer and the solvent even when the interaction parameters are independent of the temperature.
\subsection{ Theoretical description of the coil-to-globule transition}
\label{sec:khok}
To study the coil-to-globule transition in the theoretical framework, we adopt the phenomenological free energy expression for a polymer chain in an implicit solvent, which has been given by Grosberg and Kuznetsov.\cite{Grosberg1992} We choose the KW potential (see Sec~\ref{sec:widom}) to model the effective monomer-monomer interaction. The free energy expression for the system is as follows:
\begin{equation}
\frac{F}{k_{\rm B}T}=\alpha^{2}+\frac{1}{\alpha^{2}} +\frac{\sqrt{N}}{\alpha^{3}\sigma^{3}}B+
\frac{1}{\alpha^{6}\sigma^{6}}C,
\label{eq:4}
\end{equation}
where $\alpha=R_{\rm g}/(\sqrt{N}\sigma)$ characterizes the extent of the swelling,
$N$ the degree of polymerization, $\sigma$ the diameter of the monomeric unit,
$B$ and $C$ are the second and the third virial coefficients, respectively.
The first and the second terms in the above free energy are the entropic contributions with the
the coiled and the globular states given by $\alpha>1$ and $\alpha<1$, respectively.
The third and the fourth terms represent the energy contributions from the two-body and the three-body interactions, respectively.
In general, the second virial coefficient $B$ is given by
\begin{equation}
B=2\pi\int_{0}^{\infty} {\rm d}r \, r^2 \left(1-e^{-\Phi(r)/k_{\rm B}T}\right),
\label{eq:second_virial}
\end{equation}
where $\Phi(r)$ is the monomer-monomer interaction potential.
Here we assume that it has the following form:
\begin{equation}\label{eq:interaction}
\Phi(r) = \left\{
\begin{array}{ll}
\infty &: r<\sigma \\
\phi(r) &: r>\sigma
\end{array}
\right.,
\end{equation}
where $\Phi(r)$ for $r<\sigma$ corresponds to the (hard-core) excluded volume
interaction, and $\phi(r) $ for $r>\sigma$ (see Eq.~(\ref{eq:widom_potential})) is the KW solvophobic potential.
Substituting Eqs.~(\ref{eq:widom_potential}) and ~(\ref{eq:interaction}) into Eq.~(\ref{eq:second_virial}), we obtain the second
virial coefficient $B$ as
\begin{widetext}
\begin{equation}
B=\frac{2\pi \sigma^{3}}{3}\left[1+ 3\left(\frac{1+Q}{1-Q}\right)\left( \frac{1}{\ln{L}} -
\frac{2}{\left(\ln{L}\right)^{2}} +\frac{2}{\left(\ln{L}\right)^{3}} \right)\right],
\label{eq:11_g}
\end{equation}
\end{widetext}
where $L=(1-S)/(1+S)$.
Hereafter we use the dimensionless quantities such as
the temperature $\tilde{T}=k_{\rm B}T/(u-w)$,
the distance $\tilde{r}=r/\sigma$,
the solvent mediated interaction potential $\tilde{\phi}=\phi/(u-w)$,
and the second virial coefficient $\tilde{B}=3B/2\pi \sigma^{3}$.
In Fig.~\ref{fig:br}, we plot $\tilde{B}$ as a function of $\tilde{T}$ for different
$q$-values.
We observe that $\tilde{B}$ remains almost unity for low temperatures and then decreases
rapidly for higher temperatures.
This means that the monomer-monomer interaction is repulsive for lower temperatures,
while it is attractive for higher temperatures.
Moreover, the temperature corresponding to the sharp drop from the positive
(repulsive) to the negative (attractive) $\tilde{B}$-values decreases when $q$ is increased for a fixed $u-w$ value. The significance of this observation will be discussed later.
\begin{figure}
\begin{center}
\includegraphics[scale=0.35]{br}
\end{center}
\caption{Variation of $\tilde{B}$ with $\tilde{T}$ for different $q$ values.}
\label{fig:br}
\end{figure}
In order to find the equilibrium polymer conformation, we minimize Eq.~(\ref{eq:4})
with respect to $\alpha$, and obtain the equation
\begin{equation}\label{eq:minimisation}
\alpha^{5}-\alpha-\pi\sqrt{N}\tilde{B} - \frac{\tilde{C}}{\alpha^{3}}=0,
\end{equation}
where $\tilde{C}=3C/\sigma^{6}$ is the non-dimensional third virial coefficient which is related to the rigidity of the polymer chain and has contributions from the three-body interactions such as angular interactions.
We numerically solve the above equation to obtain $\alpha$ for fixed values of $q$, $N$
and $\tilde{C}$.
In Fig.~\ref{fig:ctog}, we show the variation of the swelling parameter
$\alpha$ with the temperature $\tilde{T}$ for different values of the
third virial coefficient $\tilde{C}$ when $q=5 \times 10^{5}$.
We see that the polymer chain undergoes a transition from the coiled state to
the globular state with increase in the temperature.
For low temperatures, the polymer chain is in the coiled state and hence large $\alpha$.
In this case, the first and the third terms in Eq.~(\ref{eq:minimisation}) are dominant
because $N$ is also large.
Then Eq.~(\ref{eq:minimisation}) reduces to
\begin{equation}\label{eq:11_h_1}
\alpha^{5}-\pi\sqrt{N}\tilde{B} \approx 0,
\end{equation}
and we obtain $R_{\rm g} \sim N^{3/5}$ corresponding to the scaling
of a polymer chain in a good solvent.
When the temperature is high, on the other hand, the polymer is in the globular
state and $\alpha$ becomes small.
Then the third and the fourth terms dominate in Eq.~(\ref{eq:minimisation});
\begin{equation}\label{eq:11_h}
\pi\sqrt{N}\tilde{B} + \frac{\tilde{C}}{\alpha^{3}} \approx 0.
\end{equation}
This gives the scaling $R_{\rm g} \sim N^{1/3}$ corresponding to a polymer
chain in a poor solvent.
\begin{figure}
\centering
\includegraphics[scale=0.35]{ctog}
\caption{Variation of $\alpha$ with $\tilde{T}$ for different values of $\tilde{C}$
when $q=5 \times 10^5$ and $N=10^5$.}
\label{fig:ctog}
\end{figure}
In order to understand this coil-to-globule transition in terms of the dominant interactions,
let us consider the free energy difference of a solvent molecule between the BS and US states;
\begin{equation}\label{eq:deltaf}
\Delta F = F_{\rm BS}-F_{\rm US}=U_{\rm BS}-U_{\rm US} - T(S_{\rm BS}-S_{\rm US}),
\end{equation}
where $U_{\rm BS}$ ($U_{\rm US}$) and $S_{\rm BS}$ ($S_{\rm US}$) are the
energy and the entropy of the BS (US) state, respectively.
Using the model parameters defined in Sec.~\ref{sec:widom}, the above quantity can be expressed as
\begin{equation}
\Delta F=-(u-w) +k_{\rm B}T \ln{(q-1})=k_{\rm B}T\ln x,
\label{eq:DeltaF}
\end{equation}
where $x$ is defined before in Eq.~(\ref{eq:8}). From Eq.~(\ref{eq:DeltaF}) and since $u>w$, we find $\Delta F<0$ for $x<1$ (lower temperatures), leading to the BS state being more favorable.
Given the implicit nature of the solvent in the theoretical treatment, the large value
of $\alpha$ for lower temperatures in Fig.~\ref{fig:ctog} is an indication of the large
amount of the bound solvent.
For $x>1$ (higher temperatures), on the other hand, the US state is more favorable and
the amount of the unbound solvent increases.
In this case, the attraction between the monomeric units is induced.
This is seen in Fig.~\ref{fig:br} for the larger negative value of the second virial coefficient
$\tilde{B}$.
These solvent induced interactions drive the transition from the coiled state to the
globular state.
The phase transition temperature $T^{\ast}$ determined by the condition $\Delta F=0$
is given by
\begin{equation}\label{eq:critical_temp}
T^{\ast}=\frac{u-w}{k_{\rm B}\ln{(q-1)}}.
\end{equation}
As observed in Sec.~\ref{sec:widom}, $T^{\ast}$ decreases as $q$ is increased when $u-w$ is fixed.
We further observe in Fig.~\ref{fig:ctog} that the nature of the transition changes from a discontinuous transition to a continuous one by increase in the third virial coefficient $\tilde{C}$. It is known that the third virial coefficient is larger for flexible chains.\cite{Grosberg1992,Graziano2000} This shows that the nature of transition deviates from a first-order behavior with increase in the flexibility of the chain. Another important point is that the above argument does not include any details regrading the structure of the solvent and/or polymer which indicates that the structural details of the bound solvent are not necessary for the thermoresponsive behavior to manifest.
\section{Summary and Conclusions}
\label{sec:sum}
In this paper, we have tried to understand the single chain coil-to-globule transition of a thermoresponsive polymer through simulation and theoretical approaches. In the simulations, the model comprises of a single polymer chain in an explicit solvent with temperature independent interaction parameters. The solvent is explicitly included to avoid any ad-hoc dependence of the interaction potentials on temperature. To obtain further insights, we have adopted a theoretical framework with only those interactions that have been emphasized in our simulation studies. The theoretical model describes a single chain in an implicit solvent where the effect of solvent is included into the monomer-monomer interaction potential. For the interaction between the monomers, we have used the solvophobic potential proposed by Kolomeisky and Widom.\cite{Kolomeisky1999} The temperature dependence of the solvophobic potential is not ad-hoc in nature and arises due to the underlying solvent model.\\
\indent Our simulations indicate that the LCST is dependent on the competition between
the two contributions, namely, the entropy loss of the bound solvent and the mean energy
difference between the bound and the bulk solvent. This hypothesis is supported by our theoretical calculations.
The former favors the globular state of the polymer chain, whereas the latter prefers the coiled state.
At low temperatures, solvent molecules bind to the polymer chain rather than to reside in the bulk because
the bound state is energetically favorable, and the coiled state is obtained.
At high temperatures, on the other hand, the entropy loss of the bound state is more dominant than
its energetic gain and the solvent molecules tend to leave the polymer. Such a competition of the interactions induces a solvent driven attraction between the polymeric beads
leading to the collapse of the polymer chain. These findings are supported by the experimental studies of Bischofberger and coworkers who showed that the LCST is dependent on the mean energy difference between the bound and bulk solvent.\cite{Bischofberger2014, Bischofberger2014a}\\
\indent Another common feature between the simulation and the theory is that both of them have spherically symmetric solvent and monomeric beads. This indicates that the structural arrangement of the solvent molecules around the polymer chain is not a necessary component to be considered explicitly for the coil-to-globule transition. In other words, a coarse-grained description of the solvent is sufficient to reproduce the LCST behavior, this observation is in agreement with the experimental results of Bischofberger and coworkers,\cite{Bischofberger2014, Bischofberger2014a} where they show that a coarse-grained representation of the solvent is sufficient to explain the LCST and its variation with different alcohols.\\
\indent In our simulations studies, it is shown that the LCST is dependent on the amphiphilic-solvent attraction and solvation of the solvophobic beads. This indicates that in the case of temperature-independent interaction parameters, two different kinds of beads with opposite interaction with the solvent are required in the monomeric unit to exhibit LCST. In the theoretical model, the solvent is implicit and the variation of solvent effects with temperature is included in the monomer-monomer interaction due to which one kind of bead is sufficient to exhibit LCST. A mixture of PNiPAM and water exhibits a first-order phase transition at 32$\degree$C.\cite{Wang1998, Wu1998} In our theoretical argument, we showed that the order of the transition changes from a first-order (discontinuous) to a second-order (continuous) with the increase in the chain flexibility (or the third virial coefficient). The reason why we observe only the continuous coil-to-globule transition in our simulation is because we did not take into account any angular interactions to deal with the chain flexibility. The effect of angular and dihedral interactions on the nature of the coil-to-globule transition needs to be explored further. \\
\indent The variation in behavior with temperature for different $\overline{\epsilon}_{\rm AS}$ values (see Fig~\ref{fig:a}) is analogous to different polymers in the family of poly(N,N-alkyl alkyl acrylamide). Low $\overline{\epsilon}_{\rm AS}$ values are similar to polymers with bulky side groups such as poly(N-butyl acrylamide) which are always insoluble in water. On the other hand, high $\overline{\epsilon}_{\rm AS}$ values are similar to polymers such as polyacrylamide which are always soluble in water. This indicates that our coarse-grained model is able to explain the LCST behavior for different systems in the poly(N,N-alkyl,alkyl acrylamide) family of polymers. Furthermore, the model can also be utilized to study the variation of LCST due to factors such as co-solvents and additives by examining their effect on the two dominant physical interactions namely, the entropy loss of the bound solvent and the mean energy difference between the bound and the bulk solvent. This aspect will be addressed in our future studies.
\begin{acknowledgments}
The computations were carried out at the High Performance Computing Facility at IIT Madras. S.K. acknowledges support from the Grant-in-Aid for Scientific Research on Innovative Areas ``\textit{Fluctuation and Structure}" (Grant No.\ 25103010) from the Ministry of Education, Culture, Sports, Science, and Technology of Japan, the Grant-in-Aid for Scientific Research (C) (Grant No.\ 15K05250) from the Japan Society for the Promotion of Science (JSPS), and the JSPS Core-to-Core Program ``\textit{International Research Network for Non-equilibrium Dynamics of Soft Matter}". G.S.B thanks Okamoto, Andelman, Koga and Seki for valuable discussions.
\end{acknowledgments}
|
\section{Appendix}
In the Appendix we prove all theorems presented in the main body of the paper.
\subsection{Computing general kernels with \textit{TripleSpin}-matrices}
We prove here Theorem \ref{thm:stationary} and its non-stationary analogue.
For the convenience of the Reader, we restate both theorems.
We start with Theorem \ref{thm:stationary} that we restate as follows.
\textbf{Theorem \ref{thm:stationary} (stationary kernels)}
The family of functions
\begin{align*}
\kappa_K(\textbf{x},\textbf{y})&:=\sum_{k=1}^K\alpha_k (\mathbb{E}[\cos(\textbf{g}_k^\top\cdot \textbf{x})\cos(\textbf{g}_k^\top\cdot \textbf{y})] + \mathbb{E}[\sin(\textbf{g}_k^\top\cdot \textbf{x})\sin(\textbf{g}_k^\top\cdot \textbf{y})])
\end{align*}
with $\textbf{g}_k\sim\mathcal{N}(\mu_k,\mbox{diag}((\sigma_k^1)^2,...,(\sigma_k^d)^2))$, $\mu_k,\sigma_k\in\mathbb{R}^{n}$, $\alpha_k\in\mathbb{R}$, $K\in\mathbb{N}\cup\{\infty\}$ is dense in the family of stationary real-valued kernels with respect to pointwise convergence.
\begin{proof}
Theorem 3 of \cite{samo2015generalized} states that:
``Let $h$ be a real-valued positive semi-definite, continuous, and integrable function such that $\forall\tau\in\mathbb{R}^n,h(\tau)>0$. The family of functions
\begin{equation*}
\kappa_K(\tau):=\sum_{k=1}^K\alpha_k h(\tau\odot\sigma_k)\cos(2\pi\mu_k\tau)
\end{equation*}
with $\mu_k,\sigma_k\in\mathbb{R}^{+n}$, $\alpha_k\in\mathbb{R}$, $K\in\mathbb{N}\cup\{\infty\}$ is dense in the family of stationary real-valued kernels with respect to pointwise convergence.'' Here $\mathbb{R}^{+n}=\{x\in\mathbb{R}^{n}: x_i\geq 0\mbox{ for all }i=1,...,n \}\subset\mathbb{R}^{n}$.
Let $\odot$ denote the element-wise product.
If we choose $h(\tau)=\exp(-2\pi^2\|\tau\|^2)$, as suggested in \cite{samo2015generalized}, then it follows that
\begin{align}
\kappa_K(\tau)&:=\sum_{k=1}^K\alpha_k \exp(-2\pi^2\|\sigma_k\odot\tau\|^2)\cos(2\pi\mu_k^\top\cdot \tau) \notag\\
&=\sum_{k=1}^K\alpha_k \exp(-0.5(2\pi\|\sigma_k\odot\tau\|)^2)\cos(2\pi\mu_k^\top\cdot \tau) \notag\\
&=\sum_{k=1}^K\alpha_k \int_{\mathbb{R}} \mathcal{N}(g;0,1)\cos(2\pi\|\sigma_k\odot\tau\|g)\cos(2\pi\mu_k^\top\cdot \tau)dg \label{eq:t13}\\
&=\sum_{k=1}^K\alpha_k \int_{\mathbb{R}} \mathcal{N}(g;0,\|\sigma_k\odot\tau\|^2)\cos(2\pi g)\cos(2\pi\mu_k^\top\cdot \tau)dg \label{eq:t14}\\
&=\sum_{k=1}^K\alpha_k \int_{\mathbb{R}} \mathcal{N}(g;0,\|\sigma_k\odot\tau\|^2)[\cos(2\pi g)\cos(2\pi\mu_k^\top\cdot \tau)-\sin(2\pi g)\sin(2\pi\mu_k^\top\cdot \tau)]dg \label{eq:t15}\\
&=\sum_{k=1}^K\alpha_k \int_{\mathbb{R}} \mathcal{N}(g;0,\|\sigma_k\odot\tau\|^2)\cos(2\pi (g+\mu_k^\top\cdot \tau))dg \label{eq:t16}\\
&=\sum_{k=1}^K\alpha_k \int_{\mathbb{R}} \mathcal{N}(g;0,\tau^\top\mbox{diag}((\sigma_1^k)^2,...,(\sigma_d^k)^2)\tau)\cos(2\pi (g+\mu_k^\top\cdot \tau))dg \notag\\
&=\sum_{k=1}^K\alpha_k \int_{\mathbb{R}^n} \mathcal{N}(\textbf{g};0,\mbox{diag}((\sigma_1^k)^2,...,(\sigma_d^k)^2))\cos(2\pi (\textbf{g}^\top\cdot \tau+\mu_k^\top\cdot \tau))d\textbf{g} \label{eq:t18}\\
&=\sum_{k=1}^K\alpha_k \int_{\mathbb{R}^n} \mathcal{N}(\textbf{g};\mu_k,\mbox{diag}((\sigma_1^k)^2,...,(\sigma_d^k)^2))\cos(2\pi \textbf{g}^\top\cdot \tau)d\textbf{g} \notag\\
&=\sum_{k=1}^K\alpha_k \mathbb{E}[\cos(2\pi \textbf{g}_k^\top\cdot\tau)] \notag\\
&=\sum_{k=1}^K\alpha_k \mathbb{E}[\cos(2\pi \textbf{g}_k^\top\cdot(\textbf{x}-\textbf{y}))] \notag\\
&=\sum_{k=1}^K\alpha_k (\mathbb{E}[\cos(2\pi \textbf{g}_k^\top\cdot \textbf{x})\cos(2\pi \textbf{g}_k^\top\cdot \textbf{y})] + \mathbb{E}[\sin(2\pi \textbf{g}_k^\top\cdot \textbf{x})\sin(2\pi \textbf{g}_k^\top\cdot \textbf{y})]) \notag
\end{align}
is dense in the family of stationary real-valued kernels with respect to pointwise convergence.
Equation~(\ref{eq:t13}) follows from Bochner's theorem, (\ref{eq:t14}) from integration by substitution,
(\ref{eq:t15}) since sine is an odd function, (\ref{eq:t16}) from cosine angle sum identity,
(\ref{eq:t18}) from writing $g=\tau^\top\cdot\textbf{g}$ as linear transform of $\textbf{g}$. Absorbing $2\pi$ into $\mu_k$ and $\sigma_k$ and relaxing $\mu_k,\sigma_k\in \mathbb{R}^{+n}$ to $\mu_k,\sigma_k\in\mathbb{R}^{n}$ completes the proof.
\end{proof}
Now we will show the analogous version of that result for non-stationary kernels.
\begin{theorem}[non-stationary kernels]
The family of functions
\begin{equation*}
\kappa(\textbf{x},\textbf{y}) = \sum_{k=1}^K\alpha_k( \mathbb{E}[\cos( \textbf{g}_k^\top\cdot \textbf{x})\cos(\textbf{g}_k^\top\cdot \textbf{y})]+\mathbb{E}[\sin( \textbf{g}_k^\top\cdot \textbf{x})\sin( \textbf{g}_k^\top\cdot \textbf{y})]) \Psi_k(\textbf{x})^\top\Psi_k(\textbf{y})
\end{equation*}
where $\Psi_k(\textbf{x})=\left( \begin{array}{c}
\cos(\textbf{x}^\top\cdot \textbf{w}_k^1) + \cos( \textbf{x}^\top\cdot \textbf{w}_k^2)\\
\sin(\textbf{x}^\top\cdot \textbf{w}_k^1) + \sin( \textbf{x}^\top\cdot \textbf{w}_k^2) \end{array} \right)$, with $\textbf{g}_k\sim\mathcal{N}(0,\mbox{diag}((\sigma_1^k)^2,...,(\sigma_d^k)^2)), \sigma_k\in\mathbb{R}^{n}, \textbf{w}_k^1,\textbf{w}_k^2\in\mathbb{R}^n,\alpha_k\in\mathbb{R},K\in\mathbb{N}\cup\{\infty\}$ is dense in the family of real-valued continuous bounded non-stationary kernels with respect to the pointwise convergence of functions.
\end{theorem}
\begin{proof}
Theorem 7 of \cite{samo2015generalized} states that:
``Let $(\textbf{x}, \textbf{y}) \to \kappa^*(\textbf{x},\textbf{y})$ be a real-valued, positive semi-definite, continuous, and integrable function such that $\forall \textbf{x},\textbf{y},\kappa^*(\textbf{x},\textbf{y})>0$. The family
\begin{equation*}
\kappa(\textbf{x},\textbf{y}) = \sum_{k=1}^K\alpha_k \kappa^*(\textbf{x}\odot \sigma_k,\textbf{y}\odot \sigma_k) \Psi_k(\textbf{x})^\top\Psi_k(\textbf{y})
\end{equation*}
where $\Psi_k(\textbf{x})=\left( \begin{array}{c}
\cos(2\pi \textbf{x}^\top\cdot \textbf{w}_k^1) + \cos(2\pi \textbf{x}^\top\cdot \textbf{w}_k^2)\\
\sin(2\pi \textbf{x}^\top\cdot \textbf{w}_k^1) + \sin(2\pi \textbf{x}^\top\cdot \textbf{w}_k^2) \end{array} \right)$, with $\sigma_k\in\mathbb{R}^{n+}, \textbf{w}_k^1,\textbf{w}_k^2\in\mathbb{R}^n,\alpha_k\in\mathbb{R},K\in\mathbb{N}\cup\{\infty\}$ is dense in the family of real-valued continuous bounded non-stationary kernels with respect to the pointwise convergence of functions.''
If we choose as $\kappa^*$ the Gaussian kernel:
\begin{align*}
\kappa^*(\textbf{x},\textbf{y})&=\exp(-\|\textbf{x}-\textbf{y}\|^2/2)\\
&=\mathbb{E}[\cos(\textbf{g}^\top\cdot(\textbf{x}-\textbf{y}))]
\end{align*}
with $g\sim\mathcal{N}(0,I)$ then
\begin{align*}
\kappa^*(\textbf{x}\odot \sigma_k,\textbf{y}\odot \sigma_k)&=\mathbb{E}[\cos(\textbf{g}_k^\top\cdot(\textbf{x}-\textbf{y}))]\\
\end{align*}
with $\textbf{g}_k\sim\mathcal{N}(0,\mbox{diag}((\sigma_1^k)^2,...,(\sigma_d^k)^2))$. Absorbing $2\pi$ into $w_k$ and relaxing $\sigma_k\in \mathbb{R}^{+n}$ to $\sigma_k\in\mathbb{R}^{n}$ completes the proof.
\end{proof}
\subsection{Structured machine learning algorithms with \textit{TripleSpin}-matrices}
We prove now Lemma \ref{simple_lemma}, Remark \ref{balanceness_remark},
as well as Theorem \ref{main_struct_theorem} and Theorem \ref{corollary_theorem}.
\subsubsection{Proof of Remark \ref{balanceness_remark}}
This result first appeared in \cite{ailon2006approximate}. The following proof was given in \cite{chor_sind_2016}, we repeat it here for completeness.
We will use the following standard concentration result.
\begin{lemma}(Azuma's Inequality)
Let $X_{1},...,X_{n}$ be a martingale and assume that $-\alpha_{i} \leq X_{i} \leq \beta_{i}$ for some positive constants $\alpha_{1},...,\alpha_{n}, \beta_{1},...,\beta_{n}$.
Denote $X = \sum_{i=1}^{n} X_{i}$.
Then the following is true:
\begin{equation}
\mathbb{P}[|X - \mathbb{E}[X]| > a] \leq 2e^{-\frac{a^{2}}{2\sum_{i=1}^{n}(\alpha_{i} + \beta_{i})^{2}}}
\end{equation}
\end{lemma}
\begin{proof}
Denote by $\tilde{\textbf{x}}^{j}$ an image of $\textbf{x}^{j}$ under transformation $\textbf{HD}$. Note that the $i^{th}$ dimension of $\tilde{\textbf{x}}^{j}$
is given by the formula: $\tilde{x}^{j}_{i} = h_{i,1}x^{j}_{1} + ... + h_{i,n}x^{j,n}$,
where $h_{l,u}$ stands for the $l^{th}$ element of the $u^{th}$ column of the randomized
Hadamard matrix $\textbf{HD}$.
First we use Azuma's Inequality to find an upper bound on the probability
that $|\tilde{x}^{j}_{i}| > a$, where $a=\frac{\log(n)}{\sqrt{n}}$.
By Azuma's Inequality, we have:
\begin{equation}
\mathbb{P}[|h_{i,1}x^{j}_{1} + ... + h_{i,n}x^{j,n}| \geq a] \leq 2e^{-\frac{\log^{2}(n)}{8}}.
\end{equation}
We use: $\alpha_{i} = \beta_{i} = \frac{1}{\sqrt{n}}$.
Now we take union bound over all $n$ dimensions and the proof is completed.
\end{proof}
\begin{comment}
\subsubsection{Proof of Lemma \ref{simple_lemma} - shorter way? \todo{Krzysztof}{Choose one. This may not need extra diag of $A$ is $0$ condition.}}
\begin{proof}
Let us first assume the $\textbf{G}_{circ}\textbf{D}_{2}\textbf{HD}_{1}$ setting
(analysis for Toeplitz Gaussian or Hankel Gaussian is completely analogous).
In that setting it is easy to see that one can take $\textbf{r}$ to be a Gaussian vector (this vector corresponds to the first row of $\textbf{G}_{circ}$). Furthermore linear mappings $\Phi_{i}$ are defined as: $\Phi_{i}(x_{0},x_{1},...,x_{n-1})^{T} = (x_{n-i},x_{n-i+1},...,x_{i-1})^{T}$, where operations on indices are modulo $n$.
The value of $\delta(n)$ and $p(n)$ come from the fact that matrix $\textbf{HD}_{1}$ is used as a $(\delta(n),p(n))$-balanced matrix and from Remark \ref{balanceness_remark}.
In that setting sequence $(\rho_{1},...,\rho_{n})$ is discrete and corresponds to the diagonal of $\textbf{D}_{2}$.
Thus we have: $K = 1$. To calculate $\Lambda_{F}$ and $\Lambda_{2}$, note that $\textbf{W}^{i} = \Phi^{i}$ is a permutation matrix and therefore
each $\textbf{A}^{i,j} = (\textbf{W}^{i})^T\textbf{W}^{j}$ is a permutation matrix as well. Thus we have: $\Lambda_{F} = \sqrt{n}$ and $\Lambda_{2} = 1$.
Now let us consider the setting, where the structured matrix is of the form: $\sqrt{n}\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$.
In that case $\textbf{r}$ corresponds to a discrete vector (namely, the diagonal of $\textbf{D}_{2}$).
Linear mappings $\Phi^{i}$ are defined as:
$\Phi^{i} = \sqrt{n}\textbf{D}_{h_i}$, where $h_{i}$ is the $i^{th}$ row of $\textbf{H}$.
One can also notice that the set $\{\textbf{W}^{i}\}_{i=1,...,n}$ is defined as: $\textbf{W}^{i} = \sqrt{n}\textbf{D}_{h_i}\textbf{H}$.
Let us first to compute the Frobenius norm of the matrix $\textbf{A}^{i,j}$ repeatedly using the
fact that $\|\textbf{M}\textbf{N}\|_F \le \|\textbf{M}\|_2\|\textbf{N}\|_F$ holds for any pair
of matrices $\textbf{M},\textbf{N}$. We have
\begin{equation}
\Lambda_F = \|\textbf{A}^{i,j}\|_F = \|n\textbf{H}^T\textbf{D}_{h_j}\textbf{D}_{h_i}\textbf{H}\|_F \le
n\|\textbf{H}^T\|_2\|\textbf{D}_{h_j}\|_2\|\textbf{D}_{h_i}\|_2\|\textbf{H}\|_F =
n \cdot 1 \cdot \frac{1}{\sqrt{n}} \frac{1}{\sqrt{n}} \sqrt{n} = \sqrt{n}.
\end{equation}
Similarly, it follows that
\begin{equation}
\Lambda_2 = \|\textbf{A}^{i,j}\|_2 \le n\|\textbf{H}^T\|_2\|\textbf{D}_{h_j}\|_2\|\textbf{D}_{h_i}\|_2\|\textbf{H}\|_2 = 1
\end{equation}
completing the proof.
\end{proof}
\end{comment}
\subsubsection{\textit{TripleSpin}-equivalent definition}
We will introduce here equivalent definition of the $\textit{TripleSpin}$-model that is more technical (thus we did not give it in the main body of the paper), yet more convenient to work with in the proofs.
Note that from the definition of the \textit{TripleSpin}-family we can conclude that each structured matrix $\textbf{G}_{struct} \in \mathbb{R}^{n \times n}$ from the \textit{TripleSpin}-family is a product of three main structured blocks, i.e.:
\begin{equation}
\textbf{G}_{struct} = \textbf{B}_{3}\textbf{B}_{2}\textbf{B}_{1},
\end{equation}
where matrices $\textbf{B}_{1},\textbf{B}_{2},\textbf{B}_{3}$ satisfy two conditions that we give below.
\begin{framed}
\textbf{Condition 1:} Matrices: $\textbf{B}_{1}$ and $\textbf{B}_{2}\textbf{B}_{1}$ are $(\delta(n),p(n))$-balanced isometries. \\
\textbf{Condition 2:} Pair of matrices $(B_{2},B_{3})$
is $(K,\Lambda_{F}, \Lambda_{2})$-random.
\end{framed}
Below we give the definition of $(K, \Lambda_{F}, \Lambda_{2})$-randomness.
\begin{definition}[$(K, \Lambda_{F}, \Lambda_{2})$-randomness]
A pair of matrices $(\textbf{Y},\textbf{Z}) \in \mathbb{R}^{n \times n} \times \mathbb{R}^{n \times n}$ is $(K, \Lambda_{F}, \Lambda_{2})$-random
if there exist: $\textbf{r} \in \mathbb{R}^{k}$, and
a set of linear isometries $\phi = \{\phi_{1},...,\phi_{n}\}$,
where $\phi_i : \mathbb{R}^{n} \rightarrow \mathbb{R}^{k}$, such that:
\begin{itemize}
\item $\textbf{r}$ is either a $\pm 1$-vector with i.i.d. entries
or Gaussian with identity covariance matrix,
\item for every $\textbf{x} \in \mathbb{R}^{n}$ the $j^{th}$ element $(\textbf{Zx})_{j}$ of $\textbf{Zx}$ is of the form: $\textbf{r}^{T} \cdot \phi_{j}(\textbf{x})$,
\item there exists a set of i.i.d. sub-Gaussian random variables $\{\rho_{1},...,\rho_{n}\}$ with sub-Gaussian norm at most $K$, mean $0$, the same second moments and a $(\Lambda_{F},\Lambda_{2})$-smooth set of matrices $\{\textbf{W}^{i}\}_{i=1,...,n}$ such that for every $\textbf{x} = (x_{1},...,x_{n})^{T}$ we have: $\phi_{i}(\textbf{Y}\textbf{x}) = \textbf{W}^{i} (\rho_{1}x_{1},...,\rho_{n}x_{n})^{T}$.
\end{itemize}
\end{definition}
\subsubsection{Proof of Lemma \ref{simple_lemma}}
\begin{proof}
Let us first assume the $\textbf{G}_{circ}\textbf{D}_{2}\textbf{HD}_{1}$ setting
(analysis for Toeplitz Gaussian or Hankel Gaussian is completely analogous).
In that setting it is easy to see that one can take $\textbf{r}$ to be a Gaussian vector (this vector corresponds to the first row of $\textbf{G}_{circ}$). Furthermore linear mappings $\phi_{i}$ are defined as: $\phi_{i}((x_{0},x_{1},...,x_{n-1})^{T}) = (x_{n-i},x_{n-i+1},...,x_{i-1})^{T}$, where operations on indices are modulo $n$.
The value of $\delta(n)$ and $p(n)$ come from the fact that matrix $\textbf{HD}_{1}$ is used as a $(\delta(n),p(n))$-balanced matrix and from Remark \ref{balanceness_remark}.
In that setting sequence $(\rho_{1},...,\rho_{n})$ is discrete and corresponds to the diagonal of $\textbf{D}_{2}$.
Thus we have: $K = 1$. To calculate $\Lambda_{F}$ and $\Lambda_{2}$, note first that matrix $\textbf{W}^{1}$ is defined as $\textbf{I}$ and subsequent $\textbf{W}^{i}$s are given as circulant shifts of the previous ones (i.e. each row is a circulant shift of the previous row). That observation comes directly from the circulant structure of $\textbf{G}_{circ}$. Thus we have: $\Lambda_{F} = O(\sqrt{n})$ and $\Lambda_{2} = O(1)$. The former is true since each $\textbf{A}^{i,j}$ has $O(n)$ nonzero entries and these are all $1$s. The latter is true since each nontrivial $\textbf{A}^{i,j}$ in that setting is an isometry (this comes straightforwardly from the definition of $\{\textbf{W}^{i}\}_{i=1,...,n}$).
Finally, all other conditions regarding $\textbf{W}^{i}$-matrices are clearly satisfied (each column of each $\textbf{W}^{i}$ has unit $L_{2}$ norm and corresponding columns from different $\textbf{W}^{i}$ and $\textbf{W}^{j}$ are clearly orthogonal).
Now let us consider the setting, where the structured matrix is of the form: $\sqrt{n}\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$.
In that case $\textbf{r}$ corresponds to a discrete vector (namely, the diagonal of $\textbf{D}_{2}$).
Linear mappings $\phi_{i}$ are defined as:
$\phi_{i}((x_{1},...,x_{n})^{T}) = (\sqrt{n}h_{i,1}x_{1},...,\sqrt{n}h_{i,n}x_{n})^{T}$, where $(h_{i,1},...,h_{i,n})^{T}$ is the $i^{th}$ row of $\textbf{H}$.
One can also notice that the set $\{\textbf{W}^{i}\}_{i=1,...,n}$ is defined as: $w^{i}_{a,b} = \sqrt{n} h_{i,a}h_{a,b}$.
Let us first compute the Frobenius norm of the matrix $\textbf{A}^{i,j}$ defined based on the aforementioned sequence $\{\textbf{W}^{i}\}_{i=1,...,n}$.
We have:
\begin{equation}
\|\textbf{A}^{i,j}\|_{F}^{2} = \sum_{l,t \in \{1,...,n\}}
(\sum_{k=1}^{n} w^{j}_{k,l}w^{i}_{k,t})^{2} = n^{2}
\sum_{l,t \in \{1,...,n\}} (\sum_{k=1}^{n} h_{j,k}h_{k,l}h_{i,k}h_{k,t})^{2}
\end{equation}
To compute the expression above, note first that for $r_{1} \neq r_{2}$ we have:
\begin{equation}
\theta = \sum_{k,l} h_{r_{1},k}h_{r_{1},l}h_{r_{2},k}h_{r_{2},l}
= \sum_{k} h_{r_{1},k}h_{r_{2},k} \sum_{l} h_{r_{1},l}h_{r_{2}, l} = 0,
\end{equation}
where the last equality comes from fact that different rows of $\ {H}$ are orthogonal. From the fact that $\theta = 0$ we get:
\begin{equation}
\|\textbf{A}^{i,j}\|_{F}^{2} = n^{2} \sum_{r=1,...,n} \sum_{k,l}
h_{i,r}^{2}h_{j,r}^{2}h_{r,k}^{2}h_{r,l}^{2} =
n \cdot n^{2} (\frac{1}{\sqrt{n}})^{8} \cdot n^{2} = n.
\end{equation}
Thus we have: $\Lambda_{F} \leq \sqrt{n}$.
Now we compute $\|\textbf{A}^{i,j}\|_{2}$.
Notice that from the definition of $\textbf{A}^{i,j}$ we get that
\begin{equation}
\textbf{A}^{i,j} = \textbf{E}^{i,j} \textbf{F}^{i,j},
\end{equation}
where the $l^{th}$ row of $\textbf{E}^{i,j}$ is of the form
$(h_{j,1}h_{1,l},...,h_{j,n}h_{n,l})$ and the $t^{th}$ column of
$\textbf{F}^{i,j}$ is of the form $(h_{i,1}h_{1,t},...,h_{i,n}h_{n,t})^{T}$.
Thus one can easily verify that $\textbf{E}^{i,j}$ and $\textbf{H}^{i,j}$ are isometries (since $\textbf{H}$ is) thus
$\textbf{A}^{i,j}$ is also an isometry and therefore $\Lambda_{2} = 1$. As in the previous setting, remaining conditions regarding matrices $\textbf{W}^{i}$ are trivially satisfied (from the basic properties of Hadamard matrices).
That completes the proof.
\end{proof}
\subsubsection{Proof of Theorem \ref{main_struct_theorem}}
Let us briefly give an overview of the proof before presenting it in detail. Challenges regarding proving accuracy results for structured matrices come from the fact that for any given $\textbf{x} \in \mathbb{R}^{n}$
different dimensions of $\textbf{y} = \textbf{G}_{struct}\textbf{x}$ are no longer independent (as it is the case for the unstructured setting).
For matrices from the \textit{TripleSpin}-family we can however show that with high probability different elements of $\textbf{y}$ correspond to projections of a given vector $\textbf{r}$ (see Section \ref{sec:model}) into directions that are close to orthogonal. The "close-to-orthogonality" characteristic is obtained with the use of the Hanson-Wright inequality that focuses on concentration results regarding quadratic forms involving vectors of sub-Gaussian random variables. If $\textbf{r}$ is Gaussian then from the well known fact that projections of the Gaussian vector into orthogonal directions are independent we can conclude that dimensions of $\textbf{y}$ are "close to independent". If $\textbf{r}$ is a discrete vector then we need to show that for $n$ large enough it "resembles" the Gaussian vector. This is where we need to apply the aforementioned techniques regarding multivariate Berry-Esseen-type central limit theorem results.
\begin{proof}
We will use notation from Section \ref{sec:model} and previous sections of the Appendix.
We assume that the model with structured matrices stacked vertically, each of $m$ rows is applied. Without loss of generality we can assume that we have just one block since different blocks are chosen independently.
Let $\textbf{G}_{struct}$ be a matrix from the \textit{TripleSpin}-family. Let us assume that $\textbf{G}_{struct}$ is used by a function $f$ operating in the $d$-dimensional space and let us denote by $\textbf{x}^{1}$,...,$\textbf{x}^{d}$ some fixed orthonormal basis of that space.
Our first goal is to compute: $\textbf{y}^{1} = \textbf{G}_{struct} \textbf{x}^{1},...,\textbf{y}^{d} = \textbf{G}_{struct} \textbf{x}^{d}$.
Denote by $\tilde{\textbf{x}}^{i}$ the linearly transformed version of $\textbf{x}$ after applying block $\textbf{B}_{1}$, i.e. $\tilde{\textbf{x}}^{i} = \textbf{B}_{1} \textbf{x}^{i}$.
Since $\textbf{B}_{2}$ is $(\delta(n),p(n))$-balanced), we conclude that with probability at least: $p_{balanced} \geq 1 - dp(n)$ each element of each $\tilde{\textbf{x}}^{i}$ has absolute value at most $\frac{\delta(n)}{\sqrt{n}}$. We shortly say that each $\tilde{\textbf{x}}^{i}$ is $\delta(n)$-balanced.
We call this event $\mathcal{E}_{balanced}$.
Note that by the definition of the \textit{TripleSpin}-family, each $\textbf{y}^{i}$ is of the form:
\begin{equation}
\textbf{y}^{i} = (\textbf{r}^{T} \cdot \phi_{1}(\textbf{B}_{2}\tilde{\textbf{x}}^{i}),...,\textbf{r}^{T} \cdot \phi_{m}(\textbf{B}_{2}\tilde{\textbf{x}}^{i}))^{T}.
\end{equation}
For clarity and to reduce notation we will assume that $\textbf{r}$ is $n$-dimensional.
To obtain results for vectors $\textbf{r}$ of different dimensionality $D$ it suffices to replace in our analysis and theoretical statements $n$ by $D$.
Let us denote $\mathcal{A} = \{\phi_{1}(\textbf{B}_{2}\tilde{\textbf{x}}^{1}),...,\phi_{m}(\textbf{B}_{2}\tilde{\textbf{x}}^{1}),...,\phi_{1}(\textbf{B}_{2}\tilde{\textbf{x}}^{d}),...,\phi_{m}(\textbf{B}_{2}\tilde{\textbf{x}}^{d}))\}$.
Our goal is to show that with high probability (in respect to random choices of $\textbf{B}_{1}$ and $\textbf{B}_{2}$) for all $\textbf{v}^{i},\textbf{v}^{j} \in \mathcal{A}$, $i \neq j$ the following is true:
\begin{equation}
\label{dot_product_equation}
|(\textbf{v}^{i})^{T} \cdot \textbf{v}^{j}| \leq t
\end{equation}
for some given $0 < t \ll 1$.
\begin{comment}
\textbf{Case 1:} vector \textbf{r} is Gaussian.
Let us denote $\mathcal{A} = \{\phi_{1}(\textbf{B}_{2}\tilde{\textbf{x}}^{i}),...,\phi_{m}(\textbf{B}_{2}\tilde{\textbf{x}}^{i}),...,\phi_{1}(\textbf{B}_{2}\tilde{\textbf{x}}^{d}),...,\phi_{m}(\textbf{B}_{2}\tilde{\textbf{x}}^{d}))\}$.
Assume first that we have for all $\textbf{v}^{i},\textbf{v}^{j} \in \mathcal{A}$, $i \neq j$:
\begin{equation}
\label{dot_product_equation}
|(\textbf{v}^{i})^{T} \cdot \textbf{v}^{j}| \leq t
\end{equation}
for some given $t > 0$.
We then perform Gram-Schmidt orthogonalization on the vectors from $\mathcal{A}$ and then we rescale transformed vectors so that they preserve their original lengths. Directly from the Gram-Schmidt orthogonalization procedure we get that after this transformation each $\textbf{v} \in \mathcal{A}$ was modified by adding a vector $\textbf{z}(\textbf{v})$ of length at most $\eta = t \cdot \xi(m,d)$, where $\xi(m,d)$ is some function of $m,d$ which does not depend on $n$.
The condition that (possibly after renormalization) the variances of the corresponding entries of vectors $\textbf{G}_{struct}\textbf{x}$
and $\textbf{Gx}$ are the same
implies that each element of $\textbf{G}_{struct}\textbf{x}$ has the same distribution as the corresponding element of $\textbf{Gx}$ since all these entries are Gaussian. Note also that from this fact we get:
\begin{equation}
(\textbf{G}\textbf{x}^{j})_{i} \sim \textbf{r}^{T} \cdot \textbf{x}^{j}
\end{equation}
for $j=1,...,d$ and $i=1,...,m$, where $v_{i}$ stands for the $i^{th}$ element of $\textbf{v}$.
Furthermore, since the orthogonalization was conducted in such a way that vectors from $\mathcal{A}$ did not change their lengths and from what we have just showed, we can conclude that for every given $i=1,...,m$:
\begin{equation}
\textbf{G}_{struct} \cdot \textbf{x}^{j} \sim \textbf{r}^{T} \cdot \textbf{x}^{j} + \textbf{r}^{T} \cdot \textbf{z}(\textbf{x}^{j})
\end{equation}
for some $\textbf{z}(\textbf{x}^{j})$ with $\|\textbf{z}(\textbf{x}^{j})\|_{2} \leq \eta$.
Let us write down $\textbf{G}_{struct} \cdot \textbf{x}^{i}$ as:
\begin{equation}
\textbf{G}_{struct} \cdot \textbf{x}^{i} = X_{random} + X_{noise},
\end{equation}
where: $X_{random} = \textbf{r}^{T} \cdot \textbf{x}^{j}$
and $X_{noise} = \textbf{r}^{T} \cdot \textbf{z}(\textbf{x}^{j})$.
Let us now calculate the probability that $|X_{random}| < \sqrt{\eta}$. Note that random variable $X_{random} = \textbf{r}^{T} \cdot \textbf{x}^{i}$ is Gaussian with variance $1$ (since it is a projection of the Gaussian vector on the unit $L_{2}$-norm vector $\textbf{x}^{i}$).
Thus the aforementioned probability is at most: $2\sqrt{\eta}$.
Thus, by the union bound, the probability $p_{1}$ that the scenario above does not happen for $i=1,...,d$ is at least: $p_{1} \geq 1 - 2d\sqrt{\eta}$.
Now let us calculate the probability that $|X_{noise}| \geq \eta^{\frac{1}{4}}$. Given $\textbf{z}(\textbf{x}^{j})$, $X_{noise}$
is Gaussian with variance at most $\eta$. Thus, from the inequality:
\begin{equation}
\mathbb{P}[|g| \geq x] \leq \frac{2}{\sqrt{2 \pi}} e^{-\frac{x^{2}}{2}},
\end{equation}
where $g \sim \mathcal{N}(0,1)$, we get:
\begin{equation}
\mathbb{P}[|X_{noise}| \geq \eta^{\frac{1}{4}}] \leq \frac{2}{\sqrt{2 \pi}} e^{-\frac{1}{2\sqrt{\eta}}}.
\end{equation}
Thus, by the union bound, the probability $p_{2}$ that $|X_{noise}| < \eta^{\frac{1}{4}}$ for $i=1,...,d$ is at least $p_{2} \geq 1 - \frac{4d}{\sqrt{2 \pi}} e^{-\frac{1}{2\sqrt{\eta}}}$.
We conclude, by the union bound, that with probability $p_{noise} \geq 1 - 2d\sqrt{\eta} - \frac{4d}{\sqrt{2 \pi}} e^{-\frac{1}{2\sqrt{\eta}}}$ we have:
\begin{equation}
\label{ratio_inequality}
\frac{X_{noise}}{X_{ratio}} \leq \eta^{\frac{1}{4}}.
\end{equation}
for $i=1,...,d$.
We call this event $\mathcal{E}_{2}$.
Now, if we take any vector $\textbf{v}$ in the linear space spanned by
$\{\textbf{x}^{1},...,\textbf{x}^{d}\}$ then it can be represented as:
$\textbf{v} = \alpha_{1} \textbf{x}^{1} + ... + \alpha_{d} \textbf{x}^{d}$
for some $\alpha_{1},...,\alpha_{d}$. Then from Inequality \ref{ratio_inequality}
we get that $\textbf{G}_{struct} \cdot \textbf{v} = \textbf{y} + \textbf{z}(\textbf{y})$, where: $\textbf{y} = \textbf{Gv}$, $|\textbf{z}(\textbf{y})|_{2} \leq \eta^{\frac{1}{4}} |\textbf{y}|_{2}$ and $\textbf{G}$ stands for the unstructured Gaussian matrix. To conclude that we use the well known fact that the projections of the given Gaussian vector on the orthogonal directions are independent.
\end{comment}
Fix some $t>0$. We would like to compute
the lower bound on the corresponding probability.
Let us fix two vectors $\textbf{v}^{1}, \textbf{v}^{2} \in \mathcal{A}$ and denote them as: $\textbf{v}^{1} = \phi_{i}(\textbf{B}_{2}\textbf{x})$, $\textbf{v}^{2} = \phi_{j}(\textbf{B}_{2} \textbf{y})$ for some $\textbf{x} = (x_{1},...,x_{n})^{T}$
and $\textbf{y} = (y_{1},...,y_{n})^{T}$.
Note that we have (see denotation from Section \ref{sec:model}):
\begin{equation}
\phi_{i}(\textbf{B}_{2}\textbf{x}) = (w^{i}_{11}\rho_{1}x_{1} + ... + w^{i}_{1,n}\rho_{n}x_{n},...,w^{i}_{n,1}\rho_{1}x_{1} + ... + w^{i}_{n,n}\rho_{n}x_{n})^{T}
\end{equation}
and
\begin{equation}
\phi_{j}(\textbf{B}_{2}\textbf{y}) = (w^{j}_{11}\rho_{1}y_{1} + ... + w^{j}_{1,n}\rho_{n}y_{n},...,w^{j}_{n,1}\rho_{1}y_{1} + ... + w^{j}_{n,n}\rho_{n}y_{n})^{T}.
\end{equation}
We obtain:
\begin{equation}
(\textbf{v}^{1})^{T} \cdot \textbf{v}^{2}=
\sum_{l \in \{1,...,n\}, u\in \{1,...,n\}} \rho_{l}\rho_{u}(\sum_{k=1}^{n}x_{l}y_{u}w^{i}_{k,u}w^{j}_{k,l}).
\end{equation}
We show now that under assumptions from Theorem \ref{main_struct_theorem} the expected
value of the above expression is $0$.
We have:
\begin{equation}
\mathbb{E}[(\textbf{v}^{1})^{T} \cdot \textbf{v}^{2}]=
\mathbb{E}[\sum_{l \in \{1,...,n\}} \rho_{l}^{2}x_{l}y_{l}
(\sum_{k=1}^{n}w^{i}_{k,l}w^{j}_{k,l})],
\end{equation}
since $\rho_{1},...,\rho_{n}$ are independent
and have expectations equal to $0$.
Now notice that if $i \neq j$ then from the assumption that
corresponding columns of matrices $\textbf{W}^{i}$ and $\textbf{W}^{j}$ are orthogonal we get that the above expectation is $0$. Now assume that $i = j$. But then $\textbf{x}$ and $\textbf{y}$ have to be different and thus they are orthogonal (since they are taken from the orthonormal system transformed by an isometry).
In that setting we get:
\begin{equation}
\mathbb{E}[(\textbf{v}^{1})^{T} \cdot \textbf{v}^{2}]=
\mathbb{E}[\sum_{l \in \{1,...,n\}} \rho_{l}^{2}x_{l}y_{l}
(\sum_{k=1}^{n}(w^{i}_{k,l})^{2})]= \tau w \sum_{l=1}^{n} x_{l}y_{l} = 0,
\end{equation}
where $\tau$ stands for the second moment of each $\rho_{i}$,
$w$ is the squared $L_{2}$-norm of each column of $\textbf{W}^{i}$
($\tau$ and $w$ are well defined due to the properties of the \textit{TripleSpin}-family). The last inequality comes from the fact that $\textbf{x}$ and $\textbf{y}$ are orthogonal.
Now if we define matrices $\textbf{A}^{i,j}$ as in the definition of the \textit{TripleSpin}-model then we see that
\begin{equation}
(\textbf{v}^{1})^{T} \cdot \textbf{v}^{2} =
\sum_{l,u \in \{1,...,n\}} \rho_{l}\rho_{u}T^{i,j}_{l,u},
\end{equation}
where:
$T^{i,j}_{l,u} = x_{l}y_{u}A^{i,j}_{l,u}$.
\begin{comment}
\todo{Krzysztof}{Here is an alternate, more compact calculation without explicit sums starting after Note that we have (see denotation from Section \ref{sec:model}) up to Hanson-Wright. Please choose a variant.}
Since $\textbf{v}^1 = \textbf{W}^{i}\textbf{D}_{x}\rho$
we obtain
\begin{equation}
(\textbf{v}^{1})^{T} \cdot \textbf{v}^{2}=
\rho^T\textbf{D}_{x}(\textbf{W}^{i})^T\textbf{W}^{j}\textbf{D}_{y}\rho.
\end{equation}
Observe that $\rho_{1},...,\rho_{n}$ are independent, their expectation is $0$. Let $\tau$ stand for the common second moment of each $\rho_{i}$.
For any fixed matrix $\textbf{M}$ it's easy to see that $\mathbb{E}[\rho^T\textbf{M}\rho] = \tau \cdot \mbox{tr}(\textbf{M})$, where $\mbox{tr}$ is the trace operator.
If $i \neq j$ then from the assumption that diagonal entries of $(\textbf{W}^{i})^T\textbf{W}^{j}$ are $0$ we get that $\mbox{tr}(\textbf{D}_{x}(\textbf{W}^{i})^T\textbf{W}^{j}\textbf{D}_{y}) = 0$.
Now assume that $i = j$.
But then $\textbf{x}$ and $\textbf{y}$ have to be different and thus they are orthogonal (since they are taken from the orthonormal system transformed by an isometry).
In that setting since columns of $W^{i}$ are unit length and since both $\textbf{D}_x$ and $\textbf{D}_y$ are diagonal we get
$\mbox{tr}(\textbf{D}_{x}(\textbf{W}^{i})^T\textbf{W}^{i}\textbf{D}_{y})=\mbox{tr}(\textbf{D}_{x}\textbf{I}\textbf{D}_{y})= \textbf{x}^T\textbf{y} = 0$.
Thus we have:
\begin{equation}
\mathbb{E}[(\textbf{v}^{1})^{T} \cdot \textbf{v}^{2}]= 0.
\end{equation}
\end{comment}
Now we will use the following inequality:
\begin{theorem}[Hanson-Wright Inequality]
Let $\textbf{X} = (X_{1},...,X_{n})^{T} \in \mathbb{R}^{n}$ be a random vector with independent components $X_{i}$ which satisfy: $\mathbb{E}[X_{i}] = 0$ and have sub-Gaussian norm at most $K$ for some given $K>0$.
Let $\textbf{A}$ be an $n \times n$ matrix. Then for every $t \geq 0$
the following is true:
\begin{equation}
\mathbb{P}[\textbf{X}^{T}\textbf{A}\textbf{X} - \mathbb{E}[\textbf{X}^{T}\textbf{AX}] > t] \leq
2e^{-c \min(\frac{t^{2}}{K^{4}\|\textbf{A}\|^{2}_{F}},
\frac{t}{K^{2}\|\textbf{A}\|_{2}})},
\end{equation}
\end{theorem}
where $c$ is some universal positive constant.
Note that, assuming $\delta(n)$-balancedness, we have: $\|\textbf{T}^{i,j}\|_{F} \leq \frac{\delta^{2}(n)}{n} \|\textbf{A}^{i,j}\|_{F}$ and $\|\textbf{T}^{i,j}\|_{2} \leq
\frac{\delta^{2}(n)}{n}
\|\textbf{A}^{i,j}\|_{2}$.
Now we take $\textbf{X} = (\rho_{1},...,\rho_{n})^{T}$ and $\textbf{A} = \textbf{T}^{i,j}$ in the theorem above.
Applying the Hanson-Wright inequality in that setting,
taking the union bound over all pairs of different vectors $\textbf{v}^{i},\textbf{v}^{j} \in \mathcal{A}$ (this number is exactly: ${md \choose 2}$) and the event $\mathcal{E}_{balanced}$, finally taking the union bound over all $s$ functions $f_{i}$, we conclude that with probability at least:
\begin{equation}
\label{imp_equation}
p_{good} = 1 - p(n)ds - 2{md \choose 2}s e^{-\Omega(\min(\frac{t^{2}n^{2}}{K^{4}\Lambda_{F}^{2}\delta^{4}(n)}, \frac{tn}{K^{2}\Lambda_{2} \delta^{2}(n)}))}
\end{equation}
for every $f$ any two different vectors $\textbf{v}^{i}, \textbf{v}^{j} \in \mathcal{A}$ satisfy:
$|(\textbf{v}^{i})^{T} \cdot \textbf{v}^{j}| \leq t$.
Note that from the fact that $\textbf{B}_{2}\textbf{B}_{1}$ is $(\delta(n),p(n))$-balanced and from Equation \ref{imp_equation}, we get that with probability at least:
\begin{equation}
p_{right} = 1 - 2p(n)ds - 2{md \choose 2}s e^{-\Omega(\min(\frac{t^{2}n^{2}}{K^{4}\Lambda_{F}^{2}\delta^{4}(n)}, \frac{tn}{K^{2}\Lambda_{2} \delta^{2}(n)}))}.
\end{equation}
for every $f$ any two different vectors $\textbf{v}^{i}, \textbf{v}^{j} \in \mathcal{A}$ satisfy:
$|(\textbf{v}^{i})^{T} \cdot \textbf{v}^{j}| \leq t$ and furthermore each $\textbf{v}^{i}$ is $\delta(n)$-balanced.
Assume now that this event happens.
Consider the vector
\begin{equation}
\textbf{q}^{\prime} = ((\textbf{y}^{1})^{T},...,(\textbf{y}^{d})^{T})^{T} \in \mathbb{R}^{md}.
\end{equation}
Note that $\textbf{q}^{\prime}$ can be equivalently represented as:
\begin{equation}
\textbf{q}^{\prime} =
(\textbf{r}^{T} \cdot \textbf{v}^{1},...,\textbf{r}^{T} \cdot \textbf{v}^{md}),
\end{equation}
where: $\mathcal{A} = \{\textbf{v}^{1},...,\textbf{v}^{md}\}$.
From the fact that $\phi_{i}\textbf{B}_{2}$ and $\textbf{B}_{1}$ are isometries we conclude that: $\|\textbf{v}^{i}\|_{2} = 1$ for $i=1,...$.
Now we will need the following Berry-Esseen type result for random vectors:
\begin{theorem}[Bentkus \cite{bentkus2003dependence}]
\label{clt_theorem}
Let $\textbf{X}_{1},...,\textbf{X}_{n}$ be independent vectors taken from $\mathbb{R}^{k}$ with common mean $\mathbb{E}[\textbf{X}_{i}] = 0$. Let $\textbf{S} = \textbf{X}_{1} + ... + \textbf{X}_{n}$.
Assume that the covariance operator $\textbf{C}^{2} = cov(\textbf{S})$ is invertible. Denote $\beta_{i} =
\mathbb{E}[\|\textbf{C}^{-1}\textbf{X}_{i}\|_{2}^{3}]$
and $\beta = \beta_{1} + ... + \beta_{n}$.
Let $\mathcal{C}$ be the set of all convex subsets of $\mathbb{R}^{k}$. Denote $\Delta(\mathcal{C}) = \sup_{A \in \mathcal{C}} |\mathbb{P}[S \in A]-\mathbb{P}[Z \in A]|$,
where $Z$ is the multivariate Gaussian distribution with mean $0$ and covariance operator $\textbf{C}^{2}$. Then:
\begin{equation}
\Delta(\mathcal{C}) \leq ck^{\frac{1}{4}} \beta
\end{equation}
for some universal constant $c$.
\end{theorem}
Denote: $\textbf{X}_{i} = (r_{i}v^{1}_{i},...,r_{i}v^{k}_{i})^{T}$
for $k=md$, $\textbf{r} = (r_{1},...,r_{n})^{T}$
and $\textbf{v}^{j} = (v^{j}_{1},...,v^{j}_{n})$.
Note that $\textbf{q}^{\prime} = \textbf{X}_{1} + ... + \textbf{X}_{n}$. Clearly we have: $\mathbb{E}[\textbf{X}_{i}] = 0$.
Furthermore, given the choices of $\textbf{v}^{1},...,\textbf{v}^{k}$, random vectors $\textbf{X}_{1},..,\textbf{X}_{n}$ are independent.
Let us calculate now the covariance matrix of $\textbf{q}^{\prime}$.
We have:
\begin{equation}
\textbf{q}^{\prime}_{i} = r_{1}v^{i}_{1} + ... + r_{n}v^{i}_{n},
\end{equation}
where: $\textbf{q}^{\prime} = (\textbf{q}^{\prime}_{1},...,\textbf{q}^{\prime}_{k})$.
Thus for $i_{1}, i_{2}$ we have:
\begin{equation}
\mathbb{E}[\textbf{q}^{\prime}_{i_{1}} \textbf{q}^{\prime}_{i_{2}}] =
\sum_{j=1}^{n} v^{i_{1}}_{j}v^{i_{2}}_{j}\mathbb{E}[r_{j}^{2}] + 2\sum_{1 \leq j_{1} < j_{2} \leq n} v^{i_{1}}_{j_{1}}v^{i_{2}}_{j_{2}} \mathbb{E}[r_{j_{1}}r_{j_{2}}]
= (\textbf{v}^{i_{1}})^{T} \cdot \textbf{v}^{i_{2}},
\end{equation}
where the last equation comes from the fact $r_{j}$ are either Gaussian from $\mathcal{N}(0,1)$ or discrete with entries from $\{-1,+1\}$ and furthermore different $r_{j}$s are independent.
Therefore if $i_1=i_2=i$, since each $\textbf{v}^{i}$ has unit $L_{2}$-norm, we have that
\begin{equation}
\mathbb{E}[\textbf{q}^{\prime}_{i} \textbf{q}^{\prime}_{i}] = 1,
\end{equation}
and for $i_{1} \neq i_{2}$ we get:
\begin{equation}
|\mathbb{E}[\textbf{q}^{\prime}_{i_{1}} \textbf{q}^{\prime}_{i_{2}}]| \leq t.
\end{equation}
We conclude that the covariance matrix $\Sigma_{\textbf{q}^{\prime}}$ of the distribution $\textbf{q}^{\prime}$ is a matrix with entries $1$ on the diagonal and other entries of absolute value at most $t$.
\begin{comment}
Note that if $cov(\textbf{q}^{\prime}) = \textbf{C}^{2}$ then
from the Taylor expansion we get:
\begin{equation}
\textbf{C}^{-1} = \textbf{I} - \textbf{T} + \textbf{T}^{2} - \textbf{T}^{3} + ...,
\end{equation}
where: $T = \frac{1}{2}\textbf{E}_{t} - \frac{1}{8}\textbf{E}_{t}^{2} + \frac{1}{16}\textbf{E}^{3} - ...$
and $\textbf{E}_{t}$ is a matrix with $0$s on the diagonal and all other entries of absolute value at most $t$.
\end{comment}
For $t = o_{k}(1)$ small enough
and from the $\delta(n)$-balancedness of vectors $\textbf{v}^{1},...,\textbf{v}^{k}$
we can conclude that:
\begin{equation}
\mathbb{E}[\|\textbf{C}^{-1}\textbf{X}_{i}\|^{3}_{2}] =
O(\mathbb{E}[\|\textbf{X}_{i}\|^{3}_{2}]) = O(\sqrt{(\frac{k}{n})^{3}}\delta^{3}(n)),
\end{equation}
\begin{comment}
\todo{Krzysztof}{Can we gain anything here, e.g. 2 powers of $\delta$, if true, by bounding as $\delta(n)$ times $\sum_i\mathbb{E}[\|\textbf{X}_{i}\|^{2}_{2}]$, which is hopefully easy and nice.}
\end{comment}
Now, using Theorem \ref{clt_theorem}, we conclude that
\begin{equation}
\sup_{A \in \mathcal{C}} |\mathbb{P}[\textbf{q}^{\prime} \in A] - \mathbb{P}[Z \in A]| = O(k^{\frac{1}{4}}n \cdot \frac{k^{\frac{3}{2}}}{n^{\frac{3}{2}}} \delta^{3}(n)) = O(\frac{\delta^{3}(n)}{\sqrt{n}}k^{\frac{7}{4}}),
\end{equation}
where $Z$ is taken from the multivariate Gaussian distribution with covariance matrix $\textbf{I} + \textbf{E}$. Now if we take $\eta = \frac{\delta^{3}(n)}{\sqrt{n}}k^{\frac{7}{4}}$, $\epsilon = t = o_{md}(1)$ and take $n$ large enough, the statement of the theorem follows.
\end{proof}
\subsubsection{Proof of Theorem \ref{corollary_theorem}}
\begin{proof}
This comes directly from Theorem \ref{main_struct_theorem} and Lemma \ref{simple_lemma}.
\end{proof}
\subsubsection{Proof of Theorem \ref{hopefully_last_theorem}}
\begin{proof}
For clarity we will assume that the structured matrix consists of just one block of $m$ rows and will compare its performance with the unstructured variant of $m$ rows (the more general case when the structured matrix is obtained by stacking vertically many blocks is analogous since the blocks are chosen independently).
Consider the two-dimensional linear space $\mathcal{H}$ spanned by $\textbf{x}$ and $\textbf{y}$.
Fix some orthonormal basis $\mathcal{B} = \{\textbf{u}^{1},\textbf{u}^{2}\}$
of $\mathcal{H}$.
Take vectors \textbf{q} and $\textbf{q}^{\prime}$.
Note that they are $2m$-dimensional, where $m$ is the number of rows of the block used in the structured setting.
From Theorem \ref{corollary_theorem} we conclude that will probability
at least $p_{success}$, where $p_{success}$ is as in the statement of the theorem the following holds for any convex $2m$-dimensional set $A$:
\begin{equation}
|\mathbb{P}[\textbf{q}(\epsilon) \in A] - \mathbb{P}[\textbf{q}^{\prime}]| \leq \eta,
\end{equation}
where $\eta = \frac{\log^{3}(n)}{n^{\frac{2}{5}}}$.
Take two corresponding entries of vectors $\textbf{v}_{\textbf{x},\textbf{y}}^{1}$ and $\textbf{v}_{\textbf{x},\textbf{y}}^{2}$ indexed by a pair $(\textbf{e}_{i}, \textbf{e}_{j})$ for some fixed $i,j \in \{1,...,m\}$.
Call them $p^{1}$ and $p^{2}$ respectively. Our goal is to compute
$|p^{1} - p^{2}|$.
Notice that $p^{1}$ is the probability that $h(\textbf{x}) = \textbf{e}_{i}$
and $h(\textbf{y}) = \textbf{e}_{j}$ for the unstructured setting and $p^{2}$ is that probability for the structured variant.
Let us consider now the event $E^{1} = \{h(\textbf{x}) = \textbf{e}_{i}
\land h(\textbf{y}) = \textbf{e}_{j}\}$, where the setting is unstructured.
Denote the corresponding event for the structured setting as $E^{2}$.
Denote $\textbf{q} = (q_{1},...,q_{2m})$ (for vectors of the form: $-\textbf{e}_{i}$ and $-\textbf{e}_{j}$ the analysis is exactly the same).
Assume that $\textbf{x} = \alpha_{1} \textbf{u}^{1} + \alpha_{2} \textbf{u}^{2}$ for some scalars $\alpha_{1}, \alpha_{2} > 0$.
Denote the unstructured Gaussian matrix by $\textbf{G}$.
We have:
\begin{equation}
\textbf{Gx} = \alpha_{1}\textbf{G}\textbf{u}^{1} +
\alpha_{2}\textbf{G}\textbf{u}^{2}
\end{equation}
Note that we have: $\textbf{G}\textbf{u}^{1} = (q_{1},...,q_{m})^{T}$
and $\textbf{G}\textbf{u}^{2} = (q_{m+1},...,q_{2m})^{T}$.
Denote by $A(\textbf{e}_{i})$ the set of all the points in $\mathbb{R}^{m}$ such that their angular distance to $\textbf{e}_{i}$ is at most the angular distance to all other $m-1$ canonical vectors. Note that this is definitely the convex set.
Now denote:
\begin{equation}
Q(\textbf{e}_{i}) = \{(q_{1},...,q_{2m})^{T} \in \mathbb{R}^{2m} : \alpha_{1} (q_{1},...,q_{m})^{T} + \alpha_{2} (q_{m+1},...,q_{2m})^{T} \in A(\textbf{e}_{i})\}.
\end{equation}
Note that since $A(\textbf{e}_{i})$ is convex, we can conclude that $Q(\textbf{e}_{i})$ is also convex.
Note that
\begin{equation}
\{h(\textbf{x}) = \textbf{e}_{i}\} = \{\textbf{q} \in Q(\textbf{e}_{i})\}.
\end{equation}
By repeating the analysis for the event $\{h(\textbf{y}) = \textbf{e}_{j}\}$,
we conclude that:
\begin{equation}
\{h(\textbf{x}) = \textbf{e}_{i} \land h(\textbf{y}) = \textbf{e}_{j} \} =
\{\textbf{q} \in Y(\textbf{e}_{i},\textbf{e}_{j})\}
\end{equation}
for convex set $Y(\textbf{e}_{i},\textbf{e}_{j}) = Q(\textbf{e}_{i}) \cap Q(\textbf{e}_{j})$.
Now observe that
\begin{equation}
|p^{1}-p^{2}| =
|\mathbb{P}[\textbf{q} \in Y(\textbf{e}_{i},\textbf{e}_{j})] -
\mathbb{P}[\textbf{q}^{\prime} \in Y(\textbf{e}_{i},\textbf{e}_{j})]|
\end{equation}
Thus we have:
\begin{equation}
|p^{1} - p^{2}| \leq
|\mathbb{P}[\textbf{q} \in Y(\textbf{e}_{i},\textbf{e}_{j})] -
\mathbb{P}[\textbf{q}(\epsilon) \in Y(\textbf{e}_{i},\textbf{e}_{j})]|+
|\mathbb{P}[\textbf{q}(\epsilon) \in Y(\textbf{e}_{i},\textbf{e}_{j})] -
\mathbb{P}[\textbf{q}^{\prime} \in Y(\textbf{e}_{i},\textbf{e}_{j})]|
\end{equation}
Therefore we have:
\begin{equation}
|p^{1} - p^{2}| \leq |\mathbb{P}[\textbf{q} \in Y(\textbf{e}_{i},\textbf{e}_{j})] -
\mathbb{P}[\textbf{q}(\epsilon) \in Y(\textbf{e}_{i},\textbf{e}_{j})]| + \eta.
\end{equation}
Thus we just need to upper-bound:
\begin{equation}
\xi = |\mathbb{P}[\textbf{q} \in Y(\textbf{e}_{i},\textbf{e}_{j})] -
\mathbb{P}[\textbf{q}(\epsilon) \in Y(\textbf{e}_{i},\textbf{e}_{j})]|.
\end{equation}
Denote the covariance matrix of the distribution $\textbf{q}(\epsilon)$
as $\textbf{I} + \textbf{E}$. Note that $\textbf{E}$ is equal to $0$ on the diagonal and the
absolute value of all other off-diagonal entries is at most $\epsilon$.
Denote $k = 2m$. We have
\begin{equation}
\xi = \left|\frac{1}{(2 \pi)^{\frac{k}{2}}\sqrt{\det (I+E)}} \int_{Y(\textbf{e}_{i},\textbf{e}_{j})}
e^{-\frac{\textbf{x}^{T}(\textbf{I}+\textbf{E})^{-1}\textbf{x}}{2}}d\textbf{x} -
\frac{1}{(2 \pi)^{\frac{k}{2}}} \int_{Y(\textbf{e}_{i},\textbf{e}_{j})}
e^{-\frac{\textbf{x}^{T}\textbf{x}}{2}}d\textbf{x}\right|.
\end{equation}
Expanding: $(\textbf{I}+\textbf{E})^{-1} = \textbf{I} - \textbf{E} + \textbf{E}^{2} - ...$, noticing that $|\det(I + E) - 1| = O(\epsilon^{2m})$,
and using the above formula, we easily get:
\begin{equation}
\xi = O(\epsilon).
\end{equation}
That completes the proof.
\end{proof}
\subsection{Newton sketches - details of experiments}
Key attributes of the Newton sketch approach are a guaranteed
super-linear convergence with exponentially high probability for self-concordant
functions, and a reduced computational complexity compared to the original
second-order Newton method. Such characteristics are achieved using a sketched
version of the Hessian matrix, in place of the original one. In the proposed experiment,
the unconstrained large scale logistic regression is considered. Mathematically
given a set of $n$ observations $\{(a_i,y_i)\}_{i=1..n}$, with $a_i \in
\mathbb{R}^d$ and $y_i \in \{-1,1\}$, the logistic regression problem amounts to
finding $x \in \mathbb{R}^d$ minimizing the cost function
$$ f(x) = \sum_{i=1}^n \log (1 + \exp(-y_i a_i^T x)) \enspace .$$
The Newton approach to solving this optimization problem entails solving at each iteration the least squares equation $\nabla^2 f(x^t) \Delta^t = - \nabla f(x^t) $, where $$\nabla^2 f(x^t) = A^T \mathrm{diag} \left( \frac{1}{1+\exp(-a_i^T x)} (1 - \frac{1}{1+\exp(-a_i^T x)}) \right) A \in \mathbb{R}^{d \times d}$$ is the Hessian matrix of $f(x^t)$, $A = [a_1^T a_2^T \cdots a_n^T] \in \mathbb{R}^{n \times d}$, $\Delta^t = x^{t+1} - x^t$ is the increment at iteration $t$ and $\nabla f(x^t) \in \mathbb{R}^d$ is the gradient of the cost function. Authors of~\cite{pilanci} propose to rather consider the sketched version of the least square equation, based on a Hessian square root of $\nabla^2 f(x^t)$, denoted $\nabla^2 f(x^t)^{1/2} = \mathrm{diag} \left( \frac{1}{1+\exp(-a_i^T x)} (1 - \frac{1}{1+\exp(-a_i^T x)}) \right)^{1/2} A \in \mathbb{R}^{n \times d}$. The least squares problem becomes at each iteration $t$
$$ \left( (S^t \nabla^2 f(x^t)^{1/2})^T S^t \nabla^2 f(x^t)^{1/2} \right) \Delta^t = - \nabla f(x^t) \enspace ,$$
where $S^t \in \mathbb{R}^{m \times n}$ is a sequence of isotropic sketch matrices.
Let's finally recall that the gradient of the cost function is
$$ \nabla f(x^t) = \sum_{i=1}^n \left( \frac{1}{1+\exp(-y_i a_i^T x)}-1 \right) y_i a_i \enspace . $$
\subsection{Further experiments for kernel approximation}
For kernel approximation experiments, we used also the G50C dataset which contains 550 points of dimensionality 50 ($n = 50$) drawn from multivariate Gaussians (for Gaussian kernel, bandwidth $\sigma$ is set to $17.4734$). The results are averaged over $10$ runs.
\paragraph{Results on the G50C dataset:} The following matrices have been tested: Gaussian random matrix $\textbf{G}$, $\textbf{G}_{Toeplitz}\textbf{D}_{2}\textbf{HD}_{1}$ $\textbf{G}_{skew-circ}\textbf{D}_{2}\textbf{HD}_{1}$, $\textbf{HD}_{g_{1},...,g_{n}}\textbf{HD}_{2}\textbf{HD}_{1}$ and $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$.
In Figure \ref{kernel_approx}, for the Gaussian kernel, all curves are almost identical. For both kernels, all \textit{TripleSpin}-matrices perform similarly to a random Gaussian matrix.
$\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ performs better than a random matrix and other \textit{TripleSpin}-matrices for a wide range of sizes of random feature maps.
\begin{figure}[!t]
\centering
\includegraphics[scale = 0.19]{./images/gaussiankernel.pdf}
\includegraphics[scale = 0.19]{./images/angularkernel.pdf}
\caption{Accuracy of random feature map kernel approximation}
\label{kernel_approx}
\end{figure}
\section{Computing general kernels with \textit{TripleSpin}-matrices}
Previous works regarding approximating kernels with structured matrices covered only some special kernels, namely: Gaussian, arc-cosine and angular kernels.
We explain here how structured approach (in particular our \textit{TripleSpin}-family) can be used to approximate well most kernels.
Theoretical guarantees that cover also this special case are given in the subsequent section.
For kernels that can be represented as an expectation of Gaussian random variables it is natural to approximate them using structured matrices. We start our analysis with the so-called \textit{Pointwise Nonlinear Gaussian kernels (PNG)} which are of the form:
\begin{equation}
\label{eq:main_function_simple}
\kappa_{f,\mu,\Sigma}(\textbf{x},\textbf{y}) = \mathbb{E}\left[f(\textbf{g}^\top \cdot \textbf{x})\cdot f(\textbf{g}^\top \cdot \textbf{y})\right]
\end{equation}
where: $\textbf{x},\textbf{y} \in \mathbb{R}^{n}$, the expectation is over a multivariate Gaussian $\textbf{g}\sim\mathcal{N}(\mu,\Sigma)$ and $f$ is a fixed nonlinear function $\mathbb{R} \rightarrow \mathbb{R}$ (the positive-semidefiniteness of the kernel follows from it being an expectation of dot products).
The expectation, interpreted in the Monte-Carlo approximation setup, leads to an unbiased estimation by a sum of the normalized dot products $\frac{1}{k}f(\textbf{G} \textbf{x})^\top\cdot f(\textbf{G} \textbf{y})$, where $\textbf{G} \in \mathbb{R}^{k \times n}$ is a random Gaussian matrix and $f$ is applied pointwise, i.e. $f((v_{1},...,v_{k})):=(f(v_{1}),...,f(v_{k}))$.
In this setting matrices from the $TripleSpin$-family can replace $\textbf{G}$ with diagonal covariance matrices $\Sigma$ to speed up computations and reduce storage complexity.
This idea of using random structured matrices to evaluate such kernels is common in the literature, e.g. \cite{choromanska2015binary,le2013fastfood}, but no theoretical guarantees were given so far for general PNG kernels (even under the restriction of diagonal $\Sigma$).
PNGs define a large family of kernels characterized by $f,\mu,\Sigma$.
Prominent examples include the Euclidean and angular distance \cite{choromanska2015binary}, the arc-cosine kernel \cite{NIPS2009_3628} and the sigmoidal neural network~\cite{Williams:1998:CIN:295919.295941}.
Since sums of kernels are again kernels, we can construct an even larger family of kernels by summing PNGs. A simple example is the Gaussian kernel which can be represented as a sum of two PNGs with $f$ replaced by the trigonometric functions: $\sin$ and $\cos$.
Since the Laplacian, exponential and rational quadratic kernel can all be represented as a mixture of
Gaussian kernels with different variances,
they can be easily approximated by a finite sum of PNGs. Remarkably, virtually all kernels can be represented as a (potentially infinite) sum of PNGs with diagonal $\Sigma$.
Recall that kernel $\kappa$ is stationary if $\kappa(x,y)=\kappa(x-y)$ for all $x,y \in \mathbb{R}^{n}$ and non-stationary otherwise.
Harnessing Bochner's and Wiener’s Tauberian theorem~\cite{samo2015generalized}, we show that all stationary kernels may be approximated arbitrarily well by sums of PNGs.
\begin{theorem}[stationary kernels]
\label{thm:stationary}
The family of functions
\begin{align*}
\kappa_K(\textbf{x},\textbf{y})&:=\sum_{k=1}^K\alpha_k (\mathbb{E}[\cos(\textbf{g}_k^\top\cdot \textbf{x})\cos(\textbf{g}_k^\top\cdot \textbf{y})] + \mathbb{E}[\sin(\textbf{g}_k^\top\cdot \textbf{x})\sin(\textbf{g}_k^\top\cdot \textbf{y})])
\end{align*}
with $\textbf{g}_k\sim\mathcal{N}(\mu_k,\mbox{diag}((\sigma_{k}^1)^2,...,(\sigma_k^d)^2))$, $\mu_k,\sigma_k\in\mathbb{R}^{n}$, $\alpha_k\in\mathbb{R}$, $K\in\mathbb{N}\cup\{\infty\}$ is dense in the family of stationary real-valued kernels with respect to pointwise convergence.
\end{theorem}
This family corresponds to the \textit{spectral mixture kernels} of \cite{wilson2013gaussian}. We can extend these results to arbitrary, non-stationary, kernels; the precise statement and its proof can be found in the Appendix.
Theorem \ref{thm:stationary} and its non-stationary analogue show that the family of sums of PNGs contains virtually all kernel functions. If a kernel can be well approximated by the sum of only a small number of PNGs then we can use the $TripleSpin$-family to efficiently evaluate it. It has been found in the Gaussian Process literature that often a small sum of kernels tuned using training data is sufficient to explain phenomena remarkably well \cite{wilson2013gaussian}. The same should also be true for the sum of PNGs. This suggests a methodology for learning the parameters of a sum of PNGs from training data and then applying it out of sample.
We leave this investigation for future research.
\begin{comment}
\subsection{Neural networks with structured layers}
In deep neural network architectures some fully-connected layers, usually responsible for final computations, are not learned (due to computational bottlenecks), but their weights are constructed randomly from the Gaussian distribution. Those layers can be then replaced by $3StructBlock$-matrices. Furthermore, fully connected random layers can be also used at the beginning of the overall pipeline and serve as a preprocessing mechanism reducing data dimensionality while preserving important signal from the data \cite{choromanska2015binary}. These can be also replaced by their structured counterparts, in particular by $3StructBlock$-matrices.
\subsection{Structured random projection trees}
Random projection trees (RPTs) are effective tools for partitioning data of low intrinsic dimensionality $d$ even if the explicit dimensionality $D$ is large. The rule of the split in each node is based on the following simple randomized procedure. A projection vectors $\textbf{g}$ are chosen uniformly at random from the unit sphere.
For each data point $\textbf{x} \in \mathbb{R}^{D}$ a projection $\textbf{g}^\top \cdot \textbf{x}$ is computed and the median $m$ of the set of projections $\mathcal{P} = \{\textbf{g}^\top \cdot \textbf{x} : \textbf{x} \in \mathcal{X}\}$ is calculated. A threshold $\theta$ is chosen as: $\theta = m + \frac{6r \Delta_{\mathcal{X}}}{\sqrt{D}}$, where $r$ is chosen uniformly at random from the interval $[-1,1]$
and $\Delta_{\mathcal{X}}$ is a diameter of the dataset $\mathcal{X}$. Note that random projections vectors can be computed in advance before data is seen. In the structured RPT setting for each level of the tree the same projection vector $\textbf{g}_{struct}$ is chosen. The set of this vectors is constructed as the set of rows of the structured matrix from the \textit{3StructBlock}-family. That significantly reduces the time for costly computations of projections. Matrix from the \textit{3StructBlock}-family is also used to replace another costly computation of $\Delta_{\mathcal{X}}$ by its approximation $\Delta^{struct}_{\mathcal{X}}$. The approximation can be obtained from structured projections in a way proposed in \cite{diff_rpt}.
\end{comment}
\section{Conclusions}
We showed a generic structured computational framework that can be used with a variety of machine learning algorithms to speed them up.
Some special cases of that framework (often applied as heuristics verified experimentally to work in practice) were known before, but the general picture
and a comprehensive theoretical explanation were missing. Theory developed by us in this paper enables us not only to explain the efficiency of several special structured
cases known so far, but also to propose for many applications efficient and compact models that were not considered before.
\end{comment}
\section{Experiments}
\label{sec:experiments}
Experiments have been carried out on a single processor machine (Intel Core i7-5600U CPU @ 2.60GHz, 4 hyper-threads) with 16GB RAM for the first two applications and a dual processor machine (Intel Xeon E5-2640 v3 @ 2.60GHz, 32 hyper-threads) with 128GB RAM for the last one. Every experiment was conducted using Python. In particular, NumPy is linked against a highly optimized BLAS library (Intel MKL). Fast Fourier Transform is performed using numpy.fft and Fast Hadamard Transform is using ffht from~\cite{indyk2015}. To have fair comparison, we have set up: $\mathrm{OMP\_NUM\_THREADS}=1$ so that every experiment is done on a single thread. Every parameter corresponding to a \textit{TripleSpin}-matrix is computed in advance, such that obtained speedups take only matrix-vector products into account.
All figures should be viewed in color.
\subsection{Locality-Sensitive Hashing (LSH) application}
In the first experiment, to support our theoretical results for the \textit{TripleSpin}-matrices in the cross-polytope LSH, we compared collision probabilities in Figure \ref{fig:collision} for the low dimensional case. Results are shown for one hash function (averaged over $100$ runs). For each interval, collision probability has been computed for $20 000$ points for a random Gaussian matrix $\textbf{G}$ and four other types of \textit{TripleSpin}-matrices (descending order of number of parameters): $\textbf{G}_{Toeplitz}\textbf{D}_{2}\textbf{HD}_{1}$, $\textbf{G}_{skew-circ}\textbf{D}_{2}\textbf{HD}_{1}$, $\textbf{HD}_{g_{1},...,g_{n}}\textbf{HD}_{2}\textbf{HD}_{1}$, and $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$, where $\textbf{G}_{Toeplitz}$, and $\textbf{G}_{skew-circ}$ are respectively Gaussian Toeplitz and Gaussian skew-circulant matrices.
We can see that all \textit{TripleSpin}-matrices show high collision probabilities for small distances and low ones for large distances. All the curves are almost identical. As theoretically predicted, there is no loss of accuracy (sensitivity) by using matrices from the \textit{TripleSpin}-family.
\begin{figure}[!ht]
\CenterFloatBoxes
\begin{floatrow}
\ffigbox
{\includegraphics[scale = 0.16]{./images/collision256-64.pdf}}
{\caption{Cross-polytope LSH - collision probabilities}\label{fig:collision}}
\killfloatstyle
\vspace{0pt}
\ttabbox
{\renewcommand{\arraystretch}{1.5}
\resizebox{1.0\linewidth}{!}{%
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
Matrix dimensions & $2^{9}$ & $2^{10}$ & $2^{11}$ & $2^{12}$ & $2^{13}$ & $2^{14}$ & $2^{15}$ \\ \hline
$\textbf{G}_{Toeplitz}\textbf{D}_{2}\textbf{HD}_{1}$ & x1.4 & x3.4 & x6.4 & x12.9 & x28.0 & x42.3 & x89.6 \\ \hline
$\textbf{G}_{skew-circ}\textbf{D}_{2}\textbf{HD}_{1}$ & x1.5 & x3.6 & x6.8 & x14.9 & x31.2 & x49.7 & x96.5 \\ \hline
$\textbf{HD}_{g_{1},...,g_{n}}\textbf{HD}_{2}\textbf{HD}_{1}$ & x2.3 & x6.0 & x13.8 & x31.5 & x75.7 & x137.0 & x308.8 \\ \hline
$\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ & x2.2 & x6.0 & x14.1 & x33.3 & x74.3 & x140.4 & x316.8 \\ \hline
\end{tabular}
}%
}
{\vspace{0.25cm}
\caption{Speedups for Gaussian kernel approximation}
\label{tab:speedupskernel}
\vspace{-0.25cm}}
\end{floatrow}
\end{figure}
\subsection{Kernel approximation}
In the second experiment, we compared feature maps obtained with Gaussian random matrices and specific \textit{TripleSpin}-matrices for Gaussian and angular kernels. To test the quality of the structured kernels' approximations, we compute the corresponding Gram-matrix reconstruction error using the Frobenius norm metric as in \cite{chor_sind_2016} : $ \frac{||\textbf{K} - \tilde{\textbf{K}} ||_{F}}{||\textbf{K}||_{F}} $, where $\textbf{K}, \tilde{\textbf{K}}$ are respectively the exact and approximate Gram-matrices, as a function of the number of random features. When number of random features $k$ is greater than data dimensionality $n$, we apply described block-mechanism.
We used the USPST dataset (test set) which consists of scans of handwritten digits from envelopes by the U.S. Postal Service. It contains 2007 points of dimensionality 258 ($n = 258$) corresponding to descriptors of 16 x 16 grayscale images. For Gaussian kernel, bandwidth $\sigma$ is set to $9.4338$. The results are averaged over $10$ runs.
\paragraph{Results on the USPST dataset:} The following matrices have been tested: Gaussian random matrix $\textbf{G}$, $\textbf{G}_{Toeplitz}\textbf{D}_{2}\textbf{HD}_{1}$ $\textbf{G}_{skew-circ}\textbf{D}_{2}\textbf{HD}_{1}$, $\textbf{HD}_{g_{1},...,g_{n}}\textbf{HD}_{2}\textbf{HD}_{1}$ and $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$.
In Figure \ref{kernel_approx_USPST}, for both kernels, all \textit{TripleSpin}-matrices perform similarly to a random Gaussian matrix, but $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ is giving the best results (see Figure \ref{kernel_approx} in the Appendix for additional experiments). The efficiency of the \textit{TripleSpin}-mechanism does not depend on the dataset.
\begin{figure}[!t]
\centering
\includegraphics[scale = 0.19]{./images/gaussiankernel_USPST.pdf}
\includegraphics[scale = 0.19]{./images/angularkernel_USPST.pdf}
\caption{Accuracy of random feature map kernel approximation}
\label{kernel_approx_USPST}
\end{figure}
Table \ref{tab:speedupskernel} shows significant speedups
obtained by the \textit{TripleSpin}-matrices. Those are defined as $\mbox{time}(\textbf{G})/\mbox{time}(\textbf{T})$, where $\textbf{G}$ is a random Gaussian matrix, $\textbf{T}$ is a \textit{TripleSpin} matrix
and $\mbox{time}(\textbf{X})$ stands for the corresponding runtime.
\subsection{Newton sketches}
Our last experiment covers the Newton sketch approach initially proposed in~\cite{pilanci} as
a generic optimization framework. In the subsequent experiment we show that {\em TripleSpin} matrices can be used for this purpose, thus can speed up several convex optimization problems solvers. The logistic regression problem is considered (see the Appendix for more details). Our goal is to find $x \in \mathbb{R}^d$, which minimizes the logistic regression cost, given a dataset $\{(a_i,y_i)\}_{i=1..n}$, with $a_i \in
\mathbb{R}^d$ sampled according to a Gaussian centered multivariate distribution with covariance $\Sigma_{i,j} = 0.99^{|i-j|}$ and $y_i \in \{-1,1\}$, generated at random. Various sketching matrices $S^t \in \mathbb{R}^{m \times n}$ are considered.
We first show that equivalent convergence properties are exhibited for various {\em TripleSpin}-matrices. Figure~\ref{fig:newtonSketchConvergence} illustrates the convergence of the Newton sketch algorithm, as measured by the optimality gap defined in~\cite{pilanci}, versus the iteration number. While it is clearly expected that sketched versions of the algorithm do not converge as quickly as the exact Newton-sketch approach, the figure confirms that various {\em TripleSpin}-matrices exhibit similar convergence behavior.
As shown in~\cite{pilanci}, when the dimensionality of the problem increases, the computational cost of computing the Hessian in the exact Newton-sketch approach becomes very large, scaling as $\mathcal{O}(nd^2)$. The complexity of the structured Newton-sketch approach with \textit{TripleSpin}-matrices is only $\mathcal{O}(d n \log(n) + md^2)$. Wall-clock times of computing single Hessian matrices are illustrated in Figure~\ref{fig:newtonSketchConvergence}. This figure confirms that the increase in number of iterations of the Newton sketch compared to the exact sketch is compensated by the efficiency of sketched computations, in particular Hadamard-based sketches yield improvements at the lowest dimensions.
\begin{figure}[htb]
\centering
\begin{subfigure}[b]{.5\linewidth}
\centering
\includegraphics[width=\columnwidth]{./images/optimalityVSiter}
\label{fig:newtonSketcha}
\end{subfigure}%
\begin{subfigure}[b]{.5\linewidth}
\centering
\includegraphics[width=\columnwidth]{./images/hessianCompTime}
\label{fig:newtonSketchb}
\end{subfigure}%
\caption{Numerical illustration of convergence (left) and computational complexity (right) of the Newton sketch algorithm with various {\em TripleSpin}-matrices. (left) Various sketching structures are compared in terms of convergence against iteration number. (right) Wall-clock times of {\em TripleSpin} structures are compared in various dimensionality settings.}
\label{fig:newtonSketchConvergence}
\end{figure}
\section{Introduction}
Consider a randomized machine learning algorithm $\mathcal{A}_{\textbf{G}}(\mathcal{X})$ on a dataset $\mathcal{X}$ and parameterized by the Gaussian matrix $\textbf{G}$ with i.i.d. entries taken from $\mathcal{N}(0,1)$.
We assume that $\textbf{G}$ is used to calculate Gaussian projections that are then
passed to other, possibly highly nonlinear functions.
Many machine learning algorithms are of this form. Examples include several variants of the Johnson-Lindenstrauss Transform applying random projections to reduce data dimensionality while approximately preserving Euclidean distance \cite{liberty_jlt, ailon2013}, quantization techniques using random projection trees, where splitting in each node is determined by a projection of data onto Gaussian direction \cite{dasgupta08}, algorithms solving convex optimization problems with random sketches of Hessian matrices \cite{pilanci, pilanci2}, kernel approximation techniques applying random feature maps produced from linear projections with Gaussian matrices followed by nonlinear mappings \cite{chor_sind_2016, rahimi, choromanska2015binary}, several LSH-schemes \cite{indyk2015,charikar, terasawa} (such as some of the most effective cross-polytope LSH methods) and many more.
If data is high-dimensional then computing random projections $\{\textbf{G}\textbf{x} : \textbf{x} \in \mathcal{X} \}$ for $\textbf{G} \in \mathbb{R}^{m \times n}$ in time $\Theta(mn|\mathcal{X}|)$ often occupies a significant fraction of the overall
run time. Furthermore, storing matrix $\textbf{G}$ frequently becomes a bottleneck in terms of space complexity. In this paper
we propose a ``structured variant'' of the algorithm $\mathcal{A}$, where Gaussian matrix $\textbf{G}$ is replaced by a structured matrix
$\textbf{G}_{struct}$ taken from the defined by us \textit{TripleSpin}-family of matrices. The name comes from the fact that
each such matrix is a product of three other matrices, building components, which include rotations.
Replacing $\textbf{G}$ by $\textbf{G}_{struct}$ gives computational speedups since $\textbf{G}_{struct}\textbf{x}$ can be calculated in $o(mn)$ time: with the use of techniques such as Fast Fourier Transform, time complexity is reduced to $O(n\log m)$.
Furthermore, with matrices from the \textit{TripleSpin}-family space complexity can be also substantially reduced --- to sub-quadratic, usually at most linear, sometimes even constant.
To the best of our knowledge, we are the first to provide a comprehensive theoretical explanation of the effectiveness of the structured approach. So far such an
explanation was given only for some specific applications and specific structured matrices. The proposed \textit{TripleSpin}-family contains all previously considered structured matrices as special cases, including the recently introduced $P$-model \cite{chor_sind_2016}, yet its flexible three-block structure provides mechanisms for constructing many others.
Our structured model is also the first one that proposes purely discrete Hadamard-based constructions with strong theoretical guarantees. We empirically show that these surpass their unstructured counterparts.
\begin{comment}
We show theoretically that $\mathcal{A}_{struct}$ obtained by replacing unstructured $\textbf{G}$ with a matrix $\textbf{M}$ from the \textit{TripleSpin}-family accurately approximates $\mathcal{A}$, i.e. the loss of accuracy coming from imposing a particular structure on the linear projection is minimal. In fact our experimental findings show that some structured constructions surpass their unstructured counterparts.
\end{comment}
As a byproduct, we provide a theoretical explanation of the efficiency of the cross-polytope LSH method based on the $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ structured matrix \cite{indyk2015}, where $\textbf{H}$ stands for the Hadamard matrix and $\textbf{D}_{i}$s are random diagonal $\pm 1$-matrices. Thus we solve the open problem posted in \cite{indyk2015}. These guarantees as well as theoretical results for other aforementioned applications arise from the same general theoretical principle that we present in the paper. Our theoretical methods apply relatively new Berry-Esseen type Central Limit Theorem results
for random vectors.
\section{Related work}
\label{sec:related_work}
Structured matrices were previously used mainly in the context of the Johnson-Lindenstrauss Transform (JLT), where the goal is to linearly transform
high-dimensional data and embed it into a much lower dimensional space in such a way that the Euclidean distance is approximately preserved \cite{liberty_jlt, ailon2013, ailon2006approximate}.
Most of these structured constructions involve sparse or circulant matrices \cite{ailon2006approximate,
vybiral2011variant} providing computational speedups and space compression.
Specific structured matrices were used to approximate angular distance and Gaussian kernels \cite{choromanska2015binary, feng}.
Very recently \cite{chor_sind_2016} structured matrices coming from the so-called \textit{P-model}, were applied to speed up random feature map computations of some special kernels (angular, arc-cosine and Gaussian). The presented techniques did not work for discrete structured constructions such as the $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ structured matrix since they focus on matrices with low (polylog) chromatic number of the corresponding coherence graphs and these do not include matrices such as $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ or their
direct non-discrete modifications.
The $TripleSpin$-mechanism gives in particular a highly parameterized family of structured methods for approximating general kernels with random feature maps.
Among them are purely discrete computational schemes providing the most aggressive
compression with just minimal loss of accuracy.
Several LSH methods use random Gaussian matrices to construct compact codes of data points and in turn speed up such tasks as approximate nearest neighbor search. Among them are
cross-polytope methods proposed in \cite{terasawa}. In the angular cross-polytope setup the hash family $\mathcal{H}$ is designed for points taken from the unit sphere $S^{n-1}$, where $n$ stands for data dimensionality. To construct hashes a random matrix $\textbf{G} \in \mathbb{R}^{n \times n}$ with i.i.d. Gaussian entries
is built. The hash of a given data point $\textbf{x} \in S^{n-1}$ is defined as:
$
h(\textbf{x}) = \eta(\frac{\textbf{G}\textbf{x}}{\|\textbf{G}\textbf{x}\|_{2}}),
$
where $\eta(\textbf{y})$ returns the closest vector to $\textbf{y}$ from the set $\{\pm 1 \textbf{e}_{i}\}_{1 \leq i \leq n}$, where $\{\textbf{e}_{i}\}_{1 \leq i \leq n}$ stands for the canonical basis.
The fastest known variant of the cross-polytope LSH \cite{indyk2015} replaces unstructured Gaussian matrix $\textbf{G}$
with the product $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$. No theoretical guarantees regarding that variant were known. We provide a theoretical explanation here. As in the previous setting, matrices from the $TripleSpin$ model lead to several fast structured cross-polytope LSH algorithms not considered before.
Recently a new method for speeding up algorithms solving convex optimization problems by approximating Hessian matrices (a bottleneck of the overall computational pipeline) with their random sketches was proposed \cite{pilanci, pilanci2}.
One of the presented sketches, the so-called \textit{Newton Sketch}, leads to the sequence of iterates $\{\textbf{x}^{t}\}_{t=0}^{\infty}$ for optimizing given function $f$ given by the following recursion:
\begin{equation}
\textbf{x}^{t+1} = argmin_{\textbf{x}}
\{\frac{1}{2}\|\textbf{S}^{t}\nabla^{2}f(\textbf{x}^{t})^{\frac{1}{2}}(\textbf{x}-\textbf{x}^{t})\|_{2}^{2} + (\nabla f(\textbf{x}^{t}))^{T} \cdot (\textbf{x}-\textbf{x}^{t})\},
\end{equation}
where $\{\textbf{S}^{t} : t=1,...,\}$ is the set of the so-called \textit{sketch matrices}. Initially the sub-Gaussian sketches based on i.i.d. sub-Gaussian random variables
were used. The disadvantage of the sub-Gaussian sketches $\textbf{S} \in \mathbb{R}^{m \times n}$ lies in the fact that computing the sketch of the given matrix
$\textbf{A} \in \mathbb{R}^{n \times d}$, needed for convex optimization with sketches, requires $O(mnd)$ time. Thus the method is too slow in practice.
This is the place, where structured matrices can be applied. Some structured approaches were already considered in \cite{pilanci}, where sketches based on randomized orthonormal
systems were proposed. In that approach a sketching matrix $\textbf{S} \in \mathbb{R}^{m \times n}$ is constructed by sampling i.i.d. rows of the form
$\textbf{s}^{T} = \sqrt{n} \textbf{e}^{T}_{j} \textbf{HD}$ with probability $\frac{1}{n}$ for $j=1,...,n$, where $\textbf{e}^{T}_{j}$ is chosen uniformly at random
from the set of the canonical vectors and $\textbf{D}$ is a random diagonal $\pm 1$-matrix.
We show that our class of $\textit{TripleSpin}$ matrices can also be used in that setting.
\begin{comment}
Our theoretical results show that convex optimization with structured matrices is an accurate approximation of the unstructured one. To the best of our knowledge, such guarantees
were not known before. Replacing a particular structured approach with a generic family enables us to propose even more compact
structured models than these already considered and tune the structure of the applied matrix to speed and accuracy requirements.
\end{comment}
\section{Theoretical results}
We now show that matrices from the \textit{TripleSpin}-family can replace their unstructured counterparts in many machine learning algorithms with minimal loss of accuracy.
Let $\mathcal{A}$ be a randomized machine learning algorithm operating on a dataset $\mathcal{X} \subseteq \mathbb{R}^{n}$. We assume that $\mathcal{A}$ uses certain functions $f_{1},...,f_{s}$ such that each
$f_{i}$ uses a given Gaussian matrix $\textbf{G}$ (matrix $\textbf{G}$ can be the same or different for different functions) with independent rows taken from a multivariate Gaussian distribution with diagonal covariance matrix. $\mathcal{A}$ may also use other functions that do not depend directly on Gaussian matrices but may for instance operate on outputs of $f_{1},...,f_{s}$. We assume that each $f_{i}$ applies $\textbf{G}$ to vectors from some linear space $\mathcal{L}_{f_{i}}$ of dimensionality at most $d$.
\begin{remark}
In the kernel approximation setting with random feature maps one can match each pair of vectors $\textbf{x},\textbf{y} \in \mathcal{X}$ with a different $f=f_{\textbf{x},\textbf{y}}$.
Each $f$ computes the approximate value of the kernel for vectors $\textbf{x}$ and $\textbf{y}$.
Thus in that scenario $s = {|\mathcal{X}| \choose 2}$ and $d=2$ (since one can take: $\mathcal{L}_{f(\textbf{x},\textbf{y})} = span(\textbf{x},\textbf{y})$).
\end{remark}
\begin{remark}
In the vector quantization algorithms using random projection trees one can take $s=\nobreak 1$ (the algorithm $\mathcal{A}$ itself is a function $f$) and $d = d_{intrinsic}$, where $d_{intrinsic}$ is an intrinsic dimensionality of a given dataset $\mathcal{X}$ (random projection trees are often used if $d_{intrinsic} \ll n$).
\end{remark}
Fix a function $f$ of $\mathcal{A}$ using an unstructured Gaussian matrix $\textbf{G}$ and applying it on the linear space $\mathcal{L}_{f}$ of dimensionality $l \leq d$.
Note that the outputs of $\textbf{G}$
on vectors from $\mathcal{L}_{f}$ are determined by the sequence:
$(\textbf{G}\textbf{x}^{1},...,\textbf{G}\textbf{x}^{l})$,
where $\textbf{x}^{1},...,\textbf{x}^{l}$ stands for a fixed orthonormal basis of $\mathcal{L}_{f}$. Thus they are determined by a following vector (obtained from "stacking together" vectors $\textbf{Gx}^{1}$,...,$\textbf{Gx}^{l}$):
$$\textbf{q}_{f} = ((\textbf{G}\textbf{x}^{1})^{T},...,(\textbf{G}\textbf{x}^{l})^{T})^{T}.$$
Notice that $\textbf{q}_{f}$ is a Gaussian vector with independent entries (this comes from the fact that rows of $\textbf{G}$ are independent and the observation that the projections of the Gaussian vector on the orthogonal directions are independent). Thus the covariance matrix of $\textbf{q}_{f}$ is an identity.
\begin{definition}[$\epsilon$-similarity]
We say that a multivariate Gaussian distribution $\textbf{q}_{f}(\epsilon)$ is $\epsilon$-close to the multivariate Gaussian distribution $\textbf{q}_{f}$ with covariance matrix $\textbf{I}$ if its covariance matrix is equal to $1$ on the diagonal and has all other entries of absolute value at most $\epsilon$.
\end{definition}
To measure the "closeness" of the algorithm $\mathcal{A}^{\prime}$ with the \textit{TripleSpin}-model matrices to $\mathcal{A}$, we will measure how "close" the corresponding vector $\textbf{q}^{\prime}_{f}$ for $\mathcal{A}^{\prime}$ is to $\textbf{q}_{f}$ in the distribution sense.
The following definition proposes a quantitative way to measure this
closeness.
Without loss of generality we will assume now that each structured matrix consists of just one block since different blocks of the structured matrix are chosen independently.
\begin{definition}
\label{closeness}
Let $\mathcal{A}$ be as above. For a given $\eta, \epsilon > 0$ the class of algorithms $\mathcal{A}_{\eta, \epsilon}$ is given as a set of algorithms $\mathcal{A}^{\prime}$ obtained from $\mathcal{A}$ by replacing unstructured Gaussian matrices with their structured counterparts such that for any $f$ with $dim(\mathcal{L})_{f} = l$ and any convex set $S \in \mathbb{R}^{ml}$ the following holds:
\begin{equation}
|\mathbb{P}[\textbf{q}_{f}(\epsilon) \in S] - \mathbb{P}[\textbf{q}^{\prime}_{f} \in S]| \leq \eta,
\end{equation}
where $\textbf{q}_{f}(\epsilon)$ is some multivariate Gaussian distribution that is $\epsilon$-similar to $\textbf{q}_{f}$.
\end{definition}
The smaller $\epsilon$, the closer in distribution are $\textbf{q}_{f}$ and $\textbf{q}_{f}^{\prime}$ and the more accurate the structured version $\mathcal{A}^{\prime}$ of $\mathcal{A}$ is.
Now we show that \textit{TripleSpin}-matrices lead to algorithms from $\mathcal{A}_{\eta, \epsilon}$ with $\eta, \epsilon \ll 1$.
\begin{theorem}[structured ml algorithms]
\label{main_struct_theorem}
Let $\mathcal{A}$ be the randomized algorithm using unstructured Gaussian matrices $\textbf{G}$ and let $s,d$ be as at the beginning of the section.
Replace the unstructured matrix $\textbf{G}$ by one of
its structured variants from the \textit{TripleSpin}-family defined in Section \ref{sec:model} with blocks of $m$ rows each.
Then for $n$ large enough and $\epsilon = o_{md}(1)$ with probability at least:
\begin{equation}
1 - 2p(n)sd - 2{md \choose 2}se^{-\Omega(\min(\frac{\epsilon^{2}n^{2}}{K^{4}\Lambda_{F}^{2}\delta^{4}(n)}, \frac{\epsilon n}{K^{2}\Lambda_{2} \delta^{2}(n)}))}
\end{equation}
the structured version of the algorithm belongs to the class $\mathcal{A}_{\eta, \epsilon}$, where: $\eta = \frac{\delta^{3}(n)}{n^{\frac{2}{5}}}$, $\delta(n), p(n),K, \Lambda_{F}, \Lambda_{2}$ are as in the definition of the \textit{TripleSpin}-family from Section \ref{sec:model}
and the probability is in respect to the random choices of $\textbf{M}_{1}$
and $\textbf{M}_{2}$.
\end{theorem}
Theorem \ref{main_struct_theorem} implies strong accuracy guarantees for the specific matrices from the $TripleSpin$-family. As a corollary we get for instance:
\begin{theorem}
\label{corollary_theorem}
Under assumptions from Theorem~\ref{main_struct_theorem}
the probability that the structured version of the algorithm belongs to $\mathcal{A}_{\eta, \epsilon}$ for $\eta = \frac{\delta^{3}(n)}{n^{\frac{2}{5}}}$ is at least:
$1 - 4ne^{-\frac{\log^{2}(n)}{8}}sd - 2{md \choose 2}s e^{-\Omega(\frac{\epsilon^{2}n}{\log^{4}(n)})}$
for the structured matrices $\sqrt{n}\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$, $\sqrt{n}\textbf{HD}_{g_{1},...,g_{n}}\textbf{HD}_{2}\textbf{HD}_{1}$ as well as for the structured matrices of the form $\textbf{G}_{struct}\textbf{D}_{2}\textbf{HD}_{1}$, where $\textbf{G}_{struct}$ is Gaussian circulant, Gaussian Toeplitz or Gaussian Hankel matrix.
\end{theorem}
As a corollary of Theorem~\ref{corollary_theorem}, we obtain the following result showing the effectiveness of the cross-polytope LSH with structured matrices $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ that was only heuristically confirmed before~\cite{indyk2015}.
\begin{theorem}
\label{hopefully_last_theorem}
Let $\textbf{x},\textbf{y} \in \mathbb{R}^{n}$ be two unit $L_{2}$-norm vectors.
Let $\textbf{v}_{\textbf{x}, \textbf{y}}$ be the vector indexed by all $m^{2}$ ordered pairs of canonical directions $(\pm \textbf{e}_{i},\pm \textbf{e}_{j})$, where the value of the entry indexed by $(\textbf{u},\textbf{w})$ is the probability that: $h(\textbf{x}) = \textbf{u}$ and $h(\textbf{y}) = \textbf{w}$, and $h(\textbf{v})$ stands for the hash of $\textbf{v}$. Then with probability at least:
$p_{success} = 1 - 4ne^{-\frac{\log^{2}(n)}{8}}sd - 2{md \choose 2}s e^{-\Omega(\frac{\epsilon^{2}n}{\log^{4}(n)})}$
the version of the stochastic vector $\textbf{v}^{1}_{\textbf{x},\textbf{y}}$
for the unstructured Gaussian matrix $\textbf{G}$ and its structured counterpart $\textbf{v}^{2}_{\textbf{x},\textbf{y}}$ for the matrix $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ satisfy:
$\|\textbf{v}^{1}_{\textbf{x},\textbf{y}} - \textbf{v}^{2}_{\textbf{x},\textbf{y}}\|_{\infty} \leq \log^{3}(n)/n^{\frac{2}{5}} + c \epsilon$,
for $n$ large enough, where $c > 0$ is a universal constant.
The probability above is taken with respect to random choices of $\textbf{D}_{1}$ and $\textbf{D}_{2}$.
\end{theorem}
The proof for the discrete structured setting applies Berry-Esseen-type results for random vectors (details in the Appendix) showing that for $n$ large enough $\pm 1$ random vectors $\textbf{r}$ act similarly to Gaussian vectors.
\begin{comment}
\subsection{"Almost independent" dimensions}
Let us briefly describe how we prove the above results (all details are given in the Appendix).
Theoretical challenges regarding proving accuracy results for structured matrices come from the fact that for any given $\textbf{x} \in \mathbb{R}^{n}$
different dimensions of $\textbf{y} = \textbf{G}_{struct}\textbf{x}$ are no longer independent (as it is the case for the unstructured setting).
For matrices from the \textit{TripleSpin}-family we can however show that with high probability different elements of $\textbf{y}$ correspond to projections of a given vector $\textbf{r}$ (see Section \ref{sec:model}) into directions that are close to orthogonal. The "close-to-orthogonality" characteristic is obtained with the use of the Hanson-Wright inequality that focuses on concentration results regarding quadratic forms involving vectors of sub-Gaussian random variables. If $\textbf{r}$ is Gaussian then from the well known fact that projections of the Gaussian vector into orthogonal directions are independent we can conclude that dimensions of $\textbf{y}$ are "close to independent". If $\textbf{r}$ is a discrete vector then we need to show that for $n$ large enough it "resembles" the Gaussian vector. This is where we need to apply the aforementioned techniques regarding multivariate Berry-Esseen-type central limit theorem results.
\end{comment}
\section{The \textit{TripleSpin}-family}
\label{sec:model}
We now present the \textit{TripleSpin}-family.
If not specified otherwise, the random diagonal matrix $\textbf{D}$
is a diagonal matrix with diagonal entries taken independently at random from $\{-1,+1\}$.
For a sequence $(t_{1},...,t_{n})$ we denote by $\textbf{D}_{t_{1},...,t_{n}}$ a diagonal matrix with diagonal equal to $(t_{1},...,t_{n})$.
For a matrix $\textbf{A} \in \mathbb{R}^{n \times n}$ let $\|\textbf{A}\|_{F}$ denote its Frobenius and $\|A\|_{2} = \sup_{\textbf{x} \neq 0} \frac{\|\textbf{Ax}\|_{2}}{\|\textbf{x}\|_{2}}$ its spectral norm respectively. We denote by $\textbf{H}$ the $L_{2}$-normalized Hadamard matrix. We say that $\textbf{r}$ is a random Rademacher vector if every element of $\textbf{r}$ is chosen independently at random from $\{-1,+1\}$.
For a vector $\textbf{r} \in \mathbb{R}^{k}$ and $n>0$ let $\textbf{C}(\textbf{r},n) \in \mathbb{R}^{n \times nk}$ be a matrix,
where the first row is of the form $(\textbf{r}^{T},0,...,0)$ and each subsequent row is obtained from the previous one by right-shifting in a circulant manner the previous one by $k$. For a sequence of matrices $\textbf{W}^{1},...,\textbf{W}^{n} \in \mathbb{R}^{k \times n}$ we denote by $\textbf{V}(\textbf{W}^{1},...,\textbf{W}^{n}) \in \mathbb{R}^{nk \times n}$ a matrix obtained by stacking vertically matrices: $\textbf{W}^{1},...,\textbf{W}^{n}$.
Each structured matrix $\textbf{G}_{struct} \in \mathbb{R}^{n \times n}$ from the \textit{TripleSpin}-family is a product of three main structured components, i.e.:
\begin{equation}
\textbf{G}_{struct} = \textbf{M}_{3}\textbf{M}_{2}\textbf{M}_{1},
\end{equation}
where matrices $\textbf{M}_{1}, \textbf{M}_{2}$ and $\textbf{M}_{3}$ satisfy the following conditions:
\begin{framed}
\textbf{Condition 1:} Matrices: $\textbf{M}_{1}$ and $\textbf{M}_{2}\textbf{M}_{1}$ are $(\delta(n),p(n))$-balanced isometries. \\
\textbf{Condition 2:} $\textbf{M}_{2} = \textbf{V}(\textbf{W}^{1},...,\textbf{W}^{n})\textbf{D}_{\rho_{1},...,\rho_{n}}$ for some $(\Delta_{F},\Delta_{2})$-smooth set ${\textbf{W}^{1},...,\textbf{W}^{n}} \in \mathbb{R}^{k \times n}$ and some i.i.d sub-Gaussian random variables $\rho_{1},...,\rho_{n}$ with sub-Gaussian norm $K$. \\
\textbf{Condition 3:} $\textbf{M}_{3} = \textbf{C}(\textbf{r},n)$
for $\textbf{r} \in \mathbb{R}^{k}$, where $\textbf{r}$ is random Rademacher or Gaussian.
\end{framed}
If all three conditions are satisfied then we say that a matrix $\textbf{M}_{3}\textbf{M}_{2}\textbf{M}_{1}$ is a \textit{TripleSpin}-matrix with parameters: $\delta(n),p(n),K,\Lambda_{F},\Lambda_{2}$.
Below we explain these conditions.
\begin{definition}[$(\delta(n),p(n))$-balanced matrices]
A randomized matrix $\textbf{M} \in \mathbb{R}^{n \times m}$ is $(\delta(n),p(n))$-balanced if for every $\textbf{x} \in \mathbb{R}^{m}$ with $\|\textbf{x}\|_{2} = 1$ we have: $\mathbb{P}[\|\textbf{M}\textbf{x}\|_{\infty} > \frac{\delta(n)}{\sqrt{n}}] \leq p(n)$.
\end{definition}
\begin{remark}
\label{balanceness_remark}
One can take as $\textbf{M}_{1}$ a matrix $\textbf{HD}_{1}$,
since as we will show in the Appendix, matrix $\textbf{HD}_{1}$ is $(\log(n),2ne^{-\frac{\log^{2}(n)}{8}})$-balanced.
\end{remark}
\begin{definition}[$(\Delta_{F},\Delta_{2})$-smooth sets]
A deterministic set of matrices $\textbf{W}^{1},...,\textbf{W}^{n} \in \mathbb{R}^{k \times n}$ is $(\Lambda_{F},\Lambda_{2})$-smooth if:
\begin{itemize}
\item $\|\textbf{W}^{i}_{1}\|_{2} = .. = \|\textbf{W}^{i}_{n}\|_{2}$
for $i = 1,...,n$, where $\textbf{W}^{i}_{j}$ stands for the
$j^{th}$ column of $\textbf{W}^{i}$,
\item for $i \neq j$ and $l = 1,...,n$ we have:
$(\textbf{W}^{i}_{l})^{T} \cdot \textbf{W}^{j}_{l} = 0$,
\item $\max_{i,j} \|(\textbf{W}^{j})^{T}\textbf{W}^{i}\|_{F} \leq \Lambda_{F}$ and $\max_{i,j} \|(\textbf{W}^{j})^{T}\textbf{W}^{i}\|_{2} \leq \Lambda_{2}.$
\end{itemize}
\end{definition}
\begin{remark}
If the unstructured matrix $\textbf{G}$ has rows taken from the general multivariate Gaussian distribution with diagonal covariance matrix $\Sigma \neq \textbf{I}$ then one needs to rescale vectors $\textbf{r}$ accordingly.
For clarity we assume here that $\Sigma = \textbf{I}$ and we present our theoretical results for that setting.
\end{remark}
All structured matrices previously considered are special cases of the $TripleSpin$-family (for clarity we will explicitly show it for some important special cases). Others, not considered before, are also covered by the $TripleSpin$-family. We have:
\begin{lemma}
\label{simple_lemma}
Matrices $\textbf{G}_{circ}\textbf{D}_{2}\textbf{HD}_{1}$, $\sqrt{n}\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$
and $\sqrt{n}\textbf{HD}_{g_{1},...,g_{n}}\textbf{HD}_{2}\textbf{HD}_{1}$, where $\textbf{G}_{circ}$ is Gaussian circulant, are valid $TripleSpin$-matrices for $\delta(n) = \log(n)$, $p(n) = 2ne^{-\frac{\log^{2}(n)}{8}}$, $K = 1$, $\Lambda_{F} = O(\sqrt{n})$ and $\Lambda_{2} = O(1)$. The same is true if one replaces $\textbf{G}_{circ}$ by a Gaussian Hankel or Toeplitz matrix.
\end{lemma}
\subsection{Stacking together \textit{TripleSpin}-matrices}
We described \textit{TripleSpin}-matrices as square matrices. In practice we are not restricted to square matrices. We can construct an $m \times n$ \textit{TripleSpin} matrix for $m \leq n$ from the square $n \times n$ \textit{TripleSpin}-matrix by taking its first $m$ rows. We can then stack vertically these independently constructed $m \times n$ matrices to obtain an $k \times n$ matrix for both: $k \leq n$ and $k > n$. We think about $m$ as another parameter of the model that tunes the "structuredness" level. Larger values of $m$ indicate more structured approach while smaller values lead to more random matrices (the $m=1$ case is the fully unstructured one).
\begin{comment}
Each structured matrix $\textbf{G}_{struct} \in \mathbb{R}^{n \times n}$ from the \textit{TripleSpin}-family is a product of three main structured blocks, i.e.:
\begin{equation}
\textbf{G}_{struct} = \textbf{B}_{3}\textbf{B}_{2}\textbf{B}_{1},
\end{equation}
where matrices $\textbf{B}_{1},\textbf{B}_{2},\textbf{B}_{3}$ satisfy two conditions that we give below.
\begin{framed}
\textbf{Condition 1:} Matrices: $\textbf{B}_{1}$ and $\textbf{B}_{2}\textbf{B}_{1}$ are $(\delta(n),p(n))$-balanced isometries. \\
\textbf{Condition 2:} Pair of matrices $(B_{2},B_{3})$
is $(K,\Lambda_{F}, \Lambda_{2})$-random.
\end{framed}
If both conditions are satisfied then we say that a matrix $\textbf{B}_{3}\textbf{B}_{2}\textbf{B}_{1}$ is a \textit{TripleSpin}-matrix with parameters: $\delta(n),p(n),K,\Lambda_{F},\Lambda_{2}$.
Below we explain these conditions.
\begin{definition}[$(\delta(n),p(n))$-balanced matrices]
A randomized matrix $\textbf{M} \in \mathbb{R}^{n \times n}$ is $(\delta(n),p(n))$-balanced if it represents an isometry and for every $\textbf{x} \in \mathbb{R}^{n}$ with $\|\textbf{x}\|_{2} = 1$ we have: $\mathbb{P}[\|\textbf{M}\textbf{x}\|_{\infty} > \frac{\delta(n)}{\sqrt{n}}] \leq p(n)$.
\end{definition}
\begin{remark}
\label{balanceness_remark}
One can take as $\textbf{B}_{1}$ a matrix $\textbf{HD}_{1}$,
since as we will show in the Appendix, matrix $\textbf{HD}_{1}$ is $(\log(n),2ne^{-\frac{\log^{2}(n)}{8}})$-balanced.
Under such choice of $\textbf{B}_{1}$ one can take: $\textbf{B}_{2} = \textbf{HD}_{2}$ or $\textbf{B}_{2} = \textbf{D}$.
\end{remark}
\begin{definition}[$(K, \Lambda_{F}, \Lambda_{2})$-randomness]
A pair of matrices $(\textbf{Y},\textbf{Z}) \in \mathbb{R}^{n \times n} \times \mathbb{R}^{n \times n}$ is $(K, \Lambda_{F}, \Lambda_{2})$-random
if there exist: $\textbf{r} \in \mathbb{R}^{k}$, where $k \geq n$, and
a set of linear isometries $\phi = \{\phi_{1},...,\phi_{n}\}$,
where $\phi_i : \mathbb{R}^{n} \rightarrow \mathbb{R}^{k}$, such that:
\begin{itemize}
\item $\textbf{r}$ is either a $\pm 1$-vector with i.i.d. entries
or Gaussian with identity covariance matrix,
\item for every $\textbf{x} \in \mathbb{R}^{n}$ the $j^{th}$ element $(\textbf{Zx})_{j}$ of $\textbf{Zx}$ is of the form: $\textbf{r}^{T} \cdot \phi_{j}(\textbf{x})$,
\item there exists a set of i.i.d. sub-Gaussian random variables $\{\rho_{1},...,\rho_{n}\}$ with sub-Gaussian norm at most $K$, mean $0$, the same second moments and a $(\Lambda_{F},\Lambda_{2})$-smooth set of matrices $\{\textbf{W}^{i}\}_{i=1,...,n}$ such that for every $\textbf{x} = (x_{1},...,x_{n})^{T}$ we have: $\phi_{i}(\textbf{Y}\textbf{x}) = \textbf{W}^{i} (\rho_{1}x_{1},...,\rho_{n}x_{n})^{T}$.
\end{itemize}
\end{definition}
\begin{definition}[$(\Delta_{F},\Delta_{2})$-smooth sets]
A deterministic set of matrices $\textbf{W} = \{\textbf{W}^{1},...,\textbf{W}^{n}\}$, where $\textbf{W}^{i} = \{w^{i}_{k,l}\}_{k,l \in \{1,...,n\}}$ is $(\Lambda_{F},\Lambda_{2})$-smooth if:
\begin{itemize}
\item $\|\textbf{W}^{i}_{1}\|_{2} = .. = \|\textbf{W}^{i}_{n}\|_{2}$
for $i = 1,...,n$, where $\textbf{W}^{i}_{j}$ stands for the
$j^{th}$ column of $\textbf{W}^{i}$,
\item For $i \neq j$ and $l = 1,...,n$ we have:
$(\textbf{W}^{i}_{l})^{T} \cdot \textbf{W}^{j}_{l} = 0$,
\item For a sequence of matrices $\{\textbf{A}^{i,j}\}$, $i,j \in \{1,...,n\}$ defined as:
$A^{i,j}_{l,t} = (\textbf{W}^{j}_{l})^{T} \cdot \textbf{W}^{i}_{t}$ for $1 \leq i < j \leq n$ or $1 \leq l < t \leq n$ and $A^{i,i}_{l,l} = 0$
we have: $\max_{i,j} \|\textbf{A}^{i,j}\|_{F} \leq \Lambda_{F}$ and $\max_{i,j} \|\textbf{A}^{i,j}\|_{2} \leq \Lambda_{2}.$
\end{itemize}
\end{definition}
\begin{remark}
If the unstructured matrix $\textbf{G}$ has rows taken from the general multivariate Gaussian distribution with diagonal covariance matrix $\Sigma \neq \textbf{I}$ then one needs to rescale vectors $\textbf{r}$ accordingly.
For clarity we assume here that $\Sigma = \textbf{I}$ and we present our theoretical results for that setting.
\end{remark}
All structured matrices previously considered are special cases of the $TripleSpin$-family (for clarity we will explicitly show it for some important special cases). Others, not considered before, are also covered by the $TripleSpin$-family. We have:
\begin{lemma}
\label{simple_lemma}
Matrices $\textbf{G}_{circ}\textbf{D}_{2}\textbf{HD}_{1}$, $\sqrt{n}\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$
and $\sqrt{n}\textbf{HD}_{g_{1},...,g_{n}}\textbf{HD}_{2}\textbf{HD}_{1}$, where $\textbf{G}_{circ}$ is Gaussian circulant, are valid $TripleSpin$-matrices for $\delta(n) = \log(n)$, $p(n) = 2ne^{-\frac{\log^{2}(n)}{8}}$, $K = 1$, $\Lambda_{F} = O(\sqrt{n})$ and $\Lambda_{2} = O(1)$. The same is true if one replaces $\textbf{G}_{circ}$ by a Gaussian Hankel or Toeplitz matrix.
\end{lemma}
\subsection{Stacking together \textit{TripleSpin}-matrices}
We described \textit{TripleSpin}-matrices as square matrices. In practice we are not restricted to square matrices. We can construct an $m \times n$ \textit{TripleSpin} matrix for $m \leq n$ from the square $n \times n$ \textit{TripleSpin}-matrix by taking its first $m$ rows. We can then stack vertically these independently constructed $m \times n$ matrices to obtain an $k \times n$ matrix for both: $k \leq n$ and $k > n$. We think about $m$ as another parameter of the model that tunes the "structuredness" level. Larger values of $m$ indicate more structured approach while smaller values lead to more random matrices (the $m=1$ case is the fully unstructured one).
\end{comment}
|
\section{Introduction}
\vspace{5mm}
AdS/CFT correspondence \cite{Maldacena:1997re} provides us with the possibility to investigate a quantum theory of gravity in terms of an ordinary quantum field theory.
In order to study some geometrical properties of the dual gravity theory, one is typically required to perform a quantitative analysis on strong coupling behaviors of the corresponding quantum field theory.
Recently, such an analysis has become manageable, at least for a set of physical quantities, thanks to the developments of the supersymmetric localization.
See \cite{Pestun:2014mja}\cite{Hosomichi:2014hja} for recent reviews on this topic.
ABJM theory \cite{Aharony:2008ug} provides a prototypical example of AdS$_4$/CFT$_3$ correspondence.
This is a Chern-Simons theory coupled to matters with ${\cal N}=6$ superconformal symmetry.
The analysis of this theory has been done intensively in the context of AdS/CFT correspondence as well as in relation to M2-branes.
The researches discussing the strong coupling behaviors of ABJM theory include the ones using the planar limit \cite{Suyama:2009pd,Marino:2009jd,Drukker:2010nc,Drukker:2011zy}, the M-theory limit \cite{Herzog:2010hf} and the Fermi gas formalism \cite{Marino:2011eh}, all of which are based on the localization formula for the partition function obtained in \cite{Kapustin:2009kz}.
Among them, the Fermi gas formalism has turned out to be quite powerful.
It allows us to obtain, for example, the free energy, not only to all orders in $1/N$ expansion \cite{Fuji:2011km}, but even including non-perturbative terms \cite{Hatsuda:2012dt,Hatsuda:2013gj,Hatsuda:2013oxa}.
Recently, such an analysis has been extended to Chern-Simons-matter theories whose dual theories contain orientifolds \cite{Moriyama:2015asx,Okuyama:2015auc,Honda:2015rbb,Okuyama:2016xke,Moriyama:2016xin,Moriyama:2016kqi}.
In this paper, on the other hand, we revisit the analysis of planar solutions based on the matrix model technique.
Although the reach of this technique is practically confined in the leading order of $1/N$ expansion unless the matrix model under consideration is simple enough, it can be applied to much wider family of Chern-Simons-matter theories, compared to the other methods.
The aim of our research is to investigate a pattern in the strong coupling behaviors of various Chern-Simons-matter theories so that one could find a clue to know which theory could have a possible gravity dual.
We expect that this line of research would shed some light on the underlying principle of how the space-time of the bulk gravity theory emerges from a quantum field theory.
We focus our attention on a family of ${\cal N}\ge3$ Chern-Simons-matter theories with the gauge group ${\rm U}(N_1)\times {\rm U}(N_2)$ coupled to an arbitrary number $n$ of bi-fundamental hypermultiplets.
The planar resolvents for such theories have been investigated in \cite{Suyama:2009pd}\cite{Marino:2009jd}\cite{Suyama:2010hr,Suyama:2011yz,Suyama:2013fua}, however, explicit expressions for the resolvents have not been obtained so far except for ABJM theory and ABJ theory \cite{Aharony:2008gk}.
In this paper we show that, instead of the planar resolvent itself, its derivative can be determined explicitly for all theories mentioned above.
More precisely, we define for each theory two planar resolvents which contain the information on two sets of eigenvalues in the matrix model.
We determine the derivatives of both two resolvents explicitly except for the case $n=2$.
For this exceptional case, which turns out to be the most interesting in the context of AdS/CFT correspondence, the derivative of a linear combination of the two resolvents is determined.
From these results, we derive the explicit expressions of (a linear combination of) the vevs of BPS Wilson loops \cite{Drukker:2008zx,Chen:2008bp,Rey:2008bh} of the theories.
Since the vevs can be written in terms of well-known functions, it is now straightforward to examine in which limit the vevs of the Wilson loops may diverge, the result of which provides an important hint for when a weak gravity dual might exist.
This paper is organized as follows.
In section \ref{pureCS}, we revisit the analysis of pure Chern-Simons theory in order to motivate us to consider the derivative of the planar resolvent.
In section \ref{2-node}, we investigate the Chern-Simons-matter theories specified above, and determine explicitly the derivatives of the planar resolvents and the vevs of the Wilson loops.
The analysis is done for the case $n=2$ (subsection \ref{n=2}) and for the other cases (subsection \ref{general-n}) separately.
The validity of our formula for the planar resolvents is checked in section \ref{check} by calculating the vevs of the BPS Wilson loops perturbatively.
Section \ref{discuss} is devoted to discussion.
\vspace{1cm}
\section{Chern-Simons matrix model} \label{pureCS}
\vspace{5mm}
The partition function of the Chern-Simons matrix model is \cite{Marino:2002fk}
\begin{equation}
Z\ =\ \int d^Nu\,\exp\left[ \frac{ik}{4\pi}\sum_{i=1}^N(u_i)^2 \right]\prod_{i<j}^N\sinh^2\frac{u_i-u_j}2.
\label{Z-pureCS}
\end{equation}
The overall constant which is irrelevant in the planar limit has been omitted.
In the planar limit, any relevant quantities of this model are determined by the solution of the saddle-point equations
\begin{equation}
\frac{k}{2\pi i}u_i\ =\ \sum_{j\ne i}\coth\frac{u_i-u_j}2.
\label{SPE-1}
\end{equation}
For example, the vev of the Wilson loop is given in terms of the solution as \cite{Kapustin:2009kz}
\begin{equation}
\langle W\rangle\ =\ \frac1N\sum_{i=1}^Ne^{u_i}.
\end{equation}
The symmetry of the equations (\ref{SPE-1}) implies that the distribution of $\{u_i\}$ is invariant under the reflection, that is, the equality
\begin{equation}
\{\, u_1, \cdots, u_N\ \}\ =\ \{\, -u_1, \cdots, -u_N\ \}
\label{reflection}
\end{equation}
between two sets holds.
It is convenient to introduce new variables $x_i:=e^{u_i}$ in terms of which (\ref{SPE-1}) can be written as
\begin{equation}
\log x_i\ =\ \frac tN\sum_{j\ne i}\frac{x_i+x_j}{x_i-x_j},
\label{SPE-x}
\end{equation}
where $t$ is the 't~Hooft coupling defined as
\begin{equation}
t\ :=\ \frac{2\pi iN}{k}.
\end{equation}
The condition (\ref{reflection}) for $u_i$ is translated to
\begin{equation}
\{\ x_1, \cdots, x_N\ \}\ =\ \{\ x_1^{-1}, \cdots, x_N^{-1}\ \}.
\label{inversion}
\end{equation}
The solution of (\ref{SPE-x}) can be encoded in the resolvent defined as
\begin{equation}
v(z)\ :=\ \frac tN\sum_{i=1}^N\frac{z+x_i}{z-x_i}.
\label{resolvent-pureCS}
\end{equation}
The large $z$ expansion
\begin{equation}
v(z)\ =\ t+2t\langle W\rangle z^{-1}+O(z^{-2})
\end{equation}
provides us with the interesting physical quantities.
Suppose that $t>0$.
Then the equations (\ref{SPE-x}) can be interpreted as the equations for $N$ particles, lying on the real axis in $\mathbb{C}$, interacting among them and with an external log-type force.
In this system, all $x_i>0$ are distributed around $x=1$.
In the planar limit, the distribution of the eigenvalues $x_i$ becomes dense, and form an interval $[a,b]$ with $0<a<b$.
The equality (\ref{inversion}) implies $ab=1$.
The resolvent $v(z)$ becomes a holomorphic function on $\mathbb{C}\backslash [a,b]$ with a branch cut on $[a,b]$.
As $t$ changes continuously to a complex value, the branch points $z=a,b$ move around in $\mathbb{C}$ while keeping $ab=1$ satisfied.
We denote the branch cut by $[a,b]$ even when it does not lie on the real axis.
The equations (\ref{SPE-x}) in the planar limit can be written in terms of $v(z)$ as
\begin{equation}
2\log x\ =\ v(x_+)+v(x_-), \hspace{1cm} x\in[a,b]
\label{saddle-v(z)}
\end{equation}
where $x_+$ ($x_-$) is the point in $\mathbb{C}$ slightly above (below) $x$ on the branch cut $[a,b]$.
Requiring the finiteness of $v(z)$ at the branch points and at infinity, $v(z)$ is uniquely determined as
\begin{equation}
v(z)\ =\ 2\log\frac{z+1-\sqrt{(z-a)(z-b)}}{\sqrt{a}+\sqrt{b}}.
\label{v(z)-pureCS}
\end{equation}
Note that the square-root is defined such that $\sqrt{(z-a)(z-b)}\to z$ for large positive $z$.
The definition (\ref{resolvent-pureCS}) implies $t=v(\infty)$.
This relates $t$ with $a$ as
\begin{equation}
t\ =\ 2\log\frac{\sqrt{a}+\sqrt{b}}2.
\end{equation}
\vspace{5mm}
The expression (\ref{v(z)-pureCS}) looks rather complicated compared to the resolvents of Hermitian matrix models.
The logarithmic form seems to suggest that the fundamental object of the Chern-Simons matrix model would not be $v(z)$ itself but the exponential of $v(z)$.
Indeed, the spectral curve of this model is given as
\begin{equation}
e^{v+t}-(z+1)e^{\frac12(v+t)}+e^tz\ =\ 0,
\label{curve}
\end{equation}
which plays a role in a relation between Chern-Simons theory and a topological string theory \cite{Aganagic:2002wv}.
It would be desirable if $e^{v(z)}$ could be determined directly from the saddle-point equation (\ref{saddle-v(z)}).
In fact, this can be realized for the Chern-Simons matrix model and the lens space matrix models \cite{Halmagyi:2003ze}.
However, a generalization of the techniques used in \cite{Halmagyi:2003ze} suitable for other matrix models does not look straightforward.
\vspace{5mm}
For the case of Hermitian matrix models, the spectral curve can be derived from the loop equation.
See e.g. \cite{Itoyama:2009sc} for a recent application of the loop equation to Hermitian matrix models.
It is interesting if the loop equation for the Chern-Simons matrix model would reproduce (\ref{curve}).
For Hermitian matrix models, the loop equation is nothing but the Schwinger-Dyson equations which imply that the partition function of the matrix model satisfies the Virasoro constraints \cite{Itoyama:1990mf,David:1990ge,Mironov:1990im}.
The Virasoro constraints for the Chern-Simons matrix model were studied in \cite{Nedelin:2015mio}.
According to \cite{Nedelin:2015mio}, the corresponding Schwinger-Dyson equations can be organized into
\begin{equation}
\int d^Nu\sum_{m=1}^N\frac{\partial}{\partial u_m}\left( \frac1{z-e^{u_m}}\exp\left[ \frac{ik}{4\pi}\sum_{i=1}^N(u_i)^2 \right]\prod_{i<j}^N\sinh^2\frac{u_i-u_j}2 \right)\ =\ 0.
\end{equation}
In the planar limit, this can be rewritten as
\begin{equation}
z\,\omega(z)^2-t\,\omega(z)\ =\ \log z\cdot\omega(z)-g(z), \hspace{1cm} g(z)\ :=\ \frac tN\sum_{i=1}^N\frac{\log z-u_i}{z-e^{u_i}},
\label{loop-1}
\end{equation}
where $\omega(z)$ is defined in terms of the solution of (\ref{SPE-1}) as
\begin{equation}
\omega(z)\ :=\ \frac tN\sum_{i=1}^N\frac1{z-e^{u_i}}.
\end{equation}
The function $g(z)$ turns out not to be a simple function like a polynomial, contrary to the case of Hermitian matrix models.
The function $g(z)$ is free from the square-root branch cut, but instead, it has a logarithmic branch cut.
The discontinuity along the cut is
\begin{equation}
g(x_+)-g(x_-)\ =\ 2\pi i\,\omega(x), \hspace{1cm} x\in(-\infty,0].
\end{equation}
This implies that $g(z)$ can be given in terms of $\omega(z)$ as
\begin{equation}
g(z)\ =\ \int_{-\infty}^0dx\,\frac{\omega(x)}{x-z}.
\end{equation}
Therefore, the loop equation for the Chern-Simons matrix model is not an algebraic equation but the following non-linear integral equation
\begin{equation}
z\,\omega(z)^2-t\,\omega(z)\ =\ \log z\cdot\omega(z)-\int_{-\infty}^0dx\,\frac{\omega(x)}{x-z}.
\label{loop-integral}
\end{equation}
Unfortunately, this equation looks quite difficult to solve.
It is interesting to notice that at least one can guess the analytic structure of $\omega(z)$ from the integral equation (\ref{loop-integral}).
One may find that each of two terms in the right-hand side of the equation (\ref{loop-integral}) has a logarithmic branch cut, but they cancel exactly between them.
The non-linear structure of the left-hand side suggests the existence of a square-root branch cut in $\omega(z)$.
Let $x\in(-\infty,0]$ be a point in $\mathbb{C}$ and $\tilde{x}$ be the corresponding point on the second Riemann sheet of $\omega(z)$.
Then $\omega(\tilde{x})$ is different from $\omega(x)$ appearing in the integral.
As a result, on the second Riemann sheet, the cancellation in the right-hand side is incomplete, and a logarithmic branch cut appears in $\omega(z)$.
This is indeed the expected analytic structure of $\omega(z)$ since it is related to $v(z)$ given by (\ref{v(z)-pureCS}) as
\begin{equation}
v(z)\ =\ 2tz\,\omega(z)-t.
\end{equation}
It would be very interesting to find how to solve the integral equation (\ref{loop-integral}) and its generalizations derived from various Chern-Simons-matter matrix models.
\vspace{5mm}
We have observed that the logarithmic form of $v(z)$ makes the analysis of the Chern-Simons matrix model complicated.
It is interesting to notice that, in addition to exponentiating $v(z)$, there is another way to avoid dealing with the logarithmic form of $v(z)$.
If one takes the derivative of $v(z)$, one obtains
\begin{equation}
zv'(z)\ =\ 1-\frac{z-1}{\sqrt{(z-a)(z-b)}}.
\end{equation}
The large $z$ expansion of $zv'(z)$ is
\begin{equation}
zv'(z)\ =\ -2t\langle W\rangle z^{-1}+O(z^{-2}),
\end{equation}
which preserves the information on $\langle W\rangle$ and all the higher moments.
The missing information on $v(\infty)=t$ can be recovered via
\begin{equation}
t\ =\ -\frac12\int_C\frac{dz}{2\pi i}\frac{\log z}{z} zv'(z),
\end{equation}
where $C$ is a contour encircling the branch cut $[a,b]$ counterclockwise and excluding the origin.
Therefore, it turns out to be sufficient to determine $zv'(z)$ for the investigation of the Chern-Simons matrix model in the planar limit.
One finds that $zv'(z)$ is a solution of the equation
\begin{equation}
2\ =\ \omega(x_+)+\omega(x_-), \hspace{1cm} x\in[a,b],
\end{equation}
which is obtained from the derivative of (\ref{saddle-v(z)}).
The solution $\omega(z)$ is uniquely determined by requiring that it has the following properties:
\begin{itemize}
\item $\omega(z)$ is a holomorphic function on $\mathbb{C}\backslash[a,b]$,
\item $\sqrt{(z-a)(z-b)}\,\omega(z)$ is finite at $z=a$ and $z=b$,
\item $\omega(z)=O(z^{-1})$ for large $z$, and
\item $\omega(z)$ satisfies
\begin{equation}
\omega(z^{-1})\ =\ \omega(z).
\end{equation}
\end{itemize}
Note that the last condition is a consequence of (\ref{inversion}).
\vspace{5mm}
In the next section, we determine (a linear combination of) the derivatives of the resolvents for a family of Chern-Simons-matter matrix models whose gauge group is of the form ${\rm U}(N_1)\times{\rm U}(N_2)$.
We find that the resolvents and (a linear combination of) the vevs of the BPS Wilson loops can be written explicitly for all such matrix models.
\vspace{1cm}
\section{Chern-Simons-matter matrix models with 2 nodes} \label{2-node}
\vspace{5mm}
In this section, we investigate a Chern-Simons-matter matrix model obtained via the supersymmetric localization from a Chern-Simons-matter theory with
\begin{itemize}
\item ${\cal N}\ge3$ supersymmetry,
\item the gauge group ${\rm U}(N_1)_{k_1}\times{\rm U}(N_2)_{k_2}$, and
\item $n$ bi-fundamental hypermultiplets.
\end{itemize}
The family of such theories includes ABJM theory, ABJ theory, GT theory \cite{Gaiotto:2009mv} and theories discussed in \cite{Suyama:2013fua}.
\vspace{5mm}
The partition function of the matrix model is given as \cite{Kapustin:2009kz}
\begin{equation}
Z\ =\ \int d^{N_1}u\,d^{N_2}w\,\exp\left[ \frac{ik_1}{4\pi}\sum_{i=1}^{N_1}(u_i)^2+\frac{ik_2}{4\pi}\sum_{a=1}^{N_2}(w_a)^2 \right]\frac{\prod_{i<j}^{N_1}\sinh^2\frac{u_i-u_j}2\prod_{a<b}^{N_2}\sinh^2\frac{w_a-w_b}2}{\prod_{i=1}^{N_1}\prod_{a=1}^{N_2}\cosh^n\frac{u_i-w_a}2}.
\label{Z-2node}
\end{equation}
The saddle-point equations are
\begin{eqnarray}
\frac{k_1}{2\pi i}u_i &=& \sum_{j\ne i}^{N_1}\coth\frac{u_i-u_j}2-\frac n2\sum_{a=1}^{N_2}\tanh\frac{u_i-w_a}2, \\
\frac{k_2}{2\pi i}w_a &=& \sum_{b\ne a}^{N_2}\coth\frac{w_a-w_b}2-\frac n2\sum_{i=1}^{N_1}\tanh\frac{w_a-u_i}2.
\end{eqnarray}
In terms of new variables $x_i:=e^{u_i}$ and $y_a:=-e^{w_a}$, these equations can be written as
\begin{eqnarray}
\frac{k_1}{2\pi i}\log x_i &=& \sum_{j\ne i}^{N_1}\frac{x_i+x_j}{x_i-x_j}-\frac n2\sum_{a=1}^{N_2}\frac{x_i+y_a}{x_i-y_a},
\label{2N-saddle1} \\
\frac{k_2}{2\pi i}\log(-y_a) &=& \sum_{b\ne a}^{N_2}\frac{y_a+y_b}{y_a-y_b}-\frac n2\sum_{i=1}^{N_1}\frac{y_a+x_i}{y_a-x_i}.
\label{2N-saddle2}
\end{eqnarray}
To define the planar limit in a symmetric manner, we introduce an auxiliary parameter $k$ and define
\begin{equation}
t_1\ :=\ \frac{2\pi iN_1}{k}, \hspace{1cm} t_2\ :=\ \frac{2\pi iN_2}{k}, \hspace{1cm} \kappa_1\ :=\ \frac{k_1}k, \hspace{1cm} \kappa_2\ :=\ \frac{k_2}k.
\label{parameters-2node}
\end{equation}
The planar limit is then defined as the limit $k\to\infty$ while keeping these parameters fixed.
We define two resolvents for two sets $\{x_i\}, \{y_a\}$ of eigenvalues as
\begin{equation}
v_1(z)\ :=\ \frac{t_1}{N_1}\sum_{i=1}^{N_1}\frac{z+x_i}{z-x_i}, \hspace{1cm} v_2(z)\ :=\ \frac{t_2}{N_2}\sum_{a=1}^{N_2}\frac{z+y_a}{z-y_a}.
\label{def-2Nres}
\end{equation}
In the planar limit, $v_1(z)$ becomes a holomorphic function on $\mathbb{C}\backslash[a_1,b_1]$, and $v_2(z)$ becomes a holomorphic function on $\mathbb{C}\backslash[a_2,b_2]$.
As in the previous section, $a_1b_1=a_2b_2=1$ is assumed.
In terms of these resolvents, the saddle-point equations (\ref{2N-saddle1})(\ref{2N-saddle2}) can be written as
\begin{eqnarray}
2\kappa_1\log x &=& v_1(x_+)+v_1(x_-)-n\,v_2(x), \hspace{1cm} x\in[a_1,b_1], \\
2\kappa_2\log(-y) &=& v_2(y_+)+v_2(y_-)-n\,v_1(y), \hspace{1.1cm} y\in[a_2,b_2].
\end{eqnarray}
Our observation in the previous section suggests that, instead of dealing with these equations, we should investigate the following equations
\begin{eqnarray}
{2\kappa_1} &=& xv_1'(x_+)+xv_1'(x_-)-n\,xv_2'(x),
\label{res-eq1} \\
{2\kappa_2} &=& yv_2'(y_+)+yv_2'(y_-)-n\,yv_1'(y).
\label{res-eq2}
\end{eqnarray}
It is convenient to combine the two resolvents into a vector-valued resolvent
\begin{equation}
v(z)\ :=\ (v_1(z), v_2(z)).
\end{equation}
In terms of $v(z)$, the equations (\ref{res-eq1})(\ref{res-eq2}), together with
\begin{equation}
v_1(y_+)\ =\ v_1(y_-), \hspace{1cm} v_2(x_+)\ =\ v_1(x_-),
\end{equation}
which are required by the definition (\ref{def-2Nres}), can be written as follows:
\begin{eqnarray}
\left( {2\kappa_1}, 0 \right) &=& xv'(x_+)-xv'(x_-)M_1,
\label{res-vec1} \\
\left( 0, {2\kappa_2} \right) &=& yv'(y_+)-yv'(y_-)M_2,
\label{res-vec2}
\end{eqnarray}
where
\begin{equation}
M_1\ :=\ \left[
\begin{array}{cc}
-1 & 0 \\
n & 1
\end{array}
\right], \hspace{1cm} M_2\ :=\ \left[
\begin{array}{cc}
1 & n \\
0 & -1
\end{array}
\right].
\end{equation}
The properties required for the solution of (\ref{res-vec1})(\ref{res-vec2}) are as follows:
\begin{itemize}
\item $zv'(z)$ is holomorphic on $\mathbb{C}\backslash([a_1,b_1]\cup[a_2,b_2])$,
\item $s(z)zv'(z)$ is finite at the branch points, where
\begin{equation}
s(z)\ :=\ \sqrt{(z-a_1)(z-b_1)(z-a_2)(z-b_2)},
\label{s(z)}
\end{equation}
\item $zv'(z)=O(z^{-1})$ for large $z$, and
\item $zv'(z)$ satisfies
\begin{equation}
z^{-1}v'(z^{-1})\ =\ zv'(z).
\label{inversion2}
\end{equation}
\end{itemize}
The 't Hooft couplings $t:=(t_1,t_2)$ are given as
\begin{equation}
(t_1,0)\ =\ -\frac12\int_{C_1}\frac{dz}{2\pi i}\frac{\log z}{z} zv'(z), \hspace{1cm} (0,t_2)\ =\ -\frac12\int_{C_2}\frac{dz}{2\pi i}\frac{\log z}{z} zv'(z),
\end{equation}
where $C_1$ and $C_2$ are contours encircling $[a_1,b_1]$ and $[a_2,b_2]$ counterclockwise, respectively, and excluding the origin.
The vevs of the BPS Wilson loops are obtained from the large $z$ expansion of $zv'(z)$ as
\begin{equation}
zv'(z)\ =\ -2\left( t_1\langle W_1\rangle, -t_2\langle W_2\rangle \right)z^{-1}+O(z^{-2}).
\label{WL-general}
\end{equation}
Note that there is a minus sign in front of $\langle W_2\rangle$ since
\begin{equation}
\langle W_2\rangle\ =\ \frac1{N_2}\sum_{a=1}^{N_2}e^{w_a}\ =\ -\frac1{N_2}\sum_{a=1}^{N_2}y_a.
\end{equation}
\vspace{5mm}
\subsection{The case $n=2$} \label{n=2}
\vspace{5mm}
First, we consider the case $n=2$.
The matrix model with $n=2$ corresponds to ABJM theory and ABJ theory when $\kappa_1+\kappa_2=0$.
In general ($\kappa_1+\kappa_2\ne0$), the matrix model is derived from GT theory which is expected to describe a massive Type IIA theory.
In the case $n=2$, the equations (\ref{res-vec1})(\ref{res-vec2}) can be simplified as follows.
Notice that the matrices $M_1$ and $M_2$ have a common eigenvector:
\begin{eqnarray}
\left[
\begin{array}{cc}
-1 & 0 \\
2 & 1
\end{array}
\right]\left[
\begin{array}{c}
1 \\
-1
\end{array}
\right]
\ =\ -\left[
\begin{array}{c}
1 \\
-1
\end{array}
\right] \ =\
\left[
\begin{array}{cc}
1 & 2 \\
0 & -1
\end{array}
\right]\left[
\begin{array}{c}
1 \\
-1
\end{array}
\right].
\end{eqnarray}
Multiplying this eigenvector from the right, the equations (\ref{res-vec1})(\ref{res-vec2}) become
\begin{eqnarray}
{2\kappa_1} &=& \omega(x_+)+\omega(x_-),
\label{res1} \\
-{2\kappa_2} &=& \omega(y_+)+\omega(y_-),
\label{res2}
\end{eqnarray}
where $\omega(z)$ is defined as
\begin{equation}
\omega(z)\ :=\ zv'(z)\left[
\begin{array}{c}
1 \\
-1
\end{array}
\right]\ =\ zv_1'(z)-zv_2'(z).
\end{equation}
The required properties for $zv'(z)$ is translated to those of $\omega(z)$ by this definition.
Once $\omega(z)$ is determined, one can show that $zv'(z)$ can be given in terms of integrals of $\omega(z)$.
The function $\omega(z)$ already contains a lot of information.
For example, the 't Hooft couplings are given as
\begin{equation}
t_1\ =\ -\frac12\int_{C_1}\frac{d\xi}{2\pi i}\frac{\log z}z\omega(z), \hspace{1cm} t_2\ =\ \frac12\int_{C_2}\frac{d\xi}{2\pi i}\frac{\log z}z\omega(z),
\end{equation}
and the large $z$ expansion of $\omega(z)$ gives
\begin{equation}
\omega(z)\ =\ -2(t_1\langle W_1\rangle+t_2\langle W_2\rangle)z^{-1}+O(z^{-2}).
\end{equation}
Note that the linear combination of $\langle W_1\rangle$ and $\langle W_2\rangle$ appearing above gives the vev of the half-BPS Wilson loop \cite{Drukker:2009hy}, in the case of ABJM theory and ABJ theory.
\vspace{5mm}
The solution of (\ref{res1})(\ref{res2}) with the required properties can be given as follows.
Let $\Omega(z,\xi)$ be a holomorphic function on $\mathbb{C}\backslash\{a_1,b_1,a_2,b_2,\xi\}$ with a parameter $\xi$, satisfying the following conditions:
\begin{itemize}
\item $\Omega(z,\xi)$ has a monodromy $-1$ at the points $z=a_1,b_1,a_2,b_2$, and
\item $\Omega(z,\xi)$ has a simple pole at $z=\xi$ with the residue $1$.
\end{itemize}
Using these properties of $\Omega(z,\xi)$, One can easily check that
\begin{equation}
\omega_0(z)\ :=\ \kappa_1\int_{C_1}\frac{d\xi}{2\pi i}\Omega(z,\xi)-\kappa_2\int_{C_2}\frac{d\xi}{2\pi i}\Omega(z,\xi)
\end{equation}
is a solution of (\ref{res1})(\ref{res2}).
The finiteness of $s(z)\omega_0(z)$ at the branch points suggests that an appropriate choice of $\Omega(z,\xi)$ is
\begin{equation}
\Omega(z,\xi)\ =\ \frac{h(z)}{h(\xi)}\frac1{z-\xi}\frac{s(\xi)}{s(z)}
\end{equation}
with $h(z)$ an entire function.
A convenient choice turns out to be $h(z)=z$.
One finds that $\omega_0(z)$ does not satisfy the inversion condition
\begin{equation}
\omega(z^{-1})\ =\ \omega(z)
\label{inversion3}
\end{equation}
deduced from (\ref{inversion2}).
This problem is remedied by noticing that $\omega_0(z^{-1})$ also satisfies the equations (\ref{res1})(\ref{res2}).
Therefore,
\begin{equation}
\omega(z)\ =\ \frac12\omega_0(z)+\frac1{2}\omega_0(z^{-1})
\end{equation}
is a solution which also satisfies the inversion condition (\ref{inversion3}).
One can check that this is the only solution of (\ref{res1})(\ref{res2}) which has all the required properties deduced from those of $zv'(z)$.
By deforming the integration contour, $\omega(z)$ can be written as
\begin{equation}
\omega(z)\ =\ -{\kappa_2}\left[ 1-\frac{z^2-1}{s(z)} \right]-\frac{\kappa_1+\kappa_2}2\frac{z^2-1}{s(z)}F(z).
\end{equation}
The function $F(z)$ defined as
\begin{equation}
F(z)\ :=\ \int_{C_1}\frac{d\xi}{2\pi i}\frac{s(\xi)}{\xi(\xi-z)(\xi-z^{-1})},
\end{equation}
which is absent for ABJM theory and ABJ theory ($\kappa_1+\kappa_2=0$), can be written in terms of the complete elliptic integrals.
Explicitly,
\begin{eqnarray}
F(z)
&=& \frac{8\alpha}{\sqrt{(1-\alpha^2)(1-k^2\alpha^2)}}\frac{f(\alpha^{-2})-f(k(z)^2)}{(2-z-z^{-1})(1-k(z)^2\alpha^2)},
\end{eqnarray}
where
\begin{equation}
f(z)\ :=\ -\frac{2(1-z)(k^2-z)}{\pi i z}\Pi_1(-z,k)-\frac2{\pi i}E(k)+\frac{2(1-z)k^2}{\pi iz}K(k),
\end{equation}
and
\begin{equation}
\alpha\ :=\ \frac{1+a_1}{1-a_1}, \hspace{1cm} k(z)^2\ := \frac{z+z^{-1}+2}{z+z^{-1}-2}\alpha^{-2}, \hspace{1cm} k^2\ :=\ k(a_2)^2.
\end{equation}
The planar resolvent for ABJM theory and ABJ theory was obtained in \cite{Marino:2009jd} using the result of \cite{Halmagyi:2003ze}.
The resolvent $\omega(z)$ determined above for the case $\kappa_1+\kappa_2=0$ can be derived from the result of \cite{Marino:2009jd}.
We have found that the resolvent for general $\kappa_1$ and $\kappa_2$ has a quite complicated expression compared to that for the case $\kappa_1+\kappa_2=0$.
\vspace{5mm}
The large $z$ expansion of $\omega(z)$ gives
\begin{eqnarray}
& & t_1\langle W_1\rangle+t_2\langle W_2\rangle \nonumber \\
&=& -\frac{\kappa_2}{4}(a_1+\cdots+b_2)+\frac{2\alpha(\kappa_1+\kappa_2)}{\pi i\sqrt{(1-\alpha^2)(1-k^2\alpha^2)}}\left[ \frac{1-k^2\alpha^4}{\alpha^2}\Pi_1(-\alpha^{-2},k)-E(k)+(1+k^2\alpha^2)K(k) \right]. \nonumber \\
\label{n=2 WL}
\end{eqnarray}
Recall that, in various examples of AdS/CFT correspondence, the vev of a Wilson loop diverges as
\begin{equation}
\log |\langle W\rangle|\ =\ c\lambda^\gamma,
\end{equation}
in the limit where the 't Hooft coupling $\lambda$ is large.
For example, the exponent $\gamma$ is $\frac12$ for ${\cal N}=4$ super Yang-Mills theory \cite{Erickson:2000af}\cite{Pestun:2007rz} and ABJM theory \cite{Suyama:2009pd}\cite{Marino:2009jd}, and $\gamma=\frac13$ for GT theory \cite{Suyama:2011yz}.
Therefore, one may be interested in a divergent behavior of the expression (\ref{n=2 WL}) since it would be a sign of a possible existence of a dual gravity description via AdS/CFT correspondence.
Obviously, the first term of (\ref{n=2 WL}) diverges when $b_1$ or $b_2$ diverges, or in other words, when $a_1\to0$ or $a_2\to0$.
The second term of (\ref{n=2 WL}) is divergent if $\alpha\to1$ or $k\to1$, which correspond to
$a_1\to0$ or $a_2\to a_1$, respectively.
The expression (\ref{n=2 WL}) shows that there is no other divergent behavior.
It was observed, e.g. in \cite{Suyama:2011yz}, that the simultaneous limit $\alpha\to1$ and $k\to1$ corresponds to the limit in which a weak gravity dual exists.
In another limit, say $\alpha\to1$ but $k$ is different from 1, the distribution of two sets of eigenvalues becomes hierarchical, that is, the distribution of $\{x_i\}$ becomes large while that of $\{y_a\}$ is not.
In such a situation, the two sets of equations (\ref{2N-saddle1})(\ref{2N-saddle2}) would decouple effectively, and each set of equations would become similar to the saddle-point equations (\ref{SPE-x}) for the Chern-Simons matrix model.
\vspace{5mm}
\subsection{The cases $n\ne 2$} \label{general-n}
\vspace{5mm}
Next, consider the other cases $n\ne2$.
Recall that we would like to solve the following equations
\begin{eqnarray}
\left( {2\kappa_1}, 0 \right) &=& xv'(x_+)-xv'(x_-)M_1, \\
\left( 0, {2\kappa_2} \right) &=& yv'(y_+)-yv'(y_-)M_2.
\end{eqnarray}
For the cases $n\ne2$, a constant vector $c:=(c_1,c_2)$ satisfies these equations since there exist the constants $c_1, c_2$ which satisfy
\begin{equation}
(2\kappa_1, 2\kappa_2)\ =\ (c_1, c_2)\left[
\begin{array}{cc}
2 & -n \\
-n & 2
\end{array}
\right].
\end{equation}
Define a function $\omega(z)$ such that $zv'(z)$ is given as
\begin{equation}
zv'(z)\ =\ c+\omega(z).
\end{equation}
Then, $\omega(z)$ satisfies
\begin{equation}
\omega(x_+) \ =\ \omega(x_-)M_1, \hspace{1cm}
\omega(y_+) \ =\ \omega(y_-)M_2.
\label{eq-omega}
\end{equation}
It is convenient to consider, instead of $\omega(z)$, a function $f(z)$ defined as
\begin{equation}
f(z)\ :=\ s(z)\omega(z)
\end{equation}
which is required to have the following properties:
\begin{itemize}
\item $f(z)$ is holomorphic on $\mathbb{C}\backslash([a_1,b_1]\cup[a_2,b_2])$,
\item $f(z)$ is finite at the branch points,
\item for large $z$, $f(z)$ behaves as
\begin{equation}
f(z)\ =\ -cz^2+O(z),
\label{f-infty}
\end{equation}
\item $f(z)$ satisfies
\begin{equation}
f(z^{-1})\ =\ -z^{-2}f(z).
\end{equation}
\end{itemize}
The equations (\ref{eq-omega}) can be written in terms of $f(z)$ as
\begin{equation}
f(x_+) \ =\ -f(x_-)M_1, \hspace{1cm}
f(y_+) \ =\ -f(y_-)M_2.
\label{eq-f}
\end{equation}
We will show that the solution of (\ref{eq-f}) with the above properties is uniquely determined.
\vspace{5mm}
The problem of determining $f(z)$ turns out to be a generalization of the problem in \cite{Eynard:1995nv}\cite{Eynard:1995zv} discussing the ${\rm O}(n)$ model \cite{Kostov:1988fy}\cite{Gaudin:1989vx}, and therefore, the analysis developed in \cite{Eynard:1995nv}\cite{Eynard:1995zv} can be applied to our problem with a suitable modification.
Our strategy is to map the double cover of $\mathbb{C}\backslash([a_1,b_1]\cup[a_2,b_2])$ to a torus $T^2$ by a map defined as
\begin{equation}
u(z)\ :=\ \frac{\varphi(z)}{2\varphi(b_1)}, \hspace{1cm} \varphi(z)\ :=\ \int_{a_1}^z\frac{d\xi}{s(\xi)},
\end{equation}
where the integration contour for $\varphi(b_1)$ lies above the branch cut $[a_1,b_1]$.
Note that $u(z)$ satisfies
\begin{equation}
u(z^{-1})\ =\ u(z)-\frac12.
\end{equation}
Let $\tau:=2u(a_2)$ be the modulus of $T^2$.
The function $f(z)$ becomes a function on $T^2$ by the inverse map
\begin{equation}
z(u)\ :=\ -\frac{\vartheta_1(u-u_0)\vartheta_1(u+u_0)}{\vartheta_1(u-u_\infty)\vartheta_1(u+u_\infty)},
\end{equation}
where $\vartheta_1(u):=\vartheta_1(u,\tau)$ is the theta function, and $u_z:=u(z)$.
In the following, the function $f(z(u))$ on $T^2$ is denoted simply by $f(u)$.
By the definition of the $u$-coordinate, $f(u)$ satisfies
\begin{equation}
f(u+1)\ =\ f(u),
\label{period}
\end{equation}
since the shift of $u$ by $1$ corresponds to a move around the branch cut $[a_1,b_1]$ in the $z$-plane.
The equations (\ref{eq-f}) can be written as
\begin{eqnarray}
f(-u) &=& -f(u)M_1,
\label{f1 to f2} \\
f(u+\tau) &=& f(u)M_1M_2.
\label{eq-f_2}
\end{eqnarray}
The matrix $M_1M_2$ can be diagonalized by a matrix $S$ defined as
\begin{equation}
S\ := \ \left[
\begin{array}{cc}
1 & 1 \\
-e^{\pi i\nu} & -e^{-\pi i\nu}
\end{array}
\right],
\end{equation}
where $\nu$ parametrizes $n$ as $n=2\cos\pi\nu$.
Therefore, the equations (\ref{period})(\ref{eq-f_2}) for a vector-valued function $f(u)$ can be split into two sets of equations for two scalar-valued functions.
Define $(\tilde{f}_1(u),\tilde{f}_2(u)):=f(u)S$.
Then $\tilde{f}_1(u)$ satisfies
\begin{equation}
\tilde{f}_1(u+1)\ =\ \tilde{f}_1(u), \hspace{1cm} \tilde{f}_1(u+\tau)\ =\ e^{2\pi i\nu}\tilde{f}_1(u).
\label{eq-tilde{f}}
\end{equation}
Note that $S$ also simplifies $M_1$ and $M_2$ separately as
\begin{equation}
S^{-1}M_1S\ =\ \left[
\begin{array}{cc}
0 & -1 \\
-1 & 0
\end{array}
\right], \hspace{1cm} S^{-1}M_2S\ =\ \left[
\begin{array}{cc}
0 & -e^{-2\pi i\nu} \\
-e^{2\pi i\nu} & 0
\end{array}
\right].
\end{equation}
The equation (\ref{f1 to f2}) relates $\tilde{f}_2(u)$ to $\tilde{f}_1(u)$ as
\begin{equation}
\tilde{f}_2(u)\ =\ \tilde{f}_1(-u).
\end{equation}
In the following, we determine $\tilde{f}_1(u)$ which has the required properties deduced from those of $f(z)$.
\vspace{5mm}
A solution $G(u)$ of the equations (\ref{eq-tilde{f}}) can be constructed in terms of the theta functions, although it is not uniquely determined.
Our choice of $G(u)$ is
\begin{equation}
G(u)\ :=\ \frac{\vartheta_1(u-u_\nu)\vartheta_1(u-u_\nu+\frac12)}{\vartheta_1(u-u_\infty)\vartheta_1(u+u_\infty)},
\end{equation}
where $u_\nu:=\frac12\nu+\frac14$.
An advantage of this choice is that $G(u)$ has a nice inversion property
\begin{equation}
G(u(z^{-1}))\ =\ -\frac1zG(u(z)).
\label{inversion-G}
\end{equation}
The product $g(u):=\tilde{f}_1(u)G(u)^{-1}$ then satisfies
\begin{equation}
g(u+1)\ =\ g(u), \hspace{1cm} g(u+\tau)\ =\ g(u),
\end{equation}
that is, $g(u)$ is an elliptic function.
Since $\tilde{f}_1(z)$ has a double pole at infinity and otherwise finite, $g(u)$ must have simple poles at $u=u_\infty, -u_\infty, u_\nu, u_\nu-\frac12$.
The Riemann-Roch theorem implies that elliptic functions with at most four such simple poles form a four-dimensional vector space $V$.
Therefore, $g(u)$ can be written as
\begin{equation}
g(u)\ =\ r_1g_1(u)+r_2g_2(u)+r_3g_3(u)+r_4g_4(u),
\end{equation}
when a basis of $V$ is given.
We choose a basis as
\begin{equation}
\begin{array}{rclcrcl}
g_1(u) & := & 1, & \hspace{5mm} & g_2(u) & := & \displaystyle{-\frac{\vartheta_1(u-u_0)\vartheta_1(u+u_0)}{\vartheta_1(u-u_\infty)\vartheta_1(u+u_\infty)}}, \\ [4mm]
g_3(u) & := & \displaystyle{\frac{\vartheta_1(u-u_0)\vartheta_1(u-u_\nu+\frac12)}{\vartheta_1(u-u_\infty)\vartheta_1(u-u_\nu)}}, & \hspace{5mm} & g_4(u) & := & \displaystyle{-\frac{\vartheta_1(u+u_0)\vartheta_1(u-u_\nu)}{\vartheta_1(u+u_\infty)\vartheta_1(u-u_\nu+\frac12)}}.
\end{array}
\label{basis}
\end{equation}
Due to the inversion property (\ref{inversion-G}) of $G(u)$, the elliptic function $g(u)$ is required to satisfy
\begin{equation}
g(u(z^{-1}))\ =\ \frac1zg(u(z)).
\end{equation}
This condition implies
\begin{equation}
r_1\ =\ r_2, \hspace{1cm} r_3\ =\ r_4.
\end{equation}
The remaining two coefficients, say $r_1$ and $r_3$, are fixed by requiring the asymptotic behavior (\ref{f-infty}) of $f(z)$ at infinity.
Equivalently, they are determined by requiring $f(0)=c$.
In terms of the $u$ variable, this implies
\begin{equation}
\tilde{f}_1(u_0)\ =\ \tilde{c}_1, \hspace{1cm} \tilde{f}_2(u_0)\ =\ \tilde{f}_1(-u_0)\ =\ \tilde{c}_2,
\end{equation}
where $\tilde{c}:=cS$.
The solution is
\begin{eqnarray}
r_1
&=& \frac1{g_3(-u_0)-g_4(u_0)}\left[ \tilde{c}_1\frac{g_3(-u_0)}{G(u_0)}-\tilde{c}_2\frac{g_4(u_0)}{G(-u_0)} \right],
\label{r_1} \\
r_3
&=& \frac1{g_3(-u_0)-g_4(u_0)}\left[ -\frac{\tilde{c}_1}{G(u_0)}+\frac{\tilde{c}_2}{G(-u_0)} \right].
\label{r_3}
\end{eqnarray}
Now, the elliptic function $g(u)$ has been determined completely.
The resolvent $zv'(z)$ is therefore given in terms of the theta functions explicitly as
\begin{equation}
zv'(z)\ =\ c+\left[ \frac1{s(z)}g(u(z))G(u(z)),\ \frac1{s(z)}g(-u(z))G(-u(z))\, \right]S^{-1}.
\label{resolvent-general}
\end{equation}
Note that a part of the above calculations can be applied to the case $n=2$ as long as $\kappa_1+\kappa_2=0$.
In fact, the multiplication by $S$ for the case $n=2$ corresponds to the multiplication by ${1\brack-1}$ used in subsection \ref{n=2}.
\vspace{5mm}
One can show that the 't~Hooft couplings can be written as
\begin{equation}
t_1\ =\ -\frac12\int_{C_1}\frac{dz}{2\pi i}\frac{\log z}{z}\frac{\tilde{f}_1(z)}{s(z)}, \hspace{1cm} t_2\ =\ \frac12e^{-\pi i\nu}\int_{C_2}\frac{dz}{2\pi i}\frac{\log z}{z}\frac{\tilde{f}_1(z)}{s(z)}.
\end{equation}
The large $z$ expansion (\ref{WL-general}) of the resolvent gives the vevs of the BPS Wilson loops.
Equivalently, they can be obtained from the small $z$ expansion of $zv'(z)$:
\begin{eqnarray}
zv'(z)
&=& c-f(0)-\left[ \frac{a_1+\cdots+b_2}2f(0)+f'(0) \right]z+O(z^2).
\end{eqnarray}
We imposed $f(0)=c$ to determine the elliptic function $g(u)$.
Then, the vevs of the BPS Wilson loops are
\begin{equation}
\left( t_1\langle W_1\rangle,-t_2\langle W_2\rangle \right)\ =\ \frac{a_1+\cdots+b_2}4c+\frac12f'(0).
\end{equation}
\vspace{5mm}
Note that the coefficients $r_1$ and $r_3$ given in (\ref{r_1})(\ref{r_3}) may diverge for a particular configuration of the branch cuts.
Recall that $u_0$ is a function of the positions $a_1,\cdots,b_2$.
One can show that, as functions of $u_0$, $r_1$ and $r_3$ have poles at $u_0=u_\nu,-u_\nu+\frac12$ and at values such that $g_3(-u_0)=g_4(u_0)$.
The former cases, it is easy to show that the basis functions (\ref{basis}) degenerate as
\begin{equation}
g_3(u)\ =\ \left\{
\begin{array}{cc}
1, & (u_0=u_\nu), \\ [2mm]
-g_2(u), & (u_0=-u_\nu+\frac12)
\end{array}
\right. \hspace{1cm}
g_4(u)\ =\ \left\{
\begin{array}{cc}
g_2(u), & (u_0=u_\nu), \\ [2mm]
-1. & (u_0=-u_\nu+\frac12)
\end{array}
\right.
\end{equation}
Due to these degenerations, the poles are canceled among them, and therefore, $g(u)$ is finite for generic $u$.
Since the 't~Hooft couplings and the vevs of the Wilson loops can be given in terms of contour integrals, the finiteness of $g(u)$ implies the finiteness of these quantities.
Therefore, the poles at $u_0=u_\nu,-u_\nu+\frac12$ are physically irrelevant.
The latter case corresponds to the case $z(u_\nu)=-1$.
This implies
\begin{equation}
u_\nu\ =\ \frac12\tau\pm\frac14 \mbox{ mod }\mathbb{Z}+\mathbb{Z}\tau.
\end{equation}
In terms of $\nu$, this condition is written as
\begin{equation}
\nu\ =\ \tau \mbox{ mod }\mathbb{Z}+2\mathbb{Z}\tau.
\end{equation}
When $\tau$, which is also a function of the positions $a_1,\cdots,b_2$, is chosen such that the above equation holds for a given $\nu$, then $g(u)$ satisfies
\begin{equation}
(g_3(-u_0)-g_4(u_0))g(u)\ =\ \left[ \frac{\tilde{c}_1}{G(u_0)}-\frac{\tilde{c}_2}{G(-u_0)} \right]\Bigl[ g_3(-u_0)(1+g_2(u))-g_3(u)-g_4(u) \Bigr].
\end{equation}
Since the basis functions (\ref{basis}) are linearly independent, the right-hand side is not identically zero.
Therefore, $g(u)$ diverges for generic $u$ when $z(u_\nu)=-1$ holds.
One can check that, for example when $a_1=-a_2$ holds, the quantities $t_1\langle W_1\rangle$ and $t_2\langle W_2\rangle$ diverge.
Note that the definitions of $\langle W_1\rangle$ and $\langle W_2\rangle$ imply
\begin{equation}
|\langle W_1\rangle|\ \le\ |b_1|, \hspace{1cm} |\langle W_2\rangle|\ \le\ |b_2|.
\end{equation}
Unless the two branch cuts are hierarchical, a finite $\tau$ corresponds to finite $b_1$ and $b_2$, implying that the vevs $\langle W_1\rangle$ and $\langle W_2\rangle$ are finite.
Therefore, the divergence for $z(u_\nu)=-1$ is due to the divergence of the 't~Hooft couplings $t_1$ and $t_2$.
This means that there exists a large 't~Hooft coupling limit in the parameter space of a 2-node theory with $n>2$ at which the vevs of the Wilson loops are finite.
A similar kind of behavior was observed in \cite{Suyama:2013fua} for more general theories.
\vspace{1cm}
\section{Perturbative check} \label{check}
\vspace{5mm}
In this section, we will use the planar resolvent obtained in section \ref{2-node} for the calculation of the vevs of the Wilson loops for the 2-node theories perturbatively.
The same perturbative expansion can be also obtained directly from their localization formulas.
The match between these two results provides a non-trivial check for the validity of our formulas for the planar resolvents.
\vspace{5mm}
\subsection{Expansion from the localization formula}
\vspace{5mm}
The vev of a Wilson loop can be given in terms of a finite-dimensional integral via the supersymmetric localization \cite{Kapustin:2009kz}.
For pure Chern-Simons theory, the vev $\langle W\rangle$ is given as
\begin{equation}
\langle W\rangle\ =\ \frac1Z\int d^Nu\ \exp\left[ \frac{ik}{4\pi}\sum_{i=1}^N(u_i)^2 \right]\prod_{i<j}^N\sinh^2\frac{u_i-u_j}2\cdot \frac1N\sum_{i=1}^Ne^{u_i},
\label{W-CS}
\end{equation}
where $Z$ is defined as (\ref{Z-pureCS}).
The $1/k$ expansion of $\langle W\rangle$ can be derived in a manner explained in \cite{Kapustin:2009kz}.
The idea is to relate the vev (\ref{W-CS}) to the vevs of the Gaussian matrix model whose partition function $Z_0$ is defined as
\begin{eqnarray}
Z_0\ :=\ \int d^Nu\,\exp\left[ -\frac12\sum_{i=1}^N(u_i)^2 \right]\prod_{i=1}^N(u_i-u_j)^2.
\end{eqnarray}
The partition function $Z$ can be rewritten as follows:
\begin{eqnarray}
Z
&=& 2^{-N(N-1)}\left( \frac{2\pi i}{k} \right)^{\frac12N^2}\int d^Nu\,\exp\left[ -\frac12\sum_{i=1}^N(u_i)^2 \right]\prod_{i<j}^N(u_i-u_j)^2\cdot \sum_{n=0}^\infty \left( \frac{2\pi i}k \right)^nX_n(u),
\end{eqnarray}
where $X_n(u)$ are defined such that
\begin{eqnarray}
\exp\left[ \sum_{i<j}^N2\log\left( \frac{\sinh\sqrt{\frac{2\pi i}k}\frac{u_i-u_j}2}{\sqrt{\frac{2\pi i}k}\frac{u_i-u_j}2} \right) \right]
&=& \sum_{n=0}^\infty \left( \frac{2\pi i}k \right)^nX_n(u) \nonumber \\
&=& 1+\frac{2\pi i}k\cdot\frac13\sum_{i<j}^N\left( \frac{u_i-u_j}2 \right)^2+O(k^{-2}).
\end{eqnarray}
The same rewriting can be also performed in the presence of an operator insertion.
Therefore, the vev (\ref{W-CS}) can be written as
\begin{eqnarray}
\langle W\rangle
&=& \frac{\displaystyle{\left\langle \sum_{n=0}^\infty\left( \frac{2\pi i}k \right)^nX_n(u)\sum_{m=0}^\infty\left( \frac{2\pi i}k \right)^mW_m(u) \right\rangle_0}}{\displaystyle{\left\langle \sum_{n=0}^\infty\left( \frac{2\pi i}k \right)^nX_n(u) \right\rangle_0}} \nonumber \\
&=& 1+\frac{2\pi i}k\langle W_1(u)\rangle_0+\left( \frac{2\pi i}k \right)^2\Bigl( \langle W_2(u)\rangle_0+\langle X(u)W_1(u)\rangle_0-\langle X(u)\rangle_0\langle W_1(u)\rangle_0 \Bigr)+O(k^{-3}) \nonumber \\
\label{W-CS_2}
\end{eqnarray}
where $\langle{\cal O}(u)\rangle_0$ is the vev in the Gaussian matrix model defined as
\begin{equation}
\langle {\cal O}(u)\rangle_0\ :=\ \frac1{Z_0}\int d^Nu\,\exp\left[ -\frac12\sum_{i=1}^N(u_i)^2 \right]\prod_{i=1}^N(u_i-u_j)^2\cdot {\cal O}(u),
\end{equation}
and $W_n(u)$ are defined as
\begin{eqnarray}
W_m(u) &:=& \frac1N\sum_{i=1}^N\frac{(u_i)^{2m}}{(2m)!}.
\end{eqnarray}
The vevs in (\ref{W-CS_2}) can be calculated exactly by using the Hermite polynomials.
The results are as follows:
\begin{equation}
\begin{array}{rclcrcl}
\langle W_1(u)\rangle_0 &=& \displaystyle{\frac N2}, &\hspace{5mm}& \langle X_1(u)\rangle_0 &=& \displaystyle{\frac{N(N^2-1)}{12}}, \\ [3mm]
\langle W_2(u)\rangle_0 &=& \displaystyle{\frac{2N^2+1}{24}}, &\hspace{5mm}& \langle X_1(u)W_1(u) \rangle_0 &=& \displaystyle{\frac{(N^2-1)(N^2+2)}{24}}.
\end{array}
\end{equation}
Therefore, the perturbative expansion of the vev $\langle W\rangle$ is given as
\begin{eqnarray}
\langle W\rangle
&=& 1+\frac{\pi iN}k+\frac16\left( \frac{2\pi iN}k \right)^2\left( 1-\frac1{4N^2} \right)+O(k^{-3}).
\label{expansion-pureCS}
\end{eqnarray}
Note that this is exact in $N$.
This reproduces the first three terms of the exact result \cite{Kapustin:2009kz}
\begin{equation}
\frac1Ne^{\pi iN/k}\frac{\sin\frac{\pi N}k}{\sin\frac\pi k}\ =\ 1+\frac{\pi iN}k+\frac16\left( \frac{2\pi iN}k \right)^2\left( 1-\frac1{4N^2} \right)+\frac1{24}\left( \frac{2\pi iN}k \right)^3\left( 1-\frac1{2N^2} \right)+O(k^{-4}).
\end{equation}
\vspace{5mm}
The perturbative calculation described above can be easily extended for the application to the Chern-Simons-matter matrix models with two nodes \cite{Kapustin:2009kz}.
For these models, we use vevs of the two non-interacting Gaussian matrix models whose partition function is defined as
\begin{equation}
Z_0\ :=\ \int d^{N_1}ud^{N_2}w\,\exp\left[ -\frac12\sum_{i=1}^{N_1}(u_i)^2-\frac12\sum_{a=1}^{N_2}(w_a)^2 \right]\prod_{i<j}^{N_1}(u_i-u_j)^2\prod_{a<b}^{N_2}(w_a-w_b)^2.
\end{equation}
The vev $\langle W_1\rangle$ of the Wilson loop for the ${\rm U}(N_1)$ gauge field is given as
\begin{eqnarray}
\langle W_1\rangle
&=& 1+\frac{2\pi i}{k_1}\langle W_1(u)\rangle_0+\left( \frac{2\pi i}{k_1} \right)^2\langle W_2(u)\rangle_0+\left( \frac{2\pi i}{k_1} \right)^2\Bigl( \langle Y_1(u)W_1(u)\rangle_0-\langle Y_1(u)\rangle_0\langle W_1(u)\rangle_0 \Bigr) \nonumber \\
& & +\frac{2\pi i}{k_1}\frac{2\pi i}{k_2}\Bigl( \langle Y_2(w)W_1(u)\rangle_0-\langle Y_2(w)\rangle_0\langle W_1(u)\rangle_0 \Bigr)+O(k^{-3})
\label{W-vev}
\end{eqnarray}
where $Y_1(u)$ and $Y_2(w)$, defined as
\begin{eqnarray}
Y_1(u) &:=& \frac13\sum_{i<j}^{N_1}\left( \frac{u_i-u_j}2 \right)^2-\frac n2N_2\sum_{i=1}^{N_1}\frac{u_i^2}{4}, \\
Y_2(w) &:=& \frac13\sum_{a<b}^{N_2}\left( \frac{w_a-w_b}2 \right)^2-\frac n2N_1\sum_{a=1}^{N_2}\frac{w_a^2}{4},
\end{eqnarray}
come from the one-loop part of the integrand in (\ref{Z-2node}).
The values of the vevs in (\ref{W-vev}) are
\begin{equation}
\begin{array}{rclcrcl}
\langle W_1(u)\rangle_0 &=& \displaystyle{\frac{N_1}2}, &\hspace{5mm}& \langle Y_1(u)\rangle_0 &=& \displaystyle{\frac{N_1(N_1^2-1)}{12}-\frac n8N_1^2N_2}, \\ [3mm]
\langle W_2(u)\rangle_0 &=& \displaystyle{\frac{2N_1^2+1}{24}}, &\hspace{5mm}& \langle Y_1(u)W_1(u)\rangle_0 &=& \displaystyle{\frac{(N_1^2-1)(N_1^2+2)}{24}-\frac n{16}N_1N_2(N_1^2+2)}, \\ [3mm]
\langle Y_2(w)W_1(u)\rangle_0 &=& \langle Y_2(w)\rangle_0\langle W_1(u)\rangle_0.
\end{array}
\end{equation}
Therefore, the perturbative expansion of the vev $\langle W_1\rangle$ is given as
\begin{equation}
\langle W_1\rangle\ =\ 1+\frac{\pi iN_1}{k_1}+\left( \frac{2\pi i}{k_1} \right)^2\left( \frac{4N_1^2-1}{24}-\frac n8N_1N_2 \right)+O(k^{-3}),
\end{equation}
When $k_1=k, N_1=N_2=N$ and $n=2$, this reproduces the result in \cite{Kapustin:2009kz}.
Since $(N_1,k_1)$ and $(N_2,k_2)$ appear in the partition function (\ref{Z-2node}) symmetrically, the vev $\langle W_2\rangle$ of the Wilson loop for the ${\rm U}(N_2)$ gauge field must be
\begin{equation}
\langle W_2\rangle\ =\ 1+\frac{\pi iN_2}{k_2}+\left( \frac{2\pi i}{k_2} \right)^2\left( \frac{4N_2^2-1}{24}-\frac n8N_1N_2 \right)+O(k^{-3}).
\end{equation}
In terms of the parameters (\ref{parameters-2node}), the vevs can be written as
\begin{eqnarray}
\langle W_1\rangle
&=& 1+\frac{t_1}{2\kappa_1}+\frac16\left( \frac{t_1}{\kappa_1} \right)^2\left( 1-\frac1{4N_1^2} \right)-\frac n8\frac{t_1t_2}{\kappa_1^2}+O(t^3),
\label{expansion-2node1}\\
\langle W_2\rangle
&=& 1+\frac{t_2}{2\kappa_2}+\frac16\left( \frac{t_2}{\kappa_2} \right)^2\left( 1-\frac1{4N_2^2} \right)-\frac n8\frac{t_1t_2}{\kappa_2^2}+O(t^3).
\label{expansion-2node2}
\end{eqnarray}
In the following, we will show that the planar limit of these expansions can be derived from the planar resolvent obtained in section \ref{2-node}.
\vspace{5mm}
\subsection{Expansion from the planar resolvent: pure Chern-Simons theory}
\vspace{5mm}
To illustrate how to derive the perturbative expansion from the planar resolvent, let us start with the calculation for pure Chern-Simons theory.
Recall that the resolvent $v(z)$ satisfies
\begin{equation}
zv'(z)\ =\ 1-\frac{z-1}{\sqrt{(z-a)(z-b)}},
\end{equation}
where $ab=1$ is assumed.
The 't~Hooft coupling $t$ and the vev $\langle W\rangle$ of the Wilson loop are given as
\begin{equation}
t\ =\ \frac12\int_C\frac{dz}{2\pi i}\frac{\log z}{z}\frac{z-1}{\sqrt{(z-a)(z-b)}}, \hspace{1cm} t\langle W\rangle\ =\ \frac{a+b-2}4.
\label{CS-t and W}
\end{equation}
The vev $\langle W\rangle$ depends on the coupling $t$ through the parameter $a$.
In order to derive the power series expansion of $\langle W\rangle$ in $t$, it is necessary to know which limit for $a$ corresponds to the weak coupling limit $t\to0$.
The saddle point equations (\ref{SPE-1}) imply that, for a large $k$ (small $t$), the eigenvalues are expected to be localized around the origin with a narrow width.
This implies that the limit $t\to0$ corresponds to the limit $a\to1$.
Introduce a small parameter
\begin{equation}
\delta\ :=\ -\log a.
\end{equation}
The expansion in $\delta$ will provide us with the perturbative expansion.
The integrand in (\ref{CS-t and W}) has the expansion of the following form:
\begin{equation}
\frac{\log z}{z}\frac{z-1}{\sqrt{(z-a)(z-b)}}\ =\ \sum_{n=0}^\infty f_n(z)\delta^n.
\end{equation}
Since
\begin{equation}
\frac{z-1}{\sqrt{(z-a)(z-b)}}\ =\ \left( 1+\frac{1-e^{-\delta}}{z-1} \right)^{-\frac12}\left( 1+\frac{1-e^{\delta}}{z-1} \right)^{-\frac12},
\end{equation}
the functions $f_n(z)$ have poles at $z=1$ and are holomorphic elsewhere inside $C$.
Therefore, the expansion coefficients are given by the residues of $f_n(z)$ at $z=1$.
Summing up all residues, one obtains
\begin{equation}
t\ =\ \frac14\delta^2-\frac1{96}\delta^4+\frac1{1440}\delta^6+O(\delta^8).
\end{equation}
The inverse of this relation is given as
\begin{equation}
\delta^2\ =\ 4t+\frac23t^2+\frac2{45}t^3+O(t^4).
\end{equation}
This implies that the perturbative expansion is given as
\begin{eqnarray}
\langle W\rangle
&=& \frac1t\left( \frac14\delta^2+\frac1{48}\delta^4+\frac1{1440}\delta^6+O(\delta^8) \right) \nonumber \\
&=& 1+\frac12t+\frac16t^2+O(t^3).
\end{eqnarray}
This reproduces the planar limit of (\ref{expansion-pureCS}).
\vspace{5mm}
\subsection{Expansion from the planar resolvent: 2-node theories with $n=2$}
\vspace{5mm}
The perturbative calculation for Chern-Simons-matter theories with 2-node is almost parallel with that for pure Chern-Simons theory shown in the previous subsection, as long as $n=2$.
Recall that the planar resolvent $\omega(z)$ is
\begin{equation}
\omega(z)\ =\ -\kappa_2\left[ 1-\frac{z^2-1}{s(z)} \right]-\frac{\kappa_1+\kappa_2}2\frac{z^2-1}{s(z)}\int_{C_1}\frac{d\xi}{2\pi i}\frac{s(\xi)}{\xi(\xi-z)(\xi-z^{-1})},
\label{GT-resolvent}
\end{equation}
where $s(z)$ is defined as (\ref{s(z)}).
The weak coupling limit $t_1,t_2\to0$ correspond to the limit $\delta_1,\delta_2\to0$ where
\begin{equation}
\delta_1\ :=\ -\log a_1, \hspace{1cm} \delta_2\ :=\ -\log(-a_2).
\end{equation}
In this limit, the integral in (\ref{GT-resolvent}) can be evaluated as a power series in $\delta_1$ and $\delta_2$ by evaluating residues at $\xi=1$.
Then, the 't~Hooft couplings are given in terms of $\delta_1$ and $\delta_2$ as
\begin{eqnarray}
t_1
&=& \frac{\kappa_1}4\delta_1^2-\frac{\kappa_1}{96}\delta_1^4+\frac{\kappa_2}{32}\delta_1^2\delta_2^2+\frac{\kappa_1}{1440}\delta_1^6+\frac{6\kappa_1-5\kappa_2}{1536}\delta_1^4\delta_2^2-\frac{5\kappa_2}{1536}\delta_1^2\delta_2^4+O(\delta^8), \\ [2mm]
t_2
&=& \frac{\kappa_2}4\delta_2^2-\frac{\kappa_2}{96}\delta_2^4+\frac{\kappa_1}{32}\delta_1^2\delta_2^2+\frac{\kappa_2}{1440}\delta_2^6+\frac{6\kappa_2-5\kappa_1}{1536}\delta_1^2\delta_2^4-\frac{5\kappa_1}{1536}\delta_1^4\delta_2^2+O(\delta^8).
\end{eqnarray}
The inverse of these relations is given as
\begin{eqnarray}
\delta_1^2
&=& \frac4{\kappa_1}t_1+\frac{2}{3\kappa_1^2}t_1^2-\frac{2}{\kappa_1^2}t_1t_2+\frac{2}{45\kappa_1^3}t_1^3-\frac{1}{6\kappa_1^3}t_1^2t_2+\frac{\kappa_1+2\kappa_2}{2\kappa_1^3\kappa_2}t_1t_2^2+O(t^4), \\ [2mm]
\delta_2^2
&=& \frac4{\kappa_2}t_2+\frac{2}{3\kappa_2^2}t_2^2-\frac{2}{\kappa_2^2}t_1t_2+\frac{2}{45\kappa_2^3}t_2^3-\frac{1}{6\kappa_2^3}t_1t_2^2+\frac{2\kappa_1+\kappa_2}{2\kappa_1\kappa_2^3}t_1^2t_2+O(t^4).
\end{eqnarray}
The linear combination of the vevs of the Wilson loops derived from the expansion of $\omega(z)$ is
\begin{eqnarray}
t_1\langle W_1\rangle+t_2\langle W_2\rangle
&=& -\frac{\kappa_2}4\left( e^{-\delta_1}+e^{\delta_1}-e^{-\delta_2}-e^{\delta_2} \right)-\frac{\kappa_1+\kappa_2}4\int_{C_1}\frac{d\xi}{2\pi i}\frac{s(x)}{\xi^2} \nonumber \\
&=& \frac{\kappa_1}{4}{\delta_1}^{2}+\frac{\kappa_2}{4}{\delta_2}^{2}+\frac{\kappa_1}{48}{\delta_1}^{4}+\frac{\kappa_2}{48}{\delta_2}^{4}+\frac{\kappa_1+\kappa_2}{32}{\delta_1}^{2}{\delta_2}^{2}
\nonumber \\
& & +\frac{\kappa_1}{1440}{\delta_1}^{6}+\frac{\kappa_2}{1440}{\delta_2}^{6}+\frac{\kappa_1+\kappa_2}{1536}\delta_1^4\delta_2^2+\frac{\kappa_1+\kappa_2}{1536}\delta_1^2\delta_2^4+O(\delta^8)
\nonumber \\
&=& t_1+t_2+\frac{1}{2\kappa_1}t_1^2+\frac{1}{2\kappa_2}t_2^2+\frac1{6\kappa_1^2}t_1^3-\frac1{4\kappa_1^2}t_1^2t_2-\frac1{4\kappa_2^2}t_1t_2^2
+\frac1{6\kappa_2^2}t_2^3+O(t^4). \nonumber \\
\end{eqnarray}
This reproduces the planar limit of the corresponding linear combination of (\ref{expansion-2node1}) and (\ref{expansion-2node2}) with $n=2$.
\vspace{5mm}
\subsection{Expansion from the planar resolvent: 2-node theories with $n\ne2$}
\vspace{5mm}
The planar resolvent for a 2-node theory with $n\ne2$, given in (\ref{resolvent-general}), is quite complicated.
Indeed, it is given in terms of the theta functions of $u(z)$, and $u(z)$ is given by the inverse of an elliptic function.
Therefore, the method of calculation used so far in this section does not seem to be appropriate for these general cases.
A simplification occurs if the range of the parameters is restricted such that $a_1=-a_2=:a$ holds.
In this case, the quantities $u_0$ and $u_\infty$ can be written simply as
\begin{equation}
u_0\ =\ \frac14\tau, \hspace{1cm} u_\infty\ =\ \frac14\tau-\frac12.
\end{equation}
The modulus $\tau$ can be written explicitly as
\begin{equation}
\tau\ =\ i\frac{2K(a^2)}{K(\sqrt{1-a^4})}.
\end{equation}
Inverting this relation, one obtains
\begin{equation}
1-a^4\ =\ 16q^{\frac12}\left( \frac{\sum_{n=1}^\infty q^{\frac12n(n-1)}}{1+2\sum_{n=1}^\infty q^{\frac12n^2}} \right)^4, \hspace{1cm} q\ :=\ e^{\pi i\tau}.
\end{equation}
As in the previous subsections, we introduce $\delta$ such that $a=\exp(-\delta)$.
Then, this relation implies that $q^\frac12$ can be given as a power series in $\delta$.
Explicitly,
\begin{equation}
q^\frac12\ =\ \frac{1}{4}\delta-\frac{1}{48}\delta^3-\frac{31}{7680}\delta^5+O(\delta^7).
\end{equation}
This implies that the $q^\frac12$-expansion of the resolvent gives the desired perturbative expansion.
It turns out that each coefficient of the $q^\frac12$-expansion is a linear combination of exponential functions of $u$.
Since the 't~Hooft couplings are given as
\begin{eqnarray}
t_1
&=& \varphi(b_1)\int_{-\frac12}^{+\frac12}\frac{du}{2\pi i}\frac{\log z(u)}{z(u)}g(u)G(u),
\label{t-integral} \\
t_2
&=& e^{-\pi i\nu}\varphi(b_1)\int_{-\frac12+\frac12\tau}^{+\frac12+\frac12\tau}\frac{du}{2\pi i}\frac{\log z(u)}{z(u)}g(u)G(u),
\end{eqnarray}
the integration of the coefficients can be performed easily.
To simplify the calculation further, notice that it is enough to perform the perturbative check for $(\kappa_1,\kappa_2)=(1,\pm1)$ since the resolvent for a general $(\kappa_1,\kappa_2)$ is obtained as a linear combination of the resolvents for these two special cases.
Let us focus on the cases $(\kappa_1,\kappa_2)=(1,\epsilon)$ with $\epsilon=\pm1$.
The uniqueness of the solution of the saddle point equations (\ref{res-vec1})(\ref{res-vec2}) implies
\begin{equation}
v_1'(z)\ =\ -\epsilon v_2'(-z).
\end{equation}
This equality then implies
\begin{equation}
t_1\ =\ \epsilon t_2\ =:\ t, \hspace{1cm} \langle W_1\rangle\ =\ \langle W_2\rangle\ =:\ \langle W\rangle.
\end{equation}
The integral formula (\ref{t-integral}) implies
\begin{eqnarray}
t
&=& \frac{1}{4}{\delta}^{2}+\epsilon\frac{3{e}^{\pi i\nu}-2\epsilon+3e^{-\pi i\nu}}{192}{\delta}^{4} \nonumber \\
& & +\frac{45{e}^{2\pi i\nu}-150\epsilon{e}^{\pi i\nu}+122-150\epsilon{e}^{-\pi i\nu}+45e^{-2\pi i\nu}}{46080}{\delta}^{6}+O(\delta^8).
\end{eqnarray}
Inverting this relation, one obtains
\begin{eqnarray}
\delta^2
&=& 4t-\epsilon\frac{3{e}^{\pi i\nu}-2\epsilon+3e^{-\pi i\nu}}{3}{t}^{2} \nonumber \\
& & +\frac{45{e}^{2\pi i\nu}+30\epsilon{e}^{\pi i\nu}+98+30\epsilon{e}^{-\pi i\nu}+45e^{-\pi i\nu}}{180}{t}^{3}+O(t^4).
\end{eqnarray}
The vev $\langle W\rangle$ is given as
\begin{eqnarray}
(1+\epsilon e^{\pi i\nu})t\langle W\rangle
&=& -\frac1{4\varphi(b)}\left[ g'(u_0)G(u_0)+g(u_0)G'(u_0) \right] \nonumber \\
&=& \frac{\epsilon{e}^{\pi i\nu}+1}{4}{\delta}^{2}+\frac{3{e}^{2\pi i\nu}+7\epsilon{e}^{\pi i\nu}+7+3\epsilon e^{-\pi i\nu}}{192}{\delta}^{4} \nonumber \\
& & +\epsilon\frac{45{e}^{3\pi i\nu}-15\epsilon{e}^{2\pi i\nu}+62{e}^{\pi i\nu}+62\epsilon-15{e}^{-\pi i\nu}+45\epsilon e^{-2\pi i\nu}}{46080}{\delta}^{6}+O(\delta^8) \nonumber \\
&=& \left( 1+\epsilon {e}^{\pi i\nu}\right) t+\frac{1+\epsilon{e}^{\pi i\nu}}{2}{t}^{2}-\frac{3{e}^{2\pi i\nu}-\epsilon{e}^{\pi i\nu}-1+3\epsilon e^{-\pi i\nu}}{24}{t}^{3}+O(t^4).
\end{eqnarray}
Therefore,
\begin{eqnarray}
\langle W\rangle
&=& 1+\frac12t+\left[ \frac16-\frac n8\epsilon \right]t^2+O(t^3).
\end{eqnarray}
This reproduces the planar limit of (\ref{expansion-2node1})(\ref{expansion-2node2}).
\vspace{1cm}
\section{Discussion} \label{discuss}
\vspace{5mm}
We have investigated the planar resolvents of a family of Chern-Simons-matter matrix models which are derived from ${\cal N}\ge3$ Chern-Simons-matter theories with the gauge groups of the form ${\rm U}(N_1)_{k_1}\times{\rm U}(N_2)_{k_2}$ via the supersymmetric localization.
We found that, although the resolvents themselves are not obtained in general, their derivatives can be determined explicitly.
From this result, we obtained the explicit formulas for the vevs of the Wilson loops.
We discussed the possible divergent behaviors of the vevs of the Wilson loops using the explicit formulas.
As a check of our result, we performed the perturbative calculations of the vevs of Wilson loops.
The results from the planar resolvents reproduce the results obtained directly from the localization formulas.
It is interesting to extend the analysis of this paper to a more general family of Chern-Simons-matter matrix models.
If the gauge group of a given Chern-Simons-matter theory has $g$ factors of ${\rm U}(N)$ type, the resolvent $zv'(z)$ to be determined is valued in a $g$-dimensional vector space with $g$ branch cuts.
It can be shown that the determination of $zv'(z)$ reduces to a Riemann-Hilbert problem with the monodromy matrices given in terms of the numbers of bi-fundamental hypermultiplets.
It is interesting to clarify whether some physical quantities like the vevs of the Wilson loops can be obtained in a form explicit enough to investigate their analytic properties.
We have found for the cases $n>2$ that there exists a strong 't~Hooft coupling limit in which the vevs of the Wilson loops are finite.
Similar phenomena were also observed in \cite{Suyama:2013fua} for more general theories.
It would be interesting to analyze the behavior of the physical quantities in the strong coupling limits for the cases $n>2$, and investigate the possibility for the existence of a gravity dual (see e.g. \cite{Billo:2000zr}\cite{Billo:2000zs} for a proposal for the case $n=3$).
\vspace{2cm}
{\bf \Large Acknowledgements}
\vspace{5mm}
We would like to thank H. Itoyama and T. Oota for valuable discussions.
This work was supported in part by Fujukai Foundation.
\vspace{1cm}
|
\section{Introduction}
\label{Sec1}
\setcounter{equation}{0}
Precise measurements of the anisotropies and the polarisation of the cosmic microwave background (CMB) have revolutionised cosmology from an order of magnitude science to a precision science. Especially, the latest observations with the Planck satellite~\cite{Adam:2015rua,Planck:2015xua} and also
from experiments on the ground~\cite{Naess:2014wtr,Crites:2014prc} have determined the cosmological parameters that describe the geometry and the matter content of our Universe and the initial conditions of perturbations with percent accuracy and better.
To achieve this accuracy it is very important to include the effect of gravitational lensing of CMB photons by foreground structures. Gravitational lensing has been predicted and calculated to modify the CMB power spectrum by 10\% and more on small angular scales, $\ell\gtrsim 1000$~\cite{1989MNRAS.239..195C,Seljak:1995ve,Zaldarriaga:1998ar,Hu:2000ee,
Lewis:2006fu,RuthBook}. It is therefore imperative to take it into account for precision measurements. In the mean time, lensing of the CMB has not only been detected~\cite{Smith:2007rg,Hirata:2008cb}, but it has been used to reconstruct the lensing potential to the CMB~\cite{Das:2011ak,vanEngelen:2012va,Madhavacheril:2014slf,Story:2014hni} and even a first map of the full sky lensing potential inferred from observations with the Planck satellite has been generated~\cite{Ade:2015zua}. In addition to modifying (damping and widening) the shape of the acoustic peaks in the CMB temperature and polarisation spectra, lensing also rotates E-mode polarisation into B-modes and these lensing B-modes have been detected in the CMB~\cite{Hanson:2013hsb,Ade:2014afa,Keisler:2015hfa,Array:2015xqh,Ade:2015nch}. The possibility to generate high precision maps of the lensing potential to the CMB is intriguing. The lensing potential is a weighted integral of the matter distribution out to the CMB and especially correlating it with different other surveys, for example the infrared background~\cite{Holder:2013hqu,Ade:2013aro,vanEngelen:2014zlh} or galaxy surveys~\cite{Das:2008am,Giannantonio:2015ahz,Kirk:2015dpw,Liu:2016lxy,Baxter:2016ziy}, will give us detailed information about the
matter distribution in the Universe. For this, however we must ensure that also the theoretical calculation of the CMB deflection angle is sufficiently precise and this is the goal of the present paper.
So far, the effect of lensing in the CMB has been determined using the 'Born approximation', i.e., considering the photons to move along the unperturbed geodesics when computing their deflection angle. This is strictly true only to first order and since lensing is such a significant effect, the question whether higher order terms have to be taken into account to obtain sufficient accuracy for present and future experiments is justified. In this paper we study in detail how the CMB temperature power spectrum is affected by lensing including all terms up to third order in the deflection angle.
For numerical attempts to describe post-Born effects see, for instance, \cite{Hamana:2001vz,Hamana:2004ih,Calabrese:2014gla}.
When going to second order in lensing, also the mean distance out to a fixed redshift is modified. The problem of how the mean luminosity or area distance of a 2-sphere of constant redshift is affected by aggregated lensing has been extensively studied in recent literature, considering small and intermediate redshift in~\cite{BenDayan:2012ct,BenDayan:2013gc} and going up to the last scattering surface by~\cite{Clarkson:2014pda,Bonvin:2015uha}.
It has been found, however that the effects of this average perturbation on cosmological parameter estimantions are small.
The remaining question, the one we address in this paper, is whether the effects of
'aggregated lensing' on the CMB temperature fluctuations power spectrum
are substantial and have to be taken into account.
Previous estimations of the effect of lensing at higher order on the weak lensing
power spectra found a negligibly small result~\cite{Krause:2009yr}. However, a recent attempt to evaluate the weak lensing correction of the cosmic microwave background temperature anisotropies at the next to leading order in~\cite{Hagstotz:2014qea} has obtained quite considerable corrections.
The results obtained in~\cite{Hagstotz:2014qea} are however based on the assumption that
for a stochastic deflection $\alpha$ the relation $\langle e^{i \alpha } \rangle=e^{-\langle \alpha^2 \rangle/2}$ holds. This relation is used when considering first order lensing, to go beyond the small deflection angle approximation.
It allows to re-sum all the
corrections coming from the first order deflection angle non-perturbatively.
However, it holds only if $\alpha$ is a Gaussian variable which is
(nearly\footnote{When taking into account a nonlinear matter distribution in principle also the first order deflection angle is no longer Gaussian. In the present work we focus on the pure post-Born corrections and neglect this non-Gaussianity. However, for the sake of completeness, let us underline that, in principle, the effect
of the non-Gaussianity associated to the non-linear matter distribution
can be taken in consideration by expanding $\Phi_W$ as $\Phi_W+\Phi_W^{(2)}+\Phi_W^{(3)}$ in the expressions for the deflection angles. The additional contributions
are of the same parametric order of the pure post-Born corrections. In our analysis we do not perform these corrections but in some plots we replace the liner power spectrum by a fully non-linear Halofit approximation. In this case, the above Gaussian relation to resum the first order deflection angle can in principle no longer be trusted since then the latter is no longer Gaussian.})
true for the first order deflection angle. But once we go beyond linear order in perturbation theory, the deflection angle no longer obeys Gaussian statistics and the above identity can not be used.
Actually, as we shall see, using this property corresponds to neglecting terms that are not negligible beyond linear order.
In practice, terms
of the form $\langle \theta \theta \theta \rangle$, with $\theta$ a deflection angles to any order, cannot be obtained in the above approximation which contains only even powers of deflection angles.
For this reason, in the present paper we choose a different approach. We employ the approximation of small deflection angles. With this
we find here a reliable result which takes into account the Gaussian and non-Gaussian contributions in the lensed CMB anisotropies, going beyond the correction from first order deflection angles.
Starting from the result for the deflection angle up to third order
obtained in~\cite{Fanizza:2015swa},
we investigate whether the next to leading order correction coming from foreground structures has to be
taken in account for a precise calculation of the observed, lensed CMB temperature anisotropy power spectrum.
The results of~\cite{Fanizza:2015swa} are based on the use of the so-called geodesic light-cone gauge~\cite{Gasperini:2011us}. This gauge is especially adapted to the calculation of physical observables and has already previously been applied to this goal, see~\cite{BenDayan:2012wi,Fanizza:2013doa,Marozzi:2014kua,DiDio:2014lka}.
The paper is organized as follow. In Sect.~\ref{Sec2} we present the small deflection angle approximation for CMB lensing beyond linear order, and give the expressions for the deflection angle and the amplification matrix up to
third perturbative order.
In Sect.~\ref{Sec3} we translate these results into harmonic space, '$\vl{}$--space' and we derive the deflection angle to third order in $\vl{}$--space.
In Sect.~\ref{Sec4} we evaluate the lensed power spectrum of the temperature anisotropies at higher order
and beyond the Born approximation.
In Sect.~\ref{Sec5} we introduce the Limber approximation and apply it to some of the results of the previous section.
Our main results are presented in Sect.~\ref{Sec6}. In particular, we give the Limber approximated expression for the post-Born corrections to the lensed power spectrum of the CMB temperature anisotropies, evaluating numerically the different contribution both considering a linear and a non-linear power spectrum.
In Sect.~\ref{Sec7} we conclude and add a note regarding a recent paper~\cite{Pratten:2016dsm} that appeared while we
were finalizing our manuscript.
Some lengthy expressions needed in our calculation in $\vl{}$--space are presented in Appendix~\ref{a:Al}.
\section{Weak Lensing Corrections beyond leading order}
\label{Sec2}
\setcounter{equation}{0}
We want to determine the effect of lensing on the CMB anisotropies beyond the leading order which involves only first order perturbation theory~\cite{Lewis:2006fu,RuthBook}.
Let us consider a generic scalar field $\mathcal M$, that we will identify with the CMB temperature anisotropies, and let us denotes
lensed quantities by a tilde ($\tilde{\ }$).
We can generalize the result of~\cite{Lewis:2006fu,RuthBook} and write
the following relation between the lensed and unlensed temperature fluctuations $\mathcal M$ valid up to fourth order in the deflection angles $\T{a}{i}$ (the superscript $i$ denotes the order).
\begin{eqnarray}
\label{eq:expansion}
\tilde\mathcal M(x^a)&=&\mathcal M\left(x^a+\delta \theta^{a}\right)\simeq\mathcal M(x^a)+\sum_{i=1}^4\T{b}{i}\nabla_b\mathcal M(x^a)+\frac{1}{2}\sum_{i+j\le 4}\T{b}{i}\T{c}{j}\nabla_{b}\nabla_c\mathcal M(x^a)\nonumber\\
&+&\frac{1}{6}\sum_{i+j+k\le 4}\T{b}{i}\T{c}{j}\T{d}{k}\nabla_b\nabla_c\nabla_d\mathcal M(x^a) +\frac{1}{24} \T{b}{1}\T{c}{1}\T{d}{1}\T{e}{1}\nabla_b\nabla_c\nabla_d\nabla_e\mathcal M(x^a)\,.\nonumber
\\
\label{IniExp}
\end{eqnarray}
This can be written in a more compact form as follows
\begin{eqnarray}
\tilde\mathcal M(x^a)&\simeq&
\mathcal A^{(0)}(x^a)+\sum_{i=1}^4\mathcal A^{(i)}(x^a)
+\sum_{i+j\le 4,\,1\le i\le j}\mathcal A^{(ij)}(x^a)
+\sum_{i+j+k\le 4,\,1\le i\le j\le k} \mathcal A^{(ijk)}(x^a)
\nonumber \\
& &+ \mathcal A^{(1111)}(x^a)
\,,
\label{eq:expansion-compact}
\end{eqnarray}
where
\begin{equation}
\mathcal A^{(i_1 i_2....i_n)}(x^a)=\frac{{\rm Perm}(i_1 i_2....i_n)}{n!}\T{b}{i_1}\T{c}{i_2}.....\nabla_{b}\nabla_c.......\mathcal M(x^a)
\,,
\end{equation}
where ${\rm Perm}(i_1 i_2....i_n)$ gives the number of permutation of the set $(i_1 i_2....i_n)$,
and \\ $\mathcal A^{(0)}(x^a)\equiv \mathcal M(x^a)$.
We introduce the Weyl potential $\Phi_W$ (in terms of the Bardeen potentials $\Phi$ and $\Psi$) as
\begin{eqnarray}
\Phi_W &=& \frac{1}{2}\left(\Phi+\Psi\right)\,.
\end{eqnarray}
The lensing potential $\psi$ to the last scattering surface
is then given by
\begin{eqnarray}
\psi({\bf n},z_s) &=& -\frac{2}{\eta_o-\eta_s} \int_{\eta_s}^{\eta_o} d\eta \frac{\eta-\eta_s}{\eta_o-\eta} \Phi_W(-(\eta_o-\eta){\bf n},\eta)
=-2\int_0^{r_s} dr' \frac{r_s-r'}{r_s r'}\Phi_W(-r'{\bf n},\eta_o-r') \,, \nonumber \\
\label{Lenspot}
\label{Weylpot}
\end{eqnarray}
with $\bf{n}$ the direction of propagation of photon, $\eta$ conformal time and $r$ the comoving distance, $r=\eta_o-\eta$. The index $_s$ denotes the corresponding quantity at the last scattering surface while $\eta_o$ denotes present time.
The first order deflection angle is simply the gradient of the lensing potential~\cite{Bartelmann:1999yn,RuthBook}. Taking into account also the lensing of the direction $\bf n$ on the path of the photon one
obtains the following expressions for the deflection angles up to third perturbative order~\cite{Fanizza:2015swa}
\begin{eqnarray}
\T{a}{1}\!&=&\!
-2\int_{0}^{r_s}\!dr' \frac{r_s-r'}{r_s\,r'}\nabla^a\Phi_W(r') \,,
\label{TheOrd1}
\\
\T{a}{2}\!&=&\!
-2\int_{0}^{r_s}\!dr'\frac{r_s-r'}{r_s\,r'}
\nabla_b\nabla^a\Phi_W(r')\T{b}{1}(r')\,,
\label{TheOrd2}
\\
\T{a}{3}\!&=&\!
-2\int_{0}^{r_s}\!dr'\frac{r_s-r'}{r_s\,r'}
\!\left[\nabla_b\nabla^a\Phi_W(r')\T{b}{2}(r')\!+\!\frac{1}{2}\nabla_b\nabla_c\nabla^a\Phi_W(r')\T{b}{1}(r')\T{c}{1}(r') \right]\!.
\label{TheOrd3}
\end{eqnarray}
Latin letters $a,b,c,d$ go over the two directions on the sphere.
In Eqs.~\eqref{TheOrd1} to \eqref{TheOrd3} we take into account the terms with the maximal number of transverse derivatives. We do not expand to higher order the Weyl potential, $\Psi_W$, instead we shall later consider both the linear potential and the fully non-linear one obtained by using Halofit model (see~\cite{Smith:2002dz,Takahashi:2012em}) for the non-linear fluctuations.
Let us also stress that the Taylor expansion in Eq.~(\ref{IniExp}) is valid in the approximation of small deflection angles, i.e. when the deflection angle is much smaller than the angular separations which contribute mainly to $C_\ell$. This is certainly true for $\ell \lesssim 2500$, which corresponds to an angular scale of about $4.5$ arc minutes (see \cite{Seljak:1995ve,RuthBook,Lewis:2006fu}). We shall use the small deflection angle approximation only for the second and third order deflection angles which are much smaller than this value.
To better compare the results that we will obtain in the following, and to understand discrepancy with the ones obtained in \cite{Hagstotz:2014qea}, let us define also the amplification matrix $\mathcal A^a_b$ following \cite{Fanizza:2015swa}. It is defined as the derivative of the angular coordinates of the source with respect to the angular direction of the light ray received at the observer position, namely \cite{Bartelmann:1999yn}:
\begin{equation}
\left({ \cal A}^a_b \right)= \left( \frac{\partial \theta_s^a}{\partial \theta_o^b} \right)=
\left(
\begin{array}{cc}
1 -\kappa - \gamma_1 & -\gamma_2 - \omega \\
- \gamma_2 + \omega & 1- \kappa+ \gamma_1
\end{array}
\right) \, .
\label{AmpMat}
\end{equation}
Here $\kappa$ is the convergence, $\gamma_1$ and $\gamma_2$ the two shear components and $\omega$ is the vorticity. As one sees by inspecting Eqs.~(\ref{TheOrd1}) to (\ref{TheOrd3}), the deflection angle $\delta \boldsymbol{\theta }$ can be written as the gradient of a scalar field only to first order. In full generality, introducing a general lensing potential $\varPsi$ and a curl potential $\Omega$,
we can write the deflection angles $\delta \boldsymbol{\theta}$ (to all order) as
\begin{equation}
\delta \boldsymbol{\theta }= \boldsymbol{ \nabla} \varPsi + \boldsymbol{\nabla } \times \Omega \, .
\end{equation}
It is then straightforward to show that
\begin{equation}
\kappa =- \frac{1}{2}\boldsymbol{ \nabla} \delta \boldsymbol{\theta } = -\frac{1}{2} \Delta \varPsi \qquad, \qquad \omega =- \frac{1}{2} \Delta \Omega \,
\end{equation}
and
\begin{equation}
C_\ell^{\varPsi \varPsi}=\frac{4}{\ell^4} C_\ell^{\kappa \kappa}\qquad, \qquad
C_\ell^{\Omega \Omega}=\frac{4}{\ell^4} C_\ell^{\omega \omega}\,.
\label{SpectraHO}
\end{equation}
Therefore, beyond the Born approximation, the CMB lensed power spectrum will not depend only on a lensing potential, but it has also a curl contribution, see \cite{Hagstotz:2014qea}.
In our approach, by computing directly the deflection angle $\delta \boldsymbol{\theta }$, we do not split the contribution coming from the lensing potential $\varPsi$ and from the curl potential $\Omega$, but we do include them both by construction.
We can then expand in perturbation theory and define the deformation part of the amplification matrix by subtracting the zeroth order contribution, i.e. by introducing the convenient quantity $\Psi^a_b=\delta^a_b - {\cal A}^a_b$, which is given by \cite{Fanizza:2015swa}
\begin{equation}
(\Psi^a_b)^{(n)}=-\frac{\partial \theta_s^{a (n)}}{\partial \theta_o^b},
~~~~~~~~~~~~ n \geq 1.
\label{GAMordern}
\end{equation}
Using our expressions for the deflection angle we find the following results for the deformation matrix up to third order (these were first given in
\cite{Fanizza:2015swa})
\begin{eqnarray}
(\Psi^a_b)^{(1)}&=&2\int_{0}^{r_s} d r' \frac{r_s-r'}{r_s\,r'} \,{\gamma}^{ac}\partial_c
\partial_b \psi(\eta',r', \theta^a_o) \,,
\\
(\Psi^a_b)^{(2)}&=&2\int_{0}^{r_s} d r' \frac{r_s-r'}{r_s\,r'} \,{\gamma}^{ac}
\left[\partial_c
\partial_b \partial_d \psi(r') \theta^{d (1)}- \partial_c \partial_d \psi(r') \Psi^{d (1)}_b
\right] \,,
\\
(\Psi^a_b)^{(3)}&=&2\int_{0}^{r_s} d r' \frac{r_s-r'}{r_s\,r'} \,{\gamma}^{ac}
\left[\partial_c
\partial_b \partial_d \psi(r') \theta^{d (2)}+\frac{1}{2} \partial_c \partial_b \partial_d \partial_e \psi(r') \theta^{d (1)}
\theta^{e (1)} \right.\nonumber \\
& & \left.
-\partial_c \partial_d \partial_e \psi(r') \theta^{e (1)} \Psi^{d (1)}_b
- \partial_c \partial_d \psi(r') \Psi^{d (2)}_b
\right] \,,
\end{eqnarray}
where ${\gamma}_{{ac}}$ denotes the standard metric on the two-sphere and ${\gamma}^{ac}$ is its inverse. The results above are consistent with the ones obtained in~\cite{Krause:2009yr} for the amplification matrix, while they differ from the expressions
in~\cite{Hagstotz:2014qea}.
In particular, as already underlined in~\cite{Fanizza:2015swa}, the deformed part of the amplification matrix used
in~\cite{Hagstotz:2014qea} cannot be obtained as a derivative of a deflection angle order by order, as it should be according to Eq.~(\ref{GAMordern}).
In particular, in~\cite{Hagstotz:2014qea} there seems to be a sign error in the second order expression
$(\Psi^a_b)^{(2)}$, which could be the explanation of an
overestimation of the post-Born corrections to the convergence power spectrum (due to such an error the subtle
cancellation between the terms $\langle \kappa^{(2)} \kappa^{(2)}\rangle$ and
$\langle \kappa^{(1)} \kappa^{(3)}\rangle$, which has already been obtained in~\cite{Krause:2009yr} and which also we have recovered in our framework, might be spoiled).
This could be the cause of the overestimation of the post-Born correction of the lensed power spectrum of CMB temperature anisotropies by the authors of \cite{Hagstotz:2014qea}.
As we show in the following, the post-Born corrections to
lensed power spectrum of CMB temperature anisotropies from
higher order contributions to the convergence (and vorticity) power spectrum are nearly two order of magnitude smaller than the result obtained in \cite{Hagstotz:2014qea}.
\section{Expansion in $\vl{}$-space}
\label{Sec3}
To evaluate the lensing correction to the angular power spectrum $C^{\mathcal M}_\ell$ of the CMB temperature anisotropies we consider the flat sky
limit. In this approximation, see e.g.~\cite{RuthBook}, the combination $(\ell,m)$ is replaced by a two dimensional vector $\vl{}$.
The angular position is then the 2-dimensional Fourier transform of the position in $\vl{}$ space at redshift $z$, for a generic variable $Y(z,{\bf x})$ we have
\begin{equation}
Y(z,{\bf x})=\frac{1}{2\pi}\INT{}Y(z,\vl{})e^{-i\vl{}\cdot{\bf x}}\,,
\label{Deflspace}
\end{equation}
and
\begin{eqnarray}
\langle Y(z_1,\vl{})\bar{Y}(z_2,\vl{}\,') \rangle&=&\delta\left( \vl{}-\vl{}\,' \right)C_\ell^{Y}(z_1,z_2)\,,
\end{eqnarray}
where an overline denotes complex conjugation.
We denote
$$ C_\ell^{Y}(z,z)\equiv C_\ell^{Y}(z)\,.$$
To determine the angular power spectra defined above we follow~\cite{DiDio:2013bqa, DiDio:2014lka} and introduce the (3 dimensional) initial curvature power spectrum by
\begin{equation}
\langle R_{\rm in} \left( {\bf k} \right) \bar R_{\rm in} \left( {\bf k}' \right) \rangle = \delta_D \left( {\bf k} - {\bf k}' \right) P_R \left(k \right)\,.
\end{equation}
(In both 2- and 3-dimensional Fourier transforms we use the unitary Fourier transform normalisation, hence there are no factors of $2\pi$ in this formula.)
For a given linear perturbation variable $A$ we define its transfer function $T_A(z,k)$ normalized to the initial curvature perturbation by
\begin{equation}
A \left( z, {\bf k} \right) =T_A(z,k)R_{\rm in}({\bf k}) \,.
\end{equation}
An angular power spectrum will be then given by
\begin{equation}\label{e:cAB}
C^{AB}_\ell \left( z_1, z_2 \right) = 4 \pi \int \frac{dk}{k} \mathcal{P}_R (k) \Delta^A_\ell (z_1,k)\Delta^B_\ell (z_2,k) = \frac{2}{\pi} \int dk k^2 P_R (k) \Delta^A_\ell (z_1,k)\Delta^B_\ell (z_2,k) \,,
\end{equation}
where $\mathcal{P}_R (k) = \frac{k^3}{2 \pi^2} P_R (k)$ is the dimensionless primordial power spectrum, and
$\Delta^A_\ell \left( z, k \right)$ denotes the transfer function in angular and redshift space for the variable $A$.
For example, by considering $A=B=\Phi_W$ and $A=B=\psi$ we obtain that (setting $C_\ell^{\Psi_W}(z,z')\equiv C_\ell^W(z,z')$)
\begin{eqnarray}
\!\!\!\!\!\!\!\!\!\!C_\ell^W(z,z')&=&\frac{1}{2 \pi}\int dk\,k^2\,P_R(k)
\left[ T_{\Psi+\Phi}(k,z)j_\ell\left((k r\right) \right]
\left[ T_{\Psi+\Phi}(k,z')j_\ell\left(k r'\right) \right]\,,
\\
C_\ell^\psi(z,z')&=&\frac{2}{\pi}\int dk\,k^2\,P_R(k)
\left[ \int_{0}^{r}d r_1\frac{r-r_1}{r r_1}
T_{\Psi+\Phi}(k,z_1)j_\ell\left( k r_1 \right) \right] \nonumber \\
& & \times \left[\int_{0}^{r'}d r_2 \frac{ r' -r_2 }{r' r_2}
T_{\Psi+\Phi}(k,z_2)j_\ell\left( k r_2 \right) \right] \,,
\label{Cpsi}
\end{eqnarray}
where $j_\ell$ denotes a spherical Bessel function of order $\ell$, $r= \eta_o -\eta$ is the comoving distance and analogously for $r', r_1,r_2$. Above and hereafter, we define $z=z(r)$, $z'=z(r')$, etc..
Starting from the definitions in Eq.~(\ref{Deflspace}), we can write Eq.~(\ref{IniExp}) in $\vl{}$ space in the form
\begin{eqnarray}
\tilde\mathcal M(z_s,\vl{})&\simeq&
\mathcal A^{(0)}(\vl{})+\sum_{i=1}^4\mathcal A^{(i)}(\vl{})
+\sum_{i+j\le 4,\, 1\le i\le j}\mathcal A^{(ij)}(\vl{})
+\sum_{i+j+k\le 4,\, 1\le i\le j\le k} \mathcal A^{(ijk)}(\vl{})
\nonumber
\\
& &
+\mathcal A^{(1111)}(\vl{})\,,
\label{eq:expansion-compact}
\end{eqnarray}
where, on the right hand side, we drop the redshift dependence for simplicity.
In terms of the Fourier transform of the Weyl potential, the deflections angle up to order three and out to redshift $z_s$ can be written as
\begin{eqnarray}
\T{a}{1}({\bf x})
&=&\frac{i}{\pi}
\INT{}\int_{0}^{r_s}dr' \frac{r_s-r'}{r_s\,r'}\,\ell^a
\Phi_W(r',\vl{})e^{-i\vl{}\cdot {\bf x}}
\\
\T{a}{2}({\bf x})
&=&\frac{i}{\pi^2}\INT{1}\INT{2}\int_{0}^{r_s}dr'\frac{r_s-r'}{r_s\,r'}
\left(\,\ell_1^a\ell_{1b}\Phi_W(r',\vl{1})e^{-i\vl{1}\cdot {\bf x}}\right)\nonumber\\
&&\times\int_{0}^{r'}dr'' \frac{r'-r''}{r'\,r''}\,\ell_2^b
\Phi_W(r'',\vl{2})e^{-i\vl{2}\cdot {\bf x}}
\end{eqnarray}
\begin{eqnarray}
\T{a}{3}({\bf x})
&=&\frac{i}{\pi^3}\int_{0}^{r_s}dr'\frac{r_s-r'}{r_s\,r'}\INT{1}\INT{2}\INT{3}\nonumber\\
&&\times\left[ \ell^a_1\ell_{1c}\ell_2^c\ell_{2d}\ell_3^d\Phi_W(r',\vl{1})\int_{0}^{r'}dr''\frac{r'-r''}{r'\,r''}
\Phi_W(r'',\vl{2})\right.\nonumber\\
&&\left.\times
\int_{0}^{r''}dr''' \frac{r''-r'''}{r''\,r'''}\,
\Phi_W(r''',\vl{3})\right. \nonumber\\
&&+\frac{1}{2}\ell^a_1\ell_{1c}\ell_{1d}\ell_2^c\ell_3^d\Phi_W(z',\vl{1})
\int_{0}^{r'}dr'' \frac{r'-r''}{r'\,r''}\,
\Phi_W(r'',\vl{2})\nonumber\\
&&\left.\times\int_{0}^{r'}dr''' \frac{r'-r'''}{r'\,r'''}\,
\Phi_W(r''',\vl{3}) \right]e^{-i\left( \vl{1}+\vl{2}+\vl{3}\right)\cdot {\bf x}}
\end{eqnarray}
Using these results one easily
obtains the $\vl{}$ space expressions for the terms $\mathcal A^{(i....)}$. We present all of them in Appendix~\ref{a:Al}.
\section{Lensed angular power spectrum: analytical results}
\label{Sec4}
\setcounter{equation}{0}
Let us now compute the lensed angular power spectrum $\tilde{C}^{\mathcal M}_\ell$ using the results of the previous section and of Appendix \ref{a:Al}.
First of all, one can easily see that
\begin{eqnarray}
\langle \tilde{\mathcal M}(\vl{})\bar{\tilde{\mathcal M}}(\vl{}\,')\rangle =\langle \mathcal A (\vl{}) \bar{\mathcal A}(\vl{}\,')\rangle \,,
\end{eqnarray}
where we set
\begin{equation}
\mathcal A(\vl{})= \mathcal A^{(0)}(\vl{})+
\sum_{i=1}^4\mathcal A^{(i)}(\vl{})
+\sum_{i+j\le 4,\, 1\le i\le j}\mathcal A^{(ij)}(\vl{})
+\sum_{i+j+k\le 4,\, 1\le i\le j\le k} \mathcal A^{(ijk)}(\vl{})+\mathcal A^{(1111)}(\vl{})\,.
\end{equation}
We now introduce $C_\ell^{(i\ldots ,\, j\ldots)}$ defined as follows
\begin{eqnarray}
\delta\left( \vl{}-\vl{}\,' \right) C_\ell^{(ij\ldots,ij\ldots)} &=&\langle \mathcal A^{(ij\ldots)} (\vl{})
\bar{\mathcal A}^{(ij\ldots)}(\vl{}\,')\rangle\,,
\nonumber \\
\delta\left( \vl{}-\vl{}\,' \right) C_\ell^{(ij\ldots,i'j'\ldots)} &=&
\langle \mathcal A^{(ij\ldots)} (\vl{})
\bar{\mathcal A}^{(i'j'\ldots)}(\vl{}\,')\rangle+\langle \mathcal A^{(i'j'\ldots)} (\vl{})
\bar{\mathcal A}^{(ij\ldots)}(\vl{}\,')\rangle\,,
\label{HigherOrderCll}
\end{eqnarray}
where the last definition holds when the coefficients $(ij\ldots)$ and $(i'j'\ldots)$ are different. The factor $\delta\left( \vl{}-\vl{}\,' \right)$ is a consequence of statistical isotropy.
Omitting terms of higher than fourth order in the Weyl potential and
terms that vanish as a consequence of Wick's theorem (odd number of Weyl potentials), we obtain
\begin{eqnarray}
\tilde{C}^{\mathcal M}_\ell&=& C^{{\mathcal M}}_\ell+C_\ell^{(0,11)}+C_\ell^{(1, 1)}+C_\ell^{(0,13)}+C_\ell^{(0,22)}+C_\ell^{(0,112)}+C_\ell^{(0,1111)}
\nonumber \\
& &+C_\ell^{(1, 3)}+
C_\ell^{(2, 2)}+C_\ell^{(1, 12)}+C_\ell^{(1, 111)}+C_\ell^{(2, 11)}+C_\ell^{(11, 11)}
\end{eqnarray}
where $C_\ell^{(0,0)}\equiv C^{{\mathcal M}}_\ell$ is the unlensed power spectrum and $C_\ell^{(0,11)}$ and $C_\ell^{(1, 1)}$ are the well-known leading order corrections given by
\begin{eqnarray}
C_\ell^{(0,11)} &=&-C_{\ell}^\mathcal M(z_s)\NINT{1}\,\left(\vl{1}\cdot \vl{}\right)^2\,
C_{\ell_1}^\psi(z_s,z_s)
\,,
\label{eq:firstorder1}
\\
C_\ell^{(1, 1)}&=&\NINT{1}\left[\left(\vl{}-\vl{1}\right)\cdot\vl{1}\right]^2\,C_{|\vl{}-\vl{1}|}^\psi(z_s,z_s)C_{\ell_1}^\mathcal M(z_s)
\,.
\label{eq:firstorder2}
\end{eqnarray}
The next-to-leading order corrections can be evaluated using the expressions for the deflection angle up to third order given in the previous section and inserting them in the definition of the $\mathcal A^{(i\cdots)}$'s
(see Appendix \ref{a:Al}). The expressions for these somewhat lengthy higher order corrections are also given in Appendix \ref{a:Al}.
\section{Lensed angular power spectrum: Limber approximation}
\label{Sec5}
\setcounter{equation}{0}
In order to numerically evaluate the new higher order lensing contributions to the CMB temperature anisotropies, calculated in the previous section,
we apply the Limber approximation~\cite{Limber:1954zz,Kaiser:1991qi,LoVerde:2008re}.
The Limber approximation is valid at large $\ell$ and,
because lensing is appreciable only for $\ell>100$, this is an excellent approximation for this paper.
Following~\cite{Bernardeau:2011tc}, the Limber approximation can be written as
\begin{equation}
\frac{2}{\pi}\int dk\,k^2\,f(k)
j_\ell\left(k x_1\right) j_\ell\left(k x_2\right)
\simeq \frac{\delta_D(x_1-x_2)}{x_1^2} f\left(\frac{\ell+1/2}{x_1}\right)\,,
\end{equation}
where $f(k)$ should be a smooth, not strongly oscillating function of $k$ which decreases sufficiently rapidly for $k\rightarrow\infty$. (More precisely, $f(k)$ has to decrease faster than $1/k$ for $k>\ell/x$.)
Using this approximation, one obtains Limber-approximated $C_\ell$ as follows, see also~\cite{DiDio:2015bua}
\begin{eqnarray}
C_\ell^W(z,z')&=&
\frac{1}{2 \pi}\int dk\,k^2\,P_R(k)
\left[ T_{\Psi+\Phi}(k,z)j_\ell\left( kr \right) \right]
\left[ T_{\Psi+\Phi}(k,z')j_\ell\left(k r'\right) \right]\nonumber\\
&=&\frac{\delta\left( r'-r \right)}{r^2}\,
\frac{1}{4}P_R\left(\frac{\ell+1/2}{r}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell+1/2}{r},z\right)\right]^2 \,,
\label{CWLimber}
\end{eqnarray}
and, from Eq.~(\ref{Cpsi}), we find for the power spectrum of the lensing potential
\begin{eqnarray}
C_\ell^\psi(z,z')&=&4
\int_{0}^{r}dr_1\frac{r-r_1}{r_1}
\int_{0}^{r'}dr_2\frac{r'-r_2}{r_2}\,
C_\ell^W(z_1,z_2) \nonumber\\
&=&\Theta\left( r'-r \right)\int_{0}^{r}dr_1\frac{\left(r-r_1\right)\left(r'-r_1\right)}{r\,r'\,
r_1^4}
P_R\left(\frac{\ell+1/2}{r_1}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell+1/2}{r_1},z_1\right)\right]^2\nonumber\\
&&+\,\Theta\left( r-r' \right)\int_{0}^{r'}dr_1\frac{\left(r-r_1\right)\left(r'-r_1\right)}{r\,r'\,r_1^4}
P_R\left(\frac{\ell+1/2}{r_1}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell+1/2}{r_1},z_1\right)\right]^2\,, \nonumber \\
\label{CpsiLimber}
\end{eqnarray}
where $\Theta$ is the Heaviside step function.
When $z=z'$ the result~(\ref{CpsiLimber}) simplifies to
\begin{eqnarray}
C_\ell^\psi(z,z)&=&\int_{0}^{r}dr_1\frac{\left(r-r_1\right)^2}{r^2\,r_1^4}
P_R\left(\frac{\ell+1/2}{r_1}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell+1/2}{r_1},z_1\right)\right]^2\,.
\label{CpsiLimberSameTime}
\end{eqnarray}
This is in agreement with the corresponding results of Ref.~\cite{Bernardeau:2011tc}.
In~\cite{Bernardeau:2011tc} is also shown that the Limber-approximated $C_\ell^\psi$'s at equal redshift are a very good approximation
already for $\ell>20$. We assume here that this is still true when we consider $C_\ell^\psi$ at two different redshifts.
Using the Limber-approximated $C_\ell$'s given in Eqs.~(\ref{CWLimber}) and~(\ref{CpsiLimber}) in the analytical results of the previous section,
the expressions simplify considerably.
As an example, let us study $C_\ell^{(13)}$ and $C_\ell^{(22)}$.
Inserting Eq.~\eqref{CWLimber} in Eq.~\eqref{C13}, we find
\begin{eqnarray}
C_\ell^{(0,13)}(z_s)&=&
-2\,C_{\ell}^\mathcal M(z_s)
\NINT{1}\NINT{2}\left(\vl{1}\cdot\vl{}\right)\,
\left(\vl{2}\cdot\vl{}\right)\,
\ell_2^2\,
\left(\vl{2}\cdot\vl{1}\right)\nonumber\\
&&\times
\int_{0}^{r_s}dr'\frac{r_s-r'}{r_s\,r'}
\int_{0}^{r'}dr''\frac{r'-r''}{r'\,r''}\,
C_{\ell_1}^\psi(z_s,z'')\nonumber\\
&&\times
\frac{\delta\left( r'-r'' \right)}{r'^2}\,
P_R\left(\frac{\ell_2+1/2}{r'}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell_2+1/2}{r'},z'\right)\right]^2
\nonumber
\\
& &
+2\,C_{\ell}^\mathcal M(z_s)
\NINT{1}\NINT{2}
\left(\vl{2}\cdot\vl{}\right)\,
\ell_2^2\,
\left(\vl{2}\cdot\vl{1}\right)
\left(\vl{1}\cdot\vl{}\right)
\nonumber\\
&&\times
\int_{0}^{r_s}dr'\frac{r_s-r'}{r_s\,r'}
C_{\ell_1}^\psi(z_s,z')\,
\int_{0}^{r'}dr'' \frac{r'-r''}{r'\,r''}\nonumber\\
&&\times
\frac{\delta\left( r'-r'' \right)}{r'^2}\,
P_R\left(\frac{\ell_2+1/2}{r'}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell_2+1/2}{r'},z'\right)\right]^2
\nonumber
\\
& &
+C_{\ell}^\mathcal M(z_s)
\NINT{1}\NINT{2}
\left(\vl{1}\cdot\vl{}\right)^2\,
\left(\vl{1}\cdot\vl{2}\right)^2
\nonumber
\\
&&\times
\int_{0}^{r_s}dr' \frac{r_s-r'}{r_s\,r'}\,
\int_{0}^{r_s}dr''\frac{r_s-r''}{r_s\,r''}
C_{\ell_2}^\psi(z'',z'')\nonumber\\
&&\times
\frac{\delta\left( r'-r'' \right)}{r'^2}\,
P_R\left(\frac{\ell_1+1/2}{r'}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell_1+1/2}{r'},z'\right)\right]^2\,,
\label{c13LimberBS}
\end{eqnarray}
which simplifies to
\begin{eqnarray}
C_\ell^{(0,13)}(z_s)
&=&C_{\ell}^\mathcal M(z_s)
\NINT{1}\NINT{2}
\left(\vl{1}\cdot\vl{}\right)^2\,
\left(\vl{1}\cdot\vl{2}\right)^2
\nonumber
\\
&&\times
\int_{0}^{r_s}dr' \frac{\left(r_s-r'\right)^4}{r_s^2\,r'^4}
C_{\ell_2}^\psi(z',z')
P_R\left(\frac{\ell_1+1/2}{r'}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell_1+1/2}{r'},z'\right)\right]^2.
\label{c13Limber}
\end{eqnarray}
The first two terms in Eq.~(\ref{c13LimberBS}) vanish due to the $\delta(r'-r'')$ acting on the kernel $\frac{r'-r''}{r'\,r''}$.
Similarly, using Eq.~\eqref{CWLimber} in Eq.~\eqref{C22}
we obtain
\begin{eqnarray}
C_\ell^{(0,22)}(z_s)
&=&-
C_{\ell}^\mathcal M(z_s)
\NINT{1}\NINT{2}\left(\vl{}\cdot\vl{1}\right)^2\,
\left(\vl{1}\cdot\vl{2}\right)^2
\nonumber
\\
&&\times
\int_{0}^{r_s}dr'\frac{\left(r_s-r'\right)^2}{r_s^2\,r'^4}
C_{\ell_2}^\psi(z',z')
P_R\left(\frac{\ell_1+1/2}{r'}\right)
\left[T_{\Psi+\Phi}\left(\frac{\ell_1+1/2}{r'},z'\right)\right]^2.
\end{eqnarray}
It is easy to check that $C_\ell^{(0,22)}\equiv -C_\ell^{(0,13)}$ within the Limber approximation.
Similar cancellations occur when using
Eq.~\eqref{CWLimber} to evaluate the terms
$C_\ell^{(0,1111)}$, $C_\ell^{(1, 3)}$, $C_\ell^{(2, 2)}$, $C_\ell^{(1, 111)}$, and $C_\ell^{(11, 11)}$. These cancellations are most probably the result of a consistency relation: large angle inhomogeneities cannot affect the spectrum on small angles as they just correspond to an isotropic Universe at a slightly different temperature. Some of the cancellations have also been found in~\cite{Pratten:2016dsm}. For the same reason, also the terms in $C_\ell^{(0,112)}$ cancel perfectly within the Limber approximation.
Furthermore, because of the Limber approximation, we can always use
Eqs.~(\ref{CpsiLimber}) and~(\ref{CpsiLimberSameTime}) when evaluating the power spectrum of the lensing potential.
All the remaining non-vanishing $C_\ell^{(0,ijk...)}$ and $C_\ell^{(ij,i'j')}$ are presented in the next section.
\section{Lensed angular power spectrum: final results}
\label{Sec6}
\setcounter{equation}{0}
In order to interpret properly corrections from different terms, let us classify the higher order non-null
$C_\ell^{(...)}$
in three different groups.
\subsection{First group}
\label{num-res-s1}
In this first group we collect the next-to-leading order contributions only contain first order deflection angles $\theta^{a(1)}$.
These are given by
\begin{eqnarray}
C_\ell^{(0,1111)}\!&=&
\frac{1}{4}\frac{\left( C_\ell^{(0,11)} \right)^2}{C_\ell^\mathcal M}\,,
\label{GroupOne1}\\
&&
\text{coming from } \langle \th{1}{a}\th{1}{b}\th{1}{c}\th{1}{d} \rangle
\langle \nabla_a\nabla_b\nabla_c\nabla_d\mathcal M \bar\mathcal M \rangle \,,\nonumber\\
C_\ell^{(1, 111)}\!&=&\!
-\!\!\!\NINT{1}\!\!\NINT{2}\!\left[\left(\vl{}-\vl{1}\right)\cdot\vl{1}\right]^2\left(\vl{2}\cdot\vl{1}\right)^2\,
\!C_{\ell_1}^\mathcal M(z_s)
C_{|\vl{}-\vl{1}|}^\psi(z_s,z_s)C_{\ell_2}^\psi(z_s,z_s),
\label{GroupOne2}\\
&&\text{coming from } \langle \th{1}{a}\th{1}{b}\th{1}{c}\th{1}{d} \rangle
\langle \nabla_a\mathcal M \nabla_b\nabla_c\nabla_d\bar\mathcal M \rangle \,,\nonumber\\
C_{\ell_1}^{(11, 11)}\!&=&\frac{1}{4}\frac{\left(C_\ell^{(0,11)}\right)^2}{C_\ell^\mathcal M}
+\frac{1}{2}\NINT{1}\NINT{2}\,\left\{\left[\left(\vl{}-\vl{1}+\vl{2}\right)\cdot \vl{1}\right]\,
\left(\vl{1}\cdot \vl{2}\right)\right\}^2\,
C_{\ell_1}^\mathcal M(z_s)\nonumber\\
&&\times C_{|\vl{}-\vl{1}+\vl{2}|}^\psi(z_s,z_s)C_{\ell_
2}^\psi(z_s,z_s)\,,
\label{GroupOne3}\\
&&\text{coming from } \langle \th{1}{a}\th{1}{b}\th{1}{c}\th{1}{d} \rangle
\langle \nabla_a\nabla_b\mathcal M\nabla_c\nabla_d\bar\mathcal M \rangle \,.
\nonumber
\end{eqnarray}
More generally, corrections due to the first order deflection angles can be considered beyond the small deflection angle approximation in a non-perturbative sense~\cite{Seljak:1995ve,RuthBook,Lewis:2006fu}.
Indeed, by using the Gaussianity of $\theta^{a(1)}$, exponentiation allows to sum all the corrections coming only from the first order deflection angles non-perturbatively.
This approach is based on the property that a Gaussian stochastic variable $y$ is completely determined by its 2-point statistics, and it is easy to verify that for a Gaussian variable $\langle e^{i y} \rangle=e^{-\langle y^2 \rangle/2}$.
Hence, the correction to the correlation function $\xi(r)$ due to lensing from a Gaussian $\delta\boldsymbol{\theta}$ can be taken into account as follows:
\begin{eqnarray}
\tilde\xi(r)&=&\langle \tilde\mathcal M({\bf x}) \tilde\mathcal M({\bf x}+{\bf r}) \rangle=
\langle \mathcal M({\bf x}+\delta \boldsymbol{\theta}) \mathcal M({\bf x}+{\bf r}+\delta \boldsymbol{\theta}') \rangle
\nonumber
\\
&=&\NINT{}\,C_\ell^\mathcal M e^{i\vl{}\cdot{\bf r}}\langle e^{i\vl{}\cdot\left( \delta \boldsymbol{\theta}- \delta \boldsymbol{\theta}' \right)} \rangle = \NINT{}\,C_\ell^\mathcal M e^{i\vl{}\cdot {\bf r}} e^{- \langle \left[ \vl{}\cdot\left(\delta\boldsymbol{\theta}-\delta\boldsymbol{\theta}' \right) \right]^2 \rangle /2} \,.
\label{eq:exp}
\end{eqnarray}
In the approach used in present public CMB codes like {\sc camb}{}\footnote{\url{http://camb.info}}~\cite{Lewis:1999bs} and {\sc class}{}\footnote{\url{http://class-code.net}}~\cite{Lesgourgues:2011re,Blas:2011rf}, to compute the expectation value
$\langle \left[ \vl{}\cdot\left(\delta \boldsymbol{\theta}-\delta\boldsymbol{\theta}' \right) \right]^2 \rangle$ one
assume $\delta \boldsymbol{\theta}=\boldsymbol{\theta}^{(1)}$ and
defines the matrix
\begin{eqnarray}
A_{a b} \left( {\bf r} \right) = \langle \theta_a^{(1)} \left( {\bf x} \right) \theta_b^{(1)} \left( {\bf x} + {\bf r} \right) \rangle = \int \frac{d^2 \ell}{\left( 2 \pi \right)^2} \ell_a \ell_b C^\psi_\ell e^{i {\bf r} \cdot \boldsymbol{\ell}}\,,
\end{eqnarray}
where we have used Eqs.~(\ref{Weylpot}) and (\ref{TheOrd1}).
Statistical isotropy constrains the matrix $(A_{ab})$ to be of the form
\begin{equation}
A_{a b} \left( {\bf r} \right) = \frac{1}{2} A_0 \left( r \right) \delta_{ab} - A_2 \left( r \right) \left( \frac{r_a r_b}{r^2} - \frac{1}{2} \delta_{ab} \right)\,,
\end{equation}
with
\begin{equation}
A_0 \left( r\right) = \int \frac{d \ell \ \ell^3}{ 2 \pi } C^\psi_\ell J_0 \left( r \ell \right)
\end{equation}
and
\begin{equation}
A_2 \left( r\right) = \int \frac{d \ell \ \ell^3}{ 2 \pi } C^\psi_\ell J_2 \left( r \ell \right) \, .
\end{equation}
Here $J_0$ and $J_2$ are the Bessel functions of order zero and two. Therefore we find
\begin{equation}
\langle \left[ \vl{}\cdot\left(\delta \boldsymbol{\theta}-\delta\boldsymbol{\theta}' \right) \right]^2 \rangle = \ell^2 \left( A_0 \left( 0 \right) - A_0 \left( r \right) +A_2 \left( r \right) \cos \left( 2 \phi \right) \right)\,,
\end{equation}
where $\cos \left( \phi \right) = \hat{ {\bf r}} \cdot \hat { \boldsymbol{\ell}}$. With this, the lensed power spectrum finally becomes
\begin{equation}
\tilde C^{\mathcal M\, (1)}_\ell = \int dr r J_0 \left( \ell r \right) \int \frac{d^2 \ell'}{ \left( 2 \pi \right)^2} C^{\mathcal M}_{\ell'} e^{-i \boldsymbol{\ell}' \cdot {\bf r}} \exp \left[ -\frac{\ell'^2}{2} \left( A_0 \left( 0 \right) - A_0 \left( r \right) + A_2 \left( r \right) \cos \left( 2 \phi \right) \right) \right] \,.
\label{ExpFirstOrder}
\end{equation}
The solution~(\ref{ExpFirstOrder}) captures the full correction, from first order deflection angles alone, to the unlensed $C_\ell^{\mathcal M}$ for arbitrary
large $\ell$, since the first order perturbation angle is fully re-summed \cite{Seljak:1995ve,RuthBook,Lewis:2006fu}. On the other hand, as pointed out before, since deflections angles are typically of the order of arc minutes, the small deflection
angle approximation is sufficient for about $\ell\lesssim 2500$. This is also shown in Fig.~\ref{CompFirstTerm-Resum-Order1},
where we compare the correction on the unlensed $C_\ell^{\mathcal M}$ obtained
considering only the leading correction in the small-deflection angle approximation given in Eqs.~\eqref{eq:firstorder1} and~\eqref{eq:firstorder2} with the correction obtained taking into account the re-summed contributions considering only $\th{1}{a}$ (see Eq.~\eqref{ExpFirstOrder}).
\begin{figure}[ht!]
\centering
\includegraphics{expVsFirstOrder.pdf}
\caption{Corrections due to $\th{1}{a}$. The red curve refers to first order terms given in Eqs.~\eqref{eq:firstorder1} and \eqref{eq:firstorder2}, $\Delta C_\ell=C_\ell^{(11)}+C_\ell^{(1, 1)}$, while black curve takes into account the fully re-summed first order contribution $\Delta C_\ell=\tilde C^{\mathcal M\, (1)}_\ell-C^{\mathcal M}_\ell$.
This result is well known and can be found also e.g. in~\cite{Lewis:2006fu}.}
\label{CompFirstTerm-Resum-Order1}
\end{figure}
As we see from this figure, for $\ell\gtrsim 1500$ the correction from the resummed lensing, Eq.~(\ref{ExpFirstOrder}), is about $25\%$ smaller than the one obtained at lowest order, Eqs.~(\ref{eq:firstorder1}) and~(\ref{eq:firstorder2}),
while the difference is even smaller for lower multipoles $\ell$. In general, we observe that keeping only the first term of the sum,
which corresponds to the leading term of the small-deflection angle approximation, we obtain always the right
order of magnitude of the correction for $\ell\lesssim 2500$, and it is a good approximation at even smaller
$\ell$.
Since the higher order deflections angles are so much smaller, we expect the lowest order approximation to be significantly better at higher order.
We stress again that the exact relation between the lensing potential $C^\psi_\ell$ and the lensed CMB power spectrum of Eq.~(\ref{ExpFirstOrder})
takes into account only the non-perturbative contribution coming from the first order deflection angles, and it is valid because $\th{1}{a}$ has Gaussian statistic.
Going to higher order, taking into account post-Born corrections, generates further terms which lead to non-Gaussian contributions to the deflection angle. In Ref.~\cite{Hagstotz:2014qea} post-Born corrections have been studied only via
a generalization of Eq.~(\ref{ExpFirstOrder}),
which takes into account both the higher order correction to the power spectrum of lensing potential $C_\ell^{\varPsi \varPsi}$ and of the curl potential $C_\ell^{\Omega \Omega}$.
This generalization is still based on the assumption that $\delta \boldsymbol{\theta}$
has a Gaussian statistic and uses the relation
$\langle e^{i y} \rangle=e^{-\langle y^2 \rangle/2}$ for the full non-Gaussian deflection angle. This is obviously not correct beyond
linear order in the deflection angle.
One consequence of using the generalization of Eq.~(\ref{ExpFirstOrder}) to account for the post-Born corrections to
$C_\ell^{\mathcal M}$
is that only contributions with even power of $\delta \boldsymbol{\theta}$ are considered.
Contributions related to the non-Gaussian statistics of
$\delta \boldsymbol{\theta}$, which appear in the terms
with an odd number of deflection angles generated by Taylor expanding $\langle e^{i\vl{}\cdot\left( \delta \boldsymbol{\theta}- \delta \boldsymbol{\theta}' \right)} \rangle$, are neglected.
In the following, we divide our leading post-Born corrections, which are not taken into account in the first group, into a second and a third group. The second group corresponds to the terms that are present also if $\delta \boldsymbol{\theta}$ would have a Gaussian statistics beyond linear order.
These would be the first terms of the tentative re-summation performed in~\cite{Hagstotz:2014qea}.
While the third group contains the leading terms (in the perturbative number of gravitational potentials) of the
non-Gaussian post-Born correction, which contain an odd number of deflection angles and are therefore not considered in the re-summation of Ref.~\cite{Hagstotz:2014qea}, and in general in previous research.
Following the result shown in Fig.~\ref{CompFirstTerm-Resum-Order1} for the corrections related only to $\theta^{(1)}$, we assume that the small-deflection angle approximation holds also for the post-Born higher order corrections that we determine in the following up to $\ell\simeq 2500$. Since higher order corrections are smaller than the first order ones, we believe that this is a safe assumption.
\subsection{Second group}
\label{num-res-s2}
In this second group we study the leading post-Born corrections, which come from
the deflection angles up to third order when these appear in pairs like
$\langle \th{2}{a}\th{2}{b} \rangle$ and $\langle \th{1}{a}\th{3}{b} \rangle$.
They are given by
\begin{eqnarray}
C_\ell^{(1, 3)} &=&-\NINT{1}\NINT{2}
\left[\left(\vl{}-\vl{1}\right)\cdot\vl{1}\right]^2\,
\left[\left(\vl{}-\vl{1}\right)\cdot\vl{2}\right]^2\,
C_{\ell_1}^\mathcal M(z_s)
\nonumber\\
&&\times
\int_{0}^{r_s}dr'
\frac{\left(r_s-r'\right)^2}{r_s^2\,r'^4}
C_{\ell_2}^\psi(z',z')
\nonumber\\
&&\times
P_R\left(\frac{|\vl{}-\vl{1}|+1/2}{r'}\right)
\left[T_{\Psi+\Phi}\left(\frac{|\vl{}-\vl{1}|+1/2}{r'},z'\right)\right]^2\,,\label{eq:secondgroup1}\\
&& \text{coming from }~~ \langle \th{1}{a}\th{3}{b} \rangle
\langle \nabla_a\mathcal M \nabla_b\bar\mathcal M \rangle\,,
\nonumber
\\
C_\ell^{(2, 2)} &=&
\NINT{1}\NINT{2}\,\left[\left(\vl{}-\vl{1}+\vl{2}\right)\cdot\vl{1}\right]^2\,
\left[\left(\vl{}-\vl{1}+\vl{2}\right)\cdot\vl{2}\right]^2\,C_{\ell_1}^\mathcal M(z_s)
\nonumber\\
&&\times\int_{0}^{r_S}dr'\frac{\left(r_s-r'\right)^2}{r_s^2\,r'^4}C_{\ell_2}^\psi(z',z')
\nonumber\\
&&\times P_R\left(\frac{|\vl{}-\vl{1}+\vl{2}|+1/2}{r'}\right)
\left[T_{\Psi+\Phi}\left(\frac{|\vl{}-\vl{1}+\vl{2}|+1/2}{r'},z'\right)\right]^2\,,
\label{eq:secondgroup2}\\
&& \text{coming from } ~~\langle \th{2}{a}\th{2}{b} \rangle
\langle \nabla_a\mathcal M \nabla_b\bar\mathcal M \rangle\,.
\nonumber
\end{eqnarray}
Contrary to $C_\ell^{(0,13)}$ and $C_\ell^{(0,22)}$, the two contributions above
do not exactly cancel.
On the other hand,
in the range of integration for which $|\vl{}-\vl{1}+\vl{2}|\simeq
|\vl{}-\vl{1}|$, i.e. when $\ell_2$ is small, the two above integrands are nearly identical. This leads to a significant but not exact
cancellation between
$C_\ell^{(1,3)}$ and $C_\ell^{(2,2)}$.
The physical interpretation of this is that for $|\vl{}-\vl{1}|\gg\ell_2$ a deflection angle related to the $\ell_2$ acts like a 'global rotation' which has no effect on the lensing corrections to the CMB power spectrum. Therefore, all the contributions from $|\vl{}-\vl{1}|\gg\ell_2$ cancel. This can be considered as a 'consistency relation' for the lensing corrections to the CMB power spectrum.
Similar cancellations have also been found in~\cite{Pratten:2016dsm}.
Actually, each term appearing in the lensing corrections has a 'counter term' which exactly cancels it in this limit. In Fig.~\ref{f:cancel} we show two examples of this partial cancellation which reduces the final result at large $\ell$ up to two orders of magnitude.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.6]{LogCancelationSecondExp.pdf}
\includegraphics[scale=0.6]{LogCancelationThirdExp.pdf}
\caption{Left panel: we show the partial cancellation of the second group.
$\Delta C_\ell=C_\ell^{(1,3)}$ (black curve),
$\Delta C_\ell=C_\ell^{(2,2)}$ (green curve) and the sum of them (red curve). Negative parts are dashed.
The cancellation reduces the final result by about a factor of 700 for $\ell \sim 2500$.
Right panel: we show the partial cancellation of the third group. $\Delta C_\ell=C^{(1,12)}$ (black curve),
$\Delta C_\ell=C^{(2,11)}$ (green curve),
and the sum of them (blue curve). Negative parts are dashed.
The cancellation reduces the final result by about a factor of 50 for $\ell \sim 2500$.}
\label{f:cancel}
\end{figure}
\subsection{Third group}
\label{num-res-s3}
In this third group we consider terms with three deflection angles which do not vanish due to the non-Gaussian statistic of $\theta^{a(2)}$.
These are the following two contributions
\begin{eqnarray}
C_\ell^{(1, 12)}
&=&
-2\,\NINT{1}\NINT{2}
\left(\vl{1}\cdot\vl{2}\right)\,
\left[(\vl{}-\vl{1})\cdot\vl{2}\right]\,
\left[\left(\vl{}-\vl{1}\right)\cdot\vl{1}\right]^2\nonumber\\
&&\times\,C_{\ell_1}^\mathcal M(z_s)
\int_{0}^{r_s}dr'\frac{\left(r_s-r'\right)^2}{r_s^2\,r'^4}\,
P_R\left(\frac{|\vl{}-\vl{1}|+1/2}{r'}\right)\nonumber\\
&&\times\left[T_{\Psi+\Phi}\left(\frac{|\vl{}-\vl{1}|+1/2}{r'},z'\right)\right]^2
C_{\ell_2}^\psi\left( z_s,z' \right)\,,\label{eq:thirdgroup1}\\
&&\text{coming from } ~~ \langle \th{1}{a}\th{1}{b}\th{2}{c} \rangle
\langle \nabla_a\mathcal M \nabla_b\nabla_c\bar\mathcal M \rangle\,,
\nonumber
\\
C_\ell^{(2, 11)}
&=&
2\,\NINT{1}\NINT{2}\,
\left(\vl{1}\cdot \vl{2}\right)\,
\left[\left(\vl{}-\vl{1}+\vl{2}\right)\cdot\vl{2}\right]
\left[\left(\vl{}-\vl{1}+\vl{2}\right)\cdot \vl{1}\right]^2\,
\nonumber\\
&&\times\,C_{\ell_1}^\mathcal M(z_s)
\int_{0}^{r_s}dr'\frac{\left(r_s-r'\right)^2}{r_s^2\,r'^4}
P_R\left(\frac{|\vl{}-\vl{1}+\vl{2}|+1/2}{r'}\right)\nonumber\\
&&\times
\left[T_{\Psi+\Phi}\left(\frac{|\vl{}-\vl{1}+\vl{2}|+1/2}{r'},z'\right)\right]^2
C_{\ell_2}^\psi\left( z_s,z' \right)\,,
\label{eq:thirdgroup2}\\
&&\text{coming from } ~~ \langle \th{2}{a}\th{1}{b}\th{1}{c} \rangle
\langle \nabla_a\mathcal M \nabla_b\nabla_c\bar\mathcal M \rangle \,.
\nonumber
\end{eqnarray}
We note that, also in this case, in the range of integration for which $|\vl{}-\vl{1}+\vl{2}|\simeq
|\vl{}-\vl{1}|$ the two above integrands are nearly identical and the corresponding contributions to
$C_\ell^{(1,12)}$ and $C_\ell^{(2,11)}$ partially cancel.
The physical explanation of this cancelation is the same as in the second group.
However, in this group the integral from the domain where cancellation happens, $|\vl{}-\vl{1}| \gg \ell_2$ contributes less and the cancellation is more than an order of magnitude less significant than for the second group. This is clearly visible in Fig.~\ref{f:cancel} above.
It is interesting to notice that the terms included in the second group do contribute to the lensing and curl potentials in the post-Born approximation, and they are considered already in~\cite{Hagstotz:2014qea} (modulo a possible sign error), but the terms of the third group do not appear in these potentials, they are
new contributions.
We can now write the fully corrected lensed CMB temperature anisotropy power spectrum in the form
\begin{equation}
\tilde C_\ell^{\mathcal M} =\tilde C_\ell^{\mathcal M\, (1)} + \Delta C_\ell^{(2)} + \Delta C_\ell^{(3)} \label{eq:cltot}
\end{equation}
where
\begin{eqnarray}
\Delta C_\ell^{(2)} &=& C_\ell^{(1, 3)} + C_\ell^{(2, 2)} \label{eq:cl2}\\
\Delta C_\ell^{(3)} &=& C_\ell^{(1, 12)} + C_\ell^{(2, 11)} \label{eq:cl3} \,.
\end{eqnarray}
Here the $\tilde C_\ell^{\mathcal M\, (1)} $ term denotes the well known resummed correction from the first order
deflection angle given in Eq.~\eqref{ExpFirstOrder}, while $\Delta C_\ell^{(2)}$ and $\Delta C_\ell^{(3)}$ denote the post-Born corrections from the second and third group respectively.
\subsection{Numerical Results}
\label{num-res-s4}
In this subsection we show the numerical evaluation of the results given above, considering both linear and non-linear (Halofit model \cite{Smith:2002dz,Takahashi:2012em}) power
spectra for the gravitational potential.
More precisely, here and above the figures have been generated with the following cosmological parameters $h = 0.67$, $\omega_\text{cdm}=0.12$, $\omega_b = 0.022$ and vanishing curvature. The primordial curvature power spectrum has the amplitude $A_s = 2.215 \times 10^{-9}$, the pivot scale $k_\text{pivot} = 0.05 \ \text{Mpc}^{-1}$, the spectral index $n_s = 0.96$ and no running, compatible with~\cite{Planck:2015xua}. The Bardeen transfer function $T_{\Phi+\Psi}$ has been computed with~{\sc class}{}~\cite{Blas:2011rf}, using Halofit \cite{Takahashi:2012em} for the non-linear case.
In Fig.~\ref{G-figure} we plot the relative
correction $\Delta C_\ell^{(2)}/\tilde C^{\mathcal M (1)}_\ell$ to the lensed temperature anisotropy spectrum given in Eq.~(\ref{ExpFirstOrder}) obtained when considering the
post-Born correction of the second group, Eqs.~\eqref{eq:secondgroup1} and~\eqref{eq:secondgroup2}, assuming linear and non-linear power spectra. This correction is the post-Born contribution,
coming from deflection angles up to third order, that is present in a generalization of Eq.~(\ref{ExpFirstOrder}) as
performed in~\cite{Hagstotz:2014qea}. As a consequence, this is the effect that should be compared with the result given in~\cite{Hagstotz:2014qea}.
This comparison shows how in \cite{Hagstotz:2014qea} the corresponding contribution is overestimated by two orders of magnitude. This overestimation is probably due to the sign error pointed out in
Section \ref{Sec2}.
\begin{figure}[h!]
\centering
\includegraphics[scale=1.1]{SecondLinearVsHalo.pdf}
\caption{The correction with respect to the standard lensed temperature power spectrum
$\tilde C^{\mathcal M (1)}_\ell$, given in Eq.(\ref{ExpFirstOrder}),
from the second group of terms $\Delta C_\ell^{(2)}$ (see Eq.~\eqref{eq:cl2}), for linear (red curve) and non-linear (black curve) power spectra.}
\label{G-figure}
\end{figure}
In Fig.~\ref{NG-figure} we plot the relative
correction $\Delta C_\ell^{(3)}/\tilde C^{\mathcal M (1)}_\ell$ to the lensed temperature anisotropy spectrum given in Eq.~(\ref{ExpFirstOrder})
obtained when we consider the
post-Born correction from the third group, Eqs.~\eqref{eq:thirdgroup1} and \eqref{eq:thirdgroup2}, both for linear and non-linear power spectra. This correction is the first non-Gaussian contribution
coming from $\th{2}{a}$.
This contribution is not considered in~\cite{Hagstotz:2014qea}, and in general when the next-to-leading order corrections are evaluated via the amplification matrix and using the relation
$\langle e^{i y} \rangle=e^{-\langle y^2 \rangle/2}$.
But, as can be seen comparing Figs.~\ref{G-figure} and~\ref{NG-figure}, this contribution is
more than one order of magnitude larger than the total contribution from the second group.
Each individual term of this group is about a factor of two larger than the terms of the second group.
But more importantly, the cancellation is more than one order of magnitude less pronounced at argued above.
\begin{figure}[ht!]
\centering
\includegraphics[scale=1.1]{ThirdLinearVsHalo.pdf}
\caption{
The correction with respect to the standard lensed temperature power spectrum
$\tilde C^{ \mathcal M (1)}_\ell$, given in Eq.(\ref{ExpFirstOrder}),
from the third group of terms $\Delta C_\ell^{(3)}$ (see Eq.~\eqref{eq:cl3}), for linear (red curve) and non-linear (black curve) power spectrum.
This correction is the first non-Gaussian contribution
coming from $\th{2}{a}$.
}
\label{NG-figure}
\end{figure}
Figs.~\ref{G-figure} and, especially, Fig.~\ref{NG-figure} are the main results of this paper.
Our new contribution shown in Fig.~\ref{NG-figure}, which is neglected in~\cite{Hagstotz:2014qea}, dominates
the terms of the second group by more of one order of magnitude.
The full result is presented in Fig.~\ref{All-effects}.
In this figure we show the corrections to the lensed power spectrum
$\tilde C^{\mathcal M (1)}_\ell$ from the nonlinearities of the matter power spectrum, we plot
$$\left((\tilde C^{\mathcal M (1)}_\ell)_{\rm non-lin} - (\tilde C^{\mathcal M\, (1)}_\ell)_{\rm lin}\right)/ (\tilde C^{\mathcal M (1)}_\ell)_{\rm lin}\,,
$$
from the post-Born corrections, we plot
$$
\left((\Delta C_\ell^{(2)})_{\rm lin} + (\Delta C_\ell^{(3)})_{\rm lin}\right)/ (\tilde C^{\mathcal M (1)}_\ell)_{\rm lin}\,,
$$
and from both post-Born corrections and a non-linear power spectrum, we plot
$$
\left((\tilde C^{\mathcal M (1)}_\ell)_{\rm non-lin} +(\Delta C_\ell^{(2)})_{\rm non-lin} + (\Delta C_\ell^{(3)})_{\rm non-lin}- (\tilde C^{\mathcal M\, (1)}_\ell)_{\rm lin}\right)/ (\tilde C^{\mathcal M (1)}_\ell)_{\rm lin}\,.
$$
As one can see, including a non-linear power spectrum is much more important than the post-Born correction. The post-Born correction increases from $0.01\%$ at $\ell \sim 1000$ to $0.05\%$ for $\ell \simeq 2500$, while the non-linearities give corrections of up to $1\%$ at $\ell=2500$.
\begin{figure}[ht!]
\centering
\includegraphics{All-effects.pdf}
\caption{Correction to the lensed $\tilde C^{\mathcal M (1)}_\ell$, evaluated with a linear power spectrum and given in Eq.~(\ref{ExpFirstOrder}), due to the non-linear matter power spectrum (Halofit model, red curve), the terms beyond the Born approximation ($\Delta C_\ell=\Delta C_\ell^{(2)}+\Delta C_\ell^{(3)}$, green curve) and both (blue curve). }
\label{All-effects}
\end{figure}
Finally, in Fig.~\ref{Comparing-all} we plot the corrections to the unlensed $C_\ell^\mathcal M$'s due to resummed $\th{1}{a}$, post-Born contribution from Eqs.~(\ref{eq:secondgroup1}) and~(\ref{eq:secondgroup2}), and post-Born non-Gaussian contribution from
Eqs. (\ref{eq:thirdgroup1}) and (\ref{eq:thirdgroup2}),
for both linear power spectrum and Halofit model.
As one can see the non-Gaussian part (third group) is always the most relevant post-Born correction
in going beyond the first order deflection angles contribution. Nevertheless, these post-Born corrections are below cosmic variance, as we can see from Fig.~\ref{Cosmic_Variance}. There we plot the ratio $\Delta C_\ell/\sigma_\ell$, where
\begin{equation}
\sigma_\ell\equiv\sqrt{\frac{2}{2\ell+1}}\,C^\mathcal M_\ell
\end{equation}
is the cosmic variance. We have taken into account only the linear power spectrum because we are just interested in comparing the orders of magnitude of these different effects. Lensing corrections are detectable if this ratio is larger than 1. Within the range in $\ell$ where our approximations hold ($\ell\leq2500$), post-Born corrections are smaller than cosmic variance by two orders of magnitude. In order to pass this threshold one would have to combine several hundred $\ell$-values into one bin.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.6]{Linear.pdf}
\includegraphics[scale=0.6]{Halo.pdf}
\caption{Corrections to the unlensed $C_\ell^\mathcal M$'s due to resummed $\th{1}{a}$ ($\Delta C_\ell=\tilde C^{\mathcal M\, (1)}_\ell-C^{\mathcal M}_\ell$, black curve), the second group ($\Delta C_\ell=
\Delta C_\ell^{(2)}$, red curve) and the third group ($\Delta C_\ell=
\Delta C_\ell^{(3)}$, blue curve) for the linear matter power spectrum (left panel) and the Halofit model (right panel). Dashed lines are negative parts.
As we can see from these figures, the non-Gaussian part is the most relevant
post-Born correction in going beyond the first order deflection angles contribution, for
both linear and Halofit models.}
\label{Comparing-all}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics{RelativeCosmicVariance.pdf}
\caption{The ratios of the lensing corrections and the cosmic variance $\sigma_\ell\equiv\sqrt{\frac{2}{2\ell+1}}C^{\mathcal M}_\ell$ for the resummed $\th{1}{a}$ ($\Delta C_\ell=\tilde C^{\mathcal M\, (1)}_\ell-C^{\mathcal M}_\ell$, black curve), the second group ($\Delta C_\ell=
\Delta C_\ell^{(2)}$, red curve) and the third group ($\Delta C_\ell=
\Delta C_\ell^{(3)}$, blue curve) for the linear power spectrum (dashed parts are negative values). The straight black line is the detectability threshold $\Delta C_\ell/\sigma_\ell=1$. The post-Born corrections are below the cosmic variance by two orders of magnitude up to $\ell=2500$.}
\label{Cosmic_Variance}
\end{figure}
\section{Conclusions}
\label{Sec7}
\setcounter{equation}{0}
In this work we have evaluated the impact of going beyond the Born approximation on the lensed CMB temperature
power spectrum. We postpone the study of this effect on E and B-mode polarization for a forthcoming paper.
The evaluation of the post-Born lensing correction on the unlensed $C_\ell^{\mathcal M}$ is performed
in the small deflection angle approximation. This has the drawback that it is reliable only for multipole $\ell\lesssim 2500$, but it allows us to consistently take into account the non-Gaussian nature of cosmological perturbation theory beyond the linear level.
In fact, in the previous literature \cite{Hagstotz:2014qea}, the evaluation of post-Born correction was performed
using a non-perturbative re-summation based on the hypothesis that also the non-linear part of the deflection angle $\delta \boldsymbol{\theta}$ has Gaussian statistics, needed to use the relation
$\langle e^{i y} \rangle=e^{-\langle y^2 \rangle/2}$ valid only for a Gaussian stochastic variable $y$.
This is of course no longer valid when we go beyond linear order, to take into account post-Born corrections to the deflection angle.
The contributions to the lensed temperature power spectrum coming from the non-Gaussian nature of the deflection angle are given in Eqs.~(\ref{eq:thirdgroup1}) and~(\ref{eq:thirdgroup2}). This is a new effect not taken into account in the past literature, and it turns out to be
the leading contribution when compared with the contribution given in Eqs.~(\ref{eq:secondgroup1}) and~(\ref{eq:secondgroup2}), i.e., the contribution that will be appear also in the re-summation performed in~\cite{Hagstotz:2014qea}.
From a quantitative point of view the non-Gaussian contribution from Eqs.~(\ref{eq:thirdgroup1}) and~(\ref{eq:thirdgroup2}) is more than one
order of magnitude larger of the one given in Eqs.~(\ref{eq:secondgroup1}) and~(\ref{eq:secondgroup2}), see Fig.~\ref{Comparing-all}.
This gives a correction that oscillates and becomes close to $0.05\%$ for multipoles between 2000 and 2500.
This result corrects the one of~\cite{Hagstotz:2014qea}, which is on the one hand incomplete and on the other hand overestimates the contributions from the second group by about two order of magnitude.
(The fact that the result by~\cite{Hagstotz:2014qea} is similar to the one from Eqs.~(\ref{eq:secondgroup1}) and (\ref{eq:secondgroup2}) without cancellation, hints to the fact that the overestimation is probably due to the sign error pointed out in Sect.~\ref{Sec2}.)
Looking at Fig.~\ref{All-effects} we also note that the non-linear matter power spectrum changes the lensed temperature spectrum much more (by about $1\%$ for $\ell\sim 2500$), than the post-Born corrections.
The magnitude of the correction from going beyond the Born approximation evaluated here is
relatively small and, at the present, not detectable in CMB experiments. On the other hand, its dominant contribution originates from a new
effect, neglected in~\cite{Hagstotz:2014qea}, and similar contributions could be more significant, in particular, for B-mode
polarization. This is the subject of a future paper.
\vspace{0.5cm}
{\bf Note Added.}
While we were finalizing this manuscript another study of the effect of post-Born lensing corrections
on CMB temperature anisotropies has appeared in~\cite{Pratten:2016dsm}.
Our analytical results for higher order deflection angles and the amplification matrix~\cite{Fanizza:2015swa} are in agreement with the ones used in~\cite{Pratten:2016dsm}. This means that we also agree on
the post-Born corrections to the lensing and curl potential power spectra.
As a consequence, we also obtain results consistent with the ones presented in~\cite{Pratten:2016dsm}
considering the contribution to add to the unlensed $C_\ell^\mathcal M$ coming from the terms which we call the second group, $\langle \th{1}{a}\th{3}{b} \rangle
\langle \nabla_a\mathcal M \nabla_b\bar\mathcal M \rangle$ and $ \langle \th{2}{a}\th{2}{b} \rangle
\langle \nabla_a\mathcal M \nabla_b\bar\mathcal M \rangle$,
which is the first contribution of the re-summed series probably used in~\cite{Pratten:2016dsm} to evaluate post-Born
contribution. Regarding this contribution, we also agree with~\cite{Pratten:2016dsm} that
some mistakes are present in~\cite{Hagstotz:2014qea} which overestimates the correction.
On the other hand, like~\cite{Hagstotz:2014qea}, also in~\cite{Pratten:2016dsm} the assumption that the lensing potential is Gaussian leads the authors to neglect contributions coming
from an odd number of deflection angles.
In fact, despite mentioning that non-Gaussianity from post-Born corrections can give an additional contribution, this contribution is not evaluated in \cite{Pratten:2016dsm}.
Due to the non-Gaussian nature of higher order perturbations this contribution does not vanish beyond linear order and is given by the terms which we call third group, $\langle \th{1}{a}\th{1}{b}\th{2}{c} \rangle \langle \nabla_a\mathcal M \nabla_b\nabla_c\bar\mathcal M \rangle$ and $ \langle \th{2}{a}\th{1}{b}\th{1}{c} \rangle
\langle \nabla_a\mathcal M \nabla_b\nabla_c\bar\mathcal M \rangle$.
This leads in~\cite{Pratten:2016dsm} to an underestimation
of the total correction from post-Born effects by at least one order of magnitude.
\section*{Acknowledgements}
We wish to thank Francesco Montanari for several useful discussions and suggestions.
We are grateful to the Swiss National Science Foundation for financial support.
GM wishes to thank CNPq for financial support.
GM was also partially supported by the Marie Curie IEF, Project NeBRiC - ``Non-linear effects and backreaction in classical and quantum cosmology" at the initial stage of this work.
GF is supported in part by MIUR, under grant no. 2012CPPYP7
(PRIN 2012), and by INFN, under the program TAsP ({Theoretical Astroparticle Physics}).
GF was also supported by the foundation ``Angelo Della Riccia''.
ED is supported by the ERC Starting Grant cosmoIGM and by INFN/PD51 INDARK grant.
We thank the Galileo Galilei Institute for Theoretical Physics for the hospitality and the INFN for partial support during the completion of this work.
|
\section{}
\section{Introduction}
Biological networks are often complex and it is frequently necessary to focus on subnetworks of a larger system, e.g.\ to enable a better understanding of the network properties \cite{Bhalla2003,Ackermann2012,Conradi2007}. Such subnetworks may be of interest because they carry out important biological functions or because they capture parts of the system where there is less uncertainty in the network structure or dynamical parameters. The choice of subnetwork may also be dictated by experimental data only being available for a limited number of molecular species.
There are many different methods of model reduction that have been used to simplify a large model down to one for a subnetwork \cite{Okino1998,Radulescu2012}. We have previously used projection techniques to find a systematic description of the dynamics of a subnetwork embedded in a ``bulk'' network \cite{Rubin2014d}, identifying the occurrence of memory terms as one of the key features. The resulting methods for calculating memory functions were, however, restricted to networks involving only unary and binary reactions with concentration-independent reaction rates.
While this class of networks is large, it excludes networks with reactions following Michaelis-Menten kinetics, which is commonly used to represent e.g.\ enzyme reactions. Our aim in this paper is to remove this restriction, by providing an explicit method for constructing the description of the dynamics of any subnetwork within a network consisting of unary, binary and Michaelis-Menten reactions.
An approximate way of achieving the above aim is to extend a network with Michaelis-Menten reactions to one that explicitly involves enzymes and enzyme complexes. The Michaelis-Menten terms in the kinetics are thus ``unfolded'' into unary and binary reactions so that the original projection approach \cite{Rubin2014d} can be applied. However, this is inconvenient both conceptually -- we need to include extra species not present in the original system of reaction equations -- and numerically, because the fast rates of enzyme reactions typically create a ``stiff'' system of differential equations that has to be integrated using small timesteps. The approach nevertheless provides the inspiration for the method we adopt here: we follow the above route {\em analytically}, taking the limit of fast enzyme reaction rates in such a way as to retrieve the original Michaelis-Menten kinetics exactly. The main achievement of our analysis is to show that the limit can be taken in closed form. This leads to a procedure for constructing the subnetwork dynamics directly from the original reaction network, without reference to any extra species.
In Section \ref{sec:MM} we give a summary of Michaelis-Menten dynamics and the conditions under which it is retrieved as the limit of a fast formation and dissociation of an enzyme complex; in essence these conditions are that the enzyme reactions must be fast, and the enzyme concentration low. Section \ref{sec:MMproj} details our approach to obtaining the projected equations that describe subnetwork dynamics, for systems of reaction equations that include Michaelis-Menten reactions: we temporarily add enzymes to represent these reactions, derive the projected equations, and then take the limit of fast enzyme rate and low enzyme concentration in closed form.
We describe the approach separately for
linearised (Sec.~\ref{sec:linMM}) and nonlinear (Sec.~\ref{sec:nonlinMM}) dynamics, as the nonlinear case is more complicated technically but follows the same conceptual route.
Remarkably, even though Michaelis-Menten terms are generally nonlinear, we find that in the memory terms that are characteristic of the projected equations no additional nonlinearities appear, i.e.\ the memory terms involve linear concentration fluctuations for linearised dynamics, and linear and quadratic concentration fluctuations for nonlinear dynamics. Finally in Section \ref{sec:MMcomparison} we compare predictions from the original reaction equations with Michaelis-Menten dynamics to the projected equations with either enzymes added explicitly or eliminated in closed form using the method derived in this paper.
We find that the closed form elimination is both faster to evaluate computationally, and gives a more accurate approximation to the original reaction equations.
\section{Michaelis-Menten dynamics}
\label{sec:MMdynamics}
Many biochemical reactions are catalysed by enzymes. Generally each enzyme will enable a particular reaction without being consumed. A simple model of an enzyme reaction \cite{Henri1902,Michaelis1913} is written
\begin{equation}
\label{eq:michaelismenten}
u + e \xrightleftharpoons[k^-_{{c},ue}}%{k_{-1}]{k^+_{ue,{c}}}%{k_1} c \xrightarrow{k^-_{{c},pe}}%{k_{2}} e +p
\end{equation}
where $u$ is a substrate (we use ``u'' not ``s'' here as we will need the letter ``s'' to denote ``subnetwork'' later), $e$ is an enzyme, $c$ is an enzyme-substrate complex and $p$ is a product. The reaction rates are denoted $k^+_{ue,{c}}}%{k_1, k^-_{{c},ue}}%{k_{-1}, k^-_{{c},pe}}%{k_{2}$.
These reactions describe the binding of a free enzyme with a substrate to form a substrate-enzyme complex. This complex can then dissociate into enzyme and a product. In the traditional model substrate binding is reversible but product formation is not; we will consider also the more general case below, where both reactions are reversible.
\subsection{Derivation of Michaelis-Menten equations}\
\label{sec:MM}
Let $\protein{i}$ be the concentration of species $i$. Then the set of mass action equations for system \eqref{eq:michaelismenten} is
\begin{equation}
\label{eq:MMmassaction}
\begin{split}
\frac{\partial}{\partial t}\protein{u} &= k^-_{{c},ue}}%{k_{-1}\protein{c}-k^+_{ue,{c}}}%{k_1\protein{u}\protein{e}\\
\frac{\partial}{\partial t}\protein{e} &= k^-_{{c},ue}}%{k_{-1}\protein{c}-k^+_{ue,{c}}}%{k_1\protein{u}\protein{e}+k^-_{{c},pe}}%{k_{2}\protein{c}\\
\frac{\partial}{\partial t}\protein{c} &= -k^-_{{c},ue}}%{k_{-1}\protein{c}+k^+_{ue,{c}}}%{k_1\protein{u}\protein{e}-k^-_{{c},pe}}%{k_{2}\protein{c}\\
\frac{\partial}{\partial t}\protein{p} &= k^-_{{c},pe}}%{k_{2}\protein{c}
\end{split}
\end{equation}
From these equations it follows that there is a conservation law between the enzyme and complex such that
\begin{equation}
\label{eq:MMconslaw}
\frac{\partial}{\partial t}\protein{e} + \frac{\partial}{\partial t}\protein{c} = 0 \Longrightarrow \protein{e}+ \protein{c} = x_e^{\rm{tot}}
\end{equation}
The Michaelis-Menten description of the dynamics is obtained by exploiting the fact that enzyme reactions are typically fast. This allows one to reduce the system \eqref{eq:michaelismenten} to a simpler description where the enzyme and enzyme complex no longer appear explicitly.
To achieve this simplification one uses the quasi-steady state assumption \cite{Briggs1925
. We assume all enzyme reactions are fast so that the enzyme complex is in quasi-steady state at any time. We use the qualifier ``quasi'' because this steady state depends on the concentrations of both substrate and product, which themselves generally vary in time.
For the reaction flux $v = k^-_{{c},pe}}%{k_{2}\protein{c}$ one then finds
\begin{equation}
\label{eq:MMreactionrate}
v = \frac{k^-_{{c},pe}}%{k_{2}x_e^{\rm{tot}}\protein{u}}{(k^-_{{c},ue}}%{k_{-1}+k^-_{{c},pe}}%{k_{2})/k^+_{ue,{c}}}%{k_1 + \protein{u}} = \frac{V_{\rm{max}}\protein{u}}{K_{\rm{m}}+\protein{u}}
\end{equation}
where
\begin{equation}
\label{eq:vmax}
V_{\rm{max}} = k^-_{{c},pe}}%{k_{2}x_e^{\rm{tot}}
\end{equation}
is the maximum reaction flux that can be achieved and
\begin{equation}
\label{eq:Km}
K_{\rm{m}} = \frac{k^-_{{c},ue}}%{k_{-1}+k^-_{{c},pe}}%{k_{2}}{k^+_{ue,{c}}}%{k_1}
\end{equation}
is the Michaelis constant.
The simplified description of the original reaction system can now be written in terms of the reaction flux $v$ in \eqref{eq:MMreactionrate}, as
\begin{equation}
\label{eq:MMreactioneqns}
\begin{split}
-\frac{\partial}{\partial t}\protein{u}&=\frac{V_{\rm{max}}\protein{u}}{K_{\rm{m}}+\protein{u}}
=
\frac{\partial}{\partial t}\protein{p}
\end{split}
\end{equation}
In the limit where the complex dissociates into enzyme and substrate at a much higher rate than for enzyme and product, one can achieve a further simplification known as the ``rapid equilibrium assumption'' \cite{Michaelis1913}. However, this is just a limiting case of the quasi-steady state assumption where $k^-_{{c},pe}}%{k_{2}$ is neglected against $k^-_{{c},ue}}%{k_{-1}$ in determining the Michaelis constant $K_{\rm{m}}$.
\subsection{Reversible Michaelis-Menten dynamics}
\label{sec:MMreversible}
We discuss briefly how the above analysis is modified when there is a back reaction from the enzyme and product to the complex \cite{Haldane1930,Sauro2013}. In such cases one has the more general reaction scheme
\begin{equation}
\label{eq:MMreversible}
u + e \xrightleftharpoons[k^-_{{c},ue}}%{k_{-1}]{k^+_{ue,{c}}}%{k_1} c \xrightleftharpoons[k^+_{pe,{c}}}%{k_{-2}]{k^-_{{c},pe}}%{k_{2}} e +p
\end{equation}
with the additional rate constant $k^+_{pe,{c}}}%{k_{-2}$.
The equation for the enzyme complex concentration now has an extra contribution $k^+_{pe,{c}}}%{k_{-2}\protein{e}\protein{p}$, while the equations for enzyme and product contain the same additional term with a negative sign. Using the quasi-steady state assumption one then obtains a reaction flux of the form
\begin{equation}
\label{eq:MMreactionratereverse}
v = \frac{V_{\rm{u}}\protein{u}/K_{\rm{u}} - V_{\rm{p}}\protein{p}/K_{\rm{p}}}{1 + \protein{u}/K_{\rm{u}} + \protein{p}/K_{\rm{p}}}
\end{equation}
where $V_{\rm{u}}$ and $V_{\rm{p}}$ are the maximum reaction rates for the forward and reverse reactions respectively and are given by
\begin{equation}
\label{eq:vmaxdefns}
V_{\rm{u}} = k^-_{{c},pe}}%{k_{2}x_e^{\rm{tot}},\quad
V_{\rm{p}} = k^-_{{c},ue}}%{k_{-1}x_e^{\rm{tot}}
\end{equation}
Similarly $K_{\rm{u}}$ and $K_{\rm{p}}$ are the Michaelis constants for the forward and reverse reactions and are written in terms of the mass action rate constants as
\begin{equation}
\label{eq:kmdefns}
K_{\rm{u}} = \frac{k^-_{{c},ue}}%{k_{-1} + k^-_{{c},pe}}%{k_{2}}{k^+_{ue,{c}}}%{k_1},\quad K_{\rm{p}} = \frac{k^-_{{c},ue}}%{k_{-1}+k^-_{{c},pe}}%{k_{2}}{k^+_{pe,{c}}}%{k_{-2}}
\end{equation}
\subsection{Quantitative accuracy of Michaelis-Menten approximation}
\label{sec:MMtoMA}
We outlined above how the Michaelis-Menten dynamics \eqref{eq:MMreactioneqns} -- or its generalisation \eqref{eq:MMreactionratereverse} -- emerges from a quasi-steady state approximation. We now consider under what conditions on the enzyme reaction parameters these approximations become exact, so that we can later take the appropriate limit for the enzyme parameters in the construction of the projected equations. Intuitively the enzyme reactions need to be fast; from the definition of the maximal reaction rate \eqref{eq:vmax}, which in a reaction specified from the outset in Michaelis-Menten form is given, the total enzyme concentration $x_e^{\rm{tot}}$ then needs to be small. This makes sense as the Michaelis-Menten kinetics always gives equal and opposite reaction fluxes for substrate and product, neglecting any substrate or product captured in the enzyme complex. A more formal analysis by \citet{Segel1989}, based on a singular perturbation approach to a dimensionless form of the reaction equations, gives $x_e^{\rm{tot}}/(\protein{u}+K_{\rm{m}}) \ll 1$ as the condition for validity of the quasi-steady state approximation. As $K_{\rm{m}}$ can also be regarded as fixed by the specification of a Michaelis-Menten reaction, this is consistent with the intuitive requirement of small enzyme concentration.
The discussion so far suggests that we should rewrite the enzyme reaction rate constants and total enzyme concentrations as
\begin{equation}
\label{eq:reactionratesfast}
k^+_{ue,{c}}}%{k_1 = \bar{k}^+_{ue,{c}}}%{\bar{k}_1\gamma,\ k^-_{{c},ue}}%{k_{-1} = \bar{k}^-_{{c},ue}}%{\bar{k}_{-1}\gamma,\ k^-_{{c},pe}}%{k_{2} = \bar{k}^-_{{c},pe}}%{\bar{k}_{2}\gamma,\ x_e^{\rm{tot}} = \bar{x}_e^{\rm{tot}}/\gamma
\end{equation}
where $\gamma$ is a dimensionless fast rate scale parameter.
The definitions of $V_{\rm{max}}$ \eqref{eq:vmax} and $K_{\rm{m}}$ \eqref{eq:Km} then allow us to write the remaining parameters as
\begin{equation}
\bar{x}_e^{\rm{tot}} = V_{\rm{max}}/\bar{k}^-_{{c},pe}}%{\bar{k}_{2},\
\bar{k}^+_{ue,{c}}}%{\bar{k}_1= \frac{\bar{k}^-_{{c},ue}}%{\bar{k}_{-1}+\bar{k}^-_{{c},pe}}%{\bar{k}_{2}}{K_{\rm{m}}}
\end{equation}
We now have three parameters, $\bar{k}^-_{{c},ue}}%{\bar{k}_{-1}$, $\bar{k}^-_{{c},pe}}%{\bar{k}_{2}$ and $\gamma$, that parameterise the possible enzyme kinetics underlying a given Michaelis-Menten reaction.
From the criterion of \cite{Segel1989} the Michaelis-Menten description will then become exact for $\gamma\to\infty$, irrespective of the values of $\bar{k}^-_{{c},ue}}%{\bar{k}_{-1}$ and $\bar{k}^-_{{c},pe}}%{\bar{k}_{2}$ \cite{VanSlyke1914}. The same exactness statement applies to the more general case where there is also a back reaction from product and enzyme to form an enzyme complex \cite{Li2013, Kollar2015}.
\section{Enzyme reactions in the projected equations}
\label{sec:MMproj}
The projection method as applied to protein interaction networks \cite{Rubin2014d} works with mass action kinetics for unary and binary reactions. Therefore if we
are given an interaction network that includes Michaelis-Menten reactions then {\em a priori} we need to represent these explicitly in such a form, to allow us to compute the projected equations. As explained in the introduction, this is a disadvantage both conceptually and computationally.
Our aim here is to implement this approach {\em analytically} instead: we add enzyme species to the network to get mass action kinetics and take the large enzyme rate and low enzyme concentration limit in which the mass action dynamics becomes exactly identical to Michaelis-Menten dynamics. The challenge is to understand what happens in this limit to the projected equations.
The aim of the projection method generally is to provide a description of the dynamics of a protein interaction subnetwork embedded in a bulk network. The Zwanzig-Mori projection method \cite{Mori1965} can be used to obtain such a description, specifically equations for the time evolution of the protein concentrations in a chosen subnetwork. Full details of the projection method applied to protein interaction networks are given in \cite{Rubin2014d}. We summarise below the features necessary for the analysis of enzyme dynamics.
The natural specification of the state of a given network is the vector of concentrations for all molecular species in the network, which we call $\bm{x}$. In the steady state $\bm{x}$ will fluctuate around some mean $\bm{y}$ and it will be useful to switch variables to $\bm{\delta x}=\bm{x}-\bm{y}$, which has zero mean in steady state. The protein concentrations (concentration deviations from the mean, more precisely) $\bm{\delta x}$ evolve in time according to a Fokker-Planck equation, where the stochasticity arises from copy number fluctuations. The variance of these fluctuations scales as the inverse reaction (e.g.\ cell) volume $\epsilon=1/V$. While a nonzero $\epsilon$ is needed initially to apply the Zwanzig-Mori formalism, we consider the limit of small $\epsilon$ throughout.
The time evolution of any observable $a(\bm{\delta x},t)$ is given by $(\partial/\partial t)a = \mathcal{L}a$, where $\mathcal{L}$ is the adjoint Fokker-Planck operator whose drift term encodes the mass action kinetics. We showed in \cite{Rubin2014d} that if we focus on a set of observables $\bm{z}$ containing $\bm{\delta x}$ and all products like $\prot{1}^2$ and $\prot{1}\prot{2}$, then the operator $\mathcal{L}$ can be written in {\em matrix form} $\bm{L}$ such that
\begin{equation}
\label{eq:Lmatrixeqn}
\frac{\partial}{\partial t}z_{\alpha}=\sum_\beta z_{\beta}L_{\beta\alpha} + \prot{}^3+\mathcal{O}(\epsilon)
\end{equation}
where $\prot{}^3$ indicates terms of third or higher order in $\bm{\delta x}$ while the $\mathcal{O}(\epsilon)$ terms vanish in the low noise limit.
If we now let $\{a_\alpha(\bm{\delta x})\}$ be a set of observables from the subnetwork, the projected equations have the form
\begin{equation}
\label{eq:projectedeqns}
\frac{\partial}{\partial t}a_\alpha(t) = \sum_\beta a_\beta(t)\Omega_{\beta\alpha} + \int_0^tdt'\,\sum_\beta a_\beta(t')M_{\beta\alpha}(t-t') + r_\alpha(t)
\end{equation}
We choose specifically as subnetwork observables all the subnetwork concentrations and their products, i.e.\ the entries of $\bm{z}$ only involving the subnetwork. We denote these collectively by S, and the remaining species -- all bulk concentrations (b) and concentration products involving the bulk (sb and bb) -- by the letter B.
With the observables ordered appropriately, we can then decompose the matrix $\bm{L}$ formed from the entries $L_{\beta\alpha}$ into subnetwork and bulk blocks:
\begin{equation}
\bm{L}=
\label{eq:Lmatrix}
\left(
\begin{array}{cc}
\lblockb{S}{S}& \lblockb{S}{B}\\
\lblockb{B}{S}&\lblockb{B}{B}
\end{array}
\right)
\end{equation}
The four blocks here can be broken down further according to the specific type of subnetwork and bulk observable, for example
\begin{equation}
\lblockb{B}{S} =
\left(\begin{array}{cc}
\lblockb{b}{s}&0\\
\lblockb{sb}{s}&\lblockb{sb}{ss}\\
\lblockb{bb}{s}&\lblockb{bb}{ss}\\
\end{array}\right)
\end{equation}
Here ``s'' denotes linear subnetwork observables, ``ss'' product observables from the subnetwork, ``sb'' cross-product observables between subnetwork and bulk, and so on. In this way $\lblockb{b}{s}$ contains the linear coefficients of bulk concentrations in the equations of motion for subnetwork concentrations, while the coefficients of subnetwork-bulk products in these equations are in the block $\lblockb{sb}{s}$.
In terms of the blocks of the $\bm{L}$-matrix defined as above, the ``rate matrix'' $\bm{\Omega}$ in \eqref{eq:projectedeqns} is simply \cite{Rubin2014d}
\begin{equation}
\label{eq:omega}
\bm{\Omega}=\lblockb{S}{S} =
\left(\begin{array}{cc}
\lblockb{s}{s}&0\\
\lblockb{ss}{s}&\lblockb{ss}{ss}
\end{array}\right)
\end{equation}
It contains terms from the subnetwork dynamics that are local in time. The memory function (matrix) can be written as \cite{Rubin2014d}
\begin{equation}
\label{eq:mem}
\bm{M}(\Delta t) = \lblockb{S}{B}e^{\lblockb{B}{B}\Delta t}\lblockb{B}{S}
\end{equation}
where $\Delta t=t-t'$. The entries of this matrix, which appear in \eqref{eq:projectedeqns}, are the memory functions
$M_{\beta\alpha}(\Delta t)$: they determine how strongly the past values of
the observable $a_\beta$ affect the present rate of change of
$a_\alpha$. For later use we note here the Laplace transform version
of the memory function, which is
\begin{equation}
\label{eq:memLT}
\hat{\bm{M}}(z) = \lblockb{S}{B}\left(z-\lblockb{B}{B}\right)^{-1}\lblockb{B}{S}
\end{equation}
For brevity we will use the term ``memory function'' also for $\hat{M}(z)$ when it is clear from the context that the Laplace transform is meant.
\
One important property of the memory functions is their boundary structure: if we define a boundary species as a subnetwork species that is involved in a reaction with a bulk species, then among the projected equations for subnetwork concentrations only those for boundary species contain memory terms. We also note that linear memory functions are only nonzero for boundary species influencing other boundary species. Similarly, only concentration products involving at least one boundary species appear in nonlinear memory terms \cite{Rubin2014d}.
The final term in \eqref{eq:projectedeqns}, $r(t)$, is what is known as the random force. It accounts for the fact that because of the interaction between subnetwork and bulk, the time evolution of the subnetwork observables cannot be closed. When using the projected equations to make predictions in numerical examples, we will drop the random force as there is no simple way of calculating it. The otherwise exact projected equations thus become an approximation, but one that in previous work~\cite{Rubin2014d} we showed to be remarkably accurate.
Our analysis starts from a given reaction network involving unary and binary protein reactions, and enzyme reactions described by Michaelis-Menten terms. We assume for simplicity that all enzyme reactions are reversible; the irreversible scenario can be obtained from this as the limiting case where the rate of dissociation into enzyme and product is much larger than the rate for formation of enzyme complex in the reverse direction.
The mass-action kinetics for each enzyme reaction is
\begin{equation}
\label{eq:MMdimensional}
\begin{split}
\frac{\partial}{\partial t}\prot{u} &=
-f_{ue,c} + \ldots\\
\frac{\partial}{\partial t}\prot{e} &=
-f_{ue,c}-f_{pe,c}\\
\frac{\partial}{\partial t}\prot{p} &=
-f_{pe,c} + \ldots
\end{split}
\end{equation}
where the dots indicate fluxes from other reactions. The reaction fluxes from substrate and enzyme to complex, and from product and enzyme to complex, read respectively (compare \eqref{eq:MMmassaction})
\begin{equation}
\label{eq:MMfluxes_orig}
\begin{split}
f_{ue,c} &=
\km{c}{u}{e}(-x_e^{\rm{tot}}+\y{e}+\prot{e}) +
\kp{u}{e}{c}
(\y{u}+\prot{u})
(\y{e} + \prot{e})\\
f_{pe,c} &=
\km{c}{p}{e}(-x_e^{\rm{tot}}+\y{e}+\prot{e}) +
\kp{p}{e}{c}
(\y{p}+\prot{p})
(\y{e} + \prot{e})
\end{split}
\end{equation}
Here we have used the enzyme conservation law \eqref{eq:MMconslaw} to eliminate the complex concentration via $\protein{c}=x_e^{\rm{tot}}-\protein{e}=
x_e^{\rm{tot}}-\y{e}-\prot{e}$. The enzyme steady state concentration $\y{e}$ can be found by requiring that in the steady state, where $\prot{e}=\prot{u}=\prot{p}=0$, the two fluxes must sum to zero to ensure $(\partial/\partial t)\prot{e}=0$. This gives
\begin{equation}
\label{eq:ye}
\y{e} = x_e^{\rm{tot}}\,
\frac{\km{c}{u}{e}
+ \km{c}{p}{e}}
{\km{c}{u}{e}
+ \km{c}{p}{e}
+ \kp{u}{e}{c}
\y{u}
+ \kp{p}{e}{c}
\y{p}}
\end{equation}
We now write the enzymatic reaction rate constants in terms of a fast rate
scale $\gamma$ as in \eqref{eq:reactionratesfast}, and similarly the steady
state enzyme concentration, which
from \eqref{eq:ye} must scale as the inverse
of $\gamma$, i.e.\ $\y{e}=\ytil{e}/\gamma$,
like the total enzyme concentration $x_e^{\rm{tot}}$. This gives
\begin{equation}
\label{eq:MMfluxes}
\begin{split}
f_{ue,c} &=
\kmtil{c}{u}{e}\gamma(-\bar{x}_e^{\rm{tot}}/\gamma+\ytil{e}/\gamma+\prot{e}) +
\kptil{u}{e}{c}\gamma
(\y{u}+\prot{u})
(\ytil{e}/\gamma + \prot{e})\\
f_{pe,c} &=
\kmtil{c}{p}{e}\gamma(-\bar{x}_e^{\rm{tot}}/\gamma+\ytil{e}/\gamma+\prot{e}) +
\kptil{p}{e}{c}\gamma
(\y{p}+\prot{p})
(\ytil{e}/\gamma + \prot{e})
\end{split}
\end{equation}
The scaled rate constants and steady state enzyme concentration are related to the Michaelis-Menten parameters as explained after \eqref{eq:MMreversible}, i.e.\
\begin{equation}
\label{eq:MMrev_constants_vs_rates}
\begin{split}
V_{\rm{u}} &= \frac{\kmtil{c}{p}{e}\ytil{e}(\kmtil{c}{u}{e}
+ \kmtil{c}{p}{e}
+ \kptil{u}{e}{c}
\y{u}
+ \kptil{p}{e}{c}
\y{p})}{\kmtil{c}{u}{e}
+ \kmtil{c}{p}{e}},
\quad K_{\rm{u}} = \frac{\kmtil{c}{p}{e}+\kmtil{c}{u}{e}}{\kptil{u}{e}{c}}\\
V_{\rm{p}} &= \frac{\kmtil{c}{u}{e}\ytil{e}(\kmtil{c}{u}{e}
+ \kmtil{c}{p}{e}
+ \kptil{u}{e}{c}
\y{u}
+ \kptil{p}{e}{c}
\y{p})}{\kmtil{c}{u}{e}
+ \kmtil{c}{p}{e}},\quad
K_{\rm{p}} = \frac{\kmtil{c}{p}{e}+\kmtil{c}{u}{e}}{\kptil{p}{e}{c}}
\end{split}
\end{equation}
In the above representation it is not obvious which terms have to be regarded as ``fast'' in the remainder of the analysis, and which as slow. We therefore switch to dimensionless concentration variables $\protnondim{i}=\prot{i}/\y{i}$. In terms of these we have
\begin{equation}
\label{eq:MMdimensionless}
\begin{split}
\frac{\partial}{\partial t}\protnondim{u} &=
-f_{ue,c}/\y{u} + \ldots\\
\frac{\partial}{\partial t}\protnondim{e} &=
-\gamma(f_{ue,c}+f_{pe,c})/\ytil{e} \\
\frac{\partial}{\partial t}\protnondim{p} &=
-f_{pe,c}/\y{p} + \ldots
\end{split}
\end{equation}
with enzymatic fluxes
\begin{equation}
\label{eq:MMfluxes_dimensionless}
\begin{split}
f_{ue,c} &=
\kmtil{c}{u}{e}\ytil{e} (-\bar{x}_e^{\rm{tot}}/\ytil{e}+1+\protnondim{e}) +
\kptil{u}{e}{c}\y{u}\ytil{e}
(1+\protnondim{u})
(1 + \protnondim{e})\\
f_{pe,c} &=
\kmtil{c}{p}{e}\ytil{e}(-\bar{x}_e^{\rm{tot}}/\ytil{e}+1+\protnondim{e}) +
\kptil{p}{e}{c}\y{p}\ytil{e}
(1+\protnondim{p})
(1 + \protnondim{e})
\end{split}
\end{equation}
Here one sees clearly that the enzyme evolution equation contains only fast terms that scale with $\gamma$, while the equations for substrate and product only contain slow terms. We will therefore use dimensionless concentrations throughout, and to lighten the notation we will in the following drop the tildes, as well as the bars indicating rescaling with $\gamma$. Note also that as at steady state ($\bm{\delta x}=0$) the total flux into or out of any molecular species must be zero, we can drop the constant pieces from $f_{ue,c}$ and
$f_{pe,c}$: they
have to cancel against each other in the equation for $\prot{e}$, and against the other steady state fluxes in the
equations for $\prot{u}$ and $\prot{p}$.
\begin{figure}[!ht]
\centering
\mbox{
\subfigure[\ subnetwork]{\includegraphics[scale=0.8]{enzymesinsub.eps}\label{fig:enzymesub}}
\quad
\subfigure[\ bulk enzymes away from the boundary]{\includegraphics[scale=0.8]{enzymesinbulk.eps}\label{fig:enzymebulk}}
\quad
\subfigure[\ bulk enzymes at the boundary]{\includegraphics[scale=0.8]{enzymesinbound.eps}\label{fig:enzymebound}}
}
\caption[Example enzyme networks]{The three possible locations of enzyme reactions with respect to subnetwork and bulk. (a) Substrate and product, and hence entire reaction, in the subnetwork; (b) substrate and product, and hence entire reaction, in the bulk; (c) substrate in the subnetwork and product in the bulk, in which case we assign also the enzyme and enzyme complex to the bulk.}
\label{fig:3enzymecases}
\end{figure}
Our task now is to take a mass-action (unary and binary) reaction system where every
enzyme reaction is represented as in
\eqref{eq:MMfluxes_dimensionless}, and to find closed form expressions
for the rate matrix and memory functions in the limit
$\gamma\to\infty$ where we know that this mass-action description
becomes identical to Michaelis-Menten dynamics. The effect of the
enzyme reactions depends on where they are located relative to subnetwork
and bulk, with three potentially distinct cases as shown in Fig.~\ref{fig:3enzymecases}.
If both substrate and product are in the subnetwork, also the enzyme will be located there and, as is clear from Fig.~\ref{fig:3enzymecases}a, away from the boundary of the subnetwork. We will therefore find ``fast'' equations of motion for such enzymes without any memory terms. These enzymes can therefore be kept explicitly in a first stage of our analysis, and eliminated in a second stage following the standard logic that leads to the Michaelis-Menten description.
If both substrate and product are within the bulk then the entire enzyme reaction takes place there (Fig.~\ref{fig:3enzymecases}b), contributing fast reaction rate constants to $\lblockb{B}{B}$. Accordingly we will find that such reactions only give contributions to the memory functions, not the rate matrix.
The third case is the one where the substrate is on the boundary and the product is in the bulk (or vice versa, but for reversible enzyme reactions the two cases are mathematically equivalent). We then choose to assign also the enzyme and the enzyme complex to the bulk, and refer to this situation as a (bulk) enzyme on the boundary. Reactions involving such enzymes, whose rate constants appear in $\lblockb{S}{B},\ \lblockb{B}{B}$ and $\lblockb{B}{S}$, will contribute fast terms in the memory functions that decay on a timescale of order $1/\gamma$. In the limit $\gamma\to \infty$ these terms become local in time and so turn into contributions to the rate matrix.
\section{Linearised Dynamics}
\label{sec:linMM}
In linearised dynamics we only consider terms in the mass-action kinetics up to linear order in $\bm{\delta x}$. The dimensionless scaled reaction equations for a reversible Michaelis-Menten reaction are then, from
\eqref{eq:MMdimensionless} and \eqref{eq:MMfluxes_dimensionless},
\begin{equation}
\label{eq:MMlin}
\begin{split}
\frac{\partial}{\partial t}\prot{u} &= -\km{c}{u}{e}(\y{e}/\y{u})\prot{e} - \kp{u}{e}{c}\y{e}(\prot{u} + \prot{e})+ \ldots\\
\frac{\partial}{\partial t}\prot{e} &= -\gamma[(\km{c}{u}{e}+\km{c}{e}{p})\prot{e} + \kp{u}{e}{c}\y{u}(\prot{u}+\prot{e})+ \kp{e}{p}{c}\y{p}(\prot{p} + \prot{e})]\\
\frac{\partial}{\partial t}\prot{p} &= -\km{c}{e}{p}(\y{e}/\y{p})\prot{e} - \kp{e}{p}{c}\y{e}(\prot{p} + \prot{e}) + \ldots
\end{split}
\end{equation}
Let us partition the matrix form of the adjoint Fokker-Planck matrix operator \eqref{eq:Lmatrix} so that the bulk species are split into fast and slow blocks. If e$'$ and e represent the collection of subnetwork and bulk enzymes respectively and s and b represent the other molecular species in the subnetwork and bulk, we partition as
\begin{equation}
\label{eq:LoperatorMMlin}
\bm{L} =\left(
\begin{array}{c|c}
\lblockb{S}{S}& \lblockb{S}{B}\\
\hline
\lblockb{B}{S}& \lblockb{B}{B}
\end{array}
\right)=
\left( \begin{array}{cc|cc}
\lblockb{s}{s}&\lblockb{s}{e'}&\lblockb{s}{b}&\lblockb{s}{e}\\
\lblockb{e'}{s}&\lblockb{e'}{e'}&\lblockb{e'}{b}&\lblockb{e'}{e}\\
\hline
\lblockb{b}{s}&\lblockb{b}{e'}&\lblockb{b}{b}&\lblockb{b}{e}\\
\lblockb{e}{s}&\lblockb{e}{e'}&\lblockb{e}{b}&\lblockb{e}{e}
\end{array}\right)
=\left( \begin{array}{c|cc}
m &w_1&f_1\\ \hline
w_2&w_3&f_2\\
w_4&w_5&f_3\\
\end{array}\right)
\end{equation}
where $w$ are slow terms and $f$ are fast terms; the top left block denoted $m$ contains a mixture of fast and slow terms. In writing the last equality above we have grouped s and e$'$ together; the resulting specific $3\times3$ block structure of slow and fast terms is one that we will find again in the case of the full nonlinear dynamics.
Note that because subnetwork enzymes only have interactions with subnetwork species (s and e$'$), $\lblockb{b}{e'}$ and $\lblockb{e}{e'}$ are zero. Similarly, because bulk proteins or enzymes do not interact with subnetwork enzymes, $\lblockb{e'}{b}$ and $\lblockb{e'}{e}$ vanish. This means that subnetwork enzymes do not feature at all in the calculation of the memory function \eqref{eq:memLT}, which makes intuitive sense. The vanishing of $\lblockb{b}{e'}$ and $\lblockb{e}{e'}$ is important also as these blocks contain rate constants for the time evolution of (subnetwork) enzymes, which by our construction scale with $\gamma$: if these blocks were nonzero, it would change the character of $w_2$ and $w_4$ from slow to fast.
To analyse the memory function \eqref{eq:mem} that results from \eqref{eq:LoperatorMMlin}, we note that $\lblockb{B}{B}$ has both slow and fast sub-blocks. As a result the memory function should in general have both slow contributions that decay on $O(1)$ timescales, and fast contributions that decay for time differences of $O(1/\gamma)$. As the memory function appears as a weight in an integral over the past \eqref{eq:projectedeqns}, the fast contributions only matter for $\gamma\to\infty$ if their amplitude is proportional to $\gamma$ so that the integral over all time differences remains finite. Accounting also for subleading terms in the amplitude dependence then suggests the following decomposition of the memory function:
\begin{equation}
\label{eq:memt}
M(\Delta t) = (\gamma M_f^0(\Delta\bar{t}) + M_f^1(\Delta\bar{t}) + \ldots) + (M_w^0(\Delta t) + \frac{1}{\gamma}M_w^1(\Delta t) + \ldots)
\end{equation}
where $\Delta\bar{t}=\gamma \Delta t$. In principle an arbitrary constant can be added to e.g.\ $M_f^1(\Delta\bar{t})$ and subtracted from $M_w^0(\Delta t)$; we fix this constant by requiring that all fast contributions decay to zero for large $\Delta\bar{t}$. Fig.~\ref{fig:fastmemory} shows an example of the above decomposition.
\begin{figure}[!ht]
\centering
\subfigure[]{\includegraphics[scale=0.8]{fastboundarymemplot.eps}}
\quad
\subfigure[]{\includegraphics[scale=0.8]{fastboundarymemplot2.eps}}
\caption[Memory function with enzyme on the boundary]{Example plots of a memory function when the enzyme is on the boundary, for different values of $\gamma$. (a) The memory has a slow and a fast part, with the fast part moving to shorter and shorter timescales as $\gamma$ increases.
(b)
Scaled plot of the fast part: if the fast memory is divided by $\gamma$ we see that a scaling plot is approached for large $\gamma$. This shows that the amplitude of the fast part of the memory grows proportionally to $\gamma$.}
\label{fig:fastmemory}
\end{figure}
The leading fast and slow contributions can now be extracted relatively simply from the Laplace transform of the memory function \eqref{eq:memLT}, which has the decomposition
\begin{equation}
\label{eq:memz}
\hat{M}(z) = (\hat{M}_f^0(\bar{z}) + \frac{1}{\gamma}\hat{M}_f^1(\bar{z}) + \ldots) + (\hat{M}_w^0(z) + \frac{1}{\gamma}\hat{M}_w^1(z)+\ldots)
\end{equation}
where $\bar{z}=z/\gamma$.
The leading fast term can be obtained by taking the limit
\begin{equation}
\label{eq:memfast}
\lim_{\gamma\to \infty}\hat{M}(z)\big|_{\bar{z}=\text{const}} = \hat{M}_f^0(\bar{z})
\end{equation}
because the subleading fast terms are down by powers of $1/\gamma$, and in the slow terms $z=\gamma\bar{z}\to\infty$ so that the Laplace transforms $\hat{M}_w^0(z)$ etc.\ vanish.
The leading slow part can then be found from a limit at constant $z$, namely
\begin{equation}
\label{eq:memslow}
\lim_{\gamma\to \infty}\left[\hat{M}(z)- \hat{M}_f^0(0)\right]\Big|_{z=\text{const}} = \hat{M}_w^0(z)
\end{equation}
We could have equivalently written $\hat{M}_f^0(\bar{z})$ inside the square brackets, as $\bar{z}=z/\gamma\to 0$ when $\gamma\to\infty$ at fixed $z$.
Using the method above, we can now find the fast and slow pieces of the memory function derived from the adjoint Fokker-Planck (matrix) operator in equation \eqref{eq:LoperatorMMlin}. From \eqref{eq:memLT} this memory function reads
\begin{equation}
\label{eq:MMmemLTorig}
\hat{M}(z) = \begin{pmatrix}
w_1 & f_1
\end{pmatrix}
\begin{pmatrix}
z - w_3 & - f_2\\
- w_5 & z - f_3
\end{pmatrix}^{-1}
\begin{pmatrix}
w_2 \\ w_4
\end{pmatrix}
\end{equation}
and after working out the inverse using standard block inversion identities
and simplifying we obtain
\begin{equation}
\label{eq:MMmemLT}
\begin{split}
\hat{M}(z)&= \left(w_1 -f_1(z-f_3)^{-1}(- w_5)\right)\left(z-w_3 + f_2(z-f_3)^{-1}w_5\right)^{-1}\left(w_2+f_2(z-f_3)^{-1}w_4\right)\\
&\quad + f_1(z-f_3)^{-1}w_4
\end{split}
\end{equation}
If we now write all fast blocks as $f=\gamma\bar{f}$, we can see that application of
\eqref{eq:memfast} identifies the
fast part of the memory function as
\begin{equation}
\label{eq:memzfast}
\hat{M}_f^0(\bar{z}) = \bar{f}_1(\bar{z}-\bar{f}_3)^{-1}w_4
\end{equation}
while \eqref{eq:memslow} gives for
the slow part
\begin{equation}
\label{eq:memzslow}
\hat{M}_w^0(z) = (w_1 -\bar{f}_1(\bar{f}_3)^{-1} w_5)(z-w_3 - \bar{f}_2(\bar{f}_3)^{-1}w_5)^{-1}(w_2-\bar{f}_2(\bar{f}_3)^{-1}w_4)
\end{equation}
Here we have used the fact that in the combination $z-\gamma\bar{f}_3$, the first term can be neglected when $\gamma\to\infty$ at constant $z$.
The fast part of the memory decays on an ever shorter timescale $\sim 1/\gamma$ as $\gamma$ increases. In the limit, and when it is used inside a memory function integral, it becomes equivalent to a delta function $\delta(\Delta t)$ multiplied by the area under the fast piece, which is just $ \hat{M}_f^0(0) = -\bar{f}_1(\bar{f}_3)^{-1}w_4$. The rate matrix that we obtain from \eqref{eq:LoperatorMMlin} for $\gamma\to\infty$ is therefore
\begin{equation}
\bm{\Omega} = m -\bar{f}_1(\bar{f}_3)^{-1}w_4
\label{eq:Omega_effective}
\end{equation}
while the memory function in the same limit is given by \eqref{eq:memzslow}. We can now compare to the rate matrix and memory function that would result from an adjoint Fokker-Planck operator like \eqref{eq:LoperatorMMlin} without the fast bulk variables e, i.e.\
\begin{equation}
\label{eq:L_no_e}
\bm{L}_{\setminus\rm{\enzyme}}=\left( \begin{array}{c|c}
m &w_1\\ \hline
w_2&w_3
\end{array}\right)
\end{equation}
This would give $\Omega=m$ and $\hat{M}(z)=w_1(z-w_3)^{-1}w_2$. Looking at \eqref{eq:memzslow} and \eqref{eq:Omega_effective}, we conclude that the large-$\gamma$ limit gives results that can equivalently be obtained by using an adjoint Fokker-Planck matrix without the fast bulk variables that is modified from $\bm{L}_{\setminus\rm{\enzyme}}$ to $\bm{L}_{\rm eff}=\bm{L}_{\setminus\rm{\enzyme}}+\Delta\Le$, where
\begin{equation}
\label{eq:dL}
\Delta\Le=-\left( \begin{array}{c|c}
\bar{f}_1(\bar{f}_3)^{-1} w_4 & \bar{f}_1(\bar{f}_3)^{-1} w_5 \\ \hline
\bar{f}_2(\bar{f}_3)^{-1}w_4 & \bar{f}_2(\bar{f}_3)^{-1}w_5
\end{array}\right)
\end{equation}
This is the key result of our first stage of elimination, which has removed the fast bulk variables.
\subsection{Bulk enzyme elimination as quasi-steady state method}
\label{sec:simplifyelim}
Before moving on to the second stage of also eliminating the subnetwork enzymes, we pause briefly to give a simpler form of our last result. While \eqref{eq:dL} gives a closed form for the effective $L$-matrix $\bm{L}_{\setminus\rm{\enzyme}}+\Delta\Le$ we obtain after eliminating the bulk enzymes, this form is not very intuitive. We show next that there is a much simpler statement of the result, namely that $\bm{L}_{\rm eff}=\bm{L}_{\setminus\rm{\enzyme}}+\Delta\Le$ can be obtained by treating all bulk enzymes as in quasi-steady state with the other molecular species.
To see this, we reinstate in the generic $3\times 3$ block structure of \eqref{eq:LoperatorMMlin} the specific notation for the linearised dynamics considered here, i.e.\
\begin{equation}
\bm{L}
=\left( \begin{array}{c|cc}
m &w_1&f_1\\ \hline
w_2&w_3&f_2\\
w_4&w_5&f_3\\
\end{array}\right)
=\left( \begin{array}{c|cc}
\lblockb{S}{S} & \lblockb{S}{b}& \lblockb{S}{e}\\ \hline
\lblockb{b}{S} & \lblockb{b}{b}& \lblockb{b}{e}\\
\lblockb{e}{S} & \lblockb{e}{b}& \lblockb{e}{e}
\end{array}\right)
\end{equation}
where, as before, S collects all subnetwork variables, i.e.\ subnetwork proteins s and subnetwork enzymes e$'$.
The dynamics of the system can then be written as
\begin{equation}
\begin{split}
\frac{\partial}{\partial t}\bm{\delta x}^{\text{S}^{\rm{T}}} &=
\bm{\delta x}^{\text{S}^{\rm{T}}} \lblockb{S}{S} + \bm{\delta x}^{\text{b}^{\rm{T}}} \lblockb{b}{S} + \bm{\delta x}^{\text{e}^{\rm{T}}} \lblockb{e}{S}\\
\frac{\partial}{\partial t}\bm{\delta x}^{\text{b}^{\rm{T}}} &=
\bm{\delta x}^{\text{S}^{\rm{T}}} \lblockb{S}{b} + \bm{\delta x}^{\text{b}^{\rm{T}}} \lblockb{b}{b} + \bm{\delta x}^{\text{e}^{\rm{T}}} \lblockb{e}{b}\\
\frac{\partial}{\partial t}\bm{\delta x}^{\text{e}^{\rm{T}}} &=
\bm{\delta x}^{\text{S}^{\rm{T}}} \lblockb{S}{e} + \bm{\delta x}^{\text{b}^{\rm{T}}} \lblockb{b}{e} + \bm{\delta x}^{\text{e}^{\rm{T}}} \lblockb{e}{e}
\end{split}
\end{equation}
If we now impose a quasi-steady state condition for the bulk enzymes $\bm{\delta x}^\text{e}$, this gives
\begin{equation}
\bm{\delta x}^{\text{e}^{\rm{T}}} = -\left(\bm{\delta x}^{\text{S}^{\rm{T}}} \lblockb{S}{e} + \bm{\delta x}^{\text{b}^{\rm{T}}} \lblockb{b}{e}\right) ({\lblockb{e}{e}})^{-1}
\end{equation}
Substituting this back into the equations of motion for the subnetwork species and the bulk proteins gives
\begin{equation}
\begin{split}
\frac{\partial}{\partial t}\bm{\delta x}^{\text{S}^{\rm{T}}} &=
\bm{\delta x}^{\text{S}^{\rm{T}}} \left(\lblockb{S}{S}-\lblockb{S}{e} ({\lblockb{e}{e}})^{-1}\lblockb{e}{S}\right) + \bm{\delta x}^{\text{b}^{\rm{T}}}\left( \lblockb{b}{S}-\lblockb{b}{e}({\lblockb{e}{e}})^{-1}\lblockb{e}{S}\right) \\
\frac{\partial}{\partial t}\bm{\delta x}^{\text{b}^{\rm{T}}} &=
\bm{\delta x}^{\text{S}^{\rm{T}}} \left(\lblockb{S}{b}- \lblockb{S}{e}({\lblockb{e}{e}})^{-1}\lblockb{e}{b}\right) + \bm{\delta x}^{\text{b}^{\rm{T}}}\left( \lblockb{b}{b} - \lblockb{b}{e} ({\lblockb{e}{e}})^{-1}\lblockb{e}{b}\right)
\end{split}
\end{equation}
This is exactly the dynamics that is defined by the effective $L$-matrix $\bm{L}_{\rm eff}=\bm{L}_{\setminus\rm{\enzyme}}+\Delta\Le$ derived above (see \eqref{eq:L_no_e} and \eqref{eq:dL}), hence proving our claim that this matrix can be constructed by imposing a quasi-steady state condition for the bulk enzymes.
\subsection{Michaelis-Menten terms as effective unary reactions}
\label{sec:linbulk}
The above bulk enzyme elimination can be carried out in closed form.
Each bulk enzyme can be eliminated by setting the time derivative of its concentration to zero. Here and throughout we assume that each enzyme only catalyses one reaction. (In the matrix formulation, this implies that the block $\lblockb{e}{e}$ is diagonal.) We then find for the effective dynamical equations of any substrate and product the following form:
\begin{equation}
\label{eq:MMlin_eff}
\begin{split}
\frac{\partial}{\partial t}\prot{u} &= -\lambda_{up}\prot{u} + \lambda_{\productu}(\y{p}/\y{u})\prot{p}
+ \ldots\\
\frac{\partial}{\partial t}\prot{p} &= -\lambda_{\productu}\prot{p} + \lambda_{up}(\y{u}/\y{p})\prot{u} + \ldots
\end{split}
\end{equation}
where
\begin{equation}
\begin{split}
\label{eq:unary_rates}
\lambda_{up} &= \frac{\kp{u}{e}{c}\y{e}\left(\km{c}{p}{e}+\kp{p}{e}{c}\y{p}\right)}{\km{c}{u}{e}+\km{c}{p}{e}+ \kp{u}{e}{c}\y{u}+\kp{p}{e}{c}\y{p}}\\
\lambda_{\productu} &= \frac{\kp{p}{e}{c}\y{e}\left(\km{c}{u}{e}+\kp{u}{e}{c}\y{u}\right)}{\km{c}{u}{e}+\km{c}{p}{e}+ \kp{u}{e}{c}\y{u}+\kp{p}{e}{c}\y{p}}
\end{split}
\end{equation}
are the rates for effective unary reactions converting substrate to product and back, respectively. The factors of $(\y{p}/\y{u})$ in \eqref{eq:MMlin_eff} arise because we are using dimensionless concentration variables.
The bulk enzyme elimination thus has the simple effect of replacing all bulk Michaelis-Menten reactions by unary conversion reactions with constant rates. From \eqref{eq:MMrev_constants_vs_rates} one sees that these effective rates can be expressed directly in terms of the Michaelis-Menten parameters, as
\begin{equation}
\label{eq:unary_ratesMM}
\begin{split}
\lambda_{up} &= \frac{V_{\rm{u}}/K_{\rm{u}} + (V_{\rm{u}}+V_{\rm{p}})/K_{\rm{u}}(\y{p}/K_{\rm{p}})}{\left(1 + \y{u}/K_{\rm{u}} + \y{p}/K_{\rm{p}}\right)^2}\\
\lambda_{\productu} &= \frac{V_{\rm{p}}/K_{\rm{p}} + (V_{\rm{u}}+V_{\rm{p}})/K_{\rm{p}}(\y{u}/K_{\rm{u}})}{\left(1 + \y{u}/K_{\rm{u}} + \y{p}/K_{\rm{p}}\right)^2}
\end{split}
\end{equation}
Comparing with \eqref{eq:MMreactionratereverse} shows that the rates are obtained by linearising the Michaelis-Menten reaction flux around the steady state concentrations of substrate and product. This is the closed-form procedure for bulk enzyme elimination we were after: it only requires as input the Michaelis-Menten parameters of the original network, and its steady state.
Note that the above discussion includes enzyme reactions both entirely in the bulk, or on the boundary of subnetwork and bulk (cf.\ Fig.~\ref{fig:3enzymecases}b and c). The only difference between these two cases is that for the latter group of enzymes, the effective unary reactions we have derived are between a subnetwork boundary species and a bulk species and so will contribute to the rate matrix, while for enzyme reactions entirely in the bulk the effective reactions only affect the memory function.
\subsection{Elimination of subnetwork enzymes}
\label{sec:linsub}
So far we have described how bulk enzymes can be eliminated, replacing them by effective unary conversion reactions. This allows the rate matrix and memory functions to be calculated from an effective $L$-matrix $\bm{L}_{\rm eff}=\bm{L}_{\setminus\rm{\enzyme}}+\Delta\Le$. These quantities determine the projected equations of motion for the subnetwork proteins s and the subnetwork enzymes e$'$.
The second, final stage of the elimination procedure is now to eliminate the subnetwork enzymes. Elimination of the bulk enzymes does not affect the equations of motion for the subnetwork enzymes, i.e.\ these do not acquire memory terms nor are the local-in-time terms changed.
Looking at the general formula \eqref{eq:mem} for the memory function, read in terms of the effective $L$-matrix, then confirms that the equations for the e$'$ species do not contain memory terms. This makes sense: the subnetwork enzymes are not boundary species to start with, and this is not changed by the effective unary reactions from bulk enzymes.
The projected equation for each subnetwork enzyme thus looks exactly as the original equation in \eqref{eq:MMdimensionless}. Because of the fast rate scale $\gamma$ in this, in the limit $\gamma\to \infty$ each subnetwork enzyme will be in quasi-steady state with its substrate and product. Substituting the quasi-steady state enzyme concentration into the equations for substrate and product then gives again effective unary conversion reactions, with rates as given in \eqref{eq:unary_ratesMM} for the case of bulk enzymes.
\subsection{Summary of enzyme elimination procedure for linearised dynamics}
The final procedure we have arrived at for constructing projected equations for reaction systems with Michaelis-Menten terms, within linearised dynamics, is remarkably simple: replace each Michaelis-Menten term, whether in the subnetwork, the bulk or on the boundary, by its linearisation around the steady state. This gives effective rates for unary conversion reactions among each substrate-product pair (see \eqref{eq:unary_rates}).
\section{Nonlinear Dynamics}
\label{sec:nonlinMM}
We now want to extend the above approach of eliminating enzymes from the
projected equations to the full nonlinear dynamics. Directly transplanting the results from the linearised dynamics is not possible, however: if we use the quasi-steady state assumption for the enzymes as in Section~\ref{sec:linMM}, then we get back the full Michaelis-Menten nonlinearities. As these go beyond second order in $\prot{}$, they cannot be used directly in our construction of the projected equations, which starts from reaction equations with only linear and quadratic terms as appropriate for a mass action description of unary and binary reactions.
We take as our starting point the nonlinear $L$-matrix, as shown in equation \eqref{eq:Lmatrix}, but subdivide this into smaller blocks below in order to single out contributions from enzymes.
Focussing though for now just on the distinction between linear and quadratic observables, we have two new kinds of entries. Firstly, mixed linear-quadratic elements as contained in e.g.\ $\lblockb{ss}{s}$: these are coefficients of quadratic terms in equations of motion for the concentrations (linear observables), so can be read off directly from the mass action equations. The quadratic-quadratic elements as in $\lblockb{ss}{ss}$ are coefficients from equations of motion of concentration products. For a generic product $\prot{ij}\equiv \prot{i}\prot{j}$ they are of the form
\begin{equation}
\label{eq:product_diff}
\frac{\partial}{\partial t}\prot{i}\prot{j} = \prot{j}\frac{\partial}{\partial t}\prot{i} + \prot{i}\frac{\partial}{\partial t}\prot{j}
\end{equation}
Because we are only considering terms up to quadratic order on the r.h.s., we need to insert only the linearised equations of motion for $(\partial \prot{i}/\partial t)$ and $(\partial \prot{j}/\partial t)$. All quadratic-quadratic elements of the $L$-matrix are therefore ``inherited'' from the linearised dynamics. In particular, the above structure of the equations of motion for quadratic observables means that the time evolution of any product containing at least one enzyme factor will contain fast terms.
For the purpose of eliminating the fast degrees of freedom, the nonlinear $L$-matrix can be split into four blocks mirroring the structure of the linear $L$-matrix in equation \eqref{eq:LoperatorMMlin}, viz.
\begin{equation}
\label{eq:LoperatorMMnonlin}
\bm{L}=\left(
\begin{array}{c|c}
\lblockb{S}{S}& \lblockb{S}{B}\\
\hline
\lblockb{B}{S}& \lblockb{B}{B}
\end{array}
\right)=
\left( \begin{array}{cc|cc}
\lblockb{\tilde{s}}{\tilde{s}}&\lblockb{\tilde{s}}{\tilde{e}'}&\lblockb{\tilde{s}}{\tilde{b}}&\lblockb{\tilde{s}}{\tilde{e}}\\
\lblockb{\tilde{e}'}{\tilde{s}}&\lblockb{\tilde{e}'}{\tilde{e}'}&\lblockb{\tilde{e}'}{\tilde{b}}&\lblockb{\tilde{e}'}{\tilde{e}}\\
\hline
\lblockb{\tilde{b}}{\tilde{s}}&\lblockb{\tilde{b}}{\tilde{e}'}&\lblockb{\tilde{b}}{\tilde{b}}&\lblockb{\tilde{b}}{\tilde{e}}\\
\lblockb{\tilde{e}}{\tilde{s}}&\lblockb{\tilde{e}}{\tilde{e}'}&\lblockb{\tilde{e}}{\tilde{b}}&\lblockb{\tilde{e}}{\tilde{e}}
\end{array}\right)
\end{equation}
The blocks are defined so that \~s contains the observables s and ss which have no bulk or fast factors, \~b contains the observables b, sb and bb which have at least one bulk factor but no fast factors, \~e consists of e, se, be, be$'$, ee and ee$'$ which contain at least one bulk factor and one fast factor - where the fast and bulk factor can be identical - and \~e$'$ contains e$'$, se$'$ and e$'$e$'$ where there are no bulk factors but at least one fast factor. Therefore the subnetwork block S consists of slow and fast blocks in the form of \~s and \~e$'$ respectively and similarly the bulk block B contains slow and fast contributions from \~b and \~e, respectively.
The block structure of the $L$-matrices for the linearised and nonlinear dynamics is therefore the same; however, there are some differences. The block \~e\~e contains some slow as well as fast entries, but the slow entries can be neglected in comparison in the large $\gamma$ limit.
The \~e\~e$'$ block is not zero as in the linear case due to the fact that the equation of motion for se$'$ involves products of the form be$'$ and ee$'$; importantly for our reasoning below these terms are slow, however, because of the time evolution of s in the se$'$ products.
Similarly the \~e$'$\~e block is not zero because the equation of motion for be$'$ involves slow se$'$ contributions; also in the equation for ee$'$ there are fast contributions from se$'$. The blocks \~b\~e$'$ and \~e$'$\~b remain zero, on the other hand.
We can go back to the same 3 $\times$ 3 structure for the $L$-matrix as in the linear case, by partitioning into blocks S (\~s and \~e$'$), \~b and \~e.
This 3 $\times$ 3 matrix can then be split into fast and slow blocks as in equation \eqref{eq:LoperatorMMlin}:
\begin{equation}
\def\arraystretch{1.3}
\bm{L} =\left( \begin{array}{c|cc}
\lblockb{S}{S} &\lblockb{S}{\tilde{b}}&\lblockb{S}{\tilde{e}}\\ \hline
\lblockb{\tilde{b}}{S} &\lblockb{\tilde{b}}{\tilde{b}}&\lblockb{\tilde{b}}{\tilde{e}}\\
\lblockb{\tilde{e}}{S} &\lblockb{\tilde{e}}{\tilde{b}}&\lblockb{\tilde{e}}{\tilde{e}}\\
\end{array}\right)
=\left( \begin{array}{c|cc}
m &w_1&f_1\\ \hline
w_2&w_3&f_2\\
w_4&w_5&f_3\\
\end{array}\right)
\end{equation}
Using the method of Sec.~\ref{sec:linMM} we can then find the slow and fast parts of the memory expressed in terms of these blocks, using exactly the same formulae as shown for the linearised dynamics in equation \eqref{eq:MMmemLT}. As before the result can be thought of as arising from a
modified $L$-matrix where all the fast bulk species and products contained in \~e are eliminated:
\begin{equation}
\label{eq:Leff}
\bm{L}_{\rm eff}=
\left( \begin{array}{c|c}
m-\bar{f}_1(\bar{f}_3)^{-1} w_4 & w_1-\bar{f}_1(\bar{f}_3)^{-1} w_5 \\ \hline
w_2-\bar{f}_2(\bar{f}_3)^{-1}w_4 & w_3-\bar{f}_2(\bar{f}_3)^{-1}w_5
\end{array}\right)
\end{equation}
The form of $\bm{L}_{\rm eff}$ can be derived by eliminating the fast observables in \~e directly, by assuming that they are in steady state with respect to the fast contributions from their equations of motion. We turn next to the task of actually carrying out this elimination, which is more involved than in the linear case.
\subsection{Elimination of fast bulk variables}
We can simplify matters somewhat by noting that in order to find the correct rate matrix and memory function with enzymes eliminated, we require the equations of motion with enzymes eliminated -- and hence the relevant columns of $\bm{L}_{\rm eff}$ -- for the bulk observables \~b and the \emph{linear} subnetwork observables s and e$'$. We only need the linear observables s and e$'$ because to calculate the rate matrix and memory function one never requires blocks of the form $\lblockb{\cdot}{ss}$. (The intuition is that product observables can be replaced by actual products in the low noise limit we are considering; see~\cite{Rubin2014d} for details.) A further simplification comes from the fact that the original equations of motion for \~b, s and e$'$ do not depend on ee or ee$'$ and therefore we do not need to consider these observables further: we can focus on how to eliminate the remaining fast observables se, be, be$'$ and e.
We consider first the product observables se, be and be$'$.
The equation of motion for a generic observable of type se, i.e.\ a product of the concentration of a subnetwork species $s$ with that of an enzyme $e$, reads
\begin{equation}
\frac{\partial}{\partial t}\prot{se} = \frac{\partial}{\partial t}\prot{s}\prot{e} = \prot{s}\frac{\partial}{\partial t}\prot{e}+\prot{e}\frac{\partial}{\partial t}\prot{s} = \prot{s}\frac{\partial}{\partial t}\prot{e}+ \text{slow}
\end{equation}
where we have used that $(\partial/\partial t)\prot{s}$ only contains slow terms. In finding the solution of the quasi-steady state condition $(\partial/\partial t)\prot{se}=0$, these slow terms can be neglected compared to the fast terms from $(\partial /\partial t)\prot{e}$. Writing the latter in the form
\begin{equation}
\frac{\partial}{\partial t}\prot{e} = A \prot{u}+ B\prot{p}-C\prot{e}
\end{equation}
allows us to write the leading (fast) terms in the equation for the subnetwork-enzyme product as
\begin{equation}
\frac{\partial}{\partial t}\prot{se}= \prot{s}\frac{\partial}{\partial t}\prot{e} = A\,\prot{su}+B\,\prot{sp}-C\,\prot{se}
\end{equation}
Setting this to zero shows that the quasi-steady state solution is
\begin{equation}
\label{eq:se_elimination}
\prot{se} = (A/C)\prot{su} + (B/C)\prot{sp}
\end{equation}
Comparing with the (linear) quasi-steady state solution for the enzyme concentration itself, which is $\prot{e} = (A/C)\prot{u} + (B/C)\prot{p}$, we arrive at a simple product elimination rule: products of the form se are eliminated by using the linear elimination of the enzyme, multiplying by a factor of $\prot{s}$, and then identifying $\prot{su}=\prot{s}\prot{u}$ and $\prot{sp}=\prot{s}\prot{p}$. It is straightforward to check that the same rule applies to the elimination of observables of type be and be$'$.
The only remaining fast observables that we need to eliminate are the linear bulk enzyme concentrations. Their equations of motion from \eqref{eq:MMdimensionless} and \eqref{eq:MMfluxes_dimensionless} are
\begin{equation}
\begin{split}
\frac{\partial}{\partial t}\prot{e} &= -(\gamma/\y{e})\left[
\km{c}{u}{e}\y{e} \prot{e} +
\kp{u}{e}{c}\y{u}\y{e}
(\prot{u} + \prot{e} + \prot{ue})\right.\\
&\quad+
\left.
\km{c}{p}{e}\y{e}\prot{e} +
\kp{p}{e}{c}\y{p}\y{e}
(\prot{p} + \prot{e} + \prot{pe})
\right]
\end{split}
\end{equation}
These contain product variables of the form se and be, which can now be eliminated using the method above, giving expressions in the form of equation \eqref{eq:se_elimination}. Substituting these and solving the quasi-steady state condition $(\partial/\partial t)\prot{e}=0$ then gives
\begin{equation}
\begin{split}
\prot{e} &= -\frac{1}{(\km{c}{u}{e} + \kp{u}{e}{c}\y{u} + \km{c}{p}{e} + \kp{p}{e}{c}\y{p})^2}
\left[ \kp{u}{e}{c}\y{u}\Big((\km{c}{u}{e} + \kp{u}{e}{c}\y{u}
+ \km{c}{p}{e} + \kp{p}{e}{c}\y{p})\prot{u}\right.\\
&\quad + \kp{u}{e}{c}\y{u}\prot{u\substrate} + \kp{p}{e}{c}\y{p}\prot{up}\Big)\\
&\quad +\left. \kp{p}{e}{c}\y{p}\Big((\km{c}{u}{e} + \kp{u}{e}{c}\y{u} + \km{c}{p}{e} + \kp{p}{e}{c}\y{p})\prot{p}
+ \kp{u}{e}{c}\y{u}\prot{up} + \kp{p}{e}{c}\y{p}\prot{p\product}\Big)\right]
\end{split}
\end{equation}
We can now compare to the standard Michaelis-Menten elimination of the bulk enzyme, which treats products like $\prot{ue}$ not as separate observables but identifies them with $\prot{u}\prot{e}$ and then solves $(\partial/\partial t)\prot{e} = 0$. It is straightforward to check that our above elimination formula is just this Michaelis-Menten result expanded to quadratic order. We will therefore call this result ``quadratic quasi-steady state elimination''.
The result of the first stage of elimination is therefore that we can construct equations of motion for s, e$'$, b, sb and bb by quadratic quasi-steady state elimination of the bulk enzymes. The full quadratic elimination is not needed for all observables as the equations of motion for sb and bb only contain quadratic observables: in these we use linear enzyme elimination to replace products as explained above.
We can now write down explicitly what the effective contributions to the equations of motion for a substrate and product are. One starts from \eqref{eq:MMdimensionless} and \eqref{eq:MMfluxes_dimensionless} again,
\begin{equation}
\frac{\partial}{\partial t}\prot{u} = -(1/\y{u})\left[
\km{c}{u}{e}\y{e} \prot{e}
+ \kp{u}{e}{c}\y{u}\y{e}
(\prot{u} + \prot{e} + \prot{ue})
\right] + \ldots
\end{equation}
and substitutes in the elimination formulae for $\prot{e}$ and $\prot{ue}$. After a little algebra, the
effective equation of motion for the substrate, and analogously the product, can be written in terms of unary reactions, as was the case for the linearised dynamics:
\begin{equation}
\label{eq:MMnonlin_eff}
\begin{split}
\frac{\partial}{\partial t}\prot{u} &= -\hat{\lambda}_{up}\prot{u} + \hat{\lambda}_{\productu}(\y{p}/\y{u})\prot{p}
+ \ldots\\
\frac{\partial}{\partial t}\prot{p} &= -\hat{\lambda}_{\productu}\prot{p} + \hat{\lambda}_{up}(\y{u}/\y{p})\prot{u} + \ldots
\end{split}
\end{equation}
The difference is that the reaction rates $\hat{\lambda}_{up}$ and $\hat{\lambda}_{\productu}$ are now linearly dependent on substrate and product concentrations. In terms of the (constant) reaction rates $\lambda_{\productu}$ and $\lambda_{up}$ defined in \eqref{eq:unary_ratesMM}, we can write this concentration dependence in the simple form
\begin{equation}
\label{eq:prod_ratesMM}
\begin{split}
\hat{\lambda}_{up} &= \lambda_{up}\left(1-\frac{\prot{u}\y{u}/K_{\rm{u}} + \prot{p}\y{p}/K_{\rm{p}}}{1 + \y{u}/K_{\rm{u}} + \y{p}/K_{\rm{p}}}\right)\\
\hat{\lambda}_{\productu} &= \lambda_{\productu}\left(1- \frac{\prot{u}\y{u}/K_{\rm{u}} + \prot{p}\y{p}/K_{\rm{p}}}{1 + \y{u}/K_{\rm{u}} + \y{p}/K_{\rm{p}}}\right)
\end{split}
\end{equation}
Representing every Michaelis-Menten term in the bulk or on the boundary in this form, we thus obtain a set of equations from which all bulk enzymes have been eliminated. The coefficients in these equations then define the effective $L$-matrix $\bm{L}_{\rm eff}$ (or more precisely those columns of it that we use to obtain the rate and memory matrix).
\subsection{Elimination of fast subnetwork observables}
The result of the first stage of elimination is a rate matrix and memory matrix for the projected equations of the subnetwork variables s and e$'$. The second and final stage in the elimination of fast observables is now to remove the subnetwork enzymes e$'$.
This second stage is relatively simple because the projected equations of motion for e$'$ observables are in fact just the original mass action equations. This is so because the equations of motion for e$'$ observables only contain fast contributions from e$'$ and se$'$, so do not couple to any variables that were eliminated in the first stage. Furthermore, e$'$ observables are interior subnetwork species and do not couple to the slow bulk variables. Therefore their equations of motion cannot acquire memory terms. Eliminating subnetwork enzymes is then trivial: in the limit $\gamma\to\infty$ the quasi-steady state assumption will become exact for them. Solving this gives exactly the standard Michaelis-Menten expression for the enzyme concentrations.
The final question is then where the subnetwork enzymes e$'$ feature in the projected equations for slow subnetwork species s. They appear in the rate matrix and substituting them there simply produces the usual Michaelis-Menten terms, in their full nonlinear form rather than expanded to second order as for bulk enzymes.
In the memory function e$'$ cannot appear linearly as subnetwork enzymes are not on the boundary. Such enzymes could then only appear in the memory via se$'$ terms. For this to occur would require nonzero entries in the blocks $\bm{L}_{\rm eff}^{\rm{se'},\rm{b}}$, $\bm{L}_{\rm eff}^{\rm{se'},\rm{sb}}$ or $\bm{L}_{\rm eff}^{\rm{se'},\rm{bb}}$, which are the relevant pieces of the leftmost factor $\lblock{S}{B}$ in \eqref{eq:memLT}. The block $\bm{L}_{\rm eff}^{\rm{se'},\rm{b}}$ contains the nonlinear se$'$ contributions in the equation of motion for a bulk observable; however such an equation cannot involve any se$'$ products because subnetwork enzymes do not interact with the bulk. Therefore there are no contributions to $\bm{L}_{\rm eff}^{\rm{se'},\rm{b}}$. Similarly $\bm{L}_{\rm eff}^{\rm{se'},\rm{bb}}$ must be zero as from the product rule of differentiation in \eqref{eq:product_diff} there would have to be a shared species in the first and second product index in order to obtain a nonzero contribution. For the block $\bm{L}_{\rm eff}^{\rm{se'},\rm{sb}}$ to have nonzero entries there must be a shared index, which in this case must be s. This means that nonzero elements could come only from the linearised equation of motion for the b observable; however, there are no contributions from e$'$ in the equations of motion for bulk observables and therefore $\bm{L}_{\rm eff}^{\rm{se'},\rm{sb}}$ must also be zero. In summary, this means that there are no contributions from subnetwork enzymes to any memory function.
\subsection{Summary of enzyme elimination procedure for nonlinear dynamics}
\label{subsec:summary_nonlin}
The final procedure we have arrived at for constructing projected equations for reaction systems with Michaelis-Menten terms, for nonlinear dynamics, can be split into two simple steps. The first is to construct the reduced $L$-matrix $\bm{L}_{\rm eff}$ by expanding all the Michaelis-Menten terms to second order around the steady state. From this matrix we are then able to calculate the rate matrix and memory function. Once this is done, we reinstate the full nonlinear form of the Michaelis-Menten terms from enzymes in the subnetwork.
We note as an aside that as an alternative to the above method, one could assign the subnetwork enzymes e$'$ to the bulk in the projection, on the grounds that one does not want to track them explicitly in the final projected equations. This would have the advantage of removing the need for a second stage of fast variable elimination. On the other hand, even though the subnetwork enzymes are fast variables, assigning them to the bulk and projecting remains an approximation because it would result in an additional contribution to the random force that we would discard in using the projected equations in practice. More intuitively, the approximation arises from the fact that in (\ref{eq:Lmatrixeqn}) we are neglecting cubic terms in $\delta x$. We therefore prefer to treat the subnetwork enzymes explicitly and to eliminate them separately in the second stage of the process. As this only requires re-instating the original Michaelis-Menten terms in the subnetwork, it is well worth doing in order to exclude one source of potential approximation error.
\section{Numerical Comparisons}
\label{sec:MMcomparison}
We test our results on a model of the signalling network of epidermal growth factor receptor (EGFR) developed by \citet{Kholodenko99}. It is a well-studied network and contains a number of subnetworks. The EGFR network model as shown in Fig.~\ref{fig:MMegfr} has two subnetwork Michaelis-Menten reactions and one bulk Michaelis-Menten reaction. We choose the bulk to be the protein Shc and all complexes that include Shc.
In this section our aim is to compare two versions of the projected equations: the ones obtained by adding enzyme species explicitly with finite but large enzyme reaction rate constants, and the ones we get by eliminating enzymes in closed form as explained above. We will compare both projected descriptions with the
dynamics of the full EGFR network \cite{Kholodenko99}, i.e.\ tracking explicitly the bulk degrees of freedom, to see which of the two better represents the true time courses. The bulk is assumed to be at steady state initially. For the subnetwork we choose initial conditions that maximize nonlinear effects but still give the desired steady state (i.e.\ have the appropriate value for all conserved quantities.)
We use $\delta=\left(\sum_{i=1}^N [\prot{i}(0)]^2/N\right)^{1/2}$ to quantify the initial deviation from the steady state, where $N$ is the number
of subnetwork species.
To assess the accuracy of our two approximations we define
\begin{equation}
\label{eq:errormeasure}
\Delta=\frac{1}{T}\int_0^Tdt'\,\frac{1}{N}\sum_{i=1}^N|\prot{i}(t)-\delta\hat{x}_i(t)|
\end{equation}
where $\delta\hat{x}$ is our approximation to the subnetwork time course and the total time interval for the error measurement is chosen such as to capture the transient dynamics in the approach to the overall system steady state, with $T=150$s.
To apply the projection method we analytically obtain all the relevant parts of $\bm{L}$ using the reaction rate constants and steady state values \cite{Rubin2014d}. Applying the method summarized in Sec.~\ref{subsec:summary_nonlin}, we then find the projected equations: a set of dynamical equations for the subnetwork species involving memory terms. These integro-differential equations can in principle be integrated directly, but as numerical algorithms for such equations are not in widespread use, it is advantageous to transform them to a system of differential equations for an enlarged set of variables. These can then be integrated using any standard differential equation solver.
The transformation is achieved by diagonalizing the matrix $\lblockb{B}{B}$ in the matrix exponential in \eqref{eq:mem}; we denote its eigenvalues, which should have negative real part, by $-\lambda_a$. Decomposing the matrix exponential into the contributions from these eigenvalues and corresponding (left and right) eigenvectors, and multiplying from the left and right by $
\lblockb{S}{B}$ and $\lblockb{B}{S}$ respectively, one can write the memory function in the form
\begin{equation}
M_{\beta\alpha}(\Delta t) = \sum_a
m_{\beta\alpha}^ae^{-\lambda_a \Delta t}
\end{equation}
where the $m_{\beta\alpha}^a$ are constant coefficients. The projected equations \eqref{eq:projectedeqns} can then be expressed as
\begin{equation}
\label{eq:projectedeqnssolve}
\frac{\partial}{\partial t}a_\alpha(t) = \sum_\beta a_\beta(t)\Omega_{\beta\alpha} + \sum_a c_\alpha^a(t) + r_\alpha(t)
\end{equation}
with
\begin{equation}
c_\alpha^a(t) = \int_0^tdt'\,\sum_\beta a_\beta(t')m_{\beta\alpha}^ae^{-\lambda_a(t-t')}
\end{equation}
The memory terms that appear are now pure exponentials, so the $c_\alpha^a$ can instead be obtained by solving the {\em differential} equations
\begin{equation}
\frac{\partial}{\partial t} c_\alpha^a(t) = - \lambda_a c_\alpha^a(t) + \sum_\beta a_\beta(t)m_{\beta\alpha}^a
\label{eq:c_eqns}
\end{equation}
with initial condition $c_\alpha^a(0)=0$. In our numerics we thus solve (\ref{eq:projectedeqnssolve},\ref{eq:c_eqns}), with the random force term $r_\alpha(t)$ omitted as usual. The solution gives us the desired time courses of the subnetwork species.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.7]{egfr_pic_enzymes.eps}
\caption[EGFR network with highlighted enzymes]{EGFR network as described in
\citet{Kholodenko99}, adapted to include enzyme reactions. Three added enzyme reactions (highlighted by darker colours) with enzymes denoted E1-3 and enzyme-substrate complexes ``enzyme-R'' (denoted ER), ``enzyme-PLC$\gamma$'' (denoted EP), ``enzyme-Shc'' (denoted ES) capture the Michaelis-Menten contributions to the dynamics. Lines directly connecting molecular species indicate unary conversion reactions. We show binary reactions in the style of a factor graph representation \cite{Bishop2006}: each reaction is represented by a square; the two reaction partners are connected to the square by lines and an arrow points from the square to the reaction product. (The arrow does not have a meaning beyond this, and in particular does not indicate the direction in which the reaction takes place.) Nodes in the chosen subnetwork are coloured pink and purple; the four heavy circles indicate the boundary nodes.}
\label{fig:MMegfr}
\end{figure}
\subsection{Explicit enzyme reactions}
\label{sec:EGFRMMtoMA}
For the projected equations with enzymes represented explicitly, we need to convert the three (irreversible) Michaelis-Menten reactions in the EGFR network model \cite{Kholodenko99},
\begin{equation}
\begin{split}
\text{RP} &\rightleftharpoons \text{R}_2\\
\text{PLC}\gamma\text{P}&\rightleftharpoons \text{PLC}\gamma\\
\text{ShP}&\rightleftharpoons \text{Shc}
\end{split}
\end{equation}
into mass action form. The reversible harpoons above are intended to indicate only that there is a Michaelis-Menten reaction between the two species, and not e.g.\ whether it is reversible or irreversible. For each such reaction we use the relevant $K_{\rm{m}}$ and $V_{\rm{max}}$ values together with a suitable value of the fast rate scale $\gamma$ to create a set of mass action reaction rate constants. The relevant mass action terms are then added to the equations for substrate and product, and we add a mass action equation for the time evolution of the concentrations of enzyme and enzyme complex, as in \eqref{eq:MMmassaction}. The steady states of the enzyme and enzyme complex, which we require for the construction of the projected equations, can then be found by solving the relevant equations for the substrate, enzyme and enzyme complex. We note that to solve the full system of equations, the steady state values of the enzyme-substrate complexes must be added to the relevant conservation laws to ensure that the correct steady state is reached, with the same concentrations as in the Michaelis-Menten description. From this explicit expanded mass action system we can then construct projected equations, either linearised or nonlinear, in the standard manner explained in Sec.~\ref{sec:MMproj}.
\subsection{Consistency of enzyme elimination in linearised dynamics}
To eliminate the enzymes from the projected equations, we apply the elimination procedure as described for linearised dynamics in Section \ref{sec:linMM}: we write the enzyme reactions as unary reactions (effectively conformational changes) between substrate and product and construct from the resulting $\bm{L}_{\rm eff}$ a set of projected equations for the subnetwork observables that no longer makes explicit reference to enzymes.
Fig.~\ref{fig:MMlin_nonlin_errors} shows how the predictions from the linearised projected equations with enzymes represented explicitly approach the ones from enzyme elimination as the fast enzyme reaction scale $\gamma$ grows. Specifically we show the deviation measure (\ref{eq:errormeasure}) that is obtained if we insert for $\delta x_i$ and $\delta \hat x_i$ the predictions from the two types of projected equations.
As $\gamma$ increases, the deviation decays to zero as it should. This provides an important consistency check on our enzyme elimination method, which we obtained by taking {\em analytically} the large $\gamma$ limit of the projected equations with explicit enzymes.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.8]{lin_nonlin_gammavserror2.eps}
\caption[Errors from enzymes in linearised and nonlinear dynamics]{Deviations between predictions of projected equations with explicit enzymes and with enzymes eliminated, as a function of the fast enzyme rate scale $\gamma$ used when enzymes are represented explicitly. Shown are results for an initial condition far from the steady state, with $\delta \approx 2$. The two curves show the deviations for the two separate cases of linearised and nonlinear dynamics.
}
\label{fig:MMlin_nonlin_errors}
\end{figure}
\subsection{Consistency of enzyme elimination in nonlinear dynamics}
For the full nonlinear dynamics we derive the projected equations with enzymes eliminated using the method described in Sec.~\ref{sec:nonlinMM}. As for the linearised dynamics we use the predictions from these equations as a baseline in a comparison against the projected equations for the reaction network with explicit enzymes. Figure \ref{fig:MMlin_nonlin_errors} shows that the deviation between the time courses predicted by the two sets of equations again decreases towards zero with increasing $\gamma$ as it should.
\subsection{Accuracy of projected equations}
While the results in Fig.~\ref{fig:MMlin_nonlin_errors} are useful as consistency checks, the question that we are primarily interested in is how well the projected equations perform in predicting the real dynamics of the chosen subnetwork, as given by the full system of reaction equations including Michaelis-Menten terms.
Fig.~\ref{fig:MMnonlineqnerrors} shows the approximation errors $\Delta$ of the predictions made by (a) the simpler projected equations, with enzymes explicitly dependent and therefore involving a fast enzyme rate scale parameter $\gamma$, and (b) the projected equations derived using the new enzyme elimination method. In both cases we consider the full nonlinear projected equations, and show the approximation error $\Delta$ as a function of the distance $\delta$ of the initial conditions from the steady state.
One can argue on general grounds that nonlinear projected equations should have an approximation error growing only as $\delta^3$, whereas simpler approximation methods give much larger errors of order $\delta$~\cite{Rubin2014d}. Fig.~\ref{fig:MMnonlineqnerrors} shows that the expectation of a $\delta^3$-scaling is obeyed well for the projected equations derived using the enzyme elimination method. The projected equations derived by representing enzymes explicitly, on the other hand, have significantly higher $\Delta$, even for the largest $\gamma=10^5$ that we can use while still being able to solve the resulting stiff differential equations. The approximation error scales $\Delta \sim \delta$ over a large range, indicating that the introduction of the explicit enzymes and their reaction rates causes a much larger approximation error than the projection method (with its neglect of the random force) itself. One would expect from the trends in Fig.~\ref{fig:MMnonlineqnerrors} that for larger $\gamma$ than we can access, the error would grow as $\sim \delta$ for small $\delta$ but then cross over to the $\gamma\to\infty$ (enzyme elimination) curve at larger $\delta$. The crossover value of $\delta$ has to go to zero as $\gamma$ grows to reflect the consisteny between the $\gamma\to\infty$ limit and the enzyme elimination method.
It is notable that computation times required to integrate the system of projected equations derived by enzyme elimination are shorter by roughly a factor of 2 than those for the case where enzymes are represented explicitly; this is due to the appearance of ``stiff'' terms in the latter case, which the enzyme elimination avoids. The enzyme elimination method thus is not only more accurate but also computationally more efficient.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.8]{nonlinerrorplot.eps}
\caption[Errors from enzymes in nonlinear dynamics]{Log-log plot of approximation error $\Delta$ vs initial deviation from steady state $\delta$. We compare the nonlinear projected equations with enzymes eliminated and with enzymes represented explicitly for different values of the fast rate scale parameter $\gamma$. The baseline in both cases are the time courses from the full system of reaction equations including Michaelis-Menten terms (i.e.\ with the quasi-steady state approximation, which is exact for $\gamma\to\infty$). The enzyme elimination method has an approximation error growing as $\delta^3$ as indicated by the blue dotted line; for the smallest $\delta$ the errors become too small to measure accurately.
The projected equations with explicit enzymes produce much larger errors that grow linearly in $\delta$ for small $\delta$
.
}
\label{fig:MMnonlineqnerrors}
\end{figure}
\section{Discussion}
We have considered the problem of describing subnetwork dynamics in protein interaction networks. The projection approach we have previously developed gives accurate results in this regard, but can be applied directly only to systems with unary and binary reactions. Our aim in this paper was to extend it to systems involving enzymatic reactions represented as Michaelis-Menten terms.
We summarized the conditions under which enzyme reactions represented in mass action form become equivalent to Michaelis-Menten kinetics. This suggested a convenient scaling, of fast enzyme reaction rates and simultaneously low enzyme concentrations, that exactly reproduces Michaelis-Menten equations in the limit where the relevant fast rate scale parameter $\gamma$ grows large.
Applying this construction, one can map a protein interaction network with Michaelis-Menten reactions to an extended one with only unary and binary reactions; the main task is then to understand what limit is approached in the projected description when $\gamma\to\infty$. By analysing where the fast enzyme degrees of freedom feature in the rate matrix and memory functions that define the projected equations, we showed that it is possible to construct the projected equations directly in the large $\gamma$-limit, both for linearised and nonlinear dynamics, using a quasi-steady state elimination. This gives us effective unary reaction contributions to represent the Michaelis-Menten dynamics, with concentration-dependent reaction rate constants in the nonlinear case, and so allows one to constuct the projected equations without ever introducing enzymes explicitly.
The resulting method significantly widens the range of biochemical reaction systems to which the projection approach can be applied; as one example we showed an application to a subnetwork of the EGFR reaction network of \citet{Kholodenko99}. Here we demonstrated that the enzyme elimination method produces significantly more accurate predictions for the subnetwork dynamics than can be obtained when enzymes are represented explicitly, even for the largest values of the fast rate scale parameter $\gamma$. We also found that enzyme elimination leads to gains in computational efficiency.
Our general approach to constructing projected equations for networks of nonlinear reaction equations should be more widely applicable still, and we hope in future work to consider e.g.\ Michaelis-Menten reactions with inhibition and Hill equations \cite{Murray2001}. Conceptually more demanding will be an extension to cases where copy number fluctuations from finite reaction volumes are significant, which would correspond to keeping nonzero $\epsilon$ in our notation. An elegant treatment of the case where the bulk contains only fast variables has been given in \citet{Thomas2010} and may serve as a useful point of departure in this direction.
\section*{Acknowledgements}
PS acknowledges the stimulating research environment provided by the EPSRC Centre for Doctoral Training in Cross-Disciplinary Approaches to Non-Equilibrium Systems (CANES, EP/L015854/1). KJR gratefully acknowledges a
BBSRC Quota Doctoral Training Grant.
|
\section{Introduction}
In cosmology, matter is normally treated as relativistic fluid
and its equation of motion is studied directly.
Nevertheless, in cases such as discussing symmetry,
it is more convenient to utilize Lagrangian formulation,
and consider the fluid equation via the variational principle.
Naturally,
various works have been proposed (for example~\cite{Taub})
to construct Lagrangian formulation of relativistic
perfect fluid.
However, while the choice of dynamical variables is essential to simplify the discussion,
these methods are somehow indirect, such as
representing the fluid elements by field variables that do not appear
explicitly in the equations of motion.
Such roundabout may not be desirable for both pedagogical and theoretical reasons,
{\it e.g.} building a simple cosmological model.
The generalization of variational principle to the fluid mechanics
is carried out by following the case of a rigid body~\cite{Marsden-Ratiu},
leading to Euler-Poincar\'e equations.
In such sense, the appropriate variables for the case of a perfect fluid are,
$(u,\rho,p)$, representing:
velocity vector field of a matter flow,
energy/mass density, pressure: respectively.
In fact,
such way of construction in case of non-relativistic perfect fluid
is given by Lanczos~\cite{Lanczos},
and the case of relativistic dust fluid ($p=0$) is given
in the elementary text book by Dirac~\cite{Dirac}.
Combining these two approaches, the authors presented
the Lagrangian formulation of relativistic perfect fluid
in a geometrical way~\cite{OYIT}, (section 3.3).
Nevertheless, the above formalism still had some unnecessary complications.
In the present paper,
we will present a more simplified and straightforward formulation
which will give covariant equations of motion.
This formulation is in accordance with the Eulerian specification of fluid dynamics.
\section{Variational Principle}
The Lagrangian of the perfect fluid on $(n+1)$-dimensional Lorentzian manifold
$(M,g)$ is given by
\begin{eqnarray}
L_{\rm p.f.}(u,\rho,p) &:=& L_{\rm dust}(u,\rho) + \int\ p\ \delta L (u) \ ,
\label{eq1}
\end{eqnarray}
where the first term is the Lagrangian of dust fluid, $u \in \frak{X}(M)$ is the vector field
of the fluid velocity,
and $\rho \in \Lambda^{n+1}(M)$ is the mass density $(n+1)$-form of the fluid.
It is given by
\begin{eqnarray}
L_{\rm dust}(u,\rho) &:=& \rho F(u),
\label{eq2}
\end{eqnarray}
where we consider
a Finsler metric of the form:
$F=\sqrt{g_{\mu\nu}(x)dx^\mu dx^\nu}$, and $F(u)=\sqrt{g_{\mu\nu}(x)u^\mu u^\nu}$ .
Here, $\{(x^0,x^1,x^2,\dots, x^n)\}$ are the local coordinates of $M$
and $\displaystyle u=u^\mu \bib{}{x^\mu}$.
$F$ and its derivatives are regarded as functions on $TM$,
and throughout this paper, notation such as $F(u)$
means the contraction of the vector field $u$
to $dx^\mu$ in the above sense.
We also set the speed of light $c=1$.
The second term of Eq.~(\ref{eq1}) represents the effect of pressure
$p \in C^\infty (M)$, where the integral $\int$ represents symbolically
that its variation gives $p\ \delta L (u)$.
$L (u)$ is given by
\begin{eqnarray}
L (u) &:=& \omega_g \, F(u),
\label{eq3}
\end{eqnarray}
where
$\omega_g = \sqrt{-g}dx^0\wedge dx^1\wedge \cdots\wedge dx^{n}$,
($g:= {\rm det}\big( g_{\mu\nu}(x) \big)$) is the volume element.
The action is given by the integral
\begin{eqnarray}
{\cal A}[u,\rho,p]:=\int_{\sigma} L_{\rm p.f.}(u,\rho,p),
\label{eq4}
\end{eqnarray}
where $\sigma$ represents a $(n+1)$-dimensional submanifold of $M$.
By considering its variation,
\begin{eqnarray}
\delta {\cal A}[u,\rho,p]&:=& \int_\sigma \delta L_{\rm p.f.}(u,\rho,p)
=\int_\sigma \delta L_{\rm dust}(u,\rho) + p\ \delta L(u),
\label{eq5}
\end{eqnarray}
one obtains the equations of motion.
\subsection{Equation of dust fluid}
First, we calculate $\delta L_{\rm dust}(u,\rho)$.
The fluid element shifts infinitesimally to the direction
$\xi \in \frak{X}(M)$, and
the corresponding variations of
$\rho$ and $u$ are represented by the Lie derivatives as,
\begin{eqnarray}
\delta \rho &=& {\cal L}_\xi \rho, \quad
\delta u={\cal L}_\xi u=[\xi, u]=-{\cal L}_u \xi\ , \quad
\delta u^\mu=[\xi, u]^{\mu}.
\label{eq6}
\end{eqnarray}
Then,
\begin{eqnarray}
\delta L_{\rm dust}(u,\rho) &=& (\rho+\delta \rho)F(u+\delta u)
-\rho F(u)
= \delta \rho F(u)+\rho \bib{F}{u^\mu}
\delta u^\mu
\nonumber\\
&=&({\cal L}_\xi \rho) F(u)+\rho \bib{F}{u^\mu}
\langle dx^\mu, {\cal L}_\xi u \rangle
= (d\iota_\xi \rho ) F(u)-\rho \bib{F}{u^\mu}
\langle dx^\mu,{\cal L}_u \xi \rangle
\nonumber\\
&=&
d\left\{\iota_\xi \rho F(u)\right\}
+(-1)^{n+1}
\iota_\xi \rho \w d(F(u))
-{\cal L}_u \left(\rho \bib{F}{u^\mu}\xi^\mu\right)
+\left\langle {\cal L}_u \left(\rho \bib{F}{u^\mu}dx^\mu\right),\xi
\right\rangle
\nonumber\\
&=&
d\left\{\iota_\xi \rho F(u)-\iota_u \rho \bib{F}{u^\mu}\xi^\mu
\right\}
+(-1)^{n+1}
\iota_\xi \{\rho \w d(F(u))\}
\nonumber\\
&& -\rho \langle d(F(u)), \xi \rangle
+\left\langle {\cal L}_u \left(\rho \bib{F}{u^\mu}dx^\mu\right),\xi
\right\rangle.
\label{eq7}
\end{eqnarray}
The Cartan's formula ${\cal L}_u=d\iota_u+\iota_u d$ is used,
and for notational convenience, we abbreviated
$\displaystyle \bib{F}{u^\mu}:=\bib{F}{dx^\mu}(u)$.
The bracket $\langle \, , \rangle$ denotes contraction, for instance,
\begin{eqnarray*}
\left\langle {\cal L}_u \left(\rho \bib{F}{u^\mu}dx^\mu\right),\xi
\right\rangle
:= \left({\cal L}_u \rho \bib{F}{u^\mu}\right ) \langle dx^\mu, \xi\rangle
+ \rho \bib{F}{u^\mu} \langle {\cal L}_u dx^\mu ,\xi \rangle,
\end{eqnarray*}
and $\langle dx^\mu ,\xi \rangle :=\xi^\mu$.
Then since $\rho \w d(F(u))=0$,
the variational principle tells us that
\begin{eqnarray}
0&=&{\cal L}_u \left(\rho \bib{F}{u^\mu}dx^\mu\right)-\rho d(F(u)), \nonumber\\
&=& ({\cal L}_u \rho)\bib{F}{u^\mu}dx^\mu+\rho {\cal L}_u
\left(\bib{F}{u^\mu}dx^\mu\right)-\rho d(F(u)), \nonumber\\
&=& ({\cal L}_u \rho)\bib{F}{u^\mu}dx^\mu
+\rho \left\{
\bbib{F}{x^\nu}{u^\mu}u^\nu dx^\mu+\bbib{F}{u^\alpha}{u^\mu}
u^\nu \bib{u^\alpha}{x^\nu}dx^\mu+\bib{F}{u^\alpha} \bib{u^\alpha}{x^\mu}dx^\mu
\right\} \nonumber \\
&& -\rho \left(
\bib{F}{x^\mu}dx^\mu+\bib{F}{u^\alpha}\bib{u^\alpha}{x^\mu}dx^\mu
\right), \nonumber\\
&=& ({\cal L}_u \rho)\bib{F}{u^\mu}dx^\mu
-\rho
\left[\left\{\bib{F}{x^\mu}-d\left(\bib{F}{dx^\mu}\right)\right\}(u)
\right] dx^\mu,
\label{eq8}
\end{eqnarray}
for arbitrary $\xi\in \frak{X}(M)$.
This equation (\ref{eq8}) is called the Euler-Poincar\'e equation for the dust fluid.
Therefore, clipping off $dx^\mu$, we get,
\begin{eqnarray}
0&=&\left [({\cal L}_u \rho)\bib{F}{dx^\mu}
-\rho \left\{\bib{F}{x^\mu}-d\left(\bib{F}{dx^\mu}\right)\right\} \right](u).
\label{eq9}
\end{eqnarray}
Contracting (\ref{eq9}) with $u^\mu$ gives
$ 0=({\cal L}_u \rho) F(u)$,
where we used the homogeneity
$u^\mu \bib{F}{dx^\mu}(u)= F(u)$, and its Euler-Lagrange derivative,
\begin{eqnarray}
0&=&
u^\mu \left\{ \bib{F}{x^\mu}-d\left(\bib{F}{dx^\mu}\right)\right\}(u).
\label{eq10}
\end{eqnarray}
Thus, in the case $F(u)\neq 0$, we obtain
\begin{eqnarray}
{\cal L}_u \rho=0, && \quad
\rho \left\{\bib{F}{x^\mu}-d\left(\bib{F}{dx^\mu}\right)\right\}(u)=0,
\label{eq11}
\end{eqnarray}
where the first equation represents the conservation of mass, while
the second is the Euler equation for dust fluid.
This equation is generalized to perfect fluid
in the next section.
Up to now, we have only used the homogeneous property of $F$,
and not its concrete form given by the Lorentizan metric.
Therefore, we may choose for $F$ any general Finsler metric
that satisfies ${\rm rank}\left(\bbib{F}{dx^\mu}{dx^\nu}\right)=n$ .
In that case, the equation (\ref{eq11}) could be regarded as
an equation of dust fluid on Finsler spacetime.
\subsection{Equation of perfect fluid}
Here we will calculate
the pressure term $p\, \delta L(u)$ in Eq.~(\ref{eq5}).
\begin{eqnarray}
p\, \delta L(u)
&=& p\, \delta(\omega_g F(u)) =
p\, \left\{ \omega_g \left( F(u+\delta u) - F(u) \right) +\delta\omega_g \, F(u) \right\},
\label{eq12}
\end{eqnarray}
where the first term is calculated similarly as in (\ref{eq7}),
\begin{eqnarray}
&& p \omega_g \left( F(u+\delta u) - F(u) \right) \nonumber\\
&=& p \omega_g \frac{\partial F}{\partial u^\mu} \delta u^\mu
= p \omega_g \frac{\partial F}{\partial u^\mu} \langle dx^\mu, \delta u\rangle
= p \omega_g \frac{\partial F}{\partial u^\mu} \langle dx^\mu,{\cal L}_\xi u\rangle
=-p \omega_g \frac{\partial F}{\partial u^\mu} \langle dx^\mu,{\cal L}_u \xi\rangle
\nonumber
\\
&=&
-{\cal L}_u\left( p \omega_g \bib{F}{u^\mu} \langle dx^\mu,\xi \rangle \right)
+ \left \langle {\cal L}_u \left( p\omega_g \bib{F}{u^\mu} dx^\mu \right), \xi \right \rangle
\nonumber \\
&=& -d \left\{
\iota_u \left( p\omega_g \frac{\partial F}{\partial u^\mu} \xi^\mu \right)
\right\}
+ \xi^\mu \left\{
{\cal L}_u \left( p\omega_g \right) \frac{\partial F}{\partial u^\mu}
+ p\omega_g\ \iota_u d \left( \frac{\partial F}{\partial u^\mu}\right)
+p\omega_g \frac{\partial F}{\partial u^\rho}
\bib{u^\rho}{x^\mu}
\right\},
\label{eq13}
\end{eqnarray}
and the second term of Eq.~(\ref{eq12}) becomes
\begin{eqnarray}
pF(u)\delta\omega_g &=& {\cal L}_\xi\left( pF(u)\omega_g \right)
- {\cal L}_\xi\left( pF(u) \right) \omega_g \nonumber\\
&=& d \left\{ \iota_\xi \left( pF(u) \omega_g \right) \right\}
- \xi^\mu \left\{ \frac{\partial p}{\partial x^\mu} \omega_g F(u)
+ p \omega_g \left( \frac{\partial F}{\partial x^\mu}(u)
+ \frac{\partial F}{\partial u^\rho} \frac{\partial u^\rho}{\partial x^\mu} \right)
\right\}.
\label{eq14}
\end{eqnarray}
Then,
\begin{eqnarray}
p \, \delta L(u)
&=& d \left\{ \iota_\xi\left( p\omega_gF(u) \right)
-\iota_u \left( p\omega_g \frac{\partial F}{\partial u^\mu} \xi^\mu \right) \right\}
\nonumber\\
&& + \xi^\mu \left[
{\cal L}_u(p\omega_g) \bib{F}{u^\mu}
-p\omega_g \left\{ \frac{\partial F}{\partial x^\mu}(u)
-\iota_u d\left(\frac{\partial F}{\partial u^\mu}\right) \right\}
-\frac{\partial p}{\partial x^\mu}\omega_g F(u)\right].
\label{eq15}
\end{eqnarray}
Finally, by using
Eqs.~(\ref{eq5}), (\ref{eq7}), and (\ref{eq15}),
we obtain
\begin{eqnarray}
\delta L_{\rm p.f.}(u,\rho,p)
&=& \delta L_{\rm dust}(u,\rho)
+ p\, \delta L(u) \nonumber
\\
&=& d\left\{ \iota_\xi\rho F(u) -\iota_u \rho\ \frac{\partial F}{\partial u^\mu}\xi^\mu
+p\iota_\xi \omega_gF(u)
-p\iota_u \omega_g \frac{\partial F}{\partial u^\mu}\xi^\mu \right\}
+ \xi^\mu {\cal EL}_\mu (L_{\rm p.f.}).\ \ \ \
\label{eq16}
\end{eqnarray}
By the variational principle,
the last term gives the Euler-Lagrange equation,
\begin{eqnarray}
{\cal EL}_\mu (L_{\rm p.f.})&=&0, \nonumber \\
{\cal EL}_\mu (L_{\rm p.f.}) &:=& \frac{\partial F}{\partial u^\mu}{\cal L}_u(\rho+p\omega_g)
-(\rho+p\omega_g) \left\{ \frac{\partial F}{\partial x^\mu}
- d \left(\frac{\partial F}{\partial dx^\mu}\right)
\right\}(u)
-\frac{\partial p}{\partial x^\mu} \omega_g F(u).\ \ \ \ \
\label{eq17}
\end{eqnarray}
The above equation (\ref{eq17}) is invariant with respect to the transformation
$(u,\rho,\omega_g) \to (\phi u, \phi^{-1}\rho,\phi^{-1} \omega_g)$,
where $\phi$ is an arbitrary positive definite function on $M$.
This transformation $u\to \phi u$ corresponds to changing ``time''
which is a parameter of fluid elements.
Therefore, this scale transformation means reparametrization
in the perspective of Eulerian specification of flow.
\subsection{covariant Euler equation}
In the previous sections, we have scarcely used the concrete form of the
function $F$,
which was given by the
Lorentzian metric.
In fact, only the homogeneous property of $F$
was required to carry out the above calculations.
Here, we will show that the covariant
Euler-Lagrange (Euler-Poincar\'e) equation of perfect fluid
(\ref{eq17})
will indeed give a familiar Euler form (evolution equation form).
Considering the concrete form,
\begin{eqnarray}
F(u) &:=& \sqrt{g_{\mu\nu}(x)u^\mu u^\nu}, \quad
\frac{\partial F}{\partial u^\mu} = \frac{g_{\mu\nu} u^\nu}{F(u)},\ \ \
\rho =:\mu \omega_g,
\label{eq21}
\end{eqnarray}
the factor in the second term of
Eq.~(\ref{eq17}) is rewritten as
\begin{eqnarray}
&&\left[ \bib{F}{x^\mu}-d\left(\bib{F}{dx^\mu}\right) \right](u)
=
\left[ \frac{1}{2F} \bib{g_{\alpha\beta}}{x^\mu}dx^\alpha dx^\beta
-d\left(\frac{g_{\mu\beta}dx^\beta}{F}\right) \right](u)
\nonumber\\
&=& \frac{1}{2F(u)} \bib{g_{\alpha\beta}}{x^\mu}u^\alpha u^\beta
-\frac{g_{\mu\beta}du^\beta}{F(u)}
-\frac{1}{F(u)}\bib{g_{\mu\beta}}{x^\alpha}u^\alpha u^\beta
+\frac{g_{\mu\beta}u^\beta}{F(u)^3}
\left(\frac12\bib{g_{\alpha\nu}}{x^\rho}u^\alpha u^\nu u^\rho
+g_{\alpha\nu}u^\alpha du^\nu \right)
\nonumber\\
&=&-\frac{g_{\mu\beta}}{F(u)}
\left\{ u^\nu \bib{u^\beta}{x^\nu} + {\varGamma^\beta}_{\nu \rho} u^\nu u^\rho \right\}
+\frac{g_{\mu\beta}u^\beta}{F(u)^3}
\left\{ u_\nu \bib{u^\nu}{x^\alpha} u^\alpha
+ u_\nu {\varGamma^\nu}_{\alpha \rho} u^\alpha u^\rho \right\}
\nonumber\\
&=&-\frac{g_{\mu\beta}}{F(u)}\nabla_u u^\beta
+\frac{g_{\mu\beta}u^\beta g_{\alpha\nu}u^\alpha}{F(u)^3} \nabla_u u^\nu
=
-\bbib{F}{dx^\mu}{dx^\beta}(u)
\nabla_u u^\beta.
\label{eq22}
\end{eqnarray}
Then, Eq.~(\ref{eq17}) becomes,
\begin{eqnarray}
(\rho+p\omega_g)F_{\mu\nu}(u)
(\nabla_u u^\nu)=\bib{p}{x^\mu}\omega_g F(u) -
\bib{F}{u^\mu}
{\cal L}_u (\rho+p\omega_g),
\label{eq23}
\end{eqnarray}
where we set
\begin{eqnarray}
F_{\mu\nu} &:=& \frac{\partial^2 F}{\partial dx^\mu\ \partial dx^\nu},
\quad F_{\mu\nu}(u)=\bbib{F}{dx^\mu}{dx^\nu}(u).
\label{eq24}
\end{eqnarray}
We would like to solve this equation for $\nabla_u u^\nu$.
Since $F_{\mu\nu}$ satisfies the homogeneity relation $F_{\mu\nu}dx^\nu =0$
or $F_{\mu\nu}(u) u^\nu=0$, it is not invertible.
Nevertheless, contracting Eq.~(\ref{eq23}) by $u^\mu$, we get
\begin{eqnarray}
0 &=& {\cal L}_u (\rho+p \omega_g) - ({\cal L}_u p) \omega_g,
\label{eq25}
\end{eqnarray}
where we have assumed $F(u)\neq 0$.
This is the equation of continuity for the perfect fluid on a Lorentzian manifold,
and satisfying this equation guarantees that
a particular solution for (\ref{eq23}) exists.
Using (\ref{eq25}), the equation (\ref{eq23}) is rewritten as
\begin{eqnarray}
F_{\mu\nu}(u)
(\nabla_u u^\nu) &=& \frac{1}{(\mu + p)}\left(
F(u) \bib{p}{x^\mu}-
\bib{F}{u^\mu}
u(p) \right).
\label{eq26}
\end{eqnarray}
Define,
\begin{eqnarray}
\tilde{F}^{\mu\nu}
:= Fg^{\mu\nu}-\frac{dx^\mu dx^\nu}{F},
\label{eq27}
\end{eqnarray}
which satisfies the relation:
\begin{eqnarray}
\tilde{F}^{\mu\rho}F_{\rho\nu}=\delta^\mu_\nu-\bib{F}{dx^\nu}\frac{dx^\mu}{F}, \ \ \
\tilde{F}^{\mu\rho}(u)F_{\rho\nu}(u)=\delta^\mu_\nu-
\bib{F}{u^\nu}
\frac{u^\mu}{F(u)}.
\label{eq28}
\end{eqnarray}
Now set, $c_\nu:= {\it r.h.s}$ of (\ref{eq26}), then $\tilde{F}^{\mu\nu}c_\nu$ is the
particular solution of (\ref{eq23}). Explicitly, it is
\begin{eqnarray}
\frac{\tilde{F}^{\mu\nu}(u)}{\mu+p} \left[
F(u)\bib{p}{x^\nu}-
\bib{F}{u^\nu}
u(p) \right]
=\frac{1}{\mu+p} \left(
F(u)^2
g^{\mu\nu}\bib{p}{x^\nu}-u^\mu u(p) \right).
\label{eq29}
\end{eqnarray}
The general solution would be in the form $\tilde{F}^{\mu\nu}c_\nu+\lambda u^\mu$,
where
$\lambda=\lambda(x,dx)(u)=\lambda (x^\mu,u^\mu)$
is an arbitrary and
homogeneous function of degree one in the $u^\mu$ variables.
The general solution includes one free function $\lambda (x,u)$, since
${\rm rank}(F_{\mu\nu})=n=(n+1)-1$, and $(F_{\mu\nu})$ has 0-eigen vector $u^\mu$.
Thus it is obtained as,
\begin{eqnarray}
\nabla_u u^\mu=\frac{1}{\mu+p}\left\{
F(u)^2
g^{\mu \nu}\bib{p}{x^\nu}-u^\mu u(p)\right\}+\lambda(x,u)u^\mu,
\label{eq30}
\end{eqnarray}
or
\begin{eqnarray}
\nabla_u u^\mu=\frac{F(u)^2}{\mu+p} g^{\mu\nu}
\bib{p}{x^\nu}+\lambda_1(x,u)u^\mu.
\label{eq31}
\end{eqnarray}
The
$\lambda_1(x,u)=\lambda_1(x^\mu,dx^\mu)(u)$
also is an arbitrary and
homogeneous function of degree one in
$u^\mu$.
The existence of these arbitrary functions $\lambda$, $\lambda_1$
indicates that these equations are reparametrization invariant with respect to $u$.
Similarly as in the previous section, the equations (\ref{eq25}) and (\ref{eq31})
are unchanged by the transformation $\phi$, namely
$(u,\omega_g) \to
(\phi u,\phi^{-1}\omega_g)$.
Therefore, we will call this {\it the covariant
Euler equation of curved spacetime}.
\section{Discussion}
We have presented the Lagrangian formulation of
the perfect fluid on Lorentzian manifold by using
standard and minimum variables.
The Euler-Lagrange equation (\ref{eq17}),
is derived geometrically from the Lagrangian (\ref{eq1})
defined by a specific Finsler metric given in terms of Lorentizan metric $g$.
We showed that this equation gives the Euler form (\ref{eq31})
and the equation of continuity (\ref{eq25}).
The above results are invariant
with respect to the transformation:
$(u,\rho,\omega_g) \to (\phi u,\phi^{-1} \rho,\phi^{-1} \omega_g)$,
where $\phi$ is an arbitrary positive definite function on $M$.
This invariance corresponds to the reparametrization
invariance of fluid elements with respect to the time parameter.
This allows us more flexibility to choose the variables that may be advantageous
in considering solutions or symmetries of the fluid field or energy density
on curved spacetime.
The function $\lambda (x,dx)$ in Eq.(\ref{eq30}) (or $\lambda_1(x,dx)$ in Eq.~(\ref{eq31}))
can be determined consistently with the choice of the
evolution rate.
For example, by taking
the velocity speed,
$F(u)=1,\ u_\nu u^\nu=1$, and
$u_\nu \nabla_u u^\nu =0$, we obtain,
$\displaystyle \lambda_1 (x,u)=-\frac{u(p)}{\mu+p}$.
Then we get
\begin{eqnarray}
\nabla_u u^\mu= \frac{g^{\mu\nu}-u^\mu u^\nu}{\mu+p} \bib{p}{x^\nu},
\label{eq33}
\end{eqnarray}
while the continuity equation becomes
\begin{eqnarray}
\nabla_\alpha \{(\mu+p)u^\alpha\}-u^\mu \bib{p}{x^\mu}=0,
\label{eq32}
\end{eqnarray}
which are indeed the relativistic Euler equation and
continuity equation of perfect fluid~\cite{Choquet}.
The discussion in the previous subsections, especially in
subsection A. and B.,
was established mostly without the reference to the concrete form of
the metric $F$. The important property which contributed to the theory
was only the homogeneity of $F$.
This suggests that we may choose for $F$ a more general
form of Finsler metric
instead of the specific form given by the Lorentzian metric.
We are now working on such extensions, hoping to report the results soon.
\begin{acknowledgements}
We thank Prof. M. Morikawa and the astrophysics and cosmology lab (Ochanomizu university)
for support and encouragements.
E. Tanaka thanks YITP of Kyoto University.
\end{acknowledgements}
|
\section{Introduction}
Cellular automata \cite{Kari} are defined over certain lattices where cells change their states in either a synchronous or an asynchronous way depending on a set of local rules which specify a mapping from the state of each cell's neighborhood at the current step into a cell state at the next step. In the present paper we focus on the specification and enumeration of cell neighbors.
In cellular automata theory \cite{Kari}, two kinds of cell neighborhood, von Neumann's and Moore's, are usually considered for two dimensional space. These are then generalized to multidimensional space, and extended for radius greater than one.
However, for multidimensional space, von Neumann's neighborhood can generate too sparse a topology, while Moore's is too dense. Our idea consists in introducing an adjustable parameter which allows us to span between these two neighborhoods. The parameter represents the dimension of hypercubes connecting neighboring cells. ``The On-Line Encyclopedia of Integer Sequences'' \cite{OEIS} has recently approved two new sequences (OEIS A265014) and (OEIS A266213) studied in the present paper and implemented with software \cite{Z15hmn}.
\section{Basic Notions}
Let us consider an infinite integer lattice of dimension $d$. Nodes of this lattice have coordinates $\vec{i}=(i_1,i_2,...,i_d)$, $i_j{\in}\mathbb{Z}$, $1{\leq}j{\leq}d$. According to the terminology of cellular automata \cite{Kari}, we call these nodes \textit{cells}, denoted $c_{\vec{i}}$, and consider each of them to be a unit-size $d$-hypercube with its center situated with coordinates $\vec{i}$. To study neighborhoods of a cell systematically, we recall the definitions of the following distances in multidimensional space \cite{Suther}:
\begin{itemize}
\item
Minkowsky distance \cite{Minkowsky}:
$L^{p}(\vec{i^\prime},\vec{i})
= (\sum\limits_{j}\vert\vec{i'}-\vec{i}\vert)^{p})^{1/p}$.
\item
Manhattan distance \cite{Suther}:
$L^{1}(\vec{i'},\vec{i})
= \sum\limits_{j}(\vert\vec{i'}-\vec{i}\vert)$.
\item
Chebyshev distance \cite{Chebyshev}:
$L^{\infty}(\vec{i^\prime},\vec{i})
= \max\limits_{j}(\vert\vec{i'}-\vec{i}\vert)$.
\end{itemize}
Using these definitions of the above distances, we can characterize Moore's neighborhood \cite{Moore} of a cell $c_{\vec{i}}$ as the set of cells which are situated at Chebyshev distance 1, and von Neumann's neighborhood \cite{Neumann} as the set of cells which are situated at Manhattan distance 1, from $c_{\vec{i}}$. For the 2-dimensional case, two neighborhoods are illustrated in Figure~\ref{fig-nbh}.
\begin{figure}
\center
\includegraphics[width=0.5\textwidth]{fig-nbh}
\caption{Classical neighborhoods (2-dimensional case): a) von Neumann's neighborhood; b) Moore's neighborhood.}
\label{fig-nbh}
\end{figure}
Note that the facets of a finite $d$-dimensional hypercube include $2d$ facets which are $(d-1)$-dimen\-sional hypercubes, each of them includes $2(d-1)$ facets which are $(d-2)$-dimensional hypercubes and so on; finally, there are $2^d$ $0$-dimensional hypercubes (i.e. vertices). In von Neumann's neighborhood, cells are connected only via facets which are $(d-1)$-dimensional hypercubes while in Moore's neighborhood, cells are connected via bounds which are $(d-j)$-dimensional hypercubes, $1{\leq}j{\leq}d$.
\section{$k$-neighborhood}
Let us consider an infinite $d$-dimensional lattice. In von Neumann's neighborhood, the neighboring cells are situated at a Manhattan distance of 1. The number of neighbors is $2d$ and given by the sequence (OEIS A005843). This is clear if we consider the cell coordinate difference $\Delta\vec{i}=\vec{i'}-\vec{i}$, since this difference vector can contain only one nonzero element (belonging to the set $\lbrace{-1,1}\rbrace$, which consists of two elements), and any one of the $d$ coordinates can be chosen to take this value.
In Moore's neighborhood, the neighboring cells are situated at Chebyshev distance of 1. The number of neighbors is calculated as $(3^d-1)$ and given by the sequence (OEIS A024023). Indeed, when we consider the cell coordinate difference $\Delta\vec{i}$, its elements give precisely the set $\lbrace{-1,0,1}\rbrace^d$ (with $3^d$ elements), except that the vector having all coordinates equal to $0$ is excluded.
\textbf{Definition 1.} A \textit{sharp $k$-neighborhood} is a set of cells having difference of either $-$1 or 1 in exactly $k$ coordinates with respect to the current cell.
Thus, we consider neighbors connected via $(d-k)$-cube bounds of the unit-size $d$-hypercube which represents a cell. In other words, neighbors are situated at a Chebyshev distance of 1 restricted by Manhattan distance equal to $k$. We denote this neighborhood by $S(d,k)$. Noticing that only $k$ coordinates of the difference vector $\Delta\vec{i}$ (which can be chosen in $C_d^k$ ways) are nonzero, and must belong to the set $\lbrace{-1,1}\rbrace$ consisting of two elements, gives us the following formula
\begin{equation}\label{eq-ns}
\hat{K}(d,k)=\vert{S(d,k)}\vert=2^k{C_d^k}
\end{equation}
represented by sequence (OEIS A013609). Note that, $S(d,1)$ coincides with von Neumann's neighborhood as far as $C_d^1=d$ and a union over $k$ gives us Moore's neighborhood with $\sum_{j=1}^{d} {\hat{K}(d,j)}=3^d-1$. The diagonal numbers equal $\hat{K}(d,d)=2^d$. For instance, in the 3-dimensional case, we have 6 sharp 1-neighbors connected via 2-cube bounds (facets or squares), 12 sharp 2-neighbors connected via 1-cube bounds (sides), and 8 sharp 3-neighbors connected via 0-cube bounds (vertices).
\textbf{Definition 2.} A \textit{$k$-neighborhood} is a set of cells having difference of either $-$1 or 1 in $j$ coordinates, $1 \leq j \leq k$, with respect to the current cell.
It directly follows from the definition that the number of neighbors in a $k$-neighborhood can be calculated as:
\begin{equation}\label{eq-ng}
K(d,k)=\vert{G(d,k)}\vert=\sum\limits_{j=1}^{k} \hat{K}(d,k)=\sum\limits_{j=1}^{k} 2^{j}C_d^j
\end{equation}
represented by a new sequence (OEIS A265014).
Thus, a $k$-neighborhood connects cells via cubes of dimension $(d-j)$, where $1\leq{j}\leq{k}\leq{d}$. In other words, neighboring cells are situated at Chebyshev distance 1, restricted by Manhattan distance less than or equal to $k$. Of the various $k$-neighborhoods, we obtain as particular cases: von Neumann's neighborhood for $k=1$ and Moore's neighborhood for $k=d$.
Now we are interested in efficient computation of $K(d,k)$. We know that $K(d,1)=2d$ and $K(d,d)=3^d-1$. For $\hat{K}(d,k)$, the following recurrent expression is known (OEIS A013609):
\begin{equation}\label{eq-ns-rc}
\hat{K}(d,k)=2\hat{K}(d-1,k-1)+\hat{K}(d-1,k)
\end{equation}
combined with $\hat{K}(d,0)=1$ in the original sequence (or $\hat{K}(d,1)=2d$ in our case starting from 1). Taking into consideration the fact that $K(d,k)$ represents partial sums of $\hat{K}(d,k)$ on rows, we write:
\[ K(d,k)=K(d,k-1)+\hat{K}(d,k), \qquad K(d,1)=\hat{K}(d,1). \]
Thus we can use a combined scheme, sequentially computing $K(d,k)$ after $\hat{K}(d,k)$ for each combination $(d,k)$. To obtain a completely separated scheme, we use (\ref{eq-ns-rc}) and write
\[ K(d,k)=K(d,k-1)+2\hat{K}(d-1,k-1)+\hat{K}(d-1,k). \]
Expressing $\hat{K}(d,k)$ via $K(d,k)$ in the following way
\[ \hat{K}(d,k)=K(d,k)-K(d,k-1) \]
yields
\begin{equation}\label{eq-ng}
K(d,k)=K(d,k-1)+K(d-1,k)+K(d-1,k-1)-2K(d-1,k-2)).
\end{equation}
\section{$k$-neighborhoods of Radius Greater than One}
The offsets of separate coordinates to neighboring cells are usually equal to either $-$1 or 1, which gives us cells connected with the current cell via some common $(d-k)$-cube. Such a set of direct neighbors defines the mesh (lattice) of connections studied in the previous sections. But sometimes more distant cells, which are separated from the current cell by a hypercube of direct neighbors, influence its behavior. Such an influence is usually specified using a concept of radius \cite{Kari} which extends the standard notion of neighborhood, considered as a neighborhood of radius 1. Sometimes it is of some use to distinguish sharp from usual neighborhoods of definite radius $r$ by applying the equality $L(\vec i, \vec i')=r$, or the inequality $L(\vec i, \vec i')\leq r$, respectively, in the same way as for the $k$-neighborhood discussed above. A sharp neighborhood corresponds to the surface area of the corresponding figure while the usual neighborhood corresponds to its volume. Some difficulty concerns border cells because they are included twice in the usual formula for a hypercube's surface area, $2d(2r+1)^{d-1}$. Usually, the figure of a cell neighborhood defined by some metric is convex. Following the principles outlined in the previous section we will decrement the value excluding from it the current cell. We denote the number of neighbors by capital $R$ in calculations which use parameter $r$ to specify the radius of the neighborhood. Note that the number of neighbors in a standard neighborhood represents the sum of its sharp neighborhoods:
\begin{equation}
R(r)=\sum\limits_{l=1}^{r} \hat{R}(l).\label{eq-nssum}
\end{equation}
Extending Moore's neighborhood for radius $r\geq 1$ to a set of cells situated at Chebyshev distance $r$ gives us a $d$-hypercube of size $2r+1$ with the current cell in its center (which is excluded). Then the total number of neighbors is calculated as
\begin{eqnarray}
R_{\mathit{Moore}}(d,r)=(2r+1)^d-1,\label{eq-nmrs}\\
\hat{R}_{\mathit{Moore}}(d,r)=\sum\limits_{m=1}^{d} {C_d^m}{2^m}{(2r-1)}^{d-m}\label{eq-nmr}.
\end{eqnarray}
Among the various kinds of extended neighborhood, Moore's is the most spacious. Sometimes it is convenient to think of other neighborhoods as its restrictions.
Extending von Neumann's neighborhood for radius $r\geq 1$ to a set of cells situated at Manhattan distance $r$ gives us a diamond-shaped neighborhood, illustrated for the 2-dimensional case in Figure~\ref{fig-rdm}. Let us calculate the number of neighbors in this neighborhood.
\begin{figure}
\center
\includegraphics[width=0.6\textwidth]{fig-rdm}
\caption{Diamond-shaped neighborhood of radius $r$ as a set of cells situated at Manhattan distance $r$.}
\label{fig-rdm}
\end{figure}
\begin{theorem}\label{th-der}
The number of cells situated at sharp radius $r$ in a diamond-shaped neighborhood of a $d$-dimensional lattice is calculated as
$$\hat{R}_{\mathit{diamond}}(d,r)=\sum\limits_{k=1}^{\min(d,r)} C_{r-1}^{k-1} C_d^k 2^k.$$
\end{theorem}
\begin{proof}
We organize a combinatorial choice in the following way. We consider sequentially variants having exactly $k$ nonzero coordinates. At first, we choose $k$ coordinates from the total set of $d$ coordinates, which can be done in $C_d^k$ ways. Next, we partition $r$ into a sum of exactly $k$ nonzero numbers, which can be done in $C_{r-1}^{k-1}$ ways; the choice is organized as a distribution of $r$ balls between $k$ boxes \cite{Stanley}. Finally, because the Manhattan distance uses absolute values, we can choose variants of sign distribution over the $k$ nonzero values in $2^k$ ways. Multiplying the various magnitudes and summing up the required function on the number of nonzero elements $k$, we obtain the required expression.
\end{proof}
\begin{theorem}
The number of cells situated at radius $r$ in a diamond-shaped neighborhood for a $d$-dimensional lattice is calculated as
$$R_{\mathit{diamond}}(d,r)=\sum\limits_{k=1}^{\min(d,r)} C_{r}^{k} C_d^k 2^k;$$
the numbers $D(d,r)=R_{\mathit{diamond}}(d,r)+1$, which are obtained starting the sum from zero, are known as Delannoy numbers \cite{Banderier}.
\end{theorem}
\begin{proof}
We organize a combinatorial choice in the way similar to that in the proof of Theorem~\ref{th-der}, but to cover all variants which contain $k$ nonzero components whose sum is equal to or less than $r$, we introduce a dummy box and a dummy ball which are not taken into consideration subsequently. Next, we distribute $(r+1)$ balls between $(k+1)$ boxes. Of these, $k$ boxes correspond to the vector $\vec{i}$ and the final one is a dummy used to allow distributions of all $x\leq r$ balls between $k$ boxes. The dummy box contains a surplus, the other boxes contain more than 1 ball, with the total being less than or equal to $r$ balls. Thus at this stage we obtain $C_r^k$ variants, which yields the required expression (known as Delannoy numbers \cite{Banderier}).
\end{proof}
\textbf{Corollary}. The following equality holds:
$$D(d,r)=\sum\limits_{k=1}^{\min(d,r)} C_{r}^{k} C_d^k 2^k = \sum\limits_{l=1}^{r} \sum\limits_{k=1}^{\min(d,l)} C_{l-1}^{k-1} C_d^k 2^k$$
To prove the Corollary, we use (\ref{eq-nssum}). The expression gives us one more variant for computing Delannoy numbers, showing that they can be decomposed into partial sums of a new useful sequence $\hat{R}_{\mathit{diamond}}(d,r)$ represented with (OEIS A266213). Note that, \cite{Bruck} includes a table showing the size of a diamond-shaped neighborhood of radius $r$ comparing to Moore's neighborhood of radius $r$, but without mentioning Delannoy numbers and without proof.
Delannoy numbers \cite{Banderier}, sequence (OEIS A008288), are represented as a triangle using its detour on anti-diagonals. Note that they were initially introduced to describe the number of paths in a 2-dimensional rectangular $d\times r$ lattice from its left-bottom corner with coordinates $(0,0)$ to its right-upper corner with coordinates $(d,r)$ using only the three following steps: $(1,0)$, $(0,1)$, $(1,1)$. Delannoy numbers are also known as the tribonacci triangle because of the following efficient recursive scheme for their computation using three previously computed members:
$$D(d,r)=1,~if~d=1~or~r=1,$$
$$D(d,r)= D(d-1,r)+D(d-1,r-1)+D(d,r-1).$$
Note that the same scheme is valid for $\hat{R}_{\mathit{diamond}}(d,r)$ with the following initial conditions:
$$ \hat{R}_{\mathit{diamond}}(d,0)=1, d\geq 0; \qquad \hat{R}_{\mathit{diamond}}(0,r)=0, r>0 $$
There are also other known ways for extending von Neumann's neighborhood using radius $r > 1$; for example the scheme outlined in \cite{Dutta}. This only allows differences with absolute value equal to $r$ in a single coordinate and is illustrated in Figure~\ref{fig-rnm} for the 2-dimensional case. We will call this extension a narrow von Neumann's neighborhood of radius $r$. It suits our intention to extend the $k$-neighborhood studied in previous section to extended neighborhoods of radius $r>1$.
\textbf{Definition 3}. A \textit{$k$-neighborhood of radius $r$} is a set of cells having difference $\vert\Delta\vert\leq r$ with respect to the current cell in $j\leq k$ coordinates. When using equalities instead of inequalities we say that the neighborhoods are sharp on $r$ or $k$ (or both), respectively.
Note that the parameter $k$ allows us to span neighborhoods of radius $r$ from narrow von Neumann's ($k=1$) to Moore's ($k=d$). However, using neighborhoods sharp on $k$ seems unmotivated because it contains gaps.
\begin{figure}
\center
\includegraphics[width=0.6\textwidth]{fig-rnm}
\caption{Narrow von Neumann's neighborhood of radius $r$.}
\label{fig-rnm}
\end{figure}
A $k$-neighborhood of radius $r$ is the union of all $(d-k)$-cubes of size $(2r+1)$ which have their center in the current cell. However, when calculating the number of neighbors as a sum of cells in $(d-k)$-cubes for all possible combinations of $k$-from-$d$ coordinates, we should exclude intersections of cubes which are summed up multiple times.
\begin{theorem}
The number of neighbors in a $k$-neighborhood of radius $r$ is calculated using the following expression:
\begin{eqnarray}
R(d,k,r) &=& \sum\limits_{j=0}^{k} C_{d}^{j} (2r)^j \label{eq-dkr1}
\end{eqnarray}
\end{theorem}
\begin{proof}
To prove (\ref{eq-dkr1}) we organize a combinatorial choice on the number $j$ of nonzero components of the coordinates offset vector $\Delta{\vec i}$. For each component, there are $2r$ variants since zero is excluded; thus, we obtain $(2r)^j$ combinations of nonzero coordinates. Next, $j$ nonzero coordinates are chosen from $d$ coordinates in $C_d^j$ ways which give us the corresponding multiplier.
\end{proof}
\section{Conclusions}
For $d$-dimensional cellular automata, we have introduced $k$-neighborhood based on a parameter $k$ that corresponds to connections of neighboring cells via common $(d-k)$-cubes. We distinguish sharp and usual neighborhoods, according to whether equality of inequality on the parameter $k$ is considered. Usual $k$-neighborhoods span from von Neumann's neighborhood ($k=1$) to Moore's neighborhood ($k=d$). This concept is useful for adjusting the density of neighbors, which increases with increasing $k$. The $k$-neighborhood was extended using the concept of radius $r$ greater than one. We studied both von Neumann's neighborhoods of radius $r$: the narrow corresponding to $k=1$ and the diamond-shaped corresponding to a set of cells situated at Manhattan distance $r$. Calculating the number of neighbors for the sharp case provided a new expression which partial sums equal Delannoy numbers. Two new sequences (OEIS A265014) and (OEIS A266213) have been approved by OEIS \cite{OEIS} and implemented in software \cite{Z15hmn}.
|
\section{Introduction}\label{Introduction}
An expression Quantitative Trait Locus (eQTL) is a genetic polymorphism (typically a single-nucleotide
polymorphism, abbreviated by SNP) that is associated with transcriptional expression levels
in a particular tissue. The statistical analysis of eQTLs has become increasingly important in
understanding molecular mechanisms connecting genetic variation to complex traits and disease (\cite{morley2004genetic, gilad2008revealing, grundberg2012mapping, westra2013systematic}). For example, eQTL studies can be used to plausibly link disease phenotypes analyzed in Genome-Wide Association Studies (GWAS) to gene expression, with recent work focusing on variation across tissues
\citep{gamazon2015gene, ardlie2015genotype}.
Although the underlying biology is complex, a fundamental step in many analyses is to compare the genotypes of a large number of SNPs to the expression levels of all genes, which presents challenges in computation and multiple testing \citep{wright2012computational}.
Statistical analyses of eQTLs have often been based on standard linear regression
\citep{shabalin2012matrix},
with a focus on testing and detection.
A key step, commonly considered necessary to avoid false positives, has been to normalize and transform the
expression data prior to analysis \citep{beasley2009rank}. Often normalization removes the scale of the expression data, and with it, a natural measure of effect size due to genotype.
As a consequence, eQTL effect size has often been described in terms of regression partial $R^2$ between genotype and transformed expression (see for example \cite{stranger2007population}).
However, the $R^2$ statistic can be highly sensitive to transformations of the response, and is difficult to interpret biologically. An appropriate eQTL model should reflect a coherent model of allelic contributions to expression, and provide a null hypothesis to test for dominance \citep{powell2013congruence}. Furthermore, a biologically appropriate effect-size model will improve the accuracy of hypothesis tests (as we show in this paper), and provide reliable rankings of eQTLs in terms of effect sizes instead of $p$-values alone.
In this paper we propose ACME, a model for the effect size of cis-acting eQTLs, in which the effects of genotype alleles on expression
are \underline{A}dditive \underline{C}ontributions on the original expression scale, with
\underline{M}ultiplicative \underline{E}rror.
In the ACME model the log of expression is equal to the {\em log of a linear systematic term}
(``log-of-linear") plus noise and covariate effects: this seemingly subtle difference from
standard log-linear modeling is of key importance in estimating and interpreting effect sizes.
The ACME model reflects a marked departure from standard practice
in eQTL analyses and has important implications for downstream inferences on effect sizes and dominance.
Standard normalizations meant to control false positives run against, as we will argue, the most coherent conception of eQTL action: namely that gene expression is \emph{additive} in allele count. A primary contribution of this paper is to assess the validity of this conception against the alternatives suggested by standard practices. Additionally, we provide (i) a fast, custom fitting algorithm and corresponding software package, (ii) a robustness analysis of our method, and (iii) diverse results and comparisons from a full cis-eQTL analysis using both ACME and existing models.
The organization of the paper is as follows. In the remainder of this section, we discuss
standard eQTL models. In Section \ref{Model}, we lay out the ACME
model and conduct statistical tests on real data to show its conformity to cis-eQTL
action. In Section \ref{pvalues} we analyze robustness of ACME p-values to violations of model assumptions in real data. Section \ref{Power} describes a simulation study to assess consequences of using standard normalizations when ACME is the true model. In Section \ref{RealData} we discuss results from analyses of all cis-eQTLs in nine tissues (data from the GTEx project, \citealt{lonsdale2013genotype}) using both ACME and existing methods. In Section \ref{Discussion}, we summarize our contributions and discuss future research.
\subsection{Existing approaches to gene expression modeling}\label{existing}
Gene expression data are rarely analyzed on the original scale, due to heteroskedasticity and heavy-tailed errors \citep{rantalainen2015robust}. Instead, logarithmic transformation of expression is a standard pre-processing step for many microarray platforms
\citep[e.g.][]{morley2004genetic, irizarry2003exploration} and often plays an important role in
downstream analyses such as differential expression
\citep[e.g.][]{cancer2013genomic, li2014regression}.
RNA-Seq data are inherently count-based, and statistical analyses of such data often make use of
binomial, negative-binomial, or Poisson generalized linear models \citep{mccarthy2012differential, zhou2011powerful, zwiener2014transforming}, which use logarithmic or near-logarithmic link functions.
However, count-based modeling is rarely used in eQTL analysis, due to computational requirements, and the fact that several stages of read-count normalization are usually applied to expression data. Furthermore, count-based modeling may not be necessary in studies with large sample sizes \citep{zhou2011powerful}, and eQTL analyses are often performed using linear
regression of log-transformed expression, assuming additive allelic effects on the log scale \citep[e.g.][]{myers2007survey}.
Another common transformation of expression is inverse quantile-normalization \ \citep[e.g.][]{dixon2007genome}, used to ensure normality of residuals under the null \citep{beasley2009rank, szymczak2013adaptive}. Given a vector $y$ of length $n$, the quantile-normalization (QN) transformation is the function $Q(y_i)=\Phi^{-1}((rank(y_i)/(n+1))$, mapping each value to a normal quantile corresponding to its rank. Henceforth, eQTL analysis involving linear regression of quantile-normalized gene expression will be referred to as ``QN-linear". For eQTL analyses, the QN-linear model yields $p$-values that are approximately uniform under the null of no association between
expression and genotype. However, the quantile-normalization mapping inherent to this approach erases all connection between the linear model parameters and the original gene expression. Hence, estimated coefficients from QN-linear regression do not reflect the scale of the original
data, and contain almost no information about the true allelic effect.
As a result, QN-linear model effect-size estimates from eQTLs with clearly diverse signal-to-noise ratios can yield nearly identical $p$-values (see Web Appendix \ref{qn-vs-ACME}).
\subsection{Notation and data}\label{notation}
eQTL data from $n$ samples will be written as follows.
The SNP genotype is the number of minor alleles 0, 1, or 2, rounded if using imputed data.
Genotype is contained in an $S \times n$ matrix where $S > 0$ is the number of SNP markers; we denote a
(length $n$) row of the genotype matrix by $s$.
Expression is measured by the number of mapped reads relative to the
library size (see Web Appendix \ref{data_prep}).
Read counts for expression are contained in a $T \times n$ matrix
where $T>0$ is the number of genes or transcripts; the
length $n$ row of expression matrix is denoted $c$.
Finally, the $p$ covariates (e.g. sex or batch) are stored in a $p \times n$ matrix.
The GTEx pilot data set \citep{ardlie2015genotype} is used for all analyses and investigations. Given the purported scope of the ACME model, the analysis focus is on ``cis" gene-SNP pairs, for which the SNPs are within 1 megabase upstream or downstream of the transcription start or stop sites. The GTEx pilot data contains a SNP database and expression data from nine tissues, each having sample sizes between $n = 83$ and $n = 156$. Each tissue-specific data set had $p = 19$ covariates: sex and 3 genotype principal components, which were shared across all tissues; and 15 PEER (Probabalistic Estimation of Expression Residuals) factors computed from expression data \citep{stegle2012using}.
\section{The ACME model and diagnostics}\label{Model}
This section provides a ground-up introduction of the ACME model. The first consideration is the appropriate scale of expression data for error control. As discussed in Section \ref{existing}, errors from linear models of raw gene expression data are known to be
heteroskedastic and non-normal. In Web Appendix \ref{raw-expr-resids},
we show that the non-normality observed in real-data residuals after linear regression with raw expression causes severe Type-I error discrepancies. Though the QN-linear model avoids this problem, it is unsuitable for effect-size estimation (as discussed in Section \ref{existing}). Thus we assess the commonly-used log transformation, in two ways. First, we display tests of normality and heteroskedasticity of residuals after
fitting QN-linear, standard linear, and various log-linear models (see Web Appendix \ref{normality})
to GTEx data.
We see that models using log-transformed expression perform much better than those based on
raw expression. Comparison to the QN-linear model results indicate that the log-transformation is
still somewhat subject to noise and outliers. However, we consider the resulting violations of normality and homoskedasticity
to be acceptably modest, when balanced against the ability to assess effect size with a model that respects the scale of expression data. (Section \ref{pvalues} provides a deeper look into Type-I errors from log-based methods.) Second, we assess residual normality under the box-cox transformation \citep{box1964analysis}, defined for $\lambda\in\mathbb{R}$ as
\begin{equation}\label{boxcox}
t_\lambda(y) = \begin{cases}\tfrac{y^\lambda - 1}{\lambda},&\lambda\ne0\\\log(y),&\lambda=0.\end{cases}
\end{equation}
We perform Shapiro-Wilk tests on box-cox transformed expression data from null eQTLs (as judged by QN-linear p-values) subsampled from real data. We find that the log transformation ($\lambda = 0$) consistently results in the fewest instances of significantly non-normal residuals. These results are displayed in Web Figure \ref{fig:box-cox}.
The analyses described above suggest that the log transformation puts gene expression on a natural scale for error control in eQTL effect-size analysis. We now discuss systematic components of various log-scale effect size models, including ACME (to be introduced). Henceforth, let $y_i := \log(1 + c_i)$ denote the log-transformed normalized gene read count from sample $1 \leq i \leq n$, where the addition of 1 avoids taking the logarithm of zero.
The value $c_i$ is the result of taking the original raw count for the gene in sample $i$, library-normalizing and then scaling up
to the magnitude of the original mean count (see Web Appendix \ref{data_prep}).
Let $s_i$ denote the minor allele count for the SNP in sample $i$ ($s_i \in \{0, 1, 2\}$).
Let ${\boldsymbol Z}_i$ denote the $p \times 1$ vector of covariates for sample $i$, and let $\gamma$ be
an unknown $p \times 1$ covariate coefficient vector. Finally, let $\epsilon_1, \ldots, \epsilon_n$ be independent
$N(0, \sigma^2)$ errors with positive variance $\sigma^2$.
Note that the quantities $\sigma$ and $\gamma$ may differ across gene-SNP pairs.
It is common in eQTL studies to assume that the covariate effect ${\mathbf Z}_i^T \gamma$ contributes to
expression on the same scale as the noise \citep[e.g.][]{shabalin2012matrix}. We follow this practice for all models considered in this paper (including linear regressions with both raw and quantile-normalized gene expression). Additional support for this convention can be seen from the fact that the covariates were computed from normalized data to reduce the influence of outliers \citep{ardlie2015genotype}, so it is natural for them to be residualized on the log-scale.
\subsection{Log-ANCOVA and log-linear models}\label{log-models}
We now describe two log-scale linear eQTL models. If one assumes each genotype is associated with a
distinct level of average log-expression, the associated linear model effectively includes a dominance term for
the homozygous genotype for the reference allele,
and yields what we call
the ``log-ANCOVA" model:
\begin{equation}\label{eq:logANCOVA}
y_i = \alpha_0{\mathbbm{1}}_0(s_i) + \alpha_1{\mathbbm{1}}_1(s_i) + \alpha_2{\mathbbm{1}}_2(s_i) + {\boldsymbol Z}_i^T\gamma + \epsilon_i .
\end{equation}
Here the parameters $\alpha_j$ are unknown log-expression means corresponding to the genotypes,
and ${\mathbbm{1}}_k(s_i) = 1$ if $s_i = k$, and zero otherwise. Another (simpler) log-scale model includes just one parameter for allele \emph{count}:
\begin{equation}\label{eq:logLinear}
y_i = \theta_0 + \theta_1s_i + {\boldsymbol Z}_i^T\gamma + \epsilon_i.
\end{equation}
Above, $\theta_0$ is baseline log-expression, and $\theta_1$ is the contribution to log-expression of each reference allele. This model includes one fewer degree of freedom than log-ANCOVA, due to the loss of the dominance term $\alpha_2$. Linear regression on allele count
has been heavily used in eQTL analysis \citep{ardlie2015genotype},
perhaps partly because evidence of eQTL dominance effects are scant, even in trans-analyses \citep{wright2014heritability}. Furthermore, simpler models like \eqref{eq:logLinear} are useful in that they can be used to test for dominance and, in case of allelic independence, provide appropriate estimates of eQTL action.
\subsection{The ACME Model}\label{ACME}
Despite the prevalence of the log-linear model \eqref{eq:logLinear} in eQTL analysis, it has not been subjected to careful scrutiny. The current understanding of cis-eQTL variation in humans is that it is largely allele-specific
\citep{castel2015tools}, i.e.,
the transcription of a gene in a particular chromosome is influenced primarily by one or more
SNP alleles on the same chromosome.
Thus, in the absence of feedback mechanisms, the effect of each SNP allele should
be additive on the {\it original} expression scale.
To incorporate this understanding while respecting the heavy-tailed nature of gene expression data,
we propose the following log-scale non-linear regression
(ACME):
\begin{equation}\label{eq:ACME}
y_i = \log(\beta_0 + \beta_1s_i) + {\boldsymbol Z}_i^T\gamma + \epsilon_i.
\end{equation}
Here $\beta_0$ is the baseline mean expression, and $\beta_1$ the additive contribution of
each allele.
Exponentiating each side of equation \ref{eq:ACME}, on the expression
scale we have:
\begin{equation}\label{eq:ACME-exp}
c_i + 1 = \big(\beta_0 + \beta_1s_i\big) \;\cdot\; \exp\left({\boldsymbol Z}_i^T\gamma + \epsilon_i\right).
\end{equation}
It is clear from this equation that the effect of genotype is linear in raw expression, as desired{\color{red}.}
We fit the ACME model to data via maximum likelihood, using a Gauss-Newton algorithm, which is derived
in Web Appendix \ref{Fitting}.
\subsubsection{The effect size}
The coefficients $\beta_1$ and $\beta_0$ from the ACME model operate on the original expression scale,
so they lend themselves naturally to a ``fold-change" interpretation.
In particular, the ratio
$\beta_1/\beta_0$ represents the fraction of mean increase due to a single referent allele compared to
the baseline genotype 0. We note that Equation \eqref{eq:ACME} may be written
\begin{equation}\label{eq:ACME2}
y_i = \log(\beta_0) + \log\left(1 + \frac{\beta_1}{\beta_0} s_i\right) + {\boldsymbol Z}_i^T\gamma + \epsilon_i,
\end{equation}
which separates the role of $\beta_0$ in determining baseline expression and
the role of $\beta_1/\beta_0$ in determining the effect of genotype.
Equation \ref{eq:ACME2} also plays a role in the fitting algorithm.
In what follows we use ``effect size" to refer to the ratio $\beta_1/\beta_0$.
In Web Appendix \ref{SE}
we obtain a formula for the standard error of $\beta_1/\beta_0$, using a reduced
Hessian matrix derived from the model.
We note that other notions of effect size may be of interest to biologists. For example, an alternative effect size model, studied simultaneously and independently, incorporates allele additivity through an explicit focus on a fold-change parameter \citep{mohammadi2016quantifying}.
\subsubsection{Fitting algorithm and software}
Though the ACME model can in principle be fit by brute-force likelihood maximization, we found stock implementations of this approach to be slow and unreliable in practice. We have crafted a custom fitting algorithm for ACME, provided in Web Appendix E.
We have implemented the fitting algorithm, and a parallelized wrapper for full cis-genome analysis, in a free and open-source software package called $\mathtt{ACMEeqtl}$. The package is written in the $\mathtt{R}$ statistical computing language, and is available on the CRAN repository. In Section \ref{Power:computation}, we benchmark the computation time of our software.
\subsection{Goodness-of-fit tests}\label{diagnostics}
In the previous section, we pointed out that
the log-linear model assumes allelic effect additivity on the log-expression scale, whereas ACME assumes additivity on the raw-expression scale. These assumptions can be treated as competing hypotheses, evaluations of which may be performed with goodness-of-fit tests. In this section, we carry out such tests using data from the GTEx project.
To derive the test, note that the log-linear and ACME models are each nested within the log-ANCOVA model. For instance, the log-ANCOVA model reduces to the ACME model via the parameterization
\begin{align*}
&\alpha_0(\beta) := \log(\beta_0),\\
&\alpha_1(\beta) := \log(\beta_0) + \log\left(1 + \frac{\beta_1}{\beta_0}\right),\\
&\alpha_2(\beta) := \log(\beta_0) + \log\left(1 + 2\frac{\beta_1}{\beta_0}\right).
\end{align*}
Thus, if either
model is sufficient to explain variation in gene expression, any further improvements in the log-ANCOVA fit should be small and consistent with the extra degree of freedom in that model. Conversely, if the fit of log-ANCOVA is (significantly) better than a smaller model, it suggests the smaller model is insufficient. In testing sets of coefficients in nonlinear regression models with normal errors, $F$-tests are widely used \citep{smyth2002nonlinear}, and generally better handle the degree of freedom issues posed by numerous covariates than do likelihood ratio tests.
Define $SSE_3$ as the sum of squared residuals from the fit of log-ANCOVA, and $SSE_2$ as the sum of squared residuals from the fit of any nested model with 2 degrees of freedom (e.g.\ LL and ACME). Then the goodness-of-fit test statistic is $F = \frac{SSE_2 - SSE_3}{SSE_3 / (n - p - 3)}$ which is approximately $F$-distributed with
$1$ and $n - p - 3$ degrees of freedom.
A $p$-value for the goodness-of-fit test is then obtained from the upper-tail of $F_{1,n-p-3}$.
For both the log-linear and ACME models, we applied the goodness-of-fit $F$ test to every cis-acting gene-SNP
pair in Thyroid tissue (with $n = 105$ tissue samples) from GTEx pilot data. As the log-linear and ACME models are indistinguishable under the null model ($\beta_1=0$), we examined the distribution of goodness-of-fit p-values on four bins of cis-eQTL strength (``Null", ``Weak", ``Medium", and ``Strong"), as judged by QN-linear model regression p-values (described fully in Web Appendix \ref{sampling-scheme}). To judge the distribution of the $F$-test p-values from each model, we plotted Q-Q plots on the $-\log_{10}$ scale (see Figure \ref{fig:diagnostic2}). On each plot, we also supplied the genomic inflation factor $\lambda$. The genomic inflation factor is defined by $\lambda:= \text{median}_i\{\chi^2_i\} / 0.455$,
where $\chi^2_i$ is the 1 d.f. chi-squared test statistic corresponding to the $p$-value for the $i$-th test, following the original reasoning for genomic control \citep{devlin1999genomic}.
Figure \ref{fig:diagnostic2} shows that the distribution of $F$-statistic $p$-values from the log-linear
model grow increasingly non-uniform as eQTLs become more significant, suggesting that the log-linear
model is mis-specified.
In contrast, $F$-statistic $p$-values for the ACME model are approximately uniform
for eQTLs of all strengths.
In other words, the fit of the ACME model is largely indistinguishable from that of log-ANCOVA, whereas the fit of the log-linear model is largely insufficient to explain non-null eQTLs. Overall, these results provide strong empirical support for the raw-expression allelic additivity assumption of the ACME model, and against the log-expression additivity of the log-linear model. We conclude that, among models nested within log-ANCOVA (which are all models based solely on allelic effect), ACME best conforms to the underlying eQTL signal. Similar results were observed after applying the same sub-sampling and testing pipeline to data from four other GTEx pilot tissues (see Web Figures \ref{fig:more-gof-acme} and \ref{fig:more-gof-ll}). For completeness, we also applied the above goodness-of-fit pipeline to the QN-linear model, using the corresponding ANCOVA with quantile-normalized expression. We find the same upward trend of poor fits with stronger effect size observed in the LL model. Along with the discussion in Section \ref{existing}, this further illustrates the inadequacy of quantile-normalized linear models to capture eQTL action.
\begin{figure}[!htb]
\centering
\makebox{
\includegraphics[scale = 0.16]{full_analyses/Thyroid/lrt_stats_Thyroid_ACME.png}
}
\vskip0.2cm
\makebox{
\includegraphics[scale = 0.16]{full_analyses/Thyroid/lrt_stats_Thyroid_LL.png}
}
\vskip0.2cm
\makebox{
\includegraphics[scale = 0.16]{full_analyses/Thyroid/lrt_stats_Thyroid_QN.png}
}
\caption{\label{fig:diagnostic2} Q-Q plots of likelihood ratio test $p$-values for ACME, LL, and QN models, in each sector of GTEx Thyroid sample data, $n = 105$. The grey line is where we would expect the $p$-values (represented by the red dots) to fall if they were perfectly uniform, and the green line represents the 95\% window of error around this expectation. $\lambda$ is the estimated genomic inflation factor.}
\end{figure}
\section{Model $p$-values and Type I error}\label{pvalues}
In this section we address violations of residual normality, which can affect Type I error. The large number of tests performed in eQTL analyses presents a special challenge for false positive control. For cis-analysis, the number of tests is typically on the order of $10^7$ \citep{lonsdale2013genotype}. Thus, using a Bonferroni bound to control family-wise error at 0.05 requires $p$-values
to be accurate at values of $10^{-9}$. In order to perform a test of no effect, i.e.,
$H_0:\frac{\beta_1}{\beta_0} = 0\;\;\;\text{vs.}\;\;\;H_a:\frac{\beta_1}{\beta_0}\ne 0$
for a given gene-SNP pair, we fit the ACME model and the reduced mean-model with $\beta_1 = 0$, and then
derive a p-value by comparing the resulting $F$-statistic with the $F_{1, n - p - 2}$ distribution.
Non-normality in errors can potentially result in non-uniform $F$-statistic $p$-values under the null.
To assess this, we examined the performance of ACME on simulated null data with realistic errors.
In addition, we examined the effect of skew in errors for the extremal $p$-values resulting from large numbers of tests.
\subsection{Empirical performance of the $F$ test}\label{F-test-sims}
We began our investigation of the empirical performance of the $F$ test by fitting the ACME model to null
data simulated with realistic residuals. The residuals for each simulated gene-SNP pair were obtained by re-sampling estimated residuals from ACME fits to real GTEx data (full details in Web Appendix \ref{sec5app}).
Both the ACME and log-linear models were fit to 1 million null eQTLs generated in this manner,
and $p$-values were obtained from each method using the $F$-test.
The results are shown in Figure \ref{fig:null_pvals}. The $F$-test $p$-values appear
nearly uniform for both models, as the inflation factors were in the range 0.995-1.005 ($\lambda = 1$ corresponds to no inflation).
We emphasize that these conclusions address the behavior of the ACME fit under a realistic null $-$ the earlier analyses established that ACME offers superior fit for real data when evidence of the alternative is strong. Similar results for different settings of the variance of the simulated error are shown in Web Figure \ref{fig:null_pvals-extra}.
\begin{figure}[!htb]
\centering
\mbox{
\includegraphics[scale = 0.10]{sim_plots/null/sig=1/ACME_F_pvals_sig=1.png}
\includegraphics[scale = 0.10]{sim_plots/null/sig=1/LL_F_pvals_sig=1.png}
}
\caption{\label{fig:null_pvals} $p$-value distributions from null simulated data with realistic errors and real covariate/genotype data. $\lambda$ values are inflation factors.}
\end{figure}
\subsection{Importance sampling estimates of Type I error under skew in residuals}\label{importance-sampling}
While the results above are encouraging, consideration of 1 million null pairs is not sufficient to assess
the quality of very small $p$-values under realistic errors.
We are unaware of any attempts via direct data simulation to quantify robustness in eQTL studies to
the stringent multiple testing thresholds necessary for eQTL studies (as low as $10^{-9}$ for cis-testing).
We note that the robustness investigations of \cite{rantalainen2015robust} used only $10^6$ simulations
for each investigated condition.
A computationally efficient way to assess Type-I error rates for extreme nominal $p$-values is to perform
importance sampling \citep{tokdar2010importance}, in which samples are drawn from an appropriate
alternative distribution (with $\beta_1 \ne 0$), then re-weighted to provide an estimated probability
of rejection under the null.
For regression models, skewness in the error distribution
has a major impact on false positive control, and can cause both conservative or
anti-conservative behavior \citep{zhou2015hypothesis}.
Accordingly, we carried out importance sampling using the skew-normal model \citep{azzalini1996multivariate} for the
distribution of $\epsilon$, with skewness determined by a parameter $\delta$, with $\delta=0$ corresponding
to the assumed normal error model. Simulations were performed with modest average expression
$\beta_0=100$, error variance $\sigma^2=1$, minor allele frequencies 0.025, 0.05, and 0.1, and for sample sizes
$n=$ 100, 250, and 500. Among the GTEx datasets used in this paper, most showed skewness in ACME
residuals between -0.5 and 0.5 (see Web Figure \ref{fig:skew-results} for an example using Adipose GTEx pilot data).
We chose the skew-normal parameter to correspond to skewness in this range (details in Web Appendix \ref{sec5app2}).
Target
type I error values $\alpha$ ranged from $10^{-20}$ to $10^{-1}$.
The results for the ACME $F$-test $p$-values are shown in Web Figures \ref{fig:investigate-skew1}-\ref{fig:investigate-skew3}, using the importance sampling
approach detailed in Web Appendix \ref{sec5app2}.
Some general conclusions can be drawn. For $n=100$ and
negative skewness in $\epsilon$ (with $\delta = -0.45$), the ACME $p$-values are noticeably conservative
for $\alpha< 10^{-6}$. For positive skewness, the $p$-values are slightly anti-conservative, but
more accurate than for negative skewness due to asymmetry in the behavior of the systematic
component of the ACME model. For larger $n=250$, the conservativeness under negative skewness is less extreme, and the $p$-values reasonably accurate to $\alpha=10^{-9}$ for the skewness range shown. This suggests that p-values for the ACME
model should produce acceptable Type-I error rates for most gene-SNP pairs, even for those with relatively small sample sizes. Even pairs with higher skew show acceptable type I error for sample sizes of 250 or greater, and any
deviations from ideal behavior tend to be conservative. These trends hold across the tested values of the MAF, though the p-value skew becomes more severe for lower MAF. For trans-analysis, larger sample sizes may be required due to the more stringent testing thresholds. However, \cite{wright2014heritability} suggested that sample sizes $>1000$ are necessary to reliably detect trans-eQTLs, and for such large studies we would expect robust ACME $p$-values, a separate issue from whether ACME is appropriate for trans-analysis. For small sample sizes and to serve as an ancillary approach, we describe the MCC method \citep{zhou2015hypothesis} as a fast method to obtain robust $p$-values in Web Appendix \ref{mccapp}.
\section{Power, estimation accuracy, and computation speed}\label{Power}
In this section we present simulation results which display the detection power, estimation accuracy, and computation speed of the ACME model versus existing alternatives. Recall the representation of the ACME model from equation \ref{eq:ACME2}, involving the parameter $\eta:=\beta_1/\beta_0$. We simulated 100 repetitions of the model at values of $\eta$ along the range $(-0.5, 10)$.
The sample-size was set to $n = 105$, as components of the simulations were taken from real data (as in Section \ref{F-test-sims}). At each repetition, the other components of the model were set as follows: (1)
allele counts from a randomly sampled real-data allele count vector corresponding to GTEx samples of Thyroid tissue; (2) real-data covariate matrix corresponding to Thyroid samples, constant across all repetitions and values of $\eta$; (3) noise vector ($\epsilon_{n\times 1}$) and covariate effect ($\gamma_{p\times 1}$) generated as normals with mean 0 and covariances $\sigma_\epsilon^2 I_n$ and $\sigma_\gamma^2 I_p$ (respectively), independent within and across repetitions.
We replicated the above simulation framework for various choices of $\sigma_\gamma$, with $\sigma_\epsilon$ fixed at 1. We then applied linear (RAW), quantile-normalized linear (QN), log-linear (LL), log-ANCOVA (ANCOVA), and ACME models to each instance of the simulation. For each value of $\eta$, we computed the average and standard deviation over the repetitions of the following metrics (per model). (1) $F$-test p-value for hypotheses $H_0:\eta = 0$ vs.\ $H_1:\eta\ne 0$; (2) Estimated raw expression value when reference allele count equals 1; (3) Estimated raw expression value when reference allele count equals 2.
Note that estimation with the QN model cannot provide (2) or (3), as the model is based on a rank-normalized expression. Furthermore, parameters of the LL model are not directly comparable to those of ACME or RAW (as they are additive on the log-expression scale), which motivates our choice to evaluate estimated expression rather than estimated $\beta_1$.
In Figure \ref{fig:power-sims}, we display results from the above simulation framework with $\sigma_\epsilon = \sigma_\gamma = 1$. Note that, for the top plot, the $x$-axis is transformed according to $w(\eta) := \log(1 + 2\eta)$, a scale which better displays the results when $\eta\in(-0.5, 0)$. We replicated the simulation framework with other choices of $\sigma_\gamma$ both above and below $\sigma_\epsilon$. The results of these replications are shown in Web Figure \ref{fig:power-sims2}. Our results show that the ACME model achieves the most power and estimation accuracy among the alternative methods. Hence, if the ACME model is the best representation of underlying eQTL biology in terms of allele count (as the analyses in Section \ref{diagnostics} suggest), use of the existing methods reduces accuracy and sensitivity. Additionally, in Web Figure \ref{fig:power-sims2}, we show that when $\sigma_\gamma$ is increased (i.e., the covariates have more effect), the accuracy and sensitivity of the competing methods is even further reduced.
\textbf{Remark}: The power of log-ANCOVA is quite close to that of ACME. This is expected, as log-ANCOVA contains ACME, the true simulated model. Thus, in this simulation study, the log-ANCOVA results act mainly as a reference; the larger model's practical and conceptual shortcomings compared to ACME were discussed in Section \ref{Model}. The focus of this study is mainly on single-parameter models and the consequences of misspecification.
\begin{figure}[!htb]
\centering
\mbox{
\centering
\includegraphics[width=.45\linewidth]{sim_plots/2/add1/cvrtMult=1/pvals2.png}
}
\mbox{
\includegraphics[width=.45\linewidth]{sim_plots/2/add1/cvrtMult=1/SNP1.png}
\includegraphics[width=.45\linewidth]{sim_plots/2/add1/cvrtMult=1/SNP2.png}
}
\caption{Results of large-scale simulation experiment. Middle: $-\log{10}$ $F$-test p-values as a function of $\eta$. Left and right: predicted raw expression with one and two reference alleles, respectively.}
\label{fig:power-sims}
\end{figure}
\subsection{Computation times}\label{Power:computation}
To assess computation speed, we timed various methods on every simulation instance for Figure \ref{fig:power-sims}. There were 10,000 unique values of $\eta$, and therefore 1 million simulation instances. On each instance, we recorded the computation time for: (1) the LL model with least-squares estimation; (2) the ACME model with maximum likelihood using the BFGS method implemented in $\mathtt{optim}$ from $\mathtt{R}$; (3) the ACME model with the custom fitting algorithm derived in Web Appendix \ref{Fitting}.
The timing of (1) will be of the same order as any other procedure based on least-squares (RAW, QN, and ANCOVA). The timing of procedure (2) is provided as a benchmark for procedure (3). All computations were performed on an Intel Xeon E5-2640 (2.50 GHz), using the $\mathtt{R}$, and timed with the $\mathtt{microbenchmark}$ package.
The mean and standard deviation of computation times for the procedures, over the 1 million simulation instances, were as follows: least-squares at 0.129ms (0.233), BFGS at 2.687ms (1.270), custom ACME at 0.470ms (0.285). So, the efficiency of the custom ACME algorithm is quite comparable to that of least-squares estimating equations, and outstrips stock optimization methods. The complete package implementing our algorithm, including wrappers that employ parallelization to process all cis-eQTL results from massive-scale eQTL data, is available in the $\mathtt{ACMEeqtl}$ package on the CRAN respository.
\section{Large-scale real data analysis}\label{RealData}
This section contains further comparisons between different effect-size models. Note that the most important real-data comparisons between ACME and standard models are the goodness-of-fit tests shown in Section \ref{diagnostics}. Those tests showed that ACME best fits cis-eQTL data, compared to other allele-count models. Moreover, ACME estimates correspond to a coherent biological interpretation of cis-eQTL action. In light of this, it can be asked whether effect-size \emph{rankings} from other methods, at least, correspond to those from ACME. Effect size estimates for all cis-pairs were computed from Thyroid tissue data using the QN-linear, LL, and ACME models. The top plots in Figure \ref{fig:full3} show that effect-size ordering given to the strongest eQTLs by QN and LL differ markedly from ACME. Also in Figure \ref{fig:full3} is a plot of ACME versus QN-linear regression $p$-values, shown mainly to provide a fuller comparison to \cite{ardlie2015genotype} and other studies which rely on the QN-linear model. It is clear that, while QN p-values are associated with ACME's (as one would hope), they are by no means identical in rank, and quite different in magnitude for the strongest eQTLs. Furthermore, as shown in the top row of plots, many QN effect sizes differed in sign from ACME's. All these comparisons held across tissues, as shown in Web Figure \ref{fig:full-extra}. Beyond method comparisons, we also display summaries of ACME effect sizes in Web Figure \ref{fig:full-extra2}.
\begin{figure}[!htb]
\centering
\mbox{
\includegraphics[width=.45\linewidth]{full_analyses/Thyroid/ACMEes_vs_QNes_Thyroid.pdf}
\includegraphics[width=.45\linewidth]{full_analyses/Thyroid/ACMEes_vs_LLes_Thyroid.pdf}
}
\vskip0.2cm
\mbox{
\includegraphics[width=.45\linewidth]{full_analyses/Thyroid/ACME_vs_QN_pvals_Thyroid_zoomed.pdf}
\includegraphics[width=.45\linewidth]{full_analyses/full-analysis-times.png}
}
\caption{\label{fig:full3} Results of genome-wide cis-eQTL ACME effect size estimations on Thyroid tissue ($n = 105$) from GTEx data. Top row: comparisons of ACME effect sizes (transformed with $w$ as introduced in Section \ref{Power}) with both QN and LL effect sizes. Bottom-left: QN vs ACME regression p-values. Bottom-right: Full-tissue procedure times of the Matrix-EQTL and ACME fitting softwares.}
\end{figure}
Computation times for real-data analyses were recorded for both the ACME and QN model fitting procedures, using the same PC as mentioned in Section \ref{Power:computation}. The $\mathtt{R}$ packages $\mathtt{MatrixEQTL}$ and $\mathtt{ACMEeqtl}$ provide full-tissue cis-eQTL procedures for the QN and ACME models (respectively). We timed the run of each procedure on each of the nine tissues. The results are shown in the bottom-right of Figure \ref{fig:full3}. We note that the ACME software benefits from parallelization that the Matrix-EQTL software currently does not employ. This explains why the ACME full-tissue procedure is faster than Matrix-EQTL, even though the former is based on an iterative optimization procedure.
\section{Discussion}\label{Discussion}
We have proposed ACME, a new model for the effect-size of cis-eQTLs. ACME follows a simple additive model for cis-eQTL action, and is supported by careful analysis of real data.
In particular, we show via goodness-of-fit tests that while the error distribution of gene expression is best modeled on the log-scale, cis-eQTL additive allelic effects occur on the \emph{original} scale, contrary to assumptions implied by standard transformations.
Beyond simply harmonizing some biological considerations regarding allele-specific expression, this analysis shows that a single parameter can be used in most instances to catalog the effect size of a cis-eQTL, and that dominance in cis-eQTLs appears to be rare.
Simulations in Sections \ref{pvalues} and \ref{Power} showed the robustness and the superior power of the ACME model. Real-data analyses in Section \ref{RealData} suggested that, for eQTL ranking purposes, estimates and p-values from standard models (QN and LL) are not adequate stand-ins for those from ACME. Furthermore, as seen in the goodness-of-fit tests, use of the QN-linear or LL models can create false evidence of dominance. Finally, we showed ACME estimation on all cis-pairs is computationally feasible, and have provided open-source software.
We believe the ACME model places cis-eQTL effect size analysis on a solid statistical foundation, and can be readily implemented in current eQTL studies. The results may be useful for investigations in which interpretable eQTL effect sizes and reliable rankings are relevant, such as examining enrichment and overlap with genome-wide association studies \citep{zhu2016integration}. Furthermore, our analytic standard errors for ACME effect sizes allows ACME estimates to be used directly in methods for downstream analysis of multi-tissue eQTL variation \citep{flutre2013statistical, li2013empirical}.
Though we did not explore the application of the ACME model to trans-eQTL analysis, the ACME model can in principle be applied to trans-eQTLs. However, for trans-eQTLs, dominance effects may be more plausible, while current sample sizes may be inadequate to fully investigate such effects. Thus we consider the use of the ACME model for trans-eQTLs to be exploratory. Another possible extension is a multi-SNP ACME model. However, it is not obvious that allelic additivity should hold in a multi-SNP model.
Nonetheless, to facilitate analysis, we have implemented a step-wise fitting algorithm for a multi-SNP ACME model using additivity across loci, with code included in our software package. These estimates can be used to consider preliminary ACME models that are additive in genotypes across SNPs, pending more rigorous development of a multi-SNP approach.
\backmatter
\section*{Supplementary Materials}
Web Appendices and Web Figures referenced in Sections \ref{existing}, \ref{notation}, \ref{Model}, \ref{log-models}, \ref{ACME}, \ref{diagnostics}, \ref{F-test-sims}, \ref{importance-sampling}, \ref{Power}, \ref{Power:computation}, and \ref{RealData} are available with
this paper at the Biometrics website on Wiley Online Library. Code to fit the ACME model comes with the $\mathtt{ACMEeqtl}$ package, available on the CRAN repository.\vspace*{-8pt}
\section*{Acknowledgements}
GTEx data were from dbGaP phs000424.v3.p1 (http://www.gtexportal.org), supported by the NIH Common Fund (commonfund.nih.gov/GTEx). Supported in part by NIH R01MH101819-01, HG007840, NSF DMS1310002, and EPA RD83574701. The authors acknowledge many helpful conversations and phone calls with members of the GTEx Consortium.\vspace*{-8pt}
\bibliographystyle{biom}
|
\section{Introduction. The statement of the main theorem}
The wide presence of power laws in real networks, in biology, economics, and linguistics can be explained in the framework of various mathematical models (see, for example, \cite{durett,mitzenmacher}). According to the Zipf law (\cite{Baayen}), in a list of wordforms ordered by the frequency of occurrence, the frequency of the $r$th wordform obeys a power function of~$r$ (the value $r$ is called the rank of the wordform). One can easily explain this law with the help of the so-called monkey model.
Assume that a monkey (see a detailed bibliographical review in~\cite{monkey}) types any of 26 Latin letters or the space on a keyboard with the same probability of $1/27$. We understand a word as a sequence of symbols typed by the monkey before the space. Let us sort the list of possible words with respect to probabilities of their occurrence (the empty word, whose probability equals $1/27$, will go first in this list followed by 26 one-letter words whose probabilities equal $1/27^2$ and then by $26^2$ possible two-letters words and so on). Evidently, the probability $p(r)$ of a word with the rank of $r$ satisfies the inequality
\begin{equation}
c_1 r^{-\alpha}< p(r)< c_2 r^{-\alpha},\quad \mbox{where $\alpha=\log 27/\log 26$, $c_1,c_2>0$.}
\label{eq:one}
\end{equation}
(here and below we use the symbol $\log$ if the base of the logarithm is not significant; but for the natural logarithm we use the symbol $\ln$).
Relatively recently inequality~(\ref{eq:one}) was generalized to the case of non\-equ\-ipro\-bable letters. Let $p_0$ be the probability that the monkey types the space, let $p_i$, $i=1,\ldots,n,$ denote probabilities of choosing the $i$th letter from the set of $n$ letters ($p_i>0$, $\sum_{i=0}^n p_i=1$), and let $p(r)$ be, as above, the probability of a word with a rank of $r$. Then, as is proved in \cite{Mit, we}, the following inequality analogous to~(\ref{eq:one}) takes place, namely, $\exists c_1,c_2:
0<c_1<c_2$ such that
\begin{equation}
c_1 r^{-\alpha}< p(r)< c_2 r^{-\alpha},\quad \mbox{where $\alpha=1/\gamma$}
\label{eq:two}
\end{equation}
and $\gamma$ is the root of the equation $\sum_{i=1}^n p_i^\gamma=1$ (evidently, $0<\gamma<1$). Note that inequality~(\ref{eq:two}) is equivalent to the boundedness of the difference $-\log p(r)-\alpha \log r$.
In the case, when the probability of each letter is not fixed, but depends on the previous one, words represent trajectories of a Markov chain with the absorbing state $\omega_0$ and transient states $\omega_1,\ldots,\omega_n$. Then the value $p(r)$ is the probability of the $r$th trajectory in the list of possible trajectories sorted in the non-increasing order of probabilities. In this case, the asymptotic behavior of $p(r)$ does not necessarily have a power order. Namely, in this case one of two alternatives takes place (\cite{canada,ourELA}). The first variant is that there exists the limit
\[\lim_{r\to\infty} \frac{-\log p(r)}{r^{1/m}}=c,\quad c>0,\]
where $m$ is some positive integer constant value that depends on the structure of the transition probability matrix and the structure of states, where the initial distribution of the Markov chain is concentrated. The second variant is that independently of the initial distribution there exists the following nonzero limit (the so-called \textit{weak} power law):
\[\lim_{r\to\infty} \frac{-\log p(r)}{\log r}.\]
This limit equals $1/\gamma$, where $\gamma$ is now defined with the help of the substochastic matrix~$P$ of transition probabilities where the row and the column that correspond to the absorbing state~$\omega_0$ are deleted. Namely, raising all elements of the mentioned matrix to the power of $\gamma$ would equate its spectral radius to 1.
These results were obtained independently in \cite{arXiv1,canada} and later refined in \cite{ourELA}. Namely, as appeared, the first alternative means the subexponential order of the asymptotics, i.e., in this case $\exists c_1,c_2: 0<c_1<c_2$ such that
\[
c_1 \exp(-c r^{-1/m})< p(r)< c_2 \exp(-c r^{-1/m}).
\]
The case of the second alternative is much more difficult. If the matrix~$P$ does not have the block-diagonal structure with coinciding powers such that raising elements of blocks to these powers makes the spectral radius equal 1, then one can replace the weak power law with a \textit{strong} one. Namely, in this case the asymptotic behavior of $p(r)$ has the power order, i.e. inequality~(\ref{eq:two}) is valid (with ``matrix'' $\gamma$ defined above). Therefore, inequality~(\ref{eq:two}) takes place in a ``typical'' case of letter probabilities.
However one more natural question still remains without an answer.
Inequality~(\ref{eq:two}) means that the asymptotic form has a power \textit{order} but does not imply the \textit{exact} power asymptotics. In a general case, as follows from the first example given in this section, useful properties can be established neither when letters in words are Markov-dependent no when they are independent. However, as we prove later in this paper, in a ``typical'' case, for words composed of independent letters, the asymptotic behavior of the function~$p(r)$ is exact power. The following theorem is valid.
\begin{mytheorem}[main]
\label{th:main}
Let at least one of ratios $\log p_i/\log p_j$, $i,j\in\{ 1,\ldots,n\}$, be irrational and let $\gamma$ be the root of the equation $\sum_{i=1}^n p_i^\gamma=1$. Then the limit
\begin{equation}
\label{eq:thone}
\lim_{r\to\infty} \frac{p(r)^{-\gamma}/p_0^{-\gamma}}{r}
\end{equation}
exists and equals $H(\mathbf{p}^\gamma)$, where $H(\mathbf{p}^\gamma)$ is the entropy of $\mathbf{p}^\gamma=(p_1^\gamma,\ldots,p_n^\gamma)$, i.e.,
$
H(\mathbf{p}^\gamma)=-\gamma \sum_{i=1}^n p_i^\gamma \ln p_i.
$
\end{mytheorem}
Here and below we always write the function under consideration in the numerator and do the norming (defined analytically) function in the denominator of the fraction, whose limit is to be calculated. In intermediate calculations it may be more convenient to do the opposite, but since this results only in the trivial raising of the limit constant to the power of $-1$, we sacrifice the convenience of calculations for the clarity of statements of results. Evidently, the theorem asserts that under certain assumptions there exists the nonzero limit $p(r)/r^{-\alpha}$ (where $\alpha=1/\gamma$) as $r\to\infty$.
It is equal to
$
p_0 H(\mathbf{p}^\gamma)^{-1/\gamma}.
$
Let us describe the structure of the rest part of the paper. In Section~2 we state the main theorem in terms of multinomial coefficients (of the Pascal pyramid). The proof of the theorem is reduced to the estimation of the limit behavior of the sum of these coefficients over some simplex. In Section~3 we prove an analog of this theorem with an integral in place of the sum. In this section we essentially use the Stirling formula which allows us to reduce calculations to the evaluation of a multivariate Gaussian integral. We establish an explicit formula for the determinant of the matrix of the quadratic form that defines the integrand. Finally, in Section~4 we prove that the ratio of the integral to the sum tends to 1. Here we use the general properties of the Riemann integral and uniformly distributed sequences. In conclusion we discuss possible generalizations and unsolved problems.
\section{Equivalent statements of the main theorem and the Pascal pyramid}
Let us first note that if $p_0\to 0$, then $\gamma\to 1$.
Reducing the nominator of fraction~(\ref{eq:thone}) by $p_0^{-\gamma}$, we write the following statement in this case:
\begin{mytheorem}[the case of $\gamma=1$]
\label{th:main2}
Let $p_i>0$ be the probability of the symbol~$\omega_i$, $i=1,\ldots,n,$ while $\sum_{i=1}^n p_i=1$ (there is no stop symbol). Assume that at least one of ratios $\log p_i/\log p_j$, $i,j\in\{ 1,\ldots,n\}$, is irrational. Let us consider all possible finite words (including the empty one), sort them in the non-increasing order of probabilities (we equate the probability of the empty word to 1 and calculate the probability of any other word as the product of probabilities of its letters). Let $p(r)$ be the probability of the $r$th word in the list (the word with the \textit{rank} of~$r$). Then the limit $\lim_{r\to\infty} p(r)/r^{-1}$ exists and equals $H^{-1}(\mathbf{p})$, where $H(\mathbf{p})$ is the entropy of the vector $\mathbf{p}=(p_1,\ldots,p_n),$ i.e., $H(\mathbf{p})=-\sum_{i=1}^n p_i \ln p_i$.
\end{mytheorem}
In the statement of Theorem~\ref{th:main2}, as well as in Theorem~\ref{th:main}, we use the bold font for the vector whose components are denoted by the same letter with the index ranging from~1 to~$n$. In what follows we use the bold font for analogous denotations without mentioning this fact.
One can easily see that Theorem~\ref{th:main2} is not just a particular case of Theorem~\ref{th:main}, but these theorems are equivalent. Namely, the replacement of probabilities $p_i^\gamma$ with new ones $p_i$ turns the general case into the particular one. Therefore, in what follows we neglect $p_0$, assuming (without loss of generality) that ${\sum_{i=1}^n p_i\!=\!1}$.
Fix some probability $q\in (0, 1]$ and denote by $Q(q)$ the rank of the last word whose probability is not less than $q$ in the list of all words sorted in the non-increasing order of their probabilities. Let us redefine the function $p(r)$ for noninteger $r$ as $p(r)=p(\lfloor r \rfloor)$ (here $\lfloor \cdot \rfloor$ is the integer part of a number). Evidently, functions $q=p(r)$ and $r=Q(q)$ ($q\in (0, 1]$, $r\ge 1$) are inverse (more exactly, quasi-inverse),
namely, the graph of one of hyperbola-shaped, decreasing stepwise functions turns into another one when axes $r$ and $q$ switch roles (in the first case, $q$ is the argument and $r$ is the value, and vice versa in the second case).
It can be clearly seen
that $\lim_{r\to\infty} c\, p(r)/r^{-1}=1$ is equivalent to
\[\lim_{q\to 0} c^{-1} Q(q)/q^{-1}=1.\]
Therefore the equality in the assertion of Theorem~\ref{th:main2} is equivalent to that
\[
\lim_{q\to 0} Q(q)/q^{-1}= H^{-1}(\mathbf{p}).
\]
Denote the logarithm of the denominator in the last fraction by $z=-\ln q$ (i.e., $q=e^{-z}$) and let $\tilde Q(z)=Q(e^{-z})$. In view of considerations in the above paragraph the equality in the assertion of Theorem~\ref{th:main2} is equivalent to that
\begin{equation}
\label{eq:mainlimit}
\lim_{z\to \infty} (\ln \tilde Q(z)- z)=-\ln H(\mathbf{p}).
\end{equation}
Recall the proof of inequality~(\ref{eq:two}) in~\cite{we}. It is reduced to the proof of the boundedness of the difference $\ln \tilde Q(z)- z$ for the introduced function $\tilde Q(z)$ with $z\ge 0$. Nonnegative values of $z$ form the definition domain of the function $\tilde Q(z)$ because $q\le 1 \Leftrightarrow z\ge 0$. For convenience we redefine the function $\tilde Q(z)$ by putting $\tilde Q(z)=0$ for $z<0$.
Let $a_i=-\ln p_i$. Considering all possible variants of the last letters in words, whose quantity equals the value of the function $\tilde Q$, we obtain the functional equation $\tilde{Q}(z)=\tilde{Q}(z-a_1)+\ldots+\tilde{Q}(z-a_n)+\chi(z)$, where $\chi$ is the Heaviside step (i.e., the function that vanishes with negative values of the argument and equals 1 with nonnegative values). For $z \geq M = \max\{a_1,\ldots,a_n\}$ we get the following recurrent correlation:
\begin{equation}
\label{eq:rec}
Q_n(z)=Q_n(z-a_1)+\ldots+Q_n(z-a_n),
\end{equation}
where $Q_n(z)=\tilde{Q}(z)+1/(n-1)$.
The equality $\sum_{i=1}^n p_i=1$ implies that the function $\mbox{const} \exp z$ satisfies Eq.~(\ref{eq:rec}). Since
the function $Q_n(z)$ takes a finite number of positive values
within the $[0,M]$ interval, there exist positive $c_1$ and $c_2$ such that
\begin{equation}
\label{eq:eneq}
c_1\exp z<Q_n(z)<c_2\exp z
\end{equation}
for all $0\le z\le M$.
Replacing terms in the right-hand side of the recurrent correlation~(\ref{eq:rec}) with their lower (upper) bounds, we extend the solution set of inequality~(\ref{eq:eneq}) to the domain $0\le z\le M+m$, where $m=\min\{a_1,\ldots,a_n\}$. Repeating this procedure several times, in a finite number of steps we prove that the inequality is valid for any arbitrarily large~$z$. Performing the logarithmic transformation of the inequality, we conclude that $\ln Q_n(z)-z$ is bounded, and then so is the difference $\ln\tilde Q(z)-z$.
Let us return to Theorem~\ref{th:main2}. As was mentioned above,
Theorem~\ref{th:main2} asserts (under certain assumptions) not only the boundedness of $\ln\tilde Q(z)-z$ but also the validity of equality~(\ref{eq:mainlimit}).
Let us recall the combinatory sense of the function~$\tilde Q$; it is mentioned in~\cite{we}. Evidently, all words that contain $k_1$ letters of the 1st kind, $k_2$ letters of the 2nd kind, $\ldots$, and $k_n$ letters of the $n$th kind have one and the same probability of
$\mbox{Pr}(\mathbf{k}) = p_1^{k_1}\ldots p_n^{k_n}$ (i.e., $-\ln \mbox{Pr}(\mathbf{k})=\sum_{i=1}^n k_i a_i$), ranks of these words are consecutive. The quantity of such words is defined by the multinomial coefficient
\begin{equation}
\label{eq:M}
M(\mathbf{k}) = \frac{(k_1+\ldots+k_n)!}{k_1!\ldots k_n!}.
\end{equation}
Considering the nonnegative part of the $n$-dimensional integer grid and associating the point $(k_1\ldots,,k_n)$ with the number $M(k_1,...,k_n)$, we get one of variants of the Pascal pyramid. By the definition of the function $\tilde Q$ the value $\tilde Q(z)$ equals the sum of multinomial coefficients $M(\mathbf{k})$ over all integer vectors $\mathbf{k}$ that lie inside the $n$-dimensional simplex~$S(z)=\{\mathbf{x}: \mathbf{x}\ge 0,\ \sum_{i=1}^n a_i x_i\le z\}$:
\begin{equation}
\label{eq:tildeQ}
\tilde Q(z)=\sum_{\mathbf{k}\in S(z)} M(\mathbf{k}).
\end{equation}
As a result, we obtain one more equivalent statement of the main theorem, which we are going to prove.
\begin{mytheorem}[the multinomial statement]
\label{th:main3}
Let $a_i$, $i=1,\ldots,n,$ be arbitrary positive numbers such that at least one of ratios $a_i/a_j$, $i,j\in\{ 1,\ldots,n\}$ be irrational and $\sum_{i=1}^n p_i=1$, where $p_i=\exp(-a_i)$. Let a function $\tilde Q$ obey formula~(\ref{eq:tildeQ}). Then
\[
\lim_{z\to\infty} \frac{\tilde Q(z)}{\exp(z)}=H^{-1}(\mathbf{p}),
\]
where $H(\mathbf{p})=\sum_{i=1}^n a_i p_i$.
\end{mytheorem}
\section{The proof of an analog of Theorem~\ref{th:main3} with integration instead of summation}
\subsection{Reduction of the integration to the calculation of a Gaussian integral}
The function $M(k_1,\ldots,k_n)$ is defined for integer nonnegative vectors $\mathbf{k}$. Let us redefine it for noninteger vectors by replacing (in this case) $x!$ in Definition~(\ref{eq:M}) with $\Gamma(x+1)$. In what follows we use the denotation $M(x_1,\ldots,x_n)$ (or $M(\mathbf{x})$) for the corresponding function which is continuous for nonnegative $x_i$. Further we consider this function and study its properties only for such (nonnegative)~$x_i$.
In this section we prove the following theorem.
\begin{mytheorem}[on the integral]
\label{th:main4}
Let $a_i$, $i=1,\ldots,n,$ be arbitrary positive num\-bers such that $\sum_{i=1}^n p_i=1$, where $p_i=\exp(-a_i)$. Let a function $f(z)$ obey the formula~$f(z)=\int_{\mathbf{x}\,\in S(z)} M(\mathbf{x})\,d{\mathbf{x}}$, where $d{\mathbf{x}}=\prod_{i=1}^n dx_i$. Then
\[
\lim_{z\to\infty} \frac{f(z)}{\exp(z)}=H^{-1}(\mathbf{p}).
\]
\end{mytheorem}
{\bf Proof.}\
Let us first recall some evident properties of the integrand. Note that the existence of the (Riemann) integral of $f(z)$ over the compact set $S(z)$ evidently follows from the continuity of~$M(\mathbf{x})$ in the domain under consideration.
If all components of the vector $(x_1,\ldots,x_n)$, possibly, except one component~$x_i$, equal zero, then by definition we have $M(x_1,\ldots,x_n)\equiv 1$. Let us prove that otherwise the function $M(x_1,\ldots,x_n)$ is strictly increasing in~$x_i$. Since the gamma function is positive definite, it suffices to prove that in this case the partial derivative of $\ln M(x_1,\ldots,x_n)$ with respect to $x_i$ is positive. It equals
\[ (\ln \Gamma)'(x_1+\ldots+x_n+1)- (\ln \Gamma)'(x_i+1).\]
The positiveness of this difference follows from the fact that the function $(\ln \Gamma)'$ is increasing; this property, in turn, follows from the logarithmic convexity of the gamma function (it is well known~\cite{artin} that $(\ln \Gamma)''(x)=\sum_{i=0}^\infty \frac1{(i+x)^2}>0$ with~{$x>0$}).
The proved assertion implies that the function $M(\mathbf{x})$ attains its maximum in the domain $S(z)$ at the boun\-dary $\left\langle \mathbf{a},\mathbf{x} \right\rangle=z$, where $\left\langle \mathbf{a},\mathbf{x} \right\rangle=\sum_{i=1}^n a_i x_i$. Let us calculate the exact asymptotics of the maximal value of the function $M(\mathbf{x})$ in the domain $S(z)$ with $z\to\infty$. For the vector $\mathbf x$ we denote by $x$ the sum of its components and parameterize $\mathbf x$ by the value $x$ and ratios $q_i=x_i/x$:
\[x_i=q_i x, \ q_i\ge 0,\ i=1,\ldots,n,\quad \sum_{i=1}^n q_i=1.\]
Let us use one simplest corollary of the Stirling formula~\cite{artin}, namely, the fact that with a nonnegative argument the value of the difference $\ln \Gamma(x+1)-(x \ln(x)-x+\ln (x+1)/2)$ is bounded. We obtain that with any~$x>0$,
\begin{equation}
\label{eq:lnM}
\ln M(x_1,\ldots,x_n)=x H(\mathbf{q} )+ O(\ln(x+1)),\quad \mbox{where } H(\mathbf{q} )=-\sum_{i=1}^n q_i \ln q_i
\end{equation}
(this correlation is closely connected with the so-called entropy inequality for multinomial coefficients).
We seek for the maximum of this function with $z\to\infty$ under one additional condition (namely, the requirement that the maximum is attained at the boun\-dary) $\left\langle \mathbf{a},\mathbf{x} \right\rangle=z$, where $a_i=-\ln p_i$, $0<p_i<1$, $\sum_{i=1}^n p_i=1$. Since $a_i>0$, we get $O(\ln(x+1))=O(\ln z)$. Moreover, the condition $\left\langle \mathbf{a},\mathbf{x} \right\rangle=z$ with mentioned $x_i$ gives the correlation
\begin{equation}
\label{eq:xz}
x=z H(\mathbf{q}; \mathbf{p})^{-1},\quad \mbox{where }H(\mathbf{q}; \mathbf{p})=\sum_{i=1}^n a_i q_i=-\sum_{i=1}^n q_i\ln p_i.
\end{equation}
Substituting this expression in~(\ref{eq:lnM}), we conclude that the maximum of $\ln M$ (accurate to $O(\ln z)$) is attained at a vector~$\mathbf{q}$ such that the fraction $H(\mathbf{q})/H(\mathbf{q};\mathbf{p}))$ takes on the maximal value. Recall that the difference $H(\mathbf{q};\mathbf{p})-H(\mathbf{q})$ takes on only nonnegative values and is called the Kullback--Leibler distance (divergence) $D(\mathbf{q}\,|\,\mathbf{p})$ between distributions $\mathbf{q}$ and $\mathbf{p}$~(see \cite{Kelbert}). The minimum of this difference is attained at only one value of $\mathbf{q}=\mathbf{p}$; evidently, an analogous assertion is also true for $H(\mathbf{q};\mathbf{p})/H(\mathbf{q})$: if $\mathbf{q}\neq \mathbf{p}$
\begin{equation}
\label{eq:less1}
H(\mathbf{q})/H(\mathbf{q};\mathbf{p})<1.
\end{equation}
Consequently, the maximum of the function $\ln M(\mathbf{x})$ in the domain $S(z)$ is attained (accurate to $O(\ln z)$) at the intersection of the hyperplane $\left\langle \mathbf{a},\mathbf{x} \right\rangle=z$ with the straight line $x_i=p_i x$, $i=1,\ldots n,$ where it equals $z+O(\ln z)$.
Let us now immediately prove Theorem~\ref{th:main4}. Note first that by using the L'Hopital rule we can reduce the proof to that of the formula obtained by differentiating the $f(z)/\exp(z)$ numerator and denominator with respect to~$z$ and to the proof of the equality
\begin{equation}
\label{eq:delta}
\lim_{z\to\infty} \frac{\hat f(z)}{\exp(z)}=H^{-1}(\mathbf{p}),
\end{equation}
where
$
\hat f(z)=\int_{\mathbf{x}\geq 0} M(\mathbf{x})\delta(z-\left\langle \mathbf{a},\mathbf{x} \right\rangle)\,d{\mathbf{x}},
$
and $\delta(\cdot)$ is the delta function.
Let~$\varepsilon$ be a real arbitrarily small positive value. Denote by $\Lambda_\varepsilon$ the sector consisting of points $\mathbf{x}$,
$x_i=q_i x$, $\sum_i q_i=1$, such that
\begin{equation}
\label{eq:eneqq}
p_i-\varepsilon<q_i<p_i+\varepsilon, \quad i=1,\ldots,n.
\end{equation}
With fixed $z$ on the hyperplane $\left\langle \mathbf{a},\mathbf{x} \right\rangle=z$ correlations~(\ref{eq:lnM}) and~(\ref{eq:xz}) take the form
\begin{equation}
\label{eq:lnm}
\ln M(\mathbf{x})=z H(\mathbf{q} )/H(\mathbf{q}; \mathbf{p})+ O(\ln(z)).
\end{equation}
Let us now strengthen inequality~(\ref{eq:less1}), namely, let us prove that if for $\mathbf{q}$ cor\-re\-la\-tions~(\ref{eq:eneqq}) are violated, then
\begin{equation}
\label{eq:HH}
H(\mathbf{q} )/H(\mathbf{q}; \mathbf{p})<1-C_1(\mathbf{p}) \varepsilon^2,
\end{equation}
where $C_1(\mathbf{p})$ is a positive constant independent of $\mathbf{q}$.
Since $H(\mathbf{q}; \mathbf{p})$ is a convex combination of $-\ln p_i$, it evidently is bounded:
\[0<\min\nolimits_{i} (-\ln p_i) \leq H(\mathbf{q}; \mathbf{p}) \leq \max\nolimits_{i} (-\ln p_i). \]
Consequently, formula~(\ref{eq:HH}) is equivalent to the inequality
\[
H(\mathbf{q}; \mathbf{p})-H(\mathbf{q})=D(\mathbf{q}\,|\,\mathbf{p})> C_2 (\mathbf{p}) \varepsilon^2.
\]
The latter correlation follows from the well-known property of the Kullback--Leibler divergence
\[
D(\mathbf{q}\,|\,\mathbf{p})\geq \frac14 \left( \sum_{i=1}^n |p_i-q_i| \right)^2
\]
(see, for example, lemma~3.6.10 in \cite{Kelbert}).
The proved inequality~(\ref{eq:HH}) (in view of formula~(\ref{eq:lnm})) implies that outside the domain $\Lambda_\varepsilon$ the function $M(\mathbf{x})$ is exponentially small in comparison to the maximal value inside the domain which equals $\exp(z)$. More precisely, with $x\not\in\Lambda_\varepsilon$, $\left\langle \mathbf{a},\mathbf{x} \right\rangle=z$ we get
\begin{equation}
\label{eq:eneqqq}
M(\mathbf{x})<\exp\left\{(1-C \varepsilon^2)z\right\}\quad \mbox{for some }~C>0.
\end{equation}
Note that the condition of the exponential smallness in comparison to $\exp z$ remains valid, even if $\varepsilon$ depends on $z$ and tends to 0 as~$z$ increases, though not too fast. In what follows we assume that
\[
\varepsilon=\varepsilon(z)=z^{-1/2+\delta}, \mbox{ where } \delta>0 \mbox{ is sufficiently small. }
\]
One can easily see that the same exponential upper bound as in~(\ref{eq:eneqqq}) also takes place
not only for the $M$ function but also for its integral over the domain which volume grows according to a
power law:
\[
\int_{\mathbf{x}\not\in \Lambda_{\varepsilon(z)},\mathbf{x}\geq 0} M(\mathbf{x})\delta(z-\left\langle \mathbf{a},\mathbf{x} \right\rangle)\,d{\mathbf{x}}\,<\,\exp\left\{(1-C \varepsilon^2)z\right\}
\]
with $z\to\infty$. Therefore in limit~(\ref{eq:delta}) we can treat $\hat f(z)$ as the integral
\begin{equation}
\label{eq:hatf}
\int_{\mathbf{x}\in \Lambda_{\varepsilon(z)}} M(\mathbf{x})\delta(z-\left\langle \mathbf{a},\mathbf{x} \right\rangle)\,d{\mathbf{x}}.
\end{equation}
Let us define the asymptotics~(\ref{eq:lnM}) of the function $M(\mathbf{x})$ in the domain~$\Lambda_{\varepsilon(z)}$ more precisely. Let us use the standard Stirling formula, namely, the fact that with $x\to\infty$ it holds $\ln \Gamma(x+1)=x \ln(x)-x+\ln(x)/2+\ln(2\pi)/2+R(x)$, where $0<R(x)<1/(12x)$. We obtain that in the domain~$\Lambda_{\varepsilon(z)}$,
\[
M(\mathbf{x})=\frac{1}{\sqrt{(2\pi)^{n-1}}}\exp\left\{ x H(\mathbf{q})+\ln(x)/2-\sum\nolimits_{i=1}^n \ln( x_i)/2+O(1/z)\right\}.
\]
Here, as usual, $x=\sum\nolimits_{i=1}^n x_i$; $q_i=x_i/x$. Therefore, we conclude that when considering the asymptotics of function~(\ref{eq:hatf}) we can treat $M(\mathbf{x})$ as follows:
\begin{equation}
\label{eq:widetildeM}
\widetilde M(\mathbf{x})=\frac{1}{\sqrt{(2\pi)^{n-1}}}\exp\left\{ x H(\mathbf{q})+\ln(x)/2-\sum\nolimits_{i=1}^n \ln( x_i)/2\right\}.
\end{equation}
In the latter formula we can write the exponent as
\begin{equation}
\label{eq:lefrig}
\left\{ \phantom{\sum\nolimits_{i=1}^n}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \right\}=x\ln x+\ln(x)/2-
\sum\nolimits_{i=1}^n\left( x_i \ln x_i+\ln(x_i)/2\right).
\end{equation}
Let us write the Taylor expansion up to second-order terms near the maximum point in the plane $\left\langle \mathbf{a},\mathbf{x} \right\rangle=z$, i.e., near the point $\mathbf{x}'=\mathbf{p} z / H(\mathbf{p})$ (in what follows we denote by $x'_i$ coordinates of the point $\mathbf{x}'$ and do by $x'$ the sum of these coordinates which evidently equals $z H(\mathbf{p})^{-1}$).
First of all, note that
\[
\widetilde M(\mathbf{x}')=
\frac{1}{\sqrt{(2\pi x')^{n-1}\prod_{i=1}^n p_i}}\exp(z).
\]
One can easily calculate second derivatives of expression~(\ref{eq:lefrig})
\[
\frac{\partial^2}{\partial x_i\partial x_j}\left\{ \phantom{\sum\nolimits_{i=1}^n}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \right\}=
\left\{
\begin{array}{ll}
x^{-1}-(2x^2)^{-1}, & \mbox{ if }i\neq j,\\
x^{-1}-(2x^2)^{-1}-x_i^{-1}+(2x_i^2)^{-1},& \mbox{ otherwise}
\end{array}
\right.
\]
(note that we do not use first derivatives in the Taylor expansion near the maximum point).
If $x\in\Lambda_\varepsilon$, then by formula~(\ref{eq:xz}) we have $x-x'=z (H(\mathbf{q}; \mathbf{p})^{-1}-H(\mathbf{p})^{-1})=z O(\varepsilon)$ (in the latter inequality we use the continuity of the function $H(\mathbf{q}; \mathbf{p})^{-1}$). Consequently,
\begin{equation}
\label{eq:ximxsi}
x_i-x'_i=x q_i-x' p_i=(x'+z O(\varepsilon))(p_i+O(\varepsilon))-x' p_i=z O(\varepsilon).
\end{equation}
In particular, with chosen $\varepsilon=\varepsilon(z)$ we have $|x_i-x'_i|=O(z^{1/2+\delta})$. We obtain that in the domain $\Lambda_{\varepsilon(z)}$,
\begin{eqnarray*}
\widetilde M(\mathbf{x})&=&\frac{1}{\sqrt{(2\pi x')^{n-1}\prod_{i=1}^n p_i}}\exp(z)\times \\
&&\exp\left\{\frac{\sum_{i,j=1}^n (x_i-x'_i)(x_j-x'_j)}{2 x'}-\frac{\sum_{i=1}^n(x_i-x'_i)^2}{2 p_i x'} +O(z^{-1/2+3\delta}) \right\}.
\end{eqnarray*}
Here the term $O(z^{-1/2+3\delta})$ contains both the remainder of terms of the series whose order exceeds 2 and the value of $O(z^{-1+2\delta})$ added by some omitted second-order terms. With $z\to\infty$ we can neglect the term of $O(z^{-1/2+3\delta})$. Therefore, in integral~(\ref{eq:hatf}) in place of $M(\mathbf{x})$ we should substitute the function $\widehat M(\mathbf{x})$ which differs from $\widetilde M(\mathbf{x})$ in the fact that its exponent does not contain the term of $O(z^{-1/2+3\delta})$.
Let us change variables in the integral as follows: $y_i=(x_i-x'_i)/\sqrt{x'}.$ Since the degree of homogeneity of the delta-function equals~$-1$, we obtain that limit~(\ref{eq:delta}) coincides with
\begin{equation}
\label{eq:K}
\frac{1}{\sqrt{(2\pi)^{n-1}\prod_{i=1}^n p_i}}
\int_{\mathbb{R}^n} \delta(\left\langle \mathbf{a},\mathbf{y} \right\rangle)\exp\{-\left\langle \mathcal{B} \mathbf{y},\mathbf{y} \right\rangle/2\}\,d\mathbf{y},
\end{equation}
where $\mathcal{B}$ is the $n\times n$ matrix, all whose elements equal $-1$, except diagonal components which are greater by $1/p_i$.
\subsection{Calculation of the determinant}~
\begin{mylemma}
\label{lem:one1}
Let $n\ge 2$. Consider the $n\times n$ matrix $B$, where all nondiagonal elements equal 1, while $b_{ii}=1+k_i$. Then
1. The determinant of this matrix equals $\prod_{i=1}^n k_i \left(1+\sum_{j=1}^n 1/k_j\right)$.
2. The algebraic complement of the element with indices $(i,j), i\ne j,$ equals
\[-\prod_{\ell\in [n]\setminus\{i,j\}} k_\ell,\quad \mbox{where }[n]= \{1,\ldots,n\}.\]
\end{mylemma}
\begin{myCo}
The matrix $\mathcal{B}$ in formula~(\ref{eq:K}) is degenerate.
\end{myCo}
{\bf Proof of Lemma~\ref{lem:one1}.}\
Note that the first item of Lemma~\ref{lem:one1} defines the value of the algebraic complement of the diagonal element of such a matrix. Let us prove the theorem by induction.
With $n=2$ in the formula in item~2 we get the product over the empty set; it is accepted that this product equals 1. The formula in item~1 remains valid with $n=1$. In the induction step we assume that the formula in item~1 is proved for all dimensions less than~$n$ and has to be proved for the case when the dimension equals $n$, while the formula in item~2 is proved for all dimensions not greater than $n$ and has to be proved for the $(n+1)\times (n+1)$ matrix.
For proving item~1 we can use the expansion by the last row. Multiplying the algebraic complement by the diagonal element $k_n+1$, we get the sum
\begin{eqnarray}
&\prod_{i=1}^n k_i \left(1+\sum_{j=1}^{n-1} 1/k_j\right)
+ \prod_{i=1}^{n-1} k_i\left(1+ \sum_{j=1}^{n-1} 1/k_j \right)=& \nonumber
\\
\label{eq:sum1}
&\prod_{i=1}^n k_i \left(1+\sum_{j=1}^{n-1} 1/k_j\right)
+ \prod_{i=1}^{n-1} k_i+\sum_{j=1}^{n-1}\prod_{\ell\in [n]
\setminus\{n,j\}} k_\ell.&
\end{eqnarray}
The expansion by the entire last row, taking into account the induction hypothesis for item~2, make the third part in row~(\ref{eq:sum1}) vanish. First two terms in formula~(\ref{eq:sum1}) together give the desired sum.
In order to prove item~2, let us expand the determinant considered in this item (algebraic complement of the element with $(i,j)$ indices of the
matrix~$B$ with the $(n+1)\times(n+1)$ dimension)
by the row whose number in the initial matrix of~$B$ was equal to~$j$. Generally speaking, for clarity, we use the same indices as in the numeration of the initial matrix. Since the algebraic complement considered in this item and the occurring algebraic complement for the element with indices $(j,i)$ (obtained by the expansion by a row of the determinant under consideration) have opposite signs, the value added by the element with indices $(j,i)$ equals
\[
-\prod_{r\in [n+1]\setminus \{i,j\}} k_r \left(1+\sum_{\ell\in [n+1]\setminus\{i,j\}} 1/k_\ell\right)
\]
(here we have used the induction hypothesis for item~1). The difference from the desired formula consists in the last term which equals (taking into account the first multiplier)
\[
-\sum_{\ell\in [n+1]\setminus\{i,j\}}\quad \prod_{r\in [n+1]\setminus\{i,j,\ell\}} k_r.
\]
It vanishes, when taking into account the contribution of the remaining $n-1$ elements in the $j$th row of the considered matrix.
\quad$\square$
\begin{mylemma}
\label{lem:twoo}
Let $B_1$ be the matrix mentioned in Lemma~\ref{lem:one1} (its dimension is $n\times n$, $n\geq 2$). Assume that $b_{ii}=1-1/p_i$, $i=1,\ldots n,$ where $p_i$ are arbitrary nonzero numbers. Denote by $B_2$ a matrix of the same dimension in the form $a^T a$, where $a=(a_1,\ldots a_n)$ is an arbitrary numeric row and $T$ is the transposition sign. Let $s$ be an arbitrary real number. Then
\[
\det\left( s B_2-B_1\right) =
\frac{s
\left(\left(\sum _{i=1}^n a_i p_i \right)^2-\left(\sum
_{j=1}^n p_j-1\right) \sum _{i=1}^n a_i^2
p_i\right)}{\prod _{\ell=1}^n p_\ell}+\det(-B_1).
\]
\end{mylemma}
\begin{myCo}
\label{co:conc2}
If a vector $\mathbf{p}=(p_1,\ldots,p_n)$ satisfies additional constraints $p_i>0$, $\sum_{i=1}^n p_i=1$ (i.e., $-B_1=\mathcal{B}$), while $a_i=-\ln p_i$, then
\[
\sqrt{s\ \det\left( s^{-1} B_2-B_1\right)} =
\frac{H(\mathbf{p})}{\sqrt{\prod\nolimits _{\ell=1}^n p_\ell}}.
\]
\end{myCo}
{\bf Proof of Lemma~\ref{lem:twoo}.}\
By the differentiation rule for determinants, the derivative of the determinant of an $n\times n$ matrix equals the sum of determinants of~$n$ matrices such that in the $i$th one all elements of the $i$th row are replaced with their derivatives. We obtain that $\frac{\partial^2 \mbox{\rm det\,}\left( s B_2-B_1\right)}{\partial s^2}$ is the sum of determinants of matrices each one of which contains either the zero row or two various rows of the matrix~$B_2$. Since $\mbox{rank}\, B_2=1$, we get $\frac{\partial^2 \det\left( s B_2-B_1\right)}{\partial s^2}=0$.
Thus, $\det\left( s B_2-B_1\right)$ is a linear function of~$s$, whose free term evidently equals $\det(-B_1)$. It is clear that for calculating the coefficient $\det\left( s B_2-B_1\right)$ at $s$ it suffices to summate products of each element of the matrix $B_2$ by the algebraic complement of the corresponding element of the matrix~$-B_1$. If an element has indices $(i,j)$, $i\ne j,$ then by item~2 of Lemma~\ref{lem:one1} this product equals $a_i a_j p_i p_j/\prod _{\ell=1}^n p_\ell$.
Let`s explain the positive sign in the last formula. We calculate
an algebraic complement of the $-B_1$ matrix element. The
matrix has the $n\times n$ dimension, therefore the found algebraic
complement differs from the algebraic complement of the
corresponding $B_1$ matrix element for $(-1)^{n-1}$ times. According to
item~2 of Lemma~\ref{lem:one1}, the algebraic complement of the
corresponding $B_1$ matrix element is a ``minus'' product of $n-2$
multipliers~$k_i$. In the given case each of~$k_i$ factors is negative
(equals $-1/p_i$) which results in positive sign of the last formula in
the above paragraph.
Assume that this formula is valid for all $(i,j)$. Then we get the sum
\begin{equation}
\label{eq:two2}
\sum_{i,j=1}^n a_i a_j p_i p_j/\prod _{\ell=1}^n p_\ell=
\left(\sum _{i=1}^n a_i p_i \right)^2/\prod _{\ell=1}^n p_\ell.
\end{equation}
However by item~1 of Lemma~\ref{lem:one1} the algebraic complement of the diagonal element $b_{ii}$ of the matrix $-B_1$ equals
\begin{equation}
\label{eq:thr}
(-1)^{n-1}\left( \prod _{j:j\ne i} (-1/p_j)+\sum_{j:j\ne i}\prod _{\ell\not\in\{i,j\}}(-1/p_\ell)
\right)
\end{equation}
(here and below we omit the evident requirement that values of all indices belong to the set $[n]$).
Multiplying the first term in parentheses, i.e., $\prod _{j:j\ne i} (-1/p_j)$, by $(-1)^{n-1}a_i^2$ and summing over all~$i$, we get $\sum _{i=1}^n a_i^2 p_i/\prod _{\ell=1}^n p_\ell$. Let us multiply the resting term in parentheses~(\ref{eq:thr}) by $(-1)^{n-1}a_i^2$, sum over all~$i$, and subtract the value
\[\sum _{i=1}^n a_i^2 p_i^2/\prod _{\ell=1}^n p_\ell\]
from the obtained result (note that the subtrahend was ``illegally'' included in formula~(\ref{eq:two2})).
It gives the overall contribution of the second term in formula~(\ref{eq:thr}), which equals
\[
-\sum_{j=1}^n p_j \sum _{i=1}^n a_i^2
p_i /\prod _{\ell=1}^n p_\ell .
\]
Taking into account all the calculation elements of the
desired determinant allows completing the proof of Lemma~\ref{lem:twoo}.
\quad$\square$
For completing the proof of Theorem~\ref{th:main4} let us use Corollary~\ref{co:conc2}. Let us replace the variable in integral~(\ref{eq:K}) (as was proved earlier, this integral equals the limit considered in Theorem~\ref{th:main4})
$\delta(t)=\lim_{\sigma\to 0} \frac{1}{\sqrt{2\pi}\sigma}\exp\{-\frac{t^2}{2\sigma^2}\}$. Treating the limit multiplied by the coefficient at the exponent as a multiplier in the integral, we come to the limit of the Gaussian integral
\[
\lim_{\sigma\to 0}
\frac{1}{\sigma\sqrt{(2\pi)^{n}\prod_{i=1}^n p_i}}
\int_{\mathbb{R}^n} \exp\{-\left\langle (\sigma^{-2} B_2+\mathcal{B})\, \mathbf{y},\mathbf{y} \right\rangle/2\}\,d\mathbf{y},
\]
i.e.,
\[
\lim_{\sigma\to 0}
\frac{1}{\sigma \sqrt{\prod_{i=1}^n p_i\, \det\left( \sigma^{-2} B_2+\mathcal{B}\right)}}.
\]
Immediately applying Corollary~\ref{co:conc2}, we get desired~$H^{-1}(\mathbf{p})$.
\quad$\square$
\section{The ratio between the sum and the integral}
It remains to prove that under assumptions of Theorem~\ref{th:main3}, the ratio of the integral of the function~$M$ calculated over the domain $S(z)$ to the sum of values of this function at integer points of this domain tends to 1 as $z\to\infty$. For comparing the integral of the function and the sum of its values in the same domain one usually applies the Koksma--Hlawka inequality~(see \cite{niderraiter}).
Note that usually one considers the integral over a fixed domain (as a rule, the cube $[0,1]^n$), whereas the domain in the case under consideration is varying. However, we intend only to prove the convergence of the fraction to 1 and do not need to estimate the asymptotic difference between the integral and the sum, which simplifies the task.
Evidently, it suffices to calculate the limit of the ratio for an arbitrary infinite increasing sequence $z_1,z_2,\ldots$ such that $z_i\to\infty$.
\begin{mytheorem}
\label{th:sumInt}
Let $\Omega_1,\Omega_2,\ldots$ be a sequence of Jordan measurable sets such that $\Omega_i\subset\Omega_{i+1}$ for all $i=1,2,\ldots$ Assume that $f(x)$, $x\in \Omega$, where $\Omega=\bigcup_i \Omega_i$, is an integrable and bounded on each of domains~$\Omega_i$ function such that $f(x)\ge 0$ and $\int f(x)\,d\, \Omega_i\to\infty$ as $i\to\infty$. Assume also that $K$ is a countable set of points from~$\Omega$ such that each of sets $K_i=K\cap \Omega_i$ is finite. Then if for any sufficiently small $\alpha>0$ there exists a partition of $\Omega$ onto a countable number of Jordan measurable sets $X_j=X_j(\alpha)$, $j=1,2,\ldots$, such that $\Omega_i=\bigcup_{j=1}^{n_i} X_j$ for some $n_i=n_i(\alpha)$, while
\begin{eqnarray}
\label{eq:supinf}
\sup_{x\in X_j}f(x)/\inf_{x\in X_j}f(x)<1+\alpha& &\mbox{starting with some } j,\\
\label{eq:limK}
\frac{|K\cap X_j|}{\mu X_j}\to 1& &\mbox{as } j\to\infty,
\end{eqnarray}
then in this case there exists the limit
\[
\lim_{i\to\infty} \frac{\int f(x)\,d\,\Omega_i}{\sum_{x\in K_i} f(x)}=1.
\]
\end{mytheorem}
{\bf Proof.}\
Evidently,
\[
\mu X_j \inf_{x\in X_j}f(x)\le \int f(x)\,d\,X_j \le \mu X_j \sup_{x\in X_j}f(x).
\]
Therefore, in view of~(\ref{eq:supinf}) we conclude that starting with some $j$ it holds
\[
\frac{\int f(x)\,d\,X_j/\mu X_j}{\sum_{x\in K\cap X_j} f(x)/|K\cap X_j|}\in (1-\alpha,1+\alpha)
\]
with $\alpha<1$. In accordance with~(\ref{eq:limK}) we conclude that $\frac{\mu X_j}{|K\cap X_j|}\in (1-\alpha,1+\alpha)$ for all $j$, except a finite number of values of the index. Therefore, there exists $\ell$ such that with all~$j>n_\ell$,
\[
\frac{\int f(x)\,d\,X_j}{\sum_{x\in K\cap X_j} f(x)}\in ((1-\alpha)^2,(1+\alpha)^2).
\]
Representing this correlation as a double inequality and summing it over all~$j$ from $n_\ell+1$ to $n_i$, we obtain
\[
\frac{\int f(x)\,d(\Omega_i\setminus \Omega_\ell)}{\sum_{x\in K_i\setminus K_\ell} f(x)}\in ((1-\alpha)^2,(1+\alpha)^2)
\]
with $i>\ell$.
Note that by condition the numerator in the latter fraction (different from the integral $\int f(x)\,d\,\Omega_i$ by a constant value) tends to infinity. Then the same is true for the denominator. Note that the denominator differs from $\sum_{x\in K_i} f(x)$ by a constant value.
Therefore we conclude that all limit points of the sequence $\frac{\int f(x)\,d\,\Omega_i}{\sum_{x\in K_i} f(x)}$ lie inside the interval $((1-\alpha)^2,(1+\alpha)^2)$. Due to the arbitrariness of the choice of positive~$\alpha$ Theorem~\ref{th:sumInt} is proved.
\quad$\square$
\begin{myCo}[Completion of the proof of the main theorem]
\label{co:mainconc}
Let $f(z)$ be the function mentioned in assumptions of Theorem~\ref{th:main4}
and let $\tilde Q(z)$ obey fo\-r\-mu\-la~(\ref{eq:tildeQ}). Then if at least one of ratios $a_i/a_k$, $i,k\in\{ 1,\ldots,n\}$, $i\neq k,$ is irrational, then
\[
\lim_{z\to\infty} \frac{f(z)}{\tilde Q(z)}=1.
\]
\end{myCo}
{\bf Proof.}\
For clarity we denote by~$z$ the parameter that defines the boundary of the considered domain, and do by $\zeta$ the corresponding parameter of the hyperplane that contains a certain interior point $\mathbf{x}$ of this domain, i.e., $\zeta(\mathbf{x})=\left\langle \mathbf{a},\mathbf{x} \right\rangle$.
First of all, note that considerations in Section~3.1 imply that both in the sum and in the integral we can replace $S(z)$ with the domain
\[
\widehat \Lambda(z)= S(z)\cap \Lambda_{\varepsilon(\zeta)},\quad \mbox{ where }\varepsilon(\zeta)=\zeta^{-1/2+\delta},
\]
and replace the function $M(\mathbf{x})$ with that $\widetilde M(\mathbf{x})$ defined by formula~(\ref{eq:widetildeM}).
There\-fo\-re, we need to prove that
\begin{equation}
\label{eq:inttosumz}
\frac{\int_{\mathbf{x}\,\in \widehat \Lambda(z)}
\widetilde M(\mathbf{x})\,d{\mathbf{x}}}{\sum_{\mathbf{k}\,\in \widehat \Lambda(z)}\widetilde M(\mathbf{k})}\to 1
\end{equation}
(or that the difference of logarithms of the numerator and denominator tends to zero).
In view of Theorem~\ref{th:main4} the logarithm of the numerator in the latter fraction is a uniformly continuous function of $z$, while the logarithm of the denominator evidently is a nondecreasing function.
Therefore for proving the existence of the limit with $z\to\infty$ it suffices
to prove the existence of the limit for a sequence in the form $z_n=\kappa n$, $n=1,2,\ldots$,
where $\kappa$ is an arbitrarily small positive value
(as the difference between the numerator and denominator
of the logarithms in an arbitrary point slightly differs from the
value of difference in the nearest points $z_n$ in this sequence).
Namely, just for this fixed sequence we consider the ratio from the right-hand side of~(\ref{eq:inttosumz}).
In order to apply Theorem~\ref{th:sumInt}, for an arbitrary sufficiently small positive $\alpha$ we construct a partition of $\Lambda_{\varepsilon(\zeta)}$ onto domains $X_j$ satisfying assumptions of the theorem. Namely, we construct this partition by dividing of an infinite quantity of ``flapjacks'' located between neighboring hyperplanes in the form $\zeta({\mathbf{x}})=c_r$ and $\zeta({\mathbf{x}})=c_{r+1}$, $r=1,2,\ldots,$ where $c_{r+1}=c_r+\mbox{Const},$ onto a finite number of domains~$X_j$.
Evidently, for any $\alpha\le 2\kappa$ we can choose a sequence $c_r$ such that
\[
c_{r+1}-c_r=\mbox{Const}< \frac{\alpha}{2};
\quad \mbox{for any }n\
\exists\, r: z_n=c_r.
\]
To this end, it suffices to put $c_r=\mbox{Const}\,r$, where $\mbox{Const}= \kappa/ \lceil \frac{2\kappa}{\alpha} \rceil$
(here $\lceil \cdot \rceil$ is an upward rounding to the nearest integer).
Let $C_r=\{\mathbf{x}: c_r\le\zeta(\mathbf{x})< c_{r+1} \}$. Denote by $F_r$ the $r$th ``flapjack'' $C_r \cap \Lambda_{\varepsilon(\zeta)}$. We are going to ``cut'' $F_r$ onto a finite number of domains~$X_j$. We numerate the countable number of domains $X_j$, $j=1,2,\ldots,$ so as to make domains $X_j$ obtained by ``cutting'' $F_r$ with the least~$r$ have lesser numbers, while the order of numbering inside the partition of $F_r$ plays no role.
Since $\varepsilon(\zeta)=o(\zeta)$, with $\mathbf{x}, \mathbf{y}\in F_r$ it holds: $y_i=x_i+o(x_i)$ (cf. with~(\ref{eq:ximxsi})). Consequently, with $r\to\infty$ we get $\ln y_i-\ln x_i \to 0$ and $\ln y-\ln x \to 0$.
By formula~(\ref{eq:widetildeM}),
\[
\ln \widetilde M(\mathbf{x})=\mbox{const}+
g(\mathbf{x}) +\ln(x)/2-\sum\nolimits_{i=1}^n \ln( x_i)/2,
\]
where $g(\mathbf{x})=x H(\mathbf{q})=\sum_{i=1}^n x_i \ln x_i-\left(\sum_{i=1}^n x_i\right)\ln \left( \sum_{j=1}^n x_i\right).$
We get
\[
\mbox{grad} g=\ln \mathbf{q}=(\ln q_1,\ldots,\ln q_n),\quad
\frac{\partial^2 g}{\partial x_i\partial x_j}=O(1/x),\ i,j\in\{1,\ldots,n\}.
\]
Using expansion in a series $\ln \widetilde M$ with evaluation of the second
order terms and considerations of the previous paragraph we obtain the following important observation.
If $\mathbf{x}, \mathbf{y}\in F_r$~and
\begin{equation}
\label{eq:cr}
|x_i-y_i|=o(\sqrt{x})=o(\sqrt{c_r}), \quad i=1,\ldots,n,
\end{equation}
then with sufficiently large $r$ it holds
\[
\ln \widetilde M(\mathbf{x})-\ln \widetilde M(\mathbf{y})<
|\left\langle \ln \mathbf{q}\,,\mathbf{x}-\mathbf{y} \right\rangle|+\alpha/100.
\]
Since $\mathbf{q}\to \mathbf{p}$ as $r\to\infty$, with sufficiently large~$r$ it holds
\[
|\left\langle \ln \mathbf{q}\,,\mathbf{x}-\mathbf{y} \right\rangle|\,<\,
|\left\langle \mathbf{a}\,,\mathbf{x}-\mathbf{y} \right\rangle|+\alpha/200\,<\,
(c_{r+1}-c_r)+\alpha/200.
\]
As a result, we obtain that with sufficiently small $\alpha$, starting with some~$r$, it holds
\[
\widetilde M(\mathbf{x})/\widetilde M(\mathbf{y})<1+\alpha .
\]
Therefore, dividing $F_r$ onto domains $X_j$ so as to fulfill correlation~(\ref{eq:cr}) for all points $\mathbf{x},\mathbf{y}$ that belong to one domain, we guarantee the validity of assumption~(\ref{eq:supinf}) in Theorem~\ref{th:sumInt}. Note that it suffices to fulfill condition~(\ref{eq:cr}) for all indices~$i$ except one, because the validity of this condition for the rest index follows from the fact that $\mathbf{x},\mathbf{y}\in C_r$.
Finally, let us use the irrationality of $a_{i^*}/a_{k^*}$ for some $i^*\neq k^*$. Let us denote by $I_{k^*}$ the set $\{ 1,\ldots,n\}\setminus\{ k^* \}$ and do by $I_{i^*k^*}$ the set $\{ 1,\ldots,n\}\setminus\{i^*,k^* \}$. We are going to prove that defining domains $X_j$ by inequalities
\begin{equation}
\label{eq:IK}
l_{ji}\le x_i<L_{ji},\ i\in I_{k^*}, \qquad \mbox{where } L_{ji}-l_{ji}>\mbox{const } c_r^{1/2-\delta},
\end{equation}
we fulfill condition~(\ref{eq:limK}) (with $K=\mathbb{Z}^n$). Here, as usual, $\delta$ is a sufficiently small real positive value, though in this case we can choose $\delta$ as any number in the interval $(0,1/2)$
(roughly speaking, it`s sufficient that the radius of the pieces $X_j$
used to divide ``flapjacks'' $F_r$ tend to infinity at~$r\to\infty$).
Evidently, we can divide ``almost all'' $F_r$ onto domains~$X_j$ so as to simul\-ta\-neo\-usly fulfill inequalities~(\ref{eq:cr}) and conditions~(\ref{eq:IK}) on~$l$ and~$L$ (the remaining ``cuttings'' on the edges of the domain $F_r$ which occur due to the inconsistency between the inequality $l_{ji}\le x_i<L_{ji},\ i\in I_{k^*}$ and the definition of the boundary of the domain $\Lambda_{\varepsilon(z)}$ are asymptotically small).
Evidently, $\mu X_j=\prod_{i\in I_{k^*}} (L_{ji}-l_{ji})\times (c_{r+1}-c_r)/a_{k^*}$. Since the difference $(L_{ji}-l_{ji})$ grows as $j\to\infty$, \textit{the asymptotics} of the number of ways for choosing integer $x_i$ such that
$l_{ji}\le x_i<L_{ji}$ for $i\in I_{i^*k^*}$ coincides with $\prod_{i\in I_{i^*k^*}} (L_{ji}-l_{ji})$. Here and below we understand the asymptotics as a function of~$j$ such that the ratio of the considered quantity to this function tends to 1 as $j\to\infty$. In order to complete the proof of Corollary~\ref{co:mainconc}, it remains to prove the following lemma.
\begin{mylemma}
\label{lem:finlemma}
Let the ratio $a_{i*}/a_{k^*}$ be irrational and $(L_{ji^*}-l_{ji^*})\to\infty$. Assume also that the ratio $(c_{r+1}-c_r) /a_{k^*}$ equals a constant value lesser than 1 which is independent of~$r$. Then for fixed $x_i,\ i\in I_{i^*k^*},$ the asymptotics of the number of ways to choose integer $x_i$, $i\in \{i^*,k^*\},$ such that $l_{ji^*}\le x_{i^*}<L_{ji^*}$ and $c_r\le\zeta({\mathbf x})<c_{r+1}$ simultaneously, equals $(L_{ji}-l_{ji})\times (c_{r+1}-c_r)/a_{k^*}$.
\end{mylemma}
{\bf Proof of Lemma~\ref{lem:finlemma}.}\
In what follows we need standard denotations for the fractional part $\{\cdot\}$, floor $\lfloor\cdot\rfloor$, and ceil $\lceil\cdot\rceil$ of a number.
Let $c'=\sum_{i\in I_{i^* k^*}} a_i x_i$,
$d_r=(c_r-c')/a_{k^*}$,
$D_r=(c_{r+1}-c')/a_{k^*}$, and $\theta=a_{i*}/a_{k^*}$. The condition $c_r\le\zeta({\mathbf x})<c_{r+1}$ is equivalent to that
\begin{equation}
\label{eq:fincond}
\theta\, x_{i^*}+x_{k^*}\in [d_r,D_r).
\end{equation}
If the difference $D_r-d_r$ (it equals $(c_{r+1}-c_r) /a_{k^*}$) is less than~1
(this inequality is obviously holds for sufficiently small $\alpha$) then with fixed $x_{i^*}$ the integer value $x_{k^*}$ satisfying condition~(\ref{eq:fincond}) is defined uniquely, provided that it exists. Therefore, we need to estimate the quantity of values $x_{i^*}$ in the interval $[l_{ji^*},L_{ji^*})$ such that $\{ \theta\, x_{i^*}\}\in [\{d_r\},\{ D_r\})$; here the latter correlation is understood in the sense of an interval on the unit circle, and the length of the considered interval is independent of~$r$.
Recall the definition of a well-distributed sequence~\cite[section 1.5]{niderraiter}.
``Let $(y_n)$ $n=1,2,\ldots,$ be a sequence of real numbers. For integers $N\geq 1$ and $k\geq 0$ and a subset $E$ of $[0,1)$, let $A(E,N,k)$ be the number of terms among $\{y_{k+1}\},\{y_{k+2}\},\ldots,\linebreak
\{y_{k+N}\}$ that are lying in~$E$.
The sequence $(y_n)$ $n=1,2,\ldots,$ is said to be \textit{well-distributed mod 1} if for all pairs $a$, $b$ of real numbers with $0\le a<b\le 1$ we have
\[
\lim_{N\to\infty} A([a,b);N,k)/N=b-a\quad\mbox{uniformly in }k=0,1,2,\ldots.
\]
Example. The sequence $(n\theta)$ $n=1,2,\ldots$ with $\theta$ irrational is well-distributed {mod 1}.''
The latter fact would have proved Lemma~\ref{lem:finlemma}, if the interval of the unit circle $[\{ d_r\},\{D_r\})$ was independent of~$r$. Let us clarify this property in the case of the inequality $\{D_r\}>\{ d_r\}$. In what follows we always assume that this inequality is valid (evidently, as in the definition of the well-distribution property, this leads to no loss of generality). Really, if $k$ equals $\lceil l_{ji^*} \rceil-1$ and $N$ does the difference $\lfloor L_{ji^*} \rfloor-\lceil l_{ji^*} \rceil+1$, then we obtain the uniform in $j$ convergence
\begin{equation}
\label{eq:lastlimit}
\lim_{N\to\infty} \frac{A(\,[\{ d\},\{D\});N,\lceil\, l_{ji^*} \rceil-1)}{N}=D-d,
\end{equation}
which is equivalent to the assertion of the Lemma with the fixed of $\{ d_r\}$, $\{D_r\}$.
Note that if with fixed~$j$ equality~(\ref{eq:lastlimit}) is valid for any subinterval in $[0,1)$, then we say that the corresponding sequence is \textit{uniformly distributed modulo~1}. This property follows from the property of the \textit{well-distribution modulo~1}. It is well known that (see \cite{niderraiter}[section 2.1]) for any sequence uniformly distributed modulo~1 the convergence is uniform with respect to all subintervals in $[0,1)$. Consequently, we get the uniform in~$j$ convergence
\[
\lim_{N\to\infty} \frac{A(\,[\{ d_r\},\{D_r\});N,\lceil\, l_{ji^*} \rceil-1)}{N}=\Delta,
\]
where the \textit{constant} $\Delta$ equals $D_r-d_r$. Therefore,
\[
\lim_{j\to\infty} \frac{A(\,[\{ d_r\},\{D_r\});\lfloor L_{ji^*} \rfloor-\lceil l_{ji^*} \rceil+1,\lceil\, l_{ji^*} \rceil-1)}{(L_{ji^*}-l_{ji^*})\,\Delta}=1,
\]
which coincides with the lemma assertion in a general case.
\quad$\square$
\noindent
$\square$
\section{Conclusion}
We have proved that in the monkey model the probability of words in the sorted list has the \textit{exact} power asymptotics, provided that the ratio of logarithms of probabilities of certain letters is irrational.
Note that this condition is not only sufficient, but also necessary. Really, otherwise logarithms of probabilities $a_i=-\ln p_i$, $i=1,\ldots,n,$ allow the repre\-sen\-ta\-tion $a_i=m_i v$, where $m_i$ are natural numbers and $v$ is independent of~$i$. In this case formula~(\ref{eq:rec}) defines a linear recurrent correlation on a grid with the step of $v$. This does not affect the initial constancy of the function~$\tilde Q$ in cells of the grid with the mentioned step with any value of the argument.
It should be noted that using the expression for terms of
linear recurring sequences via the corresponding powers of
roots of the characteristic equation allows clear analysis of rate
of convergence to the power law of the function~$\tilde Q$ (with a step
of~$v$ on the grid) in this degenerate case. It would be more
interesting to conduct such studies for more general case to
which the main theorem of this paper is devoted.
A generalization of results obtained in this paper to the case of the Markov dependence is of even more interest. In this case an analog of the vector $\mathbf{p}^\gamma$ is a substochastic matrix of transition probabilities where the row and column that correspond to the absorbing state are deleted, and all elements of this matrix are raised to a power of~$\gamma$ such that its spectral radius equals 1. Denote this matrix by $\mathbf{P}^\gamma$. In the case considered above all rows of the matrix $\mathbf{P}^\gamma$ coincide with $\mathbf{p}^\gamma$. In a typical case, when the strong power law takes place (see Introduction), the matrix $\mathbf{P}$ is irreducible and the matrix transposed with respect to $\mathbf{P}^\gamma$ has a positive eigenvector that corresponds to the unit eigenvalue. Let us norm this vector so as to make the sum of its components equal 1 and denote the result by~$\mathbf{w}$. In the case of the Markov chain with the transition probability matrix~$\mathbf{P}^\gamma$ this vector defines an ergodic distribution.
If all rows of the considered matrix coincide with $\mathbf{p}^\gamma$, then one can easily see that $\mathbf{w}$ coincides with $\mathbf{p}^\gamma$. It is possible that in the case of the Markov dependence with the irreducible matrix $\mathbf{P}$, an analog of Theorem~\ref{th:main} takes place, where the role of the vector $\mathbf{p}^\gamma$ is played by $\mathbf{w}$. We are going to verify this conjecture in our future research.
|
\section{Introduction}
\paragraph{ } The currently accepted standard model of elementary particles and their interactions is based on Higgs mechanism of spontaneous symmetry breaking of the local electroweak gauge symmetry that leaves the renormalizability of the model intact \cite{1p}. In this mechanism masses of the quarks, the charged leptons, the weak gauge bosons and the neutral scalar Higgs of the model are generated from the asymmetry of the vacuum.
\paragraph{ } A convenient tool to investigate the nature of the vacuum structure of a renormalizable quantum field theory and stability of the theory with spontaneous symmetry breaking at zero and finite temperature is the effective potential, which contains all quantum radiative corrections to the tree classical potential. The formalism of the effective potential as an essential tool for the investigation of the spontaneous symmetry breaking was first considered by Goldstone and et al \cite{2p} and Jona-Lasinio \cite{3p}, (for review of early works see for example ref.\cite{4p}). The physical interpretation of the effective potential as a means to explore the vacuum structure of a quantum field theory has been elaborated by Symanzik \cite{5p}, (for review see for example ref.\cite{6p} and ref.\cite{7p}). The vacuum structure of a quantum field theory involves several quantitative calculable quantities such as symmetry and asymmetry of the vacuum, stability and instability of the vacuum, symmetry breaking patterns, restoration of symmetry, induced asymmetry and vacuum energy.
\paragraph{ } The effective potential at higher order is used to investigate the vacuum structure of spontaneous symmetry breaking through Coleman-Weinberg mechanism \cite{9p} for massless standard model (SM). Starting with a theory massless in the tree-level approximation is an attractive idea where it avoids the problems associated with the negative mass square term.
\paragraph{ } The standard model contains some of free parameters, one of these parameters is the mass of the Higgs boson, as a tree level Higgs scalar coupling $\lambda$ so the mass of the Higgs boson has a priori no constraints from SM theory, and its determination self-consistently poses a challenging problem. In the present work, an attempt is made to determine the Higgs boson mass self-consistently from the renormalization group equations and the effective potential of the model. Investigation the stability of the vacuum leads to determine the Higgs boson mass from the effective potential for each order by using the renormalization condition. This paper is organized as follows, in Sec. II, we apply the renormalization group equation method with use corresponding renormalization group functions to calculate the effective potential up to three-loop order for the massless SM. In Sec. III, we drive a relation that allows to calculate the Higgs boson mass and we discuss the results and investigate the stability of the vacuum structure under radiative quantum corrections up to three-loop order in different mass scale. In addition, we compare our result with some results of other authors employing different approaches. In sec. IV, we present our conclusion.
\section{Calculations of the Effective Potential from the Renormalization Group Equation}
\paragraph{ } In the mass-independent 'tHooft-Weinberg scheme \cite{8p} the full effective potential $V(\varphi)$ satisfies renormalization group equation RGE. Thus, in general the RGE for $V(\varphi)$:
\begin{equation}
DV(\varphi)=0
\label{eq:2_1}
\end{equation}
where
\begin{equation}
D=\mu\frac{\partial}{\partial \mu}+\beta_{\lambda}\frac{\partial}{\partial \beta_{\lambda}}+m^{2}\beta_{m^{2}}\frac{\partial}{\partial m^{2}}+\beta_{g}\frac{\partial}{\partial g}+\gamma \varphi\frac{\partial}{\partial\varphi}
\label{eq:2_2}
\end{equation}
\\Here $\beta_\lambda$,$\beta_{m^{2}}$,$\beta_{g}$ and $\gamma$ are the renormalization group functions for scalar, mass and gauge couplings, respectively and $\gamma$ is the anomalous dimension, which usually calculated perturbatively. Let us now rewrite the RGE as a short form by defining the parameters of the theory, $\lambda_{p}\equiv(\lambda,m^{2},g)$ to get:
\begin{equation}
\left[\mu\frac{\partial}{\partial \mu}+\delta_{p}\beta_{p}\frac{\partial}{\partial\lambda_{p}}+\gamma\varphi\frac{\partial}{\partial\varphi}\right]V(\varphi)=0
\label{eq:2_3}
\end{equation}
\\
Where $\delta_{p}$ equals $m^{2}$ for mass coupling and 1 otherwise. Also $g$ represents all the gauge couplings $g_{1}$,$g_{2}$,and $g_{3}$ of U(1), SU(2) and SU(3) respectively, in addition to the top quark coupling $g_{t}$. The renormalization group function RGF $\beta_{p}$, which called the beta function is given by:
\begin{equation}
\beta_{p}=\frac{d\lambda_{p}}{dt},
\label{eq:2_4}
\end{equation}
with $t=\ln\frac{\mu}{\mu_{0}},\; \mu_{0}=\mu (0)$
\\
In the loop expansion, the effective potential has the expression\cite{9p}:
\begin{equation}
V_{eff}=V^{(0)}+\sum^{\infty}_{n=1}\hbar^{(n)}V^{(n)},
\label{eq:2_5}
\end{equation}
Likewise, the renormalization group functions have the expansion:
\begin{equation}\begin{aligned}
\beta_{p}=\sum^{\infty}_{n=1}\hbar^{(n)}\beta^{(n)}_{p},\\
\gamma=\sum^{\infty}_{n=1}\hbar^{(n)}\gamma^{(n)},
\end{aligned}
\label{eq:2_6}
\end{equation}
where the $\chi^{(n)}$ is the n-loop contribution to $\chi$. Then following the ref.\cite{10p}, we can write the following n-loop of $\hbar^{n}$-order as:
\begin{equation}
\mu\frac{\partial}{\partial\mu}V^{(n)}+D_{n}V^{(0)}+D_{n-1}V^{(1)}+...+D_{1}V^{(n-1)}=0,
\label{eq:2_7}
\end{equation}
where
\begin{equation}
D_{m}=\delta_{p}\beta^{(m)}_{p}\frac{\partial}{\partial\lambda_{p}}+\gamma^{(m)}\varphi\frac{\partial}{\partial\varphi},\;\;m=1,2,...,n
\label{eq:2_8}
\end{equation}
is the differential operator. Hence, one can find the n-loop contribution to the effective potential by using the recursion formula eq.(\ref{eq:2_7}) in terms of RGFs and the tree level potential.
\paragraph{ }The tree level (zero-loop) effective potential for massless SM is:
\begin{equation*}
V^{(0)}=\frac{\lambda}{3!}(\phi^{+}\phi), \text{putting}\;\;\phi^{+}\phi=\frac{\varphi^{2}}{2},
\end{equation*}
one obtains:
\begin{equation}
V^{(0)}=\frac{\lambda}{4!}\varphi^{4}
\label{eq:2_9}
\end{equation}
\\Now, applying eq.(\ref{eq:2_7}) to eq.(\ref{eq:2_9}) using eqs.(\ref{eq:2_6}) and (\ref{eq:2_8}) one gets for one-loop contribution to the effective potential:
\begin{equation}
\mu\frac{\partial}{\partial\mu}V^{(1)}+\frac{A_{1}}{4!}\varphi^{4}=0,
\label{eq:2_10}
\end{equation}
where
\begin{equation}
A_{1}=\beta^{(1)}_{\lambda}+4\lambda\gamma^{(1)}.
\label{eq:2_11}
\end{equation}
Integration of eq.(\ref{eq:2_10}) yields:
\begin{equation}
V^{(1)}=-\frac{A_{1}}{48}\varphi^{4}\ln\frac{\mu^{2}}{\mu^{2}_{0}}+C_{1}(\lambda,g,\varphi,\mu_{0}),
\label{eq:2_12}\end{equation}
\\The two-loop contribution to the effective potential:
\begin{equation}
V^{(2)}=-\frac{A}{48}\varphi^{4}\ln\frac{\mu^{2}}{\mu^{2}_{0}}-\frac{B}{96}\varphi^{4}\ln\frac{\varphi^{2}}{\mu^{2}_{0}}\ln\frac{\mu^{2}}{\mu^{2}_{0}}+\frac{B}{192}\varphi^{4}(\ln\frac{\mu^{2}}{\mu^{2}_{0}})^{2}+C_{2}(\lambda,g,\varphi,\mu_{0}),
\label{eq:2_13}\end{equation}
with
\begin{equation}
A=A_{1}\gamma^{(1)}+A_{2}, \;\;A_{2}=\beta^{(2)}_{\lambda}+4\lambda\gamma^{(2)},
\label{eq:2_14}\end{equation}\\
\begin{equation}
B=[\beta^{(1)}_{\lambda}\frac{\partial\beta^{(1)}_{\lambda}}{\partial\lambda}+8\beta^{(1)}_{\lambda}\gamma^{(1)}+16\lambda(\gamma^{(1)})^{2}+F_{1}],
\label{eq:2_15}\end{equation}\\
\begin{equation}\begin{aligned}
F_{1}=[\beta^{(1)}_{g_{1}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{1}}+4\lambda\beta^{(1)}_{g_{1}}\frac{\partial\gamma^{(1)}}{\partial g_{1}}+\beta^{(1)}_{g_{2}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{2}}+4\lambda\beta^{(1)}_{g_{2}}\frac{\partial\gamma^{(1)}}{\partial g_{2}}\\+\beta^{(1)}_{g_{3}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{3}}+4\lambda\beta^{(1)}_{g_{3}}\frac{\partial\gamma^{(1)}}{\partial g_{3}}+\beta^{(1)}_{g_{t}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{t}}+4\lambda\beta^{(1)}_{g_{t}}\frac{\partial\gamma^{(1)}}{\partial g_{t}}].
\end{aligned}
\label{eq:2_16}\end{equation}\\
For the three-loop contribution to the effective potential, one gets:
\begin{equation}\begin{aligned}
V^{(3)}=\frac{-K_{1}}{48}\varphi^{4}\ln\frac{\mu^{2}}{\mu^{2}_{0}}-\frac{K_{2}}{96}\varphi^{4}(\ln\frac{\mu^{2}}{\mu^{2}_{0}}\ln\frac{\varphi^{2}}{\mu^{2}_{0}})+\frac{K_{2}}{192}\varphi^{4}(\ln\frac{\mu^{2}}{\mu^{2}_{0}})^{2}\\-\frac{1}{192}\frac{K_{3}}{2}[\ln\frac{\mu^{2}}{\mu^{2}_{0}}(\ln\frac{\varphi^{2}}{\mu^{2}_{0}})^{2}-\ln\frac{\varphi^{2}}{\mu^{2}_{0}}(\frac{\mu^{2}}{\mu^{2}_{0}})^{2}]-\frac{K_{3}}{1152}\varphi^{4}(\ln\frac{\mu^{2}}{\mu^{2}_{0}})^{3}+C_{3}(\lambda,g,\varphi,\mu_{0}),
\end{aligned}
\label{eq:2_17}\end{equation}
where
\begin{equation}
K_{1}=[A_{1}\gamma^{(2)}+A\gamma^{(1)}+A_{3}],
\label{eq:2_18}\end{equation}
with
\begin{equation}
A_{3}=\beta^{(3)}_{\lambda}+4\lambda\gamma^{(3)},
\label{eq:2_19}\end{equation}
\\
\begin{equation}\begin{aligned}
K_{2}=[\beta^{(1)}_{\lambda}\frac{\partial A}{\partial\lambda}+\beta^{(2)}_{\lambda}\frac{\partial\beta^{(1)}_{\lambda}}{\partial\lambda}+4\beta^{(2)}_{\lambda}\gamma^{(1)}+4\beta^{(1)}_{\lambda}\gamma^{(2)}\\+\gamma^{(1)}(4A+B)+16\lambda\gamma^{(1)}\gamma^{(2)}+F_{2}+F_{3}],
\end{aligned}\label{eq:2_20}\end{equation}
with
\begin{equation}\begin{aligned}
F_{2}=[\beta^{(2)}_{g_{1}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{1}}+4\lambda\beta^{(2)}_{g_{1}}\frac{\partial\gamma^{(1)}}{\partial g_{1}}+\beta^{(2)}_{g_{2}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{2}}+4\lambda\beta^{(2)}_{g_{2}}\frac{\partial\gamma^{(1)}}{\partial g_{2}}\\+\beta^{(2)}_{g_{3}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{3}}+4\lambda\beta^{(2)}_{g_{3}}\frac{\partial\gamma^{(1)}}{\partial g_{3}}+\beta^{(2)}_{g_{t}}\frac{\partial\beta^{(1)}_{\lambda}}{\partial g_{t}}+4\lambda\beta^{(2)}_{g_{t}}\frac{\partial\gamma^{(1)}}{\partial g_{t}}],
\end{aligned}\label{eq:2_21}\end{equation}\\
\begin{equation}
F_{3}=\beta^{(1)}_{g_{1}}\frac{\partial A}{\partial g_{1}}+\beta^{(1)}_{g_{2}}\frac{\partial A}{\partial g_{2}}+\beta^{(1)}_{g_{3}}\frac{\partial A}{\partial g_{3}}+\beta^{(1)}_{g_{t}}\frac{\partial A}{\partial g_{t}},
\label{eq:2_22}\end{equation}\\
\begin{equation}
K_{3}=[\beta^{(1)}_{\lambda}\frac{\partial B}{\partial \lambda}+4B\gamma^{(1)}+F_{4}],
\label{eq:2_23}\end{equation}
with
\begin{equation}
F_{4}=\beta^{(1)}_{g_{1}}\frac{\partial B}{\partial g_{1}}+\beta^{(1)}_{g_{2}}\frac{\partial B}{\partial g_{2}}+\beta^{(1)}_{g_{3}}\frac{\partial B}{\partial g_{3}}+\beta^{(1)}_{g_{t}}\frac{\partial B}{\partial g_{t}}.
\label{eq:2_24}\end{equation}
\paragraph{ } It is worth noting that integrating renormalization group equation for the effective potential in the RGE method a constant $C$ that depending on $\varphi,m,\lambda,g$ and $\mu_{0}$ as it is clear in eqs.(\ref{eq:2_12}), (\ref{eq:2_13}) and (\ref{eq:2_17}) should be fixed. In general, to determine this constant one needs to specify renormalization conditions on the effective potential. There are several renormalization conditions. For example,
\begin{equation*}
V(\varphi^{2}=\mu^{2})=V^{(0)}(\varphi) \;\;\text{or} \;\;\frac{d^{2}V}{d\varphi^{2}}|_{\varphi=0}=m^{2} \;\;\text{and}\;\; \frac{d^{4}V}{d\varphi^{4}}|_{\varphi=\mu}=\lambda.
\end{equation*}
Here the constants $C_{1},C_{2}$ and $C_{3}$ are determined by using the condition:
\begin{equation}
\frac{d^{4}V}{d\varphi^{4}}|_{\varphi=\mu}=\lambda.
\label{eq:2_25}\end{equation}\\
Therefore, the full effective potential approximated to one-loop, two-loop and three-loop respectively becomes:
\begin{equation}
V_{1-loop\;approx}=V^{(0)}+V^{(1)}=\frac{\lambda}{24}\varphi^{4}+\frac{A_{1}}{48}\varphi^{4}[\ln\frac{\varphi^{2}}{\mu^{2}}-\frac{25}{6}],
\label{eq:2_26}\end{equation}\\
\begin{equation}\begin{aligned}
V_{2-loop\;approx}=V^{(0)}+V^{(1)}+V^{(2)}=\frac{\lambda}{24}\varphi^{4}\\+\frac{(A_{1}+A)}{48}\varphi^{4}[\ln\frac{\varphi^{2}}{\mu^{2}}-\frac{25}{6}]+\frac{B}{192}\varphi^{4}[(\ln\frac{\varphi^{2}}{\mu^{2}})^{2}-\frac{35}{3}],
\end{aligned}\label{eq:2_27}\end{equation}\\
\begin{equation}\begin{aligned}
V_{3-loop\;approx}=V^{(0)}+V^{(1)}+V^{(2)}=\frac{\lambda}{24}\varphi^{4}+\frac{I_{1}}{48}\varphi^{4}[\ln\frac{\varphi^{2}}{\mu^{2}}-\frac{25}{6}]\\+\frac{I_{2}}{192}\varphi^{4}[(\ln\frac{\varphi^{2}}{\mu^{2}})^{2}-\frac{35}{3}]+\frac{I_{3}}{1152}\varphi^{4}[(\ln\frac{\varphi^{2}}{\mu^{2}})^{3}-20],
\end{aligned}\label{eq:2_28}\end{equation}
with
\begin{equation}\begin{aligned}
I_{1}=(A_{1}+A+K_{1}),\;\; I_{2}=(B+K_{2}),\;\; I_{3}=K_{3}.
\end{aligned}\label{eq:2_29}\end{equation}\\
Now by using renormalization group functions in the standard model one can obtain the effective potential of the massless SM for one-loop eq.(\ref{eq:2_26}), two-loop order eq.(\ref{eq:2_27}) and for three-loop order eq.(\ref{eq:2_28}), these RGFs are given in the Appendix A for one, two and three level \cite{11p,12p,13p,14p,15p,16p,17p}.
\section{Vacuum Stability and the Higgs Boson Mass from the Effective Potential of Massless Standard Model}
\paragraph{ } In the massless case the scalar-field self-interacting coupling constant $\lambda$, is determined from the minimum of the potential by using the tadpole condition:
\begin{equation}
\frac{dV}{d\varphi}|_{\varphi=v}=0.
\label{eq:3_1}
\end{equation}\\
By applying eq.(\ref{eq:3_1}) to one-loop eq.(\ref{eq:2_26}), two-loop eq.(\ref{eq:2_27}) and three-loop eq.(\ref{eq:2_28}) order effective potential we get:
\begin{equation}
\frac{dV_{1-loop\;approx}}{d\varphi}|_{\varphi=v_{1}}=0\Rightarrow \lambda-\frac{11}{6}A_{1}+\frac{A_{1}}{2}(\ln\frac{v^{2}_{1}}{\mu^{2}})=0,
\label{eq:3_2}
\end{equation}
\begin{equation*}
\frac{dV_{2-loop\;approx}}{d\varphi}|_{\varphi=v_{2}}=0\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\end{equation*}
\begin{equation}\begin{aligned}
\Rightarrow\lambda-\frac{11}{6}(A_{1}+A)+24(\frac{A_{1}+A}{48}+\frac{B}{192})\ln\frac{v^{2}_{2}}{\mu^{2}}+\frac{24B}{192}[(\ln\frac{v^{2}_{2}}{\mu^{2}})^{2}-\frac{35}{3}]=0,
\end{aligned}\label{eq:3_3}
\end{equation}
\begin{equation*}
\frac{dV_{3-loop\;approx}}{d\varphi}|_{\varphi=v_{3}}=0\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\end{equation*}
\begin{equation}\begin{aligned}
\Rightarrow\lambda-\frac{11}{6}I_{1}+24(\frac{I_{1}}{48}+\frac{I_{2}}{192})\ln\frac{v^{2}_{3}}{\mu^{2}}+\frac{24I_{2}}{192}[(\ln\frac{v^{2}_{3}}{\mu^{2}})^{2}-\frac{35}{3}]+\frac{36I_{3}}{1152}(\ln\frac{v^{2}_{3}}{\mu^{2}})^{2}+\frac{24 I_{3}}{1152}[(\ln\frac{v^{2}_{3}}{\mu^{2}})^{3}-20]=0.
\end{aligned}\label{eq:3_4}
\end{equation}\\
These equations allows us to obtain the scalar Higgs coupling $\lambda$ for each order if the gauge and Yukawa couplings at the renormalization scale are known, where $v_{i}\; (i=1,2,3)$ is the vacuum expectation value. The vacuum expectation values have the expansions up to one-loop, two-loop and three-loop respectively:
\begin{equation*}
v_{1}=v_{0}+\hbar v^{(1)}, v_{0}=0,
\end{equation*}
\begin{equation*}
v_{2}=\hbar v^{(1)}+\hbar^{2} v^{(2)},
\end{equation*}
\begin{equation*}
v_{3}=\hbar v^{(1)}+\hbar^{2} v^{(2)}+\hbar^{3} v^{(3)}.
\end{equation*}\\
In perturbative theory the order n-contribution should be less than $n-1$-order to make the perturbative series convergent. So, $v^{(2)}$ and $v^{(3)}$ contributions are too small compared to the one-loop contribution value, so we consider the approximation $v_{1},v_{2},v_{3}\approx$ experiment value (246.2 GeV).\\\\
Now the Higgs boson mass is calculated from \cite{9p}:
\begin{equation}
m^{2}_{H}=\frac{d^{2}V}{d\varphi^{2}}|_{\varphi=v},
\label{eq:3_5}
\end{equation}
and satisfies the condition of eq.(\ref{eq:3_1}).\\\\
At one-loop order, using eq.(\ref{eq:2_26}), eq.(\ref{eq:3_2}) and eq.(\ref{eq:3_5}):
\begin{equation}\begin{aligned}
m^{2}_{H1-loop}=\frac{d^{2}V_{1-loop\;approx}}{d\varphi^{2}}|_{\varphi=v_{1}}=\frac{A_{1}}{6}v^{2}_{1},
\end{aligned}\label{eq:3_6}
\end{equation}\\
For two-loop order, using eq.(\ref{eq:2_27}), eq.(\ref{eq:3_3}) and eq.(\ref{eq:3_5}):
\begin{equation}\begin{aligned}
m^{2}_{H2-loop}=\frac{d^{2}V_{2-loop\;approx}}{d\varphi^{2}}|_{\varphi=v_{2}}=(\frac{A_{1}+A}{6}+\frac{B}{24}+\frac{B}{12}\ln\frac{v^{2}_{2}}{\mu^{2}})v^{2}_{2},
\end{aligned}\label{eq:3_7}
\end{equation}\\
and finally at three-loop order, using eq.(\ref{eq:2_28}), eq.(\ref{eq:3_4}) and eq.(\ref{eq:3_5}):
\begin{equation}\begin{aligned}
m^{2}_{H3-loop}=\frac{d^{2}V_{3-loop\;approx}}{d\varphi^{2}}|_{\varphi=v_{3}}=(\frac{I_{1}}{6}+\frac{I_{2}}{24}+\frac{I_{2}}{12}\ln\frac{v^{2}_{3}}{\mu^{2}}+\frac{I_{3}}{48}(\ln\frac{v^{2}_{3}}{\mu^{2}}+\ln\frac{v^{2}_{3}}{\mu^{2}})^{2})v^{2}_{3}.
\end{aligned}\label{eq:3_8}
\end{equation}\\
Equations (\ref{eq:3_6}), (\ref{eq:3_7}) and (\ref{eq:3_8}) give the Higgs boson mass in one-loop, two-loop and three-loop approximations, which are not the physical masses but the running masses.
\paragraph{ }In the massless case, we can simplify the condition in eq.(\ref{eq:3_1}) if we recall that $\mu$ is an arbitrary parameter, we are allowed to choose $\mu$ to be location of the minimum of the effective potential as has been done in ref.\cite{9p}. Equation (\ref{eq:3_2}), eq.(\ref{eq:3_3}), and eq.(\ref{eq:3_4}) become:
\begin{equation}
\frac{dV_{1-loop\;approx}}{d\varphi}|_{\varphi=v_{1}}=0\Rightarrow\lambda-\frac{11}{6}A_{1}=0,
\label{eq:3_9}
\end{equation}
\begin{equation}
\frac{dV_{2-loop\;approx}}{d\varphi}|_{\varphi=v_{2}}=0\Rightarrow\lambda-\frac{11}{6}(A_{1}+A)-\frac{35}{24}B=0,
\label{eq:3_10}
\end{equation}
\begin{equation}
\frac{dV_{3-loop\;approx}}{d\varphi}|_{\varphi=v_{3}}=0\Rightarrow\lambda-\frac{11}{6}I_{1}-\frac{35}{24}I_{2}-\frac{5}{12}I_{3}=0,
\label{eq:3_11}
\end{equation}\\
Note that, eq.(\ref{eq:3_9}) can be viewed as a generalization of the condition that relates the scalar coupling to the gauge coupling obtained in ref.\cite{9p} for the case of massless scalar quantum electrodynamics.\\\\
\newpage
The Higgs boson mass for $\mu=v$ is described by:
\begin{equation}
m^{2}_{H1-loop}=\frac{d^{2}V_{1-loop\;approx}}{d\varphi^{2}}|_{\varphi=v_{1}}=\frac{A_{1}}{6}v^{2}_{1}.
\label{eq:3_12}
\end{equation}
\begin{equation}
m^{2}_{H2-loop}=\frac{d^{2}V_{2-loop\;approx}}{d\varphi^{2}}|_{\varphi=v_{2}}=(\frac{A_{1}+A}{6}+\frac{B}{24})\;v^{2}_{2},
\label{eq:3_13}
\end{equation}
\begin{equation}
m^{2}_{H3-loop}=\frac{d^{2}V_{3-loop\;approx}}{d\varphi^{2}}|_{\varphi=v_{3}}=(\frac{A_{1}+A+K_{1}}{6}+\frac{B+K_{2}}{24})\;v^{2}_{2},
\label{eq:3_14}
\end{equation}
In present work, our initially input parameters are taken from ref.\cite{18p} at the neutral weak scale $M_{Z}$. These parameters can be computed at any scale for each loop by using the following perturbative expression:
\begin{equation}
g_{i}(t)=g_{i}(t=0)+\int (\beta^{(1)}_{g_{i}}(g_{i}(t))+\beta^{(2)}_{g_{i}}(g_{i}(t))+...)\;dt,
\label{eq:3_99}
\end{equation}
\begin{equation*}
(g_{i}=g_{1}, g_{2}, g_{3}, g_{t}),\;\; t=\ln \frac{\mu}{\mu_{0}},\; \text{with}\; \mu_{0}=M_{Z}\approx 91.19\;\text{GeV}.
\end{equation*}
For the strong coupling $\alpha_{S}$ and the top quark mass $m_{t}$ we have taken the range as in \cite{18p},\cite{14p}:
\begin{equation*}\begin{aligned}
0.1127\leq\alpha_{S}(M_{Z})\leq 0.1202,\\170\;\text{GeV}\leq m_{t}\leq 176\;\text{GeV}.
\end{aligned}\label{eq:3_122}
\end{equation*}\\
\textbf{\large Rusults and discussion}
\paragraph{ }The computations for the one-loop, two-loop and three-loop order within the $\mu-$range $(M_{Z}\leq \mu \leq 600 \text{GeV})$ are presented in Table.(\ref{tab:3_1}) for $m_{t}=170$ GeV, $\alpha_{S}(M_{Z})=0.1161$. In this table and the other tables we have used the reduced scalar coupling $\alpha_{\lambda}=\frac{\lambda}{4\pi}$, which corresponds to gauge couplings $\alpha_{g}=\frac{g^{2}}{4\pi}$, $g=g_{1}, g_{2}, g_{3}$ for U(1), SU(2) and SU(3), and appears as perturbative expansion parameter in the loop expansion of the effective potential. For the case $\mu=100$ GeV, there is a minimum of the one-loop, two-loop and three-loop effective potential (see Fig.(\ref{fig:3_1})) with too large values for Higgs scalar coupling and mass, so this result may be ruled out since it lies outside the range of validity of the approximation. It turns out that, increasing the value of $\mu$ improves the results. By considering the neutral weak mass $M_{Z}$ as a mass scale, the Higgs boson mass value is $m_{H1-loop}\approx 753$ GeV at one-loop order. This result is consistent with the recent LHC discovery \cite{19p}. However for the value, $\mu \approx 5.68\times 10^{2}$ GeV, we get $m_{H1-loop}\approx 239.2$ GeV, $m_{H2-loop}\approx 125.4$ GeV and $m_{H3-loop}\approx 181.8$ GeV. At $\mu\geq 28\times 10^{2}\approx 11.4 \;\nu$ GeV, the scalar coupling $\lambda$ at two-loop becomes negative and leads to metastable vacuum while the three-loop order is stable even at high scale energy $\approx 10^{19}$ GeV. For the larger $\mu$-range $(3 \times 10^{3} \text{GeV}\leq \mu \leq 20 \times 10^{3} \text{GeV})$ at $m_{t}=170$ GeV, $\alpha_{S}(M_{Z})=0.1161$ spontaneous symmetry breaking for one-loop and three-loop occurs (see Fig.(\ref{fig:3_3}) at approximately the same scalar coupling values as shown in Table.(\ref{tab:3_2}). In addition, we get $m_{H1-loop}\approx 126.14$ GeV at $\mu \approx 6.5 \times 10^{3}$ GeV while for three-loop $m_{H3-loop}\approx 126$ GeV at $\mu \approx 20 \times 10^{3}$ GeV.
\paragraph{ }We present our results in Table.(\ref{tab:3_3}) for the scalar Higgs coupling, obtained from eqs.(\ref{eq:3_9}),(\ref{eq:3_10}) and eq.(\ref{eq:3_11}), and Higgs boson mass in the ranges of $0.1127\leq\alpha_{S}\leq 0.1202$ and $170\;\text{GeV}\leq m_{t}\leq 176\; \text{GeV}$ for each loop order with consideration that the mass scale equals to vacuum expectation value $(\mu=v)$. Thus the Higgs boson mass for each loop is:
\begin{equation}\begin{aligned}
346.6\; \text{GeV}\leq m_{H1-loop}\leq 347.2 \;\text{GeV},\\
241.2\; \text{GeV}\leq m_{H2-loop}\leq 241.5 \;\text{GeV},\\
247.6\; \text{GeV}\leq m_{H3-loop}\leq 247.9 \;\text{GeV}.
\end{aligned}\label{eq:3_144}
\end{equation}\\
At first glance to the numerical results of one-loop order, one can find that Higgs boson mass $m_{H}$ is large for all the range under our study because the values of the scalar Higgs coupling $\lambda$ may be too large to justify the neglect of higher order contributions. However, the two-loop and three-loop results are not appreciably different from each other. The two-loop and three-loop radiative corrections have led to an improvement of the Higgs boson mass and the value of the scalar coupling.
\paragraph{ }It is interesting to note that in massless theory the absence of tree level mass term leads to large value for the scalar coupling to the limit, where the perturbation method becomes invalid, since as known the perturbation parameters should be small in order to become successively smaller at higher-order terms. In order to overcome this problem one can choose a suitable mass scale with the inclusion of higher-order contributions.
\paragraph{ }The recent discovery of Higgs-like particle does not confirm or exclude the possibility of many Higgs particles with masses larger than the one discovered in CERN. Extensions of the standard model to for example supersymmetry (SUSY) models predict the existence of families of Higgs bosons, rather than the one Higgs particle of the standard model. In all of them, there is one neutral Higgs boson with properties similar to those of the standard model Higgs boson in addition to neutral and charged Higgs bosons. For example in the Minimal Supersymmetric Standard Model (MSSM), which the simplest model of the SUSY, the Higgs mechanism yields to two Higgs doublets, leading to the existence of a quintet of scalar particles. In this model, the lightest Higgs boson mass that represent the standard model-like Higgs, is predicted to be $m_{H}<135$ GeV. Conversely, larger values of the standard model-like Higgs mass up to $\approx 250$ GeV can be obtained with the Non-Minimal Supersymmetric Standard Model (NMSSM) that can avoid the problems of the MSSM with larger particle content \cite{22p},\cite{20p}.
\paragraph{ }Recently, late last year (2015), a heavy Higgs-like particle is not exclude from the Large Hadron Collider LHC where ATLAS and CMS at LHC have reported that proton-to-proton collisions had led to create more photon pairs with energies around 750 GeV than was expected. This excess in the diphoton mass distribution ($m_{\gamma\gamma}\approx 750$ GeV), if confirmed, would indicate the presence of unexpected new elementary particles \cite{19p}. A large number of theoretical interpretations have appeared following LHC discovery \cite{21p,29p,19p}. Most of these papers explain the excess through some new boson with mass-750 GeV. For example in ref.\cite{29p}, a heavy Higgs-like boson which is six times heavier than the standard model-Higgs particle is expected to couple with new types of fermions.
\paragraph{ }In addition, there is an agreement between our result eq.(\ref{eq:3_144}) and some pervious theoretical results of other authors employing different methods \cite{23p,24p,25p,26p,27p,28p}.
\paragraph{ }A comparison between the one, two and three-loop effective potential at lowest values $m_{t}=170$ GeV and $\alpha_{S}(M_Z )=0.1127$ and largest values $m_{t}=176$ GeV and $\alpha_{S}(M_Z )=0.1202$ are shown in Figs.(\ref{fig:3_5}) and (\ref{fig:3_8}). The conclusion that can be inferred from these figures is the curves of the two-loop and three-loop effective potential are almost similar and the position of the vacuum for the three-loop order is slightly deeper than the two-loop. The one-loop effective potential is deeper at the vacuum position although it has the same trend. Furthermore, it must be noted that all these figures with different values range of $m_{t}$ and $\alpha_{S}$ are almost identical and show similar trend.
\newpage
\begin{longtable}[c]{| c | c | c | c | c | c | c | c | c | c |}
\caption{Values of the Higgs scalar coupling, $\alpha_{\lambda}=\lambda/4\pi$, and Higgs boson mass for one-loop, two-loop and three-loop order at $m_{t}=170$ GeV and $\alpha_{S}(M_{Z})=0.1161$ for the mass scale-range $100\; \text{GeV}\leq\mu\leq600\; \text{GeV}$.\label{tab:3_1}}\\
\hline
\multicolumn{1}{| c |}{$\mu\text{(GeV)}$}&\multicolumn{2}{| c |}{one-loop}&\multicolumn{2}{|c|}{two-loop}&\multicolumn{2}{| c |}{three-loop}\\
{}&$\alpha_{\lambda}$&{$m_{H}$(GeV)}&{$\alpha_{\lambda}$}&{$m_{H}$(GeV)}&{$\alpha_{\lambda}$}&{$m_{H}$(GeV)}\\ \hline
$M_{Z}$&3.75&753.1&2.14&650.71&2.67&1053.13\\
100&3.39&678.74&1.95&563.87&2.17&731.48\\
150&2.37&473.87&1.50&357.90&1.54&371.45\\
200&1.95&390.59&1.34&281.76&1.36&285.41\\
250&1.72&343.88&1.25&238.77&1.26&245.48\\
300&1.57&313.37&1.19&209.60&1.20&222.62\\
350&1.46&291.53&1.15&187.69&1.16&208.26\\
400&1.38&274.83&1.13&169.95&1.12&198.64\\
450&1.32&261.82&1.11&154.87&1.10&191.82\\
500&1.26&251.06&1.09&141.51&1.07&186.60\\
550&1.22&242.07&1.08&129.42&1.05&182.90\\
600&1.18&234.53&1.07&117.87&1.03&179.82\\
\hline
\end{longtable}
\begin{longtable}[c]{| c | c | c | c | c | c | c | c | c | c |}
\caption{Values of the Higgs scalar coupling, $\alpha_{\lambda}=\lambda/4\pi$, and Higgs boson mass for one-loop and three-loop order at $m_{t}=170$ GeV and $\alpha_{S}(M_{Z})=0.1161$ for the large mass scale-range $3\times 10^{3}\; \text{GeV}\leq\mu\leq 20\times 10^{3}\; \text{GeV}$.\label{tab:3_2}}\\
\hline
\multicolumn{1}{| c |}{$\mu\text{(GeV)}$}&\multicolumn{2}{| c |}{one-loop}&\multicolumn{2}{| c |}{three-loop}\\
{}&$\alpha_{\lambda}$&{$m_{H}$(GeV)}&{$\alpha_{\lambda}$}&{$m_{H}$(GeV)}\\ \hline
$3\times 10^{3}$&0.75&148.42&0.75&154.49\\
$6\times 10^{3}$&0.65&128.11&0.66&144.03\\
$9\times 10^{3}$&0.60&118.64&0.62&138.9\\
$12\times 10^{3}$&0.57&112.67&0.59&134.28\\
$15\times 10^{3}$&0.55&108.43&0.57&131.7\\
$18\times 10^{3}$&0.54&105.31&0.55&129.1\\
$20\times 10^{3}$&0.53&103.43&0.54&126\\
\hline
\end{longtable}
\begin{longtable}[c]{| c | c | c | c | c | c | c | c | c | c | c | c |}
\caption{Values of the Higgs scalar coupling, $\alpha_{\lambda}=\lambda/4\pi$, and Higgs boson mass for one-loop, two-loop and three-loop order with different values of the top quark mass $m_{t}$ and the strong coupling $\alpha_{S}(M_{Z})$ in the mass scale $\mu=v$.\label{tab:3_3}}\\
\hline
\multicolumn{1}{| c |}{$\alpha_{S}(M_{Z})$}&\multicolumn{1}{| c |}{$m_{t}$(GeV)}&\multicolumn{2}{c}{one-loop}&\multicolumn{2}{| c |}{two-loop}&\multicolumn{2}{| c |}{three-loop}\\
{}&{}&{$\alpha_{\lambda}$}&{$m_{H}$(GeV)}&{$\alpha_{\lambda}$}&{$m_{H}$(GeV)}&{$\alpha_{\lambda}$}&{$m_{H}$(GeV)}\\ \hline
0.1127&170&1.737&346.75&1.252&241.38&1.269&247.66\\
{}&172&1.739&346.94&1.254&241.40&1.271&247.84\\
{}&174&1.740&347.00&1.255&241.40&1.273&247.84\\
{}&176&1.741&347.16&1.256&241.24&1.275&247.84\\
\hline
0.1161&170&1.738&346.77&1.256&241.44&1.269&247.72\\
{}&172&1.738&346.81&1.254&241.46&1.271&247.74\\
{}&174&1.740&347.00&1.255&241.34&1.272&247.75\\
{}&176&1.741&347.19&1.256&241.34&1.275&247.92\\
\hline
0.118&170&1.738&346.78&1.253&241.36&1.268&247.59\\
{}&172&1.738&346.82&1.253&241.38&1.271&247.78\\
{}&174&1.740&347.01&1.255&241.34&1.272&247.79\\
{}&176&1.740&347.04&1.256&241.34&1.274&247.80\\
\hline
0.1202&170&1.736&346.63&1.252&241.51&1.268&247.64\\
{}&172&1.738&346.83&1.254&241.37&1.270&247.65\\
{}&174&1.740&347.03&1.254&241.38&1.271&247.66\\
{}&176&1.741&347.06&1.256&241.23&1.274&247.83\\
\hline
\end{longtable}
\begin{figure}
\includegraphics{afig1}
\caption{A comparison between the one-loop (dotted line), two-loop (solid line) and three-loop (dashed line) effective potential for SM at $\mu$ scale =100 GeV.}
\label{fig:3_1}
\end{figure}
\paragraph{}
\begin{figure}
\includegraphics{afig3}
\caption{A comparison between the one-loop (dotted line) and three-loop (dashed line) effective potential for SM at $\mu$ scale $=3\times 10^{3}$ GeV.}
\label{fig:3_3}
\end{figure}
\begin{figure}
\includegraphics{afig5}
\caption{A comparison between the one-loop (dotted line), two-loop (solid line) and three-loop(dashed line) effective potential for SM at $\mu$ scale =$v$.}
\label{fig:3_5}
\end{figure}
\begin{figure}
\includegraphics{afig8}
\caption{A comparison between the one-loop (dotted line), two-loop (solid line) and three-loop(dashed line) effective potential for SM at $\mu$ scale =$v$.}
\label{fig:3_8}
\end{figure}
\newpage
\section{Conclusion}
\paragraph{ } In this paper, we have used the renormalization group method to calculate the effective potential for massless standard model up to three-loop order. We have also investigated the stability of the effective potential. It is shown that while the vacuum structure for two-loop order is metastable, it is stable for three-loop order even at high energy scale. The Higgs boson mass is calculated at each loop order for a range of energy scale and a range of the top quark mass and the strong coupling. The obtained values of the Higgs boson mass are in the experimental range for the scale energy of order $20\times 10^{3}$ GeV.
|
\section{Introduction}
Minimal surfaces and the mean curvature flow in the free boundary setting are
natural extrinsic geometric elliptic and parabolic problems that have appeared
sporadically throughout the literature for some time (see Nitsche \cite{hildebrandt1979minimal,nitsche1985lectures},
Hildebrandt, Dierkes and collaborators \cite{dierkes1992minimal,dierkes2010regularity,dierkes2010global} for historical remarks).
Inspired by work on the closed hypersurfaces by Huisken \cite{huisken1984flow} and on the
Ricci flow by Hamilton \cite{hamilton1982three}, Stahl in 1994 made a fundamental contribution
\cite{thesisstahl}, establishing local and global existence plus blowup
results. Since this time, work has greatly intensified.
We say that a smooth one-parameter family of immersed disks
$F:D^n\times[0,T)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ evolves by the mean curvature flow with
free boundary on a support hypersurface $F_\Sigma:\Sigma\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ if
\begin{align}
\frac{\partial F}{\partial t} = \vec{H} = -H\nu\quad\quad&\text{ on }D^n\times[0,T)
\notag\\
\IP{\nu}{\nu_\Sigma} = 0\quad\quad\quad\quad&\text{ on }\partial D^n\times[0,T)\,,
\label{MCFwFB}\\
F(\partial D^n,t) \subset F_\Sigma(\Sigma)\,,\quad\text{and}&\quad
F(\cdot,0) = F_0(\cdot)\,.
\notag
\end{align}
Local existence follows, as demonstrated by Stahl \cite{thesisstahl}, by
writing the evolving hypersurfaces as graphs
for a short time over their initial data.
Stahl additionally gave continuation criteria: a-priori bounds on the second
fundamental form are sufficient for the global existence of a solution
\cite{stahl1996res,stahl1996convergence}.
In this work he also showed that initially convex data remains convex when the
support hypersurface is umbilic, and that in this situation the flow contracts
to a round hemishperical point (a Type I singularity). A generalisation to
other contact angles of Stahl's continuation criteria was later obtained by
Freire \cite{freire2010mean}.
Buckland studied a setting similar to that of Stahl, and focused on obtaining a
classification of singularities according to topology and type
\cite{buckland2005mcf}.
Koeller has generalised the regularity theory developed by Ecker and Huisken
\cite{ecker2004rtm,ecker1989mce,ecker1991ieh} to the setting of free boundaries
\cite{koeller2007singular}. His main regularity theorem is a criterion under
which the singular set will has measure zero.
The author has studied initially graphical mean curvature flow with free
boundary, obtaining long time existence results and results on the formation of
curvature singularities on the free boundary
\cite{vmwheeler2012rotsym,wheeler2014mean,wheeler2014meanhyperplane}.
A similar angle approach has been employed by Lambert
\cite{lambert2014perpendicular} in his work.
Edelen's work is the first systematic treatment of Type II singularities \cite{edelen2014convexity}.
Convexity estimates play a fundamental role in his work.
Regular solutions of the mean curvature flow with bounded initial area converge
as $t\rightarrow \infty$ to minimal hypersurfaces. This also occurs in the
setting of free boundary, so it is natural to consider the mean curvature flow
as a tool to study minimal surfaces.
Minimal surfaces (and hypersurfaces) are a classical topic in mathematics and as such have received enormous attention in the
literature. A review is well beyond the scope of this paper.
See for example
\cite{osserman2002survey,giusti1984minimal,schoen1970infinite,schoen1979existence,lawson1970complete,choe1990isoperimetric,choe1992sharp,colding2004space,colding2011course,nitsche1965new,andrews2015embedded}
and the references within. The studies are extensive and from many
perspectives: harmonic analysis, geometry, calculus of variations and
isoperimetry, complex analysis, partial differential equations, spectral
theory, and more.
Work in the free boundary setting is also abundant, see for example
\cite{fraser2011first,fraser2012sharp,courant2005dirichlet,jager1970behavior,hildebrandt1979minimal,gruter1981boundary}
and the references therein.
Nevertheless there remain many fundamental open questions, in particular to do
with the classification and uniqueness of minimal surfaces with free boundary.
Uniqueness for surfaces of prescribed mean curvature has been previously
treated by Vogel in \cite{vogel1988uniqueness} under certain conditions.
Minimal surfaces and capillarity surfaces of constant mean curvature in right
solid cylinders and cones have been studied before by Choe--Park, Lopez--Pyo in
\cite{choe2011capillary,lopez2014capillary,lopez2014capillary2} via
geometric and eliptic techniques. The authors have many results in these papers
and others, involving constant mean curvature surfaces with free boundary
that invite flow applications. We hope that we are able to inspire progress
in this direction.
In this paper we apply the mean curvature flow with free boundary to prove a
result in this direction (Theorem \ref{thmuniqueness}).
In particular, we prove uniqueness and non-existence results for minimal
hypersurfaces supported on oscillating or pinching cylinders (embedded double
cones) in Euclidean space.
There are no dimension, topological, or symmetry restrictions on our results.
For example, we prove:
\begin{uthm}
The only bounded smooth immersed minimal hypersurface with free boundary on a
catenoid is the flat disk supported at the origin.
\end{uthm}
\begin{uthm}
There does not exist any bounded smooth immersed minimal hypersurface with free
boundary on a cone.
\end{uthm}
This paper is organised as follows.
In Section 2 we study the mean curvature flow and prove our main result that
classifies the asymptotic behaviour of initially graphical rotationally
symmetric data by properties of the support hypersurface $\Sigma$.
When singularities develop, we additionally present some classification of their type.
We apply this in Section 3 to prove classification results for immersed
minimal hypersurfaces with free boundary.
\section{Mean curvature flow with free boundary supported on an oscillating cylinder}
The behaviour of immersions flowing by the mean curvature flow with free
boundary is largely unknown, with available results in the literature
indicating that a complete picture of asymptotic behaviour irrespective of
initial condition is extremely difficult to obtain \cite{stahl1996res,
koeller2007singular}.
Therefore the relevant question is: under which initial conditions is it
possible to obtain a complete picture of asymptotic behaviour?
Working in the class of graphical hypersurfaces is a viable strategy, so long
as the graph condition can be preserved
\cite{thesisvulcanov,wheeler2014meanhyperplane,wheeler2014mean,lambert2012constant}.
In each of these works, global results were enabled by symmetry of the initial
data and/or of the boundary.
Without such symmetries, recent work indicates that graphicality is not in
general preserved \cite{andrewswheeler} (even in the case where
$F_\Sigma(\Sigma)$ is a standard round sphere).
Let us formally set the support hypersurface
$F_\Sigma:\Sigma\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ to be rotationally symmetric and generated
by the graph of a function $\omega_\Sigma:Oz\rightarrow\ensuremath{\mathbb{R}}$ over the $Oz$ axis.
We term such a support hypersurface an \emph{oscillating cylinder}.
By convention we let $x=(x_1,\ldots,x_n)$ be a point in $\ensuremath{\mathbb{R}}^n\subset\ensuremath{\mathbb{R}}^{n+1}$,
with $n\geq 2$ and denote by $y=|x|$ the length of $x$.
With this convention the profile curve of the support surface lies in a plane generated by $Oy$ and $Oz$ axes.
We write the graph condition on $\omega_\Sigma$ as
\begin{align}
\IP{{\nu}_{\Sigma}(z)}{e_1}\ >\ C_\Sigma \geq 0,\label{Sigma_graph}
\end{align}
where $C_\Sigma$ is a global constant, ${\nu}_{\Sigma}$ the normal to
${\omega}_{\Sigma}$, and $\IP{\cdot}{\cdot}$ is the standard inner
product in ${\ensuremath{\mathbb{R}}}^{n+1}$.
Our convention is that $\nu_\Sigma$ points away from the interior of the
evolving hypersurface.
Let us now describe how a rotationally symmetric graphical mean curvature flow
with free boundary $F:D^n\times[0,T)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ satisfying
\eqref{MCFwFB} can be represented by the evolution of a scalar function (the
graph function).
Let us set $D(t)=(0,r(t))\subset\ensuremath{\mathbb{R}}$.
The Neumann boundary is at $\partial D(t)=r(t)$.
The left-hand endpoint of $D(t)$, the zero, is not a true boundary point.
It arises from the fact that the scalar generates a radially symmetric graph
that is topologically a disk.
The coordinate system degenerates at the origin and so it is artificially
introduced as a boundary point.
This is however a technicality, and no issues arise in dealing with quantities
at this fake boundary point, since by symmetry and smoothness we have that
the radially symmetric graph is horizontal at the origin.
We represent the mean curvature flow of a radially symmetric graph
$F:D^n\times[0,T)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$
by the evolution of its graph function $\omega: D(t)\times
[0,T)\rightarrow \ensuremath{\mathbb{R}}$, that must satisfy the following:
\begin{align}
\frac{\partial \omega}{\partial t}\ \ &= \frac{d^2\omega}{dy^2}\
\frac{1}{1+(\frac{d\omega}{dy})^2}+\frac{d\omega}{dy}\
\frac{n-1}{y}&&~~\text{ on }~~(0,r(t))\times[0,T),
\label{Neumannproblem}\\
\IP{{\nu}_\omega}{{\nu}_{\Sigma}} &= 0 \text{ and
}r(t)={\omega}_{\Sigma}(\omega(r(t),t))&&~~\text{ on }~~
r(t)\times[0,T),\notag\\
\lim_{y\rightarrow0}\frac{1}{y}&\frac{d\omega}{dy}(y)\text{ exists, and }\notag\\
\omega(y,0) &= \omega_0&&~~\text{ on }~~(0,r(0)).\notag
\end{align}
where $\omega_0:(0,r(0))\rightarrow \ensuremath{\mathbb{R}}$ generates the initial graph, $\omega_0
\in C^2((0,r(0)))$, that also satisfies the boundary Neumann boundary condition
$\IP{{\nu}_{\omega_0}}{{\nu}_{\Sigma}}= 0$ at $r(0)$.
Note that in this representation the graph direction for $\omega_\Sigma$ is
perpendicular to the graph direction for $\omega$.
(Contrast with \cite{vmwheeler2012rotsym}.)
The two graphs share the same axis of revolution.
Examples of this include graphs evolving inside a vertical catenoid neck or
inside the hole of a vertical unduloid.
\subsection{Existence}
We prove global existence of solutions to \eqref{Neumannproblem} by obtaining
uniform $C^1$ estimates.
The problem \eqref{Neumannproblem} is a quasilinear second-order PDE on a
time-dependent domain with a Neumann boundary condition.
The change in domain can be calculated (see \eqref{rprime}) and depends only on
$\omega_\Sigma$, $\omega'$, and $\omega''$.
The local unique existence of a solution in this setting is standard and has
been discussed in detail in \cite{thesisvulcanov,vmwheeler2012rotsym}.
We note that the uniqueness of a solution shows that the representation
\eqref{Neumannproblem} of a solution to \eqref{MCFwFB} is preserved.
Our first main result is the following.
\begin{thm}[Long time existence]
Let $\omega_\Sigma$ and ${\omega}_0$ be defined as above. Assume
\eqref{Sigma_graph}, and that
\begin{equation}
\label{nopinch}
\text{there is no point $z^*$ where $\omega_{\Sigma}(z^*)=0$}\,.
\end{equation}
We further assume for negative and positive infinity that either one of
\begin{equation}
\label{EQcondn}
\text{the limit $\lim_{z\rightarrow\pm\infty}\omega_\Sigma(z)$ does not exist}
\end{equation}
or
\begin{equation}
\label{noshrinkingatinfinity}
\text{ there exists an $|\alpha| > 0$ such that }
z\frac{d\omega_\Sigma}{dz}(z) > 0\text{ for all }
\begin{cases}
z>\alpha, \text{if }\alpha > 0,
\\
z<\alpha, \text{if }\alpha < 0,
\end{cases}
\end{equation}
hold.
Then there exists a global smooth solution $\omega: D(t)\times [0,T)\rightarrow
\ensuremath{\mathbb{R}}$ to the
problem \eqref{Neumannproblem} that converges smoothly to
$\omega_\infty:D(\infty)\rightarrow\ensuremath{\mathbb{R}}$.
The function $\omega_\infty$ is smooth and generates a minimal surface.
\label{LTE}
\end{thm}
\begin{rmk}
The class of support hypersurfaces that satisfy \eqref{noshrinkingatinfinity}
above at both positive and negative infinity are those whose derivative is monotone
outside a compact subset with the correct sign.
Examples of such include the catenoid.
The catenoid also satisfies condition \eqref{EQcondn}.
An example that satisfies \eqref{noshrinkingatinfinity} but not \eqref{EQcondn}
is
\[
\omega_\Sigma(z) = 2 - e^{-z^2}\,.
\]
More generally, any $\omega_\Sigma$ whose derivative is monotone increasing
outside a compact set and converges at infinity satisfies
\eqref{noshrinkingatinfinity} but not \eqref{EQcondn}.
Examples that satisfies \eqref{EQcondn} but not \eqref{noshrinkingatinfinity}
include the unduloids.
Examples that satisfy \eqref{EQcondn} for $z\rightarrow\infty$ and
\eqref{noshrinkingatinfinity} for $z\rightarrow-\infty$ can be pasted together
using those mentioned above; for example, mollify the positive part of a
catenoid with the negative part of $z\mapsto 2-e^{-z^2}$.
\end{rmk}
\begin{rmk}
The condition \eqref{nopinch} prevents $\omega_\Sigma$ from pinching on the
axis of rotation.
If this condition is violated, we expect that some solutions to
\eqref{Neumannproblem} develop finite-time singularities.
This case is treated in detailed in subsection 2.3.
At such points we do not require that $\omega_\Sigma$ is smooth; that is, we
allow cones.
\end{rmk}
\begin{rmk}
The condition \eqref{noshrinkingatinfinity} prevents the solution from
shrinking and sliding off to infinity.
This would happen for a solution supported in the $\Sigma$ generated by
$\omega_\Sigma(z) = 1 + e^{-z^2}$.
Clearly for such solutions we can not expect convergence.
\end{rmk}
For the proof, we use standard machinery of parabolic theory (see for example
\cite{lieberman1996second,ladyshenzkaya1968parabolic}) and its variants for
time-dependent domains as discussed in \cite{vmwheeler2012rotsym,thesisvulcanov}.
In these references a maximum principle is proved; we will apply this without
further reference.
The condition $\IP{{\nu}_\omega}{{\nu}_{\Sigma}} = 0$ can be written in a
simpler way if we take into account the fact that we are working with two
graph functions.
The outer normal to $\omega$ is given by
\[
{\nu}_{\omega}=\frac{1}{\sqrt{1+(\frac{d\omega}{dy})^2}}\big(-\frac{d\omega}{dy},1\big)\,.
\]
For the unit normal to $\omega_\Sigma$ we need to rotate and translate the axes.
We find
\[
{\nu}_{\Sigma}=\frac{1}{\sqrt{1+(\frac{d{\omega}_{\Sigma}}{dz})^2}}\big(1,-\frac{d{\omega}_{\Sigma}}{dz}\big)\,.
\]
This transforms the Neumann boundary condition into
\begin{align}
\frac{d \omega}{dy}(r(t),t)=-\frac{d {\omega}_{\Sigma}}{dz}(\omega(r(t),t))~~
\text{ for all } t\in[0,T),\label{Neumanncondition}
\end{align}
and gives us the following uniform boundary gradient estimate for $\omega$.
\begin{lem}[Uniform boundary gradient estimates]
Let $\omega_\Sigma$ and ${\omega}_0$ be defined as above.
Assume \eqref{Sigma_graph}.
Then
\begin{align*}
\bigg|\frac{d \omega}{dy}(r(t),t)\bigg|\ \leq \ \sqrt{\frac{1}{C_\Sigma}-1}
\end{align*}
for all $t\in[0,T)$.
\label{gradientestimates}
\end{lem}
\begin{proof}
The Neumann condition \eqref{Neumanncondition} gives a bound on the
gradient of $\omega$ in terms of the gradient $\omega_\Sigma$,
however the constant is not particularly clear.
To find this constant we once more look at the boundary condition. Due to the
rotational symmetry we see that the unit normal of $\Sigma$ is, on the boundary,
the same vector as the tangent vector to the evolving graphs. Thus
\begin{align*}
\nu_{\Sigma}(\omega(r(t),t))\ =\ \frac{1}{\sqrt{1+(\frac{d \omega}{dy}(r(t),t))^2}}\bigg( 1,\ \frac{d \omega}{dy}(r(t),t)\bigg),
\end{align*}
for all $t\in[0,T)$. Replacing this into the graph condition \eqref{Sigma_graph} we find
\begin{align*}
\IP{\frac{1}{\sqrt{1+(\frac{d \omega}{dy}(r(t),t))^2}}\bigg(1,\ \frac{d \omega}{dy}(r(t),t)\bigg)}{e_1} \ \geq\ C_\Sigma.
\end{align*}
Simplifying we obtain
\begin{align*}
\sqrt{1+\bigg(\frac{d \omega}{dy}(r(t))\bigg)^2}\ \leq\ \frac{1}{C_\Sigma},
\end{align*}
which yields the desired estimate.
\end{proof}
\begin{proof}[Proof of Theorem \ref{LTE}]
As we allow the boundary to possibly oscillate, height bounds are not immediate.
The maximum principle applies to $|\omega|$, yielding that $|\omega|$ is
bounded by the maximum of its boundary and initial values.
The main task is to control the value of $|\omega|$ on the Neumann boundary.
For the Hopf lemma to work in excluding new maxima on the Neumann boundary we
need to have a certain sign on the directional derivative $\frac{d
\omega}{dy}(r(t),t)$. This is not possible since this quantity changes sign with
the gradient of the $\Sigma$.
This is evident from \eqref{Neumanncondition}.
To obtain height bounds we proceed as follows. First suppose that there exist
two points $z_{sup}$ and $z_{inf}$ such that we have
\[
z_{inf}\ <\ \min_{[0,r(0)]} \omega_0<\ \max_{[0,r(0)]} \omega_0\ < z_{sup}
\]
and
\[
\frac{d\omega_{\Sigma}}{dz}(z_{sup})
=\frac{d\omega_{\Sigma}}{dz}(z_{inf})
= 0
\,.
\]
Then the initial graph is contained in
\[
\Omega = \{(x,z)\,:\,z\in(z_{inf},z_{sup}),\quad |x| < \omega_\Sigma(z)\}\,.
\]
The set $\Omega$ is bounded by the support hypersurface and a minimal
disk at each end.
These act as barriers for the flow: by the avoidance principle we find
\[
z_{inf} < \omega < z_{sup}\,.
\]
If such points $z_{inf}$ and $z_{sup}$ do not exist, then there do not exist
flat disks supported on $\Sigma$ disjoint from the initial graph $\omega_0$
that can be used as barriers.
Assume that there is no such disk in
\[
U = U^+ \cup U^-
\]
where
\[
U^+ = \{(x,z)\,:\,z \ge 0,\, z>\max_{[0,r(0)]} \omega_0\text{ or } z<\min_{[0,r(0)]}\omega_0\}
\]
and
\[
U^- = \{(x,z)\,:\,z < 0,\, z>\max_{[0,r(0)]} \omega_0\text{ or } z<\min_{[0,r(0)]}\omega_0\}.
\]
Each of $U^+$ and $U^-$ have at most two components, one finite and bounded by the plane $z=0$ and another unbounded.
Let $z$ be in the unbounded component of $U^+$.
There are two cases.
{\bf Case 1.} Condition \eqref{noshrinkingatinfinity}
is satisfied on an unbounded component $U^{++}$ of $U^+$.
On this component, the derivative $\frac{d\omega_{\Sigma}}{dz}$ has a sign.
As we know that for sufficiently large $z$, the derivative $\frac{d\omega_{\Sigma}}{dz}(z)$ is positive, in this case it
must be positive on all of $U^{++}$.
Now the boundary condition \eqref{Neumanncondition} implies that
\[
\frac{d\omega}{dy}(z)<0\,.
\]
The Hopf lemma implies that $\omega$ may never reach such a region.
Similarly, if condition \eqref{noshrinkingatinfinity} is satisfied on an
unbounded component $U^{--}$ of $U^-$, and $z$ is in the unbounded component of
$U^-$, then on this component $\frac{d\omega_{\Sigma}}{dz}$ has a
sign, and as we know that for sufficiently large $z$ the derivative
$\frac{d\omega_{\Sigma}}{dz}(z)$ is negative, in this case it must be negative
on all of $U^{--}$.
Now the boundary condition \eqref{Neumanncondition} implies that for all such $z$
\[
\frac{d\omega}{dy}(z)>0\,.
\]
The Hopf lemma again implies that $\omega$ may never reach such a region.
{\bf Case 2.} Condition \eqref{EQcondn} is satisfied on an unbounded component
$U^{++}$ of $U^+$.
As no minimal disk exists on this component, the derivative $\frac{d\omega_{\Sigma}}{dz}$ again has a sign.
If the sign is positive, then the Hopf lemma applies as in Case 1 above.
If the sign is negative, then as $z\rightarrow\infty$, the function $\omega_\Sigma$ is uniformly bounded from below (by the no pinching condition \eqref{nopinch}) and decreasing.
Therefore it converges, violating \eqref{EQcondn}.
If condition \eqref{EQcondn} is satisfied on an unbounded component $U^{--}$ of $U^{-}$ then the derivative is negative.
If it were positive, then similarly as above this is in contradiction with \eqref{EQcondn}.
Therefore in either case the evolving surfaces are contained within a compact region of $\ensuremath{\mathbb{R}}^{n+1}$, and so the graph
function $\omega$ is uniformly bounded.
Thus we are left with obtaining gradient estimates for the evolving graphs $\omega:D(t)\times[0,T)\rightarrow\ensuremath{\mathbb{R}}$.
Let us set $M_t := F(D^n,t)$.
Following \cite{ecker1989mce} we consider the quantity $v=\IP{{\nu}_{M_t}}{e_{n+1}}^{-1}$, which is modulo a tangential
diffeomorphism equal to $\sqrt{1+(\frac{d\omega}{dy})^2}$.
The function $v$ satisfies
\[
\Big(\frac{d}{dt} - \Delta_{M_t}\Big) v \leq 0
\]
and this allows us to apply the maximum principle.
Since the problem deals with evolving hypersurfaces with boundary we have that the maximum of the gradient is controlled
by the maximum between the initial values and the boundary values.
In the graphical setting, this translates to the following estimate:
\begin{align*}
\sup_{(0,r(t))} \bigg|\frac{d \omega}{dy}\bigg|
\le \max \bigg\{\sup_{(0,r(0))}\bigg|\frac{d \omega}{dy}\bigg| ,\
\sup_{s\in[0,t]}\bigg|\frac{d \omega_0}{dy}(r(s),s)\bigg| ,\ \ \sup_{s\in[0,t]}\bigg|\frac{d \omega}{dy}(0,s)\bigg|
\bigg\}\,,
\end{align*}
for all $t\in[0,T]$.
We now refine this by considering maxima at the boundary.
At the artificial boundary point ($y=0$) the gradient function vanishes due to rotational symmetry; that is,
\[
\frac{d \omega}{dy}(0,t)=0\,.
\]
Lemma \ref{gradientestimates} gives a uniform estimate for the gradient on the Neumann boundary.
We can therefore conclude that
\begin{align*}
\sup_{(0,r(t))} \bigg|\frac{d \omega}{dy}\bigg|
\le \max \bigg\{ \sup_{(0,r(0))}\bigg|\frac{d \omega_0}{dy}\bigg|,\ \sqrt{\frac{1}{C_\Sigma}-1}\bigg\}\,.
\end{align*}
Having obtained a-priori uniform $C^1$ estimates for $\omega:D(t)\times[0,T)\rightarrow\ensuremath{\mathbb{R}}$, the quasilinear parabolic
operator \eqref{Neumannproblem} may be considered to be linear with bounded coefficients, and for such a problem global
existence is standard.
Convergence to minimal hypersurfaces is guaranteed by bounded initial area:
We calculate
\[
\frac{d}{dt}\int_{D^n}\,d\mu = -\int_{D^n}|H|^2d\mu
\]
which implies
\[
\int_0^\infty\int_{D^n}H^2d\mu\,dt \le \int_{D^n}\,d\mu\bigg|_{t=0} = c\,.
\]
Finally, we have $|D(t)| \ge \inf \omega_\Sigma > 0$ by assumption, so that $\omega$ does not vanish (c.f. Theorem
\ref{thmsingularities}).
Since all derivatives are uniformly bounded, we may apply a compactness theorem to conclude that $M_t\rightarrow
M_\infty$ and that the mean curvature of $M_\infty$ is identically zero. This argument has been used before by many
authors, see for example \cite{huisken1989npm,buckland2005mcf,thesisvulcanov}.
\end{proof}
\begin{rmk}
On the free Neumann boundary, the rotational symmetry of the solution prevents tilt behaviour.
This occurs when the normal to the graph becomes parallel to the vector field of rotation for $\Sigma$.
This behaviour is explained in much greater detail in \cite{thesisvulcanov} and it is present in many situation of free
boundary problems \cite{andrewswheeler}, thus the need to use the rotationally symmetry in constructing the barriers
needed to show the elliptic results.
\end{rmk}
\subsection{Convergence}
After showing that the solution to the problem \eqref{Neumannproblem} exists for all times we are interested in studying
the precise shape that it attains in the limit as $t\rightarrow \infty$, knowing already that it is a minimal
hypersurface.
In fact, the theory of minimal hypersurfaces (note that the boundary of this disk is a circle) implies that the limit is
a flat disk.
However, we may prove this directly without requiring the general theory, and so we contribute a proof here.
We also give some related results of interest.
\begin{thm}[Convergence to flat disks]
Under the hypotheses of Theorem \ref{LTE}, the global smooth solution
$\omega:D(t)\times[0,\infty)\rightarrow\ensuremath{\mathbb{R}}$ to the problem \eqref{Neumannproblem}
satisfyies
\begin{align*}
\displaystyle \lim_{t\rightarrow \infty}\sup_ {[0,r(t)]}\bigg|\frac{d \omega}{dy}(r(t),t)\bigg|=0\,,
\end{align*}
that is, the
solution converges to a flat disk as $t\rightarrow \infty$.
\label{thmconvergence}
\end{thm}
\begin{proof}
First we prove that the gradient on the boundary vanishes.
Let us denote by $u(x)=\omega_\infty(|x|)$, $u:D_\infty\rightarrow\ensuremath{\mathbb{R}}$, $D_\infty=\{x\in\ensuremath{\mathbb{R}}^n\ :\ y=|x|\in
Dom(\omega_\infty) \}$.
The mean curvature of $u$ is
\[
H=-div\bigg(\frac{\text{D} u}{\sqrt{1+|\text{D} u|^2}}\bigg)
\]
where $\text{D}$ and $div$ are the gradient and divergence in $\ensuremath{\mathbb{R}}^n$ respectively.
We can then compute using divergence theorem and denoting by $\nu_{\partial
D_\infty}$ the outer pointing normal to the boundary of the domain $D_\infty$:
\begin{align*}
0\ &=\ -\int_{D_\infty}H\, dx=\int_{D_\infty} div\bigg(\frac{\text{D} u}{\sqrt{1+|\text{D} u|^2}}\bigg) \, dx\\
&\ =\ \int_{\partial D_\infty} \frac{\text{D} u}{\sqrt{1+|\text{D} u|^2}} \cdot \nu_{\partial
D_\infty} \, dSx\\
&\ =\ \int_{\partial D_\infty} \frac{\frac{d \omega_\infty}{dy}}{\sqrt{1+|\frac{d \omega_\infty}{dy}|^2}} \, dx\\
&\ =\ 2\pi r(\infty)\frac{\frac{d \omega_\infty}{dy}}{\sqrt{1+|\frac{d \omega_\infty}{dy}|^2}}(r(\infty))
\end{align*}
where smoothness of the solution at the rotation axis (i.e. $\frac{d\omega_\infty}{dy}(0)=0$) ensures that the second
boundary term vanishes. This implies that $\frac{d\omega_\infty}{dy}(r(\infty))\equiv0$.
Using this we can show that the gradient of $\omega_\Sigma$ vanishes everywhere:
\begin{align*}
0\ &=\ -\int_{D_\infty}Hu\, dx=\int_{D_\infty} div\bigg(\frac{\text{D} u}{\sqrt{1+|\text{D} u|^2}}\bigg) u\, dx\\
&\ =\ -\int_{D_\infty} \frac{|\text{D} u|^2}{\sqrt{1+|\text{D} u|^2}}\,dx+ \int_{\partial D_\infty} \frac{\text{D} u}{\sqrt{1+|\text{D} u|^2}} \cdot \nu_{\partial
D_\infty} u\, dSx\\
&\ =\ -\int_{D_\infty} \frac{|\text{D} u|^2}{\sqrt{1+|\text{D} u|^2}}\,dx
\end{align*}
where we have used $\frac{d\omega_\infty}{dy}(r(\infty))=0$ for the Neumann boundary and also the smoothness of the solution at the
rotation axis, $\frac{d\omega_\infty}{dy}(0)=0$ to make the boundary term vanish. This implies that $\text{D} u\equiv 0$ and
thus $\frac{d\omega_\infty}{dy}\equiv0$, that is, $\omega_\infty$ is a constant.
\end{proof}
The above calculation implies the following result, which is interesting in its own right.
\begin{lem}
Suppose $F:D^n\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ is an embedded minimal disk with boundary on an oscillating cylinder $\Sigma$
with axis of revolution $Oz$.
If
\begin{itemize}
\item $F(\partial D)$ is a circle in a plane orthogonal to $Oz$; and
\item $F$ is graphical over a disk orthogonal to $Oz$ (but not necessarily rotationally symmetric),
\end{itemize}
then $F(D^n)$ is a standard flat disk.
\label{LMrig}
\end{lem}
This implies that on the boundary, the gradient of limiting hypersurface will vanish \emph{independent of the angle
imposed by the flow problem}.
This explains the non-compactness of the flow (and consequent appearance of
translators) in cases where the support hypersurface doesn't allow this to
happen. (See \cite{altschuler1994,leithesis,guan1996mean} for further results on
flows with various contact angles.)
\begin{cor}
Suppose $F:D^n\times[0,T)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ is as in Lemma \ref{LMrig},
except that at the free boundary where the prescribed angle is $\alpha$, that
is,
\[
\IP{\nu_\omega}{\nu_\Sigma} = \cos\alpha\,.
\]
If there exists no point $z\in Oz$ such that
\[
\frac{\frac{d\omega_\Sigma}{dz}}
{\sqrt{1+\left(\frac{d\omega_\Sigma}{dz}\right)^2}}
= -\cos\alpha
\]
then the flow never reaches an equilibrium.
\end{cor}
\begin{figure}
\centering
\begin{center}
\includegraphics[trim=2cm 13cm 0.5cm 2cm,clip=true,width=0.59\textwidth]{twomin.pdf}\\
\captionstyle{\centering}
\caption{Flat discs at extrema of $\Sigma$.}
\label{fig:twomin}
\end{center}
\end{figure}
The graphs generating the contact hypersurface $\Sigma$ are in general oscillating, forming local minima and maxima as
they stretch in both directions of the axis $Oz$. The flow may in general converge to any flat disk supported at a
critical point of $\omega_\Sigma$.
Note that the case of $\Sigma$ being a cylinder has been treated previously in \cite{huisken1989npm}.
Generically, one expects the flow to converge to minimal disks with the smallest possible area.
Such disks are found at local minima of $\omega_\Sigma$; however, it appears difficult to rule out convergence to other
minimal disks for arbitrary data.
In the following we give sufficient conditions on the initial data (that $\Omega$ is contained in a shrinking neck
region, see Definition \ref{DFbellies}) to guarantee convergence to a minimal disk supported on a local minimum of
$\omega_\Sigma$.
\begin{defn}[Bellies and necks, see Figure \ref{fig:twomin}]
\label{DFbellies}
Let $\Sigma$ be an oscillating cylinder.
A \emph{region} of $\Sigma$ is any set ($z_1, z_2 \in [-\infty,\infty]$)
\[
\Theta(z_1,z_2) = \{(x,z)\,:\,|x|<\omega_\Sigma(z),\, z\in(z_1,z_2)\}\,.
\]
A \emph{shrinking neck region} of $\Sigma$ is any region $\Theta(z_1,z_2)$ where for all $z\in(z_1,z_2)$
\[
\frac{d\omega_\Sigma}{dz}(z) = 0\quad\Longrightarrow\quad
\text{$\omega_\Sigma(z)$ is a weak local minimum value for $\omega_\Sigma$}\,.
\]
A \emph{belly region} of $\Sigma$ is any region $\Theta(z_1,z_2)$ where for all $z\in(z_1,z_2)$
\[
\frac{d\omega_\Sigma}{dz}(z) = 0\quad\Longrightarrow\quad
\text{$\omega_\Sigma(z)$ is a weak local maximum value for $\omega_\Sigma$}\,.
\]
\end{defn}
\begin{thm}[Convergence in shrinking necks]
Assume the hypotheses of Theorem \ref{LTE}.
Suppose that the initial data $F_0(D^n)$ is contained in a shrinking neck region $\Theta(z_1,z_2)$.
Then the global solution $F:D^n\times[0,\infty)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ to \eqref{MCFwFB} converges to a flat disk supported
at a local minimum of $\omega_\Sigma$ in $\Theta(z_1,z_2)$.
If there is just one such minimum at $z^*\in(z_1,z_2)$ then the flat disk supported at $\omega_\Sigma(z^*)$ is the
unique limit of all solutions to \eqref{Neumannproblem} with initial data in $\Theta(z_1,z_2)$.
\label{minimumflatdiscs}
\end{thm}
\begin{proof}
Theorems \ref{LTE} and \ref{thmconvergence} yield that the solution exists for all time and converges to a flat disk.
It remains to identify which disks may serve as limits for the flow.
All minimal disks in a shrinking neck region are located at local minima of $\omega_\Sigma$ by definition.
Therefore we will be finished if we can prove that there exists a shrinking neck region $\Theta(w_1,w_2)$ such that for
all $(x,t) \in D^n\times[0,\infty)$,
\begin{equation}
\label{Claim1}
F(D^n,t) \subset \Theta(w_1,w_2)\,.
\end{equation}
We extend the given shrinking neck region in either direction until we reach a critical point of $\omega_\Sigma$.
More precisely, let us take
\[
w_1 = \sup\bigg( \bigg\{z\in(-\infty,z_1]\,:\, \frac{d\omega_\Sigma}{dz}(z) = 0 \bigg\} \cup \big\{ -\infty \big\} \bigg)
\]
and
\[
w_2 = \inf\bigg( \bigg\{z\in[z_2,\infty)\,:\, \frac{d\omega_\Sigma}{dz}(z) = 0 \bigg\} \cup \big\{ \infty \big\} \bigg)
\]
Clearly $\Theta(z_1,z_2)\subset \Theta(w_1,w_2)$, and $\Theta(w_1,w_2)$ is a region of $\Sigma$.
To see that $\Theta(w_1,w_2)$ is a shrinking neck region, let $z_0\in\Theta(w_1,w_2)\setminus\Theta(z_1,z_2)$ be a point
where
\[
\frac{d\omega_\Sigma}{dz}(z_0) = 0\text{ and }
\text{$\omega_\Sigma(z_0)$ is a local maximum value for $\omega_\Sigma$}\,.
\]
If $\omega_\Sigma'(z_1) = 0$ then $w_1 = z_1$; similarly for $z_2$.
If both conditions are satisfied, then the set $\Theta(w_1,w_2)\setminus\Theta(z_1,z_2)$ is empty, and such a $z_0$ can not exist.
Suppose otherwise. Then either $z_0\in(w_1,z_1)$ or $z_0\in(z_2,w_2)$.
Suppose the former.
Since $\omega_\Sigma'(z_0) = 0$, $z_0 \in \{z\in(-\infty,z_1]\,:\, \omega_\Sigma'(z) = 0\}$ and $z_0 > w_1$.
This contradicts the definition of $w_1$.
Similarly, $z_0\in(z_2,w_2)$ contradicts the definition of $w_2$.
Therefore such a $z_0$ can not exist.
We now claim \eqref{Claim1}.
Let us prove this by contradiction.
Suppose there exists a sequence of points in space-time $((x_n,z_n),t_n)\in F(D^n,t_n)\times[0,\infty)$ such that
$z_n \rightarrow z_\infty$ such that $z_\infty \le w_1$ or $z_\infty \ge w_2$.
Let us first bring $z_\infty\le w_1$ to a contradiction.
If $z_\infty = -\infty$, this contradicts the height bound from Theorem
\ref{LTE}.
Therefore the only way for $z_n\rightarrow z_\infty$ such that $z_\infty \le w_1$ is if $w_1$ is finite.
Then by definition of $w_1$ we have $\frac{d\omega_\Sigma}{dz}(w_1) = 0$ and so there exists a
minimal disk supported at $w_1$ that serves as a barrier for the solution.
Therefore we are left with the case where $z_\infty = w_1$.
In this case, we have for sufficiently large $n$
\[
\frac{d\omega_\Sigma}{dz}(z_n) < 0\,.
\]
(The strict sign follows from the definition of $w_1$ and the use of the flat disk as a barrier.)
The boundary condition then yields $\frac{d\omega}{dz}(r(t_n),t_n) > 0$.
Therefore there is no maximum for $\omega(\cdot,t_n)$ at it's Neumann boundary.
However, by assumption, the graphs $\omega(\cdot,t_n)$ are moving downward to the flat disk.
Therefore there must be a new minimum for $\omega(\cdot,t_n)$, or equivalently, a new maximum for
$|\omega(\cdot,t_n)|^2$.
This maximum must be either at the axis of rotation ($y = 0$) or in $(0,r(t_n))$.
The parabolic evolution equation for $\omega$ implies that
\begin{align*}
\sup_{(0,r(t))} \big|\omega \big|^2
\le \max \bigg\{\sup_{(0,r(0))}\big|\omega\big|^2 ,\
\sup_{s\in[0,t]}\big|\omega(r(s),s)\big|^2 ,\ \ \sup_{s\in[0,t]}\big|\omega(0,s)\big|^2 \bigg\}\,.
\end{align*}
We already ruled out new maxima on the Neumann boundary. The Hopf Lemma implies that new maxima are also impossible at
the axis of rotation, since there $\frac{d\omega}{dz} = 0$ by symmetry. The only case remaining is that the new maxima
occur on the interior, which is clearly a contradiction.
Therefore $z_\infty > w_1$. A similar argument shows that $z_\infty < w_2$, and so the claim \eqref{Claim1} is proved.
\end{proof}
\begin{rmk}[$\Sigma$ catenoid]
If $\Sigma$ is a catenoid or it has only one global minimum then $\Theta(-\infty,\infty)$ is a shrinking neck region.
Theorem \ref{minimumflatdiscs} above then yields that solutions converge as $t\rightarrow \infty$ to the unique flat
disk perpendicular to $\Sigma$ at this point (c.f. the analogous result in \cite{vmwheeler2012rotsym}).
\end{rmk}
If the initial data is contained in a maximal finite belly region, then it is trapped in this region, by comparison with
flat disks at either end (c.f. the proof of Theorem \ref{minimumflatdiscs} above).
If the initial data is to one side of the highest (or lowest) flat disk in the belly region, then it is also in a
shrinking neck region, and the previous theorem applies.
If the initial data \emph{intersects any flat disk} in the belly region, then the asymptotic behaviour of the flow
becomes more complicated.
In the following result we give a sufficient conditions that guarantees the flow (even if initially in a belly region
intersecting a flat disk) moves out of the belly region and converges to a flat disk in a shrinking neck region.
\begin{figure}
\centering
\begin{center}
\includegraphics[trim=3cm 13cm 0.5cm 3cm,clip=true,width=0.69\textwidth]{neck.pdf}\\
\captionstyle{\centering}
\caption{Uniqueness of limiting disks for flows with initial data contained in a shrinking neck region.}
\label{fig:neck}
\end{center}
\end{figure}
\begin{thm}[Convergence for initial data in bellies]
Assume the hypotheses of Theorem \ref{LTE}.
Suppose that the initial data $F_0(D^n)$ is contained in a belly region $\Theta(z_1,z_2)$ such that it intersects a flat
disk in $\Theta(z_1,z_2)$ and:
\begin{enumerate}
\item[(a)] $H(\cdot,0)<0$ and $\omega_0(r(0)) > \alpha$ where $\alpha$ is the $z$-coordinate of the highest minimal disk in $\Theta(z_1,z_2)$;
or
\item[(b)] $H(\cdot,0)>0$ and $\omega_0(r(0)) < \beta$ where $\beta$ is the $z$-coordinate of the lowest minimal disk in
$\Theta(z_1,z_2)$.
\end{enumerate}
Then the global solution $F:D^n\times[0,\infty)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ to \eqref{MCFwFB} converges to a flat disk supported
at a local minimum of $\omega_\Sigma$ in a shrinking neck region.
If $\Theta(z_1,z_2) \subset \Theta(z_{min},z_{max})$ where $\Theta(z_{min},z_{max})$ is a maximal belly region and $-\infty <
z_{min} < z_{max} < \infty$ then
the flat disk supported at $\omega_\Sigma(z_{max})$ is the
unique limit of all solutions to \eqref{Neumannproblem} with initial data satisfying (a) and
the flat disk supported at $\omega_\Sigma(z_{min})$ is the
unique limit of all solutions to \eqref{Neumannproblem} with initial data satisfying (b).
\label{minimumflatdiscsH}
\end{thm}
\begin{rmk}
A sign on the mean curvature does not imply that the profile $\omega$ is convex or concave.
Note that if $\alpha = \beta$ then we set $z^* = \alpha = \beta$ (see Figure \ref{fig:belly}).
If the belly region is infinite on either side, then we do not expect solutions to converge.
This is not possible here, as it would contradict the assumptions of Theorem \ref{LTE}.
\end{rmk}
Before we start the proof of the theorem we require a result on preservation of the sign of the mean curvature for mean
curvature flow with free boundary.
This is due to Stahl \cite{thesisstahl}.
\begin{figure}
\centering
\begin{center}
\includegraphics[trim=3cm 14cm 0.5cm 3cm,clip=true,width=0.59\textwidth]{belly.pdf}\\
\captionstyle{\centering}
\caption{Uniqueness of flat disks in belly regions}
\label{fig:belly}
\end{center}
\end{figure}
\begin{prop}[\cite{thesisstahl}]
Let $H(\cdot,0)\geq (\leq)\, 0$ everywhere on $M_0$.
Then $H(\cdot,t)\geq (\leq)\, 0$ for all $t\geq 0$ where $M_t$ a solution
of the mean curvature flow with free boundary.
\end{prop}
For completeness we sketch the proof.
It is based on the use of the maximum principle and the fact that on the boundary the directional derivative of the mean
curvature is equal to the mean curvature multiplied by a component of the second fundamental form of $\Sigma$ at that
point.
The Hopf Lemma then yields a contradiction, for any smooth $\Sigma$, with the appearance of a new zero (maximum or
minimum) on the boundary.
The strong maximum principle yields the strict sign for all strictly positive times.
\begin{proof}[Proof of Theorem \ref{minimumflatdiscsH}]
Theorems \ref{LTE} and \ref{thmconvergence} yield that the solution exists for all time and converges to a flat disk.
Suppose we are in situation (a).
First we translate the $Oy$ axis so that $\alpha = 0$, and $z_1 \le 0 \le z_2$.
After this translation, the definition of belly region implies
\begin{align*}
z\frac{d \omega_\Sigma}{dz}\ \leq\ 0,
\end{align*}
for all $z\in\Theta(z_1,z_2)$.
As in the proof of Theorem \ref{minimumflatdiscs}, consider the maximal belly region $\Theta(z_{min},z_{max}) \supset
\Theta(z_1,z_2)$.
By assumption, neither of $z_1$, $z_2$ may be infinite, and so by definition of $z_{min}$, $z_{max}$, there exist flat
disks on the boundary of $\Theta(z_{min},z_{max})$ supported on $\Sigma$.
We are in the case of negative mean curvature.
To show that the graphs will ascend and converge as $t \rightarrow\infty$ to the flat disk at $z=z_{max}$ we look at the
derivative of the boundary point $r(t)$.
Since
\begin{align*}
r(t)\ =\ \omega_\Sigma(\omega(r(t),t))\,,
\end{align*}
we calculate
\begin{align}
\label{rprime}
r'(t) = \frac{d\omega_\Sigma}{dz}\bigg(\frac{\partial\omega}{\partial t} + \frac{d\omega}{dy} r'(t)\bigg)\,.
\end{align}
Substituting in the boundary condition \eqref{Neumanncondition} and $\frac{\partial\omega}{\partial t}= -Hv$
yields
\begin{align*}
r'(t)\ =\ -\frac{H}{v}\frac{d\omega_\Sigma}{dz},
\end{align*}
where we have once again denoted $v=\sqrt{1+(\frac{d\omega}{dy})^2}=\sqrt{1+(\frac{d\omega_\Sigma}{dz})^2} > 0$ at the
boundary points.
Now for all $z\ge0$ (recall the translation)
\begin{align*}
\frac{d \omega_\Sigma}{dz} \le 0.
\end{align*}
This gives us that $r'(t) < 0$ which means that $r(t)$ is decreasing.
As we are in a belly region above the highest flat disk, this implies that $\omega(r(t),t)$ is monotone increasing.
Given that the graphs exist for all times and converge to a flat disk, the first such encountered by the solution is
the flat disk at $z_{max}$.
Since this disk also serves as a barrier for the solution, the proof is finished.
\end{proof}
\begin{rmk}[Height restrictions on the boundary point]
If $\alpha=\beta=z^*$, the restriction on $\omega_0(r(0)) \ge z^*$ is necessary.
This is because otherwise the mean curvature of $\omega_0$ can not be everywhere negative.
If $\omega_0(r(0)) < z^*$, then $\omega_0$ would have to turn after passing the translated $Oy$ axis so that it reaches
the axis of rotation orthogonally, creating a mean convex region.
To see this note that by \eqref{Neumanncondition}, for all $z < z^*$ we have
$\frac{d\omega_0}{dy}(r(0))\ <\ 0$.
At the rotation axis the gradient is vanishing by smoothness, that is, $\frac{d\omega_0}{dy}(0)=0$.
This implies that there exits a point $y^*\in (0,r(0)]$ such that $\frac{d^2\omega_0}{dy^2}(y^*)< 0$.
Otherwise the gradient would just increase, giving a contradiction.
At this point we calculate the mean curvature:
\begin{align*}
0 > H = -\frac{\frac{d^2\omega_0}{dy^2}}{\sqrt{1+(\frac{d\omega_0}{dy})^2}^3}-\frac{n-1}{y{\sqrt{1+(\frac{d\omega_0}{dy})^2}}} \frac{d\omega_0}{dy}
\end{align*}
and obtain that $\frac{d\omega_0}{dy}(y^*) > 0$.
Since $\frac{d\omega_0}{dy}(r(0)) < 0$, we see that there exists a point $y_2\in (0,r(0)]$ such that
$\frac{d\omega_0}{dy}(y_2)=0$.
Repeating the above by replacing the point at zero with $y_2$, we find a second point $y_3\in (y_2,r(0)]$ with the
property that the gradient vanishes at $y_2$.
Denote by $y_1=0$.
In this way we obtain a sequence of points converging $y_k\rightarrow y_\infty$ as $k\rightarrow \infty$,
such that $\frac{d\omega_0}{dy}(y_k)=0$.
If $y_\infty = r(0)$, then by smoothness of $\omega_0$ we obtain a contradiction with the strict sign by the Neumann condition \eqref{Neumanncondition}.
If $y_\infty \in (0,r(0))$ then by smoothness of $\omega_0$, in a left-neighbourhood of $y_\infty$ we have that
$\omega_0$ is flat. This implies in particular that the first two derivatives of $\omega_0$ vanish there, and so the
mean curvature vanishes also. This is in contradiction with $H < 0$.
\end{rmk}
\begin{rmk}[Initial boundary point at $z^*$]
If the initial data has Neumann boundary tangential to the flat disk at $z^*$, and is disjoint from the flat disk in the
interior, it will immediately move into a shrinking neck region and Theorem \ref{minimumflatdiscs} applies. If it is not
immediately disjoint from the flat disk, then it is either tangential or crosses the flat disk. If tangential, then the
mean curvature is zero at some interior points, and this is a contradiction with the mean curvature having a definite
sign. If it crosses the disk, then the same proof in the above remark applies to show that the mean curvature must
change sign.
\end{rmk}
\begin{figure}
\centering
\begin{center}
\includegraphics[trim=1cm 12cm 1cm 0.5cm,clip=true,width=0.59\textwidth]{singularity.pdf}\\
\captionstyle{\centering}
\caption{Examples of finite-time singularities.}
\label{fig:singularity}
\end{center}
\end{figure}
\subsection{Singularities}
In this section we treat the case when the support hypersurface $\Sigma$ pinches on its axis of rotation; that is, there
exists one or more points $z^*$ such that $\omega_\Sigma(z^*) = 0$.
We do not require that $\Sigma$ is smooth at those points so examples of such support hypersurfaces include cones,
parabolae or hypersurfaces that form cusps at the rotation axis.
\begin{defn}
Let $\omega_\Sigma:Oz\rightarrow[0,\infty)$ be a continuous function.
Assume that $\omega_\Sigma$ is smooth outside finitely many points $P =
\{w_1,\ldots,w_{n_p}\}$, where $\omega_\Sigma(w_i) = 0$; that is,
$\omega_\Sigma\in C^\infty_{loc}(Oz\setminus P)$.
Assume that there exists a compact set $K \supset P$ such that
\begin{equation*}
z\frac{d\omega_\Sigma}{dz}(z) > 0\quad\text{ for all }z\in\ensuremath{\mathbb{R}}\setminus K\,.
\end{equation*}
The function $\omega_\Sigma$ generates a smooth rotationally symmetric disconnected hypersurface
$F_\Sigma:\Sigma\rightarrow\ensuremath{\mathbb{R}}^{n+1}$, where $\Sigma$ is the disjoint union of $n_p+1$ cylinders.
We term the support hypersurface $F_\Sigma$ a \emph{pinching cylinder}.
\label{pinchingcylinder}
\end{defn}
Note that if $n_p=0$ in Definition \ref{pinchingcylinder} we are in one of the cases considered earlier in the paper.
\begin{rmk}
Although we require that $\omega_\Sigma$ be only continuous on $\ensuremath{\mathbb{R}}$, it may pinch and be smooth (or analytic) everywhere
on $\ensuremath{\mathbb{R}}$. For example, this is the case if $\omega_\Sigma$ is a non-negative polynomial in $z$ with zeros; for example,
\[
\omega_\Sigma(z) = (z-2)^2(z+2)^2\,.
\]
\end{rmk}
\begin{thm}[Flow in conical pinching cylinders]
Let $\Sigma$ be a pinching cylinder as in Definition \ref{pinchingcylinder} with $n_p=1$.
Let $w_1 = z^* = 0$.
Assume \eqref{Sigma_graph} (understood as limits from the left and right at points in $P$).
Suppose that for all $z\in\ensuremath{\mathbb{R}}\setminus\{0\}$,
\begin{equation}
\label{conelike}
z\frac{d\omega_\Sigma}{dz}(z) > 0\,.
\end{equation}
Then the maximal time of existence for any solution $\omega:D(t)\times[0,T)\rightarrow\ensuremath{\mathbb{R}}$ to \eqref{Neumannproblem}
satisfies $T < \infty$.
The hypersurfaces $F:D^n\times[0,T)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ generated by $\omega$ contract as $t\rightarrow T$ to the point
$(0,z^*)$.
\label{thmsingularities}
\end{thm}
\begin{proof}
The proof height and gradient estimates goes through exactly as in Theorem \ref{LTE} and Lemma \ref{gradientestimates}.
We do not have global existence however, as in this setting, we do not have a uniform bound on $r(t)$ from below.
We claim that the solution $\omega:D(t)\times[0,T)\rightarrow\ensuremath{\mathbb{R}}$ of \eqref{Neumannproblem} exists smoothly
for all $t\in[0,T)$, $T<\infty$, and $r(t)\rightarrow0$ as $t\rightarrow0$.
To see this, we first show $T<\infty$.
For the sake of contradiction, assume that the graphs exist for all time, that is, $T=\infty$.
Then the solutions converge to a flat disck perpendicular to the contact hypersurface $\Sigma$ as per Theorem
\eqref{thmconvergence}.
However, any flat disk must be supported on $\Sigma$ by a point where the gradient of $\omega_\Sigma$ vanishes.
Such a point (by \eqref{conelike}) does not exist.
Therefore $T< \infty$.
Since the height and gradient bounds provides us with uniform $C^1$ estimates for all time, the only posibility
preventing global existence is that $r(t)\rightarrow0$ as $t\rightarrow T$.
Therefore the solution converges to a point on the axis of rotation.
The solution however must also satisfy the Neumann condition, and so the limit point must be a point where $\Sigma$
pinches off; as there is only one such point where this occurs, we are done.
\end{proof}
\begin{rmk}[Non-rotational initial data]
Any initially bounded mean curvature flow with free boundary, irrespective of
symmetry or topological properties, exists at most for finite time when
supported on a pinching cylinder as in Theorem \ref{thmsingularities}.
This is because so long as the initial immersion is bounded, we may always
construct a rotationally symmetric graphical solution such that the initial
immersion lies between this solution and the pinchoff point $(0,z^*)$.
The flow generated by this pair of initial data remain disjoint by the
comparison principle, and as the rotationally symmetric solution contracts to a
point in finite time, the flow of immersions must either develop a curvature
singularity in finite time or contract to the same point (and possibly remain
regular while doing so).
Similarly, in a shrinking neck region, we may use the rotationally symmetric
graphical solutions as barriers to obtain that any mean curvature flow with
free boundary whose initial data is contained in a shrinking neck either exists
for all time and converges to a flat disk or develops a curvature singularity
in finite time.
\end{rmk}
Our next task is to determine the type of the singularity.
We are able to show that in most cases the singularity is Type I or better
(Type 0: that it is not a curvature singularity at all but a loss of domain).
The cases that allow us to do this are when the gradient of $\omega_\Sigma$ is bounded.
This includes cones and cusps.
We are not yet able to conclude the same for the case of parabolae, that is, when the gradient of $\omega_\Sigma$
is unbounded on $Oz\setminus P$.
\begin{defn}[Singularities]
Let $F:D^n\times[0,T)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ be a mean curvature flow with free boundary supported on a pinching cylinder.
If there exists an $\varepsilon>0$ such that for all $t\in(T-\varepsilon,T)$
\begin{itemize}
\item the second fundamental form is uniformly bounded, that is,
\[
|A|^2(x,t) \le C < \infty\,,
\]
then we say the singularity is Type 0;
\item the second fundamental form is uniformly controlled under parabolic rescaling, that is,
\[
|A|^2(x,t) \le \frac{C}{T-t}\,,
\]
then we say the singularity is Type I;
\item neither of the previous two cases apply, we say the singularity is Type II.
\end{itemize}
\end{defn}
\begin{thm}[Type 1 singularities]
Let $\omega_\Sigma$ and $\omega_0$ be as in Theorem \ref{thmsingularities}.
If there exist two constants $0<C_1<\infty$ and $C_2<\infty$ such that for $z$
sufficiently close to $z^*$ we have:
\begin{itemize}
\item Conical pinchoff
\begin{align*}
C_1 \le \bigg|\frac{d\omega_\Sigma}{dz}(z^*)\bigg| \le C_2,
\end{align*}
then the singularity from Theorem \ref{thmsingularities} is Type I;
\item Polynomial pinchoff
\begin{align*}
C_1|\omega_\Sigma(z)|^{\sigma} \le \bigg|\frac{d\omega_\Sigma}{dz}(z)\bigg| \le C_2|\omega_\Sigma(z)|^{\sigma}\,,
\end{align*}
for $\sigma<1$, then the singularity from Theorem \ref{thmsingularities} is
Type I, and in particular there exist $\hat{C}_1, \hat{C}_2$ such that for $t$
sufficiently close to $T$ we have
\[
\frac{\hat{C}_1}{T-t}
\le
|A|^2(x,t)
\le
\frac{\hat{C}_2}{T-t}
\]
\end{itemize}
\label{thmtypeI}
\end{thm}
Before starting the proof of the theorem we need to compute the norm squared of
the second fundamental form and mean curvature in terms of the profile curve
$\omega$.
\begin{lem}
For a rotationally symmetric hypersurface generated by the rotation of a graph
function $\omega$ about an axis perpendicular to the graph direction, the norm
squared of the second fundamental form and mean curvature are given by the
formulae
\begin{align*}
|A|^2 &= \frac{1}{(1+(\frac{d\omega}{dy})^2)^3} \Big(\frac{d^2\omega}{dy^2}\Big)^2
+ \frac{1}{1+(\frac{d\omega}{dy})^2}\frac{1}{y^2}\Big(\frac{d\omega}{dy}\Big)^2\notag\\
H &= -\frac{1}{\sqrt{1+(\frac{d\omega}{dy})^2}^3} \frac{d^2\omega}{dy^2}\ -\ \frac{1}{\sqrt{1+(\frac{d\omega}{dy})^2}}\frac{1}{y}\frac{d\omega}{dy}.
\end{align*}
\label{secondff}
\end{lem}
\begin{proof}
The proof is a lengthy but straightforward computation using the parametrisation for a rotationally symmetric graph,
that is, $F(x,t)=(x,\omega(|x|,t))$, where $x\in D^n$, and denoting $y=|x|$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thmtypeI}]
Given that the gradient of $\omega_\Sigma$ is uniformly bounded, as before in the proof of Theorem \ref{LTE} and the
proof of Theorem \ref{thmsingularities} we have that the $\omega$ satisfy uniform $C^1$ estimates up to the time of
singularity.
Let us denote this time by $T$. The estimates imply that there exists a constant $C_4<\infty$ depending only on the
initial data such that
\begin{align*}
\bigg|\frac{d^2\omega}{dy^2}\bigg| \leq C_4\,.
\end{align*}
Thus the second fundamental form will explode at worst as quickly as $r(t)\searrow0$, that is, there exists a
constant $C=C(C_2,C_4)<\infty$ such that
\begin{align*}
|A|^2(y,t) \le C\frac{1}{y^2}\bigg(\frac{d\omega}{dy}\bigg)^2(y,t),
\end{align*}
for all $t$ and $y$.
On the rotation boundary, that is at $y=0$, the right hand side is uniformly bounded by symmetry.
(A unique tangent plane exists at the origin.)
Everywhere else the gradient and $1/y$ is bounded by an absolute constant multiplied by it's value at the boundary.
Thus there exists a constant denoted by
abuse of notation $C=C(C,C_2,C_4)<\infty$ such that
\begin{align}
\sup_{[0,r(t)]}|A|^2(t)\ \leq\ C \frac{1}{r(t)^2}\bigg(\frac{d\omega}{dy}\bigg)^2(r(t),t)\,,
\label{estbu1}
\end{align}
for all $t$.
From here we separate the proof into the two cases.
First assume the case of cones, that is there exists a second constant $C_1>0$ such that
\begin{align*}
\bigg|\frac{d\omega_\Sigma}{dz}(z^*)\bigg| \ge C_1.
\end{align*}
From our estimates above, there exists a constant denoted by abuse of notation $C=C(C,C_2,C_4)<\infty$ such that
\begin{align}
\sup_{[0,r(t)]}|A|^2(t)\ \leq\ C \frac{1}{r(t)^2},
\label{first2}
\end{align}
for all $t$.
To compute the rate of blow up for the boundary point $r(t)$, we use the time evolution
for $r(t)$ (computed earlier in the proof of Theorem \ref{minimumflatdiscsH})
\begin{align*}
r'(t)\ =\ -\frac{H}{v}\frac{d\omega_\Sigma}{dz}.
\end{align*}
Substituting for the Neumann boundary condition \eqref{Neumanncondition} and the formula for the mean curvature in Lemma
\ref{secondff}, we obtain
\begin{align*}
r'(t)r(t)\ =\ -\frac{(\frac{d\omega}{dy})^2}{1+(\frac{d\omega}{dy})^2}\ -\ \frac{\frac{d^2\omega}{dy^2}\frac{d\omega}{dy}}{(1+(\frac{d\omega}{dy})^2)^2} r(t).
\end{align*}
Given that the gradient is bounded away from $0$ by $C_1$ (using the Neumann condition and the bound on the gradient of
$\omega_\Sigma$), and also bounded from above by $C_2$, we have that
\begin{align*}
r'(t)r(t)\ \leq\ -\frac{C_1^2}{1+C_2^2}\ -\ \frac{\frac{d^2\omega}{dy^2}\frac{d\omega}{dy}}{(1+(\frac{d\omega}{dy})^2)^2} r(t).
\end{align*}
We know $r(t)\rightarrow 0$ as $t\rightarrow T$ where $T$ is the final time of existence. We also know that the second
derivative is bounded by $C_4$. Thus we can choose $0<t^*<T$, independently of the sign of the second term above in
$r'(t)r(t)$, such that
\begin{align*}
r'(t)r(t)\ \leq\ -C_5.
\end{align*}
for some constant $0<C_5=C_5(C_1,C_2,C_4)$ for all $t^*<t<T$.
Note that $C_5$ is bounded away from $0$ is independent of $t$.
Integrating from $t<T$ to $T$ and using the fact that $r(T)=0$ we find
\begin{align*}
r^2(t)\ \geq\ 2C_5(T-t),
\end{align*}
for all $t\geq t^*$.
Substituting this into \eqref{first2} we obtain the following bound for the second fundamental form
\begin{align*}
\sup_{[0,r(t)]}|A|^2(t)\ \leq\ \frac{C}{2C_5}\frac{1}{T-t}
\end{align*}
for all $t\in (t^*,T)$, that is, the singularity is Type I.
Now consider the case of polynomial pinchoff: for $z$ sufficiently close to $z^*$ we have
\begin{align*}
C_1|\omega_\Sigma(z)|^{\sigma} \le \bigg|\frac{d\omega_\Sigma}{dz}(z)\bigg| \le C_2|\omega_\Sigma(z)|^{\sigma}\,.
\end{align*}
Using $r'(t) = -\frac{H}{v}\omega_\Sigma'(\omega(r(t),t))$ and the above we estimate:
\begin{align*}
r^{1-2\sigma}(t)r'(t)
\le Cr^{1-2\sigma}(t)-\frac{C_1}{1+C_2r^{2\sigma}}
\le -C_1/2
\end{align*}
for $t$ sufficiently close to $T$ (since $\sigma < 1$) and some $C=C(C_1,C_2,C_4)$ , and so
\begin{align*}
-r^{2-2\sigma}(t)
= \int_t^T (r^{2-2\sigma}(t))'\,ds
\le \int_t^T -C_1(1-\sigma)\,ds
= -C_1(1-\sigma)(T-t)
\,.
\end{align*}
This implies
\[
\frac{1}{r^{2-2\sigma}(t)} \le \frac{1}{C_1(2-2\sigma)}\frac{1}{T-t}\,.
\]
Estimating as above (beginning at estimate \eqref{estbu1} earlier) we find
\begin{align*}
\sup_{[0,r(t)]}|A|^2(t)
&\le C\frac{1}{r^2(t)}\bigg(\frac{d\omega}{dy}\bigg)^2(r(t),t)
\\
&= C\frac{1}{r^2(t)}\bigg(\frac{d\omega_\Sigma}{dz}\bigg)^2(\omega(r(t),t))
\\
&\le C\frac{1}{r^{2-2\sigma}(t)}
\\
&\le C(\sigma) \frac{1}{T-t}
\,.
\end{align*}
Therefore the singularity is Type I.
Now as the assumption is two-sided, we find that (for a different constant
$C(\sigma)$) the same estimate above for the second fundamental form holds, but
from below.
Therefore the singularity is no better and no worse than Type I, and the
statement follows.
\end{proof}
\begin{rmk}
Conical pinchoff is a special case of polynomial pinchoff.
For polynomial pinchoff, it isn't possible to satisfy all condition of the theorem for $\sigma\ge1$.
For $\sigma>0$, the pinchoff is \emph{convex} and for $\sigma<0$ the pinchoff is \emph{concave}.
These names come from the following examples:
\[
\omega_\Sigma(z) = z^\alpha
\]
satisfies $\omega_\Sigma'(z) = \alpha \omega_\Sigma^{1-\frac1\alpha}(z)$.
Therefore $\alpha > 1$ corresponds to $\sigma \in (0,1)$ and $\alpha < 1$ corresponds
to $\sigma < 0$.
Clearly all asymptotically polynomial pinchoffs are allowed by the condition $\sigma < 1$.
Concave pinchoff is related to the singularity resulting from mean curvature
flow with free boundary supported in the sphere, studied by Stahl
\cite{stahl1996res}.
\end{rmk}
\begin{thm}[Type 0 singularities]
Let $\omega_\Sigma:Oz\rightarrow\ensuremath{\mathbb{R}}$ be the profile curve of a rotationally
symmetric hypersurface satisfying \eqref{Sigma_graph} and
\[
\lim_{z\rightarrow\infty}\omega_\Sigma(z) = 0\,,\quad
\bigg|\frac{d\omega_\Sigma}{dz}(z)\bigg| \le C|\omega_\Sigma|^{1+\sigma}(z)\,,\quad \sigma>0\,.
\]
Then the maximal time of existence for any solution $\omega:D(t)\times[0,T)\rightarrow\ensuremath{\mathbb{R}}$ to \eqref{Neumannproblem}
satisfies $T = \infty$.
The hypersurfaces $F:D^n\times[0,T)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ generated by $\omega$ satisfy
\[
||A||_\infty^2(t) \rightarrow \alpha_0\quad\text{as $t\rightarrow\infty$}\,,
\]
and so either
\begin{itemize}
\item $F(D^n,t)$ converges smoothly to a flat disk; or
\item Modulo translation, $F(D^n,t)$ converges to a flat point, that is, a singularity of Type 0.
\end{itemize}
\label{thmtype0}
\end{thm}
\begin{proof}
First, a uniform a-priori gradient bound follows by applying Lemma
\ref{gradientestimates}.
Note that the difference
\[
\sup\{|\omega(y_1,t) - \omega(y_2,t)|\,:\,y_1,y_2\in[0,r(t)]\}
\]
is uniformly bounded, since if it weren't, this would contradict the uniform
gradient bound.
Therefore, the translated flow $\hat\omega(y,t) := \omega(y,t) - \omega(0,t)$
has uniformly bounded height, and so, exists for all time.
Note importantly that the domain of $\hat\omega$ is equal to the domain of
$\omega$, that is, $r(t)$ is invariant under translation.
For the original solution, we have $T = \infty$ and global existence, however the
height may become unbounded.
We calculate, as in the proof of Theorem \ref{thmtypeI} above,
\begin{align*}
\sup_{[0,r(t)]}|A|^2(t)
&\le C\frac{1}{r^2(t)}\bigg(\frac{d\omega}{dy}\bigg)^2(r(t),t)
\\
&\le C\frac{1}{r^2(t)}\bigg(\frac{d\omega_\Sigma}{dz}\bigg)^2(\omega(r(t),t))
\\
&\le C\frac{|\omega_\Sigma|^{2+2\sigma}(r(t),t)}{r^2(t)}
\\
&\le Cr^{2\sigma}(t)\,.
\end{align*}
Therefore the claim follows with $\displaystyle \alpha_0 = C\lim_{t\rightarrow\infty}r^{2\sigma}(t)$.
If $r(t)\rightarrow0$ and $|\omega(0,t)|\rightarrow\infty$ then the limit is a
flat point, and if $r(t)\rightarrow r_\infty>0$ then the proof of Theorem
\ref{thmconvergence} applies and the limit is a flat disk.
\end{proof}
\begin{rmk}
Examples of support hypersurfaces with profile curves satsifying the conditions
of Theorem \ref{thmtype0} include exponentials and reciprocal polynomials, such as
\[
\omega_\Sigma(z) = e^{-z}
\]
and a (monotone) mollification of
\[
\omega_\Sigma(z) =
\begin{cases}
\frac1z\,,\quad\text{for }z>1
\\
-z+2\,,\quad\text{for }z\le1\,.
\end{cases}
\]
\end{rmk}
\section{Uniqueness results for minimal hypersurfaces with free boundary}
In this section we apply the parabolic results proved earlier to the uniqueness
problem for minimal hypersurfaces with free boundary.
We emphasize that the results in this section hold for immersed minimal
hypersurfaces with free boundary, that is, without any restrictions on topology, symmetry,
or graphicality.
In this section we assume $\Sigma$ to be a pinching oscillating cylinder, as in Section 2.3.
Examples of this include catenoids, unduloids, cones, parabolae, and so on.
Generically, an oscillating cylinder decomposes into belly regions and
shrinking neck regions (if maximal, these have non-trivial overlap).
In belly regions, there may exist flat minimal disks that are not rotationally
symmetric with respect to the $Oz$ axis; for example, if part of the belly
region is spherical, then there exist infinitely many such tilted flat disks.
Clearly these disks serve as barriers for the mean curvature flow with free boundary.
Unfortunately, there does not exist a mean curvature flow with free boundary
that is asymptotic (in positive time) to such slanted disks.
For the case of a shrinking neck region however, solutions are asymptotic to
flat disks, and these disks have $Oz$ as their axis of rotation.
Our result is the following:
\begin{thm}[Uniqueness in shrinking necks]
Let $F:M^n\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ be an immersed bounded smooth minimal
hypersurface with free boundary on a pinching cylinder
$F_\Sigma:\Sigma\rightarrow\ensuremath{\mathbb{R}}^{n+1}$.
If $F(M)\subset\Theta(z_1,z_2)$ where $\Theta(z_1,z_2)$ is a shrinking neck
region, then $F(M)$ is a standard flat disk.
\label{thmuniqueness}
\end{thm}
\begin{proof}
We squeeze the minimal hypersurface between two rotationally symmetric graphical
solutions to the mean curvature flow with free boundary.
First, consider the maximal shrinking neck region $\Theta(z_{min},z_{max}) \supset \Theta(z_1,z_2)$.
Suppose that $\Theta(z_{min},z_{max})$ is bounded.
Then by maximality there exists flat minimal disks supported on $F_\Sigma$ at $z_{min}$ and $z_{max}$.
As $F(M)\subset\Theta(z_1,z_2)\subset\Theta(z_{min},z_{max})$, there exist
graphical rotationally symmetric smooth hypersurfaces
$f^1,f^2:D^n\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ supported on $F_\Sigma$ and disjoint from the
minimal disks at $z_{min}$, $z_{max}$, and $F(M)$ such that $f^1 < F(M) <
f^2$.\footnote{We say a hypersurface $M$ is less than a hypersurface $N$ in
$\Theta(a_1,a_2)$ if along each vertical line from $\{z=a_1\}$ to $\{z=a_2\}$,
the intersection point with $M$ is lower than the intersection point with $N$.}
Now take $f^1$ and $f^2$ as initial data for the mean curvature flow with free boundary,
generating flows $F^1,F^2:D^n\times[0,\infty)\rightarrow\ensuremath{\mathbb{R}}^{n+1}$.
By the results of Section 2, each of these flows converge to minimal disks.
By the comparison principle (see for example \cite{thesisstahl,thesisvulcanov})
we have that the hypersurfaces $F^1(D^n,t)$, $F^2(D^n,t)$ are disjoint from
each other, as well as disjoint from $F(M)$.
This is because if they were to intersect at any point, it would be a point of
tangency, and then, as $F(M)$ is minimal and $F^i(D^n,t)$ is not minimal for
all $t\in[0,\infty)$, this would be a contradiction. (This is the only part of
the argument where we require any smoothness of the minimal immersion $F$.)
Now the region these flows foliate is
\[
\Omega = \bigcup \Big\{ F^i(D^n,t)\,:\,t\in[0,\infty)\Big\}\,.
\]
The minimal hypersurface $F(M)$ must be disjoint from this region, and it may
not lie above $f^2$ or below $f^1$ by construction.
Therefore, as $F(M)$ lies in a shrinking neck region $\Theta(z_1,z_2)$, it is supported in a purely cylindrical portion of $\Theta(z_1,z_2)$.
We extend our foliation through this cylindrical region by translation, that is, we take one further flow $g:D^n\times[0,1]\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ where
\[
\partial_tg = \nu\,,\quad
g(D^n,0) = \lim_{t\rightarrow\infty} F^1(D^n,t)\,,\quad
g(D^n,1) = \lim_{t\rightarrow\infty} F^2(D^n,t)\,.
\]
This flow completes the foliation of the region containing $F(M)$.
The flow $g$ is a flow of minimal hypersurfaces.
Therefore, at a first point and time $t^*\in[0,1]$ of tangency, we must have $F(M) = g(D^n,t^*)$, that is, $F(M)$ is a flat disk.
Suppose now that $\Theta(z_{min},z_{max})$ is unbounded on one or both sides; say $z_1 = -\infty$.
Now the boundedness hypothesis on $F(M)$ implies that there exists a
rotationally symmetric graphical smooth hypersurface
$f^1:D^n\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ supported on $F_\Sigma$ and disjoint from $F(M)$
such that $f^1 < F(M)$.
We can use this as initial data and proceed using the argument above.
Similarly, if $z_2 = \infty$, boundedness of $F(M)$ implies that there exists a
rotationally symmetric graphical smooth hypersurface
$f^2:D^n\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ supported on $F_\Sigma$ and disjoint from $F(M)$
such that $F(M) < f^2$.
Again, we can use this $f^2$ in the argument above.
In all cases $F(M)$ is a flat disk, and so we are finished.
\end{proof}
This theorem implies in particular:
\begin{cor}[Non-existence of immersed minimal free boundary hypersurface with
topological type other than that of a disk]
There exists no smooth bounded immersed minimal $n$-dimensional hypersurface
supported on $F_\Sigma$ in any topology other than that of the flat disk.
\label{thmnonexistence1}
\end{cor}
\begin{cor}[The catenoid case]
The only bounded smooth immersed minimal hypersurface with free boundary on a
catenoid is the flat disk supported at the origin.
\end{cor}
Similar results are provable using the same method as above.
We have not attempted to give an exhaustive list.
When there is no minimal disk supported on $F_\Sigma$, then one may guess that
there is no minimal hypersurface supported on $F_\Sigma$.
One situation where this holds is the following:
\begin{prop}[Non-existence of minimal hypersurfaces in cones, parabolae]
Let $F_\Sigma$ be as in Theorem \eqref{thmsingularities}.
There does not exist an immersed bounded smooth minimal hypersurface
$F:M^n\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ supported on $F_\Sigma$.
\label{thmnonexistence3}
\end{prop}
\begin{proof}
As the proof is similar to that of Theorem 3.1, we only give a brief outline.
In Theorem \ref{thmsingularities}, there exists precisely one point of pinching
for $F_\Sigma$.
Without loss of generality we can assume that $F_\Sigma$ lies either above or
below this point of pinching.
In either case, we can construct a rotationally symmetric graphical smooth
hypersurface $f:D^n\rightarrow\ensuremath{\mathbb{R}}^{n+1}$ supported on $F_\Sigma$ such that $F(M)$ lies between $f$ and the pinching point of $\Sigma$.
We use $f$ to generate a mean curvature flow with free boundary.
This yields a foliation of the region between $f$ and the pinching point, which
must by smoothness have a first point and time of tangency with $F$.
However $F$ is minimal and the flow generated by $f$ is not; this yields a
contradiction.
\end{proof}
\section*{acknowledgements}
The author is supported by Australian Research Council Discovery grant DP150100375 at
the University of Wollongong. The author is grateful to the Korea Institute
for Advanced Study and Hojoo Lee for his hospitality and interesting discussions related to this work.
\bibliographystyle{plain}
|
\section{Introduction}
Flavour physiscs plays a fundamental role in the search for New Physics (NP) in particle accelerators. The existence
of new particles can be tested indirectly through the effects they might induce on processes at low-energies. Among
flavour physics processes,
$\Delta F=2$ transitions provide very strong constrains on NP. The most general $\Delta F=2$ weak effective Hamiltonian
can be constructed in terms of the following complete set of parity even (PE) and parity odd (PO) 4-quark operators,
\begin{align}
\textrm{PE}:&\qquad Q_{k}^{\pm} \in \left\{ Q_{VV+AA}^{\pm}, Q_{VV-AA}^{\pm},
Q_{SS-PP}^{\pm}, Q_{SS+PP}^{\pm},
2Q_{T\tilde T}^{\pm}\right\},\nonumber\\
\textrm{PO}:&\qquad \mathcal{Q}_{k}^{\pm} \in \left\{ \mathcal{Q}_{VA+AV}^{\pm},\mathcal{Q}_{VA-AV}^{\pm},
-\mathcal{Q}_{SP-PS}^{\pm}, \mathcal{Q}_{SP+PS}^{\pm},
2\mathcal{Q}_{T\tilde T}^{\pm}\right\}.
\label{eq:pe_po_ops}
\end{align}
Here, a 4-fermion operator with a particular Dirac structure and with four generic
flavours of quarks is given by
\begin{equation}
O_{XY}^{\pm} = {1\over2}\left[
(\overline\psi_{1}\Gamma_{X}\psi_{2})(\overline\psi_{3}\Gamma_{Y}\psi_{4})
\pm (2\leftrightarrow4)\right].
\end{equation}
From the operators in Eq.(\ref{eq:pe_po_ops}), only $Q_{1}$ and $\mathcal{Q}_{1}$ correspond to SM processes. All
the others appear only in beyond SM processes.
Regularizations which break explicitly chiral symmetry generally induce complicated renormalization patterns
for composite operators since they allow for
mixing among operators of different naive chirality. Indeed, when considering Wilson-fermions,
all PE operators in Eq.(\ref{eq:pe_po_ops}) mix under renormalization.
On the other hand, the PO operators renormalize
as in chirally preserving regularizations, namely \cite{M4}
\begin{equation}
[\mathcal{Q}_{1}]_{R}=\mathcal{Z}_{11}\mathcal{Q}_{1},\quad
\left[
{\begin{array}{c}
\mathcal{Q}_{2} \\
\mathcal{Q}_{3} \
\end{array} }
\right]_{R}=
\left[
{\begin{array}{cc}
\mathcal{Z}_{22} & \mathcal{Z}_{23} \\
\mathcal{Z}_{32} & \mathcal{Z}_{33}\
\end{array} }
\right]
\left[
{\begin{array}{c}
\mathcal{Q}_{2} \\
\mathcal{Q}_{3} \
\end{array} }
\right],\quad
\left[
{\begin{array}{c}
\mathcal{Q}_{4} \\
\mathcal{Q}_{5} \
\end{array} }
\right]_{R}=
\left[
{\begin{array}{cc}
\mathcal{Z}_{44} & \mathcal{Z}_{45} \\
\mathcal{Z}_{54} & \mathcal{Z}_{55}\
\end{array} }
\right]
\left[
{\begin{array}{c}
\mathcal{Q}_{4} \\
\mathcal{Q}_{5} \
\end{array} }
\right].
\label{eq:po_ren}
\end{equation}
In practice, one can avoid the renormalization of PE operators following the strategies in \cite{M5} and \cite{M6}, for which
only renormalized matrix elements of PO operators are needed.
In this work we thus focus on the renormalization of PO operators.
We first construct a suitable set of 3-point functions in
the $\chi$SF in order to define the renormalization conditions. Thanks to the mechanism of automatic $O(a)$ improvement all renormalization factors considered will be affected only by $O(a^{2})$ lattice artefacts,
even without $O(a)$ improving the operators or the action in the bulk. We then expand the renormalization conditions
obtained to 1-loop order in perturbation theory, and perform an exploratory study that sets the basis for future non-perturbative studies.
\section{A few words on the $\chi$SF}
Chirally rotated boundary conditions for a flavour doublet of fermionic fields $\psi$ and $\overline{\psi}$ take the form \cite{XSF}
\begin{equation}
\left.\widetilde{Q}_{+}\psi(x)\right|_{x_{0}=0} = \left.\widetilde{Q}_{-}\psi(x)\right|_{x_{0}=T} =0,
\qquad \left.\overline{\psi}(x)\widetilde{Q}_{+}\right|_{x_{0}=0} = \left.\overline{\psi}(x)\widetilde{Q}_{-}\right|_{x_{0}=T}=0,
\label{eq:boundary_conditions}
\end{equation}
with the projectors $\widetilde{Q}_{\pm}=\frac{1}{2}(1\pm i\gamma_{0}\gamma_{5}\tau^{3})$ and where $\tau^{i}$ are
the Pauli matrices.
These boundary conditions are related to the standard SF boundary
conditions through the non-anomalous chiral rotation
\begin{equation}
\psi\rightarrow R(\alpha)\psi,\quad \overline{\psi}\rightarrow\overline{\psi}R(\alpha),
\quad R(\alpha)=\exp(i\alpha\gamma_{5}\tau^{3}/2),
\quad \alpha = \pi/2.
\label{eq:rotation}
\end{equation}
A dictionary relating correlation functions in both setups can be established through the mapping
\begin{equation}
\langle\mathcal{O}[\psi,\overline{\psi}]\rangle_{\chi\textrm{\small SF}}=
\langle\mathcal{O}[R(-\pi/2)\psi,\overline{\psi}R(-\pi/2)]\rangle_{\textrm{\small SF}}.
\label{eq:corr_rel}
\end{equation}
The boundary conditions in Eq.(\ref{eq:boundary_conditions}) are invariant under the rotated version of
parity\footnote{$P_{5}:\psi(x)\rightarrow i\gamma_{0}\gamma_{5}\tau^{3}\psi(\widetilde{x}),\quad
P_{5}:\overline{\psi}(x)\rightarrow -\overline{\psi}(\widetilde{x})i\gamma_{0}\gamma_{5}\tau^{3},\quad
\widetilde{x}=(x_{0},-{\bf x})$}
$P_{5}$,
i.e. $[\widetilde{Q}_{\pm},\gamma_{0}\gamma_{5}\tau^{3}]=0$.
The $P_{5}$ transformation can thus be used to distinguish between
$P_{5}$-even and $P_{5}$-odd correlation functions and invoke the mechanism
of automatic $O(a)$ improvement: on the lattice, all bulk $O(a)$ effects are absent from $P_{5}$-even correlation functions, while these are contained in the $P_{5}$-odd observables which are thus pure lattice artefacts.
For automatic $O(a)$ improvement to be at work one needs to set the quark masses to their critical value and tune the
coefficient of a dimension 3 boundary counterterm, in order to obtain the correct symmetries of
the $\chi$SF in the continuum limit.
Note that additional $O(a)$ effects originating from the boundaries are present in any correlation function. These can be eliminated introducing
a couple of $O(a)$ counterterms at the space-time boundaries.
In the following, we consider the lattice set-up described in \cite{PV2014,DB2014}, to where we refer for details on
the action and on the renormalization and improvement conditions for determining the critical mass and boundary counterterms.
\section{Correlation functions of 4-fermion operators in the $\chi$SF}
In the following we define a set of correlation functions for PO 4-fermion operators in the $\chi$SF. These are
obtained by rotating correlation functions of PE operators in the standard SF through Eq.(\ref{eq:corr_rel}).
We start defining two different types of 3-point correlation functions in the standard SF, with an arbitrary
choice of flavours,
\begin{equation}
F_{i}(x_{0})=\langle O_{5}^{'21}Q_{i}^{1234}(x_{0})O_{5}^{43} \rangle,\qquad
K_{i}(x_{0})={1\over3}\sum_{k=1}^{3}\langle O_{k}^{'21}Q_{i}^{1234}(x_{0})O_{k}^{43} \rangle,
\label{eq:SF_4f_corrs}
\end{equation}
where $i\in[1,5]$ labels the basis of PE operators in Eq.(\ref{eq:pe_po_ops}) and where
$O_{5}$, $O'_{5}$, $O_{k}$ and $O'_{k}$ are standard SF boundary operators.
In the above equations we choose the flavour combination $f_{1}=f_{2}=f_{4}=u$ and $f_{3}=d$.
After applying the chiral rotation Eq.(\ref{eq:rotation}) one obtains the mapping
\begin{align}
Q_{1}^{\pm}\rightarrow -i \mathcal{Q}_{1}^{\pm,uudu},\textrm{ }
Q_{2}^{\pm}\rightarrow -i \mathcal{Q}_{2}^{\mp,uudu},\textrm{ }
Q_{3}^{\pm}\rightarrow -i \mathcal{Q}_{3}^{\mp,uudu},\textrm{ }
Q_{4}^{\pm}\rightarrow -i \mathcal{Q}_{4}^{\pm,uudu},\textrm{ }
Q_{5}^{\pm}\rightarrow -i \mathcal{Q}_{5}^{\pm,uudu}.
\end{align}
Hence, by applying the rotation Eq.(\ref{eq:rotation}) to Eq.(\ref{eq:SF_4f_corrs}) we obtain correlation functions of
PO operators in the $\chi$SF,
\begin{align}
F_{i}(x_{0})\rightarrow G_{i}(x_{0})=\langle \mathcal{O}_{5}^{'uu}\mathcal{Q}_{i}^{uudu}(x_{0})\mathcal{O}_{5}^{ud} \rangle,\quad
K_{i}(x_{0})\rightarrow L_{i}(x_{0})={1\over3}\sum_{k=1}^{3}\langle \mathcal{O}_{k}^{'uu}\mathcal{Q}_{i}^{uudu}(x_{0})\mathcal{O}_{k}^{ud} \rangle.
\label{eq:XSF_4f_corrs}
\end{align}
Here $\mathcal{O}_{5}^{12}$, $\mathcal{O}_{5}^{'12}$, $\mathcal{O}_{k}^{12}$ and $\mathcal{O}_{k}^{'12}$
are the boundary fermion bilinears in the $\chi$SF \cite{DB2016}.
We also consider the boundary to boundary correlation functions $g_{1}^{12}$ and $l_{1}^{12}$, and the
boundary to bulk correlation functions $g_{\tilde{V}}^{12}(x_{0})$ and $l_{\tilde{V}}^{12}(x_{0})$,
where $\tilde{V}_{\mu}^{12}$ is the conserved vector current. These are renormalized\footnote{
Note that since the vector current $\tilde{V}_{\mu}^{12}$ is exactly conserved, there is no renormalisation factor for the fermion bilinear in $[g_{\tilde V}^{12}]_{R}$ and $[l_{\tilde V}^{12}]_{R}$.} by multiplying them
with the proper power of renormalization factors of the boundary fields $Z_{\zeta}$, i.e.
\begin{align}
[g_{1}^{12}]_{R}=Z_{\zeta}^{4}g_{1}^{12},\quad
[l_{\tilde V}^{12}]_{R}=Z_{\zeta}^{2}l_{\tilde V}^{12},\quad
[g_{\tilde V}^{12}]_{R}=Z_{\zeta}^{2}g_{\tilde V}^{12},\quad
[l_{1}^{12}]_{R}=Z_{\zeta}^{4}l_{1}^{12}.
\label{eq:ren_g}
\end{align}
Due to the presence of the boundary fields in Eq.(\ref{eq:XSF_4f_corrs}), the renormalized $G_{i}$ and $L_{i}$ also need the renormalization factor $Z_{\zeta}$. This can be eliminated by normalising $G_{i}$ and $L_{i}$ with a suitable combination of boundary to boundary or boundary to bulk correlation functions. To this end, we define the ratios
\begin{equation}
\mathcal{G}_{i} = G_{i}/\mathcal{N}(\alpha,\beta,\gamma),\qquad
\mathcal{L}_{i} = L_{i}/\mathcal{N}(\alpha,\beta,\gamma).
\label{eq:G_L_norm}
\end{equation}
The normalization $\mathcal{N}(\alpha,\beta,\gamma)$ is given by
\begin{equation}
\mathcal{N}(\alpha,\beta,\gamma) =
\left\{ {g_{1}^{ud}\over g_{1}^{ud(0)}}\right\}^{\alpha}
\left\{ {l_{1}^{ud}\over l_{1}^{ud(0)}}\right\}^{\beta}
\left\{ {g_{\tilde V}^{ud}\over g_{\tilde V}^{ud(0)}}\right\}^{2\gamma}
\left\{ {l_{\tilde V}^{uu}\over l_{\tilde V}^{uu(0)}}\right\}^{2(1-\alpha-\beta-\gamma)},
\label{eq:normalization}
\end{equation}
where $\alpha$, $\beta$ and $\gamma$ are real coefficients satisfying the condition $\alpha+\beta+\gamma \le 1$,
and where the superscript $(0)$ denotes the tree-level correlation functions.
The exponents in the different
terms of Eq.(\ref{eq:normalization}) are chosen such that the correct power of the renormalization
factor $Z_{\zeta}$ is eliminated in Eq.(\ref{eq:G_L_norm}).
\section{Renormalization conditions}
Renormalization conditions can be obtained by demanding some renormalized correlation functions to be equal
to their tree level values at a given scale $\mu=L^{-1}$, where $L$ is the size of the system in the spatial directions. In order to specify a particular renormalization scheme
one has to: a) choose between the observables $\mathcal{G}_{i}^{\pm}$
and $\mathcal{L}_{i}^{\pm}$ for imposing the renormalization conditions, b) fix the parameters
$\alpha$, $\beta$ and $\gamma$, and c) fix the dimensionless parameters $\theta$, $\rho=T/L$
and the timeslice $x_{0}$ at which correlations are evaluated.
In the following we will always consider
$\rho=1$ and $x_{0}=T/2$.
It is convenient to write explicitly the parameter dependance on the matrix elements in Eq.(\ref{eq:G_L_norm})
and introduce the following notation,
\begin{equation}
\mathcal{H}_{i,{\bf c}}^{1,\pm}(\theta) = \mathcal{G}_{i}^{\pm}(x_{0}=T/2),\qquad
\mathcal{H}_{i,{\bf c}}^{2,\pm}(\theta) = \mathcal{L}_{i}^{\pm}(x_{0}=T/2),\qquad
{\bf c}=(\alpha,\beta,\gamma).
\label{eq:H1_H2}
\end{equation}
This way, renormalization conditions for the operator which does not mix are specified by
\begin{equation}
\mathcal{Z}_{11}^{\pm}(g_{0},a\mu)\mathcal{H}_{1,{\bf c}}^{s,\pm}(\theta)
= \left.\mathcal{H}_{1,{\bf c}}^{s,\pm}(\theta) \right|_{g_{0}=0},
\qquad s=1,2.
\label{eq:Q1_cond}
\end{equation}
For the pairs of operators which mix, we impose renormalization conditions on each
$2\times2$ matrix by forming combinations of the form \cite{MP2014}
\begin{equation}
\left(
{\begin{array}{cc}
\mathcal{Z}_{22} & \mathcal{Z}_{23} \\
\mathcal{Z}_{32} & \mathcal{Z}_{33}\
\end{array} }
\right)
\left(
{\begin{array}{cc}
\mathcal{H}_{2,{\bf c}_{1}}^{s_{1}}(\theta_{1}) & \mathcal{H}_{2,{\bf c}_{2}}^{s_{2}}(\theta_{2}) \\
\mathcal{H}_{3,{\bf c}_{3}}^{s_{1}}(\theta_{1}) & \mathcal{H}_{3,{\bf c}_{4}}^{s_{2}}(\theta_{2})\
\end{array} }
\right)=
\left.\left(
{\begin{array}{cc}
\mathcal{H}_{2,{\bf c}_{1}}^{s_{1}}(\theta_{1}) & \mathcal{H}_{2,{\bf c}_{2}}^{s_{2}}(\theta_{2}) \\
\mathcal{H}_{3,{\bf c}_{3}}^{s_{1}}(\theta_{1}) & \mathcal{H}_{3,{\bf c}_{4}}^{s_{2}}(\theta_{2})\
\end{array} }
\right)\right|_{g_{0}^{2}=0},
\label{eq:mat_cond}
\end{equation}
and similarly with the conditions for the operators $\mathcal{Q}_{4}$ and $\mathcal{Q}_{5}$.
In order to define sensible renormalization conditions the tree-level matrix of the r.h.s of
Eq.(\ref{eq:mat_cond}) must be invertible. Note that the source type, the parametes $\bf{c}$ and $\theta$ must be the same within the different columns.
The amount of parameters appearing in the renormalization conditions Eqs.(\ref{eq:Q1_cond})
and (\ref{eq:mat_cond}) leaves quite some freedom in choosing particular renormalization schemes.
It has been pointed out \cite{MP2014} that the anomalous dimensions (AD) of the four fermion operators that mix
can be large.
A possible criteria to choose a renormalization scheme is to look for those for which the ADs are not too large.
At the present order In perturbation theory this is equivalent to look for combinations of parameters
that lead to an RG-evolution for which the NLO contribution is close to the LO.
With this in mind, we describe the
RG-evolution of a matrix $\mathcal{Z}$ of Z-factors
$\mathcal{Z}(\mu_{1})=U(\mu_{1},\mu_{2})\mathcal{Z}(\mu_{2})$ as \cite{MP2014}
\begin{equation}
U(\mu_{2},\mu_{1})=\textrm{T}\exp\left\{
\int_{\overline{g}(\mu_{1})}^{\overline{g}(\mu_{2})}
\frac{\gamma(g)}{\beta(g)}dg
\right\}\equiv\tilde{U}^{-1}(\mu_{1})\tilde{U}(\mu_{2}).
\label{eq:formal_RG}
\end{equation}
The matrix $\tilde{U}(\mu)$ is expanded in perturbation theory as
\begin{equation}
\tilde{U}(\mu)=
\left[\frac{\overline{g}^{2}(\mu)}{4\pi}\right]^{-\frac{\gamma_{0}}{2\beta_{0}}}
\left[\textbf{1}+\overline{g}^{2}(\mu)J(\mu)+O(\overline{g}^{4})\right],
\end{equation}
where the matrix $J(\mu)$ contains the NLO evolution and satisfies
\begin{equation}
\frac{\partial}{\partial\mu}J(\mu)=0,\qquad
J-\left[\frac{\gamma_{0}}{2\beta_{0}},J\right]=
\frac{\beta_{1}}{2\beta_{0}^{2}}\gamma_{0}-\frac{1}{2\beta_{0}}\gamma_{1}.
\label{eq:J_eqs}
\end{equation}
The matrix $J(\mu)$ depends explicitly on the AD at NLO $\gamma_{1}$. A criteria in choosing a particular
renormalization scheme is thus to demand the norm of $J$ to be as small as possible.
This is equivalent to ask the NLO scale evolution of the Z-factors to be as close as possible to the LO evolution.
One can thus choose appropriate combinations
of observables and parameters in Eqs.(\ref{eq:Q1_cond}) and (\ref{eq:mat_cond}) to minimize the norm of $J$.
Schemes obtained following this strategy are desirable in the matching at high energy.
On the other hand, minimizing $|J|$ does not guarantee a good behaviour of the perturbative expansion.
Ultimately, only the comparison of the non-perturbative and perturbative
evolutions will determine which schemes allow for a reliable matching at high-energy.
The idea is thus to use 1-loop perturbation theory to identify a set of potentially
good schemes and most importantly avoid schemes with particularly large NLO
ADs. Finally, having several different candidate schemes allows to compare the RGI
operators obtained and thus have an estimate on the goodness of the matching.
\section{Results}
A given renormalization factor $\mathcal{Z}_{ij}$ in Eq.(\ref{eq:po_ren}) is expanded in perturbation theory as
$\mathcal{Z}_{ij}\simeq 1 + \overline{g}^{2}\mathcal{Z}_{ij}^{(1)}+O(\overline{g}^{4})$. The 1-loop coefficient $\mathcal{Z}^{(1)}_{ij}$
has an asymptotic form in $a/L$ given by
\begin{equation}
\mathcal{Z}^{(1)}_{ij}= \sum_{n=0}^{\infty}\left(r_{n,ij}+s_{n,ij}\ln(a/L)\right)(a/L)^{n}
\label{eq:Z_as}
\end{equation}
Here, $r_{0,ij}$ is the finite asymptotic part of $\mathcal{Z}^{(1)}_{ij}$, $s_{0,ij}$ is the universal LO anomalous dimension,
and the $r_{n,ij}$ and $s_{n,ij}$ coefficients with $n>0$ correspond to $O(a^{n})$ cutoff effects. Due to automatic $O(a)$
improvement, the coefficients $s_{1,ij}$ should be zero regardless of whether the action or the operators have been improved in the bulk. The coefficients $r_{1,ij}$ are zero once boundary $O(a)$ improvement is implemented.
In order to calculate the different $\mathcal{Z}^{(1)}_{ij}$ we expand
Eqs.(\ref{eq:Q1_cond}) and (\ref{eq:mat_cond}) to
$O(\overline{g}^{2})$ in perturbation theory
and compute numerically the Feynman diagrams contributing to the different correlation functions involved.
Details on the gauge fixing procedure and on the expansion of the correlation functions will be
reported elsewhere. In this way we obtain explicit expressions for the coefficients $\mathcal{Z}^{(1)}_{ij}$ in terms of the
parameters $\bf{c}$ \footnote{The rest of the parameters appearing in Eqs.(\ref{eq:Q1_cond}) and (\ref{eq:mat_cond}) must be fixed explicitely.}.
From this, it is easy to apply the criteria introduced in the previous section to find appropriate renormalization schemes. For all operators we
obtain the correct value of the LO anomalous dimension, which is a strong check on the calculation. Moreover, we find all
$r_{1,ij}$ and $s_{1,ij}$ coefficients to be consistent with 0, confirming the absence of $O(a)$ effects.
Examples for the determination of the finite part for some schemes considered are shown in Figure \ref{fig:dZ}.
\begin{figure}[!ht]
\centering
\includegraphics[clip=true,scale=0.5]{dZ.eps}
\caption{Convergence to the continuum limit of the 1-loop renormalization factors $\mathcal{Z}^{\pm,(1)}_{11}$
(left pannel) and $\mathcal{Z}^{+,(1)}_{44}$, $\mathcal{Z}^{+,(1)}_{45}$, $\mathcal{Z}^{+,(1)}_{54}$
$\mathcal{Z}^{+,(1)}_{55}$ (right pannel), for a specific scheme choosen for illustrative purposes, after substracting the corresponding finite parts and logarithmic divergencies. The
coloured discontinuous lines are fits to the data excluding the coarsest lattice spacings. }
\label{fig:dZ}
\end{figure}
For the opeators $\mathcal{Q}_{1}^{\pm}$, Eqs.(\ref{eq:formal_RG} - \ref{eq:J_eqs}) are scalar and it is
possible to find schemes for which $J(\mu)$ is as small as desirable.
For each pair of operators that mix, we build renormalization conditions considering all possible combinations
of observables with either $\theta=0.0$ or $0.5$. For each condition obtained in this way, we find the sets
of parameters $(\alpha_{i},\beta_{i},\gamma_{i})$ that minimize $|J|$. Since the value of $|J|$ remains stable in the
neighbourhood of the minimum, we can identify regions in parameter space for which the NLO
contributions are not unreasonably large. Any choice of parameters outside these regions
can otherwise lead to arbitrarily large NLO contributions.
As an example,
in Figure \ref{fig:U_running} we show the LO and NLO
evolution of $\tilde{U}$ in
terms of $\overline{g}^{2}$, in the basis in which $\gamma_{0}$ is diagonal, for 3 schemes $S_{i}$ ($i=1,2,3$)
for the pair $\left\{\mathcal{Q}_{4}^{+},\mathcal{Q}_{5}^{+}\right\}$. The values of $(\alpha_{i},\beta_{i},\gamma_{i})$
have been chosen to minimize $|J|$ in the three schemes. The LO is universal, with only the diagonal
terms being non-zero in the chosen basis.
The difference between LO and NLO evolution is more evident in some of the matrix elements, but which
matrix element differs the most depends on the scheme. The overall behaviour of the scale evolution looks however
comparable for all schemes once
$(\alpha_{i},\beta_{i},\gamma_{i})$ are chosen close to the values which minimise the norm of $J$.
\begin{figure}[!ht]
\centering
\includegraphics[clip=true,scale=0.5]{Utilde_several.eps}
\caption{Components of tha matrix $\tilde{U}(\mu)$ in the basis in which $\gamma_{0}$ is diagonal. as a function of $\overline{g}^{2}$ for the operators $\{\mathcal{Q}^{+}_{4},\mathcal{Q}^{+}_{5}\}$, in three different renormalization schemes.}
\label{fig:U_running}
\end{figure}
\section{Conclusions}
In this work we have set up the $\chi$SF for the renormalization of parity odd $\Delta F=2$ 4-fermion operators.
We have defined a set of 3 point correlation functions for building renormalization conditions. From these, we expect
an improvement of the statistical error
in non-perturbative calculations with respect to the standard SF, where 4 point functions
have to be used \cite{MP2014}.
Thanks to automatic $O(a)$ improvement, bulk operator improvement can be avoided. Likely, this will allow to extrapolate the lattice SSF of these operators as $O(a^{2})$, increasing significantly the precision on the continuum results.
The flexibility of our renormalization conditions allows us to seek for schemes for which the NLO is as close as possible to the LO RG-evolution. In fact, we were able to identify potentially good schemes in this respect. The strategy developed seems promising, and deeper studies in parameter space are on the way. Ultimately of course only a full non-perturbative study can show which are the most suitable schemes for the determination of the relevant RGI operators. An additional important point to investigate is the size of cutoff effects in the corresponding SSFs. These results will then set the ground for future non-perturbative studies using the $\chi$SF.
Numerical calculations were performed at the Galileo machine at CINECA. We gratefully thank Stefan Sint and Tassos Vladikas for insightful discussions.
|
\section{Introduction}
\label{intro.sec}
The 2.16-m reflector is an English equatorial mount telescope at Xinglong
Observatory and it obtained the first light in 1989, which was
built/developed by Chinese ourselves independently.
The effective aperture of the telescope is 2.16 meter, and the focal ratio of
the primary mirror is f/3 \citep{szy89}. The effective aperture of the
secondary mirror is 0.717 m. Currently, there are two available foci for
mounting astronomical observing instruments, the Cassegrain focus and the
Coud\'{e} focus. For the Cassegrain focus, which is a R-C system, the focal
ratio is f/9, and the scale on the focal plane is $10\farcs61$/mm. While for
the Coud\'{e} system, it is f/45, and the scale on the focal plane is
$2\farcs12$/mm.
The most special character of the optical system is that the Cassegrain
system and the Coud\'{e} system share the same secondary mirror and there is
a relay mirror in the Coud\'{e} system, and both systems can
sufficiently eliminate the spherical aberration and coma aberration
\citep{szy89}.
Since its first light in 1989, the 2.16-m telescope has been used in various
scientific research fields in Galactic and extragalactic astrophysics,
including the determination of stellar parameters (e.g.,
abundances, surface gravity, temperatures) of a large sample of stars,
the discoveries of substellar and planetary companions of stars, studies of
active galaxy nuclei (AGN), including the identifications of high-redshift
quasars, discoveries and studies of supernovae, e.g., 1993J \citep{wh94},
as well as time-domain science (e.g., supernovae, gamma-ray bursts,
stellar tidal disruption events, and variable stars).
Although the 2.16-m telescope is the third largest optical telescope in
China right now (smaller than LAMOST and the Lijiang 2.4-m telescopes, which
were installed in 2008 and 2007, respectively), it plays an important role in
Chinese astronomical observations. Every year there are hundreds of Chinese
astronomers applying for observing time at the 2.16-m telescope, including
significant numbers of new telescope users and graduate students.
Therefore, it is of
great importance and useful to describe the specific parameters and the
observing ability of the telescope and its instruments to the telescope users,
which could be very helpful for planning and carrying out observations
and for the data reduction. This paper is
organized as follows: the Xinglong Observatory is introduced in
Sect.\ref{obs.sec}; the telescope and its instruments are described in
Sect.\ref{tel.sec} and the efficiencies of the telescope and its instruments
also have been estimated in this section. The relevant science and
projects based on observational data obtained by the 2.16-m telescope are
described in Sect.\ref{sci.sec}; finally a brief summary is given and some
future facilities are discussed in Sect.\ref{sum.sec}.
\section{The Xinglong Observatory}
\label{obs.sec}
The Xinglong Observatory of National Astronomical Observatories, CAS (NAOC)
(IAU code 327, coordinates: $40^{\circ}23'39''$ N, $117^{\circ}34'30''$ E)
was founded in 1968. At present, it is one of most primary observing stations
of NAOC. As the largest optical astronomical observatory site in the continent
of Asia, it harbors 9 telescopes with effective aperture greater
than 50\,cm. These are LAMOST, the 2.16-m reflector, a 1.26-m optical \&
near-infrared reflector, a 1\,m Alt-Az reflector, a 85\,cm reflector
(NAOC-Beijing Normal University Telescope, NBT), a 80\,cm reflector (Tsinghua
University-NAOC,
TNT), a 60/90\,cm Schmidt telescope, a 60\,cm reflector and a 50\,cm reflector.
The average altitude of the Xinglong Observatory is $\sim960$ m and it is
located at the South of the main peak of the Yanshan Mountains, in the
Xinglong county, Hebei province, which is $\sim120$km northeast of Beijing.
The mean and median seeing values of the Xinglong Observatory are
$2\farcs0$ and $1\farcs8$, respectively, and on average, there are 117
photometric nights and 230 useful nights per year based on the data of
2007-2014 \citep{zhang15}. For most of the time, the wind speed is less
than 4 m s$^{-1}$ (the mean value is 2 m s$^{-1}$), and the sky
brightness is $\sim$21.1 mag arcsec$^{2}$ in $V$ band at the zenith
\citep{zhang15}.
Each year, more than a hundred astronomers use the telescopes of
Xinglong Observatory to perform the observations for the studies on Galactic
sciences (stellar parameters, extinction measurements, Galactic
structures, exoplanets, etc.) and extragalactic sciences (including nearby
galaxies, AGNs, high-redshift quasars), as well as time-domain astronomy
(supernovae, gamma-ray bursts, stellar tidal disruption events, and different
types of variable stars).
In recent year, besides the basic daily maintenance of the telescopes,
new techniques and methods have been explored by the engineers and
technicians of Xinglong Observatory to improve the efficiency of observations.
Meanwhile, the Xinglong Observatory is also a National popular-science and
education base of China for training students from graduate schools,
colleges, high schools and other education institutions throughout China,
and it has hosted a number of international workshops and summer schools.
\section{The 2.16-m reflector and its instruments}
\label{tel.sec}
As shown in Figures~\ref{fig1} and \ref{fig2}, the telescope is a R-C system
on an English equatorial mount. The effective aperture of the primary mirror
is 2.16 meter. Currently, there are three primary instruments
available for the Cassegrain focus: BFOSC, OMR and the fiber-fed HRS.
Previously,
a high resolution spectrograph was mounted on the Coud\'{e} focus before the
fiber-fed HRS was mounted in 2010. Since then, the Coud\'{e} focus has
only been used for a few special experiments, such as adaptive optics
tests. As shown in Figure~\ref{fig1}, the diameter of the dome is 22 meter,
which is relative large for a 2m-class telescope, due to its design. The
rotation part of the telescope weights 91 tons and the pointing accuracy
which is modified with pointing model is $rms=10\farcs0$. While the
tracking precision of the telescope is $rms=20\farcs0$ within 10 minutes, and
$rms=50\farcs0$ within 1 hour. The position precision of the telescope on the
sky under guiding is $rms<0\farcs15$ \citep{huang15}.
\begin{figure}
\centerline{
\includegraphics[scale=0.5,angle=-90]{fig1.ps}}
\caption[]{Cartoon of the 2.16-m telescope at Xinglong Observatory,
which is an English equatorial mount, in a large dome with diameter of 22
meter.}
\label{fig1}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[scale=0.4,angle=0]{fig2.ps}}
\caption[]{The Xinglong 2.16-m telescope at Xinglong Observatory.}
\label{fig2}
\end{figure}
\subsection{The BFOSC instrument}
The BFOSC is one of the primary instruments of the telescope, for which
the design and processing of mechanism and electronics control, the design of
grims and assembling and debugging of the whole system were done by
University of Copenhagen and ESO provided consult and optical design of the
focal reducer. It is available for the f/9 Cassegrain focus. The scale of
the focal plane is
$10\farcs61$/mm. As discussed in Sect.~\ref{intro.sec}, it can
be used for both imaging mode and spectroscopy mode, which is switchable.
Figure~\ref{fig3} shows the optical layout of the BFOSC instrument, including
the aperture wheel, filter wheel, grism/echelle wheel, calibration device,
guiding device, collimator, shutter, camera and the CCD detector.
\begin{figure}
\centerline{
\includegraphics[scale=0.8,angle=0]{fig3.ps}}
\caption{The optical and mechanical layout of the BFOSC instrument
from \citet{hwl}.}
\label{fig3}
\end{figure}
There are eight positions on the aperture wheel: three of them for direct
imaging, coronagraph mask and focal adjusting plate, respectively and the
other five for long/short slit plates. For the long slits, the lengths are
all $9'.4$ and there are nine options for the slit widths ($0\farcs6$,
$0\farcs7$, $1\farcs1$, $1\farcs4$, $1\farcs8$, $2\farcs3$, $3\farcs6$,
$7\farcs0$ and $14\farcs0$). While for the
short slits, the slit lengths are various for different slit width:
$3\farcs5$ for slit width of $0\farcs6$, $4\farcs0$ for slit width of
$1\farcs0$, $3\farcs6$ for slit width of $1\farcs6$ and $3\farcs7$ for
both slit widths of $2\farcs3$ and $3\farcs2$. The coronagraph mask is used
to block out the intense light from bright sources near the observing targets,
with circular spots of which the diameters are $2\farcs0$, $3\farcs0$,
$4\farcs0$, $6\farcs0$, $9\farcs0$ and $12\farcs0$.
For the filter wheel, there are several sets of filters available. The
Johnson-Cousins $UBVRI$ filters are for the broadband photometry.
In addition, there are special filters used for the spectroscopic observing:
the $Z$ band for the transmitting spectral region $\lambda\ge910$ nm, and the
$385LP$ band for removing the 2nd-order spectrum in the wavelength region
$\lambda\le385$ nm. The transmission curves are shown in Figure~\ref{fig4}.
\begin{figure}
\centerline{
\includegraphics[scale=0.7,angle=-90]{fig4.ps}}
\caption[]{The transmission curves of the BFOSC filters, including the
Johnson-cousins $UBVRI$ bands as well as the $385LP$ and $Z$ bands for
spectroscopic observations.}
\label{fig4}
\end{figure}
Besides, a series of interference narrow band filters covering [OIII]
(rest-frame wavelength of $\lambda$5007A), $\rm H{\alpha}$ and HeII
(rest-frame wavelength $\lambda$4686A) are also available for the BFSOC
imaging mode. For series of [OIII] band observations, there are 8 filters of
which the central wavelengths are
between 500.9-536.0 nm with FWHM of 6 nm, corresponding to redshifts of
0-20000 km s$^{-1}$. While for the series of $\rm H_{\alpha}$ bands, there
are 11 filters of which the central wavelengths are between 656.2-706.0 nm
with FWHM of 7 nm, corresponding to redshifts of 0-22000 km s$^{-1}$. For
the He bands, there are three filters of which the central wavelengths are
between 447.1-468.6 nm with FWHM of 6 nm. In addition, an [OIII] filter with
FWHM of 13 nm and an $\rm H_{\alpha}$ filter with FWHM of 14 nm are also
provided for a redshift of z=0. These filters can be used for the study
of star-forming regions of nearby galaxies at different redshifts.
Table~\ref{t1.tab} presents the parameters of the grisms/prisms/echelle for
the BFOSC instrument. From left to right in columns are names, working spectral
orders, reciprocal linear dispersions, dispersions and wavelength coverages of
the various grisms/prisms/echelle, respectively. The configurations can be
chosen by the users depending on the different requirements of the projects.
For grisms G3/G6/G10, the blue ends of wavelength coverages are limited by the
cutoff of the atmospheric window, while for grisms G5/G8, the red ends are
limited by the size of the CCD. The low dispersion grisms G10/G11/G12 are also
used for the cross-disperser when mounted on the filter wheel. The echelle E13
is only used for measuring the velocity field of extended sources with the 3rd
order spectrum, and the $V$ band filter is recommended to be applied together
with to remove the other order of the spectrum.
\begin{table}[ht!!!]
\small
\centering
\begin{minipage}[]{100mm}
\caption[]{The Parameters of Grisms/Prisms/Echelle for the BFOSC Instrument
from \citet{hwl}.}
\label{t1.tab}
\end{minipage}
\tabcolsep 3mm
\begin{tabular}{ccccc}
\hline\noalign{\smallskip}
Name & Spec.Ord. & Rec. Lin. Disp. & Disp. & Wav. Range \\
& & (\AA/mm) & (\AA/pix) & (\AA) \\
\hline\noalign{\smallskip}
P1 & & $573-2547$ & $8.6-38.2$ & $4000-5600$ \\
\hline
G3 & 1 & 139 & 3.12 & $3300-6400$ \\
G4 & 1 & 198 & 4.45 & $3850-7000$ \\
G5 & 1 & 199 & 4.47 & $5200-10120$ \\
\hline
G6 & 1 & 88 & 1.98 & $3300-5450$ \\
G7 & 1 & 95 & 2.13 & $3870-6760$ \\
G8 & 1 & 80 & 1.79 & $5800-8280$ \\
\hline
G10 & 1 & 392 & 8.80 & $3300-6400$ \\
G11 & 1 & 295 & 6.63 & $3900-7300$ \\
G12 & 1 & 837 & 18.8 & $5200-10200$ \\
\hline
E9+G10 & $21-11$ & $16.8-38.4$ & $0.38-0.86$ & $3300-6400$ \\
E9+G11 & $18-9$ & $21.0-47.9$ & $0.47-1.076$ & $3900-7300$ \\
E9+G12 & $14-6$ & $29.0-73.2$ & $0.65-1.64$ & $5200-10200$ \\
\hline
E13+V & 3 & 33.1 & 0.666 & $4980-5990$ \\
\noalign{\smallskip}\hline
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
Table~\ref{t2.tab} presents the spectral resolutions of some frequently used
BFSOC grisms at a minimum slit width $0\farcs6$ and a slit width of
$\sim2\farcs3$ used in typical seeing condition, which are estimated with
the emission lines of the planetary nebula (PN) IC4997. The observations were
taken on November 24, 2014 and February 26, 2016, with gratings of G4, G6,
G7 and G8 on BFOSC. Throughout the nights of the observations, the weather were
clear and the seeing were between $\sim2\farcs0 - 2\farcs4$. The exposure time
was 0.2-10 seconds depending on the dispersions of the different grisms.
The shortest exposure of 0.2 second is considerably longer than the shutter
speed of 1.5 milliseconds.
\begin{table}[ht!!!]
\small
\centering
\begin{minipage}[]{100mm}
\caption[]{Spectral resolutions of several frequently used grisms of BFSOC at\
a slit width of $0\farcs6$ and $2\farcs3$, which is around the typical seeing.}
\label{t2.tab}
\end{minipage}
\tabcolsep 3mm
\begin{tabular}{c|cccc|cccc}
\hline\noalign{\smallskip}
Wavelength & G4 & G6 & G7 & G8 & G4 & G6 & G7 & G8 \\
\hline\noalign{\smallskip}
({\AA}) & \multicolumn{4}{c|}{$0\farcs6$} & \multicolumn{4}{c}{$2\farcs3$} \\
\hline\noalign{\smallskip}
4341 & 531 & 1423 & 1204 & \nodata & 246 &463 &478 & \nodata \\
4363 & 557 & 1503 & 1252 & \nodata & \nodata &\nodata &\nodata & \nodata \\
4861 & \nodata &\nodata &\nodata &\nodata & 262 & 538 & 521 & \nodata \\
4959 & & & & & 265 & 540 & 524 & \nodata \\
5007 & 620 & 1413 & 1328 & \nodata & 265 & 555 & 532 & \nodata \\
6563 & 824 & \nodata & \nodata & 2245 & 321 & \nodata & \nodata & 820 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
For the calibration device, four lamps can be mounted at most. One is for
the flat-fielding correction and the other lamps can be used for wavelength
calibration. The lights of the lamps illuminate the integrating sphere at
first, then modified to be f/9 light beam and finally reflected to the
focal plane of the BFOSC instrument. The Fe lamp and Ne lamp are frequently
used for the wavelength calibration.
On the guiding device, a mirror reflects the view outside of the observing
field of view (FOV) to an ICCD, which is movable in 3D directions to find a
proper guide star and also to adjust the focus. The FOV of the guiding device
is $10'\times20'$, which is large enough to find a suitable guide star.
In 2010 an E2V $55-30-1-348$ back illuminated CCD, AIMO was installed on the
spectrograph and the CCD controller was made by the Lick Observatory. The
size of the CCD is $1242\times1152$ pixels with pixel size of 22.5 $\mu$m.
The pixel scale is $0\farcs457$ and the FOV is $9'.46 \times8'.77$ according
to the size of CCD. Figure~\ref{fig5} shows the quantum efficiency (Q.E.) of
the CCD. It can be seen that the maximum quantum efficiency is higher than
90\% around 5700 {\AA} of the wavelength.
\begin{figure}
\centerline{
\includegraphics[scale=0.7,angle=0]{fig5.ps}}
\caption[]{The quantum efficiency of the BFOSC CCD camera.}
\label{fig5}
\end{figure}
Table~\ref{t3.tab} lists the gains, readout noises and readout speed of
the BFOSC CCD at various readout times. Observers can choose different options
for the specific observations and the slow readout speed is applied in
most of the time.
\begin{table}[ht!!!]
\small
\centering
\begin{minipage}[]{100mm}
\caption[]{Gains, Readout Noises and readout Speed of BFOSC CCD from
\citet{hwl}.}
\label{t3.tab}\end{minipage}
\tabcolsep 3mm
\begin{tabular}{ccccc}
\hline\noalign{\smallskip}
Readout Speed & Mode & Gain & Readout Noise & Readout Time \\
& & ($e^-$/ADU) & ($e^-$/pix.) & (second) \\
\hline\noalign{\smallskip}
Fast & 0 & 99.58 & 132.5 & 5 \\
Fast & 1 & 49.44 & 103.7 & 5 \\
Fast & 2 & 24.48 & 87.35 & 5 \\
Median & 0 & 2.33 & 6.80 & 8 \\
Median & 1 & 1.22 & 5.93 & 8 \\
Median & 2 & 0.50 & 6.06 & 8 \\
Slow & 0 & 2.43 & 3.25 & 28 \\
Slow & 1 & 1.13 & 2.58 & 28 \\
Slow & 2 & 0.50 & 2.46 & 28 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
There are six observing modes for the BFOSC instrument:
(1) direct imaging; (2) long slit spectroscopy; (3) slitless
spectroscopy; (4) echelle grism spectroscopy; (5) coronograph mask;
and (6) multiple object Spectroscopy. Figure~\ref{fig6} shows the
optical layout of the lens and light path for the direct imaging mode
and the spectroscopy observing mode, which could be switched in a few minutes.
\begin{figure}
\centerline{
\includegraphics[scale=0.9,angle=0]{fig6.ps}}
\caption{The layout of the lens of two different working modes of the BFOSC
instrument from \citet{hwl}.}
\label{fig6}
\end{figure}
Since 2012, a multi-object spectroscopy (MOS) observing mode of BFSOC is
available for observers by placing a multiple-aperture mask on the
aperture wheel \citep{zhou14}. In this mode, 10-20 objects can be observed
simultaneously, depending on the spatial distribution density of the targets.
The MOS improves the observing efficiency of multiple-object observations,
such as star-forming activities of HII regions of nearby galaxies,
star clusters, or groups and clusters of galaxies.
The total efficiency of the observing system, including the atmosphere,
telescope and its instruments, is what the observers are mostly concerned
about. In fact, there are a number of factors to be considered for calculating
the total efficiency of the system, such as
atmospheric extinction, reflectivity of primary and secondary mirror,
transmissions of filters and the quantum efficiency of the CCD. It can be
estimated through observing standard stars in the Equation~\ref{eq1},
\begin{equation}
\label{eq1}
\eta (\lambda)=\frac{F_{ADU} \cdot G}{F_{\lambda} \cdot \delta \lambda \cdot S_{tel}}
\end{equation}
In this formula, $F_{\rm ADU}$ is the observed number counts of a standard
star per second (ADU s$^{-1}$); $G$ is the gain of the CCD ($e^{-1}$ ADU$^{-1}$);
$F_{\lambda}$ is theoretical photon flux of a standard star derived from its
AB mag ($photon s^{-1} cm^{-2}$ \AA$^{-1}$) ; $\delta \lambda$ is the
effective band width of the filter in imaging observations or the dispersion of
the grating for spectroscopic observations (\AA);
$S_{\rm tel}$ is the effective area of the primary mirror of the telescope
($cm^2$) and $\lambda$ is the effective wavelength of the filter or the
wavelength at which the efficiency is to be computed for the spectroscopy (\AA).
In order to estimate the total efficiency of the 2.16-m telescope, a Landolt
standard star PG2336+004B was observed on November 21, 2014 in the broadband
$UBVRI$ bands of the BFOSC photometric system by using Equation~\ref{eq1}.
The seeing was $\sim2\farcs0-3\farcs0$ throughout the whole night of
observations and the airmass is $\sim1.3$. Figure~\ref{fig7} is a plot of
the total efficiency in different
bands as a function of central wavelength. The atmospheric extinction,
reflectivity of primary and secondary mirrors, transmissions of filters,
Q.E. of CCD as well as other factors were not corrected in the calculation.
It can be seen that the total efficiency is relative low in the $U$ band
($\sim2$\%) and $B$ band ($\sim7$\%), but it is relatively high in $VRI$ bands
($\sim15$\%).
\begin{figure}
\centerline{
\includegraphics[scale=0.5,angle=0]{fig7.ps}}
\caption[]{The total efficiency of the BFOSC photometric system including the atmosphere of the 2.16-m telescope in the $UBVRI$ bands at an airmass of $\sim1.3$.}
\label{fig7}
\end{figure}
The limiting magnitude is also an important quantity to evaluate the observing
ability of a telescope. Actually at the same night on November 21, 2014,
a Landolt standard star PG2336+004 was observed in the $UBVRI$ bands. The
airmass was $\sim1.3$. The limiting magnitude (signal-to-noise
ratio, $SNR=5$) as a function of exposure time in the $UBVRI$ bands are shown
in Figure~\ref{fig8} and the same relations but for the limiting magnitudes
of $SNR=10$ are shown in Figure~\ref{fig9}. It is noted that in the 640s
exposure the limiting magnitudes of the $B/V/R$ bands could reach $\sim22$ mag
in $SNR=5$ and $\sim21$ mag in $SNR=10$. The limiting magnitude in the $U$
band is the lowest because the sensitivity of the observing system in the blue
band is relatively low, further affected by the effect of significant
atmospheric extinction in blue band. However, although the efficiency in the
$I$ band is similar to those of the $VR$ bands, the night-sky background is
much brighter ($\sim2$ times) than in the $VR$ bands. In addition, the fringe
of the CCD in the $I$ band contributes significant noise. Although we have
tried the de-fringing method introduced by \citet{c87}, it has not been
improved significantly. Therefore, these factors together make the limiting
magnitudes $\sim1$ mag shallower than those in the $BVR$ bands.
\begin{figure}
\centerline{
\includegraphics[scale=0.5,angle=0]{fig8.eps}}
\caption[]{The limiting magnitudes in the $UBVRI$ bands for the
signal-to-noise ratio of $SNR=5$ at an airmass of $\sim1.3$.}
\label{fig8}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[scale=0.5,angle=0]{fig9.eps}}
\caption[]{The same as in Figure~\ref{fig8}, but for a signal-to-noise ratio
of $SNR=10$.}
\label{fig9}
\end{figure}
In order to estimate the efficiency of the spectroscopic system of BFOSC,
the two ESO standard stars HR153 and HR9087 were observed on November 24,
2014, with several frequently used grisms (G4, G5, G6, G7 and G8)
of the BFOSC instrument. The weather was clear and the seeing was
$\sim2\farcs4$. The exposure time was 0.2-10 seconds according to the
dispersions of the grisms.
To make sure that most of the flux can be obtained and measured for
estimating the efficiency, the slit was configured as $7\farcs0$.
The total efficiency, which is defined above, including
atmospheric extinction and instruments, was calculated with
Equation~\ref{eq1} and shown in Figure~\ref{fig10}. For the grisms G4 and G5,
the peak efficiencies are $\sim$15\% and $\sim$10\% respectively,
while for the other three grisms, the total efficiencies of the peak
are $\sim$5\%. The filter $385LP$ was used for removing the 2nd-order
spectrum of wavelength $\lambda \ge$385 nm.
\begin{figure}
\centerline{
\includegraphics[scale=0.55,angle=0]{fig10.eps}}
\caption[]{The total efficiencies estimated for grisms G4, G5, G6, G7 and G8
of the BFOSC instrument.}
\label{fig10}
\end{figure}
For the limiting magnitude in the spectroscopic observations, the
previous observations show that typically it could reach $V=20$ mag of
$SNR=5$ in the grism G6, with exposure of 1 hour, given by \citet{hwl}.
\subsection{The Spectrograph made by Optomechanics Research inc. (OMR)}
The OMR is another low-intermediate resolution spectrograph available on
the Cassegrain focus of the 2.16-m telescope. It was made by Optomechanics
Research Inc. (Arizona, USA) at the end of 1994 and was tested on the 2.1m
telescope at Kitt Peak. Then, it was installed on the 2.16-m telescope of
Xinglong Observatory in 1995 and was available to astronomers in 1996.
The optical layout of the OMR is shown in Figure~\ref{fig11}. The system is
composed of slit \& decker assembly, filters, spectrograph, calibration
system, collimator, gratings, CCD camera, guiding CCD camera, console and data
collecting system.
\begin{figure}
\centerline{
\includegraphics[scale=0.9,angle=-90]{fig11.ps}}
\caption[]{The optical layout of the OMR spectrograph.}
\label{fig11}
\end{figure}
The slit \& decker assembly is composed of decker, the slit width adjuster,
calibration reflector, and driving device. The top surface of the slit is
aluminized stainless steel with high reflecting power while the slit
edge was polished to be very sensitive and should not be touched. The
slit is tilted by $20^{\circ}$ with respect to telescope optical axis and the
projected slit width
can be adjusted remotely through the main console in the range of 0.05-1.0mm
(corresponding to $0\farcs5-10\farcs6$), which is displayed on the main
console via the voltage value. The size of the slit jaws is 32.8mm$\times$38mm
and the effective length of the slit is 28.8mm (corresponding to $5'.1$).
The slit jaw reflects the light to the guiding system.
There are six positions on the filter wheel for the filters of Clear, Corning
$4-71$, Schott $BNG-37$, $BG-39$, $GG-475$ and $RG-695$. All the filters have
the same size of 25mm$\times$25mm. The central wavelengths of filters could be
matched with the 1st or 2nd order of the spectrum by the softwares
automatically.
There are three wavelength calibration lamps (Fe-Ne, Fe-Ar and He-Ar)
and a flat-fielding lamp for the calibration device. All of lamps can be
controlled remotely. For the Fe-Ne lamp, a standard 1.5-inch ISTC Model
\#WL-22810 is supplied, and actually other types of lamps can
be used as well, such as Hamamatsu and Starna. The maximum working current is
20 mA and the normal working current is 10-12 mA. The Fe-Ar lamp ISTC Model
is \#WL-22611. For the safety of He-Ar lamp, the DC power should
be used for the supply. As for the flat-fielding lamp, the power is the
standard 12V DC, 1.5A halogen tungsten lamp (Sylvania\#808-301550) and it
usually can be used for 1000-2000 hours. Meanwhile, there are four condensers
for the calibration lamps, which are made of quartz or pyrex glass. The
diameters are 25 mm and the focal lengths are $f=38$ mm. The field lens is
made of quartz, with a diameter of 38 mm and a focal length of 64 mm.
The mirror of the collimator is an off-axis parabolic aluminized reflector,
made of pyrex glass. The aperture of the mirror is $D=110$ mm and the focal
length is $f=674$ mm. The off-axis angle is 8.1$^{\circ}$ and the focus can
be adjusted remotely.
At present there are six reflecting blazed gratings mounted on the OMR
spectrograph. The gratings are made of aluminized pyrex glass and
can be switched manually. The parameters of the blazed gratings are given in
Table~\ref{t4.tab}, including the number of the gratings, the grooves, groove
areas, reciprocal linear dispersions, dispersions, blaze wavelengths of the
1st order spectra and the blaze angles. The overall working
wavelength coverage of the gratings is $\sim$3700-10000 {\AA}. For the
specific grating, the wavelength coverage is adjustable and it can be
estimated through parameters of Table~\ref{t4.tab} and the size of the CCD. For
instance, the wavelength coverage of gratings of 1200 lines/mm, is $\sim1380$
{\AA} while for the grating of 300 line/mm, it is $\sim5420$ {\AA}. The SPEC
software, {which is the camera controlling software designed and installed with the
PI Spec-10 CCD camera}, can recognize the gratings and the wavelength coverage
can be adjusted via SPEC according to the requirements of the observers.
The camera is a Schmidt-Cassegrain system. The useful aperture is $D=100$ mm
and the focal length is $f=150$ mm. The CCD mounted is a PI Spec-10 PIXIS
$1340{\times}400$ scientific CCD detector, which delivers the highest
sensitivity possible and $>16-$bit dynamic range for spectroscopy
applications. The pixel size is 20$\times$20 $\mu$m pixels and the CCD size
is $26.8{\times}8.0$ mm$^2$. The deepest cooling temperature is $-75^{\circ}$
C. The quantum efficiency of the CCD is shown in Figure~\ref{fig12}, which is
midband.
\begin{figure}
\centerline{
\includegraphics[scale=0.7,angle=0]{fig12.ps}}
\caption[]{The quantum efficiency of the CCD camera of the OMR Spec-10, which is midband.}
\label{fig12}
\end{figure}
The guiding system of the OMR is composed of a reflecting system and a CCD
Camera. In the reflecting system, the slit reflecting mirror is aluminized and
the r.m.s. of the surface flatness is 1/4 $\lambda$. For the transfer lens,
the diameter is 50 mm, with focal length of $f=260$ mm. While for focal lens,
the diameter is 36 mm, with focal length of $f=85$ mm. Both of the lenses have
been coated. The CCD camera is an Alta U42 CCD, made by Apogee Imaging Systems,
Inc. The sensor is E2V CCD$42-40$ and the size is 2048$\times$2048 pixels.
The Gain is 1.3e$^-$/count and the maximum digitized well capacity is 82k
e$^-$. The dark current is 0.39e$^-$/pixel/s.
\begin{table}[ht!!!]
\small
\centering
\begin{minipage}[]{100mm}
\caption[]{The Current Parameters of the OMR Gratings.}\label{t4.tab}\end{minipage}
\tabcolsep 3mm
\begin{tabular}{ccccccc}
\hline\noalign{\smallskip}
No. & Grooves & Gro. Area & Rec. Lin. Disp. & Disp. & Blz. Wav. & Blz. Ang. \\
& (l/mm)&(mm) & (\AA/mm) & (\AA/pix) &(1st order) (\AA) & ($^{\circ}$)\\
\hline\noalign{\smallskip}
1 & 150 & $90\times90$ & 400 & 8.0 & 5000 & 2.2 \\
2 & 300 & $90\times90$ & 200 & 4.0 & 8000 & 6.5 \\
3 & 600 & $102\times102$ & 100 & 2.0 & 6500 & 11.3 \\
4 & 1200 & $102\times102$ & 50 & 1.0 & 8000 & 28.7 \\
5 & 300 & $102\times102$ & 200 & 4.0 & 4224 & 3.6 \\
6 & 1200 & $102\times102$ & 50 & 1.0 & 4000 & 13.9 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
Similarly, the Landolt standard star Feige 66 was observed with two blazing
gratings 1200B and 1200R of the OMR instrument and the total efficiencies
(including the atmosphere extinction, the reflecting rate of mirrors, Q.E.
of CCD, etc.) have been computed through Equation~\ref{eq1}. As described
above, the central wavelength and coverage is adjustable, Figure~\ref{fig13}
shows the total efficiencies of two gratings 1200B and 1200R, in two different
central wavelength and wavelength coverages. It can be seen that the
efficiency ranges from $\sim 1\%$ to $\sim 2\%$, which is relatively lower than
that of the BFOSC gratings.
\begin{figure}
\centerline{
\includegraphics[scale=0.55,angle=0]{fig13.eps}}
\caption[]{The total efficiency estimates with blazing gratings 1200B/1200R of
the OMR, in two different central wavelengths and wavelength coverages.}
\label{fig13}
\end{figure}
Typically, for a star of $V=17.3$ mag in an exposure time of 1800s, the
signal-to-noise ratio is $SNR=12$ at wavelength of $\lambda$=5500 {\AA};
for a star of SDSS $g=19.8$ mag in an exposure time of 3600s,
the signal-to-noise ratio is $SNR=5$ on average for the whole band.
\subsection{The fiber-fed High Resolution Spectrograph (HRS)}
Previously an echelle spectrograph was mounted on the Coud\'{e} focus of the
telescope for high resolution spectroscopic observations.
The special design of the optical system is to allow the Cassegrain
system and the Coud\'{e} system to share the same secondary mirror and there is
a relay mirror in the Coud\'{e} system. When switching between the two
systems, the Coud\'{e} system can eliminate spherical aberration and coma
aberration sufficiently just by slightly moving the secondary mirror
\citep{szy89}. The resolution power was between R=16,000 to 170,000 (79 gr/mm)
in the blue beam from 330nm to 580nm
and R=13,000 to 170,000 (31.6 gr/mm) in the red beam from 520nm to 1100nm.
For an exposure time of 1 hour, the limiting magnitude could reach $V=9.5$ in
the red band and $V=7.2$ mag in the blue band with a signal-to-noise ratio
$S/N=100$ \citep{zhao00,zl01}.
Since 2010, a new fiber-fed High Resolution Spectrograph (HRS) has
been developed by the Nanjing Institute of Astronomical Optics \&
Technology (NIAOT) to satisfy the scientific requirement of, for instance,
exoplanets surveys, the study of stellar abundances, and stellar
magnetic activities. The optical layout and light path are shown in
Figure \ref{fig14}. The HRS system is available for the Cassegrain focus and
composed of four parts: 1) the Cassegrain connector of the telescope; 2)
the fiber connector and the micro-optical module for focal-ratio changing;
3) the main body of the spectrograph; 4) the data collecting part of the CCD
camera.
The fiber-fed HRS observing mode could be switched to other instruments (such
as BFOSC or OMR) conveniently through the Cassegrain focal interface
of the telescope within only a few minutes. The calibration system
of HRS, the $\rm I_2$ cell and its heating system, as well as the tip/tilt
system are mounted on the interface of the spectrograph. The whole main
body is sealed in a protective box and the temperature variation is less
than $r.m.s.=0.5^\circ$C for a week. At present, a $2\farcs4$ aperture fiber
is configured and the tip/tilt system is working to improve the efficiency of
the system when the seeing is ideal. The fiber of $1\farcs6$ is at the
commissioning stage and it could be available very soon. An environmental
controlling system can accurately keep the stability of both temperature and
humidity for the HRS system. Furthermore, a sub-controlling system of
pressure will be installed at this system in the future.
The basic parameters of HRS are listed in Table~\ref{t5.tab}. The working
wavelength coverage is 360-1000 nm and the instrumental efficiency of
the spectrograph is $\ge 34\%$ for the peak at a wavelength coverage of
640-790 nm ($RI$ band), and $\ge10\%$ for $\lambda > 4500$ {\AA} (the whole
working band), based on the tests of November 23, 2010. The spectral resolution
is 32,000-106,000 for the spectrograph and it is $R=49800$ at a fixed slit
width of 0.19 mm (corresponding to 1\farcs8) based on the test of April 12,
2011. The stability of
the instrument for the velocity measurement is $r.m.s.=\pm6$ m/s and the
temperature is quite stable even for two weeks. The CCD camera is a
back-illuminated first-order red-sensitive E2V CCD 203-82. The size is
$4096\times4096$ pixels with a pixel size of 12 $\mu$m. The typical quantum
efficiency of the CCD is $>90\%$ under the temperature of $-100^{\circ}$C in
the wavelength coverage of $\sim500-650$ nm, which is shown as the solid
line in Figure~\ref{fig15}.
The liquid nitrogen (LN) holding time of the system is $\sim20$ hours, and the
cooling temperature is $-106^{\circ}$C. For $Gain=1.01$ e$^-$/ADU, the readout
noise (RON) is 2.84 e$^-$ at the readout speed of 50k, 4.29 e$^-$ at readout
speed of 100k and 7.88 e$^-$ at readout speed of 200k. The distance between
the nearby orders of the spectrum is $\ge20$ pixels and the two fibers (
$2\farcs4$ aperture one and $1\farcs6$ aperture one) can work at the
same time. The aperture of the fiber is 100 $\mu$m ($2\farcs4$) and the fiber
length is $\sim19$ m. The FOV for the guiding camera is $3'\times3'$ and in
the guiding plate the aperture of pupil is $4\farcs0$ in front of the fiber.
The temperature variation of the system is $r.m.s.=\pm0.05^{\circ}$C for a whole
night and $r.m.s.=\pm0.34^{\circ}$C for a week. The tip/tilt system is working
normally on the Cassegrain focus. The system is currently in its
commissioning phase, and its radial velocity precision dose not yet reach
its design goal of a few $\sim$cm/s. It is expected that this goal will be
achieved once commissioning is complete.
The guiding CCD camera is a GC1380, and the software $AVTUniCamViewer$ is
used for CCD controlling, exposure time configuration and data collection. The
telescope controlling software can monitor the guiding uncertainty of the
guiding system.
An astro-frequency comb calibration system, which is developed by Peking
University (PKU), also was installed and is being tested on the 2.16-m
telescope. The full spectral wavelength range of the astro-frequency comb is
160nm and the central wavelength is $640\pm20$ nm. The observed spacing of the
comb teeth is 29.01 GHz. Once working normally, it can greatly improve the
measuring accuracy of the stellar radial velocity. The system has significant
advantages over the $\rm I_2$ cell \citep{wi10}. It consists of a series of
discrete, equally spaced spectral lines with equal intensity and it is
repeatable in long time scale. Right now, the system has been set up and the
first spectrum simultaneously from the comb and the flat-fielding lamp has been
obtained during the engineering run. Meanwhile, a 25GHz AstroComb Optical
Frequency calibration system manufactured by the MenloSystems company, also
has been installed on the HRS instrument system recently.
The spectral coverage is $\sim$450-720 nm, with the central wavelength of
$540\pm30$ nm.
The flatness of the spectrum is r.m.s. $<5$ dB and the observed spacing of the
comb teeth is 25 GHz. The luminous power of the system for the full spectral
wavelength range is $>10\mu$W. The two new calibration systems are at the
commissioning stage. The two new systems are supposed to be applied in
the end of this year.
\begin{figure}
\centerline{
\includegraphics[scale=1.,angle=0]{fig14.ps}}
\caption[]{The optical layout and the light path of the fiber-fed HRS
spectrograph.}
\label{fig14}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[scale=0.7,angle=0]{fig15.ps}}
\caption[]{The quantum efficiency of the basic midband CCD camera used on the
HRS at -100$^{\circ}$C, which is shown by the solid line}
\label{fig15}
\end{figure}
\begin{table}[ht!!!]
\small
\centering
\begin{minipage}[]{100mm}
\caption[]{The Current Instrumental Parameters of the HRS.}\label{t5.tab}\end{minipage}
\tabcolsep 3mm
\begin{tabular}{ccc}
\hline\noalign{\smallskip}
Tech. Index & Parameters & Notes \\
\hline\noalign{\smallskip}
Wave. Cov. & 3650-10,000 {\AA} & ThAr lamp \\
\hline
Aper. of Fiber & $2\farcs4$ / $1\farcs6$ & \\
\hline
Spec. Res. & 32,000-106,000 & tested on Apr. 12, 2011 \\
\hline
Stab. of inst. & $\pm6~m/s$ & \\
\hline
Effi. & $34\%$ at peak value & \\
& $\ge10$ \% for $\lambda > 4500$ {\AA} & tested on Nov. 23, 2010 \\
\hline
& E2V 4k$\times$4k 12$\mu$m & \\
& Scientific chip, class 1 & \\
& On-chip Binning & \\
CCD Camera & 50-200 KHz & 2$\times$1 binning \\
& LN holding time: 20 hours & mode suggested \\
& (w/LN auto-filling) & \\
& LN cooling & \\
\hline
FOV of Guid. Cam. & $3'.3\times3'.3$ & \\
\hline
Temp. var. / day & $\pm0.05^{\circ}$ & \\
\hline
Temp. var. / week & $\pm0.34^{\circ}$ & \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
In order to estimate the total efficiency (including the atmospheric
extinctions and reflectivity of the primary/secondary mirrors, among others),
a number of standard stars were observed with HRS on June 30, 2015.
By using Equation~\ref{eq1}, the final result shows that it is
$>$2\% at 5500{\AA}.
In addition, the limiting magnitudes are estimated. Figure~\ref{fig16}
is a plot of the signal-to-noise ratio (SNR) at $\lambda$=5500 {\AA} as a
function of exposure time for the stars from $V=5$ mag to $V=9$
mag for the HRS system, based on the observing data. For a $V=9$ mag star,
the typical signal-to-noise ratio $SNR=100$ for an exposure time of
3000-3600 seconds. A few observing tests give the SNR in the typical observing
conditions of 2012, for a star of $V=8.85$ mag with an exposure time of
3600s, the signal-to-noise ratio is $S/N=80$ at wavelength of
$\lambda=6000$ {\AA}, and for a star of $V=7.83$ mag in 2400s exposure, the
signal-to-noise ratio is $SNR=150$ at wavelength of $\lambda$=6000 {\AA}
(provided by Xiaoling Yang and Yuqin Chen).
\begin{figure}
\centerline{
\includegraphics[scale=1.,angle=0]{fig16.eps}}
\caption[]{The signal-to-noise ratio at $\lambda$=5500 {\AA} as a function of
exposure time for stars from $V=5$ mag to $V=9$ mag by using the fiber-fed
HRS spectrograph.}
\label{fig16}
\end{figure}
\section{Scientific Projects}
\label{sci.sec}
Since 1989, when the 2.16-m telescope saw its first light, a great many
scientific research projects have been carried out
in various fields, including: the study of nearby galaxies (star
formation rates, gas and dust content), AGNs and their supermassive black
holes (SMBHs), quasars, stellar parameter determinations, exoplanets,
supernovae (SNe), gamma-ray bursts (GRBs), and follow-ups of tidal disruption
events. When LAMOST was built, the 2.16-m telescope has also
been used for the LAMOST follow-up and stellar library observations.
Some of the scientific highlights obtained with the 2.16-m telescope
are as follows. In 1993, \citet{wh94} observed the spectra of supernova 1993J
with the 2.16-m telescope and found the blue-shifted oxygen lines only four
months after the optical discovery of the supernova, which is different from
the case of 1987A. $\sim100$ G-type giant stars of
$V\sim6$ mag have been monitored with the Coud\'{e} HRS system \citep{zl01}.
In a joint planet-search program between China and Japan by using the Xinglong
2.16-m telescope and Okayama Astrophysical Observatory (OAO), a number of
substellar companions of intermediate-mass giant stars has been discovered
by \citet{liu08,liu09}. Later, a brown dwarf companion candidate was discovered
by \citet{wang12} and a long-period giant planet has been discovered by
\citet{wang14}, which is actually the first exoplanet discovered jointly
with the Subaru telescope, the OAO 1.88m telescope and the Xinglong 2.16-m
telescope.
In recent years, in order to improve the efficiency of the telescope and to
maximize the scientific output, 7-8 key projects with the telescope
have been supported for 3-5 years. Each project usually owns $\sim20-40$
nights per year, which allows to carry out long-term observations and
large-sample surveys.
For instance, a number of quasars at intermediate redshift $2.2 < z < 3$ have
been identified by \citet{wu11,wu13}. An H$\alpha$ imaging survey of
$\sim$1400 nearby ALFALFA galaxies is carried out to study the star formation
rate and the stellar distributions. From 1997 to 2002, the spectra of
$\sim100$ blue compact galaxies (BCGs), which was the largest sample before
SDSS, have been observed with the 2.16-m telescope, and the metallicity,
extinction of dust, star formation rate of the sample galaxies have been
derived by \citet{kc02,kong02,kong04,shi05,kong02}. A group of Peking
University has discovered a high redshift quasar ($z=5.06$), J2202+1509, in
November 2014 with the 2.16-m telescope, which is SDSS $i=18.79$ mag
(the paper is in preparation). The redshift
is derived from the emission lines, such as Ly$\alpha$ at wavelength of 1216
{\AA}, and Nv at wavelength of 1240 {\AA}. This demonstrates that the 2.16-m
telescope is able to be used to discover and study high redshift quasars.
\section{Observing Time Application}
\label{pro.sec}
Each year, the call for proposals begins around 20 October for the
period of one month. Astronomers who are willing to use the telescope can
submit proposals through the website http://astrocloud.china-vo.org before
the deadline. After that, the proposals will be collected and
reviewed by the Time Allocation Committee (TAC) of the 2.16m telescope.
The probability of observing nights obtained for observers highly depends on
the mark ranking. The ToO (Target-of-Opportunity) follow-up observations,
like transient, SNe and GRBs, is supported, which are allowed to be applied
and observed when the transient sources happen. The turn-around time
(i.e. the time interval between a ToO alert and the start of the first
observation) of the 2.16-m telescope depends on the system and configure
being used when it happens. It concludes the following parts: the read-out
time of last frame ($\sim$30-50 seconds); the time for changing the grisms and
adjusting slit
width and sometimes it even takes longer time when the grism is not on the
spectragraph ($\sim$5-10 minutes); the pointing time of the telescope
($\sim$1-5 minutes); the possible time for changing instrument and focusing
($\sim$5-10 minutes), etc. Therefore, it takes $\sim$10-30 minutes
in all for starting the ToO observations.
Since the first light of the 2.16-m telescope, there are astronomers from
France, Japan, Taiwan, Hongkong, etc., beyond China Mainland, applied and
used the telescope and publish papers. Until now, there are more than 150
SCI papers published with the data obtained with the 2.16-m telescope since
the first light. All these papers have been peer-reviewed and most of them
were published on the high-impact magazines in astrophysics, e.g., $ApJ$, $AJ$,
$MNRAS$ and $A\&A$, and the impact factors are around $\sim4-6$, including the
articles published in $Nature$. When the key projects are concluded, there
will be numerous high impact journals published in the future.
\section{Summary and Discussion}
\label{sum.sec}
As a 2-meter class optical astronomical telescope in China, the 2.16-m telescope
of Xinglong Observatory of NAOC, plays an important role in observational
astronomy today. There are currently three primary instruments, the BFOSC, OMR
and the fiber-fed HRS, available for the telescope users.
When the 2.16-m telescope saw the first light in 1989, various scientific
projects have carried out, based on observations of the telescope in the
research areas of nearby galaxies, AGN, supernovae, GRBs, exoplanets, Galactic
stars, time-domain astronomy, and many other sciences. Since the LAMOST spectroscopic
survey started, the 2.16-m telescope has also been used for the LAMOST
follow-ups, in order to identify interesting objects, and for stellar library
observations. A great many remarkable studies have been done with the telescope,
including the spectroscopic research of the supernova 1993J \citep{wh94}.
In recent years, in order to improve the utilizing efficiency
of the telescope and the scientific output further, 7-8 key observing
projects have been set up, in which the observers own 20-40 nights/year
to carry our long-term observations and serial research.
In the future, a number of potential new instruments will be available
to the 2.16-m telescope users to improve its observing abilities further, such
as an imaging photopolarimeter $RoboPol$ \citep[please see,][]{king14}, a
prototype of which is currently in use at the 1.3-m telescope of the Skinakas
Observatory in Crete, Greece, where it works well. A similar photo-polarimeter
will probably be installed at the 2.16-m telescope in the next one or two
years, which will be the first astronomical photo-polarimeter in China. It
could be used for polarization measurements, for instance of Seyfert galaxies,
blazars, and GRBs. Further, an intermediate resolution spectrograph (IRS) is
being investigated. If it is installed and committed in the future,
the projects of intermediate-resolution spectroscopic
observations (e.g., X-ray binaries, bright stars, nearby galaxies, and
a large number of fainter Galactic stars) could be carried out. In addition,
the testing of the astro-frequency comb calibration system is at the
commissioning stage.
When it is finished, the precision of the radial velocity measurements
of Galactic stars could be improved greatly, even to a few cm s$^{-1}$.
Recently a number of adaptive optics experiments were performed on the 2.16-m
telescope by engineers from various Chinese institutes, which can improve
the spatial resolution and signal-to-noise ratio of targets to a large extent.
The exoplanet detection technology group from the Nanjing Institute of
Astronomical Optics \& Technology, NAOC and the California State University,
Northridge collaborate and built the High Performance Portable Adaptive Optics
(HPAO) which is mounted on the Coud\'{e} focus of 2.16-m telescope and succeed
after the testing observations. Although the limiting magnitude of the current
HPAO system is faint to 3.8 mag in $H$ band, the system will be upgraded
in the next step and the results can be improved significantly.
\begin{acknowledgements}
This research was supported by the National Natural Science Foundation of
China (NFSC) through grants 11003021, 11373003, 11273027, 11303042 and
National Key Basic Research Program of China (973 Program) 2015CB857002.
Z.F. acknowledges a Young Researcher Grant of the National Astronomical
Observatories, Chinese Academy of Sciences. We thank Yuqin Chen, Liang Wang
and Xiaoying Yang for providing the limiting magnitude measurements of the
HRS instrument and thank Yingwei Chen for providing the picture of the
2.16-m telescope.
\end{acknowledgements}
|
\section*{Acknowledgments}
This work is supported by the NSAF Grant Nos.~U1330201 and U1530401, the NSFC Grant Nos.~91421102 and 60836001, and the MOST 973 Program Grant Nos.~2014CB848700, 2014CB921401 and 2011CBA00304. F.N. is partly supported by the RIKEN iTHES Project, MURI Center for Dynamic Magneto-Optics, the ImPACT Program of JST, and Grant-in-Aid for Scientific Research (S). J.S.T. is partially supported by the Japanese Cabinet Office's ImPACT project. We thank Yu-xi Liu, P.-M. Billangeon for valuable discussions, K. Kusuyama for technical assistance, and Anton F. Kockum for a critical reading of the manuscript.
{\it Note added}. Recently, we became aware of a work by Novikov {\it et al.}~\cite{Novikov}, which also studies EIT in a similar setup.
\begin{figure}
\begin{center}
\includegraphics[scale=0.9]{figure4.eps}
\caption{(color online) The transition spectroscopy between states $|0\rangle$ and $|2\rangle$ of the tunable 3D transmon at a flux bias giving $E_J/E_C=27.5$. (a) One-photon transition process between states $|0\rangle$ and $|1\rangle$, and two-photon transition process between states $|0\rangle$ and $|2\rangle$. (b) One-photon transition process between states $|0\rangle$ and $|2\rangle$.}\label{fig4}
\end{center}
\end{figure}
\section*{Appendix: One-photon transition between states $|0\rangle$ and $|2\rangle$}
In the experiment, we have also measured the one-photon transition spectroscopy between the eigenstates $|0\rangle$ and $|2\rangle$ of the 3D transmon at another flux bias giving $E_J/E_C=27.5$. Figure~\ref{fig4}(a) shows the transition spectroscopy for the one-photon process between the eigenstates $|0\rangle$ and $|1\rangle$ of the transmon as well as the two-photon process between the eigenstates $|0\rangle$ and $|2\rangle$. The corresponding transition frequencies are measured to be $\omega_{10}/2\pi=5.202$ GHz and $\frac{1}{2}\omega_{20}/2\pi=5.014$ GHz, respectively. For comparison, the transition spectroscopy for the one-photon process between states $|0\rangle$ and $|2\rangle$ is also shown in Fig.~\ref{fig4}(b), with the resonance exactly at the transition frequency $\omega_{20}/2\pi=10.028$~GHz. Note that these transition frequencies are still far detuned from the cavity frequency ($8.2169$ GHz). Moreover, as shown in Fig.~\ref{fig5}, we have measured the Rabi oscillations between states $|0\rangle$ and $|2\rangle$ by using a driving field with frequency $10.028$~GHz. It gives that the Rabi oscillation period is $56.8$~ns and the oscillation decay time is $130.6\mathrm{ns}$ by fitting with the experiment. These results reveal that one-photon transition between $|0\rangle$ and $|2\rangle$ indeed occurs in our 3D transmon system.
\begin{figure}
\includegraphics[scale=0.5]{figure5.eps}\label{fig5}
\caption{(color online) The measured Rabi oscillations between states $|0\rangle$ and $|2\rangle$ of the 3D trnasmon by using a $10.028 \mathrm{GHz}$ driving field. The black circles are experimental data and the red curve is an exponentially damped sinusoidal fit. The Rabi oscillation period is $56.8$~ns and the oscillation decay time is $130.6$~ns from the fit.}\label{fig5}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=1.1]{figure6.eps}
\caption{(color online) Electric- and magnetic-dipole transition matrix elements of the 3D transmon as a function of the ratio $E_J/E_C$. (a) and (b) correspond to the electric-dipole one-photon transitions for $n_g=0.5$ and $0.25$, respectively. (c) and (d) correspond to the magnetic-dipole one-photon transitions for $n_g=0.5$ and $0.25$, respectively. }\label{fig6}
\end{center}
\end{figure}
The Hamiltonian of the transmon can be written as
\begin{equation}
H=4E_c(n-n_g)^2-E_J\cos\varphi.
\label{tranH}
\end{equation}
The electric-dipole transition matrix element between eigenstates $|i\rangle$ and $|j\rangle$ of the transmon is proportional to $\langle i|n|j\rangle$, due to the field-induced time-dependent variation of the charge bias $n_g$. For the 3D transmon, the charge bias $n_g$ is not fixed but floated, so it can change for a different experimental setup. In Fig.~\ref{fig6}(a), we calculate $|\langle 0|n|1\rangle|$, $|\langle 0|n|2\rangle|$ and $|\langle 1|n|2\rangle|$ at $n_g=0.5$. These calculated transition matrix elements are similar to the results obtained in Ref.~\cite{Koch}. The result of $|\langle 0|n|2\rangle|=0$ indicates that no electric-dipole one-photon transition occurs between states $|0\rangle$ and $|2\rangle$. When $n_g\neq 0.5$ (e.g., $n_g=0.25$), $|\langle 0|n|2\rangle|$ can be nonzero, but it is much smaller than $|\langle 0|n|1\rangle|$, especially around $E_J/E_c=16.99$ (i.e., the ratio given in our experiment for the EIT) [see Fig.~\ref{fig6}(b)]. Therefore, the electric-dipole one-photon transition between $|0\rangle$ and $|2\rangle$ is still too weak in the case of $n_g\neq 0.5$.
For the 3D transmon used in our experiment, the single Josephson junction is replaced by a symmetric SQUID with an effective Josephson coupling
\begin{equation}
E_J=2E_{J0}\cos\left(\frac{\pi\Phi_x}{\Phi_0}\right),
\end{equation}
where $E_{J0}$ is the Josephson coupling energy of each junction in the SQUID.
In addition to the static magnetic flux $\Phi_x$, when a weak time-dependent magnetic flux $\Phi_a(t)$ is applied (i.e., $|\pi\Phi_a(t)/\Phi_0|\ll 1$), the Hamiltonian becomes
\begin{equation}
H(t)=H+I\Phi_a(t),
\end{equation}
where
\begin{equation}
I=\frac{2\pi E_{J0}}{\Phi_0}\sin\left(\frac{\pi\Phi_x}{\Phi_0}\right)\cos\varphi
\end{equation}
is the circulating current in the SQUID loop~\cite{You2001}. The magnetic-dipole transition matrix element between eigenstates $|i\rangle$ and $|j\rangle$ of the transmon is proportional to $\langle i|\cos\varphi|j\rangle$. In Figs.~\ref{fig6}(c) and (d), we show the calculated $|\langle 0|\cos\varphi|1\rangle|$, $|\langle 0|\cos\varphi|2\rangle|$ and $|\langle 1|\cos\varphi|2\rangle|$. At $n_g=0.5$, while both $|\langle 0|\cos\varphi|1\rangle|=0$ and $|\langle 1|\cos\varphi|2\rangle|=0$, $|\langle 0|\cos\varphi|2\rangle|$ is nonzero [see Fig.~\ref{fig6}(c)], indicating that only the magnetic-dipole one-photon transition between eigenstates $|0\rangle$ and $|2\rangle$ is allowed in this case. When $n_g\neq 0.5$ (e.g., $n_g=0.25$), both $|\langle 0|\cos\varphi|1\rangle|$ and $|\langle 1|\cos\varphi|2\rangle|$ become nonzero, but they are smaller than $|\langle 0|\cos\varphi|2\rangle|$. In particular, they are much smaller than $|\langle 0|\cos\varphi|2\rangle|$ at $E_J/E_c=16.99$ [see Fig.~\ref{fig6}(d)].
Because $E_J\gg E_c$ in the transmon, when the field-induced time-dependent variation of the charge bias $n_g$ is small, the magnetic-dipole one-photon transition between $|0\rangle$ and $|2\rangle$ can become as important as the electric-dipole one-photon transitions $|0\rangle\rightarrow|1\rangle$ and $|1\rangle\rightarrow|2\rangle$. This explains the observation of the one-photon transition between states $|0\rangle$ and $|2\rangle$ in our tunable 3D transmon.
|
\section{INTRODUCTION}
The exciting and ingenious idea of holographic dark energy that has recently attracted many researchers is capable to interpret current cosmic acceleration. \cite{Li}-\cite{Farajollahi}. The idea initiated from the cosmological application of
the more fundamental holographic principle and despite some objections it reveals the dynamical nature of the vacuum energy by relating it to
cosmological volumes. The holographic principle states that due to the limit set by the formation of a black hole, in effective field theory, the UV Cut-off, $\Lambda_{uv}$, is related to the IR Cut-off $L$ as $L^3\Lambda_{uv}^4\leq L M_p^2$ where $M_p$ is reduced Planck mass
. The effective field theory describes all states of system except those already collapsed to a black hole and the
vacuum energy density via quantum fluctuation is given by $\rho_{vac}\sim \Lambda_{uv}^4\sim M_p^2 L^{-2}$ where $L$ is characteristic length scale of the universe
. From vacuum energy density the dark energy density caused via quantum fluctuation is given by $\rho_{de}\sim 3d^2 M_p^2 L^{-2}$ where $d$ is a model parameter. By taking different characteristic length scale we can construct various holographic dark energy models.
On the other hand, a large amount of current research heads towards higher dimensional gravity and in particular brane cosmology \cite{Antoniadis},\cite{Randall2},
in which assumes our Universe as a brane embedded in higher dimensional spacetime. A well-known example of brane cosmological model is the Dvali-Gabadadze-Porrati (DGP) braneworld model \cite{Dvali}, in which the 4D world is a FRW brane
embedded in a 5D Minkowski bulk. On the 4D brane the gravity action is proportional to $M_P^2$ whereas in the bulk it is proportional to the corresponding
quantity in 5D, $M_5^3$. The total energy-momentum is confined on a
3D brane embedded in a 5D infinite volume
Minkowski bulk. There are two different ways to confine the 4D brane into the 5D spacetime;
the DGP model has two separate branches denoted by $\epsilon=\pm 1$
with distinct features. The $\epsilon$ = +1 branch is capable to interpret the current cosmic acceleration without any need to introduce dark
energy, whereas for the $\epsilon$ = -1 branch, dark energy is needed \cite{Deffayet},\cite{Chimento}.
Holographic dark energy in the context of DGP brane cosmological models have been investigated in \cite{Wu2} and \cite{Setare3}. But, in both of them a 4D holographic dark energy model has been used. Recently, the extended 5D holographic dark energy model has been considered in \cite{Kim}, but it is grounded on unstable and non-physical arguments. For more studies about this topic see \cite{Saridakis} and \cite{Saridakis2}.
The main motivation of 5D holographic dark energy is that in a 5D spacetime, the natural framework for the cosmological application in connection with dark energy of holographic principle is the space of the bulk and not the 4D brane, because the attributes of gravitational theory or quantum field, such as cut-off's and vacuum energy can be determined by the maximal uncompactified space of the model. For example, in braneworld models, black holes are generally D-dimensional \cite{Kanti},\cite{Cavaglia}, irrespective of their 4D effective effects. Consequently, although the holographic principle is applicable to any dimensions \cite{Bousso},\cite{Iwashita}, its cosmological application on the subject of dark energy must be considered in the maximally-dimensional subspace, i.e. in the bulk. Afterwards, this 5D holographic dark energy results in an effective 4D dark energy component in the Friedmann equation on the brane with inherent
holographic nature.
In this work we present the 5D holographic dark energy in DGP brane cosmology. We will see that under this discussion we can generate an effective 4D holographic dark energy from the Hubble horizon and so some of the holographic dark energy model problems, related to future event horizon as the length scale, like causality problem and circular logic problem are removed, automatically. A simple and not very similar work at this topic has been done in a different way in \cite{Liu}. Here, we will constrain the cosmological parameters in the model, both analytically and also using observational data.
Also, in \cite{Bamba}, the authors have shown the equivalent description of different
theoretical models of dark energy. They have examined the fluid and scalar dark energy descriptions of various models such as holographic dark energy. We should notice that our model follows this organism, as well and we can see this from the numerical results. But at a special case we will show the relation between our model and the $\Lambda$CDM model, when the dark energy dominated regime is considered.
\section{HOLOGRAPHIC DARK ENERGY IN THE BULK}
In this section we review the basic results of a 5D bulk holographic dark energy. For a 5D spherically symmetric and uncharged black hole, its mass $M_{BH}$ is related to its Schwarzschild radius $r_s$ via $M_{BH} = 3\pi M_5^3r_s^{2}/8$ where $M_5$ is the 5D Planck mass and is related to the 5D gravitational
constant $G_5$ with $M_5 = G_5^{-\frac{1}{3}}$. Moreover, the 4-dimensional Planck mass $M_P$ given by $M_P^2 = M_5^{3}V_{1} $ where $V_{1}$ is the volume of the extra-dimensional space. For the holographic dark energy in the 5D bulk we have $\rho_{\Lambda 5}V(S^{3})\leq M_{BH}$ where $\rho_{\Lambda 5}$ is the 5D bulk vacuum energy, and $V(S^{3})$ is the volume of the maximal hypersphere in a 5D spacetime,
given as $V(S^{3}) = \pi^2r^{4}/2$. The dark energy density, $\rho_{\Lambda 5}$, can be viewed as holographic dark energy,
\begin{equation}\label{HDE}
\rho_{\Lambda 5} = c^2M_{BH}V^{-1}(S^{3})=\frac{3c^2M_5^{3}}{4\pi}L^{-2},
\end{equation}
with $L$ as a large distance and $c^2<1$. Similar to 4-dimensional case, the distance $L$ can be Hubble radius, particle horizon, or the most appropriate future event horizon, for a flat Universe.
Next, we study the holographic dark energy in 5D DGP-brane cosmology with the action
of the form
\begin{eqnarray}\label{action}
S &=& \frac{1}{16\pi}M_5^3\int_{{\cal M}_5}d^5x\sqrt{-^5g}(^5R+{\cal L}_m) \nonumber\\ &+& \frac{1}{16\pi}M_P^2\int_{\sum}d^4x\sqrt{-^4g}(^4R+{\cal K}+{\cal L}_m),
\end{eqnarray}
where the ${\cal M}_5$ and $\sum$ indicate respectively bulk (B) and brane (b) and ${\cal K}$ is the trace of extrinsic curvature. The extra term in the
boundary introduces a a cross-over length scale $r_c = \frac{M_P^2}{2M_5^3}$ which separates two different
regimes of the theory. For distances much smaller than $r_c$ one would expect the solutions to be well approximated
by general relativity and at larger distances the modifications takes into account. In FRW cosmology, the Friedmann equation on the brane is \cite{Phong}
\begin{equation}\label{fried}
H^2 = \frac{8\pi\rho_b}{3M_P^2}-\frac{k}{a^2}+\frac{\epsilon}{r_c}\sqrt{H^2-\frac{4\pi\rho_B}{3M_5^3}
+\frac{k}{a^2}+\frac{\xi}{a^4}}.
\end{equation}
We assume a flat universe, $k = 0$, and a vanishing last term $\xi = 0$. If also consider the matter density in the bulk a holographic dark energy, $\rho_B = \rho_{\Lambda5}$ and the matter on the brane as a cold dark matter $\rho_b = \rho_m$, then we have
\begin{equation}\label{friedmann}
H^2 = \frac{8\pi\rho_m}{3M_P^2}+\frac{\epsilon}{r_c}\sqrt{H^2-\frac{4\pi\rho_{\Lambda5}}{3M_5^3}}.
\end{equation}
Alternatively, we may rewrite the Friedmann equation in the conventional form
\begin{equation}\label{traditional}
H^2 = \frac{8\pi\rho_m}{3M_P^2} + \frac{8\pi\rho_\Lambda}{3M_P^2}
\end{equation}
where the 4D dark energy is
\begin{equation}\label{4DDE}
\rho_\Lambda\equiv\rho_{\Lambda4}=\frac{3M_P^2\epsilon}{8\pi r_c}\sqrt{H^2-\frac{4\pi \rho_{\Lambda5}}{3M_5^3}}.
\end{equation}
From (\ref{HDE}) we find the 5D bulk holographic dark energy, $\rho_{\Lambda5}$, and finally arrive at the following form for
the effective 4D holographic dark energy
\begin{equation}\label{effective4DDE}
\rho_\Lambda = \frac{3M_P^2\epsilon}{8\pi r_c}\sqrt{H^2-c^2L^{-2}}.
\end{equation}
In a flat 4D spacetime, the future event horizon is taken as the the most appropriate cut-off scale that fits holographic statistical physics \cite{Li}. Similarly, in \cite{Saridakis} the author shows that in the 5D extension of 4D, still future Event horizon is preferable. In this work, we show that in case of a 5D DGP holographic model of dark energy the Hubble radius can be taken as a cut-off scale that suitably fit the observational data. With the Hubble radius as the horizon, i.e. $L=H^{-1}$ and by using (\ref{traditional}), the dark energy and matter densities then become
\begin{equation}\label{4DDEapp}
\rho_\Lambda = \frac{3M_P^2\epsilon H}{8\pi r_c}\sqrt{1-c^2}
\end{equation}
\begin{equation}\label{traditionalapp}
\rho_m = \frac{3M_P^2}{8\pi}(H^2 - \frac{\epsilon H}{r_c}\sqrt{1-c^2})
\end{equation}
From (\ref{4DDEapp}), one immediately finds two new constraints on $c$ and $\epsilon$, i.e. $0\leq c\leq1$ and for positive $\rho_\Lambda$, $\epsilon$ must be $1$. The second constraint can also be obtained from a thermodynamic method. In terms of density parameter, $\Omega_\Lambda$, equation (\ref{4DDEapp}) can be rewritten as
\begin{equation}\label{Homega}
H\Omega_\Lambda = \frac{\epsilon\sqrt{1-c^2}}{r_c}.
\end{equation}
By differentiating of the above equation with respect to $t$ we yield
\begin{equation}\label{Hdot}
\dot H = \frac{-\epsilon\sqrt{1-c^2}\dot{\Omega_\Lambda}}{r_c\Omega_\Lambda^2}.
\end{equation}
On the other hand, the Gibbons-Hawking entropy is proportional to the squared radius of the universe that is taken as the Hubble horizon, i.e.
\begin{equation}\label{GH}
S \sim A \sim L^2 \sim H^{-2}.
\end{equation}
Since entropy should increase during expansion phase, we obtain $\dot H \leq0$. From (\ref{Hdot}) and by assuming that $\Omega_\Lambda$ increases with time we again find $\epsilon = +1$. We should notice that the two constraints on $c$ and $\epsilon$ are the characteristics for this model only if Hubble radius is taken as the horizon. By using $\Omega_\Lambda = 1-\Omega_m = 1-\Omega_{m0}H_0^2H^{-2}(1+z)^{3}$ and $\Omega_{r_c} = 1/(4r_c^2H_0^2)$ in (\ref{Homega}), the Friedmann equation becomes
\begin{equation}\label{BFequation}
E^2 =2\sqrt{\Omega_{r_c}(1-c^2)}E+\Omega_{m0}(1+z)^{3}.
\end{equation}
where we have used the definition $E^2 = H^2/H_0^2$. This equation at $z=0$ yields the constraint
\begin{equation}\label{Friedcons}
1-\Omega_{m0}=2\sqrt{\Omega_{r_c}(1-c^2)}.
\end{equation}
In the next section we are going to constraint the model parameters $c$, $\Omega_{m0}$ and $\Omega_{r_c}$ with the observational data by employing $\chi^2$ method. However, with attention to constraint (\ref{Friedcons}) we implement this in two different ways.
{\bf A:}
Using (\ref{Friedcons}) we can neglect $c$ and $\Omega_{r_c}$ in favor of $\Omega_{m0}$ in (\ref{BFequation}) and obtain
\begin{equation}\label{1stFried}
E^2 =(1-\Omega_{m0})E+\Omega_{m0}(1+z)^{3}.
\end{equation}
After best-fitting the value of $\Omega_{m0}$, with attention to constraints on $c$, (i.e., $0\leq c\leq1$) it seems that we can obtain some constraints for $\Omega_{r_c}$. After rewriting (\ref{Friedcons}) as
\begin{equation}\label{rccons}
\Omega_{r_c}=\frac{(1-\Omega_{m0})^2}{4(1-c^2)}
\end{equation}
we see that for $c=0$ where shows no holographic dark energy we can find for $\Omega_{r_c}$ a lower limit as $\Omega_{r_c}=(1-\Omega_{m0})^2/4$. Also, for $c=1$ where in our model stands for a universe filled with matter (see (\ref{4DDEapp}) and (\ref{traditionalapp})) we find $\Omega_{r_c}\rightarrow\infty$.
{\bf B}:
From (\ref{Friedcons}) we obtain $\Omega_{m0}$ in terms of $c$ and $\Omega_{r_c}$ and after replacing in (\ref{BFequation}) we reach to
\begin{equation}\label{2ndFried}
E^2 =2\sqrt{\Omega_{r_c}(1-c^2)}E+(1-2\sqrt{\Omega_{r_c}(1-c^2)})(1+z)^{3}.
\end{equation}
Then, we can best-fit our model parameters $c$ and $\Omega_{r_c}$ in such a way that the condition $\Omega_{m0}\geq0$ is satisfied. It is clear from (\ref{Friedcons}) that for any permitted values of $c$ and $\Omega_{r_c}$ in our model, $\Omega_{m0}$ can not be larger than one.
\subsection{EQUIVALENCE WITH $\Lambda$CDM MODEL}
We know that various dark energy models have been proposed to explain late time acceleration of the universe. Some of them are due to modifications in the matter content of the universe. For example, models with some scalar fields such as quintessence, phantom and tachyon and also models with some kinds of fluid with special equation of states such as Chaplygin gas. On the other hand another set of dark energy models is due to modifications in geometry of spacetime such as $f(R)$ and $f(T)$ gravity models. In a recent work \cite{Bamba}, the authors have shown that any dark energy model may be expressed as the
others.
In this work, we have proposed another dark energy model, thus there should be equivalent descriptions of it in any other models of dark energy. In an attempt to show this equivalence, we assume that we are in a dark energy dominated universe. So, we can neglect the first term at the right hand side of (\ref{traditional}) and then replace $\rho_\Lambda$ by (\ref{4DDEapp}). Thus we reach to
\begin{equation}\label{equivalent}
H = \frac{\epsilon}{r_c}\sqrt{1-c^2}.
\end{equation}
We see that at late time, the Hubble parameter in our model reduce to a constant value. So, in dark energy dominated regime our model is parallel to the $\Lambda$CDM model, the simplest dark energy model which can be interpreted both as a modification in geometry of space-time or in matter content of the universe. See, figure (\ref{fig:3}).
\section{OBSERVATIONAL CONSTRAINTS AND COSMOLOGICAL TEST}
There are a variety of observational data to fit the parameters in cosmological models. Here, we use
Sne Ia which consists of 557 data points belonging to the Union sample \cite{Amanullah}, the baryonic acoustic oscillation (BAO) distance ratio and the cosmic microwave background (CMB) radiation. The constraints from a combination of SNe Ia, BAO and CMB can be obtained by minimizing $\chi^2_{SNe}+\chi^2_{BAO}+\chi^2_{CMB}$. We should notice that to best-fit by SNe Ia data, following \cite{Wu} we have performed a marginalization on the present value of a cosmological parameter called distance moduli $\mu_0$.
For the model, using the first approach we obtain $\Omega_{m0}=0.246^{+0.014 +0.030 +0.047}_{-0.013 -0.027 -0.040}$ for 1$\sigma$, 2$\sigma$ and 3$\sigma$, respectively with $\chi^2_{min}=586.316$. Using (\ref{rccons}) the lower limit for $\Omega_{r_c}$ is obtained as $\Omega_{r_c}\geq0.142$. In figure \ref{fig:1}, left panel we have shown the evolution of $\chi^2$ in terms of $\Omega_{m0}$. The horizontal black dash lines indicate 1$\sigma$, 2$\sigma$ and 3$\sigma$ confidence intervals of $\Omega_{m0}$ where have calculated by adding 1, 4 and 9 to the minimum value of $\chi^2$, respectively. Also, in the right panel we have shown the probability with respect to $\Omega_{m0}$. To this aim we have used the popular definition of probability as P $=\exp(\frac{\chi^2_{min}-\chi^2}{2})$.
\begin{figure}[h]
\centering
\includegraphics[width=0.48\textwidth]{chi.eps}
\includegraphics[width=0.48\textwidth]{likelihood.eps}
\caption{left: The evolution of $\chi^2$ with respect to $\Omega_{m0}$. right: The probability distribution with respect to $\Omega_{m0}$.}\label{fig:1}
\end{figure}
The results have been shown in table \ref{table:1}. By $dof$ we mean degrees of freedom, where equals the number of observational data minus the number of parameters we are fitting to data. Note that at the first approach we just have one parameter.
\begin{table}[h]
\caption{bestfit values of the first approach}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline\hline
observational data & SNe & SNe + BAO & SNe + CMB & SNe + BAO + CMB \\
\hline\hline
$\Omega_{m0}$ & 0.175 & 0.168 & 0.246 & 0.246 \\
\hline
$\chi^2_{min}$ & 543.270 & 544.456 & 584.383 & 586.316 \\
\hline
$\chi^2_{min}/dof$ & 0.977 & 0.976 & 1.049 & 1.049 \\
\hline
\end{tabular}
\label{table:1}
\end{table}\
At the second approach we obtain the best-fit values $c=0.318$ and $\Omega_{r_c}=0.158$ with $\chi^2_{min}=586.287$. Then, using (\ref{Friedcons}) we can calculate $\Omega_{m0}=0.246$ where is exactly the same as the result of the first approach. The results have been shown in table \ref{table:2}.
\begin{table}[h]
\caption{bestfit values of the second approach}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline\hline
observational data & SNe & SNe + BAO & SNe + CMB & SNe + BAO + CMB \\
\hline\hline
c & 0.884 & 0.382 & 0.786 & 0.318 \\
\hline
$\Omega_{r_c}$ & 0.778 & 0.202 & 0.370 & 0.158 \\
\hline
$\chi^2_{min}$ & 543.110 & 544.300 & 584.346 & 586.287 \\
\hline
$\chi^2_{min}/dof$ & 0.979 & 0.977 & 1.051 & 1.051 \\
\hline
\end{tabular}
\label{table:2}
\end{table}\
Fig. (\ref{fig:2}) shows the confidence contours of the parameters $\Omega_{r_c}$ and $c$ for combination of SNe Ia, BAO and CMB. The black curve and the gray area show $\Omega_{m0}=0$ and $\Omega_{m0}>0$, respectively.
\begin{figure}[h]
\centering
\includegraphics[width=0.48\textwidth]{confidence.eps}
\caption{The contour plot for 1$\sigma$(red), 2$\sigma$(green) and 3$\sigma$(blue) confidence regions using best-fitted parameters by SNe Ia, BAO and CMB combination. The black circle indicates the best fit point. We have removed the $\Omega_{m0}<0$ region (white area) by the black curve that shows $\Omega_{m0}=0$.}\label{fig:2}
\end{figure}
Using the best-fitted constrained model parameters, figure (\ref{fig:3}), left panel, shows the evolutionary curve of the holographic dark energy EoS parameter for the best fitted values of our model. One can see that the parameter becomes tangent to $-1$ in the future. If we assume that the holographic dark energy is conserved on the brane as
\begin{equation}\label{conservation}
\dot\rho_\Lambda + 3H(1+w_\Lambda)\rho_\Lambda = 0,
\end{equation}
then using (\ref{4DDEapp}) and the definition of $E$ we obtain
\begin{equation}\label{eos}
w_\Lambda = -1+\frac{(1+z)}{3E}\frac{dE}{dz}\cdot
\end{equation}
Also, we can express the total EoS parameter generally as
\begin{equation}\label{eostot}
w_{tot} = -1+\frac{2(1+z)}{3E}\frac{dE}{dz}\cdot
\end{equation}
In figure (\ref{fig:3}), the evolutionary curve of DE and total EoS parameter for the best fitted values has been shown.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{2w.eps}
\caption{The evolutionary curve of DE and total EoS parameter with respect to redshift.}\label{fig:3}
\end{figure}
At the end of this section we have a comparison with \cite{Wu2}. As we mentioned in introduction, at this work the authors have used a 4D holographic dark energy model in the DGP brane cosmology. They have investigated all kinds of length scale, i.e. Hubble horizon, particle and future event horizon. They have shown that if the Hubble horizon is considered as the IR cut-off, in $\epsilon=+1$ branch, the best fit values from SNe Ia+BAO data, prefer a pure DGP model with negligible vacuum energy. But in our article, we have shown that in a 5D holographic dark energy in DGP brane cosmology, choosing Hubble horizon as the length scale causes an effective 4D holographic dark energy which in spite of describing late time acceleration of the universe, removes important problems of an ordinary 4D holographic dark energy model, such as causality and circular logic problems.
\section{SUMMARY AND REMARKS}
We proposed a 5D holographic dark energy scenario in DGP-Brane cosmology. We fitted the model parameters in two different ways by using Sne Ia, Sne Ia+BAO and Sne Ia+CMB and Sne Ia+BAO+CMB dataset. It is noteworthy that the model fits different combination of dataset equally well with the $\chi^2_{min}\rightarrow 1$. However, the combination of Sne Ia+BAO+CMB dataset constrain $\Omega_{m0}$ slightly better.
The fitting model strongly suggest that the accelerated expansion is caused by the 5D holographic dark energy which transferred to an effective 4D dark energy. Without CDM in the 4D matter, i.e., considering a dark energy dominated Universe we found that our model is equivalent with the $\Lambda$CDM model. Also, here we considered Hubble radius as the length scale where caused some special constraints on our model parameters, $c$ and $\epsilon$.
|
\section{Introduction}
Ergodicity plays a fundamental role in the formulation of statistical
physics. This property is usually stated by saying that ensemble average and
time average of observables are equals, the last one being taken in the long
time (infinite) limit. In contrast with thermodynamical systems, where the
lack or ergodicity is induced by a spontaneous symmetry breaking \cit
{goldenfeld}, the disparity between ensemble and time averages may also be
found as an emergent property of complex systems. Named as weak ergodicity
breaking (EB) \cite{bouchad}, this feature is induced by the power-law
nature of the statistical distributions associated to the observables and
their dynamics \cite{bouchad,lutz}.
Time averages in presence of weak EB remain random even in the long time
limit. Their statistics, termed as weakly non-ergodic statistical physics
\cite{reben,barkai}, define a still very active line of research. Continuous
time random walk characterized by divergent trapping times is a natural
frame where weak EB was studied \cite{reben,barkai,bel,saa,radons,dentz}. In
addition, diverse kinds of complex anomalous diffusion processes are a
natural partner of weak EB. Analysis were performed for particles embedded
in heterogenous media \cite{cherstvy}, periodic potentials \cite{widera},
and in homogeneous disordered media \cite{garcia}. Geometric \cite{peters},
escaled \cite{safdari}\ and ultraslow \cite{bodrova} Brownian motions, as
well as diffusion induced by the combined action of different driven noises
\cite{godec,gbel,igor}, convoluted memory processes \cite{fuli}\ and
Langevin dynamics \cite{kessler} also were characterized from a similar
perspective.
In addition to its theoretical interest, weak EB was also found in different
physical systems such as deterministic dynamics \cit
{golan,albers,filho,akimoto} and blinking nanocrystals \cit
{nanocrystal,margolin}; also in molecular transport \cite{manzo} and
tracking of biological single molecules \cite{burovS,burov,jeon,unkel,simon
\ such as lipid granules \cite{unkel}, and diffusion in the plasma membrane
of living cells \cite{simon}. Weak EB also arises in complex networks \cit
{geneston,west}, fluid turbulence \cite{turbulence} and brain dynamics \cit
{ross}.
Weak EB can be studied in systems that have associated a stationary state,
such as for example random walks on finite domains, and also in
non-stationary systems such as unbounded diffusive ones (see for example
Refs. \cite{reben} and \cite{burovS} respectively). Independently of the
dimensionality, weak EB is in general associated or related to some
underlying self-similiar (effective) mechanism characterized by power-law
distributions. The main goal of this paper is to demonstrate that systems
whose dynamics involves global memory effects may also develop EB.
Furthermore, we establishes that the lack of ergodicity may happens even in
absence of statistical properties (residence times) characterized by
dominant power-law distributions.
Global memory (or correlation) effects refer to systems whose stochastic
dynamics at a given time depends on its whole previous temporal history
(trajectory). These kinds of dynamics has been studied previously \cit
{gunter,gandi,kenkre,katja,boyer,kursten,kim,esguerra,hanel}, mainly as a
mechanism that induces superdiffusion. In contrast, here we study random
walk processes defined over a finite set of states where the persistence in
a given state or the transition to another one depends on the previous
system trajectory. The random times between consecutive steps is defined by
a set of waiting time distributions with finite average times. In addition,
our main results rely on alternative memory mechanisms. They are related to
a P\'{o}lya urn dynamics \cite{feller,norman,pitman,queen,budini}, which is
one of the simplest models of contagion process, being of interest in
various disciplines \cite{norman}. In contrast to other global correlation
mechanisms, the urn-like dynamics is able to induce weak EB. Interestingly,
the departure from ergodicity arises even when the (average) residence times
in each state are finite.
The paper is organized as follows. In Sec. II, we introduce the globally
correlated random walk model. The probability density of time-averaged
observables is obtained in general. In Sec. III, we study three different
global memory mechanisms: the elephant random walk model, a random walk
driven by an urn-like dynamics, and an imperfect case of the last one. In
Sec. IV, for all models, we obtain the probability density of the residence
times. Sec. V is devoted to the Conclusions. Analytical calculations that
support the main results are presented in the Appendixes.
\section{Finite random walk with global memory effects}
In this section we introduce the globally correlated random walk model and
study its properties. The probability density of time-averaged observables
is also obtained.
\subsection{Model}
The system is characterized by a finite set of states $\mu =1,\cdots L.$ To
each state $\mu $ we assign a waiting time distribution $w_{\mu }(t),$ which
gives the statistics of times between consecutive steps of the stochastic
dynamics. We assume that all average time
\begin{equation}
\tau _{\mu }\equiv \int_{0}^{\infty }dtw_{\mu }(t)t, \label{TauMedio}
\end{equation
are finite, $\tau _{\mu }<\infty .$
The stochastic dynamics is as follows. At the beginning (initial time), each
state is selected in agreement with a set of probabilities $\{p_{\mu
}\}_{\mu =1}^{L},$ $0\leq p_{\mu }\leq 1,$ normalized as $\sum_{\mu
=1}^{L}p_{\mu }=1.$ Given that a state $\mu $ is selected, the system
remains in it during a random time selected in agreement with the waiting
time distribution $w_{\mu }(t).$ After this step, the system may remain in
the same state or jump to another one. Hence, it may \textit{persists} in
the same state, remaining an extra time interval chosen in agreement with
the same waiting time distribution, or \textit{jump} to a different state
with a different waiting time distribution. This dynamic repeats itself in
time after each step, where step refers to the process of selecting the next
state.
The state corresponding to the next step is chosen in agreement with a
conditional probability $\mathcal{T}_{n}(\{n_{1},n_{2}\cdots n_{L}\}|\mu )$
[denoted as $\mathcal{T}_{n}(\{n_{\nu }\}|\mu )].$ Here, $n$ indicates the
number of steps performed up to the present time, while $n_{\nu }$ gives the
number of times that each state $\nu $ was chosen previously. Then,
n=\sum_{\nu =1}^{L}n_{\nu }.$\ The dependence of the process on the whole
previous trajectory (global correlation) is given by the dependence of
\mathcal{T}_{n}(\{n_{\nu }\}|\mu )$ on the set $\{n_{\nu }\}_{\nu =1}^{L}.$
The previous definitions completely characterize the stochastic dynamics in
terms of the initial probabilities $\{p_{\mu }\}_{\mu =1}^{L},$ the waiting
time distributions $\{w_{\mu }(t)\}_{\mu =1}^{L}$ and the conditional (or
transition) probabilities $\mathcal{T}_{n}(\{n_{\nu }\}|\mu ).$
For the studied models [see Eqs. (\ref{BAmodel}), (\ref{PolyaModel}), and
\ref{MixedModel})], as a consequence of the memory effects, the following
property is observed. In the long time limit $(t\rightarrow \infty ),$ which
also correspond to a divergent number of steps $(n\rightarrow \infty ),$ the
fraction
\begin{equation}
f_{\mu }=\lim_{n\rightarrow \infty }\frac{n_{\mu }}{n}, \label{fracciones}
\end{equation
$\sum_{\mu =1}^{L}f_{\mu }=1,$ may become random variables whose values
depend on each particular realization. Their probability density is denoted
by $\mathcal{P}(\{f_{\mu }\}),$ which satisfies the normalization condition
\int_{\Lambda }df_{1}\cdots df_{L-1}\mathcal{P}(\{f_{\mu }\})=1.$ Here,
\Lambda $ is the region defined by the condition $\sum_{\nu =1}^{L}f_{\mu
}=1.$ The average of $f_{\mu }$ over an ensemble realizations, denoted by
\langle \cdots \rangle ,$ i
\begin{equation}
\langle f_{\mu }\rangle =\int_{\Lambda }df_{1}\cdots df_{L-1}\ f_{\mu
\mathcal{P}(\{f_{\nu }\}). \label{qu}
\end{equation}
At a given time $t,$ with $P_{\mu }(t)$ we denote the (ensemble) probability
$[\sum_{\mu =1}^{L}P_{\mu }(t)=1]$ that the system is in the (arbitrary)
state $\mu .$ This object is characterized in Appendix \ref{stationary} from
the dynamics defined previously. The stationary probability reads $P_{\mu }^
\mathrm{st}}\equiv \lim_{t\rightarrow \infty }P_{\mu }(t).$\ It can be
written in terms of $\mathcal{P}(\{f_{\nu }\})$ a
\begin{equation}
P_{\mu }^{\mathrm{st}}=\left\langle \frac{f_{\mu }\tau _{\mu }}{\sum_{\mu
^{\prime }=1}^{L}f_{\mu ^{\prime }}\tau _{\mu ^{\prime }}}\right\rangle ,
\label{PEqui}
\end{equation
where $\tau _{\mu }$ is defined by Eq. (\ref{TauMedio}). In Appendix \re
{stationary} we also derive this result. Basically it say us that in each
realization the system reaches a (random) stationary state defined by the
weights $(f_{\mu }\tau _{\mu })/\sum_{\mu ^{\prime }=1}^{L}f_{\mu ^{\prime
}}\tau _{\mu ^{\prime }}.$ In consequence, $P_{\mu }^{\mathrm{st}}$ depends
on which memory mechanism drives the stochastic dynamics.
\subsection{Time-averaged observables}
To each state $\mu ,$ we assign an observable with value $\mathcal{O}_{\mu
}. $ Hence, each realization of the random walk defines a corresponding
trajectory $\mathcal{O}(t).$ In the stationary regime, its \textit{ensemble
average} $\langle \mathcal{O}\rangle _{\mathrm{st}}\equiv \lim_{t\rightarrow
\infty }\langle \mathcal{O}(t)\rangle =\lim_{t\rightarrow \infty
}\sum\nolimits_{\mu =1}^{L}P_{\mu }(t)\mathcal{O}_{\mu },$ i
\begin{equation}
\langle \mathcal{O}\rangle _{\mathrm{st}}=\sum\nolimits_{\mu =1}^{L}P_{\mu
}^{\mathrm{st}}\mathcal{O}_{\mu }, \label{OEstacion}
\end{equation
where the weights follows from Eq. (\ref{PEqui}). On the other hand, its
\textit{time average} is defined as $\mathcal{O}\equiv \lim_{t\rightarrow
\infty }(1/t)\int_{0}^{t}dt^{\prime }\mathcal{O}(t^{\prime }),$ which leads
t
\begin{equation}
\mathcal{O}=\lim_{t\rightarrow \infty }\sum_{\mu =1}^{L}\Big{(}\frac{t_{u}}{
}\Big{)}\mathcal{O}_{\mu }. \label{Observable}
\end{equation
Here, $t_{u}$\ is the total residence time in the state $\mu $ in the
interval $(0,t).$ Hence, $\sum\nolimits_{\mu =1}^{L}t_{u}=t.$
Even when a long time limit is present in the previous definition, the
observable $\mathcal{O}$ may be a random object that depends on each
particular realization. Its probability density can be written a
\begin{equation}
P(\mathcal{O})=\lim_{t\rightarrow \infty }\left\langle \delta \Big{(
\mathcal{O}-\sum\nolimits_{\mu =1}^{L}\frac{t_{u}}{t}\mathcal{O}_{\mu
\Big{)}\right\rangle , \label{P(O)}
\end{equation
where, as before, $\left\langle \cdots \right\rangle $ denotes average over
an ensemble of realizations and $\delta (x)$ is the Dirac delta function.
Now, our goal is to calculate this object for the dynamics defined
previously.
Given that the waiting time distributions are characterized by a finite
average time $\tau _{\mu },$ Eq. (\ref{TauMedio}), after invoking the law of
large numbers, in the long time limit the total residence time $t_{u}$ in
each state can be approximated as $t_{u}\simeq n_{\mu }\tau _{\mu }.$
Consistently, the present time is $t\simeq \sum\nolimits_{\mu =1}^{L}n_{\mu
}\tau _{\mu }.$ Hence, we can writ
\begin{equation}
\lim_{t\rightarrow \infty }\frac{t_{u}}{t}\simeq \lim_{n\rightarrow \infty
\frac{n_{\mu }\tau _{\mu }}{\sum\nolimits_{\mu ^{\prime }=1}^{L}n_{\mu
^{\prime }}\tau _{\mu ^{\prime }}}=\frac{f_{\mu }\tau _{\mu }}
\sum\nolimits_{\mu ^{\prime }=1}^{L}f_{\mu ^{\prime }}\tau _{\mu ^{\prime }}
,
\end{equation
where the last relation follows from Eq. (\ref{fracciones}). Taking into
account that the fractions $\{f_{\mu }\}_{\mu =1}^{L}$ are characterized by
the distribution $\mathcal{P}(\{f_{\mu }\}),$ Eq.~(\ref{P(O)}) become
\begin{eqnarray}
P(\mathcal{O}) &=&\int_{\Lambda }df_{1}\cdots df_{L-1}\mathcal{P}(\{f_{\mu
}\}) \notag \\
&&\times \delta \Big{(}\mathcal{O}-\sum\nolimits_{\mu =1}^{L}\frac{f_{\mu
}\tau _{\mu }}{\sum\nolimits_{\mu ^{\prime }=1}^{L}f_{\mu ^{\prime }}\tau
_{\mu ^{\prime }}}\mathcal{O}_{\mu }\Big{)}. \label{Distribution}
\end{eqnarray
Therefore, $P(\mathcal{O})$ can be completely characterized after knowing
the distribution $\mathcal{P}(\{f_{\mu }\}).$ Notice that the specific
structure of the waiting time distributions $\{w_{\mu }(t)\}_{\mu =1}^{L}$
only appears through the average times $\{\tau _{\mu }\}_{\mu =1}^{L},$ Eq.
\ref{TauMedio}).
\subsection{Ergodicity and localization}
For an ergodic dynamics the fractions $f_{\mu }$ [Eq. (\ref{fracciones})]
must be characterized by their ensemble average, Eq. (\ref{qu}). Hence
\begin{equation}
\mathcal{P}(\{f_{\mu }\})=\delta (f_{1}-\langle f_{1}\rangle )\delta
(f_{2}-\langle f_{2}\rangle )\cdots \delta (f_{L}-\langle f_{L}\rangle ).
\label{Delta}
\end{equation
Inserting this expression into Eq. (\ref{Distribution}), it follows the
distributio
\begin{equation}
P(\mathcal{O})=\delta (\mathcal{O-}\langle \mathcal{O}\rangle _{\mathrm{st
}), \label{DeltaErgodico}
\end{equation
where $\langle \mathcal{O}\rangle _{\mathrm{st}}$\ is given by Eq. (\re
{OEstacion}) with the weight
\begin{equation}
P_{\mu }^{\mathrm{st}}=\frac{\langle f_{\mu }\rangle \tau _{\mu }}{\sum_{\mu
^{\prime }=1}^{L}\langle f_{\mu ^{\prime }}\rangle \tau _{\mu ^{\prime }}}.
\end{equation
From Eqs. (\ref{PEqui}) and (\ref{Delta}), we note that these weights
correspond to the stationary probabilities of each state $\mu $ in the
ergodic case. Hence, time averages and ensemble averages do in fact coincide.
The maximal departure with respect to ergodicity happens when the dynamics
localize, that is, the system remains in the initial condition. This case
corresponds t
\begin{equation}
\mathcal{P}(\{f_{\mu }\})=\sum_{\mu =1}^{L}p_{\mu }\delta (f_{1})\cdots
\delta (f_{\mu }-1)\cdots \delta (f_{L}). \label{Localizado}
\end{equation
Hence, Eq. (\ref{Distribution}) become
\begin{equation}
P(\mathcal{O})=\sum_{\mu =1}^{L}p_{\mu }\delta (\mathcal{O}-\mathcal{O}_{\mu
}). \label{PLocalization}
\end{equation
These limits are reached by the following memory mechanisms.
\section{Examples}
In the examples worked below, the stochastic dynamics may reach both the
ergodic and localized regimes Eqs. (\ref{DeltaErgodico}) and (\re
{PLocalization}) respectively. The distribution $\mathcal{P}(\{f_{\mu }\})$
can be explicitly calculated and then the non-ergodic properties
characterized through Eq. (\ref{Distribution}).
\subsection{Elephant random walk model}
This correlation model has been studied extensively in the recent literature
as a mechanism for inducing superdiffusion \cite{gunter,kim,kursten}. In the
present context, it is defined by the transition probabilit
\begin{equation}
\mathcal{T}_{n}(\{n_{\nu }\}|\mu )=\varepsilon q_{\mu }+(1-\varepsilon
\frac{n_{\mu }}{n}. \label{BAmodel}
\end{equation
The positive weights $0<q_{\mu }<1$ are extra parameters normalized as
\sum_{\mu =1}^{L}q_{\mu }=1.$ The parameter $\varepsilon $ assumes values in
the interval $[0,1].$ The stochastic dynamics can be read as follows. With
probability $\varepsilon ,$ and independently of the previous history, the
new state is chosen in agreement with the probabilities $\{q_{\mu }\}_{\mu
=1}^{L}.$ On the other hand, with probability $(1-\varepsilon )$ each state
is chosen in agreement with the weights $\{n_{\mu }/n\}_{\mu =1}^{L},$ which
in fact depend on the whole previous history of the process.
For $\varepsilon =1,$ the selection of the new state is completely random
and independent of the previous history. Therefore, the system is ergodic in
this case, Eq. (\ref{DeltaErgodico}). On the other hand, for $\varepsilon =0$
the dynamics localize, that is, the system remains in the initial condition,
Eq. (\ref{PLocalization}).
Even when the dynamics reaches the ergodic and localized regime, for
intermediates values $0<\varepsilon <1$ the dynamics is ergodic. This
property is demonstrated in Appendix \ref{elephant}. In fact, the
distribution $\mathcal{P}(\{f_{\mu }\})$ is delta distributed, Eq. (\re
{Delta}), with $\langle f_{\mu }\rangle =q_{\mu }.$
\subsection{Random walk driven by an urn-like dynamics}
In the P\'{o}lya urn dynamics \cite{feller,norman} (initially) an urn
contains many balls that, for example, are characterized by $L$ different
possible colors. At each step, one determine the color of one ball taken at
random and put into the urn one extra ball of the same color. A similar
process can be defined by starting the urn with only one ball \cit
{pitman,queen,budini} (Blackwell-MacQueen urn). Its dynamics is defined by
the following conditional probability, which is taken as the driving memory
mechanism.
For the random walk over the $\mu =1,\cdots L$ states, we take the
conditional probability \cite{pitman,queen
\begin{equation}
\mathcal{T}_{n}(\{n_{\nu }\}|\mu )=\frac{\lambda q_{\mu }+n_{\mu }}
n+\lambda }. \label{PolyaModel}
\end{equation
As before, the set of parameters $\{q_{\mu }\}_{\mu =1}^{L}$ is normalized
to one. Instead, $\lambda $ is a positive free parameter. For $\lambda
\rightarrow \infty $ the dynamics loses any dependence on the previous
history achieving in consequence an ergodic regime, Eq. (\ref{DeltaErgodico
). On the other hand, for $\lambda =0,$ a localized regime is achieved, Eq.
\ref{PLocalization}). Hence, the intermediate values of $\lambda $ avoid
this regime and in consequence one can define a nontrivial dynamics starting
from $n=1.$
For arbitrary values of $\lambda ,$ the probability density of the
(asymptotic) fractions (\ref{fracciones}) is derived in Appendix \re
{dirichlet}. It can be written a
\begin{equation}
\mathcal{P}(\{f_{\mu }\})=\left\{ \sum_{\nu =1}^{L}\frac{p_{\nu }}{q_{\nu }
f_{\nu }\right\} D(\{f_{\mu }\}|\{\lambda q_{\mu }\}), \label{PPolya}
\end{equation
where $D(\{f_{\mu }\}|\{\lambda _{\mu }\})$ is a Dirichlet distribution \cit
{pitman,queen}
\begin{equation}
D(\{f_{\mu }\}|\{\lambda _{\mu }\})\equiv \frac{\Gamma (\lambda )}
\prod_{\mu ^{\prime }}\Gamma (\lambda _{\mu ^{\prime }})}\prod_{\mu }f_{\mu
}^{\lambda _{\mu }-1}. \label{Dirichlete}
\end{equation
Here, $\lambda =\sum\nolimits_{\mu =1}^{L}\lambda _{\mu }.$ The (ensemble)
average fraction reads $\langle f_{\mu }\rangle =(q_{\mu }\lambda +p_{\mu
})/(\lambda +1).$ When $p_{\nu }=q_{\nu },$ due to the normalization
\sum_{\nu =1}^{L}f_{\nu }=1,$ the first factor in Eq. (\ref{PPolya}) does
not contribute, and $\langle f_{\mu }\rangle =q_{\mu }.$
We notice that $\mathcal{P}(\{f_{\mu }\})$ [Eq. (\ref{PPolya})] depends on
the initial conditions $\{p_{\mu }\}_{\mu =1}^{L}.$ This property arises
from the strong memory effects that drive the underlying stochastic
dynamics. Nevertheless, this dependence is not able to cancel any of the
stationary fractions. In consequence, the initial conditions are not
relevant for breaking or not ergodicity. In fact, given that $\mathcal{P
(\{f_{\mu }\})$ departs from Eq. (\ref{Delta}), this model leads to EB. The
distribution $P(\mathcal{O})$ [Eq. (\ref{Distribution})] can be evaluated
from Eq. (\ref{PPolya}).
As an example, we consider a \textit{two-level system}, where the observable
is defined by $\{\mathcal{O}_{\mu }\}\rightarrow (\mathcal{O}_{2},\mathcal{O
_{1}),$ with $\mathcal{O}_{1}\leq \mathcal{O\leq O}_{2}.$ After integration,
we ge
\begin{equation}
P(\mathcal{O})=\frac{1}{\mathcal{N}}\frac{[\omega _{2}(\mathcal{O}_{2}
\mathcal{O})]^{\lambda _{1}-1}[\omega _{1}(\mathcal{O}-\mathcal{O
_{1})]^{\lambda _{2}-1}}{[\omega _{2}(\mathcal{O}_{2}-\mathcal{O})+\omega
_{1}(\mathcal{O}-\mathcal{O}_{1})]^{\lambda _{1}+\lambda _{2}}},
\label{PtimePolya}
\end{equation
where for shortening the expression we introduced the parameters $\lambda
_{1}\equiv \lambda q_{1},$\ $\lambda _{2}\equiv \lambda q_{2},$ and the
weight
\begin{equation}
\omega _{1}\equiv \frac{\tau _{1}}{\tau _{1}+\tau _{2}},\ \ \ \ \ \ \ \ \
\omega _{2}\equiv \frac{\tau _{2}}{\tau _{1}+\tau _{2}}.
\end{equation
Here, $\tau _{1}$ and $\tau _{2}$ are the average times corresponding to the
two waiting time distributions $w_{1}(t)$ and $w_{2}(t)$ respectively [Eq.
\ref{TauMedio})]. The normalization constant reads $\mathcal{N}^{-1}=
\mathcal{O}_{2}-\mathcal{O}_{1})\omega _{1}\omega _{2}\Gamma (\alpha
_{1}+\alpha _{2})/\Gamma (\alpha _{1})\Gamma (\alpha _{2}).$ For simplicity,
in the previous expressions we assumed the initial condition $p_{\mu
}=q_{u}.\ $The case $p_{\mu }\neq q_{u}$ can be recovered from these
expressions [see Eqs. (\ref{PPolya}) and (\ref{Dirichlete})].
The model (\ref{PolyaModel}) demonstrates that global memory effects may
lead to EB. This result has a close relation with the breakdown of the
standard central limit theorem for globally correlated random variables \cit
{budini}. On the other hand, as shown in Sec. IV, depending on the values of
$\lambda ,$ here EB arises because the residence times in each state may be
divergent, that is, their probability density is characterized by power-law
tails. The next modified dynamics also develops EB, but does not involve
power-law statistics.
\subsection{Imperfect urn-like Model}
Here, we consider a model that can be seen as an imperfect case of the
previous one. We consider the possibility of having random state selections
that do not depend on the previous\ system history. The transition
probability read
\begin{equation}
\mathcal{T}_{n}(\{n_{\nu }\}|\mu )=\varepsilon q_{\mu }+(1-\varepsilon
\frac{\lambda q_{\mu }+M_{\mu }}{M+\lambda }. \label{MixedModel}
\end{equation
The set $\{q_{\mu }\}_{\mu =1}^{L}$ is normalized as before, $0\leq
\varepsilon \leq 1,$ and $\lambda \geq 0.$ Hence, with probability
\varepsilon $ each state $\mu ,$\ independently of the previous trajectory,
is chosen with weight $q_{\mu }.$ Complementarily, with probability
(1-\varepsilon )$ the state is chosen in agreement with the urn mechanism
Eq. (\ref{PolyaModel}). In fact, here $M_{\mu }$ is the number of times that
the state $\mu $ was chosen \textit{with} the urn dynamics. Furthermore, $M$
is the number of times that the urn mechanism was applied, $M=\sum_{\mu
=1}^{L}M_{\mu }.$ In contrast with the elephant model [Eq. (\ref{BAmodel})],
here the contribution proportional to $\varepsilon $ can be think as an
error in the application of the urn dynamics.
In order to clarify the stochastic dynamics induced by Eq. (\ref{MixedModel
), in Fig. 1 we plot two realizations (upper panels) for a two-level system
with $\mathcal{O}_{2}=1$ and $\mathcal{O}_{1}=-1.$ Hence, the observable
realizations switch between these two values. The waiting time distributions
are exponential one
\begin{equation}
w_{\mu }(t)=\gamma _{\mu }\exp [-\gamma _{\mu }t],
\label{ExponentialWaiting}
\end{equation
with $\mu =1,2.$ In the lower panels we plotted the conditional probability
\mathcal{T}_{n}(\{n_{\nu }\}|\mu )$ as a function of $n.$ For clarity, each
value is continued in the real interval $(i-1,i).$ The left panels
corresponds to $\varepsilon =0.1,$ Eq. (\ref{MixedModel}), while the right
panels to $\varepsilon =0,$ that is, Eq. (\ref{PolyaModel}). In both cases
\mathcal{T}_{n}(\{n_{\nu }\}|\mu )$ attains stationary values for increasing
$n$ [Eq. (\ref{fracciones})]. Nevertheless, in the case $\varepsilon =0.1$
at random values of $n$ the conditional probability collapses to the value
q_{\mu }.$ This effect gives the error or imperfection\ with respect to the
case $\varepsilon =0.$
\begin{figure}[tbp]
\includegraphics[bb=45 810 700 1110,angle=0,width=8.6cm]{Break_Figura1.eps}
\caption{Realizations of a two-level systems (upper panels) with observable
\{\mathcal{O}_{2}=1,\mathcal{O}_{1}=-1\}$ driven by an urn-like dynamics,
jointly with the corresponding conditional probabilities $\mathcal{T
_{n}(\{n_{\protect\nu }\}|\protect\mu )$ [Eqs. (\protect\ref{PolyaModel})
and (\protect\ref{MixedModel})] as a function of $n$ (lower panels). The
parameters are $\protect\lambda =2,$ $p_{1}=q_{1}=0.4,$ and
p_{2}=q_{2}=0.6. $ The waiting time distributions are exponential functions
[Eq. (\protect\ref{ExponentialWaiting})] with $\protect\gamma _{1}=\protec
\gamma _{2}=\protect\gamma .$ In (a) and (b) $\protect\varepsilon =0.1,$
while in (c) and (d) we take $\protect\varepsilon =0.$}
\end{figure}
The probability distribution of the asymptotic fractions [Eq. (\re
{fracciones})] associated to Eq. (\ref{MixedModel}) is given b
\begin{eqnarray}
\mathcal{P}(\{f_{\mu }\}) &=&\left\{ \sum_{\nu =1}^{L}\frac{p_{\nu }}{q_{\nu
}}\frac{f_{\nu }-\varepsilon q_{\nu }}{1-\varepsilon }\right\} \frac{1}
(1-\varepsilon )^{L-1}} \notag \\
&&D\Big{(}\Big{\{}\frac{f_{\mu }-\varepsilon q_{\mu }}{1-\varepsilon
\Big{\}}|\{\lambda q_{\mu }\}\Big{)}, \label{DirReducida}
\end{eqnarray
where $D(\{f_{\mu }\}|\{\lambda _{\mu }\})$ is the Dirichlet distribution
Eq. (\ref{Dirichlete}). Furthermore, each fraction is restricted to the
domai
\begin{equation}
\varepsilon q_{\mu }\leq f_{\mu }\leq 1-\varepsilon (1-q_{\mu }).
\label{DominioF}
\end{equation
In this case, the average fraction read
\begin{equation}
\langle f_{\mu }\rangle =\frac{q_{\mu }(\lambda +\varepsilon )+p_{\mu
}(1-\varepsilon )}{(\lambda +1)}. \label{FraccionMedias}
\end{equation}
Eq. (\ref{DirReducida}) is related to Eq. (\ref{PPolya}) by the change of
variables $f_{\mu }\rightarrow \varepsilon q_{\mu }+(1-\varepsilon )f_{\mu
}. $ This relation follows by considering the asymptotic limits of Eqs. (\re
{MixedModel}) and (\ref{PolyaModel}), and by using that the law of large
numbers applies to the error mechanism. For $\varepsilon =0$ the previous
expressions recover the previous case, Eq. (\ref{PPolya}). Interestingly,
the effect of introducing the imperfect mechanism is to reduce the domain of
each fraction $f_{\mu }$, Eq. (\ref{DominioF}).
From Eqs. (\ref{Distribution}) and (\ref{DirReducida}) we can calculate the
distribution of the time-averaged observable. Below we consider a \textit
two-level system }with $\mathcal{O}_{1}<\mathcal{O}<\mathcal{O}_{2}$ and
initial condition $p_{\mu }=q_{u}.$ This case straightforwardly allows us to
reconstruct the case $p_{\mu }\neq q_{u}.$ We ge
\begin{eqnarray}
P(\mathcal{O}) &=&\frac{1}{\mathcal{N}_{\varepsilon }}[\omega _{2}(\mathcal{
}_{2}-\mathcal{O})-\omega _{1}^{\varepsilon }(\mathcal{O}-\mathcal{O
_{1})]^{\lambda _{1}-1} \notag \\
&&\times \lbrack \omega _{1}(\mathcal{O}-\mathcal{O}_{1})-\omega
_{2}^{\varepsilon }(\mathcal{O}_{2}-\mathcal{O})]^{\lambda _{2}-1}
\label{PTimeMixto} \\
&&\times \frac{1}{[\omega _{2}(\mathcal{O}_{2}-\mathcal{O})+\omega _{1}
\mathcal{O}-\mathcal{O}_{1})]^{\lambda _{1}+\lambda _{2}}}. \notag
\end{eqnarray
The possible values of the time-averaged observable is restricted to the
domain $\mathcal{O}_{\min }\leq \mathcal{O}\leq \mathcal{O}_{\max },$ wher
\begin{equation}
\mathcal{O}_{\max }\equiv \frac{\mathcal{O}_{2}+\mathcal{O}_{1}\omega
_{1}^{\varepsilon }\omega _{2}^{-1}}{1+\omega _{1}^{\varepsilon }\omega
_{2}^{-1}},\ \ \ \ \ \mathcal{O}_{\min }\equiv \frac{\mathcal{O}_{1}
\mathcal{O}_{2}\omega _{2}^{\varepsilon }\omega _{1}^{-1}}{1+\omega
_{2}^{\varepsilon }\omega _{1}^{-1}}. \label{ObsLimits}
\end{equation
Furthermore, we introduced the parameter
\begin{equation}
\omega _{1}^{\varepsilon }\equiv \omega _{1}\frac{\varepsilon q_{1}}
1-\varepsilon q_{1}},\ \ \ \ \ \ \ \ \omega _{2}^{\varepsilon }\equiv \omega
_{2}\frac{\varepsilon q_{2}}{1-\varepsilon q_{2}},
\end{equation
while the normalization constant is $\mathcal{N}_{\varepsilon }^{-1}=
\mathcal{O}_{2}-\mathcal{O}_{1})\omega _{1}\omega _{2}[\Gamma (\lambda
)/\Gamma (\lambda _{1})\Gamma (\lambda _{2})](1-\varepsilon )^{-(\lambda
-1)}(1-\varepsilon q_{1})^{\lambda _{1}-1}(1-\varepsilon q_{2})^{\lambda
_{2}-1}.$ Consistently, for $\varepsilon =0,$ Eq. (\ref{PTimeMixto})
recovers the previous case, Eq. (\ref{PtimePolya}). From the previous
expression it become clear that the error mechanism introduced in Eq. (\re
{MixedModel}) lead to a shrinking of the probability density of the
time-averaged observable.
\begin{figure}[tbp]
\includegraphics[bb=45 580 710 1105,width=8.6cm]{Break_Figura2.eps}
\caption{Probability density of the time-averaged observable $\mathcal{O}$.
We take a two-level system driven by an urn-like dynamics with different
values of $\protect\lambda $ and $\protect\varepsilon .$ The full lines
correspond to the analytical expressions Eqs. (\protect\ref{PtimePolya}) and
(\protect\ref{PTimeMixto}). The waiting time distributions are exponential
functions [Eq. (\protect\ref{ExponentialWaiting})] with $\protect\gamma _{1}
\protect\gamma _{2}=\protect\gamma .$ In all plots we take
p_{1}=p_{2}=q_{1}=q_{2}=1/2.$ The (red) circles ($\protect\varepsilon =0.5$)
and (blue) squares ($\protect\varepsilon =0$) correspond to numerical
simulations. $\protect\lambda $\ is indicated in each plot.}
\end{figure}
In order to check these results, in Fig. 2, we plot the distribution (\re
{PTimeMixto}) for a two-level system where as before we take $\mathcal{O
_{2}=1,$ $\mathcal{O}_{1}=-1,$ and the exponential waiting time
distributions (\ref{ExponentialWaiting}). For each value of $\lambda ,$\ we
plot the cases $\varepsilon =0.5$ [Eqs. (\ref{MixedModel}) and (\re
{PTimeMixto})] and $\varepsilon =0$ [Eqs. (\ref{PolyaModel}) and (\re
{PtimePolya})]. Consistently, a higher $\varepsilon $ leads to a shrinking
of the density $P(\mathcal{O}),$\ which confirms that for $\varepsilon
\rightarrow 1$ an ergodic regime is achieved, $P(\mathcal{O})=\delta
\mathcal{O}).$ The same happen for increasing $\lambda .$ On the other hand,
the plots show that $P(\mathcal{O})$ may develops different forms such as $U$
and bell shapes, or even uniform ones. Similar dependences arise when
studying renewal random walks with divergent average trapping times \cit
{reben}.
In all cases, the numerical simulations (circles and squares) follows from a
time average performed on a time interval with $n=10^{3}$ steps and $10^{5}$
realizations. The theoretical results fit very well the numerical ones.
\section{Probability density of residence times}
In contrast to the elephant random walk model, the previous urn models
develop weak EB. Here, we explore if this property is induced, or not, by a
power-law statistics. In fact, for continuous-time random walks with renewal
events, EB is induced by the divergence of the average residence time in
each state \cite{reben}. The residence times are the random times that the
system stays or remains in a given state before jumping to another one (see
Fig. 1). Here, for the models introduced previously, we calculate their
probability density. The calculations are valid for arbitrary number of
states $L.$
We consider a single trajectory in the \textit{long time limit}, such that
the fractions $\{f_{\mu }\}_{\mu =1}^{L}$ [Eq. (\ref{fracciones})] can be
described by their associated probability density $\mathcal{P}(\{f_{\mu }\})$
[see Eqs. (\ref{PPolya}) and (\ref{DirReducida})]. At the beginning of the
residence in a given state $\mu $ the first time interval is chosen in
agreement with its waiting time distribution $w_{\mu }(t).$ In each step,
the system remains in the same state with probability $f_{\mu },$ which add
a new random time interval also defined from $w_{\mu }(t).$ The residence
time ends when a different state $\nu \neq \mu $ is chosen. This change
occurs with probability $(1-f_{\mu }).$ Therefore, the probability $W_{\mu
}(\{f\}|\tau )d\tau $ of leaving the state $\mu $ after a residence time
\tau $ can be written in the Laplace domain $[g(s)=\int_{0}^{\infty }d\tau
g(\tau )e^{-s\tau }]$ a
\begin{equation}
W_{\mu }(\{f\}|s)=(1-f_{\mu })w_{\mu }(s)\sum_{n=0}^{\infty }f_{\mu }^{n}\
w_{\mu }^{n}(s).
\end{equation
Here, $w_{\mu }(s)$ is the Laplace transform of the waiting time
distribution $w_{\mu }(t)$ associated to the state $\mu .$ The previous
expression takes into account all possible way of leaving the state $\mu $
after a given number of steps. It can be rewritten a
\begin{equation}
W_{\mu }(\{f\}|s)=(1-f_{\mu })\frac{w_{\mu }(s)}{1-f_{\mu }w_{\mu }(s)}.
\label{Wu}
\end{equation}
The density\ $W_{\mu }(\{f\}|\tau )$ is a conditional object. In fact, it is
defined for a particular realization with random values of the fraction
f_{\mu }.$ Therefore, the probability density of the residence time $W_{\mu
}(t)$ is obtained after averaging over realizations, $W_{\mu
}(t)=\left\langle W_{\mu }(\{f\}|t)\right\rangle ,$ which is equivalent to
an average over the distribution $\mathcal{P}(\{f_{\nu }\})$ of the set of
fractions $\{f_{\nu }\}_{\nu =1}^{L}.$ Therefore, we ge
\begin{equation}
W_{\mu }(\tau )=\int_{\Lambda }df_{1}\cdots df_{L-1}\mathcal{P}(\{f_{\nu
}\})W_{\mu }(\{f\}|\tau ), \label{WFinal}
\end{equation
where $W_{\mu }(\{f\}|\tau )$ follows from Eq. (\ref{Wu}) after Laplace
inversion. The average persistence time $T_{\mu }$\ is defined b
\begin{equation}
T_{\mu }\equiv \int_{0}^{\infty }d\tau \ \tau W_{\mu }(\tau ).
\label{AverageResidence}
\end{equation}
The previous two expressions can be evaluated for arbitrary waiting time
distributions and memory models. For an exponential waiting time
distribution $w_{\mu }(t)=\gamma _{\mu }\exp [-\gamma _{\mu }t]$ [Eq. (\re
{ExponentialWaiting})] with mean value $\tau _{\mu }=1/\gamma _{\mu }$ [Eq.
\ref{TauMedio})], it follows $w_{\mu }(s)=\gamma _{\mu }/(s+\gamma _{\mu }).$
From Eq. (\ref{Wu}), we get $W_{\mu }(\{f\}|s)=(1-f_{\mu })\gamma _{\mu
}/[s+(1-f_{\mu })\gamma _{\mu }],$ which can be inverted a
\begin{equation}
W_{\mu }(\{f\}|\tau )=(1-f_{\mu })\gamma _{\mu }\exp [-(1-f_{\mu })\gamma
_{\mu }\tau ].
\end{equation
For an ergodic system, characterized by the probability density $\mathcal{P
(\{f_{\nu }\})$ given by Eq. (\ref{Delta}), from Eq. (\ref{WFinal}) we ge
\begin{equation}
W_{\mu }(\tau )=(1-\left\langle f_{\mu }\right\rangle )\gamma _{\mu }\exp
[-(1-\left\langle f_{\mu }\right\rangle )\gamma _{\mu }\tau ]. \label{WErgo}
\end{equation
This result is consistent with the definition of the underlying stochastic
process that in each step allows the persistence in the same state. In fact,
the average persistence time is $T_{\mu }=1/[\gamma _{\mu }(1-\left\langle
f_{\mu }\right\rangle )],$ indicating an increasing of the average
persistence time with an increasing of the weight $\left\langle f_{\mu
}\right\rangle .$ On the other hand, in the localized regime [Eq. (\re
{Localizado})], due to the absence of transitions, it is not possible to
define $W_{\mu }(\tau ).$
Taking exponential waiting time distributions Eq. (\ref{ExponentialWaiting
), for a \textit{two-level system} $[\mu =1,2]$ characterized by the
conditional probability (\ref{MixedModel}) (imperfect urn model), after a
simple change of variables, Eqs. (\ref{DirReducida}) and (\ref{WFinal})
delive
\begin{equation}
W_{\mu }^{\varepsilon }(\tau )=\frac{1}{\mathcal{N}}\int_{0}^{1}df\ \varphi
_{\varepsilon }\exp [-\varphi _{\varepsilon }\ \tau ]c_{\mu }\ f^{\lambda
_{\mu }-1}(1-f)^{\lambda _{\mu ^{\prime }}-1}, \label{W_TLS}
\end{equation
where the super-index denotes the dependence on the parameter $\varepsilon .$
$\lambda _{\mu }=\lambda q_{\mu }$ $[\mu =1,2]$ while $\lambda _{\mu
^{\prime }}$ $[\mu ^{\prime }=2,1]$ corresponds to the other system state,
\lambda _{\mu ^{\prime }}=\lambda q_{\mu ^{\prime }}=\lambda (1-q_{\mu }).$
The initial conditions appears through the contributio
\begin{equation}
c_{\mu }\equiv \frac{p_{\mu }}{q_{\mu }}f+\frac{1-p_{\mu }}{1-q_{\mu }}(1-f).
\end{equation
The decay rate $\varphi _{\varepsilon }$ i
\begin{equation}
\varphi _{\varepsilon }\equiv \gamma _{\mu }[1-\varepsilon q_{\mu
}-(1-\varepsilon )f],
\end{equation
while the normalization constant reads $\mathcal{N}^{-1}=\Gamma (\lambda
_{1}+\lambda _{2})/\Gamma (\lambda _{1})\Gamma (\lambda _{2}).$
Straightforwardly, the average persistence time, $T_{\mu }^{\varepsilon
}=\int_{0}^{\infty }d\tau \ \tau W_{\varepsilon }(\tau ),$ from Eq. (\re
{W_TLS}) can then be written a
\begin{equation}
T_{\mu }^{\varepsilon }=\frac{1}{\mathcal{N}}\int_{0}^{1}df\ \frac{1}
\varphi _{\varepsilon }}c_{\mu }f^{\lambda _{\mu }-1}(1-f)^{\lambda _{\mu
^{\prime }}-1}. \label{TResidenceTLS}
\end{equation}
Consistently, for $\varepsilon =1$ Eq. (\ref{W_TLS}) recovers Eq. (\re
{WErgo}) with $\left\langle f_{\mu }\right\rangle =q_{\mu }$ [Eq. (\re
{FraccionMedias})]. Hence, $\gamma _{\mu }T_{\mu }^{1}=1/(1-q_{\mu }).$ The
same results arise when $\lambda \rightarrow \infty .$ For arbitrary
\varepsilon $ and $\lambda ,$ from Eqs. (\ref{W_TLS}) and (\re
{TResidenceTLS})\ explicit analytical expressions can be found for both
W_{\mu }^{\varepsilon }(\tau )$ and $T_{\mu }^{\varepsilon }$ [see Appendix
\ref{analytical}].
Interestingly, for $0<\varepsilon \leq 1$ (and any initial condition) the
average residence time $T_{\mu }^{\varepsilon }$ is finite [see Eq. (\re
{TResAn})]. This is the main result of this section. In fact, this result
demonstrates that weak EB may arise even in the absence of power-law
statistical distributions with divergent average residence times. On the
other hand, for the case $\varepsilon =0,$ that is, the dynamics defined by
the conditional probabilities (\ref{PolyaModel}), the average residence time
$T_{\mu }^{0},$ depending on the parameter values, may be finite or
infinite. From Eqs. (\ref{TResidenceTLS}) and Eq. (\ref{TResAn}) we ge
\begin{equation}
\gamma _{\mu }T_{\mu }^{0}=\frac{\lambda -\left( \frac{1-p_{\mu }}{1-q_{\mu
}\right) }{\lambda (1-q_{\mu })-1},\ \ \ \ \ \ \ \ \ \lambda >\frac{1}
(1-q_{\mu })}>1.
\end{equation
Consistently, for increasing $\lambda $ this expression recovers the ergodic
case [Eq. (\ref{WErgo})], $\lim_{\lambda \rightarrow \infty }\gamma _{\mu
}T_{\mu }^{0}=1/(1-q_{\mu }).$ In the complementary region of possible
values of $\lambda ,$ the average residence time is divergent
\begin{equation}
\gamma _{\mu }T_{\mu }^{0}=\infty ,\ \ \ \ \ \ \ \ \ \lambda \leq \frac{1}
(1-q_{\mu })}. \label{infinito}
\end{equation
This last regime indicates that the density $W_{\mu }^{\varepsilon }(\tau )$
develops power-law tails. In fact, for long residence times, $\gamma _{\mu
}\tau \gg 1,$ from Eqs. (\ref{W_TLS}) and (\ref{ResidentGeometric}) it can
be approximated a
\begin{equation}
W_{\mu }^{0}(\tau )\approx \gamma _{\mu }C_{\mu }^{0}\Big{(}\frac{1}{\gamma
_{\mu }\tau }\Big{)}^{\lambda (1-q_{\mu })+1}, \label{PowerAprox}
\end{equation
which defines the previous finite and infinite average time regimes. The
dimensionless constant reads $C_{\mu }^{0}=(p_{\mu }/q_{\mu })(1-q_{\mu
})\Gamma (1+\lambda )/\Gamma (q_{\mu }\lambda ).$ When $p_{\mu }=0$ $(p_{\mu
^{\prime }}=1)$ the asymptotic behavior becomes $W_{\mu }^{0}(\tau )\approx
(1/\gamma _{\mu }\tau )^{\lambda (1-q_{\mu })+2},$ while for $W_{\mu
^{\prime }}^{0}(\tau )$ is given by Eq. (\ref{PowerAprox}). We remark that
in general $W_{\mu }^{\varepsilon }(\tau )$ $(\varepsilon >0)$ may also
develop power-law behaviors. Nevertheless, a multiplicative exponential
factor always leads to finite averages times [see for example Eq. (\re
{WpartMU}) below]
\begin{figure}[tbp]
\includegraphics[bb=55 26 425 556,width=7.5cm]{Break_Figura3.eps}
\caption{Probability distribution $W_{\protect\mu }^{\protect\varepsilon }
\protect\tau )$ $[\protect\mu =1,2]$ of the residence times for a two-level
system. The full lines correspond to the analytical result Eq. (\protect\re
{WpartMU}). The waiting time distributions are exponential functions [Eq.
\protect\ref{ExponentialWaiting})] with $\protect\gamma _{1}=\protect\gamma
_{2}=\protect\gamma .$ In both curves, $p_{1}=p_{2}=q_{1}=q_{2}=1/2,$ and
\protect\lambda =2.$ The (red) circles correspond to a numerical simulation
with $\protect\varepsilon =0.5,$ while the (blue) squares to $\protec
\varepsilon =0.$ The dotted line is the asymptotic power-law behavior
\protect\ref{PowerAprox}) of Eq. (\protect\ref{WCeroPart}).}
\end{figure}
For particular values of the characteristic parameters, the integral results
Eqs. (\ref{W_TLS}) and (\ref{TResidenceTLS}) lead to simple expressions.
Taking $p_{1}=q_{1}=1/2,$ $p_{2}=q_{2}=1/2,$ and $\lambda =2$ [Fig. (2b)]
the density of residence times become
\begin{equation}
W_{\mu }^{\varepsilon }(\tau )=\frac{\exp (-\gamma _{\varepsilon }^{+}\tau
)(1+\gamma _{\varepsilon }^{+}\tau )-\exp (-\gamma _{\varepsilon }^{-}\tau
)(1+\gamma _{\varepsilon }^{-}\tau )}{\gamma _{\mu }\tau ^{2}(1-\varepsilon
}, \label{WpartMU}
\end{equation
where for shortening the expression we introduced the rates $\gamma
_{\varepsilon }^{+}\equiv \gamma _{\mu }\varepsilon /2$ and $\gamma
_{\varepsilon }^{-}\equiv \gamma _{\mu }(1-\varepsilon /2).$ In the case
\varepsilon =1$ (ergodic dynamics), we get $W_{\mu }^{\varepsilon }(\tau
)=(\gamma _{\mu }/2)T_{\mu }^{1}\exp [-(\gamma _{\mu }/2)\tau ].$ Hence,
T_{\mu }^{1}=2/\gamma _{\mu }.$ In the case $\varepsilon =0$ it reduces t
\begin{equation}
W_{\mu }^{0}(\tau )=\frac{1}{\gamma _{\mu }\tau ^{2}}[1-(1+\gamma _{\mu
}\tau )\exp (-\gamma _{\mu }\tau )], \label{WCeroPart}
\end{equation
which explicitly shows the presence of dominant power-law tails. The average
residence time [Eq. (\ref{TResidenceTLS})], for arbitrary $\varepsilon $
read
\begin{equation}
\gamma _{\mu }T_{\mu }^{\varepsilon }=\frac{2\mathrm{arctanh}(1-\varepsilon
}{(1-\varepsilon )}=\frac{\mathrm{\ln }\left( \frac{2-\varepsilon }
\varepsilon }\right) }{(1-\varepsilon )}, \label{TResParti}
\end{equation
where $\mathrm{arctanh}[x]=\ln \sqrt{\frac{1+x}{1-x}}$ for $x\in (-1,1).$
Thus, $T_{\mu }^{\varepsilon }$ is finite for $0<\varepsilon \leq 1.$
Consistently with Eqs. (\ref{infinito}) and (\ref{WCeroPart}), it diverges
for $\varepsilon =0,$ $T_{\mu }^{0}=\lim_{\varepsilon \rightarrow 0}T_{\mu
}^{\varepsilon }=\infty .$
In order to check the previous results we determined the distribution
W_{\mu }^{\varepsilon }(\tau )$ from a set of realizations such as those
shown in Fig. 1. For the same system than in Fig. 2, the results are shown
in Fig. 3. Furthermore, we take $w_{1}(t)=w_{2}(t)=\gamma \exp (-\gamma t),$
which implies $W_{1}^{\varepsilon }(\tau )=W_{2}^{\varepsilon }(\tau ).$
Consistently with the previous analytical results [Eq. (\ref{WpartMU})], for
$\varepsilon =0.5$ [Fig. 3(a)] asymptotically the density of residence times
$W_{\mu }^{\varepsilon }(\tau )$ is not dominated by power-law behaviors.
Instead for $\varepsilon =0$ [Fig. 3(b)] an asymptotic power-law behavior is
clearly observed [Eq. (\ref{WCeroPart})]. The numerical and theoretical
results are consistent between them.
The numerical probability densities of Fig. 3 were obtained from a set of
equally sampled realizations. This means that the same number of data for
the random residence times are taken from each realization. We took $5\times
10^{3}$ realizations with a total length of $n=5\times 10^{5}$ steps.
Furthermore, after running the dynamics during $10^{3}$ steps (long time
limit), $5\times 10^{3}$ random residence times were taken from each
realization.
\section{Summary and Conclusions}
We have introduced a random walk dynamics characterized by global memory
mechanisms. Given a finite set of states, in each step the system may remain
in the same state of jump to another one. These alternative events are
chosen from a conditional probability that depends on the whole previous
history of the system. The time between consecutive steps is determinate by
a set of waiting time distributions, all of them characterized by a finite
average time.
We focused the analysis on the ergodic properties of the stochastic
dynamics. Hence, we characterized the probability density of time-averaged
observables, [Eq. (\ref{Distribution})]. By analyzing different memory
mechanisms, we conclude that global correlations are not a sufficient
condition for breaking ergodicity, such as for example in the elephant
random walk model \ [Eq. (\ref{BAmodel})]. On the other hand, alternative
urn-like memory mechanisms [Eqs. (\ref{PolyaModel}) and (\ref{MixedModel})]
do in fact break ergodicity. In these cases, considering a two-level
dynamics, the distribution of time-averaged observables can be found in an
explicit analytical way [Eqs. (\ref{PtimePolya}) and (\ref{PTimeMixto})].
For random walks dynamics over a finite set of states, EB may be induced by
a divergent average residence time in each state. In order to cheek this
possibility for the present models, we calculated the probability density of
the residence times [Eq. (\ref{WFinal})], and the corresponding average
residence time [Eq. (\ref{AverageResidence})]. In general, the distributions
do not develop asymptotic power-law behaviors consistent with a divergent
average residence time. Hence, we conclude that global memory effects are in
fact an alternative mechanism that leads to EB. This main conclusion was
explicitly checked for two-level dynamics [Eqs. (\ref{W_TLS} ) and (\re
{TResidenceTLS})]. Only for a particular set of values, the residence times
have a divergent average. All previous results were confirmed by numerical
simulations [see Figs. (2) and (3)].
In conclusion, we established that weak EB may arise in systems
characterized by global memory effects. This property may emerge even when
the relevant variables are not characterized by power-law statistical
behaviors.
\section*{Acknowledgments}
This work was supported by Consejo Nacional de Investigaciones Cient\'{\i
ficas y T\'{e}cnicas (CONICET), Argentina.
|
\section*{Introduction and formulation of the main results}
\addcontentsline{toc}{section}{Introduction and formulation of the problems}
\label{section0}
Consider the model domain $\Omega_\alpha$ which is the plane angle of magnitude $\alpha$ between the half axes $\bR^+$ and the beam $\bR_\alpha$ turned by the angle $\alpha$ from $\bR^+$ (see Fig. 1). The corresponding boundary is a model curve:
\begin{eqnarray}\label{e0.1}
\begin{array}{c}
\Gamma_\alpha:=\pa\Omega_\alpha=\bR^+\cup\bR_\alpha,\qquad \bR^+=[0,\infty),\qquad 0<\alpha<2\pi,\\[2mm]
\bR_\alpha:=\{e^{i\alpha}t=(t\cos\,\alpha,t\sin\,\alpha)\;:\; t\in\bR^+\}.
\end{array}
\end{eqnarray}
Note, that the case $\alpha=\pi$ is already well treated in the literature.
The unit normal vector field $\{\nub(x)\}_{x\in\Gamma_\alpha}$ on the boundary $\Gamma_\alpha$ is defined by the equality
\begin{eqnarray}\label{e0.2}
\nub(x)=\left\{\begin{array}{ll}
(0,-1)^\top & \hbox{\rm for}\quad x\in\bR^+\\[3mm]
(-\sin\,\alpha,\cos\,\alpha)& \hbox{\rm for}\quad x\in\bR_\alpha.\end{array}\right.
\end{eqnarray}
and defines the following normal derivative $\pa_{\nub}$ on the boundary:
\begin{eqnarray}\label{e0.3}
\partial_{\nub(t)}=\left\{\begin{array}{ll}
- \dst\lim_{(x_1,x_2)\to{t=(\tau,0)}}\pa_{x_2} & \text{for}\;\; t\in \bR^+,\\[3mm]
\dst\lim_{(x_1,x_2)\to t=(\tau\cos\,\alpha,\tau\sin\,\alpha)}[-\sin\, \alpha\,\pa_{x_1}+\cos\,\alpha\,\pa_{x_2}] & \text{for}\;\;t\in\bR_\alpha. \end{array}\right.
\end{eqnarray}
\setlength{\unitlength}{0.4mm}
\hskip-3mm
\begin{picture}(300,140)
\put(0,-20){\epsfig{file=angular_1.pdf,height=60mm, width=100mm}}
\put(0,30){\makebox(0,0)[lc]{$0$}}
\put(155,60){\makebox(0,0)[bc]{$t=(\tau\,\cos\,\alpha,\tau\,\sin\,\alpha)^\top$}}
\put(110,110){\makebox(0,0)[bc]{$\nub(t)=(-\sin\,\al,\cos\,\al)^\top$}}
\put(140,22){\makebox(0,0)[bc]{$t=(\tau,0)^\top$}}
\put(162,-15){\makebox(0,0)[bc]{$\nub(t)=(0,-1)^\top$}}
\put(83,35){\makebox(0,0)[bc]{$\alpha$}}
\put(150,37){\makebox(0,0)[bc]{$\Omega_\alpha$}}
\put(230,23){\makebox(0,0)[bc]{$\bR^+$}}
\put(200,121){\makebox(0,0)[bc]{$\bR_\alpha$}}
\end{picture}
\vskip-3mm
\hskip95mm{\rm Fig. 1}
\vskip9mm
Many problems in mathematical physics e.g., cracks in elastic media, electromagnetic scattering by surfaces etc., are formulated in the form of boundary value problems for elliptic partial differential equations in domains with angular points at the boundary. In the recent paper \cite{BDKT13} is described how such BVPs can be investigated with the help of their local representatives-model problems in planar angles
$\Omega_{\alpha_j}$ of magnitude $0<\alpha_j<2\pi$, $j=1,\ldots,m$. The purpose of the present paper is to study the model mixed boundary value problem for the Helmholtz equation in the model domain $\Omega _\alpha$
\begin{eqnarray}\label{e0.4}
\left\{\begin{array}{ll}
\Delta u(x)+k^2u(x)=f(x),\qquad & x\in\Omega_\alpha, \\[0.2cm]
u^+(t)=g(t), \qquad & t\in\mathbb{R}_\alpha, \\[0.2cm]
(\partial_\nub u)^+(t)=h(t),\qquad & t\in\mathbb{R}^+,
\end{array}\right.
\end{eqnarray}
for a complex parameter ${\rm Im}\,k\not=0$ Here $u^+$ and $(\partial_{\nub_\Gamma}u)^+$ denote respectively the Dirichlet and the Neumann traces on the boundary.
Let us recall short definitions of the Bessel potential $\mathbb{H}^s_p(\Omega_\alpha)$, $\widetilde{\mathbb{H}}^s_p( \Omega_\alpha)$, $\mathbb{H}^r_p(\bR^+)$, $\widetilde{\mathbb{H}}^r_p(\bR^+)$ and Sobolev-Slobode\v{c}kii $\widetilde{\mathbb{W}}^r_p(\bR^+)$, $\mathbb{W}^r_p(\bR_\alpha)$ etc. spaces for $r\in\bR$, $1<p<\infty$. The spaces $\mathbb{H}^r_p(\Gamma_\alpha)$, $\widetilde{\mathbb{H}}^r_p(\Gamma_\alpha)$, $\mathbb{W}^r_p(\Gamma_\alpha)$ and $\widetilde{\mathbb{W}}^r_p(\Gamma_\alpha)$ can only be defined for $-2+1/p<r<1+1/p$. For detailed definitions and properties of these spaces we refer to the classical source \cite{Tr95} and also to \cite{CD01,Du84a,Du01,DS93,Hr85,HW08}.
Bessel potential space $\bH^s_p(\bR^n)$ is defined as a subset of the space of Schwartz distributions $\bS'(\bR^n)$ and is endowed with the following norm (see \cite{Tr95}):
\[
||u\big|\bH_p^s(\bR^n)||:=||\langle D\rangle^su\big| L_p(\bR^n)||,\quad
\mbox{ where }\quad \langle D\rangle^s:=\cF^{-1}(1+|\xi|^2)^{\frac s2}\cF.
\]
$\cF$, $\cF^{-1}$ are the Fourier transforms. For the definition of the Sobolev-Slobode\v{c}kii space $\bW_p^s(\bR^n)=\bB_{p,p}^s(\bR^n)$ see \cite{Tr95}.
The space $\wt {\bH}_p^s(\Omega)$, where $\Omega\subset\bR^n$, is defined as the subspace of $\bH_p^s(\bR^n)$ of those functions (or distributions if $s<0$) $\vf\in \bH_p^s(\bR^n)$, which are supported in the subset $\Omega$, $\supp\vf\subset\ov{\Omega}$, whereas $\bH_p^s(\Omega)$ denotes the quotient space $\bH_p^s(\Omega):=\bH_p^s(\bR^n)\Big/\wt{\bH}_p^s(\Omega^c)$ and $\Omega^c:=\bR^n\setminus\ov{\Omega}$ is the complemented domain. The space $\bH_p^s(\Omega)$ can be identified with the space of distributions $\vf$ on $\Omega$ which admit extensions $\ell\vf\in\bH_p^s(\bR^n)$. Therefore $r_\Omega\bH_p^s(\bR^n)=\bH_p^s(\Omega)$, where $r_\Omega$ denotes the restriction from $\bR^n$ to the domain $\Omega$.
Worth noting that for an integer $m=1,2,\ldots$ the spaces $\bH^m_p(\bR^n)$ and $\bW^m_p(\bR^n)$ coincide and are known as the Sobolev spaces, endowed with the following equivalent norm (see \cite{Tr95}):
\[
||u\big|\bW_p^m(\bR^n)||:=\sum_{|\alpha|\leqslant m}||\pa^\alpha u\big| L_p(\bR^n)||.
\]
Let $\cS:=\partial\Omega$ be the smooth boundary and consider $\widetilde{\mathbb{H}}^{-1}_0(\Omega)$, a subspace of $\widetilde{\mathbb{H}}^{-1}(\Omega)$, orthogonal to
\[
\widetilde{\mathbb{H}}^{-1}_\cS (\Omega):=\left\{f \in\widetilde{\bH}^{-1} (\Omega)\;:\;\langle f,\varphi\rangle = 0 \;\text{for all}\;\varphi\in C^1_0(\Omega)\right\},
\]
where $\langle\cdot,\cdot\rangle$ denotes the pairing between the adjoint spaces and coincides with the usual scalar product for regular functions.
$\widetilde{\mathbb{H}}^{-1}_\cS(\Omega)$ consists of those distributions on $\Omega$, belonging to $\widetilde{\mathbb{H}}^{-1}(\Omega)$ which have their supports just on $\cS$ and $\widetilde{\mathbb{H}}^{-1} (\Omega)$ can be decomposed into the direct sum of subspaces which are orthogonal to each-other:
\[
\widetilde{\mathbb{H}}^{-1}(\Omega)= \widetilde{\mathbb{H}}^{-1}_\cS(\Omega)\oplus
\widetilde{\mathbb{H}}^{-1}_0(\Omega).
\]
The space $\widetilde{\bH}^{-1}_\cS(\Omega)$ is non-empty (see \cite[\S\, 5.1]{HW08}) and excluding it from $\widetilde{\mathbb{H}}^{-1}(\Omega)$ is necessary to make BVPs uniquelly solvable (cf. \cite{HW08} and the next Theorem \ref{t0.1}).
Let us commence with the existence results for a solution to BVP \eqref{e0.4}. Lax-Milgram Lemma applied to these BVPs gives the following solvability result.
\begin{theorem}\label{t0.1}
The Mixed BVP \eqref{e0.4} has a unique solution in the classical weak setting:
\begin{equation}\label{e0.5}
u\in\bH^1(\Omega_\alpha),\qquad f\in\widetilde{\mathbb{H}}^{-1}_0(\Omega_\alpha),\qquad g\in\mathbb{H}^{1/2}(\bR_\alpha),\qquad h\in\mathbb{H}^{-1/2}(\bR^+).
\end{equation}
\end{theorem}
{\bf Proof:} The proof is verbatim to the proof of similar Theorem 2.1 (also see Remark 2.2 and Remark 2.3) in \cite{DTT14}. The operator treated in \cite{DTT14}, is very similar: Laplace-Beltarmi with $0$ order summand on a compact surface. The main tool,the Lax-Milgram Lemma applies equally successful for non-compact domain. \QED
As we see from Theorem \ref{t0.1} the BVP \eqref{e0.4} has a unique solution in the classical setting \eqref{e0.5} independent of the values of angular points on the boundary. This property changes dramatically as soon as we consider the BVP \eqref{e0.4} in the following non-classical setting
\begin{eqnarray}\label{e0.6}
\begin{array}{r}
u\in\mathbb{H}^s_p(\Omega_\alpha),\quad f\in\widetilde{\mathbb{H}}^{s-2}_p(\Omega_\alpha)\cap
\widetilde{\mathbb{H}}^{-1}_0(\Omega_\alpha),\quad g\in\mathbb{W}^{s-1/p}_p(\bR_\alpha),\\[3mm]
h\in\mathbb{W}^{s-1-1/p}_p(\bR^+), \qquad 1<p<\infty, \quad
\dst\frac1p<s<1+\dst\frac1p.
\end{array}
\end{eqnarray}
(see Remark \ref{r0.4} below). This indicates that the derivative of a solution has singularities depending on the angles on the boundary; although this is knows long ago, we find for the first time all values of "forbidden angles" in Theorem \ref{t0.3} below.
Note that, the upper constraint in $\dst\frac1p<s<1+\dst\frac1p$ ensures the invariant definition of the Bessel potential and Sobolev-Slobode\v{c}kii spaces, while the lower constraint ensures the existence of the trace $u^+$ on the boundary.
Moreover, from Theorem \ref{t0.1} we can not even conclude that a solution is continuous in the closed domain $\overline{\Omega}_\alpha$. If we can prove that there is a solution $u\in\mathbb{H}^1_p(\Omega_\alpha)$ for some $2<p<\infty$, we can enjoy even a H\"older continuity of $u$ in $\overline{\Omega}_\alpha$. It is very important to know maximal smoothness of a solution in some problems, for example in approximation methods.
Along with the non-classical settings \eqref{e0.6} the trace of a solution $u^+$ on the boundary $\Gamma_\alpha$ satisfies the following compatibility conditions:
\begin{eqnarray}\label{e0.7}
\begin{array}{c}
u^+_+ - \J_\alpha u^+_\alpha\in\wt{\bW}^{s-1/p}_p(\bR^+),\\[3mm]
(\pa_\nub u)^+_+ +\J_\alpha (\pa_\nub u)^+_\alpha\in\wt{\bW}^{s-1/p-1}_p(\bR^+),
\end{array}
\end{eqnarray}
where $v^+_+$ denotes the trace on $\bR^+$, $v^+_\alpha$ denotes the trace on $\bR_\alpha$ and
\begin{eqnarray}\label{e0.8}
\J_\alpha\vf(t)=\vf(t\,\cos\,\alpha,t\,\sin\,\alpha),\qquad t\in\bR^+
\end{eqnarray}
is the pull back operator from $\bR_\alpha$ to $\bR^+$. These compatibility conditions are direct consequences of the inclusion
$u\in\bH^s_p(\Omega_\alpha)$ and the properties of traces on the boundary
$\Omega_\alpha$.
To formulate the main theorem of the present work we need the following definition.
\begin{definition}\label{d0.2}
The BVP \eqref{e0.4}, \eqref{e0.5} is Fredholm if the homogeneous problem $f=g=h=0$ has a finite number of linearly independent solutions and the BVP has a solution if and only if the data $f,g,h$ satisfy a finite number of orthogonality conditions.
\end{definition}
Next we formulate the main theorem of the present paper, which is proved in \S\, \ref{sect5}.
\begin{theorem}\label{t0.3}
Let $\alpha\in(0,2\pi)$, $1<p<\infty$ and $\dst\frac1p<s<1+\dst\frac1p$. The model mixed BVP \eqref{e0.4} is Fredholm in the non-classical setting \eqref{e0.6} if and only if:
\begin{eqnarray}\label{e0.10}
e^{4\pi(s-1/p)i}\sin^2\pi\left(s-\dst\frac2p\right)+\cos^2(\pi -\alpha)\left(s-\dst\frac2p\right)\not=0.
\end{eqnarray}
In details the condition \eqref{e0.10} is written as follows:
{
\begin{subequations}
\begin{eqnarray}\label{e0.10a}
&&\hskip-17mm 1)\;\; s\not=\dst\frac1p+\dst\frac n4\quad n=1,2,3\quad
\text{and}\quad s\not=\dst\frac2p+n,\quad n=0,-1,\\
\label{e0.10b}
&&\hskip-17mm 2)\;\;\text{If}\quad s=\dst\frac2p-1,
\quad\text{then}\quad\alpha\not=\frac\pi2,\,\frac{3\pi}2,\\
\label{e0.10c}
&&\hskip-17mm 3)\;\;\text{If}\;\;s=\frac1p+\dst\frac12,\;\;
\text{then}\;\; p\not=\dst\frac2{n-2m}\;\; \text{or}\;\;\alpha\not=\dst\frac{2k+1}{2m}\pi,\\
\label{e0.10d}
&&\hskip-19mm \begin{array}{l}
4)\;\;\text{If}\quad s=\dst\frac1p+\dst\frac n2+\dst\frac14,\quad n=0,1, \quad\text{then}\\
\quad\; \alpha\not=2\pi p\dst\frac{2k+1}{2np+p-4}\quad \text{and}
\quad \alpha\not=4\pi\dst\frac{np-kp-2}{2np+p-4},
\end{array}
\end{eqnarray}
\end{subequations}}
where $k,m=0,\pm1,\ldots.$
In particular, the BVP \eqref{e0.4} has a unique solution in the settings \eqref{e0.6} if:
{
\begin{eqnarray}
\begin{array}{l} \label{e0.11}
\dst\frac1p+\dst\frac14<s\leqslant\dst\frac2p\quad\text{for}\quad
1<p\leqslant2,\\[3mm]
\dst\frac2p\leqslant s<\dst\frac1p+\dst\frac34,\quad\text{for}\quad
2\leqslant p<\infty.
\end{array}
\end{eqnarray}}
\end{theorem}
\begin{remark}\label{r0.4}
From \eqref{e0.11} follows directly that the BVP \eqref{e0.4} has a unique solution in the setting \eqref{e0.6} but for $p=2$ if
\begin{eqnarray}\label{e0.11a}
\dst\frac34<s<\dst\frac54.
\end{eqnarray}
As we already know from Theorem \ref{e0.1} and as it follows from \eqref{e0.11a}, for $p=2$, $s=1$ (the classical setting) the BVP \eqref{e0.4} has no "forbidden angles" $\alpha$ and the problem is uniquely solvable for all values of $\alpha$ (cf. Theorem \ref{t0.1}). {The "forbidden angles" $\alpha$ might only emerge when $p\not=2$ or $s\not=1$ and for these angles the problem is not even Fredholm.}
\end{remark}
Theorem \ref{t0.3} is a corollary of the next two Theorem \ref{t0.5} and Theorem \ref{t0.6}.
\begin{theorem}\label{t0.5}
Let $1<p<\infty$ and $\dst\frac1p<s<1+\dst\frac1p$. Let $g_0\in\bW^{s-1/p}_p(\Gamma_\alpha)$ and $h_0\in\bW^{s-1-1/p}_p(\Gamma_\alpha)$ be some fixed extensions of the boundary conditions $g\in\bW^{s-1/p}_p(\bR_\alpha)$ and $h\in\bW^{s-1-1/p}_p(\bR^+)$ (non-classical case), defined initially
on parts of $\Gamma_\alpha$.
A solution to the BVP \eqref{e0.4} is represented by the formula
\begin{eqnarray}\label{e0.12}
u(\cx)=\N_\cC f(\cx)+\W_\Gamma(g_0+\varphi_0)(\cx)-\V_\Gamma(h_0
+\psi_0)(x),\qquad x\in\Omega_\alpha,
\end{eqnarray}
where $\varphi_0$ and $\psi_0$ are solutions to the following
system of pseudodifferential equations
\begin{eqnarray}\label{e0.13}
&&\left\{\begin{array}{ll}
\dst\frac12\varphi_0(t)+r_{\bR^+}\V_{\Delta+k^2,-1}\psi_0(t)
=G_+(t),\qquad & t\in\bR^+,\\[3mm]
\dst\frac12\psi_0(t) - r_{\bR_\alpha}\V_{\Delta+k^2,+1}\varphi_0(t)
=H_-(t)\qquad & t\in\bR_\alpha, \end{array}\right. \\[2mm]
&&\hskip15mm\varphi_0\in\wt{\bW}^{s-1/p}_p(\bR^+), \qquad
\psi_0\in\wt{\bW}^{s-1-1/p}_p(\bR_\alpha),\nonumber\\[2mm]
&&\hskip15mm G_+:=r_{\bR^+}G_0\in\bW^{s-1/p}_p(\bR^+),\qquad
H_-=r_{\bR_\alpha}H_0\in\bW^{s-1-1/p}_p(\bR_\alpha),\nonumber\\
&& G_0:=(\N_{\Delta+k^2}f)^+-\dst\frac12g_0+\W_{\Delta+k^2,0}g_0
-\V_{\Delta+k^2,-1}h_0\in\bW^{s-1/p}_p(\Gamma_\alpha),\nonumber \\
&& H_0:=(\pa_\nub\N_{\Delta+k^2} f)^+-\dst\frac12h_0
+\V_{\Delta+k^2,+1}g_0-\W^*_{\Delta+k^2,0}h_0\in\bW^{s-1-1/p}_p
(\Gamma_\alpha).\nonumber
\end{eqnarray}
Vice versa: if $u$ is a solution to the BVP \eqref{e0.4},
$g:=r_{\bR_\alpha}u^+$, $h:=r_{\bR^+}(\pa_\nub u)^+$ and
$g_0\in\bW^{s-1/p}_p(\Gamma_\alpha)$, $h_0\in\bW^{s-1-1/p}_p(\Gamma_\alpha)$
are some fixed extensions of $g$ and $h$ to $\Gamma_\alpha$, then $\vf_0:=u^+ - g_0$, $\psi_0:=(\pa_t u)^+ - h_0$ are solutions to the system \eqref{e0.13}.
The system of boundary pseudodifferential equations \eqref{e0.13} has a unique pair of solutions $\varphi_0,\psi_0\in\wt{\bW}^{-1/2}(\Gamma_\alpha)
=\wt{\bH}^{-1/2}(\bR^+)$ in the classical setting $p=2$, $s=1$.
\end{theorem}
The proof of Theorem \ref{t0.5} is exposed in \S\, \ref{sect3}.
For the system \eqref{e0.13} we can remove the constraint $\dst\frac1p<s<1+\dst\frac1p$ and consider two different settings for arbitrary $r\in\bR$:
\begin{subequations}
\begin{eqnarray}\label{e0.15a}
&\vf_0\in\wt{\bW}{}^r_p(\bR^+), \quad \psi_0\in\wt{\bW}{}^{r-1}_p(\bR^+),
\quad G\in\bW^r_p(\bR^+),\quad H \in\bW^{r-1}_p(\bR^+),\\[3mm]
\label{e0.15b}
&\vf_0\in\wt{\bH}{}^r_p(\bR^+), \quad \psi_0\in\wt{\bH}{}^{r-1}_p(\bR^+), \quad G\in\bH^r_p(\bR^+),\quad H\in\bH^{r-1}_p(\bR^+)
\end{eqnarray}
\end{subequations}
\begin{theorem}\label{t0.6}
Let $1<p<\infty$, $r\in\bR$.
The system of boundary integral equations \eqref{e0.13} is Fredholm in both the Sobolev-Slobode\v{c}kii \eqref{e0.15a} and the Bessel potential \eqref{e0.15b} space settings if and only if:
{
\begin{eqnarray}\label{e0.16}
e^{4\pi ri}\sin^2\pi\left(\dst\frac1p-r-1\right)+
\cos^2(\pi -\alpha)\left(\dst\frac1p-r-1\right)\not=0.
\end{eqnarray}}
In details the condition \eqref{e0.16} is written as follows:
{
\begin{subequations}
\begin{eqnarray}\label{e0.16a}
&&\hskip-17mm 1)\;\; r\not=\dst\frac1p-n\quad \text{and}\quad r\not=\dst\frac n4;\\
\label{e0.16b}
&&\hskip-17mm 2)\;\;\text{If}\quad r=\frac1p-n,\;\; n\not=0,\quad
\text{then}\quad\alpha\not=\frac{2k+1}{2n}\pi,\\
\label{e0.16c}
&&\hskip-17mm 3)\;\;\text{If}\;\;r=\dst\frac n2,\;\;\text{then}\;\; p\not=\dst\frac2{n+2m+2}\;\; \text{or}\;\;\alpha\not=\dst\frac{2k+1}{2m}\pi,\\
\label{e0.16d}
&&\hskip-19mm \begin{array}{l}
4)\;\;\text{If}\quad r=\dst\frac{2n+1}4, \quad\text{then}\\
\quad\; \alpha\not=2\pi p\dst\frac{2k-1}{4-2np-3p}\quad \text{and}
\quad \alpha\not=2\pi\frac{2-np-2kp}{2-2np-3p},
\end{array}
\end{eqnarray}
\end{subequations}}
where $k,m,n=0,\pm1,\ldots.$
The system \eqref{e0.13} has the unique pair of solutions in the space settings \eqref{e0.15a} and \eqref{e0.15b} if:
\begin{eqnarray}
\begin{array}{l} \label{e0.17}
-\dst\frac34<r\leqslant\dst\frac1p-1 \quad\text{for}\quad
1<p\leqslant2,\\
{\dst\frac1p-1\leqslant r<-\dst\frac14\quad\text{for}}\quad 2\leqslant p<\infty.
\end{array}
\end{eqnarray}
\end{theorem}
The proof of the Theorem \ref{t0.6} is exposed in \S\, \ref{sect5}.
Investigations of the boundary integral equations run into difficulties due to the absence of results on Mellin convolution equations \eqref{e0.13} in the Sobolev-Slobode\v{c}kii \eqref{e0.15a} and the Bessel potential \eqref{e0.15b} space settings. In the present paper we apply the results on Mellin convolution equations with meromorphic kernels in the Bessel potential and Sobolev-Slobode\v{c}kii spaces obtained recently by R. Duduchava \cite{Du15} and V. Didenko \& R. Duduchava \cite{DD16}. We write explicitly the symbol $\cM^s_{\alpha,p}(\omega)$ of the corresponding operator as a function on the infinite rectangle $\mathfrak{R}$, {and this symbol is responsible, as usual,} for the Fredholm property and the index of the operator.
Major contribution to BVPs for elliptic equations in two and multidimensional domains with edges and cones on the boundary was made by V. Kondratjev by his celebrated paper \cite{Ko67}. The method is based on Mellin transformation and allows to find asymptotic of solutions. The approech was very popular and used intensively in the literature, see papers and monographs by P. Grisward \cite{Gr85}, M. Dauge \cite{Da88}, V. Kozlov, V. Mazya, J. Rossman \cite{KMR01}, B.W. Schulze \cite{Sc92} and many others. The investigations are mostly performed in special Kondratjev's weighted spaces, adapted to the geometry of domains with singularities.
In the recent papers \cite{CK08,CK10,CK13,CK15} L. Castro \& D. Kapanadze reduce BVP \eqref{e0.4} in the $\bH^{1+\ve}(\Omega_\alpha)$ space settings to an equivalent equation with Wiener-Hopf $\pm$ Hankel operators, by manipulating with the even and odd extensions and the reflection operators. The obtained equations were investigated in $\bL_2(\bR^+)$ space and, in the last paper \cite{CK15}, in the special potential space, defined by the Mellin {transformations.}
In a series of papers \cite{Kru98,Kru01,Kru07,Kru09} P.A. Krutitskii investigated Boundary value problems for the Helmholtz equation in a planar 2D unbounded domain $\Omega$ outer to a finite number of finite domains and cuts with different boundary conditions. Unique solvability was proved in classical strong setting $u\in C^1(\overline{\Omega})\cap C^2(\Omega)$.
Rigorous analytical solution of the model boundary value problems with
different boundary conditions is crucial for understanding elliptic boundary value problems in Lipschitz domains (see \cite{KS03,KMR01,No58} and \cite{Me87} for the physical background and early references). In \cite{BDKT13} is described how the modern localization technique can be applied to the investigation {of BVPs in domains} with the Lipschitz boundary by reducing {them to several local} Boundary
value problems in model domains.
Model BVPs for rational angles in the classical setting are solved explicitly in \cite{ENS13a,ENS13b}. Other known results are either very limited to special situations such as the rectangular case \cite{CST04,CST06,MPST93} or rather complicated in what concerns the analytical methods \cite{KMM05,ZM00} or not describing appropriate function spaces, see, e.g., \cite{Ma59,Uf03}. For the historical survey and for further references we recommend \cite{CK13,ZM00,Va00}.
Yet another approach, which can also be applied, is the limiting absorption principle, which is based on variational formulation and Lax-Milgram Lemma and its generalizations (see e.g., in \cite{BT01,BCC12a,BCC12b}) but, again, BVPsa are considered in the classical setting $p=2$ only.
In the 1960’s there was suggested to solve canonical diffraction problems in Sobolev spaces, based on results on pseudodifferential equations in domains with corners and, more generally, in Lipschitz domains (see papers of E. Meister \cite{Me85,Me87}, E. Meister and F.-O. Speck \cite{MS79}, W.L. Wendland \cite{WSH79}, A. Ferreira dos Santos \cite{ST89} etc.) In the book of Vasil’ev \cite{Va00} one find a considerable list of references.
There are many other papers where concrete diffraction problems are studied. We confine ourselves only with the rederence to some of them: \cite{CK06}, \cite{CK08}, \cite{CK10} \cite{CST03}, \cite{Kru01}, \cite{Kru09}, \cite{KMM05}, \cite{Ma59}, \cite{MPST93}, \cite{MPST98}, \cite{MR96}, \cite{MSV11}, \cite{MSB11}, \cite{MSS97}, \cite{ZM00}.
\section{Boundary poseudodifferential operators}
\label{Sec1}
\setcounter{equation}{0}
Let $\cH_k(x)$ be the fundamental solution to the Helmholtz equation
\begin{eqnarray}\label{e1.2}
\Delta\cH_k(x) + k^2\cH_k(x)=\delta(x),\qquad x\in\bR^2,
\end{eqnarray}
which coincides with the Hankel function of the first kind and order $0$
$\cH_k(x)=\dst\frac14H^{(1)}_0(k|x|)$. The Hankel function decays
exponentially at the infinity and has the following asymptotic
{(see \cite{GR94,Kru98}):}
\begin{eqnarray}\label{e1.4}
H^{(1)}_0(|z|)=\dst\frac2{\pi}\ln\,|z|+{\rm const} { + \cO(|z|^2\ln\,|z|)
\quad\hbox{\rm as}}\quad |z|\to0.
\end{eqnarray}
It is important to note, that {the asymptotic equality \eqref{e1.4} remains valid after taking any finite number of derivatives.}
Consider standard layer potential operators on the model domain $\Omega_\alpha$, the Newton, the Single and the Double layer potentials respectively
\begin{eqnarray}\label{e1.1}
\begin{array}{l}
\N_{\Delta+k^2}\vf(x):=\dst\int_{\Omega_\alpha}
\cH_k(x-y)\vf(y)\,dy,\\[3mm]
\V_{\Delta+k^2}\vf(x):=\dst\int_{\Gamma_\alpha}\cH_k(x-\tau)\vf(\tau)d\sigma,\\[3mm]
\W_{\Delta+k^2}\vf(x):=\dst\int_{\Gamma_\alpha}\pa_{\nub(\tau)}\cH_k(x-\tau)\vf(\tau)d\sigma,
\qquad x\in\Omega_\alpha.
\end{array}
\end{eqnarray}
The potential operators, defined above, have standard boundedness properties in the Bessel potential spaces (see, e.g., \cite{DNS95,Du01,HW08}):
\begin{eqnarray}\label{e1.5}
\begin{array}{rcl}
\N_{\Delta+k^2} &:&\bH_p^s(\Omega_\alpha)\longrightarrow
\bH^{s+2}_p(\Omega_\alpha),\qquad s\in\bR,\quad 1<p<\infty,\\
\V_{\Delta+k^2} &:&\bH_p^r(\Gm_\alpha)\longrightarrow
\bH^{r+1+\frac1p}_p(\Omega_\alpha),\\
\W_{\Delta+k^2} &:&\bH_p^r(\Gm_\alpha)\longrightarrow
\bH^{r+\frac1p}_p(\Omega_\alpha),\qquad \dst\frac1p-2<r<\dst\frac1p+1,\quad 1<p<\infty.
\end{array}
\end{eqnarray}
It is well known that any solution $u\in\bH^1(\Omega_\alpha)$ to the BVP \eqref{e0.4} in the space is represented as {follows
\begin{eqnarray}\label{e2.1}
u(x)=\N_{\Delta+k^2}f(x)+\W_{\Delta+k^2}u^+(x)
-\V_{\Delta+k^2}[\pa_\nub u]^+(x)\qquad x\in\Omega_\alpha,
\end{eqnarray}
(see \cite{DNS95,Du01}) where densities represent} the Dirichlet $u^+$ and the Neumann $[\pa_\nub u]^+$ traces of the solution $u$ on the boundary.
Let us remind the Plemelji formulae
\begin{eqnarray}\label{e2.2}
\begin{array}{rcl}
(\W_{\Delta+k^2}\vf)^\pm(t)=\pm\dst\frac12\vf(t)+\W_{{\Delta+k^2},0}\vf(t),\\[3mm]
(\pa_\nub\V_{\Delta+k^2}\psi)^\pm(t)=\mp\dst\frac12
\psi(t)+\W^*_{{\Delta+k^2},0}\psi(t),\\[3mm]
(\pa_\nub\W_{\Delta+k^2}\psi)^\pm(t)=\V_{{\Delta+k^2},+1}\psi(t),\\[3mm]
(\V_{\Delta+k^2}\vf)^\pm(t)=\V_{{\Delta+k^2},-1}\vf(t) \qquad t\in\Gamma_\alpha:=\pa\Omega_\alpha,
\end{array}
\end{eqnarray}
where the pseudodifferential operators ($\Psi$DO)
\begin{eqnarray}\label{e2.3}
\begin{array}{l}
\V_{\Delta+k^2,-1}\vf(t):=\dst\int_{\Gamma_\alpha}\cH_k(t-\tau)\vf(\tau)d\sigma,\\[3mm]
\W_{\Delta+k^2,0}\vf(t):=\dst\int_{\Gamma_\alpha}\pa_{\nub(\tau)}\cH_k(t-\tau)\vf(\tau)d\sigma,\\[3mm]
\W^*_{\Delta+k^2,0}\vf(t):=\dst\int_{\Gamma_\alpha}\pa_{\nub(t)}\cH_k(t-\tau)\vf(\tau)d\sigma,\\[3mm]
\V_{\Delta+k^2,+1}\vf(t):=\dst\int_{\Gamma_\alpha}\pa_{\nub(t)}\pa_{\nub(\tau)}\cH_k(t-\tau)
\vf(\tau)d\sigma, \qquad t\in\Gamma_\alpha
\end{array}
\end{eqnarray}
of orders $-1$, $0$, $0$ and $+1$, are associated with the layer potentials of the Helmholtz equation. Due to the asymptotic \eqref{e1.4}, the operator $\V_{\Delta+k^2,-1}$ has weakly singular kernel and the integral exists in the Lebesgue sense, while the operators $\W_{\Delta+k^2,0}$ and $\W^*_{\Delta+k^2,0}$ have singular kernel of order $-1$ and the integrals exists in the Cauchy Mean Value sense.
To explain in which sense is understood the hypersingular integral operator $\V_{\Delta+k^2,+1}\vf(t)$, let us recall the following equality
\begin{eqnarray}\label{e2.4a}
\Delta+k^2=\partial^2_1+\partial^2_2+k^2=\partial^2_\nub+\partial^2_\ell+k^2,
\end{eqnarray}
{
where $\nub(t)$ is the unit normal vector field, $\partial_\nub$ is the normal derivative (see \eqref{e0.1}--\eqref{e0.3}),
From \eqref{e1.2} and \eqref{e2.4a} follows the equality
\[
\delta=(\Delta+k^2)\cH_k=\partial^2_\nub\cH_k +(\partial^2_\ell+k^2)\cH_k,
\]
which we use to prove the following:
\begin{eqnarray}\label{e2.4c}
\partial_{\nub(x)}\partial_{\nub(y)}\cH_k(x-y)=-\partial^2_{\nub(y)}\cH_k(x-y)
=-\delta(x-y)+(\partial^2_{\ell(y)}+k^2)\cH_k(x-y).
\end{eqnarray}
Due to the equality \eqref{e2.4c} and the integration by parts formula for the tangential differential operator (see \cite{Du01,DMM06})
\begin{eqnarray}\label{e1.10}
\dst\int_{\Gamma_\alpha}\pa_{\ell(\tau)}\psi(\tau)\vf(\tau)d\sigma =-\dst\int_{\Gamma_\alpha}\psi(\tau)\pa_{\ell(\tau)}\vf(\tau)d\sigma.
\end{eqnarray}
the hypersingular operator $\V_{\Delta+k^2,+1}$ is represented as
\begin{eqnarray}\label{e2.5}
\V_{\Delta+k^2,+1}\vf(t)&\hskip-3mm:=&\hskip-3mm\dst\int_{\Gamma_\alpha}
\pa_{\nub(t)}\pa_{\nub(\tau)}\cH_k(t-\tau)\vf(\tau)d\sigma\nonumber\\
&\hskip-3mm=&\hskip-3mm-\vf(t)+\dst\int_{\Gamma_\alpha}(\pa^2_{\ell(\tau)}
+k^2)\cH_k(t-\tau)\vf(\tau)d\sigma\nonumber\\
&\hskip-3mm=&\hskip-3mm-\vf(t)-\dst\int_{\Gamma_\alpha}\pa_{\ell(\tau)}
\cH_k(t-\tau)\pa_{\ell(\tau)}\vf(\tau)d\sigma \nonumber\\
&&+k^2\dst\int_{\Gamma_\alpha}\cH_k(t-\tau)\vf(\tau)d\sigma, \quad
t\in\Gamma_\alpha.
\end{eqnarray}
Accoding to the obtained equality \eqref{e2.5} the operator $\V_{\Delta+k^2,+1}$ is a sum of singular integral operator applied to the tangential derivative $\pa_\ell\vf$ of the density and the regular integral applied to $\vf$ itself.
The following pseudodifferential operators
\begin{eqnarray}\label{e2.6}
\begin{array}{l}
\V_{\Delta,-1}\vf(t):=\dst\frac1{2\pi}\int_{\Gamma_\alpha}\ln|t-\tau|\vf(\tau)d\sigma,\\[3mm]
\W_{\Delta,0}\vf(t):=\dst\frac1{2\pi}\dst\int_{\Gamma_\alpha}\pa_{\nub(\tau)}\ln|t-\tau|\vf(\tau)d\sigma,\\[3mm]
\W^*_{\Delta,0}\vf(t):=\dst\frac1{2\pi}\dst\int_{\Gamma_\alpha}\pa_{\nub(t)}\ln|t-\tau|\vf(\tau)d\sigma,\\[3mm]
\V_{\Delta,+1}\vf(t):=\dst\frac1{2\pi}\dst\int_{\Gamma_\alpha}\pa_{\nub(t)}\pa_{\nub(\tau)}\ln|t-\tau|
\vf(\tau)d\sigma, \qquad t\in\Gamma_\alpha
\end{array}
\end{eqnarray}
of orders $-1$, $0$, $0$ and $+1$, are associated with the Laplace equation (see \cite{DNS95,Du01,HW08})
\begin{eqnarray*}
\Delta u(x) = 0, \qquad x\in\bR^2.
\end{eqnarray*}
which has the logarithmic fundamental solution
\begin{eqnarray*}
\cK_\Delta(x):=\frac1{2\pi}\ln|x|,\qquad \Delta\cK_\Delta(x) = \delta(x), \qquad x\in\bR^2,
\end{eqnarray*}
The pseudodifferential operators defined above have standard mapping properties (see \cite{DNS95,Du01,HW08}):
\begin{eqnarray}\label{e2.7}
\begin{array}{rcl}
\V_{\Delta+k^2,-1},\ \V_{\Delta,-1} &:&\bH_p^s(\Gm_\alpha)\longrightarrow
\bH^{s+1}_p(\Gm_\alpha),\\[3mm]
\W_{\Delta+k^2,0},\ \W_{\Delta,0} &:&\bH_p^s(\Gm_\alpha)\longrightarrow
\bH^s_p(\Gm_\alpha),\\[3mm]
\W^*_{\Delta+k^2,0},\ \W^*_{\Delta,0} &:&\bH_p^s(\Gm_\alpha)\longrightarrow
\bH^s_p(\Gm_\alpha),\\[3mm]
\V_{\Delta+k^2,+1},\ \V_{\Delta,+1} &:&\bH_p^s(\Gm_\alpha)\longrightarrow
\bH^{s-1}_p(\Gm_\alpha)
\end{array}
\end{eqnarray}
for $1<p<\infty$, $\dst\frac1p-2<r<\dst\frac1p+1$.
\begin{lemma}[Lemma 1.1, \cite{DD16}]\label{l1.1}
Let $1<p<\infty$, $\dst\frac1p-2<r<\dst\frac1p+1$ and either $\bR_0=\bR^+$, $\bR_1=\bR_\alpha$ or vice versa $\bR_0=\bR_\alpha$, $\bR_1=\bR^+$. Let, respectively, $r_j\;:\;\bH_p^s(\Gamma_\alpha)\to\bH_p^s(\bR_j)$, $j=0,1$, be the corresponding restriction operators. Then the differences
\begin{eqnarray}\label{e2.7a}
\begin{array}{rcl}
\T_1:=r_1[\V_{\Delta+k^2,-1} - \V_{\Delta,-1}]r_0 &:&\wt{\bH}_p^s(\bR_0)\longrightarrow
\bH^{s+1}_p(\bR_1),\\[3mm]
\T_2:=\W_{\Delta+k^2,0} - \W_{\Delta,0} &:&\bH_p^s(\Gm_\alpha)\longrightarrow
\bH^s_p(\Gm_\alpha),\\[3mm]
\T_3=\T_2^*:=\W^*_{\Delta+k^2,0} - \W^*_{\Delta,0} &:&\bH_p^s(\Gm_\alpha)\longrightarrow
\bH^s_p(\Gm_\alpha),\\[3mm]
\T_4:=r_1[\V_{\Delta+k^2,+1} - \V_{\Delta,+1}]r_0 &:&\wt{\bH}_p^{s+1}(\bR_0)\longrightarrow
\bH^s_p(\bR_1)
\end{array}
\end{eqnarray}
are locally compact operators: the operators $v\T_j$, $j=1,2,3,4$, are compact for arbitrary function $v\in C^\infty_0(\Gamma_\alpha)$ with a compact support.
\end{lemma}
Next we will write {some pseudodifferential} operators (PsDOs) in \eqref{e2.6} in explicit form for the later use in \S\, 3.. First let us consider the PsDOs $r_{\bR^+}\V_{\Delta,+1}r_{\bR_\alpha}$. By applying the equality
\[
\partial_{\nub(x)}\partial_{\nub(y)}\ln|x-y|=-\partial^2_{\nub(y)}\ln|x-y|
=-\delta(x-y)+\partial^2_{\ell(y)}\ln|x-y|,
\]
proved similarly to \eqref{e2.5}, we get:
\begin{eqnarray*}
&&\V_{\Delta,+1}\vf(t):=\dst\int_{\Gamma_\alpha}\pa_{\nub(t)}\pa_{\nub(\tau)}
\ln|t-\tau|\vf(\tau)d\sigma=-\vf(t)+\dst\int_{\Gamma_\alpha}\pa^2_{\ell
(\tau)}\ln|t-\tau|\vf(\tau)d\sigma\nonumber\\
&&\hskip30mm=-\vf(t)-\dst\int_{\Gamma_\alpha}\pa_{\ell(\tau)}\ln|t-\tau|
\pa_{\ell(\tau)}\vf(\tau)d\sigma, \quad t\in\Gamma_\alpha.
\end{eqnarray*}
By using the parametrization $x=(x_1,x_2)^\top=(t,0)^\top$ of {$\bR^+$, the} parametrization $y=(y_1,y_2)^\top=(\tau\cos\,\alpha,\tau\sin\,\alpha)^\top$ of $\bR_\alpha$, recalling that $\bR_\alpha$ is oriented from $-\infty$ to $0$ and using the equality \eqref{e0.3}, equalities
\begin{eqnarray}\label{e1.16}
{\pa_{\ell(y)}}=-\cos\,\alpha\,\pa_{y_1}-\sin\,\alpha\,\pa_{y_2},\quad
\ln|x-y|=\frac12\ln\big[(x_1-y_1)^2 + (x_2-y_2)^2\big]
\end{eqnarray}
for $t\in\bR^+$, $y\in\Gamma_\alpha$, we proceed as follows:
\begin{align*}
\pa_{\nub(t)} & \pa_{\nub(y)}\cK_\Delta(t\!-\!y)=\pa_{\nub(x)}
\pa_{\nub(y)}\cK_\Delta(x\!-\!y)\Big|_{\begin{array}{l}\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}}\\
&=-\!\pa_{x_2}(-\sin\,\alpha\,\pa_{y_1}+\cos\,\alpha\,\pa_{y_2})
\cK_\Delta(x\!-\!y)\Big|_{\begin{array}{l}\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}}\\
&=\left\{-\sin\,\alpha\,\pa_{y_1}\pa_{y_2}+\cos\,\alpha\,
\pa_{y_2}^2\right\}\cK_\Delta(x\!-\!y)\Big|_{\begin{array}{l}
\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}}\\
&=\left[\cos\,\alpha\Delta\cK_\Delta(x-y)-\pa_{y_1}\left\{\cos\,\alpha\,
\pa_{y_1}+\sin\,\alpha\,\pa_{y_2}\right\}\cK_\Delta(x-y)
\right]\Big|_{\begin{array}{l}\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}}\\
&=\left[\cos\,\alpha\delta(x-y)+\pa_{y_1}\pa_{\ell(y)}\cK_\Delta(x-y)\right]
\Big|_{\begin{array}{l}\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}}\\
&=\left[\cos\,\alpha\,\delta(0)+\frac1{4\pi}\pa_{\ell(y)}\pa_{y_1}
\ln\big[(x_1-y_1)^2 + (x_2-y_2)^2\big]\right]
\Big|_{\begin{array}{l}\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}}\\
&=\left[\cos\,\alpha\,\delta(0)-\frac1{2\pi}\pa_{\ell(y)}\frac{x_1-y_1}
{2\pi[(x_1-y_1)^2 + (x_2-y_2)^2]}\right]
\Big|_{\begin{array}{l}\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}}
\end{align*}
Now integrating by parts (see \eqref{e1.10}) we continue as follows:
\begin{align}\label{e1.17}
r_{\bR^+}\V_{\Delta,+1}r_{\bR_\alpha}v(t)&=\frac1{2\pi}
r_{\bR^+}\dst\int_{\bR_\alpha}\frac{(x_1-y_1)\pa_{\ell(y)}v(y)d\sigma}
{(x_1-y_1)^2 + (x_2-y_2)^2} \Big|_{\begin{array}{l}\scriptstyle x=(t,0)\\[-2mm]
\scriptstyle y=(\tau\cos\,\alpha,\tau\sin\,\alpha)\end{array}} \hskip-10mm\nonumber\\
&=-\frac1{2\pi}\dst\int_0^\infty\frac{t-\tau\cos\,\alpha}
{t^2 + \tau^2 - 2t\tau\,\cos\alpha}(\J_\alpha\pa_\ell v)(\tau)d\tau
\nonumber\\
&=-\frac1{4\pi}\int_0^\infty\left[\dst\frac1{
t-e^{i\alpha}\tau}+\frac1{t-e^{-i\alpha}\tau}\right]
(\J_\alpha\pa_\ell v)(\tau)d\tau,\nonumber\\
&=\frac14\left[\K_{e^{i\alpha}}
+\K_{e^{-i\alpha}}\right]\pa_\tau v_1(t),\qquad t\in\bR^+,
\end{align}
since $(\J_\alpha\pa_\ell v)(\tau)=-(\pa_\tau v_1)(\tau)$, where $v_1:=\J_\alpha v$.
The formula
\[
\J_\alpha r_{\bR_\alpha}\V_{\Delta,+1}r_{\bR^+}w(t)
=-\frac14\left[\K_{e^{i\alpha}}+\K_{e^{-i\alpha}}
\right](\pa_\tau w)(t), \quad t\in\bR^+
\]
is proved similarly.
Now we look to the singular integral operators $r_{\bR^+}\pa_\ell\V_{\Delta,-1}r_{\bR_\alpha}$ and $r_{\bR_\alpha}\pa_\ell\V_{\Delta,-1}r_{\bR^+}$. We proceed as in \eqref{e1.17}:
\begin{eqnarray}\label{e1.19}
\J_\alpha r_{\bR_\alpha}\pa_\ell\V_{\Delta,-1}r_{\bR^+}w(t)&\hskip-3mm=&\hskip-3mm\frac1{2\pi}\J_\alpha
r_{\bR_\alpha}\dst\int_{\bR^+}\pa_{\ell(x)}\ln|x-y|\,w(y)d\sigma\nonumber\\
&\hskip-3mm=&\hskip-3mm-\frac1{2\pi}\J_\alpha r_{\bR_\alpha}\dst\int_{\bR^+}\frac{\cos\,\alpha(x_1 - \tau)
+x_2\sin\,\alpha}{(x_1-\tau)^2 + x^2_2}w(\tau)\,d\tau\nonumber\\
&\hskip-3mm=&\hskip-3mm-\frac1{2\pi}\dst\int_0^\infty\frac{\cos\,\alpha(t\cos\,\alpha-\tau) + t\,\sin^2\alpha}
{(t\,\cos\alpha-\tau)^2 + t\sin^2\alpha}w(\tau)d\tau\nonumber\\
&\hskip-3mm=&\hskip-3mm-\frac1{2\pi}\dst\int_0^\infty\frac{t-\tau\,\cos\alpha}
{(t\,\cos\alpha - \tau)^2 + t^2\sin^2\alpha}w(\tau)d\tau\nonumber\\
&\hskip-3mm=&\hskip-3mm-\frac1{4\pi}\int_0^\infty\left[\dst\frac1{t-e^{i\alpha}\tau}
+\frac1{t-e^{-i\alpha}\tau}\right]w(\tau)d\tau,\nonumber\\
&\hskip-3mm=&\hskip-3mm-\frac14\left[\K_{e^{i\alpha}}+\K_{e^{-i\alpha}}\right]w(t),
\quad t\in\bR^+.
\end{eqnarray}
The formulae
\begin{eqnarray}\label{e2.12a}
&&\hskip-20mm r_{\bR^+}\pa_\ell\V_{\Delta,-1}r_{\bR_\alpha}w(t)
=\frac14\left[\K_{e^{i\alpha}}+\K_{e^{-i\alpha}}
\right]\J_\alpha w(\tau), \quad t\in\bR^+
\end{eqnarray}
is proved similarly.
For the singular integral operator $\W_{\Delta,0}$ we proved the following:
\begin{eqnarray}\label{e2.8a}
&&\hskip-20mm r_{\bR^+}\W_{\Delta,0}r_{\bR_\alpha}\vf(t)
=- \J_\alpha r_{\bR_\alpha}\W_{\Delta,0}r_{\bR^+}\vf(t)
=\frac1{4i}\left[e^{i\alpha}\K_{e^{i\alpha}}-e^{-i\alpha}
\K_{e^{-i\alpha}}\right]\vf_1(t),\\
\label{e2.8b}
&&\hskip-20mm r_{\bR^+}\W_{\Delta+k^2}r_{\bR^+}=r_{\bR^+}\W_{\Delta}r_{\bR^+}
=r_{\bR_\alpha}\W_{\Delta+k^2}r_{\bR_\alpha}
=r_{\bR_\alpha}\W_{\Delta}r_{\bR_\alpha}=0,\\
&&\hskip60mm \vf_1(t):=(\J_\alpha\vf)(t), \qquad t\in\bR^+, \nonumber
\end{eqnarray}
where $\J_\alpha$ is the pull back operator (see \eqref{e0.8}) and
$r_{\bR^+}$ and $r_{\bR_\alpha}$ are the restriction operators to the
spaces on the corresponding subsets $\bR^+$ and $\bR_\alpha$. {We drop the proofs of \eqref{e2.8a} and \eqref{e2.8b} because these formulae are not applied in the present manuscript.}
For the dual operator $\W^*_{\Delta,0}$ we get:
\begin{eqnarray}\label{e2.9a}
&&\hskip-15mm r_{\bR_\alpha}\W^*_{\Delta,0}r_{\bR^+}\vf(t)=- \J_\alpha
r_{\bR^+}\W^*_{\Delta,0}r_{\bR^+}\vf(t)
=\frac1{4i}\left[\K_{e^{i\alpha}}-\K_{e^{-i\alpha}}\right]\vf(t),
\quad t\in\bR^+,\\
\label{e2.9b}
&&\hskip-15mm r_{\bR^+}\W^*_{\Delta+k^2}r_{\bR^+}=r_{\bR^+}
\W^*_{\Delta}r_{\bR^+}=r_{\bR_\alpha}\W^*_{\Delta+k^2}r_{\bR_\alpha}
=r_{\bR_\alpha}\W^*_{\Delta}r_{\bR_\alpha}=0.
\end{eqnarray}
\section{Boundary integral equation of the model problem}
\label{sect3}
\setcounter{equation}{0}
{\bf Proof of Theorem \ref{t0.5}:} The boundary data $g\in\bW^{s-1/p}_p(\bR_\alpha)$ and $h\in\bW^{s-1-1/p}_p(\bR^+)$ of the BVP \eqref{e0.4} in the non-classical formulation \eqref{e0.6} are defined initially on the parts of the boundary $\bR_\alpha$ and $\bR^+$, respectively. Let $g_0\in\bW^{s-1/p}_p(\Gamma_\alpha)$ and $h_0\in\bW^{s-1-1/p}_p(\Gamma_\alpha)$ be some fixed extensions of these boundary data to the entire boundary $\Gamma_\alpha=\bR^+\cup\bR_\alpha$. We remind, that the spaces $\wt{\bW}^{s-1/p}_p(\bR_\alpha)$ and $\wt{\bW}^{s-1/p}_p(\bR^+)$ are subsets of $\bW^{s-1/p}_p(\Gamma_\alpha)$ and functions from $\wt{\bW}^s_p(\bR^+)$ and $\wt{\bW}^s_p(\bR_\alpha)$ are extended by $0$ to $\bR_\alpha$ and to $\bR^+$, respectively. The difference between of two such extensions belong to the spaces $\wt{\bW}^{s-1/p}_p(\bR_\alpha)$ and $\wt{\bW}^{s-1-1/p}_p(\bR^+)$ respectively. Therefore, we should look for two unknown functions $\varphi_0\in\wt{\bW}^{s-1/p}_p(\bR^+)$ and $\psi_0\in\wt{\bW}^{s-1-1/p}_p(\bR_\alpha)$, such that for $g_0+\varphi_0$ and $h_0+\psi_0$ the boundary conditions in \eqref{e0.4} hold on the entire boundary. i.e., for any solution $u(x)$ to the BVP \eqref{e0.4}, there holds
\begin{eqnarray}\label{e3.1}
\begin{array}{c}
u^+(t)=g_0(t)+\varphi_0(t)=\left\{\begin{array}{ll} g_0(t)+\varphi(t)\quad &{\rm if}\quad
t\in\bR^+,\\[3mm]
g_0(t)\quad &{\rm if}\quad t\in\bR_\alpha,
\end{array}\right.\\ \\
(\pa_\nub u)^+(t)=h_0(t)+\psi_0(t)=\left\{\begin{array}{ll} h_0(t)
\quad &{\rm if}\quad t\in\bR^+,\\[3mm]
h(t)+\psi_0(t)\quad &{\rm if}\quad t\in\bR_\alpha.
\end{array}\right.
\end{array}
\end{eqnarray}
By introducing the boundary values of a solution \eqref{e3.1} of the BVP \eqref{e0.4} into the representation formula \eqref{e2.1} we get the
following representation of a solution:
\begin{eqnarray}\label{e3.2}
u(x)=\N_{\Delta+k^2} f(x)+\W_{\Delta+k^2}[g_0+\varphi_0](x)-\V_{\Delta+k^2}[h_0+\psi_0](x),\qquad x\in\cC.
\end{eqnarray}
The known and unknown functions in \eqref{e3.1} and \eqref{e3.2} belong to the following spaces
\begin{eqnarray}\label{e3.3}
g_0\in\bW^{s-1/p}_p(\Gamma_\alpha), \quad h_0\in\bW^{s-1-1/p}_p(\Gamma_\alpha), \quad
\varphi_0\in\wt{\bW}^{s-1/p}_p(\bR^+), \quad \psi_0\in\wt{\bW}^{s-1-1/p}_p(\bR_\alpha).
\end{eqnarray}
By applying the boundary conditions from \eqref{e0.4} to \eqref{e3.2} and the Plemelji formulae \eqref{e2.2} we get the following:
\[
\begin{array}{r}
\left\{\begin{array}{l}
g_0(t)+\varphi_0(t)=u^+(t)={(\N_{\Delta+k^2} f)^+}+\dst\frac12(g_0(t)+\varphi_0(t))\\
\hskip20mm+\W_{\Delta+k^2,0}[g_0+\varphi_0](t)-\V_{\Delta+k^2,-1}[h_0+\psi_0](t),\\[2mm]
h_0(t)+\psi_0(t)=(\pa_\nub u)^+(t)={(\pa_\nub \N_{\Delta+k^2} f)^+}+\V_{\Delta+k^2,+1}[g_0+\varphi_0](t)\\
\hskip20mm+\dst\frac12(h_0(t) +\psi_0(t))-\W^*_{\Delta+k^2,0}[h_0+\psi_0](t),\qquad t\in\Gamma.
\end{array}\right.
\end{array}
\]
Rearranging the known and unknown functions the system acquires the following form
\begin{eqnarray}\label{e3.4}
\left\{\begin{array}{ll}\dst\frac12\varphi_0 - \W_{\Delta+k^2,0}r_{\bR^+}
\varphi_0 + \V_{\Delta+k^2,-1}r_{\bR_\alpha}\psi_0=G_0,\\[3mm]
\dst\frac12\psi_0 + \W^*_{\Delta+k^2,0}r_{\bR_\alpha}\psi_0
-\V_{\Delta+k^2,+1}r_{\bR^+}\varphi_0=H_0
\qquad&\text{on}\quad\Gamma_\alpha=\partial\Omega_\alpha, \end{array}\right.
\end{eqnarray}
where $G_0$ and $H_0$ are exposed in \eqref{e0.13} and we used the properties $r_{\bR^+}\varphi_0=\varphi_0$, $r_{\bR_\alpha}\psi_0=\psi_0$.
By applying the restriction $r_{\bR^+}$ to the both parts of the first equation in \eqref{e3.4} and the restriction $r_{\bR_\alpha}$ to the second one and by recalling the equalities $r_{\bR^+}\W_{\Delta+k^2}r_{\bR^+}
=r_{\bR_\alpha}\W^*_{\Delta+k^2}r_{\bR_\alpha}=0$ (cf. \eqref{e2.8a}, \eqref{e2.9a}) we arrive to the system \eqref{e0.13}.
There is the full equivalence between the solvability of the system \eqref{e3.4} and the sol\-vability of the BVP \eqref{e0.4}, given by the representation formula \eqref{e3.2}. Therefore the unique solvability of the BVP \eqref{e0.4} implies the unique solvability of the system \eqref{e3.4} and vice versa, The concluding assertion of Theorem \ref{t0.5} follows then from Theorem \ref{t0.1}. \QED
In the formulation and the proof of the next Lemma \ref{l3.1} we use
localization and quasi-localization principle. A quasi-localization means "freezing coefficients" and changing underling contours and surfaces by an isomorphic but simpler ones. For details of a quasi-localization we refer the reader to the papers \cite{Si65} and \cite{CDS03}, where the quasi-localization is well described for singular integral operators and for BVPs, respectively. We also refer to \cite[\S\, 3]{Du15}, where is exposed a short introduction to quasi-localization.
Let us agree to understand under local equivalence and local quasi-equivalence of equations the local equivalence of the corresponding operators in the corresponding spaces.
\begin{lemma}\label{l3.1}
System of the pseudodifferential equation \eqref{e0.13} and the system
\begin{eqnarray}\label{e3.6}
\left\{\begin{array}{ll}\dst\frac12\varphi_+(t) + r_{\bR^+}\V_{\Delta,-1}
r_{\bR_\alpha}\psi_-(t)=G_1(t),\qquad & t\in\bR^+,\\[3mm]
\dst\frac12\psi_-(t) - r_{\bR_\alpha}\V_{\Delta,+1}r_{\bR^+}\varphi_+(t)
=H_1(t),\qquad & t\in\bR_\alpha, \end{array}\right. \\[3mm]
\varphi_+\in\wt{\bW}^{s-1/p}_p(\bR^+), \qquad
\psi_-\in\wt{\bW}^{s-1-1/p}_p(\bR_\alpha),\nonumber\\[2mm]
G_1\in\bW^{s-1/p}_p(\bR^+), \qquad
H_1\in\bW^{s-1-1/p}_p(\bR_\alpha).\nonumber
\end{eqnarray}
are locally equivalent at $0$.
At any other point $x\in\Gamma_\alpha\setminus{0}\cup\{+\infty,e^{i\alpha}\infty\}$, including the both infinity points $+\infty$ and $x=e^{i\alpha}\infty:=\lim_{y\to+\infty}
e^{-i\alpha}y$, the system \eqref{e0.13}
is locally quasi-equivalent to the trivial system
\begin{eqnarray}\label{e3.7}
\begin{array}{l}
\dst\frac12\varphi=H_2, \quad \varphi,H_2\in\bW^{s-1/p}_p(\bR)
\qquad \text{for}\quad x\in\bR^+,\\[4mm]
\dst\frac12\psi=G_2, \quad \psi,G_2\in\bW^{s-1/p-1}_p(\bR)\qquad
\text{for}\quad x\in\bR_\alpha.
\end{array}
\end{eqnarray}
\end{lemma}
{\bf Proof:} The systems \eqref{e0.13} and \eqref{e3.6} are locally equivalent at $0$, because the differences
\[
\begin{array}{rcl}
\T_1:=r_{\bR^+}[\V_{\Delta+k^2,-1} - \V_{\Delta,-1}]r_{\bR_\alpha}&:&
\wt{\bW}_p^r(\bR_\alpha)\longrightarrow\bW^{r+1}_p(\bR^+),\\[3mm]
\T_4:=r_{\bR_\alpha}[\V_{\Delta+k^2,+1} - \V_{\Delta,+1}]r_{\bR^+}
&:&\wt{\bW}_p^{r+1}(\bR^+)\longrightarrow\bW^r_p(\bR_\alpha)
\end{array}
\]
are locally compact for all $r\in\bR$ due to Lemma \ref{l1.1} and compact operators are locally equivalent to $0$.
Now let us describe the local quasi-equivalent systems of \eqref{e0.13} at $x\in\bR^+\cup\{+\infty\}$. Operators $\A_1:=\dst\frac12r_{\bR_\alpha}$
$\A_2:=r_{\bR^+}\V_{\Delta+k^2,+1}r_{\bR_\alpha}$, $\A_3:=r_{\bR_\alpha}\V_{\Delta+k^2,+1}r_{\bR^+}I$ and are locally
quasi-equivalent to $0$ since $v_x\A_1=\A_1v_xI=v_x\A_3=\A_2v_xI=0$ while the operators $v_x\A_2$ and $\A_3v_xI$ are compact for all $v_x\in C^\infty(\bR)$, $O\not\in\supp\,v_x$. Compact operator are, as mentioned already, locally quasi-equivalent to $0$. The identity operator $\dst\frac12r_{\bR^+}$ is locally quasi equivalent to the identity $\dst\frac12I$ in the space on the entire axes $\bR$.
Thus, the local quasi-equivalence of the system \eqref{e0.13} and the first equation in \eqref{e3.7} at $x\in\bR^+$ follows.
The local quasi-equivalence of the system \eqref{e0.13} and the second
equation in \eqref{e3.7} at $x\in\bR_\alpha$ is proved similarly, by
using the pull back operator $\J_\alpha$ (see \eqref{e0.8}). \QED
\begin{lemma}\label{l3.2}
The system of pseudodifferential equation \eqref{e0.13} is Fredholm if
and only if the system of pseudodifferential equation
\begin{eqnarray}\label{e3.10}
&&\left\{\begin{array}{ll}
\vf(t)-\dst\frac12\left[\K_{e^{i\alpha}}+\K_{e^{-i\alpha}}\right]\psi(\tau)
d\tau=G(t),\\[3mm]
\psi(t)+\dst\frac12\left[\K_{e^{i\alpha}}+\K_{e^{-i
\alpha}}\right]\vf(\tau)d\tau=H(t), \qquad t\in\bR^+,
\end{array}\right.\\[2mm]
&& \varphi,\; \psi\in\wt{\bW}^{s-1-1/p}_p(\bR^+),\qquad G,\;
H\in\bW^{s-1-1/p}_p(\bR^+) \nonumber
\end{eqnarray}
is locally invertible at $0$, where
\begin{equation}\label{e3.11}
\mathbf{K}^1_c\phi(t):=\displaystyle\frac1\pi\int\limits_0^\infty
\frac{\phi(\tau)\,d\tau}{t-c\,\tau}, \qquad 0<|\arg\,c|<2\pi, \quad\phi\in\bL_p(\bR^+).
\end{equation}
is the Mellin convolutions operator (see \cite{Du79,Du84b,Du86,Du82}).
\end{lemma}
{\bf Proof:} Due to the main principle of the quasi-localization
(see Proposition 3.4 in \cite{Du15}) the system \eqref{e0.13}
is Fredholm if and only if locally quasi-equivalent systems
(equations) is locally invertible at each point of the compactification of $\Gamma_\alpha$ which includes the infinite points, i.e., for each
$x\in\Gamma_\alpha\cup\{+\infty\}\cup\{e^{\i\alpha}\infty\}$.
The systems \eqref{e3.7} is obviously uniquely solvable (the corresponding operators are invertible).
Thus, the system \eqref{e0.13} is Fredholm if and only if the system
\eqref{e3.6} is locally invertible at $0$.
Equivalence of the local solvability of the systems \eqref{e3.6} and
\eqref{e3.10} is proved as follows.
Multiply both equations in \eqref{e3.6} by 2, apply to the first equation the differentiation $\pa_t$, replace $\vf:=\pa_t\vf_0$, apply to the second equation the $\J_\alpha$ (see \eqref{e0.8}) and replace $\psi=\J_\alpha\psi_0$, also under the integral. Now the system \eqref{e3.10} is derived easily from \eqref{e3.6} with the help of formulae \eqref{e1.17} and \eqref{e1.19}.
To prove the local equivalence at $0$ of the systems \eqref{e3.6} and
\eqref{e3.10} note, that the multiplication by $2$ and the pull back
operator $\J_\alpha$ are invertible. As for the differentiation
\[
\pa_t:=\dst\frac d{dt}\;:\;\bW^r_p(\bR^+)\to\bW^{r-1}_p(\bR^+),\qquad
\pa_t\;:\;\wt\bW^r_p(\bR^+)\to \wt\bW^{r-1}_p(\bR^+)
\]
it is locally invertible at any finite point $x\in\bR$ because the operators
\[
\pa_t-iI\;:\;\bW^r_p(\bR^+)\to\bW^{r-1}_p(\bR^+),\qquad
\pa_t+iI\;:\;\wt\bW^r_p(\bR^+)\to \wt\bW^{r-1}_p(\bR^+)
\]
are isomorphisms (represent the Bessel potentials; see \cite[Lemma 5.1]{Du79}). On the other hand, the embeddings
\[
iI\;:\;\bW^r_p(\bR^+)\to\bW^{r-1}_p(\bR^+),\qquad
iI\;:\;\wt\bW^r_p(\bR^+)\to \wt\bW^{r-1}_p(\bR^+)
\]
are locally compact due to the Sobolev's embedding theorem and the
compact perturbation can not influence the local invertibility. \QED
\section{Mellin convolution operators in the Bessel potential spaces}
\label{sect4}
\setcounter{equation}{0}
The results of the foregoing two sections together with the results on a Banach algebra generated by Mellin and Fourier convolution operators (see \cite{Du87}) allow the investigation of the Fredholm properties of lifted Mellin convolution operators. For this we write the symbol of a model operator
\begin{equation}\label{e4.1}
\mathbf{A}:=d_0I + \sum_{j=1}^nd_j\mathbf{K}^1_{c_j},\qquad 0<\arg\,c_j<2\pi, \quad d_0,d_j\in\bC,\quad j=1,\ldots,n,
\end{equation}
in the Bessel potential spaces setting
$\A\;:\;\mathbb{H}^s_p(\mathbb{R}^+)\to \mathbb{H}^s_p( \mathbb{R}^+)$ which is compiled of the identity $I$ and Mellin convolution operators $\mathbf{K}^1_{c_1},\ldots,\mathbf{K}^1_{c_n}$ with meromorphic kernels.
To expose the symbol of the operator \eqref{e4.1}, consider the infinite
clockwise oriented ``rectangle'' $\mathfrak{R}:=\Gamma_1\cup\Gamma_2^-\cup\Gamma_2^+ \cup\Gamma_3$, where (cf. Figure 2)
\[ \Gamma_1:=\{\infty\}\times\overline{\mathbb{R}},\qquad\Gamma^\pm_2
:=\overline{\mathbb{R}}^+\times\{\pm\infty\},\qquad
\Gamma_3:=\{0\}\times\overline{\mathbb{R}}.
\]
\setlength{\unitlength}{0.4mm}
\vskip7mm
\hskip25mm
\begin{picture}(300,140)
\put(-00,40){\epsfig{file=Rectangle.pdf,height=40mm, width=80mm}}
\put(80,50){\makebox(0,0)[lc]{$(0,\xi)$}}
\put(80,130){\makebox(0,0)[lc]{$(\infty,\xi)$}}
\put(80,33){\makebox(0,0)[lc]{$\Gamma_3$}}
\put(80,145){\makebox(0,0)[lc]{$\Gamma_1$}} \put(-13,90){\makebox(0,0)[lc]{$\Gamma^-_2$}}
\put(8,90){\makebox(0,0)[lc]{$(\eta,-\infty)$}}
\put(203,90){\makebox(0,0)[lc]{$\Gamma^+_2$}}
\put(160,90){\makebox(0,0)[lc]{$(\eta,+\infty)$}}
\put(-10,145){\makebox(0,0)[lc]{$(\infty,-\infty)$}}\
\put(170,35){\makebox(0,0)[lc]{$(0,+\infty)$}}
\put(-10,35){\makebox(0,0)[lc]{$(0,-\infty)$}}\
\put(170,145){\makebox(0,0)[lc]{$(\infty,+\infty)$}}
\put(-20,15){\makebox(0,0)[lc]{Fig. 2. The domain $\mathfrak{R}$ of definition of the symbol $\mathcal{A}^s_p(\omega)$.}}
\end{picture}
Now we recall the symbol $\mathcal{A}^s_p$ of the operator $\mathbf{A}$ written in [DD16]:
\begin{subequations}
\begin{equation}\label{e4.2a}
\mathcal{A}^s_p(\omega):=d_0\mathcal{I}^s_p(\omega)
+\sum_{j=1}^nd_j\mathcal{K}^{1,s}_{c_j,p}(\omega),
\qquad \omega\in\mathfrak{R}.
\end{equation}
The symbols $\mathcal{I}^s_p(\omega)$ and $\mathcal{K}^{1,s}_{c_j,p}(\omega)$
in \eqref{e4.2a} are defined as follows:
\begin{eqnarray}\label{e4.2b}
\mathcal{I}^s_p(\omega)&\hskip-3mm:=&\hskip-3mm\begin{cases}
g^s_p(\infty,\xi), & \omega=(\infty,\xi)\in\overline{\Gamma}_1,
\\[1ex]
\left(\displaystyle\frac{\eta-\gamma}{\eta+\gamma}\right)^{\mp s}, &
\omega=(\eta,\pm\infty)\in\Gamma^\pm_2, \\[1ex]
e^{\pi si}, &\omega==(0,\xi)\in\overline{\Gamma}_3,\end{cases}
\end{eqnarray}
\begin{eqnarray}\label{e4.2c}
\mathcal{K}^{1,s}_{c,p}(\omega)&\hskip-3mm:=&\hskip-3mm\begin{cases}
\displaystyle \frac{e^{-i\pi\left(\frac1p-i\xi-1\right)}c^{\frac1p-i\xi-s-1}}{ \sin\pi(\frac1p-i\xi)},&\omega=(\infty,\xi)\in\overline{\Gamma}_1,\\[1ex]
0, &\omega=\eta,\pm\infty)\in\Gamma^\pm_2,\\[1ex]
\displaystyle \frac{e^{-i\pi\left(\frac1p-i\xi-1\right)}c^{\frac1p-i\xi-s-1}}{ \sin\pi(\frac1p-i\xi)},&\omega=(0,\xi)\in\overline{\Gamma}_3,
\end{cases}\\[1.5ex]
g^s_p(\infty,\xi)&\hskip-3mm:=&\hskip-3mm\frac{e^{2\pi si}+1}2
-\frac{e^{2\pi si}-1}{2i}\cot\pi\Big(\frac1p-i\xi\Big)=e^{\pi si}\frac{\sin\pi\Big(\frac1p-s-i\xi\Big)}
{\sin\pi\Big(\frac1p-i\xi\Big)},\quad\xi\in\mathbb{R},\nonumber
\end{eqnarray}
\end{subequations}
where
\[
0<\arg\,c<2\pi,\quad-\pi<\arg(c\,\gamma)<0,\quad 0<\arg\gamma<\pi,\quad c^\gamma=|c|^\gamma e^{i\gamma\arg\,c}.
\]
\begin{proposition}[\cite{DD16}, Theorem 5.4, \cite{Du15}, Theorem 4.14]\label{p4.1}
Let $1<p<\infty$, $s\in\mathbb{R}$. The operator
\begin{equation}\label{e4.3}
\mathbf{A}:\widetilde{\mathbb{H}}{}^s_p(\mathbb{R}^+)\longrightarrow
\mathbb{H}^s_p (\mathbb{R}^+)
\end{equation}
defined in \eqref{e4.1} is Fredholm if and only if its symbol $\mathcal{A}^s_p(\omega)$ defined in \eqref{e4.2a}--\eqref{e4.2c}, is elliptic. If $\mathbf{A}$ \, is Fredholm, then
\begin{equation*}
{\rm Ind}\mathbf{A}=-{\rm ind}\det\mathcal{A}^s_p.
\end{equation*}
The operator $\A$. defined in \eqref{e4.1}, is locally invertible at $0$ in the setting \eqref{e4.3} if and only if its symbol $\mathcal{A}^s_p(\omega)$ defined in \eqref{e4.2a}--\eqref{e4.2c}, is elliptic on $\Gamma_1$:
\begin{equation}\label{e4.4}
\inf_{\omega\in\Gamma_1}\det\mathcal{A}^s_p(\omega)\not=0.
\end{equation}
\end{proposition}
\begin{proposition}[\cite{Du15,DD16}]\label{p4.2}
Let $1<p<\infty$, $s\in\mathbb{R}$ and let $\mathbf{A}$ be defined by \eqref{e4.1}. If the operator $\mathbf{A}\;:\;\widetilde{\mathbb{H}}{}^s_p (\mathbb{R}^+) \longrightarrow\mathbb{H}^s_p (\mathbb{R}^+)$ is Fredholm (is invertible) for all $a\in(s_0,s_1)$ and $p\in(p_0,p_1)$, where $-\infty<s_0<s_1 <\infty$, $1<p_o<p_1<\infty$, then
\[
\mathbf{A}\;:\;\widetilde{\mathbb{W}}{}^s_p (\mathbb{R}^+) \longrightarrow
\mathbb{W}^s_p (\mathbb{R}^+),\qquad s\in(s_0,s_1), \quad p\in(p_0,p_1)
\]
is Fredholm (is invertible, respectively) in the Sobolev-Slobode\v{c}kii spaces $\mathbb{W}^s_p$ and has the same index
\[
{\rm Ind}\,\mathbf{A}=-{\rm ind}\,\det\,\mathcal{A}^s_p.
\]
\end{proposition}
\begin{proposition}[\cite{Du15,DD16}]\label{p4.3}
Let $1<p<\infty$, $s\in\mathbb{R}$ and let $\mathbf{A}$ be defined by \eqref{e4.1}. If the operator $\mathbf{A}\;:\;\widetilde{\mathbb{H}}{}^s_p (\mathbb{R}^+) \longrightarrow\mathbb{H}^s_p (\mathbb{R}^+)$ is Fredholm (is invertible) for all $a\in(s_0,s_1)$ and $p\in(p_0,p_1)$, where $-\infty<s_0<s_1 <\infty$, $1<p_o<p_1<\infty$, then
\[
\mathbf{A}\;:\;\widetilde{\mathbb{W}}{}^s_p (\mathbb{R}^+) \longrightarrow
\mathbb{W}^s_p (\mathbb{R}^+),\qquad s\in(s_0,s_1), \quad p\in(p_0,p_1)
\]
is Fredholm (is invertible, respectively) in the Sobolev-Slobode\v{c}kii spaces $\mathbb{W}^s_p$ and has the equal index
\[
{\rm Ind}\,\mathbf{A}=-{\rm ind}\,\det\,\mathcal{A}^s_p.
\]
\end{proposition}
\section{Investigation of the Boundary integral equation of model problem}
\label{sect5}
\setcounter{equation}{0}
\noindent
{\bf Proof of Theorem \ref{t0.6}:} Due to Lemma \ref{l3.2} the boundary pseudodifferential equation \eqref{e0.13} of the model mixed boundary value problem is Fredholm if the pseudodifferential equation \eqref{e3.10} is locally invertible at $0$.
Let us investigate the boundary integral equation \eqref{e3.10}. For this it is convenient to rewrite it as an operator equation
\begin{eqnarray}\label{e5.1}
&\M_\alpha\Phi=\F, \\[3mm]
&\Phi:=\left(\begin{array}{c}\vf\\ \psi\end{array}\right)\in
\widetilde{\bH}{}^r_p(\bR^+),\qquad {\bf
F}:=\left(\begin{array}{c}G\\
H\end{array}\right)\in\bH^r_p(\bR^+)\nonumber \end{eqnarray}
where the operator $\M_\alpha\;:\;\widetilde{\bH}{}^r_p(\bR^+)
\to{\bH}^r_p(\bR^+)$ has the form
\begin{eqnarray*}
\M_\alpha:=\left[\begin{array}{cc} I & -\dst\frac12[\K^1_{e^{i\alpha}}
+\K^1_{e^{i(2\pi-\alpha)}}]\\
\dst\frac12[\K^1_{e^{i\alpha}}+\K^1_{e^{i(2\pi-\alpha)}}] & I \end{array}\right]
\end{eqnarray*}
since $\K^1_{e^{i(-\alpha)}}=\K^1_{e^{i(2\pi-\alpha)}}$. Now
Propositions \ref{p4.1} and \ref{p4.2} can be applied to $\M_\alpha$.
We investigate the equation \eqref{e5.1} in the Bessel potential space setting \eqref{e0.15b}. The proof for the Sobolev-Slobode\v{c}kii spaces $\mathbb{W}^s_p$ follows then from Proposition \ref{p4.2} and we leave the details of the proof to the reader.
Since
\begin{eqnarray*}
&&\hskip-15mm\dst\frac{e^{-i\pi(\Xi-1) + i\alpha(\Xi-r-1)} +
e^{-i\pi(\Xi-1)+i(2\pi - \alpha)(\Xi-r-1)}}{
2\sin\pi\Xi}\\[3mm]
&&=e^{-i\pi(\Xi-1)+i\pi(\Xi-r-1)}\dst\frac{e^{i(\pi-\alpha)(\Xi-r-1)}
+e^{-i(\pi-\alpha)(\Xi-r-1)}}{2\sin\pi\Xi}\\[3mm]
&&=e^{-\pi ri}\dst\frac{\cos[(\pi-\alpha)(\Xi-r-1)]}{\sin\pi\Xi},
\end{eqnarray*}
using formula \eqref{e4.2a}-\eqref{e4.2c} we write the symbol of the operator $\M_\alpha$:
\begin{eqnarray}\label{e5.2}
\cM^r_{\alpha,p}(\omega)&\hskip-3mm=&\hskip-3mm\left[\begin{array}{cc} \mathcal{I}^r_p(\omega)
&-\dst\frac12[\mathcal{K}^{1,r}_{e^{i\alpha},p}+\mathcal{K}^{1,r}_{
e^{i(2\pi-\alpha)},p}](\omega)\\ \dst\frac1{2}[\mathcal{K}^{1,r}_{e^{i\alpha},p}+\mathcal{K}^{1,r}_{e^{i(2\pi-\alpha)},p}](\omega) & \mathcal{I}^r_p(\omega) \end{array}\right] \nonumber\\
&\hskip-3mm=&\hskip-3mm\left[\begin{array}{cc} e^{\pi ri}
\dst\frac{\sin\pi(\Xi-r)}{\sin\pi\Xi}&\hskip-7mm -e^{-\pi ri}
\dst\frac{\cos[(\pi-\alpha)(\Xi-r-1)]}{\sin\pi\Xi}\\[3mm]
e^{-\pi ri}\dst\frac{\cos[(\pi-\alpha)(\Xi-r-1)]}{\sin\pi\Xi}
&e^{\pi ri}\dst\frac{\sin\pi(\Xi-r)}{\sin\pi\Xi}
\end{array}\right],\\[3mm]
&&\hskip25mm\text{for}\quad \omega=(\infty,\xi)\in\overline{\Gamma_1},
\quad\xi\in\bR,\quad\Xi:=\dst\frac1p-i\xi,\nonumber
\end{eqnarray}
We did not write the symbol on $\Gamma_2^\pm$ and $\Gamma_3$, because we are only interested in the local invertibility of $\A$ at $0$ (see Theorem \ref{t0.5}, Lemma \ref{l3.1}, Lemma \ref{l3.2} and Proposition \ref{p4.1}).
From \eqref{e5.2} follows:
\begin{eqnarray*}
\det\,\cM^r_{\alpha,p}(\omega)=e^{-2\pi ri}
\dst\frac{e^{4\pi ri}\sin^2\pi(\Xi-r)+
\cos^2[ (\pi -\alpha)(\Xi-r-1)]}{\sin^2\pi\Xi},\\
\omega=(\infty,\xi)\in\overline{\Gamma_1}.
\end{eqnarray*}
Since $e^{4\pi ri}=cos(4\pi r)+i\sin(4\pi r)$, from the latter formula follows that the symbol $\cM^r_{\alpha,p}(\omega)$ {\bf is elliptic} on $\Gamma_1$ if:
{
\begin{eqnarray}\label{e5.3}
&&\hskip-10mm\sin(4\pi r)\sin^2\pi\left(\dst\frac1p-r\right)
=\sin(4\pi r)\sin^2\pi\left(\dst\frac1p-r-1\right)\not=0
\quad\text{or}\nonumber\\[3mm]
&&\hskip-10mm\cos(4\pi r)\sin^2\pi\left(\dst\frac1p-r\right) +
\cos^2\left[(\pi-\alpha)\left(\dst\frac1p - r-1\right)\right]=\\[3mm]
&&\hskip-10mm=\cos(4\pi r)\sin^2\pi\left(\dst\frac1p-r-1\right) +
\cos^2\left[(\pi-\alpha)\left(\dst\frac1p - r-1\right)\right]\not=0.\nonumber
\end{eqnarray}}
From \eqref{e5.3} we derive the following conditions of the ellipticity:
{
\begin{itemize}
\item[1)]
$r\not=\dst\frac n4$ and $r\not=\dst\frac1p-n-1$, $n=0\pm1,\ldots$, and the condition coincides with \eqref{e0.16a}.
\item[2)]
If $\dst\frac1p-r-1=n$, then $\cos(\pi-\alpha)n\not=0,$ i.e., $(\pi-\alpha)n\not=\dst\frac\pi2+\pi k$, $n=\pm1,\pm2,\ldots$, $k=0,\pm1,\ldots$. This condition coincides with \eqref{e0.16b}.
\item[3)]
If $r=\dst\dst\frac n2$, then
\[
\sin^2\pi\left(\dst\frac1p-\dst\frac n2-1\right)+\cos^2(\pi-\alpha)
\left(\dst\frac1p - \dst\frac n2-1\right)\not=0
\]
and the ellipticity condition is
\[
\dst\frac1p-\dst\frac n2-1\not=m\quad\text{or}\quad(\pi-\alpha)
\left(\dst\frac1p - \dst\frac n2-1\right)=(\pi-\alpha)m\not=\frac\pi2(2k+1).
\]
The condition coincides with \eqref{e0.16c}.
\item[4)]
If $r=\dst\dst\frac n2-\dst\dst\frac14$, then
\begin{eqnarray*}
&&\sin^2\pi\left(\dst\frac1p-\dst\frac n2 -\dst\frac34\right)
-\cos^2(\pi-\alpha)\left(\dst\frac1p - \dst\frac n2-\dst\frac34\right)\\
&&=\cos^2\pi\left(\dst\frac1p-\dst\frac n2-\frac14\right)
-\cos^2(\pi-\alpha)\left(\dst\frac1p - \dst\frac n2-\dst\frac34\right)\not=0.
\end{eqnarray*}
Then the ellipticity condition is
\begin{eqnarray*}
&&\pi\left(\dst\frac1p-\dst\frac n2-\dst\frac14\right)
-(\pi-\alpha)\left(\dst\frac1p - \dst\frac n2-\dst\frac34\right)
=\frac\pi2+\alpha\left(\dst\frac1p - \dst\frac n2-\dst\frac34\right)\not=\pi k\\
&&\text{and}\quad\pi\left(\dst\frac1p-\dst\frac n2-\dst\frac14\right)
+(\pi-\alpha)\left(\dst\frac1p - \dst\frac n2-\dst\frac34\right)\\
&&=\pi+2\pi\left(\dst\frac1p-\dst\frac n2\right)
-\alpha\left(\dst\frac1p - \dst\frac n2-\dst\frac34\right)
\not=\pi(k+1)
\end{eqnarray*}
and coincides with \eqref{e0.16d}.
\end{itemize}}
Concerning the unique solvability conditions \eqref{e0.17} of the system \eqref{e0.13} in the non-classical setting \eqref{e0.15b}.
If the conditions \eqref{e0.17} hold, one of the conditions \eqref{e0.16a}-\eqref{e0.16c} hold as well and, therefore, the symbol of the system \eqref{e0.13} is elliptic. Moreover, for $r=-\dst\frac12$, $p=2$ and arbitrary $0<\alpha<2\pi$ the system \eqref{e0.13} has a unique solution (cf. the concluding assertion of Theorem \ref{t0.5}). But $r=-\dst\frac12$, $p=2$ and arbitrary $0<\alpha<2\pi$ also satisfy the conditions \eqref{e0.17} and, due to Proposition \ref{p4.3} the system of boundary integral equations \eqref{e0.13} has a unique solution for all values of the parameters $\alpha$, $r$ and $p$ which satisfy the conditions \eqref{e0.17}. \QED.
\noindent
{\bf Proof of Theorem \ref{t0.3}:} Due to the Theorem \ref{t0.5} the BVP \eqref{e0.4} is Fredholm in the non-classical setting \eqref{e0.6} if the system \eqref{e0.13} in the setting \eqref{e0.15a} is, provided $r=s-1-\dst\frac1p$. The ellipticity condition \eqref{e5.3} for the BVP \eqref{e0.4} acquires the form \eqref{e0.10} and can also be written in the form:
{
\begin{eqnarray}\label{e5.9}
&&\hskip-10mm\sin\,4\pi\left(s-\dst\frac1p\right)
\sin^2\pi\left(s-\dst\frac2p\right)\not=0\quad\text{or}\\[2mm]
&&\hskip-10mm\cos\,4\pi\left(s-\dst\frac1p\right)\sin^2\pi
\left(s-\dst\frac2p\right) +\cos^2(\pi-\alpha)\left(s -\dst\frac2p\right)\not=0.\nonumber
\end{eqnarray}
From \eqref{e5.3} we get the following conditions of the ellipticity (we have to take into the account the constraint $\dst\frac1p<s<1+\dst\frac1p$):
\begin{itemize}
\item[1)]
$s\not=\dst\frac1p+\dst\frac n4$ and $s\not=\dst\frac2p+n$, $n=0\pm1,\ldots$, and this condition coincides with \eqref{e0.10a}.
\item[2)]
If $s=\dst\frac2p+n$, then $\cos(\pi-\alpha)n\not=0,$ i.e., $(\pi-\alpha)n\not=\dst\frac\pi2+\pi k$, $n=\pm1,\pm2,\ldots$, $k=0,\pm1,\ldots$. This condition coincides with \eqref{e0.10b}.
\item[3)]
If $s=\dst\frac1p+\dst\frac n2$, then
$\sin^2\pi\left(\dst\frac n2-\dst\frac1p\right)+\cos^2(\pi-\alpha)
\left(\dst\frac n2-\dst\frac1p\right)\not=0$, the ellipticity condition is
\[
\dst\frac n2-\dst\frac1p\not=m\quad\text{or}\quad(\pi-\alpha)
\left(\dst\frac n2-\dst\frac1p\right) =(\pi-\alpha)m\not=\frac\pi2(2k+1)
\]
and coincides with \eqref{e0.10c}.
\item[4)]
If $s=\dst\frac1p+\dst\frac n2+\dst\frac14$, then
\begin{eqnarray*}
&&\sin^2\pi\left(\dst\frac n2 +\dst\frac14-\dst\frac1p\right)
-\cos^2(\pi-\alpha)\left(\dst\frac n2+\dst\frac14-\dst\frac1p\right)\\
&&=\cos^2\pi\left(\dst\frac n2-\dst\frac14-\dst\frac1p\right)
-\cos^2(\pi-\alpha)\left(\dst\frac n2+\dst\frac14-\dst\frac1p\right)\not=0
\end{eqnarray*}
and the ellipticity condition is
\begin{eqnarray*}
&&\pi\left(\dst\frac n2-\dst\frac14-\dst\frac1p\right)
-(\pi-\alpha)\left(\dst\frac n2+\dst\frac14-\dst\frac1p\right)
=-\frac\pi2+\alpha\left(\dst\frac n2+\dst\frac14-\dst\frac1p\right) \not=\pi k\\
&&\text{and}\quad\pi\left(\dst\frac n2-\dst\frac14-\dst\frac1p\right)
+(\pi-\alpha)\left(\dst\frac n2+\dst\frac14-\dst\frac1p\right)\\
&&=2\pi\left(\dst\frac n2-\frac1p\right)
-\alpha\left(\dst\frac n2+\dst\frac14-\dst\frac1p\right)
\not=\pi k
\end{eqnarray*}
and coincides with \eqref{e0.10d}.
\end{itemize}}
The further proof is similar to the proof of Theorem \ref{t0.6}.
\QED
\baselineskip=12pt
|
\section{Introduction}
The fundamental physical processes which governs the evolution of electron flows with velocity gradient are of great interest in wide range of research areas in astrophysical and laboratory contexts.
In astrophysical scenario, the relativistic jets which are observed across wide range of astrophysical scales from micro-quasars to Gamma Ray Bursts (GRBs), supernovas etc., \cite{J.H.N,GRB,GRB1,SUPERNOVA} would have sheared flow of electrons.
In laser plasma experiments also, there are many situations where the sheared electron flow configuration is inevitable. For instance
experiments on fast-ignition scheme of laser-driven inertial confinement fusion involve electron beam propagation inside a plasma which would invariably
result in a sheared configuration of electron flow. When a high intensity laser irradiates a solid surface and/or a compressed plasma it generates
electron beam at the critical density surface of the plasma by the wave breaking mechanism \cite{malka,Modena,joshi}. This beam typically propagates inside the high density region of the plasma exciting reverse shielding background
electron currents. The forward and reverse currents spatially separate
by Weibel instability leading to a sheared electron flow configuration. However, since the transverse extent of the beam is finite compared to the plasma width, being
commensurate with the laser focal spot, the sheared configuration of electron flow automatically exists between the beam and the background
stationary electrons at the edge of the propagating beam \cite{chandra} even before
Weibel destabilization process. In such a scenario the Kelvin-Helmholtz (KH) instability develops immediately at the edge of the beam and
does not require a Weibel destabilization process to preempt it.
The KH instability is a well known instability and has been widely studied in the context of hydrodynamic fluid. However,
the sheared-electron velocity flow encountered in laboratory and astrophysical cases, mentioned above, differs from the hydrodynamic fluid flows in
many respects. For instance, the sheared-flow configuration of electron fluid invariably has currents and sheared current flows associated with it.
Consequently, the evolution of the magnetic field associated with it becomes an integral part of the
dynamics. Development of charge imbalance is another aspect in the evolution. Though the equilibrium
charge balance is provided by the neutralizing static background of electrons, compressible electron flow
during evolution can easily lead to charge imbalance as the ions would not respond at fast electron
time scale phenomena. This would lead to electrostatic field generation
which has added influence in the dynamics. Lastly,
the flow of electrons in most cases is relativistic. Thus, to summarize
the KH instability in this case has additional effects due to the presence of electromagnetic features,
compressibility leading to electrostatic fields, relativistic effects etc. In the non -relativistic limit the electromagnetic effects on KH instability
in the context of sheared electron flows have been investigated in detail by employing the Electron Magnetohydrodynamic (EMHD) model
\cite{bul,amita,amita1, Pegoraro}. This model neglects the displacement currents and space charge effects
and assumes stationary ions which provide the neutralizing background.
The relativistic effects on K-H instability in compressible neutral hydrodynamic fluid has been studied by Bodo $\emph{et al.}$ \cite{Bodo,Drazin}. Recently,\cite{sita} Sundar $\emph{et al.}$ have incorporated
relativistic effects on sheared-electron flows. This study points out crucial
role of shear on the relativistic mass factor due to sheared velocity configuration. The
effect due to displacement current was retained in the relativistic regime. It was, however, shown that for the weakly relativistic case the effects due to
displacement current were negligible. However, in these studies, the space charge effects which may arise when compressibility
of the electron fluid are present, have not been incorporated.
The present article aims at
exploring these features using a PIC simulation.
We have carried out a 2.5D relativistic electromagnetic Particle-in-Cell simulations to study the electron shear flow instability in both cases of weak and strong relativistic flows. By 2.5D we mean
that all three components of the fields are taken into consideration, however, their spatial
variations are confined in 2-D plane only.
When the flow is weakly relativistic, we observe the development of electromagnetic
KH instability at the location of shear which ultimately develops into vortices.
These vortices merge subsequently forming longer scales, in conformity
with the inverse cascade phenomena observed in typical 2-D fluid systems. The density perturbations are
observed to be weak in this case. The results in this case are thus very similar to the predictions
of the EMHD fluid behaviour.
When the relativistic effects are mild (and not weak), the KH instability occurs at a slower time scales. The
KH vortices are observed initially, which are soon overwhelmed by compressibility effects which
introduce magnetized non-linear electrostatic oscillations (non-linear upper hybrid oscillations )in plasma transverse to flow.
In strongly relativistic regime the electrostatic oscillations dominate right from the very beginning.
The amplitude of the oscillations increases leading to phenomena of wave breaking.
In the nonlinear regime, the spectra is observed to be broad in all the three cases which implies
turbulence.
The paper is organized as follows. In section II, we describe our simulation methodology.
The results of PIC simulations and their implications are presented in section III. It is seen that in
strong relativistic case compressibility effects seem to dominate resulting in electrostatic oscillations
transverse to the flow. These electrostatic oscillations are understood on the basis of a simplified
one dimensional model in section IV. Section V contains the description of the power spectrum of the
fields in the nonlinear regime.
Section VI contains the summary and conclusions.
\section{Description of Simulation }
We choose the electron to have a flow velocity along $\hat{y}$ with a following sheared flow configuration
as equilibrium
\begin{equation}
V_{0y}(x)=V_{0} \left[tanh((x-L_x/4)/\epsilon)+tanh((3L_x/4-x)/\epsilon)\right]-V_{0},
\label{velocity}
\end{equation}
where $\epsilon$ is width of shear layer, $L_x$ is total length of simulation box in transverse direction of
flow and $V_{0}$ is the maximum amplitude of the flow velocity. This flow structure is shown schematically in
Fig.~\ref{fig:sc}.
The electron flow is responsible for current and produces an equilibrium magnetic field in the $B_0\hat{z}$
direction.
During the simulations, ions
are kept at rest and merely provide for the neutralizing background. In order to satisfy the condition for
equilibrium force balance on electrons, there is a need to displace the electrons and ions slightly
in space, so that an equilibrium electric field $\vec{E}_0$ gets created. This is chosen in such a fashion
so as to satisfy the condition of
\begin{equation}
\vec{E}_0 + \frac{V_{0y}\hat{y} \times \vec{B}_0 \hat{z}}{c} = 0
\end{equation}
This ensures that the Lorentz force on electrons vanishes everywhere. This clearly indicates the necessity
for having an equilibrium electric field along $\hat{x}$. The electron and ion charges are thus
displaced in an appropriate fashion so as to satisfy the
Gauss's law
\begin{equation}
\nabla.\bold E= \frac{\partial E_{x}}{\partial x}=-\frac{1}{c}\frac{\partial \left ( B_{0z}V_{0y}\right)}{\partial x}=4\pi e\left ( n_{0i}-n_{0e}\right),
\end{equation}
here $n_{0i}$ and $n_{0e}$ are unperturbed ion and electron number densities respectively in equilibrium, e is charge of electron and c is speed of light.
To maintain equilibrium in system we have thus arranged the electron particle number density according to following relationship \cite{coroniti},
\begin{equation}
n_{0e}=n_{0i}+\frac{1}{4\pi ec}\frac{\partial \left ( B_{0z}V_{0y}\right)}{\partial x}.
\label{electron_den}
\end{equation}
The ions are distributed uniformly with a density $n_{0i}$ of $3.18\times10^{18}$cm$^{-3}$ and
$n_{0e}$ is adjusted as per Eq.~(\ref{electron_den}).
The area of the simulation box $R$ is chosen to be $6\times5$ $( c/ \omega_{0e})^2$ corresponding to
$600\times500$ cells; where $\omega_{0e}=\sqrt{4\pi n_{0i}e^2/m_e}$ is electron plasma frequency corresponding
the uniform plasma at the background density of ions.
Also, $c/ \omega_{0e}= d_{e}=3.0\times 10^{-4}$cm is the skin depth. We have used 128 particles per cell
for both ion and electron in our simulation. To resolve the underlying physics at the scale which is
smaller than the skin depth, we have chosen
a grid size of 0.01$d_{e}$. The time step $\Delta t$, decided by the Courant condition, is 0.035 femtosecond.
We have considered four different set of parameters for our investigation. In all cases, velocity profile of electron is assigned by eq.~\ref{velocity}. For the first case, we choose the flow velocity of
electron in the weakly relativistic regime and chose the shear width to be less than the plasma skin depth.
We would refer this as Case (a) which has the following parameters $V_{0}=0.1c$, $\epsilon=0.05$ $c/\omega_{0e}$. This is the weakly relativistic case where the EMHD fluid description is supposed
to work pretty well.
We consider then in case (b), the dependence of KH instability on shear width. We do this by
changing the value of shear width in comparison to skin depth. As per the EMHD description
the growth rate decreases when the shear width is shallow compared to the skin depth.
We illustrate this by specifically choosing a value of $\epsilon=1.5$ $c/\omega_{0e}$.
In the third and fourth cases (c) and (d) a mild and strong relativistic limit with parameters $V_{0}=0.5c$,
$\epsilon=0.05$ $c/\omega_{0e}$ and $V_{0}=0.9c$, $\epsilon=0.05$ $c/\omega_{0e}$ are respectively
considered.
\section{PIC Simulation Results }
In the three subsections we discuss the results of
(I) Weakly relativistic (II) Mild relativistic (III) Strong relativistic cases.
\subsection{ I. Weakly relativistic}
We choose a the value of $V_0 = 0.1 c$ for electron velocity to study the weakly relativistic case.
We observe a destabilization of the sheared flow configuration. The instability is tracked
by plotting the evolution of the
perturbed kinetic energy(PKE) of the electrons in the system. This is shown in
Fig.~\ref{fig:kh1_gr}. The initial steep rise is due to numerical noise. Thereafter,
the instability grows from the noise spectrum. Since, the noise would lack the
exact eigen mode structure of any particular mode, initially a combination
of unstable modes start growing.
Subsequently, as the mode with
fastest growth dominates a linear rise in the plot of PKE can be clearly observed.
It should be noted that evolution follows the EMHD fluid predictions of
the growth rate being higher for the case (a)
when the shear width is sharper than the skin depth. In case (b) the growth is observed to be small and the
saturation also occurs quite fast.
For a closer look at the instability development the color contour plot of the evolution of magnetic
field (Fig.~\ref{fig:mag1}), vorticity (Fig.~\ref{fig:vortices}) and the two components of
Electric field (Fig.~\ref{fig:elec1}) has been shown at various times. From (Fig.~\ref{fig:mag1}) magnetic field evolution, one can
observe that the magnetic perturbations start at the location where velocity shear is maximum.
These perturbations grow forming magnetic vortices which subsequently merge to form bigger structures.
The merging process of magnetic field is along expected lines of 2-D inverse cascade EMHD depiction of
the problem. The fluid vorticity also shows similar traits, however, at later times
$t = 59.60$ (in normalized units) the long scale vorticities show signs of disintegration. The two components of electric fields
also show emergence of KH structures and merging. A comparison of normalized amplitudes of electric and
magnetic field shows that the electric fields are much weaker than the magnetic fields.
We also show the plot of
electron density
in the nonlinear regime of the KH instability at $t = 36.75$ in Fig.~\ref{fig:kh1_density}.
We observe that the density also acquires distinct structures of KH like vortices in the shear region.
The density perturbations in the weakly relativistic case is observed to be weak. The maximum
observed value of $\tilde{ne}/n_{e0} \sim 1.2$. On the other hand we would see in the strongly
relativistic case this is as large as $8$ to $10$.
This suggests that in the weakly relativistic regime the instability
essentially has an electrostatic character.
\subsection{ II. Mild relativistic case}
In the mild relativistic case where $V_0 = 0.5 c$, the KH is observed to be considerably weak.
The vorticity plots shown in Fig.~\ref{fig:vor2} shows an initial tendency towards developing the KH rolls.
The KH rolls in this case are fewer in number. For case(a) they were $5$ here they are only around $3$.
This again suggests that the growth rate for relativistic case gets confined towards longer scale as per the predictions of EMHD model.
The fluid analysis carried out earlier also suggests that the cut off wavenumber of the KH moves towards
longer scales in mildly relativistic cases.
The KH rolls are observed to be very soon overwhelmed by certain oscillations transverse to the flow. The oscillations
transverse to the flow are also clearly evident in the electron density plots of Fig.~\ref{fig:elec2}.
The density oscillations in this case are pretty strong with $\tilde{n_e}/n_{e0} \sim 4$.
The KH suppression and the appearance of these upper
electrostatic oscillations can be understood as follows.
As the relativistic effect increases the $\vec{V} \times \vec{B}/c$ force becomes dominating.
Thus a small perturbed magnetic field $\tilde{B}$, induces a strong $V_0\times \tilde{B}/c$ force
along $x$, which is responsible for the upper hybrid electrostatic oscillations.
\subsection{ III. Strong relativistic}
We now choose $V_0 = 0.9c$ for understanding the strongly relativistic case.
The time evolution of PKE in this case shows the linear growth of the instability. However, the
instability is dominated by the
upper hybrid electrostatic oscillations which are observed right from the very beginning. Thus the development of the
rolls typical of the KH instability are not very clearly evident in this case.
Representing the initial distribution of flowing and the static electrons by different colors (red
and blue respectively) we show the snapshots of their displacement in space in Fig.~\ref{fig:par9}.
The electron compressibility is clearly evident, so much so and white regions
devoid of electrons are created
(snapshot at $\omega_{0e}t= 3.5$). The Electric fields due to background ions, however, pull
these electrons back which results in a large amplitude excitation of nonlinear upper hybrid electrostatic plasma
oscillations. These oscillations are discussed in detail in the next section.
A comparative value of the growth rate obtained from the slope of the evolution of PKE in the table below for all the cases studied by us.
\begin{center}
{\bf{TABLE I}} \\
The maximum growth rate ($\Gamma_{gr} max.$) of K-H instability evaluated from slope of perturbed kinetic energy\\
\vspace{0.2in}
\vspace{0.2in}
\begin{tabular}{c c c c c c c c c c c c c ll}
\hline
\hline
&$V_{0}/c $ \hspace{0.3in} & $ \varepsilon/(c/ \omega_{0e})$ \hspace{0.3in} &
$\Gamma_{max}/(V_0\omega_{0e}/c)$\\
\hline
&0.1 \hspace{0.3in} & 0.05 \hspace{0.3in} &0.7 \\
&0.1 \hspace{0.3in} & 1.5 \hspace{0.3in} &0.0 \\
&0.5 \hspace{0.3in} & 0.05 \hspace{0.3in} &0.34 \\
& 0.9 \hspace{0.3in} & 0.05 \hspace{0.3in} &0.23 \\
\hline
\end{tabular}
\end{center}
Since, classically the KH instability typically scales with the fluid flow velocity
we have chosen to divide the growth rate with $V_0$ for a better appreciation of the comparison.
The comparison clearly, shows that $\Gamma_{max}/V_0$ decreases due to relativistic effects in agreement with the earlier fluid analysis by Sundar {\it et al}. Thus the distinction between the PIC and EMHD fluid simulations finally
boils down to the appearance and dominance of electrostatic oscillations transverse to the flow direction.
We study the transverse oscillations in detail in the next section.
\section{Nonlinear upper hybrid electrostatic oscillations}
One of the main observations is the appearance of strong upper hybrid electrostatic oscillations
triggered from the edge of the flow region with increasing relativistic effects.
We show in Fig.~\ref{fig:wavekhpic} the amplitude of these oscillations as a function of time at $y = 2.5$ $c/\omega_{0e}$ for the strongly
relativistic case of $V_0 = 0.9c$. It can be seen that the density perturbations acquire a very high
amplitude fairly rapidly
$\tilde{n_e}/n_{e0} \sim 8$. This is a very nonlinear regime for the oscillations where wave breaking
and trajectory crossing would occur. This is indeed so as the particle distribution of Fig.~\ref{fig:par9} shows clear
crossing of blue and red electrons.
In order to understand the dynamics behind this phenomenon, we model the phenomena by a one-dimensional
magnetized relativistic electron fluid equations for electrostatic disturbances.
Thus the governing equations of the model are expressed as
\begin{eqnarray}
&&\left(\frac{\partial }{\partial t} +v_{ex}\frac{\partial}{\partial x} \right)n_{e}=-n_{ex}\frac{\partial v_{ex}}{\partial x}, \\
\label{con1}
&&\left(\frac{\partial }{\partial t} +v_{ex}\frac{\partial}{\partial x} \right)p_{ex}=-eE_{x}-\frac{ev_{ey}B_0(x)}{c}, \\
\label{mom1}
&&\left(\frac{\partial }{\partial t} +v_{ex}\frac{\partial}{\partial x} \right)p_{ey}=\frac{ev_{ex}B_0(x)}{c}, \\
\label{mom2}
&&\left(\frac{\partial }{\partial t} +v_{ex}\frac{\partial}{\partial x} \right)E_x=4\pi en_{0i}v_{ex}, \\
\label{Max1}
\end{eqnarray}
where $p_{e\alpha}=\gamma m_{e}v_{e\alpha}$ is $\alpha$-component of momentum;$\alpha$ is subscript for x and y, $\gamma=[1+p^2/m^2_ec^2]^{1/2}$ is relativistic factor and $n_{0i}$ is background ion density.
The inhomogeneous magnetic field $B_0(x)$ is the equilibrium magnetic field generated from the
equilibrium electron flow considered in our PIC simulations. For the double tangent hyperbolic profile
it will have the following form:
\begin{eqnarray}
&& B_0(x)=\frac{4\pi n_{0e}e}{c}\left(V_0\epsilon\log \left(\cosh\left(0.25L_x -x\right)\right)+V_0\epsilon\log \left(\cosh\left(x-0.75L_x\right)\right)-V_0 x \right).\nonumber\\
\label{B1}
\end{eqnarray}
We have solved the above equations numerically with initial profile of $v_{ey}$ using the eq.~(\ref{velocity}). For the weakly relativistic case of $V_0 = 0.1c$ the electrostatic oscillations
that get generated are quite small and continue to remain so indefinitely (See Fig.~\ref{fig:wavekhnon}).
However, when the value of $V_0$ is increased to a high value of
$V_0=0.9c$, large-amplitude non-linear oscillations in electron density (see Fig.~\ref{fig:wavekhrel})
can be clearly seen. This is similar to the results of our PIC simulations.
The upper hybrid frequency $\omega_{UH}$ is given by
\begin{equation}
\omega_{UH}^2=\omega_{0e}^2+\omega_{ce}^2
\label{uh}
\end{equation}
In our simulations, since the magnetic field is nonuniform, the upper hybrid oscillations occur against an inhomogeneous magnetic field background.
For comparing the observed oscillation frequency with that of the upper hybrid
oscillations we have chosen to consider an average magnetic field. Thus
$\omega_{ce}$=$eB_{r.m.s}/m_e c$, $B_{r.m.s}$ is root mean square value of magnetic field.
We calculate the upper hybrid frequency from dispersion relation eq.~\ref{uh}, from PIC simulation of $\omega_{UH}$ and from 1D model and have tabulated it in table II for the two cases of mild and strong relativistic flows.
It can be seen that all the two approaches (simplified dispersion equation and 1 D model) yield comparable estimates for the observed electrostatic oscillations in the PIC simulations.
\begin{center}
{\bf{TABLE II}} \\
The table for upper hybrid frequency obtained from dispersion relation eq.~\ref{uh} $\omega_{UH}(anal.)$, from PIC simulation $\omega_{UH}(PIC)$ and from 1d model $\omega_{UH}$(1d model) for various different value of $V_0$\\
\vspace{0.2in}
\vspace{0.2in}
\begin{tabular}{c c c c c c c c c c c c c ll}
\hline
\hline
&$V_{0}/c $ \hspace{0.3in} & $\omega_{UH}(anal.)/\omega_{0e}$ \hspace{0.3in} & $\omega_{UH}(PIC)/\omega_{0e}$ \hspace{0.3in} & $\omega_{UH}$(1d model)$/\omega_{0e}$
\\
\hline
&0.5 \hspace{0.3in} &1.09 \hspace{0.3in} &1.06 \hspace{0.3in} &1.06 \\
&0.9 \hspace{0.3in} &1.27 \hspace{0.3in} &1.26 \hspace{0.3in} &1.25 \\
\hline
\end{tabular}
\end{center}
The upper hybrid frequency obtained from various method are match very well and affirm the existence of upper hybrid mode in sheared electron flow.
\section{Nonlinear regime}
The nonlinear regime of the simulation shows evidence of turbulence generation for both
weak and strong relativistic cases. We have plotted the spectra of magnetic and electric fields
as a function of $k_y$ defined by the following relationship.
\begin{equation}
S_F(k_{y})=\frac{1}{L_{x}}\int_{0}^{L_{x}}F^2(x, k_{y}) dx,
\label{Ey}
\end{equation}
where $ S_F(k_y)$ is one dimensional longitudinal energy spectra of the field F, where F is the x-
dependent longitudinal Fourier transform of any of the electric and magnetic fields
(represented by $R$ here) given by
\begin{equation}
F(k_{y}, x)=\frac{1}{L_{y}}\int_{0}^{L_{y}}R(x, y)exp(-ik_y y) dy,
\end{equation}
We observe that in both strong and weak relativistic cases the spectra is broad and has a power law
behaviuor (see Fig.~\ref{fig:spectra3}, Fig.~\ref{fig:spectrab1} and Fig.~\ref{fig:spectrae1} ). The spectral scaling index is found to be close to $-4$.
In the strong relativistic case, we observe that the power law extends towards the longer wavelength
region of $kd_e \sim 1$ whereas this is not so for the weak relativistic case.
It appears that it is easier to generate longer scales in the strongly relativistic case.
\section{Summary and Conclusion}
A detailed PIC simulation was carried out to study the instability of sheared relativistic electrons
against a background of neutralizing ions. Our studies on weakly relativistic case show good agreement with
the observations based on EMHD fluid approximation. For instance the observation of instability getting
driven when the shear scale is sharper than the skin depth, the development of KH vortices in the shear
region which ultimately merge to form longer structures etc., are all in conformity of the fluid
EMHD theory.
In the strong relativistic case the compressibility effects are dominant and one observes a
characteristic
electrostatic oscillations transverse to the flow direction. This
overwhelms the KH in the system. The nonlinear regime in all cases
shows a broad power spectra of magnetic field which is indicative of turbulence.
\clearpage
\newpage
\bibliographystyle{unsrt}
|
\section{Introduction}
Nonautonomous chain rules formulas in $BV$
have been successfully used in the study of
semicontinuity properties of integral functionals
(see \cite{DC,dcfv,DCFV2,dcl}) and
conservation laws with discontinuous flux of the form
\begin{equation}\label{f:sc1}
u_t + \Div \boldsymbol B(x, u) = 0,
\qquad (t,x)\in (0,+\infty)\times\R^{N}{R}^N
\end{equation}
(see \cite{CD,CDD,CDDG}
and also \cite{vol,vol1} in the autonomous case).
In this paper we shall restrict our attention only to this second kind of application.
In order to clarify the connection between chain rule formulas and
uniqueness results for the Cauchy problems
associated with \eqref{f:sc1},
it will be convenient to recall some previous results.
In \cite{CDD} the authors considered a flux $\boldsymbol B$ such that
$\boldsymbol B(\cdot,z)$ is a special function of bounded variation (SBV)
and of class $C^1$ with respect to the second variable.
A uniqueness result for~\eqref{f:sc1} is then obtained in the class of $BV$ functions
by using the chain rule formula proven in \cite{ACDD} for the
composite function ${\bf v}(x) := \boldsymbol B(x, u(x))$.
(For the sake of completeness we recall that,
under the same structural hypotheses on the flux,
a similar uniqueness result has been recently obtained in \cite{CDDG}
for weak entropy solutions, without the $BV$ regularity requirement.)
On the other hand, Panov proved in \cite{Pan1}
an existence result of entropy solutions
in the case of discontinuous fluxes $\boldsymbol B(x,z)$ such that $\boldsymbol B(\cdot,z)$
is a vector field whose distributional divergence $\Div_x \boldsymbol B(\cdot, z)$
is a measure (see \cite{Anz,cf,ChenTorres} for a general theory of bounded divergence-measure vector field).
This assumption on $\Div_x \boldsymbol B(\cdot, z)$,
rather than requiring $\boldsymbol B(\cdot, z) \in SBV$,
is indeed natural when looking for entropy solutions of \eqref{f:sc1}.
The structure of the proof of the uniqueness result in \cite{CDD}
can be adapted to this more general situation,
provided that one can prove a
suitable chain rule formula.
This is exactly the aim of this paper:
in Section~\ref{s:div} we shall prove a
nonautonomous chain rule formula for
the divergence of the vector field
${\bf v}(x) := \boldsymbol B(x, u(x))$,
where $\boldsymbol B(\cdot, t)$ is a
divergence--measure vector field,
of class $C^1$ with respect to the second variable,
and $u\colon\R^{N}{R}^N\to\R^{N}{R}$ is a function of bounded variation.
Then, we can mimic the proof in \cite{CDD} in order to obtain,
under these assumptions on $\boldsymbol B$,
a uniqueness result for $BV$ solutions
of the Cauchy problems associated with \eqref{f:sc1}
(see Section~\ref{s:cons}).
We stress that this is not a genuine well-posedness
result, since uniqueness of solutions has been proven
in a class of functions
which is smaller
than the one for which existence
has been obtained by Panov.
\medskip
Before stating our results in a more precise way,
let us recall the state of the art
about chain rule formulas,
starting from the autonomous (i.e., independent of $x$) case.
The first result concerning distributional derivatives
is the one proved by
Vol'pert in \cite{vol} (see also \cite{vol1}), in view of applications to the study of quasilinear hyperbolic equations.
He established a chain rule formula for distributional derivatives of the composite function $v(x)=B(u(x))$, where $u:\Omega\to\R^{N}{R}$ has bounded variation in the open subset $\Omega$ of $\R^{N}{R}^{N}$ and $B:\R^{N}{R}\to\R^{N}{R}$ is continuously differentiable.
He proved that $v$ has bounded variation and its distributional derivative
$Dv$ (which is a Radon measure on $\Omega$) admits an explicit representation in terms of the classical derivative $B'$ and of the distributional derivative $Du$\,.
More precisely, the following equality holds
\begin{equation}\label{chainDv}
Dv= B' (u)\nabla u \ \mathcal L^{N}+ B'(\widetilde u) D^cu+[B(u^+)-B(u^-)]\, \nu_u\, \mathcal H^{N-1}
\res{J_{u}}\,,
\end{equation}
in the sense of measures,
where
\[
Du=\nabla u \ \mathcal L^{N}+D^cu+
(u^+ - u^-)
\nu_u\, \mathcal H^{N-1}
\res{J_{u}}\,
\]
is the decomposition of $Du$ into its absolutely continuous part
$\nabla u\, {\mathcal L}^N$ with respect to the Lebesgue measure $\mathcal L^{N}$, its Cantor part $D^cu$ and its jump part, which in turn is a measure concentrated on
the $\mathcal{H}^{N-1}$--rectifiable jump set $J_u$ of $u$.
Here, $\nu_u$ denotes the measure theoretical unit normal to $J_u$, $\widetilde u$ is the approximate limit
of $u$ and $u^+$, $u^-$ are the traces of $u$ on $J_u$.
(Here and in the following we refer to Chapter~3 of \cite{AFP} for notations and the basic facts
concerning $BV$ functions.)
An identity similar to \eqref{chainDv} holds also in the vectorial case (see Theorem 3.96 in \cite{AFP}), namely when ${\bf u}:\R^{N}{R}^{N}\to\R^{N}{R}^{h}$ has bounded variation and $B:\R^{N}{R}^{h}\to\R^{N}{R}$ is continuously differentiable.
In this case, \eqref{chainDv} can be written as
\begin{equation}\label{chainDv2}
D v=\nabla B({\bf u})\nabla {\bf u} \ \mathcal L^{N} + \nabla B(\widetilde {\bf u}) D^c{\bf u}+
[B({\bf u}^+)-B({\bf u}^-)] \nu_{\bf u}\, \mathcal H^{N-1}
\res{J_{{\bf u}}}\,.
\end{equation}
A further extension, that we are not going to use in the present paper,
concerns the case when $B$ is only a Lipschitz continuous function.
In this case, a general form of the formula was proved by Ambrosio and Dal Maso in \cite{adm}
(see also \cite{LM}, Theorem 3.99 in \cite{AFP} for the scalar case
and \cite{DCinpreparation} for the nonautonomous case).
\medskip
Recently, analogous chain rule formulas have been obtained
in the case of an explicit dependence with respect to the space variables $x$, especially in view of applications to semicontinuity results for convex integral nonautonomous
functionals (see \cite{DC,dcfv,DCFV2,dcl}) and to conservation laws with discontinuous flux (see \cite{CD,CDD,CDDG}).
This amounts to describe the
distributional derivative of the composite function ${\bf v}(x)=\boldsymbol B(x,{\bf u}(x))$,
where $\boldsymbol B(x,\cdot)$ is continuously differentiable and
$\boldsymbol B(\cdot,{\bf z})$ and ${\bf u}$ are functions with low regularity
(which will be specified later).
These formulas contain another derivation term due to the presence of the explicit dependence on $x$.
In the case of ${\bf u}$ and $\boldsymbol B$ regular functions, the classical chain rule formula
\[
\nabla {\bf v}(x)=\nabla_{x}\boldsymbol B (x,{\bf u}(x))+ \nabla_{{\bf z}}\boldsymbol B(x,{\bf u}(x))\cdot\nabla {\bf u}(x)\,,\quad x\in\R^{N}{R}^{N}\,,
\]
is a pointwise identity and the derivatives here occurring are the classical ones.
Clearly, when $\boldsymbol B(\cdot, {\bf z})$ or ${\bf u}$ (or both) are not regular functions,
then also ${\bf v}$ need not be regular and a number of extra terms will appear,
as in the previous formulas \eqref{chainDv} and \eqref{chainDv2}.
In the main result of the paper,
stated in Theorem~\ref{chainb3}, we assume that
$h=1$, the function $\boldsymbol B(x,\cdot)$ is $C^1$,
$\Div_x\boldsymbol B(\cdot,z)$
is a Radon measure,
and $u$ is a scalar function of bounded variation.
As we shall see, in order to obtain the formula in such generality,
we need to assume \textsl{a-priori} the existence of the strong traces $\boldsymbol B^\pm(\cdot, z)$ of $\boldsymbol B(\cdot, z)$
on a $\mathcal{H}^{N-1}$--rectifiable universal singular set
$\mathcal{N}$, independent of $z$ (see Section~\ref{s:div}).
Nevertheless, we think that this restriction should not be
too much severe in view of applications to conservation laws with
discontinuous flux (see Section~\ref{s:cons} and Remark~\ref{r:triang}).
Clearly, in this case one can expect a chain
rule formula only for the divergence of the
composite function \({\bf v}(x) = \boldsymbol B(x, u(x)))\).
Namely,
we shall prove that
the distributional divergence of ${\bf v}$
is given by
\begin{equation}\label{lolo5}
\begin{split}
\Div {\bf v}(x))=&\left.\Div_x\boldsymbol B(x,t)\right|_{t=u(x)}\\
+&\pscal{\boldsymbol b(x, \widetilde{u}(x))}{\widetilde Du}
+\pscal{
\boldsymbol B^*(x,u^+)
-\boldsymbol B^*(x,u^-)
}{\nu_u}
{\mathcal H}^{N-1}\res {\mathcal N\cup J_u}
\end{split}
\end{equation}
in the sense of measure, where $\boldsymbol B^*(x,z)=\frac12[\boldsymbol B^+(x,z)+\boldsymbol B^-(x,z)]$ and the measure $\Div_x\boldsymbol B(x,z)$, depending on the parameter $z$, is computed in ${z=u(x)}$ in a suitable sense (see Remark \ref{remrem}).
The proof is based on the regularization argument used in \cite{DCFV2}.
We shall see that, when $u$ is a function of bounded variation,
this explicit formula for $\Div {\bf v}$ can be used to obtain uniqueness results
for \eqref{f:sc1}.
This improves the analogous uniqueness result previously obtained in \cite{CDD},
using the chain rule formula for ${\bf v}$,
assuming $\boldsymbol B(\cdot, z)$ a special function of
bounded variation.
We emphasize that, interestingly enough,
the proof of the uniqueness result of \cite{CDD}
can be retraced with minor modifications in
this new setting, since what is really needed
is only the chain rule formula for
$\Div {\bf v}$.
\medskip
The structure of the paper is the following.
First of all, in Sections~\ref{s:ureg} and~\ref{s:BV} we review
some known chain rule formulas that we believe will help the
reader understanding the more general formula of Section~\ref{s:div}.
More precisely, in Section~\ref{s:ureg} we recall the results proved
in the case $\boldsymbol B(\cdot, {\bf z})$ not regular,
but ${\bf u}$ regular enough (i.e., in the Sobolev space $W^{1,1}$).
These results have been proved in \cite{dcl} and \cite{DCFV2}, assuming
that $\Div_x \boldsymbol B(\cdot, {\bf z})$ belongs to $L^1$ or
to the space of Radon measures respectively.
In Section~\ref{s:BV} we review the chain rule formulas in the case
when both $\boldsymbol B(\cdot, {\bf z})$ and ${\bf u}$ are of bounded variation,
recalling the results proved in \cite{DCFV2} and \cite{ACDD}
in the case of scalar and vector functions respectively.
In Section~\ref{s:div} we prove the main result of the paper,
stated in Theorem~\ref{chainb3},
concerning the distributional divergence of the composite function ${\bf v}(x)=\boldsymbol B(x,u(x))$.
Finally, in Section~\ref{s:cons}
we consider some applications of Theorem~\ref{chainb3}
to the uniqueness issue for the Cauchy problems
for multidimensional scalar conservation laws with discontinuous flux.
\section{Nonautonomous chain rules for $u\in W^{1,1}$}
\label{s:ureg}
As we have already said in the Introduction,
this section and the next one are devoted to the review
of some known chain rule formulas.
In the following, $\Omega$ will always denote a nonempty open subset
of $\R^{N}{R}^N$.
\subsection{Vectorial case $\mathbf{u}\in W^{1,1}(\Omega;\R^{N}{R}^{h})$}
The first formula of this type is established in \cite{dcl} for functions $u\in W^{1,1}(\R^{N}{R}^{N};\R^{N}{R}^{h})$ by assuming that, for every ${\bf z}\in \R^{N}{R}^{h}$, $\boldsymbol B(\cdot,{\bf z})$ is an $L^1$ function whose distributional divergence belongs to $L^1$.
(In particular, this condition holds if $\boldsymbol B(\cdot,{\bf z})\in W^{1,1}(\R^{N}{R}^{N};\R^{N}{R}^{h})$.)
Let us consider the space
\[
L^{1}({\operatorname*{div};}\Omega)=\left\{ \mathbf{u}\in L^{1}(\Omega
;\R^{N}{R}^{N}):\operatorname*{div}\mathbf{u}\in L^{1}(\Omega)\right\} ,
\]
where $\operatorname*{div}\mathbf{u}$ denotes the distributional divergence of \({\bf u}\).
We recall that an $\mathcal{H}^{N-1}$--measurable set $A\subset\R^{N}{R}^N$ is said to be
countably $\mathcal{H}^{N-1}$--rectifiable
if it can be covered, up to a set of vanishing $\mathcal{H}^{N-1}$ measure,
by a sequence of $C^1$ hypersurfaces.
The set $A$ is said to be purely $(N-1)$-unrectifiable
if $\Haus{N-1}(A\cap\Gamma)=0$ whenever $\Gamma$ is countably
$\Haus{N-1}$-rectifiable.
\begin{theorem}
[$L^{1}({\operatorname*{div};}\Omega)$-dependence]\label{weak theorem}
Let $\boldsymbol B:\Omega\times\R^{N}{R}^{h}\rightarrow\R^{N}{R}^{N}$ be a Borel function.
Assume
that there exist an $\mathcal{L}^{N}$-null set $\mathcal{N}_0\subset\Omega$ and
a purely $\mathcal{H}^{1}$-unrectifiable set $\mathcal{M}\subset\R^{N}{R}%
^{h}$ such that
\begin{enumerate}
\item[(i)] for all $\mathbf{z}\in\R^{N}{R}^{h}$ the function
$\boldsymbol B\left(
\cdot,\mathbf{z}\right) \in L_{\operatorname*{loc}}^{1}({\operatorname*{div}
;}\Omega)$;
\item[(ii)] for all $x\in\Omega\setminus\mathcal{N}_0$ the function
$\operatorname*{div}\nolimits_{x}\boldsymbol B(x,\cdot)$ is approximately continuous in
$\R^{N}{R}^{h}$;
\item[(iii)] for all $x\in\Omega\setminus\mathcal{N}_0$ the function $\boldsymbol B\left(
x,\cdot\right) $ is differentiable in $\R^{N}{R}^{h}\setminus\mathcal{M}$
and approximately continuous in $\mathcal{M}$;
\item[(iv)] for every $\Omega^{\prime}\times D\subset\subset\Omega
\times\R^{N}{R}^{h}$ there exist $g\in L^{1}(\Omega)$ and $L>0$ such
that
\[
\left| \boldsymbol B(x,\mathbf{z})\right| +\left| \operatorname*{div}\nolimits_{x}%
\boldsymbol B(x,\mathbf{z})\right| \leq g(x)
\]
for $\mathcal{L}^{N}$--a.e.\ $x\in\Omega^{\prime}$ and for all
$\mathbf{z}\in D,$ and
\[
\left| \nabla_{\mathbf{z}}\boldsymbol B(x,\mathbf{z})\right| \leq L
\]
for $\mathcal{L}^{N}$--a.e.\ $x\in\Omega^{\prime}$ and for all
$\mathbf{z}\in D\setminus\mathcal{M}$.
\end{enumerate}
Then for every ${\bf u}\in W^{1,1}(\Omega;\R^{N}{R}^{h})\cap
L_{\operatorname*{loc}}^{\infty}(\Omega;\R^{N}{R}^{h})$ the function
$\mathbf{v}:\Omega\rightarrow\R^{N}{R}^{N}$, defined by
\[
\mathbf{v}(x)=\boldsymbol B(x,{\bf u}(x))\quad x\in\Omega,
\]
belongs to $L_{\operatorname*{loc}}^{1}({\operatorname*{div};}\Omega)$ and
\[
\operatorname*{div}\mathbf{v}(x)=\operatorname*{div}\nolimits_{x}%
\boldsymbol B(x,{\bf u}(x))+\operatorname*{tr}\left( \nabla_{{\bf z}}%
\boldsymbol B(x,{\bf u}(x))\,\nabla{\bf u}(x)\right)
\]
for $\mathcal{L}^{N}$--a.e.\ $x\in\Omega$, provided
$\nabla_{{\bf z}}\boldsymbol B(x,\mathbf{u}(x))\,\nabla\mathbf{u}(x)$ is interpreted to be zero whenever
$\nabla\mathbf{u}(x)=0$, irrespective of whether $\nabla_{{\bf z}}\boldsymbol B(x,\mathbf{u}(x))$ is defined.
\end{theorem}
\subsection{Scalar case $u\in W^{1,1}(\Omega)$}
An important special case is given when $\boldsymbol B$ has the form
\begin{equation}
\boldsymbol B(x,u)=\int_{0}^{u}\boldsymbol b(x,s)\,ds, \label{A}%
\end{equation}
where $\boldsymbol b\colon\Omega\times\R^{N}{R}\rightarrow\R^{N}{R}^{N}$.
Clearly, when $\boldsymbol B$ is of class $C^1$ with respect to the second variable,
it is not a real restriction assuming that it is of the form \eqref{A}
for some vector field $\boldsymbol b$ which is continuous with respect to the same variable.
On the other hand, some assumptions in Theorems~\ref{chainb3reg}
and~\ref{chainb2} below can be stated in a more polished way in terms of $\boldsymbol b$.
In this case a formula has been established in \cite{DCFV2} (see Theorem 3.4) for scalar functions
$u\in W^{1,1}(\Omega)$ by assuming that, for every $t\in \R^{N}{R}$, $\boldsymbol B(\cdot,t)$ is an $L^\infty_{\rm{loc}}$ function whose distributional divergence is a Radon measure.
In the following, we shall denote by ${\mathcal{DM}^{\infty}}(\Omega)$ the space of all
vector fields belonging to $L^\infty(\Omega;\R^{N}{R}^N)$ whose divergence
in the sense of distribution is a Radon measure with finite total variation.
\begin{theorem}[$\mathcal{DM}^{\infty}$-dependence]\label{chainb3reg}
Let
$\boldsymbol b:\Omega\times\R^{N}{R}\rightarrow\R^{N}{R}^N$ be a locally
bounded Borel function. Assume that
\begin{enumerate}
\item[(i)] for ${\mathcal L}^N$-a.e.\ $x\in
\Omega$ the function $\boldsymbol b(x,\cdot)$ is continuous in $\R^{N}{R}$;
\item[(ii)] for ${\mathcal L}^1$-a.e.\ $t\in
\R^{N}{R}$ the function $\boldsymbol b(\cdot,t) $ belongs to $\mathcal{DM}^{\infty}(\Omega)$;
\item[(iii)] for any compact set $H\subset\R^{N}{R}$,
$$
\int_H|{\rm div}_x\boldsymbol b(\cdot,t)|(\Omega)\,dt<+\infty\,.
$$
\end{enumerate}
Then, for every $u\in W^{1,1}(\Omega)\cap L^{\infty}(\Omega)$,
the function ${\bf v}:\Omega\rightarrow\R^{N}{R}^N$,
defined by
\[
{\bf v}(x):= \boldsymbol B(x, u(x)) = \int_{0}^{u(x)}\!\boldsymbol b(x,t)\,dt\,,
\]
belongs to $\mathcal{DM}^{\infty}(\Omega)$ and for any
$\phi\in C_0^1(\Omega)$ we have
\[
\begin{split}
\int_\Omega\pscal{\nabla\phi(x)}{{\bf v}(x)}dx=-&\int_{-\infty}^{+\infty}\!\!dt\!\!\int_\Omega{\rm
sgn}(t)\chi^*_{\Omega_{u,t}}\phi(x)d\diver_{x}\boldsymbol b(x,t)
\\
-&\int_\Omega\phi(x)\pscal{\boldsymbol b(x,u(x))}{\nabla
u(x)}dx
\end{split}
\]
where
\begin{equation}
\label{f:out}
\Omega_{u,t} :=
\left\{x\in\Omega:\ \text{$t$ belongs to the segment of
endpoints $0$ and $u(x)$}\right\}
\end{equation}
and for a.e. $t$ the function $\chi^*_{\Omega_{u,t}}$ is the precise representative of the $BV$ function
$\chi_{\Omega_{u,t}}$.
\end{theorem}
\section{Nonautonomous chain rule for $u\in BV$}
\label{s:BV}
\subsection{The scalar case $u\in BV(\R^{N}{R}^{N})$}
The case of a scalar function $u\in BV(\R^{N}{R}^{N})$ is studied in the papers \cite{dcfv} and \cite{DCFV2}, where it is considered
again in the case $B = B(x,u)$ of the form \eqref{A}.
In the first paper the authors have established the validity of the chain rule by requiring a $W^{1,1}$ dependence with respect to the variable $x$,
while in the second one it is assumed only a $BV$ dependence with respect to the variable $x$\,.
\begin{theorem}[$BV$-dependence, see \cite{DCFV2}]
\label{chainb2}
Let
$b:\Omega\times\R^{N}{R}\rightarrow\R^{N}{R}$ be a locally
bounded Borel function. Assume that
\begin{enumerate}
\item[(i)] for ${\mathcal L}^1$-a.e.\ $t\in
\R^{N}{R}$ the function $b(\cdot,t) \in BV(\Omega)$;
\item[(ii)] for any compact set $H\subset\R^{N}{R}$,
\[
\int_H|D_xb(\cdot,t)|(\Omega)\, dt < +\infty\,.
\]
\end{enumerate}
Then, for every $u\in BV(\Omega)\cap L_{\rm loc}^{\infty}(\Omega)$,
the function $v:\Omega\rightarrow\R^{N}{R}$,
defined by
\[
v(x):=\int_{0}^{u(x)}\!b(x,t)\,dt\,,
\]
belongs to $BV_{\rm loc}(\Omega)$ and for any
$\phi\in C_0^1(\Omega)$ we have
\begin{equation}\label{chainrule}
\begin{split}
\int_{\Omega}\!\nabla\phi(x) v(x)\,dx = {} &
- \int_{-\infty}^{+\infty}\!\!dt\!\!\int_{\Omega}{\rm sgn}(t)
\chi^*_{\Omega_{u,t}}(x)\phi(x)\,dD_xb(x,t)\\
& -
\int_{\Omega}\!\phi(x)b^*(x,\widetilde
u(x))\nabla u(x)\,dx
-\int_{\Omega}\phi(x)b^*(x,\widetilde
u(x))\,dD^cu(x)
\\
& -
\!\int_{J_u}\!\!\phi(x)\nu_u(x)\,d\mathcal{H}^{N-1}(x)\!\!
\int_{u^-(x)}^{u^+(x)}\!\!b^*(x,t)\,dt,
\end{split}
\end{equation}
where $\Omega_{u,t}$ is the set defined in \eqref{f:out},
$J_u$ is the jump set of $u$,
and $\chi^*_{\Omega_{u,t}}$ and
$b^*(\cdot,t)$ are, respectively, the precise representatives of
$\chi_{\Omega_{u,t}}$ and $b(\cdot,t)$.
\end{theorem}
Notice that if $b(x,t)\equiv b(t)$, then (\ref{chainrule}) reduces
to the well known chain rule formula for the composition of $BV$
functions with a Lipschitz function, while, in the special case that
$b(x,t)\equiv b(x)$, (\ref{chainrule}) gives the
formula for the derivative of the product of two $BV$ functions.
\subsection{The vectorial case ${\bf u}\in BV_{\rm loc}(\R^{N}{R}^{N};\R^{N}{R}^h)$}
More recently, a very general formula has been proven in \cite{ACDD} (see also \cite{CD} for $N=1$) for vector functions ${\bf u}\in BV(\R^{N}{R}^{N},\R^{N}{R}^{h})$.
Here the function $B\colon\R^{N}{R}^{N}\times\R^{N}{R}^h\to\R^{N}{R}$
is required to satisfy the following assumptions:
\begin{enumerate}
\item[(a)] $x\mapsto B(x,{\bf z})$ belongs to $BV_{\rm loc}(\R^{N}{R}^{N})$ for all
${\bf z}\in\R^{N}{R}^h$;
\item[(b)] ${\bf z}\mapsto B(x,{\bf z})$ is continuously differentiable in $\R^{N}{R}^h$ for almost every $x\in\R^{N}{R}^{N}$.
\end{enumerate}
We will use
the notation $C_B$ to denote a Lebesgue negligible set of points
such that $B(x,\cdot)$ is $C^1$ for all $x\in\R^{N}{R}^{N}\setminus C_B$.
We assume that $B$ satisfies, besides (a) and
(b), the following \emph{structural assumptions}:
\begin{enumerate}
\item[(H1)]For some constant $M$, $|\nabla_{\bf z} B(x,{\bf z})|\le M$ for all
$x\in\R^{N}{R}^{N}\setminus C_B$ and ${\bf z}\in\R^{N}{R}^h$.
\item[(H2)]For any compact set $H\subset\R^{N}{R}^h$ there exists a modulus of continuity
$\tilde\omega_H$ independent of $x$ such that
\[
|\nabla_{\bf z} B(x,{\bf z})-\nabla_{\bf z} B(x,{\bf z}')|\leq \tilde \omega_H(|{\bf z}-{\bf z}'|)
\]
for all ${\bf z},\,{\bf z}'\in H$ and $x\in\R^{N}{R}^{N}\setminus C_B$.
\item[(H3)]For any compact set $H\subset\R^{N}{R}^h$ there exist a positive Radon measure
$\lambda_H$ and a modulus of continuity $\omega_H$ such that
\[
|\widetilde{D}_x B(\cdot,{\bf z})(A)-\widetilde{D}_x B(\cdot,{\bf z}')(A)|\le
\omega_H(|{\bf z}-{\bf z}'|)\lambda_H(A)
\]
for all ${\bf z},\,{\bf z}'\in H$ and $A\subset\R^{N}{R}^{N}$ Borel.
\item[(H4)] The measure
\begin{equation}\label{defsigma}
\sigma\defeq \bigvee_{{\bf z}\in\R^{N}{R}^h} |D_x B(\cdot,{\bf z})|,
\end{equation}
(where $\bigvee$ denotes the least upper bound in the space of
nonnegative Borel measures) is finite on compact sets, i.e.\ it is a
Radon measure.
\end{enumerate}
We can now canonically build a countably $\Haus{N-1}$-rectifiable
set $\mathcal N$ containing all jump sets of $B(\cdot,{\bf z})$ as follows.
Indeed, we define
\begin{equation}\label{bsigma}
\mathcal N=\bigl\{x: \limsup_{r\downarrow 0}
\frac{\sigma(B_r(x))}{\omega_{n-1}r^{n-1}}
>0\bigr\}.
\end{equation}
It can be checked that $\mathcal N$ is $\sigma$-finite with
respect to $\Haus{N-1}$ and it is countably $\Haus{N-1}$-rectifiable
(see \cite{ACDD}, Section~2, for details).
\begin{theorem}\label{chain ruleqwqw}
Let $B$ be satisfying (a), (b), (H1)-(H2)-(H3)-(H4) above.
Then
for any function ${\bf u}\in BV_{\rm loc}(\R^{N}{R}^{N};\R^{N}{R}^h)$,
the function ${\bf v}(x) := B(x,{\bf u}(x))$ belongs to $
BV_{\rm loc}(\R^{N}{R}^{N})$ and the following chain rule holds:
\begin{itemize}
\item[(i)] (diffuse part) $|D{\bf v}|\ll\sigma+|D{u}|$ and, for any Radon measure $\mu$
such that $\sigma+|D{u}|\ll\mu$, it holds
\begin{equation}\label{diffuse}
\frac{d\widetilde{D}{\bf v}}{d\mu} =\frac{d\widetilde{D}_x B(\cdot,\tilde
{\bf u}(x))}{d\mu} +\nabla_{\bf z}\tilde B(x,\tilde {\bf u}(x))\frac{d\widetilde{D}
{\bf u}}{d\mu}\qquad\text{$\mu$-a.e. in $\R^{N}{R}^{N}$.}
\end{equation}
\item[(ii)] (jump part)
$J_{\bf v}\subset \mathcal N\cup J_{\bf u}$
and, denoting by $u^\pm(x)$ and $B^\pm(x,{\bf z})$ the one-sided traces of
$u$ and $B(\cdot,{\bf z})$ induced by a suitable orientation of $\mathcal
N\cup J_{\bf u}$, it holds
\begin{equation}\label{jump}
D^j {\bf v}=\big(B^+(x,{\bf u}^+(x))-B^-(x,{\bf u}^-(x)\big)\nu_{\mathcal N\cup
J_{\bf u}}\Haus{N-1}\res \mathcal (\mathcal N\cup J_{\bf u})
\end{equation}
in the sense of measures.
\end{itemize}
Moreover for a.e.\ $x$ the map $y\mapsto B(y,{\bf u}(x))$ is
approximately differentiable at $x$ and
\begin{equation}\label{eq:appdif}
\nabla {\bf v}(x)=\nabla_xB(x,{\bf u}(x))+\nabla_\mathbf{z}B(x, {\bf u}(x))\nabla {\bf u}(x) \qquad
\text{$\Leb{N}$-a.e.\ in $\R^{N}{R}^{N}$\,.}
\end{equation}
\end{theorem}
Here the expression
\[
\frac{d\widetilde{D}_x B(\cdot,\tilde {\bf u}(x))}{d\mu}
\]
means the pointwise density of the measure $\widetilde{D}_x B(\cdot,{\bf z})$
with respect to $\mu$, computed choosing ${\bf z}=\tilde {\bf u}(x)$ (notice
that the composition is Borel measurable thanks to the
Scorza-Dragoni Theorem and Lemma 3.9 in \cite{ACDD}). Analogously,
the expression $\tilde B(x,{\bf z})$ is well defined at points $x$ such
that $x\notin S_{B(\cdot,{\bf z})}$ and it can be proved that
$\nabla_{\bf z}\tilde B(x,{\bf z})$ is well defined for all ${\bf z}$ out of a
countably $\Haus{N-1}$-rectifiable set of points $x$.
\section{A generalization to divergence--measure fields}
\label{s:div}
This section is devoted to the proof of the chain rule formula
for divergence--measure vector fields
(see Theorem~\ref{chainb3}).
More precisely, we shall consider the composition
${\bf v}(x) := \boldsymbol B(x, u(x))$ where
$u$ is a function of bounded variation,
$\boldsymbol B(x, \cdot)$ is of class $C^1$
and $\boldsymbol B(\cdot, t)$ is a divergence--measure field.
(See Section~\ref{ss:assumptions} for the complete list of assumptions.)
In order to obtain the formula in this general setting,
we need to assume \textsl{a-priori} the existence of strong traces of $\boldsymbol B$
on a countable
$\mathcal{H}^{N-1}$--rectifiable universal jump set $\mathcal{N}$.
We remark that this assumption is satisfied in the BV setting
recalled in Section~\ref{s:BV}.
Before stating our result,
in Section~\ref{ss:div}
we recall some basic facts on divergence--measure fields.
Then, after listing all the assumptions in Section~\ref{ss:assumptions},
we prove some preliminary results in Section~\ref{ss:prel}.
Finally, in Section~\ref{ss:div2} we prove our main
uniqueness result.
\subsection{Divergence--measure fields }
\label{ss:div}
In what follows,
since the problem is local we shall assume that $\Omega=\R^{N}{R}^{N}$.
Moreover, we shall denote by ${\mathcal{DM}^{\infty}}$ the space of all
vector fields $\boldsymbol{A}\in L^\infty_{\rm loc}(\R^{N}{R}^N;\R^{N}{R}^N)$ whose divergence
in the sense of distribution is a Radon measure with locally finite total variation.
For a vector field \(\boldsymbol{A}\in{\mathcal{DM}^{\infty}}\)
we shall use the usual decomposition of a measure
\[
\Div \boldsymbol{A}
=: \Div^a\boldsymbol{A}\,{\mathcal L}^N+\Div^s\boldsymbol{A},
\]
where
$\diver^a\boldsymbol{A}\,{\mathcal L}^N$ is the absolutely continuous part and
$\diver^s\boldsymbol{A}$ is the singular part of $\diver\boldsymbol{A}$ with respect to the Lebesgue measure.
On the other hand, Chen and Frid in \cite{cf} proved that if $\boldsymbol{A}\in {\mathcal{DM}^{\infty}}$
then $\Div\boldsymbol{A}\ll\mathcal H^{N-1}$,
hence the singular part can be further decomposed as
\[
\Div^s \boldsymbol{A}
=: \diver^c\boldsymbol{A}+\diver^j\boldsymbol{A},
\]
where
$\diver^c\boldsymbol{A}(C) = 0$ if $C$ has $\sigma$--finite
$\mathcal{H}^{N-1}$--measure.
We shall also use the diffuse part $\widetilde{\Div}\boldsymbol{A}:=\diver^a\boldsymbol{A}\,{\mathcal L}^N + \diver^c\boldsymbol{A}$
of the measure $\Div\boldsymbol{A}$.
In the following, we shall denote by
\(\mathcal{SDM}^{\infty}\) the space of all
vector fields \(\boldsymbol{A}\in{\mathcal{DM}^{\infty}}\) such that
\(\Div^c\boldsymbol{A} = 0\).
Given a domain $\Omega\subset\R^{N}{R}^N$ of class $C^1$,
we can define the
trace of the normal component of $\boldsymbol{A}$ on $\partial\Omega$
as a distribution as follows:
\begin{equation}\label{f:ntrace}
\pscal{\text{Tr}(\boldsymbol{A}, \partial\Omega)}{\varphi}
:= \int_{\Omega} \nabla\varphi \cdot \boldsymbol{A}\, dx + \int_{\Omega} \varphi\;
d \Div\boldsymbol{A},
\qquad \forall\varphi\in C^\infty_c(\R^{N}{R}^N)
\end{equation}
(see \cite{Anz,cf}).
This notion of distributional trace can be extended to any
$\mathcal{H}^{N-1}$--rectifiable set $\mathcal{J} \subset \R^{N}{R}^N$.
In particular, it is possible to define the traces
\(\text{Tr}^\pm(\boldsymbol{A}, \mathcal{J})\) in such a way that
\(\text{Tr}^-(\boldsymbol{A}, \partial\Omega) = \text{Tr}(\boldsymbol{A}, \partial\Omega)\)
for every $\Omega \Subset \R^{N}{R}^N$ of class $C^1$
(see \cite{ACM}, Definition~3.3).
Unfortunately, these distributional traces are in general too weak
to be used in a chain rule (see the discussion in \cite{ACM} and \cite{ADM}).
For our purposes we need the notion of (strong) traces given below.
\begin{definition}[Traces]\label{def:tracce}
Let $u\in L^\infty_{\rm loc}(\R^{N}{R}^{N})$ and
let $\mathcal J\subset \R^{N}{R}^{N}$ be a countably $\H^{N-1}$-rectifiable set oriented by a normal vector
field $\nu$.
We say that two Borel functions \(u^{\pm}\colon\mathcal{J}\to\R^{N}{R}\)
are the traces of \(u\) on \(\mathcal{J}\)
if for \(\H^{N-1}\)-almost every \(x\in\mathcal{J}\)
it holds
\[
\lim_{r\to 0^+} \int_{B_r^\pm(x)} |u(y) - u^\pm(x)|\, dy = 0,
\]
where
$B_r^\pm(x) := B_r(x) \cap \{ y\in\R^{N}{R}^{N} : \pm \langle y-x,\nu(x)\rangle \geq 0 \}$.
\end{definition}
The following result is a particular case of Lemma~3.1 in \cite{CDDG}.
\begin{lemma}\label{lemma:campo2}
Let $\mathcal J$ be a countably $\mathcal{H}^{N-1}$--rectifiable set
with oriented normal vector $\nu$.
Let \(\boldsymbol{A}\in \mathcal{DM}^{\infty}\)
and assume that $\boldsymbol{A}$ admits traces $\boldsymbol{A}^\pm$ on $\mathcal J$
for \(\mathcal H^{N-1}\) almost every $z\in\mathcal J$.
Then
it holds
\[
\Div \boldsymbol{A}\res \mathcal J = \pscal{\boldsymbol{A}^+ - \boldsymbol{A}^-}{\nu}\, \mathcal H^{N-1}\res \mathcal J\,.
\]
\end{lemma}
\medskip
\noindent
\subsection{Assumptions on the vector field \(\boldsymbol b\).}
\label{ss:assumptions}
Let $\boldsymbol b:\R^{N}{R}^{N}\times\R^{N}{R}\rightarrow\R^{N}{R}^N$ be a locally
bounded Borel function satisfying the following conditions:
\begin{enumerate}
\item[(i)] for ${\mathcal L}^N$-a.e.\ $x\in
\R^{N}{R}^{N}$ the function $\boldsymbol b(x,\cdot)$ is continuous in $\R^{N}{R}$,
uniformly w.r.t.\ $x$;
\item[(ii)] for ${\mathcal L}^1$-a.e.\ $t\in
\R^{N}{R}$ the function $\boldsymbol b(\cdot,t) \in {\mathcal{DM}^{\infty}}$;
\item[(iii)] the measure
\begin{equation}\label{defsigmabarra}\nonumber
\sigma\defeq \bigvee_{t\in \R^{N}{R}} |\Div_x \boldsymbol b(\cdot,t)|
\end{equation}
is a Radon measure;
\item[(iv)] for any compact set $H\subset\R^{N}{R}$ there exist a positive Radon measure
$\lambda_H$ and a modulus of continuity $\omega_H$ such that
\[
|\widetilde\Div_x \boldsymbol b(\cdot,t)(C)-\widetilde\Div_x \boldsymbol b(\cdot,w)(C)|\le
\omega_H(|t-w|)\lambda_H(C)
\]
for all $t,w\in H$ and $C\subset\R^{N}{R}^{N}$ Borel.
\end{enumerate}
Let us define the singular set
\begin{equation}
\label{f:N}
{\mathcal N}=\Big\{x\in \R^{N}{R}^{N}: \liminf_{r \to 0} \frac{\sigma(B_r(x))}{r^{N-1}}>0\Big\}.
\end{equation}
We assume that
\begin{itemize}
\item[(v)]
$\mathcal N$ is a countably $\mathcal{H}^{N-1}$--rectifiable set and
$\H^{N-1}(\mathcal N\cap K)<+\infty$
for every compact set $K\subseteq \R^{N}{R}^{N}$.
\end{itemize}
In the following, \(\nu\) will always denote an oriented normal vector field on \(\mathcal{N}\).
We remark that, in the setting of Theorem~\ref{chain ruleqwqw},
the rectifiability of the singular set $\mathcal N$
follows from (i)--(iv).
Furthermore we assume that
\begin{itemize}
\item[(vi)]
for every $t\in\R^{N}{R}$ and for \(\H^{N-1}\) almost every $x\in\R^{N}{R}^{N}\setminus \mathcal N$ there exists the limit
\begin{equation}\label{f:vi}
\widetilde \boldsymbol b(x,t)=\lim_{r\to 0} \mean{B_r(x)} \boldsymbol b(y,t)\,dy.
\end{equation}
\end{itemize}
Without loss of generality we shall always assume that
$\boldsymbol b(x,t)=\widetilde \boldsymbol b(x,t)$
on points \((x,t)\) where \eqref{f:vi} holds.
By using (vi), as in \cite[Section 3]{ACDD},
we can prove that there exists a set $\mathcal N_0 $ with \(\H^{N-1}(\mathcal N_0)=0\) such that for every point \(x\in\R^{N}{R}^{N}\setminus (\mathcal N\cup \mathcal N_0)\) and every ${\bf w}\in \R^{N}{R}$ there exists the limit
\[
\widetilde \boldsymbol b(x,t)=\lim_{r\to 0} \mean{B_r(x)} \boldsymbol b(y,t)\,dy.
\]
Moreover we consider the following assumption on the traces of the vector field \(\boldsymbol b\).
\begin{itemize}
\item[(vii)]
For every \(t\in\R^{N}{R}\),
the function \(\boldsymbol b(\cdot, t)\) admits (strong) traces
\(\boldsymbol b^\pm(\cdot,t)\) on \(\mathcal{N}\).
\end{itemize}
In what follows, we shall use the notation
\(\beta^\pm(x,t) := \pscal{\boldsymbol b^\pm(x,t)}{\nu(x)}\),
\(t\in\R^{N}{R}\), \(x\in\mathcal{N}\).
\begin{remark}\label{anzellotti}
By assumption (ii) it follows that, for every \(t\in\R^{N}{R}\),
the vector field \(\boldsymbol b(\cdot, t)\) admits distributional traces
\(\text{Tr}^\pm(\boldsymbol b(\cdot, t), \mathcal{N})\) on \(\mathcal{N}\)
in the sense of Anzellotti
(see \cite{Anz,cf}).
Nevertheless, this notion of trace is too weak
in order to obtain the chain rule formula.
\end{remark}
\medskip
As in \cite[Prop.~3.2(ii)]{ACDD},
it can be proved that there exists
a Borel set \(\mathcal{N}_1\subseteq \mathcal{N}\)
such that, for every \(x\in\mathcal{N}_1\),
the traces \(\boldsymbol b^\pm(x,t)\) are defined for every \(t\in\R^{N}{R}\) and are continuous in \(t\).
By (vii) and Proposition \ref{lemma:campo2} we have
\begin{equation}
\label{mmm}
\frac{d\Div_x^j\boldsymbol b(\cdot,t)}{d{\mathcal H}^{N-1}}(x)=\beta^+(x,t)-\beta^-(x,t)
\end{equation}
for every ${\bf w}\in \R^{N}{R}$ and for $\H^{N-1}$ a.e. $x\in {\mathcal N}$.
If assumptions (i)--(vii) hold, then for every $t\in\R^{N}{R}$ the decomposition formula
\begin{equation}
\label{decomp}
(\Div_x \boldsymbol b)(\cdot,{\bf w})=(\Div^a_x \boldsymbol b)(x,{\bf w})\,{\mathcal L}^{N}+\frac{\Div_x^c\boldsymbol b(\cdot,t)}{d\sigma}(x)\sigma+
\left[\beta^{+}(x,{\bf w})-\beta^{-}(x,{\bf w})\right]\, {\mathcal H}^{N-1} \res{{\mathcal N}}
\end{equation}
holds in the sense of measures and there exists a Borel set ${\mathcal N_2}\subseteq \R^{N}{R}^{N}$ with $\sigma({\mathcal N_2})=0$ such that
the following limit
\[
\lim_{r\downarrow 0}\frac{\widetilde\Div_x\boldsymbol b(\cdot,t)(B_r(x))}{\sigma(B_r(x))}
=\frac{d\widetilde\Div_x\boldsymbol b(\cdot,t)}{d\sigma}(x)
\]
exists for every $x\in \R^{N}{R}^{N}\setminus {\mathcal N_2}$ and for every $t\in\R^{N}{R}$ and this equality holds,
where $\frac{d\Div_x\boldsymbol b(\cdot,t)}{d\sigma}(x)$ is the Radon-Nikod\'ym derivative at $x$ of the measure $\Div_x\boldsymbol b(\cdot,t)$ w.r.t.\ $\sigma$.
In particular we have that there exists a Borel set ${\mathcal N_3}\subseteq \R^{N}{R}^{N}$ with ${\mathcal L}^N({\mathcal N_3})=0$ such that
the following limit
$$
\lim_{r\downarrow 0}\frac{\Div^a_x\boldsymbol b(\cdot,t)(B_r(x))}{{\mathcal L}^N(B_r(x))}=\frac{d\Div^a_x\boldsymbol b(\cdot,t)}{d{\mathcal L}^N}(x)
$$
exists for every $x\in \R^{N}{R}^{N}\setminus {\mathcal N_3}$ and for every $t\in\R^{N}{R}$ and this equality holds,
where $\frac{d\Div^a_x\boldsymbol b(\cdot,t)}{d\sigma}(x)$ is the Radon-Nikod\'ym derivative at $x$ of the measure $\Div^a_x\boldsymbol b(\cdot,t)$ w.r.t.\ ${\mathcal L}^N$.
Similarly, there exists a Borel set ${\mathcal N_4}\subseteq \R^{N}{R}^{N}$ with $\sigma({\mathcal N_4})=0$ such that
the following limit
$$
\lim_{r\downarrow 0}\frac{\Div^c_x\boldsymbol b(\cdot,t)(B_r(x))}{\sigma(B_r(x))}=\frac{d\Div^c_x\boldsymbol b(\cdot,t)}{d\sigma}(x)
$$
exists for every $x\in \R^{N}{R}^{N}\setminus {\mathcal N_4}$ and for every $t\in\R^{N}{R}$ and this equality holds,
where $\frac{d\Div^c_x\boldsymbol b(\cdot,t)}{d\sigma}(x)$ is the Radon-Nikod\'ym derivative at $x$ of the measure $\Div^c_x\boldsymbol b(\cdot,t)$ w.r.t.\ $\sigma$.
\subsection{Preliminary results}
\label{ss:prel}
Let us define
\begin{equation}\label{defB}
\boldsymbol B(x,t)=\int_0^t \boldsymbol b(x,w)\,dw\,.
\end{equation}
\begin{proposition}\label{kkkk}
Let $\boldsymbol b:\R^{N}{R}^{N}\times\R^{N}{R}\rightarrow\R^{N}{R}^N$ be a Borel function satisfying {\rm (i)--(vii)}.
We have that
\begin{enumerate}
\item[(a)]
for every $t\in\R^{N}{R}$ the function $x\mapsto \boldsymbol B(x,t)\,dt$ belongs to $\mathcal{DM}^{\infty}\,;$
\item[(b)] for every $t\in\R^{N}{R}$
$$
\Div_x\boldsymbol B(x, t)<\!\!<\sigma\,;
$$
\item[(c)] the equality
$$\Div_x\boldsymbol B(x, t)=
\left[\int_{0}^{t}
\frac{d\Div_x\boldsymbol b}{d\sigma}(x,w)\,dw\right]\,d\sigma
$$
holds in the sense of measures for every $t\in\R^{N}{R}$.
\end{enumerate}
\end{proposition}
\begin{proof} Since
$$
\left|\frac{d \Div_x\boldsymbol b}{d \sigma}(x,w)
\right|\leq 1\,,
$$
for every test function $\phi\in C^1_0(\R^{N}{R}^{N})$, for
$\H^{N-1}$ a.e. $x\in\R^{N}{R}^{N}$
and for every $t\in\R^{N}{R}$, we have
\begin{equation}\nonumber%
\begin{split}
& \int_{\R^{N}{R}^{N}}\nabla\phi(x)\boldsymbol B(x, t)\,dx=
\int_{0}^{t}\left[\int_{\R^{N}{R}^{N}}\nabla\phi(x)
\boldsymbol b(x,w)\,dx\right]dw
\\
=&-\int_{0}^{t}\left[\int_{\R^{N}{R}^{N}}\phi(x)
d \Div_x\boldsymbol b(\cdot,w)\right]dw
=-\int_{0}^{t}\left[\int_{\R^{N}{R}^{N}}\phi(x)
\frac{d \Div_x\boldsymbol b}{d \sigma}(x,w)\,d \sigma\right]dw
\\
=&-\int_{\R^{N}{R}^{N}}\phi(x)\left[\int_{0}^{t}
\frac{d \Div_x\boldsymbol b}{d \sigma}(x,w)\,dw\right]d \sigma.
\qedhere
\end{split}
\end{equation}
\end{proof}
\begin{corollary} \label{bbbb} Under the previous assumptions, we have
\[
\Div^a_x\boldsymbol B(x, t) =
\int_{0}^{t}
\frac{d\Div^a_x\boldsymbol b}{d{\mathcal L}^N}(x,w)\,dw
\quad\forall x\in\R^{N}{R}^N\setminus\mathcal{N}_3,
\]
and
$$\Div^c_x\boldsymbol B(x, t)=
\left[\int_{0}^{t}
\frac{d\Div^c_x\boldsymbol b}{d\sigma}(x,w)\,dw\right]\,d\sigma
$$
in the sense of measures.
\end{corollary}
\begin{corollary}\label{lololo}
Let $\boldsymbol b$ be a Borel function satisfying {\rm (i)--(vii)}
and let \(\boldsymbol B\) be the vector field defined in \eqref{defB}.
Then it holds:
\begin{enumerate}
\item[(a)]
For every \(x\in{\mathcal N}\setminus {\mathcal N_1}\)
and for every \(w\in\R^{N}{R}\)
one has
\[
\lim_{r\downarrow 0}
\mean{B^\pm_r(x)}
|\boldsymbol B(y,w) - \boldsymbol B^\pm(x,w)|\, dy = 0,
\]
where
\[
\boldsymbol B^\pm(x,w) := \int_0^w \boldsymbol b^{\pm}(x,t)\, dt.
\]
\item[(b)]
For every $x\in\R^{N}{R}^{N}\setminus({\mathcal N}\cup\mathcal N_1)$
and every \(w\in\R^{N}{R}\) one has
\[
\lim_{r\downarrow 0} \mean{B_r(x)}\left|\boldsymbol B(y,w)-\widetilde{\boldsymbol B}(x,w)\right|\,dy=0,
\]
where
\[
\widetilde{\boldsymbol B}(x,w) := \int_0^w \boldsymbol b(x,t)\, dt.
\]
\item[(c)]
The equality
\[
\begin{split}
\Div_x \boldsymbol B(x,t)
= {}&\left[\int_{0}^{w}
\nabla_x\boldsymbol b(x,t)\,dt\right]
\,d{\mathcal L}^N
+\left[\int_{0}^{w}
\frac{dD_x^c\boldsymbol b}{d\sigma}(x,t)\,dt\right]
\,d\sigma
\\
& +
\pscal{\boldsymbol B^+(x,t)-\boldsymbol B^-(x,t)}{\nu(x)}
\,d{\mathcal H}^{N-1}\res{{{\mathcal N}}}
\end{split}
\]
holds in the sense of measures.
\end{enumerate}
\end{corollary}
\begin{proof}
(a) By (vii) for every $x\in {\mathcal N}\setminus\mathcal N_1$ and for every $w\in\R^{N}{R}$ we have
\begin{equation}\nonumber%
\begin{split}
\lim_{r\downarrow 0} &
\mean{B^\pm_r(x)}\left|\int_{0}^{w}\boldsymbol b(y,t)\,dt-\int_{0}^{w}\boldsymbol b^\pm(x,t)\,dt\right|dy\\
\leq &\int_{0}^{w}\lim_{r\downarrow 0} \mean{B^\pm_r(x)}|\boldsymbol b(y,t)-\boldsymbol b^\pm(x,t)|\,dy\, \,dt=0.
\end{split}
\end{equation}
\noindent
(b) Similarly for every $x\in\R^{N}{R}^{N}\setminus( {\mathcal N}\cup\mathcal N_1)$ and for every $w\in\R^{N}{R}$
we have
\begin{equation}\nonumber%
\begin{split}
\lim_{r\downarrow 0} & \mean{B_r(x)}\left|\int_{0}^{w}\boldsymbol b(y,t)\,dt-\int_{0}^{w}\widetilde \boldsymbol b(x,t)\,dt\right|dy\\
\leq &\int_{0}^{w}\lim_{r\downarrow 0} \mean{B_r(x)}|\boldsymbol b(y,t)-\widetilde \boldsymbol b(x,t)|\,dy\, \,dt=0.
\end{split}
\end{equation}
\noindent
(c) Follows from (a), (b) and Corollary~\ref{bbbb}.
\end{proof}
\subsection{Main result}
\label{ss:div2}
\begin{theorem}[$\mathcal{DM}^{\infty}$-dependence]\label{chainb3}Let $\boldsymbol b:\R^{N}{R}^{N}\times\R^{N}{R}\rightarrow\R^{N}{R}^N$ be a locally
bounded Borel function satisfying conditions (i)--(vii). Then, for every $u\in BV_{\rm loc}(\R^{N}{R}^{N})\cap L^{\infty}_{\rm loc}(\R^{N}{R}^{N})$,
the function ${\bf v}:\R^{N}{R}^{N}\rightarrow\R^{N}{R}^N$,
defined by
\[
{\bf v}(x):=\boldsymbol B(x,u(x)
\,,
\]
belongs to $\mathcal{DM}^{\infty}$ and for any
$\phi\in C_0^1(\R^{N}{R}^{N})$ we have
\[
\begin{split}
\int_{\R^{N}{R}^{N}} & \pscal{\nabla \phi(x)}{{\bf v}(x)}
\, dx = {} \\
-&\int_{\R^{N}{R}^{N}}\phi(x)\Div_x^a\boldsymbol B(x,u(x))\,dx
-\int_{\R^{N}{R}^{N}}\phi(x)\frac{\Div_x^c\boldsymbol B}{d\sigma}(x,\widetilde u(x))\,d\sigma
\\
-&\int_{\R^{N}{R}^{N}}\!\phi(x)\pscal{\boldsymbol b(x,\widetilde u(x))}{\nabla u(x)}\,dx
-\int_{\R^{N}{R}^{N}}\phi(x)\pscal{\boldsymbol b(x, \widetilde{u}(x))}{\frac{D^c u}{|Du|} (x)}\, d\, |Du|(x)
\\ -&
\int_{{\mathcal N\cup J_u}}\phi(x)\pscal{
\boldsymbol B^+(x,u^+(x))
-\boldsymbol B^-(x,u^-(x))
}{\nu(x)}
\,d{\mathcal H}^{N-1}\,\,.
\end{split}
\]
\end{theorem}
\begin{proof}
\textsl{Step~1.}
Let us fix a test function \(\phi\in C^1_0(\R^{N}{R}^{N})\).
We claim that
\[
\begin{split}
\int_{\R^{N}{R}^{N}} \pscal{\nabla \phi(x)}{{\bf v}(x)} \, dx = {}
& -\int_{\R^{N}{R}} dt \int_{\R^{N}{R}^{N}} \sign(t) \chi^*_{\Omega_{u,t}} \phi(x) \, d\, \Div_x \boldsymbol b(x,t)
\\ &
- \int_{\R^{N}{R}^{N}} \phi(x)\, \pscal{\boldsymbol b(x, \widetilde{u}(x))}{\nabla u(x)}\, dx
\\ &
- \int_{\R^{N}{R}^{N}} \phi(x)\, \pscal{\boldsymbol b(x, \widetilde{u}(x))}{\frac{D^c u}{|Du|} (x)}\, d\, |Du|(x)
\\ & -
\frac12\int_{\mathcal N \cup J_u} \phi(x)
\left[\int_{u^-(x)}^{u^+(x)} \left[\beta^+(x,t) + \beta^-(x,t)\right]dt\right]\, d\H^{N-1}\,,
\end{split}
\]
where
$\Omega_{u,t}=\{x\in\R^{N}{R}^{N}:\,t$ belongs to the segment of
endpoints $0$ and $u(x)\}$ and $\chi^*_{\Omega_{u,t}}$ is the precise representative of the $BV$ function
$\chi_{\Omega_{u,t}}$.
In order to prove the claim,
it is enough
to use a regularization argument as in \cite{DCFV2}.
More precisely, if \(\boldsymbol b_\epsilon(\cdot, t) := \rho_\epsilon \ast \boldsymbol b(\cdot, t)\)
denotes the standard regularization of \(\boldsymbol b(\cdot, t)\),
and \({\bf v}_\epsilon(x) := \int_{0}^{u(x)} \boldsymbol b_\epsilon(x,t)\, dt\), then
\[
\begin{split}
\int_{\R^{N}{R}^{N}} \pscal{\nabla\phi(x)}{{\bf v}_\epsilon(x)}\, dx = {} &
-\int_{\R^{N}{R}^{N}} \phi(x) \, \int_0^{u(x)} \Div_x \boldsymbol b_\epsilon(x,t)\, dt
\\ & -
\int_{\R^{N}{R}^{N}} \phi(x)\, \pscal{\boldsymbol b_\epsilon(x, u)}{\nabla u}\, dx
\\ & -
\int_{\R^{N}{R}^{N}} \phi(x)\, \pscal{\boldsymbol b_\epsilon(x, \widetilde{u}(x))}{\frac{D^c u}{|Du|} (x)}\, d\, |Du|(x)
\\ & -
\int_{\mathcal N \cup J_u} \phi(x)\,
\int_{u^-(x)}^{u^+(x)} \pscal{\boldsymbol b_\epsilon(x,t)}{\nu(x)}\, dt
\, d\H^{N-1}(x)\,,
\end{split}
\]
and the claim follows by passing to the limit as \(\epsilon\to 0^+\), observing that,
by (vii),
for every \(t\in\R^{N}{R}\) it holds
\[
\lim_{\epsilon\to 0^+} \pscal{\boldsymbol b_\epsilon(x,t)}{\nu(x)} =
\frac{\beta^+(x,t)+\beta^-(x,t)}{2}
\qquad
\text{for}\ \H^{N-1}-a.e.\ x\in\mathcal{N}.
\]
(For details see the proofs of Theorems 1.1 and 3.4 in \cite{DCFV2}.)
\medskip\noindent
\textsl{Step~2.}
We assume for simplicity that $u\geq 0$.
We claim that
\begin{equation}
\begin{split}\nonumber
&\int_{\R^{N}{R}}\left[\int_{\R^{N}{R}^{N}}{\rm sgn}(t)\phi(x)
\chi^*_{\Omega_{u,t}}(x)\,d\Div_x \boldsymbol b((x,t)\right]
\,dt
\\ \nonumber
& =\int_{\R^{N}{R}^{N}}\phi(x)\Div_x^a\boldsymbol B(x,u(x))\,dx
-\int_{\R^{N}{R}^{N}}\phi(x)\frac{d\Div_x^c\boldsymbol B}{d\sigma}(x,\widetilde u(x))\,d\sigma
\\ \nonumber
&\quad +\frac12\int_{{\mathcal N\cup J_u}}\phi(x)\left[\int_{0}^{u^+(x)}[\beta^+(x,t)-\beta^-(x,t)]\,dt
+\int_{0}^{u^-(x)}[\beta^+(x,t)-\beta^-(x,t)]\,dt\right]
\,d{\mathcal H}^{N-1}\,.
\nonumber
\end{split}
\end{equation}
By using Lemma 2.2 in \cite{DCFV2} and by the decomposition formula \eqref{decomp}, we have
\begin{equation}
\begin{split}\nonumber
& \int_{-\infty}^{+\infty} \left[\int_{\R^{N}{R}^{N}}{\rm sgn}(t)\phi(x)
\chi^*_{\Omega_{u,t}}\,d\Div_x\boldsymbol b
(x,t)\right]
\,dt =
\int_{0}^{+\infty}\left[\int_{\R^{N}{R}^{N}}\phi(x)\chi^*_{\{{u>t\}}}\,d\Div_x\boldsymbol b(x,t)\right]\,dt
\\ \nonumber
& =\int_{0}^{+\infty}\left[\int_{\R^{N}{R}^{N}}\phi(x)
\chi_{\{u>t\}}\,\Div^a_x\boldsymbol b(x,t)dx\right]
\,dt
+\int_{0}^{+\infty}\left[\int_{\R^{N}{R}^{N}}\phi(x)
\chi_{\{\widetilde u>t\}}\,\frac{d\Div_x^c\boldsymbol b}{d\sigma}(x,t)\,d \sigma\right]
\,dt
\\ \nonumber
&\quad +\int_{0}^{+\infty}\left[\int_{{\mathcal N}\cup J_u}\phi(x)
\frac12\left[\chi_{\{u^+>t\}}+\chi_{\{u^->t\}}\right]\,[\beta^+(x,t)-\beta^-(x,t)]d{\mathcal H}^{N-1}\right]
\,dt
\,.
\nonumber
\end{split}
\end{equation}
By using Corollary \ref{bbbb}, the claim is proved.
\medskip
Finally,
the thesis follows from Step 1 and Step 2, observing that
\[
\begin{split}
&\frac12\int_{u^-(x)}^{u^+(x)} \left[\beta^+(x,t) + \beta^-(x,t)\right]dt\\
&+ \frac12\left[\int_{0}^{u^+(x)}[\beta^+(x,t)-\beta^-(x,t)]\,dt
+\int_{0}^{u^-(x)}[\beta^+(x,t)-\beta^-(x,t)]\,dt\right]\\
& =\int_{0}^{u^+(x)}\beta^+(x,t)\,dt
-\int_{0}^{u^-(x)}\beta^-(x,t)\,dt
\,.
\qedhere
\end{split}
\]
\end{proof}
\smallskip
\begin{corollary}\label{c:div}
Let \(h\in C^1(\R^{N}{R})\) be a function with bounded derivative,
let \({\bf{A}}\in \mathcal{DM}^{\infty}\) and
let
\[
{\mathcal N}=\Big\{x\in \R^{N}{R}^{N}: \liminf_{r \to 0} \frac{|\Div {\bf{A}} | (B_r(x))}{r^{N-1}}>0\Big\}.
\]
Assume that:
\begin{itemize}
\item[(a)]
$\mathcal N$ is a countably $\mathcal{H}^{N-1}$--rectifiable set and
$\H^{N-1}(\mathcal N\cap K)<+\infty$
for every compact set $K\subseteq \R^{N}{R}^{N}$.
\item[(b)]
For every \(x\in\R^{N}{R}^N\setminus\mathcal{N}\),
one has \({\bf{A}}(x) = \widetilde{{\bf{A}}}(x)\).
\item[(c)]
The vector field \({\bf{A}}\) admits strong traces
\({\bf{A}}^\pm\) on \(\mathcal{N}\).
\end{itemize}
Then, for every $u\in BV(\R^{N}{R}^{N})\cap L^{\infty}(\R^{N}{R}^{N})$,
the function ${\bf v}:\R^{N}{R}^{N}\rightarrow\R^{N}{R}^N$,
defined by
\[
{\bf v}(x):={\bf{A}}(x)h(u(x)
\,,
\]
belongs to $\mathcal{DM}^{\infty}$ and for any
$\phi\in C_0^1(\R^{N}{R}^{N})$ we have
\[
\begin{split}
&\int_{\R^{N}{R}^{N}} \pscal{\nabla \phi(x)}{{\bf v}(x)}
\, dx = {} \\
-&\int_{\R^{N}{R}^{N}}\phi(x)\Div^a{\bf{A}}(x)h(u(x))\,dx
-\int_{\R^{N}{R}^{N}}\phi(x)\frac{\Div^c{\bf{A}}}{d\sigma}(x)h(\widetilde u(x))\,d\sigma
\\
-&\int_{\R^{N}{R}^{N}}\!\phi(x)h^\prime(\widetilde u(x))\pscal{{\bf{A}}(x)}{\nabla u(x)}\,dx
-\int_{\R^{N}{R}^{N}}\phi(x)h^\prime(\widetilde u(x))\pscal{{\bf{A}}(x)}{\frac{D^c u}{|Du|} (x)}\, d\, |Du|(x)
\\ -&
\int_{{\mathcal N\cup J_u}}\phi(x)
\left[h(u^+(x))\pscal{{\bf{A}}^+(x)}{\nu(x)}-h(u^-(x))\pscal{{\bf{A}}^-(x)}{\nu(x)}
\right]
\,d{\mathcal H}^{N-1}\,\,.
\end{split}
\]
\end{corollary}
\begin{remark}
The theory of divergence--measure vector fields is due
to G.~Anzellotti \cite{Anz}
(see also G.--Q.~Chen and
H.~Frid \cite{cf} for its generalization
and \cite{ScSc}).
He introduced the ``dot product" of a bounded vector field $\boldsymbol{A}$, whose divergence is a Radon measure, and the gradient $Du$ of $u\in BV(\Omega)$ through a pairing $(\boldsymbol{A},Du)$ which defines a Radon measure.
He also defined the normal trace of a vector field through the boundary and establish a generalized Gauss--Green formula.
Consider now $\mu=\hbox{\rm div}\,\boldsymbol{A}$ with $\boldsymbol{A}\in{\mathcal{DM}^{\infty}}$ and let $u\in
BV_{\rm loc}(\R^{N}{R}^N)\cap L^\infty_{\rm{loc}}(\R^{N}{R}^N)$.
The distribution defined by the following expression
\[
\langle(\boldsymbol{A},Du),\varphi\rangle=-\int u^*\varphi\,d\mu-\int
u\boldsymbol{A}\cdot\nabla\varphi,\quad \varphi\in C_0^\infty(\R^{N}{R}^N)\,,
\]
is actually a Radon measure and its total variation $\vert (\boldsymbol{A}, Du) \vert$ is absolutely
continuous with respect to the measure $\vert Du \vert$.
Therefore the following Anzellotti formula holds
\begin{equation}\label{Anzellotti}
\Div(u\boldsymbol{A})=u^*\Div\boldsymbol{A}+(\boldsymbol{A},Du).
\end{equation}
in the sense of measures.
Using the distributional normal trace
defined in \eqref{f:ntrace},
the following Green formula holds
\begin{equation}\label{GreenI}
\int_{\Omega} u^* \, d\mu + \int_{\Omega} (\boldsymbol{A}, Du) =
\int_{\partial \Omega}
\text{Tr}(\boldsymbol{A}, \partial\Omega) \,
u \ d\mathcal H^{N-1}\,.
\end{equation}
We remark that,
if we apply \eqref{GreenI} to a vector field $\boldsymbol{A} \in \mathcal{DM}^{\infty}(\Omega)$ and the constant $u\equiv 1$, since $(\boldsymbol{A}, Du)=0$ we obtain
\begin{equation}\label{GreenII}
\int_{\Omega} \Div\boldsymbol{A} =
\int_{\partial \Omega}
\text{Tr}(\boldsymbol{A}, \partial\Omega)
\ d\mathcal H^{N-1}\,.
\end{equation}
\end{remark}
\begin{remark}\label{remrem}
The case \(u\in W^{1,1}\) in Theorem~\ref{chainb3} has been already treated
in \cite{DCFV2} (see also \cite{dcl}).
The representation formula in Theorem~\ref{chainb3} can be written as the following equality in the sense of measures
\begin{equation}\label{lolo}
\begin{split}
\Div(\boldsymbol B(x,u(x)))=&\left.\Div_x\boldsymbol B(x,t)\right|_{t=u(x)}\\
+&\pscal{\boldsymbol b(x, \widetilde{u}(x))}{\widetilde Du}
+\pscal{
\boldsymbol B^*(x,u^+)
-\boldsymbol B^*(x,u^-)
}{\nu}
{\mathcal H}^{N-1}\res {\mathcal N\cup J_u}
\end{split}
\end{equation}
with the compact notation
$$
\left.\Div_x \boldsymbol B(\cdot,t)\right|_{t=u(x)}=\frac12\left[\Div_x \boldsymbol B(\cdot,u^+(x))+\Div_x \boldsymbol B(\cdot,u^-(x))\right]\,.
$$
\end{remark}
\begin{remark}
The formula in Corollary~\ref{c:div} can be written as the following equality in the sense of measures:
\[
\Div({\bf{A}}(x)h(u(x)))=\Div{\bf{A}}(x)\,h(u)^*(x)+h^\prime(u(x))\pscal{{\bf{A}}(x)}{Du}.
\]
If $u\in BV$ and $u{\bf{A}}\in \mathcal{DM}^{\infty}$, then $$\pscal{{\bf{A}}(x)}{Du}=-u^* \Div{\bf{A}}+\Div(u{\bf{A}}).$$
Hence the following formula holds
\[
\begin{split}
\Div({\bf{A}}(x)h(u))=&\Div{\bf{A}}\,h(u)^*+h^\prime(u)[-u \Div{\bf{A}}+\Div(u{\bf{A}})]
\\
=&[h(u)^*-uh^\prime(u)]\Div{\bf{A}}+h^\prime(u)\Div(u{\bf{A}}).
\end{split}
\]
(For a similar formula when $u$ is not a function of bounded variation
see \cite{ADM}.)
If $\boldsymbol B(x,u(x))={\bf{A}}(x)u(x)$, then we obtain the Anzellotti formula \eqref{Anzellotti}.
\end{remark}
\section{Applications to conservation laws}
\label{s:cons}
Let us consider the multidimensional scalar conservation law with discontinuous flux
\begin{equation}\label{f:scalar}
u_t + \Div \boldsymbol B(x,u) = 0,
\qquad (t,x)\in (0, +\infty)\times\R^{N}{R}^N,
\end{equation}
where, as in the previous section,
the vector field \(\boldsymbol B\) is defined as in \eqref{defB} by
\(\boldsymbol B(x,t) := \int_0^t \boldsymbol b(x,w)\, dw\).
Here the Borel function \(\boldsymbol b\colon\R^{N}{R}^N\times\R^{N}{R}\to\R^{N}{R}^N\)
satisfies slightly stronger assumptions compared with
(i)--(vii) of Section~\ref{s:div}.
We are interested in proving the Kato contraction property for
entropy solutions, in the sense of Definition~\ref{d:entrsol}, of \eqref{f:scalar},
see Theorem~\ref{thm:uniquenessentropic}.
We stress here that, as in \cite{CDD}, our definition of entropy solution is restricted to
BV functions.
For the sake of completeness we collect the assumptions on \(\boldsymbol b\) here.
(A prime denotes the fact that the assumption has been modified
with respect to the corresponding one listed in
Section~\ref{ss:assumptions}.)
\begin{enumerate}
\item[(i)] for ${\mathcal L}^N$-a.e.\ $x\in\R^{N}{R}^{N}$ the function $\boldsymbol b(x,\cdot)$ is continuous in $\R^{N}{R}$
uniformly w.r.t.\ $x$;
\item[(ii')] for ${\mathcal L}^1$-a.e.\ $t\in
\R^{N}{R}$ the function $\boldsymbol b(\cdot,t)$ belongs to $\mathcal{SDM}^{\infty}$;
\item[(iii)] the measure
$\sigma$ defined in \eqref{defsigmabarra}
is a Radon measure.
\item[(iv')]
there exists a function \(g_1\in L^1_{\rm loc}(\R^{N}{R}^N)\) such that
\[
|\Div^a_x \boldsymbol b(x,t)-\Div^a_x \boldsymbol b(x,w)|\leq
g_1(x) \, |t-w|
\]
for all $t,w\in \R^{N}{R}$ and $x\in\R^{N}{R}^{N}$;
\item[(v')]
the set $\mathcal N$, defined in \eqref{f:N}, is a countably $\mathcal{H}^{N-1}$--rectifiable set and
$\H^{N-1}(\mathcal N)<+\infty$;
\item[(vi)] for every $t\in\R^{N}{R}$ and for \(\H^{N-1}\) almost every $x\in\R^{N}{R}^{N}\setminus \mathcal N$ there exists the limit~\eqref{f:vi};
\item[(vii)] for every \(t\in\R^{N}{R}\),
the function \(\boldsymbol b(\cdot, t)\) admits (strong) traces
\(\boldsymbol b^\pm(\cdot,t)\) on \(\mathcal{N}\).
\end{enumerate}
As in Section~\ref{s:div}, we shall use the notation
\(\beta^\pm(x,t) := \pscal{\boldsymbol b^\pm(x,t)}{\nu(x)}\),
\(t\in\R^{N}{R}\), \(x\in\mathcal{N}\).
In our context,
Theorem~\ref{chainb3} reads as follows:
For every \(u\in BV(\R^{N}{R}^{N})\) the composite function \({\bf v}(x)=\boldsymbol B(x ,u(x))\) belongs to \(\mathcal{DM}^{\infty}\) with
\begin{equation}\label{eq:stimab}
|\Div {\bf v}|\le \sigma+M|Du|
\end{equation}
and
\begin{align}
\label{eq:chainrule1b}
\widetilde{\Div}\, {\bf v} & =
\Div_x^a \boldsymbol B (x,\widetilde u(x))\, \Leb{N}+
\pscal{\boldsymbol b (x,\widetilde u(x))}{\widetilde{D} u}
\\
\label{eq:chainrule2b}
\Div^j {\bf v} &=
\pscal{\boldsymbol B^+(x,u^+(x))-\boldsymbol B^{-}(x,u^-(x))}{\nu(x)}\,\mathcal H^{N-1}\res (J_u\cup \mathcal N).
\end{align}
\begin{remark}\label{lincei}
Let us point out that our hypotheses include (and actually are modeled on) the case \(\boldsymbol B(x,u)=\widehat\boldsymbol B (k(x),u)\) where \(k\in S\mathcal{DM}^{\infty}\cap L^\infty(\R^{N}{R}^{N}; \R^{N}{R}^{N}) \), \(\H^{N-1}(J_{k})<+\infty\) and
\(\widehat \boldsymbol B\in C^1(\R^{N}{R}^{N}\times \R^{N}{R},\R^{N}{R}^{N})\cap{\rm Lip} (\R^{N}{R}^{N}\times \R^{N}{R},\R^{N}{R}^{N}) \).
\end{remark}
\begin{remark}\label{r:triang}
The assumption (vii) of existence of traces
may appear
too strong to be useful for applications.
On the other hand, a situation we have in mind is the following.
Let us consider the system
\begin{equation}\label{f:triang}
\begin{cases}
\Div \boldsymbol{A}_1(u) = 0,\\
\Div \boldsymbol{A}_2(u,v) = 0,
\end{cases}
\end{equation}
where \(u,v\colon \R^{N}{R}^{N}\to \R^{N}{R}\),
and the fluxes
\(\boldsymbol{A}_1\colon\R^{N}{R}\to\R^{N}{R}^{N}\), \(\boldsymbol{A}_2\colon \R^{N}{R}\times\R^{N}{R}\to\R^{N}{R}^{N}\)
are regular functions.
In general, a solution \(u\) of the first equation need not be of bounded variation.
Nevertheless, if \(\boldsymbol{A}_1\) is genuinely nonlinear, then
\(u\) has a quasi-BV structure, in the sense of
De Lellis, Otto and Westdickenberg (see \cite{DLOW}).
In particular, there exists a $\mathcal{H}^{N-1}$--rectifiable set \(\mathcal{N}\)
such that
\(u\) has left and right traces \(u^\pm\) on \(\mathcal{N}\),
and it has vanishing mean oscillation at every
\(x\not\in\mathcal{N}\).
The second component \(v\) is then a solution of the equation
\[
\Div \boldsymbol B(x, v) = 0
\]
where \(\boldsymbol B(x, v) := \boldsymbol{A}_2(u(x), v)\).
The vector field \(\boldsymbol B\) admits traces on \(\mathcal{N}\), given by
\(\boldsymbol B^{\pm}(x, v) = \boldsymbol B(u^\pm(x), v)\), \(x\in \mathcal{N}\), \(v\in\R^{N}{R}\).
In particular, assumption (vii) is satisfied.
Unfortunately, the quasi-BV structure of \(u\)
is not enough to have Assumption~\ref{ss:assumptions} (vii) satisfied,
since \(\boldsymbol B(\cdot, v)\) is only of bounded mean oscillation at every
point \(x\not\in \mathcal{N}\).
Nevertheless, we think that our analysis can be a good starting point
in order to obtain uniqueness results for the Cauchy problems
related to the evolutionary version of the triangular system \eqref{f:triang}.
\end{remark}
\subsection{Uniqueness of entropy solutions}
\begin{definition}[Convex entropy pair]
We say that $(S,\boldsymbol \eta)$ is a convex entropy pair
if $S\in C^2(\R^{N}{R})$ is a convex function,
and $\boldsymbol \eta = (\eta_1, \ldots, \eta_N)$ is defined by
\begin{equation}\label{f:eta}
\eta_{i}(x,v):=\int_{0}^{v} b_{i}(x,w)S'(w)dw\,,
\qquad i=1,\ldots,N.
\end{equation}
\end{definition}
In the above definition and in the sequel,
\(b_i=\boldsymbol b\cdot e_i\) are the components of \(\boldsymbol b\).
Note that according to the previous discussion, \(\boldsymbol \eta(\cdot, v) \in S\mathcal{DM}^{\infty}\) for every \(v\in \R^{N}{R}\) and its divergence is given by
\[
\begin{split}
\Div_x\boldsymbol \eta(\cdot,v)&=
\left(\int_0^v \Div^a_x\boldsymbol b(x,w)\,S'(w)dw\right) \Leb{N}
\\ &
+\left(\int_0^v
(\beta^+(x,w) - \beta^-(x,w))\, S'(w)\, dw
\right) \,\mathcal H^{N-1}\res \mathcal N\,.
\end{split}
\]
\begin{definition}[Entropy solutions]\label{d:entrsol}
A function
\[
u\in C([0,T]; L^1(\R^{N}{R}^{N})) \cap L^{\infty}((0,T)\times \R^{N}{R}^{N})\cap
BV((0,T)\times\R^{N}{R}^N)
\]
is an entropy solution of \eqref{f:scalar}
if
$u$ is a solution to \eqref{f:scalar} in the sense of distributions,
and there exists a (everywhere defined) Borel representative
$\hat{u}$ of $u$ with \(|\hat{u}(t,x)|\le \|u\|_{\infty}\) such that, for every
convex entropy pair $(S,\boldsymbol \eta)$,
one has
\begin{equation}\label{f:distr}
\begin{split}
\partial_t S(u)&+\Div \big(\boldsymbol \eta(x,u)\big)\\
&-\Div \boldsymbol \eta(x,v)\Big|_{v=\hat{u}(t,x)}+S'(\hat{u})\Div \boldsymbol B(x,v)\Big|_{v=\hat{u}(t,x)}\le 0
\end{split}
\end{equation}
in the distributional sense. Here, by \(\Div \boldsymbol B(x,v)\big|_{v=\hat{u}(t,x)}\)
we mean the measure whose action on a bounded and Borel function \(\varphi=\varphi(t,x)\) is given by
\begin{multline}\label{def:misuraresto}
\int _0^T dt \int_{\R^{N}{R}^{N}} \varphi(t,x)
\Div^a_x \boldsymbol B(x,\hat{u}(t,x)) \, dx
\\
+\int _0^T dt \int_{\mathcal N} \varphi(t,x)
\pscal{\boldsymbol B^+(x,\hat{u}(t,x))-\boldsymbol B^-(x,\hat{u}(t,x))}{\nu(x)}
\, d\H^{N-1}(x),
\end{multline}
and the same for \( \Div \boldsymbol \eta(x,v)\big|_{v=\hat{u}(t,x)}\).
\end{definition}
With these definitions at hand we can now restate our main result:
\begin{theorem}\label{thm:uniquenessentropic}
Let \(\boldsymbol b\) satisfy
the assumptions listed at the beginning of the Section,
and let \(u_1\) and \(u_2\) be two entropy solutions of \eqref{f:scalar}, then
\begin{equation}\label{eq:contrattivo}
\int_{\R^{N}{R}^{N}}|u_1(T,x)-u_2(T,x)|\,dx\le \int_{\R^{N}{R}^{N}}|u_1(0,x)-u_2(0,x)|\,dx.
\end{equation}
\end{theorem}
\subsection{Sketch of the proof of Theorem~\ref{thm:uniquenessentropic}}
The proof follows the lines of the one of Theorem~3.5 in \cite{CDD}.
We recall here only the main points,
giving references to the details in \cite{CDD}.
As a first reduction, it is not restrictive to consider
only non-negative solutions.
\medskip\noindent
\textbf{Kinetic formulation.}
Let us consider the measure-valued vector field
\[
\boldsymbol{a}(\cdot, v) :=
(\boldsymbol b(\cdot, v)\, {\mathcal L}^N_x \, , \, -\Div_x \boldsymbol B(\cdot, v)).
\]
Note that \(\boldsymbol{a}\) is a Radon measure and
\(\Div_{x,v} \boldsymbol{a} = 0\).
By a kinetic solution we mean a function
\[
u\in C([0,T]; L^1(\R^{N}{R}^{N})) \cap L^{\infty}((0,T)\times \R^{N}{R}^{N})\cap
BV((0,T)\times\R^{N}{R}^N)
\]
which is a distributional solution of \eqref{f:scalar},
and satisfies the following property:
there exists a (everywhere defined) Borel representative
$\hat{u}$ of $u$ with \(|\hat{u}(t,x)|\le \|u\|_{\infty}\),
and a positive measure \(m = m(t,x,v)\) with
\(m((0,T)\times\R^{N}{R}^{N+1}) < +\infty\), such that the function
\((t,x,v) \mapsto \chi(v, \hat{u}(t,x))\) satisfies
\begin{equation}
\label{f:kinetic}
\partial_t \chi( v, \hat{u}(t,x)) +
\Div_{x,v} [ \boldsymbol{a}(x,v) \chi(v, \hat{u}(t,x))]
= \partial_v m(t,x,v)
\end{equation}
in the sense of distributions.
Here the function \(\chi\) is defined by
\begin{equation}\label{defchi}
\chi(v,u):=
\begin{cases}
1
&\text{if $v<u$},\\
1/2
&\text{if $v=u$},\\
0
&\text{if $u<v$}.
\end{cases}
\end{equation}
The first step consists in proving that $u$
is an entropy solution of \eqref{f:scalar} if and only if it is
a kinetic solution.
The proof is almost the same of the one of Theorem~3.9 in \cite{CDD}.
We mention only that, if $u$ is an entropy solution, the kinetic measure $m$
is defined first of all as a distribution by
\begin{equation}\label{eq:8}
\begin{split}
\pscal{m}{\psi} & =
- \int_{(0,T)\times\R^{N}{R}^{N+1}} dt\, dx\, dv \, \partial_t \psi(t,x,v)\int_0^v \chi(w,u(t,x))\, dw \\
& -\int_{(0,T)\times\R^{N}{R}^{N+1}} dt\, dx\, dv \, \nabla_x \psi(t,x,v) \int_0^v \boldsymbol b(x,w) \chi(w,u(t,x)) \, dw
\\
& + \int_{(0,T)\times\R^{N}{R}^{N+1}} dt\, dx\, dv
\, \boldsymbol B(x,v) \chi(v, \hat{u}(t,x)) \nabla_x \psi(t,x,v).
\end{split}
\end{equation}
Reasoning as in \cite{CDD} we can prove that $m$ is a positive measure, with finite mass.
\medskip\noindent
\textbf{Kato contraction property.}
By Cavalieri's principle,
it is enough to prove the following contraction
property
(see \cite[Thm.~5.1]{CDD}).
\begin{theorem}\label{kin:uniqueness}
Let $u_1, u_2$ be two
entropy solution of \eqref{f:scalar},
with corresponding everywhere defined Borel
representatives $\hat{u}_1, \hat{u}_2$.
Setting $f_i(t,x,v) := \chi(v,\hat{u}_i(t,x))$, $i=1,2$, we have that
\[
\int_{\R^{N}{R}^{N+1}}\
|f_1-f_2|(T,x,v)
\ dx\,dv
\leq \int_{\R^{N}{R}^{N+1}}\
|f_1-f_2|(0,x,v)
\ dx\,dv \,.
\]
\end{theorem}
The proof of this theorem is obtained as a consequence of
the following intermediate result
(see \cite[Prop.~5.3]{CDD}).
\begin{proposition}\label{uniqueness1}
Let $u_1, u_2$ be two
entropy solution of \eqref{f:scalar},
with corresponding representatives $\hat{u}_1, \hat{u}_2$.
Setting $f_i(t,x,v) := \chi(v,\hat{u}_i(t,x))$, $i=1,2$, we have that
\begin{equation}\label{f:contr}
\begin{split}
\int_{\R^{N}{R}^{N+1}}\
|f_1-f_2|(T,x,v)
\ dx\,dv
&\leq \int_{\R^{N}{R}^{N+1}}\
|f_1-f_2|(0,x,v)
\ dx\,dv \\
&+\int_0^T \int_{\mathcal N} W(u_1,u_2)d\H^{N-1}dt
\,,
\end{split}
\end{equation}
where
\begin{equation}\label{57bis}
\begin{split}
W(u_1,u_2)
:= {}
& \pscal{\boldsymbol B^+(u_1^+)}{\nu}\big[-2\chi(u_1^+,u_2^+)+2\chi(u_1^-,u_2^-)\big]\\
+ &
\pscal{\boldsymbol B^+(u_2^+)}{\nu}\big[-2\chi(u_2^+,u_1^+)+2\chi(u_2^-,u_1^-)\big]\,.
\end{split}
\end{equation}
\end{proposition}
The proof of this proposition is long and
technical, but can be done exactly as in \cite{CDD}
using the new chain rule proved
in Section~\ref{s:div}.
Finally, Theorem~\ref{kin:uniqueness} follows
from Proposition~\ref{uniqueness1}
by proving that $W(u_1, u_2) \leq 0$
on $(0,T)\times\mathcal{N}$.
This is a purely algebraic fact that follows from the
entropy condition and the existence of traces,
see the proof of Theorem~5.1 in \cite{CDD} for details.
\bigskip
\textsc{Acknowledgments.}
The authors would like to thank
Guido De Philippis for some useful discussions
during the preparation of the manuscript.
|
\section{Introduction}
From the point of view of predictive statistical mechanics which is based on the maximum information entropy principle, with the exception of quantum mechanical probabilities, there is no reason to consider some particular probability distribution as the true distribution describing the system \cite{jaynes1}. Such a view is in a marked contrast to the interpretation that defines probability only in terms of the limit of a relative frequency of the outcome in an infinite sampling sequence, where the probabilities are therefore factual properties of the observed system \cite{feller}. From the law of large numbers it follows that the relative frequency of success in a sequence of e.g. Bernoulli trials (in a sequence of repeated independent trials of an experiment with only two possible outcomes) converges to the theoretical probability. For example, a fair coin toss is a Bernoulli trial where the theoretical probability that the outcome will be heads is equal to $1/2$.
According to the law of large numbers, the proportion of heads in a large number of fair coin tosses will converge to $1/2$ as the number of tosses approaches infinity. This means convergence in probability in the weak form of the law and convergence with probability one in the strong form, where the strong form of the law always implies the weak form of the law \cite{law of large numbers}. Accordingly, the relative frequency is a factual property of the real world that can be measured by repeating a large number of trials, or estimated from the theoretical probability. Probability, on the other hand, is something that we assign to individual events, or we calculate it for the composite events according to the rules (axioms) of probability theory, from the previously assigned probabilities of individual events.
In different applications of statistical mechanics, we try to predict the results of, or draw inferences from, some experiment that can be repeated indefinitely under what appears to be identical conditions (i.e. on the ensemble of identically prepared systems). Although traditional expositions of statistical mechanics such as \cite{penrose} define the probability as the limiting relative frequency in independent repetitions of a statistical experiment, the relation between frequencies and probabilities, implied by the law of large numbers, in statistical mechanics becomes very complex, because in reality for a macroscopic system, we do not measure the relative frequency of the occurrence of its individual microscopic states in a sequence of infinite or a large number of trials.
In the frequentist interpretation probabilities are always experimentaly verifiable, and consequently, one of the foundational problems of statistical mechanics would be to derive and to justify the probabilities of microscopic events, in the sense of frequencies in the ensemble of indentically prepared systems, from the first principles i.e. from equations of motion. This is the main problem of ergodic theory approach to statistical mechanics \cite{farquhar,dorfman}. Jaynes presented the opposite view, that if we choose to represent only the degree of our knowledge about the individual system, then there can not be anything physically real in the frequencies in the corresponding ensemble of a large number of systems, nor there is any sense in asking which ensemble is the only correct one \cite{jaynes2}. In the interpretation given by Jaynes, what we call different ensembles corresponds in reality to different degrees of knowledge about the individual system, or about some physical situation. In the argumentation of this viewpoint, Jaynes referred to the statement by Gibbs, according to which the ensembles are chosen only to illustrate the probabilites of events in the real word \cite{jaynes2,gibbs}.
The simplest interpretation of the Gibbs method of ensembles and the MaxEnt formalism follows from the fact that by maximizing the information entropy, which is also known as the uncertainty represented by a probability distribution, subject to given macroscopic constraints, one predicts just the macroscopic behaviour that can happen in the greatest number of microscopic realizations (i.e. greatest multiplicity) compatible with those constraints \cite{jaynes2,jaynes3,jaynes4,jaynes5}. Without going deeper into the problem of interpretation of probabilites, which is even more pronounced in the case of nonequilibrium states, it is more important that the distributions obtained from the application of the principle of maximum information entropy depend only on the available information and do not depend on arbitrary assumptions related to missing information. If we refer only to predictions, from the same viewpoint one can speak about the objectivity only in the extent in which the incompleteness of information about the system is taken into account. Consistent with this way of thinking, by applying the principle of maximum information entropy, we come to the relevant statistical distributions, and this is the subject of the paper.
The structure of the paper is as follows. Section \ref{sec_2} is a brief introduction on the Shannon's concept of information entropy \cite{shannon}, and on the principle of maximum information entropy and MaxEnt formalism formulated by Jaynes \cite{jaynes6,jaynes7}. Section \ref{sec_3} deals with the interpretation of MaxEnt formalism in statistical mechanics as given by Jaynes \cite{jaynes6} and Grandy \cite{grandy,grandy1}. Section \ref{sec_4} introduces the independent interpretation of probabilities in statistical mechanics on the basis of the principle of maximum information entropy. We modify here and extend the analysis given by Jaynes in \cite{jaynes8} and show that it has important consequences for the interpretation of probabilities. Section \ref{sec_5} is the conclusion summarizing the main results of the paper.
\section{Information entropy - measure of uncertainty - and the principle of maximum information entropy} \label{sec_2}
In Shannon's information theory \cite{shannon} the quantity of the form
\begin{equation}
H(p_1, \dots, p_n) = -K\sum_{i=1}^n p_i\log p_i \ , \label{eq24d}
\end{equation}
has a central role as a measure of information, choice and uncertainty for different probability distributions $p_1, \dots, p_n$. Starting from the understanding that the problem of constructing a communication device depends on the statistical structure of the information that is to be communicated (i.e. on the probabilities $p_1, p_2, \dots, p_n$ of the symbols $A_1, A_2, \dots, A_n$ of some alphabet) Shannon gave until that time the most general definition of the measure of amount of information. Sequences of symbols or "letters" may form the set of "words" of certain length, and the amount of information is measured analogously. Positive constant $K$ in (\ref{eq24d}) depends on the choice of a unit for the amount of information. In real applications expression (\ref{eq24d}), with the logarithmic base $2$ and $K=1$, represents the expected number of bits per symbol necessary to encode the random signal forming a memoryless source. But most importantly, Shannon's interpretation of the function (\ref{eq24d}) is not dependent on the specific context of information theory. He defined the function (\ref{eq24d}) as the measure of {\itshape uncertainty} related to the occurrence of possible events, or more specifically, as a measure of uncertainty {\itshape represented by the probability distribution} $p_1, p_2, \dots, p_n$. This is substantiated by three reasonable properties that are required from such a measure $H(p_1, \dots, p_n)$ that are sufficient to uniquely determine the form of this function: continuity, monotonic increase with number of possibilities in case when all probabilities are equal, and the unique and consistent composition law for the addition of uncertainties when mutually exclusive events are grouped into composite events. Shannon called the function (\ref{eq24d}) the entropy of the set of probabilities $p_1, p_2, \dots, p_n$.
However, we have still not answered an open question on how to determine or to choose the appropriate probability distribution for a particular problem or a system. The principle of maximum information entropy (MaxEnt) was formulated by Jaynes \cite{jaynes6,jaynes7} as a general criterion for construction of the probability distribution when the available information is not sufficient for the unique determination of the distribution. This principle is based on the following rationale: maximization of the information entropy (the uncertainty) subject to given constraints includes in the probability distribution only the information represented by these constraints. Therefore, predictions derived from such a probability distribution depend only on the available information and do not depend on arbitrary assumptions related to missing information.
The mathematical formulation of this principle is known as the MaxEnt algorithm. Let's consider it on the following example. Let the variable $x$ takes $n$ values $\{x_1, \dots, x_n\}$ with probabilities $\{p_1, \dots, p_n\}$ and the only data available are given by the expectation values of the functions $f_k(x)$:
\begin{equation}
F_k = \langle f_k (x)\rangle = \sum_{i = 1}^{n} p_i f_k (x_i) \ , \qquad k = 1, 2, \dots, m < n \ . \label{eq_expt_Fk_p}
\end{equation}
Probability distribution must also satisfy the normalization condition
\begin{eqnarray}
\sum _{i=1}^n p_i = 1 \ , \qquad \qquad \left (p_i \geq 0, \quad i = 1, 2, \dots, n \right ) \ . \label{eq_cond_norm_nonneg}
\end{eqnarray}
In most cases the available information given by the set of equations (\ref{eq_expt_Fk_p}) is far less then sufficient for the unique determination of the set of probabilities $\{p_1, \dots, p_n\}$, i.e. $m << n -1$. In such cases, probability distribution $\{p_1, \dots, p_n\}$ is determined by applying the MaxEnt principle. Probability distribution $\{p_1, \dots, p_n\}$ for which the information entropy (\ref{eq24d}) is maximum subject to the constraints (\ref{eq_expt_Fk_p}) is found by the method of Lagrange multipliers, i.e. by maximizing the function
\begin{eqnarray}
I & = & - \sum_{i=1}^n p_i \log p_i - (\lambda_0 - 1)\left (\sum _{i=1}^n p_i -1\right ) \nonumber\\
& & - \sum_{k = 1}^m \lambda _k \left (\sum_{i = 1}^{n} p_i f_k (x_i) - F_k\right ) \ ,
\end{eqnarray}
where $\lambda _0 - 1$, $\lambda _k$, $k=1, 2, \dots, m$, are the Lagrange multipliers. In this way we obtain the MaxEnt probability distribution
\begin{equation}
p_i = \frac{1}{Z}\exp \left \{ - \sum_{k = 1}^m \lambda _k f_k (x_i) \right \} \ , \qquad i = 1,2, \dots, n \ . \label{eq_maxent_p_i_Z_lk}
\end{equation}
The normalization factor $Z = e^{\lambda _0}$ which is also known as the partition function is given by
\begin{equation}
Z \equiv Z(\lambda _1 , \dots , \lambda _m) = \sum_{i = 1}^{n}\exp \left \{ - \sum_{k = 1}^m \lambda _k f_k (x_i) \right \} \ . \label{eq_maxent_part_fun_l_k}
\end{equation}
The expectation values of the functions $\langle f_k (x)\rangle = F_k $, $k = 1,2, \dots, m$, given by the conditions (\ref{eq_expt_Fk_p}), are equivalently given also by
\begin{equation}
F_k = \langle f_k (x)\rangle = - {\partial \log Z(\lambda _1 , \dots , \lambda _m) \over \partial \lambda _k} \ , \quad k = 1,2, \dots, m \ . \label{eq_F_k_expct_f_k_dlogZ_dl_k}
\end{equation}
Let's assume that set of $m + 1$ equations consisting of $m$ equations (\ref{eq_expt_Fk_p}) and the equation (\ref{eq_cond_norm_nonneg}) is consistent and that these equations are linearly independent. Then, using (\ref{eq_maxent_p_i_Z_lk}) and solving this set of equations, one can determine the Lagrange multipliers $\lambda _k$, $k=1,2, \dots, m$, as single-valued functions $\lambda _k(F)$ of the expected values $F = (F_1, \dots , F_m) $. The proof is given in \cite{kuic}. Then, by introducing the MaxEnt probability distribution (\ref{eq_maxent_p_i_Z_lk}) in the expression (\ref{eq24d}) for information entropy, the maximum of information entropy subject to the conditions (\ref{eq_expt_Fk_p}) and (\ref{eq_cond_norm_nonneg}) is obtained as the function of the expected values $F = (F_1, \dots , F_m) $:
\begin{equation}
(S_I)_{\mathrm{max}} = \log Z(\lambda _1 , \dots , \lambda _m) + \sum_{k = 1}^m \lambda _k F_k = S(F_1 , \dots, F_m) \ . \label{eq_S_I_max_eq_S_F}
\end{equation}
Assuming that the functions $\lambda _k(F)$, $k=1,2, \dots, m$, are continuously differentiable (or at least piecewise smooth), from (\ref{eq_F_k_expct_f_k_dlogZ_dl_k}) and (\ref{eq_S_I_max_eq_S_F}) it follows that
\begin{equation}
\lambda _k = {\partial S (F_1 , \dots, F_m) \over \partial F_k }\ , \qquad k = 1,2, \dots, m \ . \label{eq_L_eq_dS_F_dF}
\end{equation}
From equations (\ref{eq_F_k_expct_f_k_dlogZ_dl_k}), (\ref{eq_S_I_max_eq_S_F}) and (\ref{eq_L_eq_dS_F_dF}) it is obvious that the functions $\log Z(\lambda _1 , \dots, \lambda _m)$ and $S(F_1 , \dots, F_m)$ are mutually related by a Legendre transformation. Functions related in this way contain the same information but it is expressed through different variables.
Furthermore, functions $\log Z(\lambda _1 , \dots, \lambda _m)$ and $S(F_1 , \dots, F_m)$ give, in a simple way, the variances and covariances of the functions $f_k(x)$, $k = 1,2, \dots, m$. Using (\ref{eq_maxent_p_i_Z_lk}), (\ref{eq_maxent_part_fun_l_k}) and (\ref{eq_F_k_expct_f_k_dlogZ_dl_k}) one obtains
\begin{eqnarray}
{\partial^2 \log Z(\lambda _1 , \dots , \lambda _m) \over \partial \lambda_l \partial \lambda_k } & = & -{\partial F_k \over \partial \lambda_l } = -{\partial F_l \over \partial \lambda_k } \nonumber\\
& = & \langle f_k (x) f_l(x) \rangle - \langle f_k (x)\rangle \langle f_l(x) \rangle \nonumber\\
& = & - A_{kl} , \quad k,l =1,2, \dots, m , \label{eq_dlogZ_dL_ldL_k_eq_A_kl}
\end{eqnarray}
where $A$ is a symmetric matrix, $A_{kl} = A_{lk}$. In a similar way, using (\ref{eq_L_eq_dS_F_dF}) one obtains
\begin{eqnarray}
{\partial^2 S(F_1 , \dots, F_m)\over \partial F_l \partial F_k} & = & {\partial \lambda_k \over \partial F_l} = {\partial \lambda_l \over \partial F_k} \nonumber\\
& = & B_{kl}\ , \qquad k,l =1,2, \dots, m \ , \label{eq_dS_F_dF_k_dF_l_eq_B_lk}
\end{eqnarray}
where $B$ is also a symmetric matrix, $B_{kl} = B_{lk}$. Then, from (\ref{eq_dlogZ_dL_ldL_k_eq_A_kl}), (\ref{eq_dS_F_dF_k_dF_l_eq_B_lk}) and the chain rule for derivatives, it follows that
\begin{equation}
{\partial \lambda _j \over \partial \lambda _l} = \sum_{k=1}^m {\partial \lambda _j \over \partial F_k}{\partial F_k \over \partial \lambda _l } = B_{jk} A_{kl} = \delta _{jl}\ , \quad j,l =1,2, \dots, m \ ,
\end{equation}
and similarly,
\begin{equation}
{\partial F _j \over \partial F _l} = \sum_{k=1}^m {\partial F _j \over \partial \lambda _k}{\partial \lambda _k \over \partial F_l } = A_{jk} B_{kl} = \delta _{jl}\ , \quad j,l =1,2, \dots, m \ .
\end{equation}
Therefore, the matrices given by (\ref{eq_dlogZ_dL_ldL_k_eq_A_kl}) and (\ref{eq_dS_F_dF_k_dF_l_eq_B_lk}) are inverses, $A^{-1} = B$.
Elements of the matrix $A$ are the second partial derivatives of the functions $\log Z(\lambda _1 , \dots , \lambda _m)$ and represent the measure of the expected dispersion and mutual correlation of the functions $f_k(x)$, $k = 1,2, \dots, m$. Diagonal elements of the matrix $A$ give as the notion about the deviation of the variables $f_k(x)$ from their expectation values $\langle f_k (x) \rangle$. Furthermore, from (\ref{eq_maxent_p_i_Z_lk}), (\ref{eq_maxent_part_fun_l_k}) and (\ref{eq_F_k_expct_f_k_dlogZ_dl_k}), it follows that the covariance of some other function $g(x)$ with the function $f_k(x)$ is obtained as
\begin{equation}
- {\partial \langle g(x) \rangle \over \partial \lambda _k} = \langle g(x)f_k(x) \rangle - \langle g(x)\rangle \langle f_k(x) \rangle , \ \ \ k = 1,2, \dots, m .
\end{equation}
\section{Interpretation of MaxEnt formalism in statistical mechanics} \label{sec_3}
It clear that the MaxEnt probability distribution (\ref{eq_maxent_p_i_Z_lk}) has the same form as Gibbs ensemble probability distributions from equilibrium statistical mechanics. This is not surprising since the rationale of the Gibbs method of constructing ensembles was to assign that probability distribution which, while agreeing with what is known (i.e. the data given by constraints), gives the least value of the average index (logarithm) of probability of phase i.e. $\sum_{i=1}^n p_i\log p_i$ \cite{jaynes2,gibbs}. This procedure has lead Gibbs to the canonical ensemble for closed systems in thermal equilibrium with the environment, the grand canonical ensemble for open systems, and an ensemble for a system rotating at a fixed angular velocity. However, MaxEnt formalism represents a general method of statistical inference which is applicable, in principle, to all problems where only incomplete and partial information about the problem is available. Equations from the last section represent the generic form of the MaxEnt formalism. To give them a physical interpretation they should be put in the context of some specific physical situation. Since the Lagrange multipliers $\lambda = (\lambda _1 , \dots, \lambda _m)$, under certain conditions, are single-valued functions of the expected values $F = (F_1 , \dots, F_m)$, and at the same time the only parameters in the MaxEnt probability distribution, physical interpretation of these quantities is of special pertinence in that sense.
It will now be shown that the physical interpretation of the Lagrange multipliers follows from the relation describing the changes of the expected values. Values of the functions $f_k(x_i)$, $k = 1,2, \dots, m$, associated with the values $x_i$ of the variable $x$, $i = 1,2, \dots, n$, can represent the eigenvalues of some specific physical quantities, for example energy eigenvalues $E_i$, or eigenvalues of the quantities from the set of compatible quantities. Let us assume that the small change in the expectation values $\langle f_k(x) \rangle$ is done by the small change of the functions $f_k(x_i)$ and the probabilities $p_i$,
\begin{equation}
\delta \langle f_k(x) \rangle = \sum_{i=1}^{n} p_i \delta f_k(x_i) + \sum_{i=1}^{n} f_k(x_i)\delta p_i \ , \ \ k = 1,2, \dots, m \ . \label{eq_d_expt_f_k_eq_var}
\end{equation}
Here, $\delta \langle f_k(x) \rangle$ is the change of the expectation value $\langle f_k(x) \rangle$ and $\langle \delta f_k(x)\rangle = \sum_i p_i \delta f_k(x_i) $ is the expectation value of the change of $f_k(x) $. Their difference depends on the changes in the probabilities $\delta p_i$,
\begin{equation}
\delta \langle f_k(x) \rangle - \langle \delta f_k(x)\rangle = \sum_{i=1}^{n} f_k(x_i)\delta p_i \ , \qquad k = 1,2, \dots, m \ . \label{eq_df_k_expt_expt_df_k_ind_term}
\end{equation}
The change of information entropy $S_I$ is equal to
\begin{equation}
\delta S_I = - \sum_{i=1}^{n} \delta p_i \log p_i \ . \label{eq_dS_I_eq_var}
\end{equation}
Introducing the MaxEnt probabilities (\ref{eq_maxent_p_i_Z_lk}) for $\{p_i\}$ in (\ref{eq_dS_I_eq_var}) and using (\ref{eq_cond_norm_nonneg}) and (\ref{eq_df_k_expt_expt_df_k_ind_term}), one obtains
\begin{eqnarray}
\delta S & = & \sum_{k = 1} ^m \sum_{i=1}^{n} \lambda _k f_k(x_i)\delta p_i \nonumber\\
& = & \sum_{k = 1} ^m \lambda _k \left (\delta \langle f_k(x) \rangle - \langle \delta f_k(x)\rangle\right ) \ . \label{eq_dS_max_eq_dp_i_f_k}
\end{eqnarray}
Assuming that $\{p_i + \delta p_i\}$ is also a MaxEnt probability distribution, equation (\ref{eq_dS_max_eq_dp_i_f_k}) then gives the change of the maximum of information entropy due to the change in expected values (i.e. the constraints). The meaning of (\ref{eq_dS_max_eq_dp_i_f_k}) is simple to understand, if we introduce
\begin{eqnarray}
&& \delta Q_k = \sum_{i=1}^{n} f_k(x_i)\delta p_i = \delta \langle f_k(x) \rangle - \langle \delta f_k(x)\rangle \ , \cr\nonumber\\
&& k = 1,2, \dots, m \ , \label{eq_Q_k_d_expt_f_k_expt_d_f_k}
\end{eqnarray}
and then using this write $\delta S$ in the form
\begin{equation}
\delta S = \sum_{k = 1} ^m \lambda _k \delta Q_k \ . \label{eq_dS_sum_k_L_k_Q_k}
\end{equation}
Equation (\ref{eq_Q_k_d_expt_f_k_expt_d_f_k}) suggests the interpretation that was given by Jaynes \cite{jaynes6} and Grandy \cite{grandy,grandy1}. The expectation value $\langle \delta f_k(x)\rangle $ of the change $\delta f_k(x)$ is the corresponding generalized work. The remaining part of the change $\delta \langle f_k(x) \rangle$ of the expectation value $\langle f_k(x) \rangle$ comes from the change in the probability distribution $\{p_i\}$ and represents the generalized heat $\delta Q_k$ for the quantity $f_k(x)$. If the function $f_k(x)$ is such that $f_k(x_i) = E_i$ for all $i$, then $\delta Q_k$ is the heat in the usual sense. Grandy \cite{grandy,grandy1} interpreted the equations (\ref{eq_Q_k_d_expt_f_k_expt_d_f_k}) as the general rule in the probability theory, whose special case is the first law of thermodynamics. Indeed, for a macroscopic system, if $f_k(x_i) = E_i$ for all $i$, then the corresponding equation (\ref{eq_Q_k_d_expt_f_k_expt_d_f_k}) has the form of the first law of thermodynamics
\begin{equation}
\delta Q = \delta \langle E \rangle - \langle \delta E \rangle = \delta U - \delta W \ , \label{eq_F_L_T_dQ_eq_dU_dW}
\end{equation}
where $\langle E \rangle = U$ is the internal energy of the system and $\delta W = \langle \delta E \rangle$ is the work done on the system. According to \cite{grandy,grandy1}, the heat $\delta Q$ is the energy transfered through the degrees of freedom over which we don't have control, while the work $\delta W$ is the energy transfered through the degrees of freedom which we do control. In such an interpretation, the generalized $\delta Q_k$ is the part of the change of the corresponding expectation value $\delta \langle f_k(x) \rangle$ related to the change in the probability distribution by equation (\ref{eq_Q_k_d_expt_f_k_expt_d_f_k}). Equations (\ref{eq_Q_k_d_expt_f_k_expt_d_f_k}) and (\ref{eq_dS_sum_k_L_k_Q_k}) explictily show that the change in the maximum of information entropy comes from the change in the probability distribution related to $\delta Q_k$. Furthermore, Grandy has brought generalized terms $\delta Q_k$ into connection with the change of the macroscopic constraints brought by means of the external influences on the system. Based on that, Grandy \cite{grandy,grandy1,grandy2,grandy3} has developed a generalized approach which, along with the generalization of the Liouville--von Neumann equation for the density matrix through the application of the MaxEnt formalism, leads to the derivation of the macroscopic equations of motion.
Let us consider now the quasistatic change of the energy of macroscopic system, for which we specify only that it is a closed system (i.e. the system that can exchange energy, in the form of work or heat, with the environment, but not particles). From equations (\ref{eq_dS_sum_k_L_k_Q_k}) and (\ref{eq_F_L_T_dQ_eq_dU_dW}) then it follows that
\begin{equation}
\delta S = \lambda \delta Q \ ,\label{eq_dS_eq_L_dQ_q_s_e}
\end{equation}
and
\begin{equation}
\delta U - \delta W = \delta Q = \frac{1}{\lambda }\delta S \ . \label{eq_FLT_dU_dW_eq_dQ_L_dS}
\end{equation}
If we write the first law of thermodynamics in the form in which the thermodynamic entropy $S_e$ explicitly appears,
\begin{equation}
dU - \delta W = \delta Q = TdS_e \ ,
\end{equation}
then the Lagrange multiplier $\lambda $ in the analogous equation (\ref{eq_FLT_dU_dW_eq_dQ_L_dS}) can be identified as
\begin{equation}
\lambda = \frac{1}{kT} \ . \label{eq_ident_L_eq_kT-1}
\end{equation}
The change $\delta S$ in the maximum of information entropy given by (\ref{eq_dS_eq_L_dQ_q_s_e}) is thus related to the total differential of thermodynamic entropy $dS_e$ by
\begin{equation}
k d S = d S_e = {\delta Q \over T} \ , \label{eq_kdS_eq_dS_e_eq_Q_T-1}
\end{equation}
where $T$ is the temperature, and $1/T$ is the integrating factor for heat $\delta Q$. The choice of the unit for temperature (Kelvin), and respectively for entropy (Joule Kelvin$^{-1}$) is reflected in the appearance of the Boltzmann constant $k$ in the previous expressions.
The confirmation that the identification given by (\ref{eq_ident_L_eq_kT-1}) is correct comes by introducing the value of the Lagrange multiplier $\lambda = (kT)^{-1}$ in the MaxEnt probability distribution corresponding to the case considered here. In this way we obtain
\begin{equation}
p_i =\frac{1}{Z} \exp \left (- \frac{E_i}{kT}\right ) \ , \label{eq_Gibbs_kan_E_i}
\end{equation}
which is known in statistical mechanics as the Gibbs canonical distribution, describing the closed system of known temperature in equilibrium with the environment. The normalization factor of the canonical distribution, the partition function $Z$, is equal to
\begin{equation}
Z = \sum_{i=1}^{n} \exp \left (- \lambda E_i\right ) = \sum_{i=1}^{n} \exp \left (- \frac{E_i}{kT}\right ) \ . \label{eq_part_fun_Z_L_Z_kT_-1}
\end{equation}
By considering the open system (i.e. the system that can exchange energy and particles with the environment) in analogous way it is shown that the MaxEnt probability distribution, in the case when along with the expected value of energy, the expected value of the number of particles is known, corresponds to the Gibbs grand canonical distribution \cite{jaynes6,grandy}. Furthermore, it is important that the generic MaxEnt relations from the previous and this section become, in the special cases considered here, the well known equations of equilibrium statistical mechanics.
However, recent work \cite{hanggi} on the Crooks fluctuation theorem \cite{crooks} and Jarzynski equality \cite{jarzynski} indicates further insights. When these important relations of nonequilibrium statistical mechanics are extended to quantum systems strongly coupled with their environments, the thermodynamic entropy of the system of interest in such cases is related to the maximum of the information entropy of the total system (including the system of interest and its environment) minus the information entropy of the environment:
\begin{equation}
S_{e \ \mathrm{of\ the\ system}} = k\left (S_{I\ \mathrm{of\ the\ total\ system}}- S_{I\ \mathrm{of\ the \ environment}}\right )_{\mathrm{max}},
\end{equation}
where $k$ is the Boltzmann constant. The reason for this is that, unlike in the cases considered in this paper, for systems strongly interacting with their environments the correlation between the system and the environment degrees of freedom can not be neglected.
\section{MaxEnt and the interpretation of probabilities} \label{sec_4}
In this section we modify and extend the analysis given by Jaynes in \cite{jaynes8} and show how this leads to the independent interpretation of probabilities which is based on the maximum information entropy principle. Let's consider a proposition $A(n_1, \dots, n_m)$ which is a function of the sample numbers $n_i$, $i = 1, 2, \dots, m$. In the context of statistical mechanics, the sample numbers can represent, for example, the distribution $\{n_1, \dots, n_m\}$ of the number of systems from the ensemble of $n = \sum_{i=1}^m n_i$ identical systems found in $m$ different microscopic states comprising the discrete sample space. The proposition $A(n_1, \dots, n_m)$ can represent, for example, the expected value of energy of the individual system, or the expected values of some set of compatible quantities. Relative frequencies are then given by $f_i = n_i/n$, $i = 1, 2, \dots, m$. The number of outcomes for which the proposition $A$ is true is given by the sum over different distributions of sample numbers $\{n_1, \dots, n_m\}$,
\begin{equation}
M(n, A) = \sum_{\{n_i\} \in R} W(n_1, \dots, n_m) \ , \label{eq_M_sum_W}
\end{equation}
where $R$ is the region of the sample space for which the proposition $A$ is true and $W$ is the multinomial coefficient
\begin{equation}
W(n_1, \dots, n_m) = \frac{n!}{n_1 ! \cdots n_m !} \ .
\end{equation}
The greatest term (multiplicity) in the sum (\ref{eq_M_sum_W}) over the region $R$ is
\begin{equation}
W_{\mathrm{max}} = \mathrm{Max}_{R}W(n_1, \dots, n_m) \ .
\end{equation}
If $T(n, m)$ is the number of terms in the sum (\ref{eq_M_sum_W}), then it is true that
\begin{equation}
W_{\mathrm{max}} \leq M(n, A) \leq W_{\mathrm{max}}T(n, m) \ ,
\end{equation}
and
\begin{eqnarray}
&& \frac{1}{n} \log W_{\mathrm{max}} \cr\nonumber\\
&& \leq \frac{1}{n} \log M(n, A) \cr\nonumber\\
&& \leq \frac{1}{n} \log W_{\mathrm{max}} + \frac{1}{n} \log T(n, m) \ . \label{eq_logM_ineq}
\end{eqnarray}
From combinatorial arguments it follows that
\begin{equation}
T(n, m) = {n + m - 1 \choose n} = \frac{(n + m - 1)! }{n! (m - 1)!} \ .
\end{equation}
Then as $n \rightarrow \infty $
\begin{equation}
T(n, m) \sim \frac{n^{m - 1}}{(m - 1)!} \ .
\end{equation}
Therefore, as $n \rightarrow \infty $, $\log T(n, m)$ grows less rapidly than $n$,
\begin{equation}
\frac{1}{n} \log T(n, m) \rightarrow 0 \ , \label{eq_lim_1/n_logT}
\end{equation}
and from (\ref{eq_logM_ineq}) and (\ref{eq_lim_1/n_logT}) it follows that
\begin{equation}
\frac{1}{n} \log M(n, A) \rightarrow \frac{1}{n} \log W_{\mathrm{max}} \ , \label{eq_log_M_Wmax}
\end{equation}
as $n \rightarrow \infty $. The multinomial coefficient $W$ grows so rapidly with $n$ that the maximum term $W_{\mathrm{max}}$ dominates, in the sense given by (\ref {eq_log_M_Wmax}), the total multiplicity $M(n, A)$ given by the sum (\ref{eq_M_sum_W}).
However, the limit we really want is the one in which the sample frequencies $n_i/n$ tend to certain (but not yet specified) constant values $f_i$ as $n \rightarrow \infty $. Therefore, we want the limit of
\begin{equation}
\frac{1}{n}\log W = \frac{1}{n}\log \left [\frac{n!}{(nf_1) ! \cdots (nf_m) !} \right ] \ ,
\end{equation}
as $n \rightarrow \infty $. Using the Stirling asymptotic approximation
\begin{equation}
\log{n!} \sim n \log n - n + \log \sqrt{2\pi n} + O\left (\frac{1}{n} \right ) \ ,
\end{equation}
we find as $n \rightarrow \infty $ that in this limit we have
\begin{equation}
\frac{1}{n} \log W \rightarrow H \equiv - \sum_{i=1}^{m} f_i \log f_i \ , \label{eq_lim_logW_H}
\end{equation}
and this gives the information entropy of the relative frequency distribution $\{f_1, \dots, f_m \}$. So, from (\ref{eq_log_M_Wmax}) and (\ref{eq_lim_logW_H}), it follows that in a such limit we also have that
\begin{equation}
\frac{1}{n} \log M(n, A) \rightarrow \frac{1}{n} \log W_{\mathrm{max}} = H_{\mathrm{max}} \ . \label{eq_log_M_Wmax_H}
\end{equation}
Therefore, for very large $n$, the maximum multiplicity $W_{\mathrm{max}}$ is the one that dominates the total multiplicity $M(n, A)$ and maximizes the information entropy $H$ subject to the constraints that define the region of the sample space for which the proposition $A$ is true. Furthermore, it is straightforward to show that the probability of obtaining the relative frequency distribution $\{f_1, \dots, f_m \}$ which corresponds to the maximum multiplicity $W_{\mathrm{max}}$ approaches $1$ in the limit of large $n$, because from (\ref{eq_log_M_Wmax}) in this limit we have
\begin{equation}
{W_{\mathrm{max}} \over M(n,A)} \sim 1 \ .
\end{equation}
Therefore, in the limit of large $n$, without any other additional constraints except that the proposition $A$ is true, we can assume with certainty that the relative frequencies
to be used are the ones that maximize the multiplicity $W$ and, because of (\ref{eq_lim_logW_H}) and (\ref{eq_log_M_Wmax_H}), maximize the information entropy $H$. Therefore, according to the weak law of large numbers, the relative frequencies in the limit of a large number of trials ($n \rightarrow \infty $) should correspond to the MaxEnt probabilities.
So, in this context, can we now examine the frequency interpretation of probabilities as factual properties of the real world, if, as in this example, the corresponding probabilites (obtained in the limit of a large number of trials) actually follow from the principle of maximum information entropy, and therefore, depending only on the available information (i.e. on the proposition $A$), depend on our state of knowledge? This question comes naturally as the above result implies that under constraints representing the information that is available, the relative frequencies in the limit of a large number of trials tend with certainty to the corresponding MaxEnt probabilities.
\section{Conclusion} \label{sec_5}
We have shown how the probabilities in statistical mechanics can not be simply interpreted in the frequentist context. Probabilities, at least in the Gibbs formalism of statistical mechanics, are not simply relative frequencies in the ensemble of a large number of identical systems. Actually they depend on the available information about the individual system and therefore are the description of a degree our knowledge about it. The ensembles of identically prepared systems are chosen in the Gibbs formalism only to illustrate that the information we have about the individual system is incomplete, which means that it isn't sufficiently detailed to specify the exact microscopic state of a macroscopic system, nor its exact evolution in time.
Furthermore, in the case of nonequilibrium systems and processes that are irreversible on the macroscopic level, justification of nonequilibrium ensembles in the frequentist sense as a physical fact, using only first principles, via equations of motion and ergodic theorems, becomes permeated with technical and, more importantly, conceptual difficulties \cite{jaynes9}. For example, the applications of ergodic theorems for that purpose would require an infinite or large time intervals, and this is not in general always available for nonequilibrium systems that are continuously evolving and changing its macroscopic state with time. This is well exemplified in the work of Zubarev and his coworkers, who introduce a hierarchy of time scales, with different sets of quantities that are relevant for the description of a nonequilibrium system on different time scales \cite{zubarev1,zubarev2}. More important than ergodicity is the concept of a mixing system, originally introduced by Gibbs \cite{gibbs}. Mixing implies ergodicity, and hopefully can provide a mechanical foundation of both nonequilibrium and equilibrium statistical mechanics, if we can prove it for realistic systems \cite{dorfman}. However, there are differing opinions about its importance since transport coefficients and dissipativity, an essential property of macroscopic systems, can not be derived only from mixing \cite{balescu}. On the other hand, as we have shown here, MaxEnt formalism is an independent logical extension of the Gibbs method, and leads to statistical distributions which depend only on the available information. If that information is relevant for the description of a system at a macroscopic level, then accordingly, the obtained statistical distributions should be relevant for describing its macroscopic state, its properties and time evolution \cite{jaynes2,jaynes4,jaynes6,jaynes7,grandy,grandy1,grandy2,grandy3,zubarev1,zubarev2,kuic,kuic1,kuic2,kuic3}.
|
\subsubsection*{Preliminaries}
Let us assume a routing change, e.g., an announcement of a new prefix by an AS, in the network at time $t_{0}=0$. Our goal is to calculate the \textit{BGP convergence time}, i.e., the time needed till all ASes/routers in the network have the final (i.e., shortest, conforming to policies) BGP routes for this prefix.
To this end, using Assumption~\ref{assumption:exponential-lambda}, we can model the dissemination of the BGP updates in the network with the Markov Chain (MC) of Fig.~\ref{fig:mc-nodes}, where each state corresponds to the number of ASes/routers that have the final BGP routes. At time $t_{0}=0$ the system is at state $0$, while the state $N$ denotes the BGP convergence. When an AS in the SDN cluster receives the BGP update, all the nodes in the SDN cluster are informed (through the controller); thus, we have a transition, e.g., from state $i$ to state $k+i$. The transition rates in the MC, as we discuss in detail later, depend on the network topology.
The Markov Chain of Fig.~\ref{fig:mc-nodes} is transient, and the BGP convergence time is the time needed to move from state $0$ to state $N$.
For notation brevity, in the remainder we use the MC of Fig.~\ref{fig:mc-steps}, which is equivalent to the MC of Fig.~\ref{fig:mc-nodes}. Here, the states represent the \textit{number of transitions} in the MC of Fig.~\ref{fig:mc-nodes}. For example, the state/step $1$ corresponds to the state $1$ or $k$ of the MC of Fig.~\ref{fig:mc-nodes}, while the state/step $i$ corresponds to the state $i$ or $k+i-1$ in the MC of Fig.~\ref{fig:mc-nodes}. The states $0$ are equivalent in both MCs, while the state/step $C$ denotes the BGP convergence, and, thus, corresponds to the state $N$ in the MC of Fig.~\ref{fig:mc-nodes}.
If we denote with $x$ the step at which -for the first time- an AS in the SDN cluster receives the BGP update, then the transitions rates $\lambda_{i}^{'}$ in the MC of Fig.~\ref{fig:mc-steps} are given by
\begin{equation}
\lambda_{i}^{'} = \left\{
\begin{tabular}{ll}
$\lambda_{i,i+1}+\lambda_{i,i+k}$ &$, i\leq x$\\
$\lambda_{k+i-1, k+i}$ &$, i>x$
\end{tabular}
\right.
\end{equation}
\begin{figure}
\subfigure[Markov Chain (number of nodes)]{\includegraphics[width=\linewidth]{./figures/MC_nb_of_nodes1.eps}\label{fig:mc-nodes}}
\subfigure[Markov Chain (number of transitions, or \textit{steps})]{\includegraphics[width=\linewidth]{./figures/MC_steps.eps}\label{fig:mc-steps}}
\caption{Markov Chains where the states correspond to (a) the number of nodes that have updated BGP routes, and (b) the number of transitions, or \textit{steps}, of the BGP update dissemination process.}
\label{fig:markov-chains}
\end{figure}
We now proceed to calculate the rates $\lambda_{i}^{'}$.
The ASes that have received the BGP updates, will then send BGP updates to some of their neighboring ASes, according to their routing policies. We refer to a neighbor to which the update will be forwarded as a \textit{bgp-eligible} neighbor.
\begin{definition}\label{def:bgp-degree}
We define as the \texttt{bgp-degree} at step $i$, $D(i)$, the number of the ASes that are bgp-eligible neighbors with \underline{any} of the ASes that have received the BGP updates at step $i$.
\end{definition}
Although an AS might receive the same BGP update from more than one neighbors, the final BGP route will correspond to only one of the received updates (i.e., shortest path). Hence, to calculate the transition rate $\lambda_{i}^{'}$, we take into account only one BGP connection (corresponding to the shortest path) per bgp-eligible neighbor. Since the BGP update times are exponentially distributed with rate $\lambda$ (see, Assumption~\ref{assumption:exponential-lambda}), it follows that $\lambda_{i}^{'}$ will be given by\footnote{The transition time is the minimum of $D(i)$ i.i.d. exponentially distributed times.}
\begin{equation}\label{eq:lambda-prime}
\lambda_{i}^{'} = \lambda\cdot D(i)
\end{equation}
Knowing the rates $\lambda_{i}^{'}$, we can calculate the transition delays in each step. Adding the delays in each step, we can derive Theorem~\ref{thm:expected-delay-generic}, which gives the BGP convergence time, i.e., the time to move from state $0$ to state $C$.
\begin{theorem}\label{thm:expected-delay-generic}
The expectation of the BGP convergence time $T$ in a hybrid SDN/BGP inter-domain topology is given by
\begin{equation}\label{eq:theorem-expected-delay-generic}
E[T] = \frac{1}{\lambda}\cdot \sum_{x=0}^{N-k}\sum_{i=1}^{N-k}\frac{1}{D(i|x)}\cdot P_{sdn}(x)
\end{equation}
where $D(i|x)$ is the \texttt{bgp-degree } of the network at step $i$ given that the SDN cluster receives the update at step $x$, and $P_{sdn}(x)$ is the probability that the SDN cluster receives the update at step $x$.
\end{theorem}
\begin{proof}
To calculate $E[T]$, we first apply the conditional expectation
\begin{equation}\label{eq:ET-sum-ETx}
E[T] = \sum_{x=0}^{N-k}E[T|x]\cdot P_{sdn}(x)
\end{equation}
where $E[T|x]$ denotes the expected convergence time, given that the SDN cluster receives the update at step $x$. Under this condition, the bgp-degrees at each step are $D(i|x)$, $i\in[1,N-k]$. Hence, taking also into account \eq{eq:lambda-prime}, it follows that the transition delay from a step $i$ to a step $i+1$, $T_{i,i+1}$, is exponentially distributed with rate $\lambda_{i}^{'} = \lambda\cdot D(i|x)$, and its expectation is given by
\begin{equation}
E[T_{i,i+1}|x] = \frac{1}{\lambda\cdot D(i|x)}
\end{equation}
\underline{Remark:} The state/step $0$ does not correspond to a real state of the system; it is only used for presentation purposes. Thus, we set $T_{0,1}=0$.
As mentioned earlier, the BGP convergence delay is the time needed to move from step $0$ to step $C$, and thus it is given by the sum of the transition delays of all the intermediate steps, i.e.,
\begin{align}\label{eq:ETx-sum-Dix}
E[T|x] &= E\left[\sum_{i=1}^{N-k}T_{i,i+1}|x\right] \nonumber\\
&= \sum_{i=1}^{N-k}E[T_{i,i+1}|x] \nonumber\\
&= \sum_{i=1}^{N-k}\frac{1}{\lambda\cdot D(i|x)}
\end{align}
Now, the expression of \eq{eq:theorem-expected-delay-generic} follows by substituting \eq{eq:ETx-sum-Dix} to \eq{eq:ET-sum-ETx}.
\end{proof}
In the following sections we calculate the quantities ${D(i|x)}$ and $P_{sdn}(x)$ for important network topologies.
\subsection{Full-Mesh Network Topology}
We first consider a basic topology: a full-mesh network, where every AS-pair is connected.
\begin{theorem}\label{thm:P-sdn}
The probability that the SDN cluster receives the update at step $x$ is given by
\begin{equation}\label{eq:P-sdn}
P_{sdn}(x) = \frac{k}{N-x}\cdot \prod_{j=0}^{x-1}\left(1-\frac{k}{N-j}\right)
\end{equation}
\end{theorem}
\begin{proof}
The SDN cluster comprises $k$ (out of the total $N$) ASes. Since we consider that the prefix announcement is made by a (random) AS in the network, the probability that the announcing AS is in the SDN cluster (and thus $x=0$) is
\begin{equation}
P_{sdn}(0) = \frac{k}{N}
\end{equation}
If the announcing AS is not in the SDN cluster, then $x>0$, and thus
\begin{equation}\label{eq:Psdn-x-larger-0}
P_{sdn}(x>0) = 1-P_{sdn}(0) = 1-\frac{k}{N}
\end{equation}
The probability $P_{sdn}(1)$ is given by
\begin{equation}
P_{sdn}(1) = P_{sdn}(1|x>0)\cdot P_{sdn}(x>0)
\end{equation}
where $P_{sdn}(1|x>0)$ denotes the probability that the SDN cluster receives the BGP update at step $1$, given that it has not received it before. If $x>0$, then at step $1$ the remaining ASes without the update are $N-1$, of which $k$ belong to the SDN cluster. Since the BGP update processes are distributed with the same rate $\lambda$, the probability that the next AS to get the update belongs to the SDN cluster is $\frac{k}{N-1}$. Therefore, and taking into account \eq{eq:Psdn-x-larger-0}, it holds that
\begin{equation}
P_{sdn}(1) = P_{sdn}(1|x>0)\cdot P_{sdn}(x>0) = \frac{k}{N-1}\cdot \left(1-\frac{k}{N}\right)
\end{equation}
and, respectively,
\begin{align}
P_{sdn}(x>1) &= \left(1-P_{sdn}(1|x>0)\right)\cdot P_{sdn}(x>0) \nonumber\\
&= \left(1-\frac{k}{N-1}\right)\cdot \left(1-\frac{k}{N}\right)
\end{align}
Proceeding similarly for the next steps $i=2,...,N-k$, it can be shown that
\begin{equation}
P_{sdn}(i|x>i-1) = \frac{k}{N-i}
\end{equation}
and
\begin{align}
P_{sdn}(x>i-1) &= \left(1-P_{sdn}(i-1|x>i-2)\right)\cdot ... \cdot P_{sdn}(x>0) \nonumber\\
& = \left(1-\frac{k}{N-(i-1)}\right)\cdot...\cdot \left(1-\frac{k}{N}\right) \nonumber\\
& = \prod_{j=0}^{i-1}\left(1-\frac{k}{N-j}\right)
\end{align}
and, therefore,
\begin{align}
P_{sdn}(i) &= P_{sdn}(i|x>i-1)\cdot P_{sdn}(x>i-1)\nonumber\\
&= \frac{k}{N-i}\cdot \prod_{j=0}^{i-1}\left(1-\frac{k}{N-j}\right)
\end{align}
which is the expression of \eq{eq:P-sdn}.
\end{proof}
Theorem~\ref{thm:Dix-full-mesh} gives the bgp-degrees $D(i|x)$ in a mesh network as a function of $n(i|x)$, which is defined as the number of nodes with updated BGP information at step $i$, given that the SDN cluster received the update at step $x$
\begin{equation}
n(i|x) = \left\{
\begin{tabular}{ll}
$i$ & $, i\leq x$ \\
$i+k-1$ & $, i>x$
\end{tabular}
\right.
\end{equation}
\begin{theorem}\label{thm:Dix-full-mesh}
The \texttt{bgp-degree } $D(i|x)$, $i\in[1,N-k], x\in[0,N-k]$, in a full-mesh network topology is given by
\begin{equation}
D(i|x) = N-n(i|x)
\end{equation}
\end{theorem}
\begin{proof}
In a mesh network, since every AS-pair is directly connected, only the BGP messages sent by the announcing AS (i.e., shortest path) need to be considered. In step $i$, the announcing AS has $N-n(i|x)$ neighbors that have not received the BGP updates, and thus, it follows that $D(i|x) = N-n(i|x)$.
\end{proof}
\subsection{Random Graph Network Topologies}
In networks that are not full-meshes, ASes can be connected in many different ways and policies. Since it is not possible to study every single topology, we use two classes of random graphs to capture the effects of routing centralization in non full-mesh networks. In this first approach, we consider unconstrained routing policies
\subsubsection{Poisson (Erdos-Renyi) Graph}
We first consider the case of a Poisson random graph, where a link between each AS-pair exists with probability $p$. Varying the value of $p$ we can capture different levels of sparseness.
Using similar arguments as in the full-mesh case, it is easy to show that the probabilities $P_{sdn}(x)$ are given by Theorem~\ref{thm:P-sdn}. The \textit{expected} bgp-degrees, which can be used (as an approximation) instead of $D(i|x)$ in Theorem~\ref{thm:expected-delay-generic}, are given by the following Theorem.
\begin{theorem}\label{thm:Dix-poisson}
The expectation of the \texttt{bgp-degree } $D(i|x)$, $i\in[1,N-k], x\in[0,N-k]$, in a Poisson graph network topology is
\begin{equation}
E[D(i|x)] = \left(N-n(i|x)\right) \cdot \left(1-(1-p)^{n(i|x)}\right)
\end{equation}
\end{theorem}
\begin{proof}
In a non full-mesh network, some ASes are not directly connected to the announcing AS. Thus, in the calculation of the $D(i|x)$ we need to consider the bgp-eligible neighbors of \textit{all} the ASes that have received the update.
Let assume that we are at step $i$, and $n(i)$ nodes have received the BGP updates; we denote the set of these nodes as $S_{i}$. A node $m\notin S_{i}$ is a bgp-eligible neighbor with a node $j\in S_{i}$ with probability
\begin{equation}
P(m,j) = p
\end{equation}
since every pair of nodes is connected with probability $p$ (by the definition of a Poisson graph). The probability that $m$ is \textit{not} a bgp-eligible neighbor with \textit{any} of the nodes $j\in S_{i}$, is given by
\begin{align}
1-P(m,S_{i}) = \prod_{j\in S_{i}}(1-P(m,j)) = \prod_{j\in S_{i}}(1-p) = (1-p)^{n(i)}
\end{align}
since $|S_{i}| = n(i)$. It follows easily that the complementary event, i.e., $m$ is a bgp-eligible neighbor with \textit{any} of the nodes $j\in S_{i}$, happens with probability
\begin{align}
P(m,S_{i}) = 1- (1-p)^{n(i)}
\end{align}
There are $N-n(i)$ ASes without the update, with each of them being a bgp-eligible neighbor with any of the nodes $j\in S_{i}$ with (equal) probability $P(m,S_{i})$. Hence, the total number of bgp-eligible neighbors (or, as defined in Def.~\ref{def:bgp-degree}, the \textit{bgp-degree} $D(i)$) is a binomially distributed random variable, whose expectation is given by
\begin{equation}
E[D(i)] = (N-n(i))\cdot (1-(1-p)^{n(i)})
\end{equation}
\end{proof}
\begin{corollary}
Using the expectation of $D(i|x)$ in Theorem~\ref{thm:expected-delay-generic}, underestimates the BGP convergence time, i.e.,
\begin{equation}
E[T] \geq \frac{1}{\lambda}\cdot \sum_{x=0}^{N-k}\sum_{i=1}^{N-k}\frac{1}{E[D(i|x)]}\cdot P_{sdn}(x)
\end{equation}
\end{corollary}
\begin{proof}
The bgp-degree $D(i|x)$ in non full-mesh networks, is a random variable that can take different values, depending on the BGP updates dissemination process (i.e., the exact set of nodes that have received the BGP updates at step $i$, and their links to the rest of the nodes). Thus, we can write for the transition delay
\begin{equation}\label{eq:transition-delay-generic}
E[T_{i,i+1}|x] = \sum_{y} \frac{1}{y} \cdot P\{D(i|x) = y\} = E\left[\frac{1}{D(i|x)}\right]
\end{equation}
Calculating the exact value of the expectation $E\left[\frac{1}{D(i|x)}\right]$ is difficult, thus, we use a well known approximation in \eq{eq:transition-delay-generic}:
\begin{equation}
E[T_{i,i+1}|x] = E\left[\frac{1}{D(i|x)}\right] \approx \frac{1}{E[D(i|x)]}
\end{equation}
where the calculation of $E[D(i|x)]$ is much easier (see, e.g., proof of Theorem~\ref{thm:Dix-poisson}). Then the BGP convergence delay is approximately given by
\begin{equation}
E[T] \approx \frac{1}{\lambda}\cdot \sum_{x=0}^{N-k}\sum_{i=1}^{N-k}\frac{1}{E[D(i|x)]}\cdot P_{sdn}(x)
\end{equation}
However, applying Jensen's bound for the expectation of a convex function (here, $f(x) = \frac{1}{x}$) of a random variable (here, $D(i|x)$) on \eq{eq:transition-delay-generic}, gives
\begin{equation}
E[T_{i,i+1}|x] = E\left[\frac{1}{D(i|x)}\right] \geq \frac{1}{E[D(i|x)]}
\end{equation}
which proves the Corollary.
\end{proof}
\subsubsection{Arbitrary Degree Sequence Random Graph}
The structure of networks where the degrees (i.e., the number of connections) of the ASes are largely heterogeneous, e.g., power-law graphs, can be better described with a Configuration-Model Random Graph (CM-RG) rather than a Poisson graph. In the CM-RG model, a random graph is created by connecting randomly the nodes (i.e., ASes), whose degrees are given~\cite{Newman:Networks-book}. Hence, we can use the CM-RG to model a network with \textit{any arbitrary degree sequence} with mean value $\mu_{d}$ and variance $\sigma_{d}^{2}$ (and, $CV_{d} = \frac{\sigma_{d}}{\mu_{d}}$).
If the participation of an AS in the SDN cluster is independent of its degree, then the probabilities $P_{sdn}(x)$ are given by Theorem~\ref{thm:P-sdn}, and the bgp-degrees $D(i|x)$ are given by the following Result\footnote{We use the notation ``Result'', instead of ``Theorem'', because the provided expression is an approximation.}.
\begin{result}
The expectation of the \texttt{bgp-degree } $D(i|x)$, $i\in[1,N-k], x\in[0,N-k]$, in a CM-RG network topology is given by
\begin{equation}\label{eq:Dix-CM}
E[D(i|x)] = D(1|x)\cdot \prod_{j=1}^{i-1}A(j|x) + \sum_{j=1}^{i-1}\left(\mu_{d}(j|x)-1\right)\cdot \prod_{m=j+1}^{i-1}A(m|x)
\end{equation}
where
\begin{align}\label{eq:CM-D1x}
D(1|x)&= \left\{
\begin{tabular}{ll}
$\mu_{d}$ & $, x>0$ \\
$(N-k)\cdot \mu_{d} \cdot \ln\left(\frac{N}{N-k}\right)$ & $, x=0$
\end{tabular}
\right.\\
\mu_{d}(j|x) &= \mu_{d}\cdot \prod_{m=1}^{j-1}\left(1-\frac{CV_{d}^{2}}{N-n(m|x)-1}\right) \label{eq:average-out-degree}\\
A(j|x) &= 1-\frac{\mu_{d}(j|x)}{N-n(j|x)-1}
\end{align}
\end{result}
\begin{proof}
The main difference with the Poisson case is that in the CM-RG case, it is more probable that the ASes with the higher degrees will receive the BGP updates faster. For instance, let us assume that the announcing AS, e.g., AS-1, does not belong to the SDN cluster. If we denote with $d_{1},d_{2}$, and $d_{3}$ the degrees of AS-1, AS-2 and AS-3 (where AS-2 and AS-3, have not received yet the BGP update), a property of a CM-RG says that AS-1 is directly connected with AS-2 and AS-3 with probabilities
\begin{equation}
P(1,2) = c\cdot d_{1}\cdot d_{2}~~~~and~~~P(1,3) = c\cdot d_{1}\cdot d_{3}
\end{equation}
respectively, where $c$ a normalizing constant. In other words, the AS with with the higher degree has a higher probability to be directly connected to AS-1. Consequently, ASes with higher degrees have a higher probability to get the BGP update faster.
Now, let us first derive \eq{eq:CM-D1x}. If $x>0$, the announcing AS does not belong to the SDN cluster ($n(1|x>0)=1$), and thus the bgp-degree will be equal to the degree of the announcing AS. Since the average degree of a node is $\mu_{d}$, it follows easily that expectation of the bgp-degree in this case, is given by
\begin{equation}
E[D(1|x>0)] = \mu_{d}
\end{equation}
If $x=0$, the announcing AS belongs to the SDN cluster, and, thus, all the $k$ nodes in the SDN cluster have the BGP updates. In this case, the bgp-degree is the number of all bgp-eligible neighbors of these $k$ nodes. Let us denote with $S_{1}$ the set of nodes in the SDN cluster. Since the fact that a node belongs to the SDN cluster and its degree are independent, the probability that an edge coming out from a node $m\notin S_{1}$ is connected to a node $j\in S_{1}$, is equal to $\frac{k}{N}$. Hence, a node $m\notin S_{1}$, with degree $d_{m}$, is not connected to any of the $k$ nodes in the SDN cluster with probability
\begin{equation}
1-P(m,S_{1}) = \left(1-\frac{k}{N}\right)^{d_{m}}
\end{equation}
and, respectively
\begin{equation}
P(m,S_{1}) = 1-\left(1-\frac{k}{N}\right)^{d_{m}}
\end{equation}
The above equation holds $\forall m\notin S_{1}$; the degrees $d_{m}$ can have different values for each $m$. Since, there are $N-k$ nodes that do not belong to the SDN cluster (and, do not have the BGP updated route), the expected bgp-degree is given by
\begin{equation}
E[D(1|0)] = (N-k)\cdot E\left[1-\left(1-\frac{k}{N}\right)^{d}\right]
\end{equation}
where the expectation is taken over $d$, i.e., over all the degrees $d_{m}, m\notin S_{1}$. To calculate this expectation is difficult, thus we approximate it using a Taylor series approximation, i.e,
\begin{align}
E\left[1-\left(1-\frac{k}{N}\right)^{d}\right] &= 1-E\left[\left(1-\frac{k}{N}\right)^{d}\right] \nonumber\\
&\approx 1-\left(1+E[d]\cdot \ln\left(1-\frac{k}{N}\right)\right) \nonumber\\
&= - E[d]\cdot \ln\left(1-\frac{k}{N}\right)\nonumber\\
&= E[d]\cdot \ln\left(\frac{N}{N-k}\right) \nonumber \\
&= \mu_{d}\cdot \ln\left(\frac{N}{N-k}\right)
\end{align}
which completes the derivation of \eq{eq:CM-D1x}.
To compute the bgp-degrees $D(i|x)$ of the steps $i=2,...,N-k$, we follow a methodology similar to~\cite{pavlos-conf-model}. Let $D(i-1)$ be the bgp-degree at step $i-1$ and $\mu_{d}(i-1)$ the average degree of the nodes that have not received the BGP update by step $i-1$ (i.e., the nodes $m, m\notin S_{i-1}$). The average degree $\mu_{d}(i-1)$ is not equal to $\mu_{d}$, in general; this is due to the fact that nodes with higher degrees receive faster the BGP updates and thus the remaining nodes are nodes with lower degrees (for a more detailed argumentation see~\cite{pavlos-conf-model}). Following similar arguments as in~\cite{pavlos-conf-model}, it can be shown that the average degrees $\mu_{d}(i)$ are approximately given by \eq{eq:average-out-degree}.
We calculate the bgp-degree of the step $i$, based on the quantities $D(i-1)$ and $\mu_{d}(i-1)$, as follows
\begin{equation}\label{eq:Di_Di-1}
D(i) = D(i-1) - 1 + \mu_{d}(i-1)\cdot \left(1- \frac{D(i-1)}{N-n(i-1)}\right)
\end{equation}
The term $D(i-1) - 1$ is the bgp-degree of the previous step minus $1$, which denotes the $i^{th}$ node that received the BGP update (i.e., at the transition between step $i-1$ and $i$). To this quantity, we need to add the number of nodes that are bgp-eligible neighbors of the $i^{th}$ node, but are not bgp-eligible neighbors of any of the nodes $\in S_{i-1}$. The total nodes without the updated BGP routes at step $i-1$ are $N-n(i-1)$; $D(i-1)$ of these nodes are bgp-eligible neighbors of at least one node $\in S_{i-1}$. Hence, the probability that a bgp-eligible neighbor of the $i^{th}$ node is not a bgp-eligible neighbor of any of the nodes $\in S_{i-1}$, is given by $\left(1- \frac{D(i-1)}{N-n(i-1)}\right)$. Since, the $i^{th}$ node has (on average) $\mu_{d}(i-1)$ bgp-eligible neighbors, it follows that the number we need to add to the quantity $(D(i-1) - 1)$ is
\begin{equation}
\mu_{d}(i-1) \cdot \left(1- \frac{D(i-1)}{N-n(i-1)}\right)
\end{equation}
Finally, calculating recursively \eq{eq:Di_Di-1}, after some algebraic manipulations, we derive \eq{eq:Dix-CM}.
\end{proof}
\section{Introduction}\label{sec:intro}
\input{Abstract}
\section{Model}\label{sec:model}
\input{Model}
\section{Analysis: BGP Convergence Time}\label{sec:analysis}
\input{Analysis}
\section{Validation and Discussion}\label{sec:discussion}
\input{Discussion}
\section{Case Study: BGP Convergence at the Internet Core Network}\label{sec:case-study}
\input{CaseStudy}
\bibliographystyle{ieeetr}
\subsection{Network Model}
\textbf{Network Model.} We consider a network (e.g., the Internet) composed of $N$ \textit{domains} or \textit{autonomous systems} (ASes). We represent each AS as a single node, i.e., a single BGP router (similarly to~\cite{Kotronis-Routing-Centralization-ComNets-2015}). Such an abstraction allows to hide the details of the internal structure of ASes, and focus on inter-domain routing.
We assume that $k\in[1,N]$ ASes cooperate in order to centralize their inter-domain routing: there exists a \textit{multi-domain SDN controller}, which is connected to the BGP routers of these $k$ ASes\footnote{This system abstraction can capture the main functionality of most of the previously proposed approaches.}. In the remainder, we refer to the set of the $k$ ASes, as the \textit{SDN cluster}.
\textbf{BGP Updates.} As in the Internet, ASes use BGP to exchange information and establish routing paths.
When a BGP edge router of an AS receives a BGP update, it (i) calculates the updates (if any) for its BGP routing table, (ii) sends updates to the other BGP edge routers within the same AS (e.g., with iBGP), and (iii) sends updates to the BGP routers of the neighboring ASes. The time needed for this process may vary a lot among different connections since it depends on a number of factors, like the employed technology (hardware/software), routers' configuration (e.g., MRAI timers), intra-domain network, etc.
To this end, in order to be able to analytically study the BGP updates dissemination (given the uncertainty and complexity), we model the time between the reception and forwarding of a BGP update in a probabilistic way.
\begin{assumption}\label{assumption:exponential-lambda}
The time between the reception of a BGP update in an AS/router and its forwarding to a neighbor AS/router, is exponentially distributed with rate $\lambda$.
\end{assumption}
Despite the simplicity of the above assumption, our results can capture the behavior of real/emulated networks (see Section~\ref{sec:discussion}).
\textbf{Inter-domain SDN routing.} Each AS belonging to the SDN cluster informs the SDN controller upon the reception of a BGP update. The SDN controller, which is aware of the topology of the SDN cluster (neighbors, policies, paths, etc.), calculates the changes in the routing paths and installs the updated routes in each router/AS belonging to the SDN cluster. ASes react to updates from the SDN controller, as in regular BGP updates, and, thus, forward them to their (non SDN) neighbors.
Let $T_{sdn}$ be the time needed for an AS to inform the SDN controller and the controller to install the updated routes in every AS in the SDN cluster. This time can be expected to be in the order of few seconds~\cite{Kotronis-Routing-Centralization-ComNets-2015}, and much lower than the BGP updating process (cf., the default value for MRAI timers in Cisco routers is $30sec.$), thus, for simplicity, we assume here that $T_{sdn}=0$.
|
\section{Introduction}
A great deal of astronomical evidence suggests that approximately 25\% of the energy density of the Universe today is composed of cold, non-baryonic ``dark matter'' (DM) \cite{planck,particledm}. Amongst the plethora of dark matter candidates, thermal relic particles with self-annihilation cross sections set by the electroweak scale are thought to be well motivated as they can lead naturally to the observed abundance and because the hierarchy problem suggests that we expect new physics to show up at the energy scale. Direct detection experiments aim to detect such particles via the nuclear recoils caused by their elastic scattering off nuclei \cite{dd,guo}.
In the near future, the target masses of DM detectors will be increased to the ton-scale \cite{ton}, and neutrino-nucleon scattering from solar neutrinos will become detectable \cite{billard}. This can be viewed as a positive development, potentially shedding light on new aspects of solar physics and neutrino physics beyond the standard model \cite{strigari,davissn,cerdeno}. However it is a hindrance to the potential detection of light dark matter because these solar neutrinos act as a background which produce signals very similar to dark matter candidates. In recent years, several strategies have been suggested to distinguish between neutrino and dark matter signals. These involve directional detectors \cite{philipp,ohare}, annual modulation of both the dark matter and solar neutrinos signals \cite{davis}, data combination from detectors composed of different target materials \cite{ruppin}, and detectors with improved energy resolution \cite{dent}.
In this work we present another strategy, that of a polarised helium-3 target where the cross section for neutrino scattering is a strong function of the angle between the neutrino momentum and the axis of polarisation. Assuming the detector is not close to a reactor, most low energy background neutrinos come from the Sun. We analyse the dependence of the neutrino signal distribution to the orientation of the incoming neutrinos in a polarised detector and use it to reduce this background.
\begin{figure}[b]
\centering
\includegraphics[width=0.46\textwidth]{Ermax.eps}
\caption{Maximum recoil energy of the target nucleus versus the incident
WIMP mass for helium-3 and xenon-129. For helium-3, the recoil energy range for all the possible dark matter events is higher-bounded.}
\label{fig:ermax}
\end{figure}
\section{Helium-3}
The use of helium-3 for dark matter detection is very appealing for many reasons \cite{moulin1,moulin2}: being a spin 1/2 nucleus, it is sensitive to the axial interactions with WIMPs; for massive WIMPs the maximum recoil energy depends very weakly on its mass and, therefore, is higher-bounded (see figure \ref{fig:ermax}); it has a low Compton cross section to $\gamma$-rays reducing by several orders of magnitude the natural radioactive background; it has no intrinsic X-rays; the capture process $n+ ^3\text{He}\rightarrow p+ ^3\text{H}+\text{764 keV}$ gives a clear signal for neutron rejection; it can be polarised and used as a target \cite{jlab}.
There are two leading methods to polarise a significant amount of nuclei in gaseous helium-3: spin-exchange optical pumping (SEOP) \cite{walker,bouchiat} and metastability exchange optical pumping (MEOP) \cite{batz,colegrove}. The latter method is not well suited for our purposes because it needs very low pressure ($\approx$ 1 mbar), implying a huge volume for the detector. In the first method, which can operate at typical pressures of 1 - 10 bar, alkali metal vapour is polarised by optical pumping and then transfers its electronic polarisation to the helium-3 nuclei via spin-exchange collisions. Originally, a pure rubidium vapor was used in this process. However, since potassium is much more efficient than rubidium at transferring its polarisation to helium-3 nuclei, hybrid mixtures of these elements are used, leading to degrees of polarisation up to 70\% \cite{singh}.
We note that the use of any such methods might risk contamination of the final helium-3 gas with potassium or rubidium isotopes. The use of other alkalis have been proposed, including non-radioactive elements such as sodium \cite{babcock}.
\begin{comment}In its condensed form, the simplest idea is to apply a high magnetic field at low temperature. Pure liquid helium-3 is a Fermi degenerate system of relatively low magnetic susceptibility below 0.4 K, corresponding roughly to its Fermi temperature. Therefore, the degree of polarisation is small and would not increase much at lower temperature \cite{leduc}. However, this method is useful for solid helium-3, obtained at low temperature by high compression.
For gaseous helium-3, one can use either a process called spin-exchange optical pumping (SEOP) \cite{walker,bouchiat} or metastability exchange optical pumping (MEOP) \cite{batz,colegrove}.
\end{comment}
\section{Neutrino Background}
A standard model neutrino is thought to scatter simultaneously off all nucleons in a nucleus in phase when the wavelength of the momentum transfer is larger than the radius of the target nucleus. This coherent neutrino-nucleus scattering leads to a cross section enhanced by a factor of $[N-(1-4\text{sin}^2\theta_W)Z]^2$, where $N$ and $Z$ are the number of target neutrons and protons respectively, and $\theta_W$ is the weak mixing angle \cite{freedman}. This process has never been observed in standard neutrino detectors due to the tiny cross section and very low nuclear recoil energies. The coherent nature of the interaction depends upon the vectorial $V$ part of the $V-A$ standard model coupling. The axial $A$ part also leads to a scattering upon any net spin of the nucleus, but this will not be coherently enhanced since in general there will be only one unpaired nucleon. For a target like xenon for example, the axial coupling will give rise to a much smaller spin-dependent cross-section than the coherent spin-independent cross section arising from the vector current.
The maximum recoil kinetic energy in coherent neutrino-nucleus scattering is
\begin{equation}
E_{r,\text{max}}=\frac{2E_\nu^2}{m_N+2E_\nu},
\end{equation}
where $E_\nu$ is the incident neutrino energy, and $m_N$ is the mass of the target nucleus. The three-momentum exchange is related to the recoil energy by $q\approx\sqrt{2m_NE_r}$. For neutrino energies below 20 MeV and nuclear targets from ${}^3$He to ${}^{132}$Xe, the maximum recoil energy ranges between 280 keV and 6 keV, meaning that the maximum possible $q$ is quite small, $<$ 1 $\text{fm}^{-1}$. Typical nuclear radii, $R$, are 3-5 fm, then the product $qR<1$. In this regime, the neutrino scatters coherently off the nucleus.
The most important contribution to the neutrino flux in the lowest energy range capable of giving a detectable nuclear recoil (around a keV) comes from the various nuclear fusion and decay processes occurring in the Sun. For this work, we consider only the fluxes of solar neutrinos. Atmospheric neutrinos and neutrinos from the diffuse supernova background are produced at higher energies and with much lower rates, only being detectable in future multi-ton detectors \cite{billard}.
In figure \ref{fig:fluxes} we show the solar neutrino fluxes. Events induced by neutrinos from the grey coloured fluxes will not give events above the threshold considered in this paper. Solar neutrinos produced in the proton-proton chain are in red, while those produced in the CNO cycle are in green.
\section{Neutrino Background Distribution}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{fluxes.eps}
\caption{Solar neutrinos fluxes. The grey colored fluxes will not give events above thresholds considered in this paper. Solar neutrinos are produced in the proton-proton chain reaction (red) and CNO cycle (green).}
\label{fig:fluxes}
\end{figure}
The differential cross section for a neutrino scattering off a polarised nucleon is
\begin{equation}
\begin{split}
\frac{d\sigma}{d\Omega}=&\frac{G_F^2E_\nu^{\prime 2}}{16\pi^2E_\nu^2m_N^2}\bigg\{c_V^2\Big\{(pp_N)(p^{\prime}p_N^{\prime})+(pp_N^{\prime})(p^{\prime}p_N)\\&+m_N\Big[(pS)[(p^{\prime}p_N)-(p^{\prime}p_N^{\prime})]-(p^{\prime}S)[(pp_N)-(pp_N^{\prime})]\Big]\\&-m_N^2(pp^{\prime})\Big\}+c_A^2\Big\{(pp_N)(p^{\prime}p_N^{\prime})+(pp_N^{\prime})(p^{\prime}p_N)\\&-m_N\Big[(pS)[(p^{\prime}p_N)+(p^{\prime}p_N^{\prime})]-(p^{\prime}S)[(pp_N)+(pp_N^{\prime})]\Big]\\&+m_N^2(pp^{\prime})\Big\}+2c_Vc_A\Big\{(pp_N)(p^{\prime}p_N^{\prime})-(pp_N^{\prime})(p^{\prime}p_N)\\&-m_N\Big[(pS)(p^{\prime}p_N^{\prime})+(p^{\prime}S)(pp_N^{\prime})\Big]\Big\}\bigg\}.
\end{split}
\end{equation}
\noindent Here $G_F$ is the Fermi constant, $E_\nu$ is the incoming neutrino energy, $E_\nu^\prime$ is the outgoing neutrino energy, and $c_V$ and $c_A$ are the effective couplings. The neutrino and nucleus initial (final) four-momenta are respectively $p$ ($p^{\prime}$) and $p_N$ ($p_N^{\prime}$), and $S$ is the nuclear spin four-vector. Because the recoil energies are much less than the nuclear masses and neutrino energies, this expression reduces to
\begin{equation}
\begin{split}
\frac{d\sigma}{d\Omega}=&\frac{G_F^2E_\nu^2}{16\pi^2}\{c_V^2+3c_A^2+(c_V^2-c_A^2)\text{cos}\psi\\&-2c_A[(c_V-c_A)\hat{v}.\hat{s}+(c_V+c_A)\hat{v}^\prime.\hat{s}]\},
\end{split}
\label{eq:dsdo}
\end{equation}
where $\psi$ is the scattering angle in the lab frame of the outgoing neutrino with respect to the incoming neutrino direction, $\hat{v}$ and $\hat{v}^\prime$ are respectively the directions of the incoming and outgoing neutrinos and $\hat{s}$ is the direction of the nuclear spin.
\begin{comment}
In the single-particle shell model, the nuclear spin is solely due to the spin of the
single unpaired nucleon. Therefore, for the helium-3, the two protons are paired, their spins cancel out and the unpaired neutron is responsible for the nucleus 1/2 spin. Then, the coherent neutrino-nucleus scattering amplitude is
\begin{equation}
\begin{split}
\mathcal{M}=&\sum_AG\bar{\nu}\gamma^\mu(1-\gamma^5)\nu\bar{n}\gamma_\mu(c_V-c_A\gamma^5)n\\
=&(Nc_V^{\text{neutron}}+Zc_V^{\text{proton}})G\bar{\nu}\gamma^\mu(1-\gamma^5)\nu\bar{n}\gamma_\mu n\\
&-c_A^{\text{unpaired nucleon}}G\bar{\nu}\gamma^\mu(1-\gamma^5)\nu\bar{n}\gamma_\mu\gamma^5n.
\end{split}
\end{equation}
\end{comment}
\begin{figure*}
\begin{minipage}[t]{0.48\linewidth}
\includegraphics[width=.9\linewidth]{01.eps}
\end{minipage}\hfill
\begin{minipage}[t]{0.48\linewidth}
\includegraphics[width=.9\linewidth]{02.eps}
\end{minipage}
\caption{Neutrino-nucleus differential cross section for helium-3 (left) and xenon-129 (right) for $E_\nu$=6.4 MeV and $E_r=E_{r,\text{max}}/2$ as a function of $\theta$, the angle between the incoming neutrino and the nuclear spin directions.}
\label{fig:cross}
\end{figure*}
For low-energy interactions, atomic nuclei behave, in a good approximation, as a collection of independent nucleons, which yields
\begin{equation}
c_V^\text{nucleus}=Zc_V^p+Nc_V^n
\end{equation}
and
\begin{equation}
c_A^\text{nucleus}=c_A^\text{unpaired nucleon}
\label{eq:ca}
\end{equation}
for a DM-nucleus coupling, where $c_V^{p,n}$ and $c_A^{p,n}$ are respectively the effective vector and axial neutrino couplings to protons and neutrons. We take the values shown in Table \ref{tab:couplings}. In equation \ref{eq:ca} it is assumed that the nuclear spin is solely due to the spin of the single unpaired nucleon.
\begin{table}[h!]
\centering
\begin{tabular}{c|c|c}
& $c_V$ & $c_A$ \rule[-1.2ex]{0pt}{0pt}\\ \hline
Proton & $1-4\text{sin}^2\theta_W$ & 1.26 \T\rule[-1.2ex]{0pt}{0pt}\\ \hline
Neutron & -1 & -1.26 \T\\
\end{tabular}
\caption{Effective neutral-current couplings.}
\label{tab:couplings}
\end{table}
From scattering kinematics, it can be found that
\begin{equation}
\text{cos}\psi=1-\frac{E_r}{m_N}\left(\frac{E_\nu+m_N}{E_\nu}\right)^2.
\label{eq:psi}
\end{equation}
Then it follows from that
\begin{equation}
\frac{d\sigma}{dE_r}=\int_0^{2\pi}\frac{d\sigma}{d\Omega}\frac{d(\text{cos}\psi)}{dE_r}d\phi=\frac{2\pi}{m_N}\left(\frac{E_\nu+m_N}{E_\nu}\right)^2\frac{d\sigma}{d\Omega}.
\end{equation}
Figure \ref{fig:cross} shows $\frac{d\sigma}{dE_r}$ for a neutrino scattering off a helium-3 and a xenon-129 nucleus, the most abundant spin-1/2 xenon isotope. In the helium-3 case the coherent spin-independent contribution doesn't overshadow the spin-dependent one.
We can quantify the magnitude of this effect considering the relative amplitude
\begin{equation}
\alpha=\frac{1}{2}\left|\frac{\frac{d\sigma}{dE_r}(0)-\frac{d\sigma}{dE_r}(\pi)}{\frac{d\sigma}{dE_r}(\pi/2)}\right|.
\end{equation}
Its value ranges from 0 (no angular dependence) to 1 (maximum angular dependence). Table \ref{tab:amp} below shows the values of $\alpha$ for different elements taking $E_\nu$=6.4 MeV and $E_r=E_{r,\text{max}}/2$.
\begin{table}[h!]
\centering
\begin{tabular}{c|c}
& $\alpha$ \\ \hline
\T
$^3$He & 0.97 \T\rule[-1.2ex]{0pt}{0pt} \\ \hline
$^{13}$C & 0.41\T\rule[-1.2ex]{0pt}{0pt} \\ \hline
$^{15}$N & 0.36 \T\rule[-1.2ex]{0pt}{0pt}\\ \hline
$^{19}$F & 0.22\T\rule[-1.2ex]{0pt}{0pt} \\ \hline
$^{129}$Xe & 0.04\T \\
\end{tabular}
\caption{Relative amplitude of the differential cross section for different elements.}
\label{tab:amp}
\end{table}
The number of neutrino events is calculated as an integral over the differential rate and the energy dependent detection efficiency $\epsilon(E_r)$:
\begin{equation}
N_\nu=\int_{E_{\text{thr}}}^{E_{\text{max}}}\epsilon(E_r)\frac{dR_\nu}{dE_r}dE_r.
\end{equation}
The differential rate is
\begin{equation}
\frac{dR_\nu}{dE_r}=n_T\; T\int_{E_\nu^\text{min}}^{E_\nu^\text{max}}\frac{dN_\nu}{dE_\nu}\frac{d\sigma(E_\nu,E_r,\theta)}{dE_r}dE_\nu,
\end{equation}
with $n_T$ the number of target nuclei in the detector, $T$ is the exposure time, $\frac{dN}{dE_\nu}$ the neutrino flux, and $\frac{d\sigma}{dE_r}$ the differential cross section, where $\theta$ is the angle between $\hat{v}$ and $\hat{s}$. Rigorously, the solar neutrino flux depends on the annual variation in the Earth-Sun distance. The flux of solar neutrinos vary by around 3\% over a year and the shape of the spectrum is not affected. As a first approximation, we take this flux to be time-independent.
The total number of solar neutrinos per 100 kg-year of helium-3 assuming ideal energy efficiency is shown in Table \ref{tab:ne}. When the detector is fixed at $\theta=\pi$, this number is very low. We can then consider a polarised helium-3 dark matter detector that keeps the angle $\theta$ fixed at $\pi$ to minimise the number of solar neutrinos events.
\begin{comment}
For a threshold energy of 2 keV, the contributions from solar neutrinos are from the boron-8 and hep reactions from the proton-proton chain and the decays of ${}^{15}$O and ${}^{17}$F from the CNO cycle. For 1 keV, the pep reaction and the decay of ${}^{13}$N are included. For thresholds energies of 0.5 keV and 0.2 keV, the ${}^{7}$Be 861 keV line contributes as well.
\end{comment}
For a number of spin 1/2 nuclei, the level of polarisation $P$ is defined as
\begin{equation}
P=\frac{N^\uparrow-N^\downarrow}{N^\uparrow+N^\downarrow},
\end{equation}
where $N^\uparrow$ and $N^\downarrow$ are respectively the number of nuclear spin states +1/2 and -1/2.
In a 100 kg unpolarised detector with a threshold energy of 0.2 keV, around 43 solar neutrinos are expected to be detected every year. The same detector with a 70\% degree of polarisation and fixed optimal angle would detect roughly 13 solar neutrinos each year. In the ideal case where this detector is completely polarised, 1 solar neutrino would be expected each year.
\begin{table}[h!]
\centering
\begin{tabular}{c|c|c|c|c}
$E_\text{thr}$ (keV) &2 & 1 & 0.5 &0.2 \rule[-1.2ex]{0pt}{0pt}\\ \hline
$N_{\text{E}}(0)$ & 3.23& 7.67 & 23.26& 85.45 \T\rule[-1.2ex]{0pt}{0pt} \\ \hline
$N_{\text{E}}(\pi)$ & 0.07& 0.10 &0.19& 0.85 \T\rule[-1.2ex]{0pt}{0pt}\\ \hline
$N_{\text{E,unpolarised}}$ & 1.65& 3.89 &11.73& 43.15 \T\\
\end{tabular}
\caption{Number of solar neutrino events per 100 kg-year of helium-3.}
\label{tab:ne}
\end{table}
\section{Dark Matter}
For dark matter particles the polarisation modulated amplitude is small because it is velocity suppressed \cite{kamionkowski}, so the detector's orientation is largely irrelevant for those events, even though there is a preferred arrival direction for the fastest dark matter particles due to the motion of the solar system through the Galaxy.
Direct dark matter detection results put tight constraints on a $Z^\prime$ with vector couplings to light quarks due to the coherent enhancement in such models. Conversely, models where dark matter scattering on nuclei is spin-dependent are very weakly constrained \cite{kg,axialdm}. For such a class of models, helium-3 detectors would be more efficient than xenon-based detectors since 1 kg of helium-3 has the same number of unpaired neutrons as 90 kg of xenon.
\begin{figure}[b]
\centering
\includegraphics[width=0.49\textwidth]{dm.eps}
\caption{Spin-dependent exclusion curve for an exposure of 100 kg-year, threshold energies of 0.2 keV (blue) and 2 keV (red), and different levels of polarisations. The neutrino floor for xenon (green) has been rescaled by the appropriate $2A^2$ factor, where the factor of two accounts for the fact that roughly half of xenon nuclei have an unpaired neutron. In yellow, the 90\% C.L. limit on SD DM-neutron cross section from LUX \cite{lux}.}
\label{fig:exclusion}
\end{figure}
\begin{comment}
Table \ref{tab:cross} shows the spin-dependent DM-nucleus cross sections necessary to observe the same number of DM events as solar neutrinos for an unpolarised helium-3 detector and for different levels of polarisation with $E_\text{thr}$=0.2 keV and $\theta=\pi$, the optimal angle. None of the situations shown have been ruled out by LUX \cite{lux}.
\begin{table}[h]
\centering
\begin{tabular}{ccccc}
\multicolumn{1}{c|}{$m_\chi$ (GeV)} & \multicolumn{1}{c|}{5} & \multicolumn{1}{c|}{10} & \multicolumn{1}{c}{20} \\ \hline
\multicolumn{1}{c|}{unpolarised} & \multicolumn{1}{c|}{$4.24\times 10^{-41}$} & \multicolumn{1}{c|}{$8.27\times 10^{-41}$}& \multicolumn{1}{c}{$1.64\times 10^{-40}$} \\ \hline
\multicolumn{1}{c|}{70\%} & \multicolumn{1}{c|}{$1.34\times 10^{-41}$} & \multicolumn{1}{c|}{$2.62\times 10^{-41}$}& \multicolumn{1}{c}{$5.18\times 10^{-41}$} \\ \hline
\multicolumn{1}{c|}{80\%} & \multicolumn{1}{c|}{$9.29\times 10^{-42}$} & \multicolumn{1}{c|}{$1.81\times 10^{-41}$}& \multicolumn{1}{c}{$3.59\times 10^{-41}$} \\ \hline
\multicolumn{1}{c|}{90\%} & \multicolumn{1}{c|}{$5.15\times 10^{-42}$} & \multicolumn{1}{c|}{$1.00\times 10^{-41}$}& \multicolumn{1}{c}{$1.99\times 10^{-41}$} \\ \hline
\multicolumn{1}{c|}{100\%} & \multicolumn{1}{c|}{$1.00\times 10^{-42}$} & \multicolumn{1}{c|}{$1.97\times 10^{-42}$}& \multicolumn{1}{c}{$3.90\times 10^{-42}$} \\
\end{tabular}
\caption{DM-nucleus cross sections in cm$^2$ needed to obtain the same number of events as solar neutrinos for an unpolarised helium-3 detector and for different levels of polarisation with $E_\text{thr}$=0.2 keV and $\theta=\pi$.}
\label{tab:cross}
\end{table}
\end{comment}
Figure \ref{fig:exclusion} shows the limits on the spin dependent cross section that we expect to be able to obtain with an exposure of 100 kg-year for a helium-3 detector for thresholds of 0.2 keV (blue) and 2 keV (red). One can expect gamma ray backgrounds with very good shielding of about 1 event keV$^{-1}$kg$^{-1}$day$^{-1}$. To go below this one needs advanced strategies for background rejection, for example self-shielding or analysis of charge to light ratio in xenon detectors. The limits come from performing Monte Carlo simulations of distinguishing signal (DM+neutrino) from background (neutrino only) over bins of 0.2 keV width with an upper energy threshold limit of 20 keV. The addition of polarisation increases the sensitivity by nearly an order of magnitude. In the figure, we don't assume any particular coherent enhancement between dark matter and the nuclei, so the limits represent the spin-dependent constraint, although any coherence effects would be minor for helium-3.
This figure looks quite different to normal dark matter sensitivity plots where the neutrino floor due to solar neutrinos for low mass dark matter is present. The reason for this is that for such a light target, the recoil energies expected from dark matter are very small indeed, while the recoil energies from the solar neutrinos are much larger, resulting in neutrinos being a background for much heavier dark matter. At the masses we plot here, the neutrino background includes contibutions from many different reactions in the Sun. The result is an overall diminishment of sensitivity, rather than one which is only significant at low dark matter masses.
\section{Conclusions}
In this work we looked at the neutrino floor problem for future direct dark matter searches, when neutrino backgrounds from coherent neutrino-nucleus elastic scattering are important. These neutrinos interact either via the vector or axial current, the former of which depends roughly on $N^2$, where N is the number of neutrons, and the latter of which depends on the net spin of the nucleus. Hence for lighter elements such as helium-3 the two currents, labeled spin-independent and spin-dependent respectively, are of similar size.
We investigated how a polarised detector can in theory help reduce this background. We considered a detector based on helium-3 so that the spin-dependent contribution is not suppressed relative to the coherent cross section and the effect of polarisation is maximised.
We showed that the neutrino-nucleus cross section has a strong dependence on the angle between the incoming neutrino for helium-3 and the nuclear spin. This property can be used in the construction of a polarised detector, which is kept at an angle that minimises the number solar neutrinos detected. Such a detector operating at full polarisation would be able to rule out 98\% of solar neutrino events. For dark matter this effect is negligible since it's velocity suppressed and these particles are non-relativistic.
We showed that such discrimination could reduce the neutrino background from the Sun for light dark matter by almost an order of magnitude and performed estimates of the potential performance of such a detector with and without polarisation for different energy thresholds.
The cost of helium-3 is very high relative to other materials which are used for dark matter detection, and we are aware of the difficulties that would be encountered in setting up such an experiment in practice. The purpose of this note was to point out the nice features that such a detector would have if we were able to polarise the nuclei, especially with regards to their interactions with neutrinos. We note that this nice feature of helium-3 comes hand in hand with higher recoil energies for neutrino events, meaning there are more neutrino events above a certain threshold. Detailed studies to optimise this trade off and perhaps to look at other target nuclei would be natural extensions of this work.
Finally, a drop in price of helium-3 would be helpful in making the realisation of such a device feasible.
\section*{Acknowledgment}
We are extremely grateful for comments from Chamkaur Ghag, Mark Kamionkowski, Dan McKinsey, and Jim Wild. TF thanks support from CNPq SwB grant. MF acknowledge support from the STFC and funding from the European Research Council under the European Union's Horizon 2020 program (ERC Grant Agreement no.648680 DARKHORIZONS).
|
\section{Details of the diagrammatic expansion for the Gutzwiller wave function technique}
\subsection{Formal expansion}
The variational analysis of the topological
Anderson lattice Hamiltonian, including the non-local
effects of the onsite interaction, begins with the formulation of the exact
form of the expectation value with the full Gutzwiller wave function (GWF)
of any operator $\mathcal{O}_I$
that involves sites $I=\big\{\g i_1,\g i_2, \dots,\g i_M\big \}$,
\begin{equation}
\begin{split}
\big\langle \, &\mathcal{O}_{I}\,\big\rangle_G = \frac{1}{\langle\psi_G|\psi_G\rangle}
\Big\langle \,{\mathcal{O}}^G_{I}\;\prod_{\g l\not \in I}\,
\mathcal{P}_{\g l}^2 \, \Big \rangle_0 \\
&= \frac{1}{\langle\psi_G|\psi_G\rangle}\sum_{k=0}^\infty\, \fract{x^k}{k!}\, \sum_{\g l_1,...,\g l_k}\!\!\!\!{}^{'}
\Big\langle \,{\mathcal{O}}^G_{I}\, d_{\g l_1}\,d_{\g l_2}\dots d_{\g l_k}\,\Big
\rangle_0\; ,
\label{expeS}
\end{split}
\end{equation}
where $\mathcal{O}^G_{I}\equiv \prod_{i\in I}\,\mathcal{P}_{\g i}^\dagger\;\mathcal{O}_I\;
\prod_{j\in I}\,\mathcal{P}_{\g j}^{\phantom{\dagger}}$.
The expectation values in Eq.~(\ref{expeS}) are evaluated by means of the
Wick's theorem. They can be visualized by a sum of diagrams connecting sites
in the real space with lines representing single particle density matrices,
\begin{equation}
\begin{split
C_{\g{l},\g{l'}}^{ff} &\equiv \langle f_{\g{l}\sigma}^\dagger
f_{\g{l'}\sigma}^{\phantom{\dagger}} \rangle_0,\\
C_{\g{l},\g{l'} }^{fc}&\equiv \langle c_{\g{l}\sigma}^\dagger
f_{\g{l'}-\sigma}^{\phantom{\dagger}} \rangle_0,\\
C_{\g{l},\g{l'}}^{cc}&\equiv \langle c_{\g{l}\sigma}^\dagger
c_{\g{l'}\sigma}^{\phantom{\dagger}} \rangle_0,
\label{linesS}
\end{split}
\end{equation}
where $\g l,\g l'$ are
lattice indices and $\sigma$ denotes (pseudo)spin index for
$c$ ($f$) electrons. In order to evaluate the expectation value of the topological
Anderson lattice Hamiltonian
first we derive the expressions for the following projected operators,
\begin{equation}
\begin{split}
\mathcal{ P}_{\g i} d_{\g i} \mathcal{ P}_{\g i}&=\lambda_d^2[2n_{0f}\ n^{\rm HF}_{\g i}+(1-xd_0)\ d_{\g i} +n_{0f}^2\ \mathcal{ P}_{\g i}^2], \vspace{2pt} \\
\mathcal{ P}_{\g i} n^f_{\g i\sigma} \mathcal{ P}_{\g i}&=(1+xm)\ n^{\rm HF}_{\g i}+\gamma\ d_{\g i}+n_{0f}\ \mathcal{ P}_{\g i}^2, \vspace{2pt} \\
\mathcal{ P}_{\g i} f_{\g i\sigma}^{(\dagger)} \mathcal{ P}_{\g i}&=\alpha\ f_{\g i\sigma}^{(\dagger)}+
\beta\ f_{\g i\sigma}^{(\dagger)} n^{\rm HF}_{\g i}, \label{A11S}
\end{split}
\end{equation}
where additionally we define
\begin{equation}
\begin{split}
n^{\rm HF}_{\g i\sigma}&\equiv n^f_{\g i\sigma} - \langle n^f_{\g i\sigma}\rangle_0 = n^{\rm HF}_{\g i-\sigma} \equiv n^{\rm HF}_{\g i}\\
n_{0f}&\equiv\langle n^f_{\g i\sigma}\rangle_0 \\
\beta&\equiv\lambda_{s}( \lambda_d-\lambda_0),\\
\alpha&\equiv\lambda_s \lambda_0+\beta n_{0f},\\
\gamma&\equiv x(1-2n_{0f}),\\
m&\equiv n_{0f}(1-n_{0f}),\label{A4S}
\end{split}
\end{equation}
where superscript HF stands for the Hartree-Fock form of the operator.
Here, the parameters $\{\lambda_0,\lambda_s,\lambda_d\}$ have the following meaning.
The general, diagonal Gutzwiller correlator acting on the $f$
electrons can be written as,
\begin{equation}
\mathcal{ P}_{\g i}=\sum_\Gamma \lambda_\Gamma\mid\Gamma\rangle_{\g i}\langle\Gamma\mid_{\g i},
\end{equation}
with variational parameters
$\lambda_\Gamma \in \{\lambda_0,\lambda_\uparrow,\lambda_\downarrow,\lambda_d\}$ that
characterize the occupation probabilities for the four possible atomic Fock $f$-states
${\mid\Gamma\rangle_{\g i}\in\{\mid 0\rangle_{\g i},\mid\uparrow\rangle_{\g i},
\mid\downarrow\rangle_{\g i},\mid\uparrow\downarrow\rangle_{\g i}\}}$.
By constraining the correlator with Eq.~(2) (main manuscript) we obtain
\begin{equation}
\begin{split}
\lambda_0^2&=1+xn_{0f}^2,\\
\lambda_\sigma^2=\lambda_{- \sigma}^2\equiv\lambda_s^2&=1-xn_{0f}(1-n_{0f}),\\
\lambda_d^2&=1+x(1-n_{0f})^2. \label{ddS}
\end{split}
\end{equation}
The diagrammatic sums for the operators resulting from (\ref{A11S}) are expressed as
\begin{equation}
S=\sum_{k=0}^\infty \frac{x^k}{k!}S(k), \end{equation}
where $ S\in\{ T^{(1,1)}_{\g i,\g i+\g x},
T^{(1,3)}_{\g i,\g i+\g x},I^{(4)},I^{(2)}\}$,
and the k-th order contributions $S(k)$ read,
\begin{equation}
\begin{split}
T^{(1,1[3])}_{\g i,\g i+\g x}(k)&\equiv\sum_{\g l_1,...,\g l_k}\langle c_{\g i\sigma}^\dagger
[ n^{\rm HF}_{\g i,\sigma}] f_{\g i+\g x,-\sigma} d_{\g l_1,...,\g l_k}\rangle_0^c\\
I^{(4)}(k)&\equiv\sum_{\g l_1,...,\g l_k}\langle d_{\g i}
d_{\g l_1,...,\g l_k}\rangle_0^c,\\
I^{(2)}(k)&\equiv\sum_{\g l_1,...,\g l_k}\langle n_{\g i}^{\rm HF} d_{\g l_1,...,\g l_k}\rangle_0^c.
\end{split}
\end{equation}
Here superscript $c$ denotes only fully connected diagrams \cite{buenemann2012S},
because, with the help of the linked-cluster theorem waiving the summation restriction,
GWF norm cancels out all the disconnected diagrams.
In the Eqs. (\ref{A11S}) we have expressed the projected operators by their HF or relative forms as it significantly speeds up the
convergence of the numerical results concerning the summation of the diagrams \cite{buenemann2012S}.
It is attributed to the fact that by such a construction, all the two-operator averages for a single site and $f$-orbital
(so-called {\it Hartree bubbles}) automatically vanish.
\subsection{Conservation of the number of electrons}
Before the explicit evaluation of the expectation value of the topological Anderson lattice Hamiltonian with the GWF one issue,
particularly important for the modelling of Kondo insulators,
still needs to be resolved.
Due to the fact that the Gutzwiller correlator has a
diagonal form and acts only on the $f$ degrees of freedom it commutes
not only with the operator of the total number of particles but also with those of $f$ and $c$ number of electrons separately.
This yields following equities,
\begin{equation}
\begin{split}
\langle c_{\g{i}\sigma}^\dagger c^{\phantom{\dagger}}_{\g{i}\sigma} \rangle_G&=\langle c_{\g{i}\sigma}^\dagger c^{\phantom{\dagger}}_{\g{i}\sigma} \rangle_0,\\
\langle f_{\g{i}\sigma}^\dagger f^{\phantom{\dagger}}_{\g{i}\sigma} \rangle_G&=
\langle f_{\g{i}\sigma}^\dagger f^{\phantom{\dagger}}_{\g{i}\sigma} \rangle_0.
\end{split}
\end{equation}
It is not automatically satisfied in our approach
because of the real-space and the order cut-offs introduced \cite{Kaczmarczyk2015S}. However,
it can be straightforwardly enforced by the construction.
Due to the odd hybridization, introducing symmetric
cancellation of the diagrams in the real space, the
averages of the pairs of $c$ operators for all $\g i$~and $\g j$, are
the same when taken with either $|\psi_0\rangle$ or $|\psi_G\rangle$,
\begin{equation}
\langle c_{\g{i}\sigma}^\dagger c^{\phantom{\dagger}}_{\g{j}\sigma} \rangle_G=
\langle c_{\g{i}\sigma}^\dagger c^{\phantom{\dagger}}_{\g{j}\sigma} \rangle_0.
\end{equation}
Therefore, what remains, is to ensure that ${\langle f_{\g{i}\sigma}^\dagger f_{\g{i}\sigma}^{\phantom{\dagger}} \rangle_G=
\langle f_{\g{i}\sigma}^\dagger f^{\phantom{\dagger}}_{\g{i}\sigma} \rangle_0}$. The {\it correlated}
(averaged with the GWF) number of $f$ electrons can be rewritten in the following form (cf. Eq.~(\ref{A11S})),
\begin{equation}
\begin{gathered}
\langle f_{\g{i}\sigma}^\dagger f^{\phantom{\dagger}}_{\g{i}\sigma} \rangle_G= n_{0f}+(1+xm)I^{(2
)}+\gamma I^{(4)}.
\label{n0nfS}
\end{gathered}
\end{equation}
The conservation of the total number of particles can be now enforced by the construction by
imposing the cancellation
of the sum of the two last terms
in Eq. (\ref{n0nfS}) \cite{buenemann2012S}, which reduces to the following relation
\begin{equation}
I^{(2)}=\frac{-\gamma}{1+xm}I^{(4)}.
\end{equation}
As a result we calculate diagrammatically only $I^{(4)}$, which due to its structure introduces less
error connected with the real-space cut-off than $I^{(2)}$.\cite{Kaczmarczyk2015S}
\subsection{Expectation value of the Hamiltonian and the derivation of its single-particle effective correspondant}
The resulting expectation value of the topological Anderson lattice Hamiltonian
with the GWF can be readily expressed as following ($\sigma$ and lattice summations are already executed),
\widetext
\begin{equation}
\begin{split}
\frac{\langle{\mathcal{H}}\rangle_G}{L} = 8
t C^{cc}_{\g{i},\g{i}+\g{x}}-2\mu C^{cc}_{\g{i},\g{i}}-2(\mu-\varepsilon_f)n_{0f}
+U\lambda_d^2\Big(n_{0f}^2+\frac{\lambda_s^2\lambda_0^2}{1+mx} I^{(4)}\Big)
+16 V \Big(\alpha T^{(1,1)}_{\g i,\g i+\g x} +\beta T^{(1,3)}_{\g i,\g i+\g x}\Big),
\end{split}
\end{equation}
where L denotes the number of the lattice sites and $\g x$ is a vector in the $\g x$-direction with
the length of the lattice constant. Calculated diagrammatically
$\langle{\mathcal{H}}\rangle_G$ is first optimized with respect to the
variational parameter $x$. Then the optimization of Slater determinant is proceeded what leads to the construction of
the effective single-particle Hamiltonian $\mathcal{ H}^{\rm eff}$. Effectively,
it reduces to fulfilling the condition that its optimal
expectation value with $|\psi_0\rangle$
is coincident with that calculated diagrammatically for $\mathcal{H}$ with
$|\psi_G\rangle$,
\begin{equation}
\begin{gathered}
\delta\langle \mathcal{ H}^{\rm eff}\rangle_0
(C_{\g{l},\g{l'}}^{\alpha,\beta},n_{0f})=
\delta\langle \mathcal{ H}\rangle_{\rm G} (C_{\g{l},\g{l'}}^{\alpha,\beta},n_{0f})
=\sum_{\g l,\g{l'},\alpha,\beta} \Big{(}\frac{\partial \langle \mathcal{ H}\rangle_{\rm G}}
{\partial C_{\g{l},\g{l'}}^{\alpha,\beta}}\delta
C_{\g{l},\g{l'}}^{\alpha,\beta}
+\frac{\partial \langle \mathcal{ H}\rangle_{\rm G}}{\partial n_{0f}}\delta
n_{0f}\Big),
\end{gathered}
\end{equation}
where $\alpha,\beta \in\{f,c\}$. This yields following structure of $\mathcal{ H}^{\rm eff}$ presented in the main text of the manuscript,
\begin{eqnarray}
\mathcal{H}^{\rm eff}\!\!\!\!&=& \sum_{\g i,\g j}
t_{\g i\g j} \, \hat c_{\g i }^\dagger\, \hat c_{\g j }^{\phantom{\dagger}}
+\sum_{\g i,\g j}t_{\g i \g j}^{f}\, \hat f_{\g i }^\dagger \hat f_{\g j}^{\phantom{\dagger}}\nonumber \\
&&\!\!\!\!\!\!\!\!\!+\!\!\!\sum_{\langle\g i,\g j\rangle,\alpha=x,y}\bigg[\,iV_{\g i \g j}^{\rm eff}\,\Big(\hat c_{\g i}^\dagger\,\sigma_\alpha\,
\hat f_{\g j}^{\phantom{\dagger}} +\hat f_{\g i}^\dagger\,\sigma_\alpha\,
\hat c_{\g j}^{\phantom{\dagger}}\Big) +{\rm H.c.}\,\bigg] + \sum_{\langle\langle\g i,\g j\rangle\rangle}\bigg( iV_{\g i \g j}^{\rm eff}\Big[\hat c_{\g i}^\dagger\big(\sigma_x-a\sigma_y\big)
\hat f_{\g j}+\hat f_{\g i}^\dagger\big(\sigma_x-a\sigma_y\big)
\hat c_{\g j}\Big]+\rm{H.c.}\bigg)\nonumber\\
&\equiv& \sum_{\mathbf{k}}\, \hat{\Psi}_\mathbf{k}^\dagger\, \mathcal{\hat{H}}^\text{eff}(\mathbf{k})\,
\hat{\Psi}_\mathbf{k}^{\phantom{\dagger}},\label{neeS}
\end{eqnarray}
The effective microscopic parameters
determining $\mathcal{ H}^{\rm eff}$ read (note that the $c$-electron hopping remains unchanged)
\begin{equation}
\begin{split}
t_{\g i\g j}^{f }\Big|_{\g i\neq \g j}&= \frac{\partial \langle\mathcal{H}\rangle_G}
{\partial F_{\g{i},\g{j}}}, \ \ \ \ \ \ \ \ \
V_{\langle \g i\g j\rangle}^{\rm eff}= \frac{\partial \langle\mathcal{H}\rangle_G}
{i\ \partial W_{\langle \g i,\g j\rangle_x}},\\
t_{\g i\g i}^{f}&=
\frac{\partial \langle\mathcal{H}\rangle_G}
{\partial n_{0f}}, \ \ \ \ \ \
V_{\langle\langle \g i,\g j\rangle\rangle}^{\rm eff} = \frac{\partial \langle\mathcal{H}\rangle_G}
{(i-1) \partial W_{\langle\langle \g i,\g j\rangle\rangle_{x=y}}}.\label{teffS}
\end{split}
\end{equation}
\endwidetext
\subsection{General scheme}
Finding equilibrium groundstate can be described in the following steps:\\
1. Diagrammatic evaluation of $\langle\mathcal{ H}\rangle_{\rm G}$ for the chosen
$|\psi_0\rangle$.\\
2. Minimization of $\langle\mathcal{ H}\rangle_{\rm G}$ with respect to $x$. \\
3. Determination of the single-particle Hamiltonian $\mathcal{ H}^{\rm eff}$ by
optimization of $\langle\mathcal{ H}\rangle_{\rm G}$ with $|\psi_0\rangle$.\\
4. Determination the ground state, $|\psi_0'\rangle$ of $\mathcal{
H}^{\rm eff}$. \\
5. Repeating steps 1-4 in a self-consistent loop until a satisfactory
convergence for $|\psi_0\rangle=|\psi_0'\rangle$ is reached.
\section{Convergence with respect to the order of expansion}
\label{sec2S}
In Fig. \ref{fig2S} we present the resulting indirect gap value for
$\varepsilon=-0.5$ (cf. Fig. 3(a) - main manuscript) for different orders of the diagrammatic expansion.
We do not show results for $k=0$ as on the recovered there mean-field
level the topological Anderson lattice model with initially non-dispersive $f$-states
is capable to describe only semi-metallic state with a zero gap.
In the subsequent orders, $k>0$ the nontrivial behavior of a gap can
be seen. Starting from the second order, $k\geq2$ the metallic state
appears in the strongly correlated regime $U\gtrsim10$.
The practically overlaping values of the gap
for $k=4$ and $k=5$ indicate that at this level of truncation of the
order the satisfactory convergence is reached. Hence all results in the main manuscript
are presented for $k=4$. \ \ \ \ \ \ \phantom{a}
\begin{center}\vspace{-1cm}
\begin{figure}[h]
\includegraphics[width=0.47 \textwidth]{figS1}
\caption{Value of the indirect gap, $\Delta_g$ with respect to the interaction strength
$U$. The negative value of the gap denotes the range of the mutual
energy overlap between the upper and the lower hybridization bands in the metallic state.
The points for $k=4$ and $k=5$ practically overlap indicating that satisfactory convergence with respect to order is reached.}
\label{fig2S}
\end{figure}
\end{center}
\vspace{0.11cm}
|
\section{Introduction}
When the Lifelong Kindergarten group at MIT designed and built an online community for users of the Scratch programming language in 2006 and 2007, the system was designed as a platform to help young people learn to program through remixing \cite{monroy-hernandez_scratchr:_2007}. Every project shared publicly on the Scratch website is released under a license -- explained in child-friendly terms -- that allows unrestricted reuse and modification. Not an easy or obvious decision at the time, the MIT team embraced remixing with the hope that novice users might achieve their goals by downloading, studying, and modifying existing projects created by more skilled users, and through that process, they would learn programming skills.
\begin{figure}[t!]
\centering
\includegraphics[width=0.8\columnwidth]{scratch_code_parallelism.png}
\caption{Scratch code that will cause clicking on a visual object (i.e., sprite) to grow in size, and then shrink, while showing a speech bubble at the same time.
}
\label{fig:scratchcode}
\end{figure}
The idea that remixing and appropriation of content can promote learning has broad currency in the literatures on social computing and youth and media. Appropriation or remixing has been described as an important literacy or skill for young people by Bruckman \cite{bruckman_community_1998}, Jenkins \cite{jenkins_confronting_2006}, Ito \cite{ito_hanging_2009}, Lessig \cite{lessig_remix:_2008}, Manovich \cite{manovich_remix_2005}, and others. One of the most important arguments made in favor of remixing is that it allows users to engage with material created by others with different skills, knowledge, and experiences, and that this exposure promotes learning. The idea that remixing can promote learning has been particularly influential in studies of programming. Scratch designers' embrace of ``creative appropriation'' \cite{monroy-hernandez_scratchr:_2007} is only one example of a strong current in constructionist approaches to learning \cite{papert_mindstorms:_1980} that attempt to place the learning of ``computational thinking'' concepts \cite{wing_computational_2006,wing_computational_2008} in a social context \cite{papert_poetic_1976, bruckman_community_1998}.
This paper empirically tests the theory that young programmers can increase their skills and learn computational thinking concepts through remixing other programmers' code. Using a longitudinal dataset of 2,426,894 projects created by 1,068,502 users over the first five years of the Scratch online community, we examine the association between engagement in remixing and the level and speed with which users demonstrate new computational thinking concepts. Although limited in several ways, our results provide broad support for the idea that young people learn through remixing.
\section{Background}
Remixing has been defined as the reworking and combination of existing creative artifacts, usually in the form of music, video, and other interactive media. The phenomenon is widespread, culturally significant, and controversial. Lessig has suggested that remixing reflects a broad cultural shift spurred by the Internet and a source of enormous creative potential \cite{lessig_remix:_2008}. Manovich has called remixing ``a built-in feature of the digital networked media universe'' \cite{manovich_remix_2005}. Benkler has cited remixing as an example of the power and potential of a new mode of social production \cite{benkler_wealth_2006}.
Remixing online is described as important for at least three reasons. First, theorists like Lessig \cite{lessig_remix:_2008} take a normative position that participation in cultural production is \emph{prima facie} socially beneficial and have suggested that remixing represents a low-cost and accessible form of participation. Second, scholars like Benkler \cite{benkler_wealth_2006} have pointed to remixing as a possible path toward high quality information goods through the mass aggregation of many small contributions in ways that are similar, in both process and effect, to Wikipedia or free/open source software. Finally, theorists including Jenkins \cite{jenkins_confronting_2006} have advanced an argument that remixing leads to learning as a form of legitimate peripheral participation \cite{lave_situated_1991}. Of course, remixing has not been without its detractors. Keen \cite{keen_cult_2007} and Lanier \cite{lanier_you_2010} have both suggested that remixing systematically falls short of its promise, in all three senses, and that remixes are largely derivative, uninteresting, and of poor quality.
Driven by excitement about the promise of remixing, a large body of empirical research has been generated on the subject. Much of this work has focused on mapping dynamics within remixing communities to understand why some users engage in remixing \cite{cheliotis_analysis_2009, yew_social_2009} or why some artifacts are remixed while most never become sites for collaboration \cite{hill_remixing_2013, cheliotis_antecedents_2014, oehlberg_patterns_2015}. Because remixing involves appropriation, another body of work has looked at authorship and credit-giving \cite{luther_edits_2010, monroy-hernandez_computers_2011, kim_appropriate_2015} and to intersections between remixing practice and copyright law \cite{lessig_remix:_2008, hemphill_remix_2009, seneviratne_remix_2010, fiesler_remixers_2014}. In line with theory, empirical work has also considered remixing and its relationship to quality in order to explore if, and when, remixing leads to better results than what is achieved by creators working alone \cite{yu_cooks_2011, hill_cost_2013}.
While empirical remixing research has explored means of supporting and promoting remixing, tested the reception of remixes by users, and tested theories about the quality of remixed outputs, we know of no research that has empirically tested the theory that remixing acts a pathway to learning. That said, the idea that remixing can promote learning has been influential to designers of remixing systems. The creators of the Scratch online community cited Jenkins \cite{jenkins_convergence_2008} to describe their ideal model of ``active engagement'' with content, where ``members of the community can share or appropriate the original building blocks of the Scratch projects they interact with,'' and where users of the Scratch programming language can learn through remixing others' content \cite{monroy-hernandez_scratchr:_2007}. Outside of remixing communities, research on professional software engineers has documented examples of appropriation as a pathway to learning \cite{brandt_two_2009, hartmann_hacking_2008}.
The idea of a community that would use remixing to support learning about programming finds particular support in the theory of constructionism \cite{papert_mindstorms:_1980, resnick_pianos_1996} that motivated the design of Scratch itself \cite{resnick_scratch:_2009}. Constructionist theories hold that we learn best by designing and constructing ``public entities,'' and that this learning is even more meaningful and effective in a social context \cite{papert_poetic_1976}. As a result, research on learning in Scratch has focused on social interactions and relationships within the community \cite{brennan_more_2011,brennan_imagining_2013,fernando_online_2014} as well as on appropriation within the Scratch website \cite{hill_responses_2010,monroy-hernandez_computers_2011,hill_cost_2013,hill_remixing_2013}.
Although users of Scratch learn a wide range of 21st century literacies \cite{jenkins_confronting_2006}, the system was designed to support young people in learning to program. While constructionism has been primarily focused on epistemology and ways of learning, other research on computational thinking (CT) \cite{wing_computational_2006,wing_computational_2008} has influenced and described many of the concepts that Scratch hopes its users will learn. Wing, who coined the term, defines computational thinking as ``the thought processes involved in formulating problems and their solutions so that the solutions are represented in a form that can be effectively carried out by an information-processing agent'' \cite{wing_computational_2010}.
Recent work by Brennan and Resnick \cite{brennan_new_2012} has unpacked CT in terms of concepts, practices, and perspectives. In this treatment, concepts reflect the most basic building blocks of computation (like sequences, loops, and conditionals) while practices and perspectives involve higher level strategies and world-views (like debugging, or reflections on the learning process itself).
\begin{figure*}[t!]
\centering
\includegraphics[width=1.0\textwidth]{scratch_remix.png}
\caption{Remix of a Scratch ``pong'' game (original project on the left), with modification of graphic elements, as well as more sophisticated code for added functionality.}
\label{fig:scratch_remix}
\end{figure*}
Our goal is to test the theory that remixing or appropriation of code is a mechanism through which individuals can be exposed to and learn computational thinking concepts. Attempts to measure learning are always challenging and controversial and doing so in informal learning environments is even more difficult. Several previous studies have looked at learning in Scratch, although most have been case-study based and qualitative \cite{maloney_programming_2008, dahotre_qualitative_2010, peppler_uncovering_2011, scaffidi_skill_2011, brennan_best_2013}. We are particularly inspired by recent work by Yang et al. \cite{yang_uncovering_2015} that improves on previous quantitative studies of learning in informal environments by looking at user ``trajectories.''
Yang et al.\ models learning as growth in the cumulative repertoire of programming tokens, similar to a measure of demonstrated vocabulary, which may grow more or less quickly over time. Building on this approach, we seek to examine how, \emph{ceteris paribus}, a learner's repertoire of programming concepts increases when she engages in remixing others' projects that use unfamiliar blocks. This leads us to our first hypothesis (H1): \emph{changes in a user's programming repertoire will be larger when she has engaged in more remixing activity.}
We are also inspired by work by Scaffidi and Chambers who seek to measure learners' ``breadth'' in terms of the number of types of concepts a user has demonstrated \cite{scaffidi_skill_2011}, and by Brennan and Resnick's description of particular CT concepts \cite{brennan_new_2012}. As a result, we offer a narrower group of hypotheses focused on the relationship between remixing and individual CT concepts. Our second hypothesis (H2) is that \emph{a user who remixes more projects that use a particular computational thinking concept will be more likely to use that concept for the first time in a de novo project than a user who has remixed fewer of these projects.} We test this theory as it applies to each of Brennan and Resnick's CT concepts separately.
\section{Empirical Setting}
The current work contributes to the growing body of previous work in social computing on remixing and learning using the Scratch online community as its empirical setting. Scratch is a public and freely accessible website\footnote{\url{https://scratch.mit.edu}} where a large community of users create, share, and remix interactive media with the Scratch programming environment.
With more than 7 million users, 10 million projects, and 2.6 million remixes,
Scratch is the largest community dedicated to remixing as well as the largest informal community for young people to learn programming.
The Scratch programming environment is visual, block-based, and designed for novice users \cite{resnick_scratch:_2009}. Programs in Scratch are constructed by dragging and dropping visual blocks together (see Figures \ref{fig:scratchcode} and \ref{fig:scratch_remix}). Blocks are analogous to tokens or symbols in the source code of traditional computer programs, and can be used to do things like update a variable, move an object on the screen, play a sound, or repeat a sequence of other blocks. An analogy can be drawn between blocks and lines of code, though blocks are more granular. Scratch includes more than 120 distinct blocks which can be combined together to form ``scripts.'' Figure \ref{fig:scratchcode}, for example, shows two scripts with 4 and 2 blocks side-by-side.
Blocks define the behavior of on-screen graphical objects called ``sprites'' which can interact with each other and the user. Projects on the Scratch website vary enormously in subject matter as well as in complexity in terms of both code and media. Visitors to the Scratch website must create accounts to share projects or to contribute in other ways such as commenting, showing support (i.e., giving ``loveits''), tagging, or flagging projects as inappropriate. Most of the community’s users self-report their ages ranging between 8 and 17 with 12 being the median age for new accounts. As of May 2013 the Scratch programming environment has been fully integrated into the web browser and online community. Our analysis covers the Scratch community during a window from marzo -d, 2007 to abril -d, 2012, before the introduction of the web-based editing environment.
\section{Data and Measures}
\begin{table*}[ht]
\centering
\begin{tabular}{lp{1.4in}p{4.5in}}
\toprule
Concepts & Measure & Scratch Blocks \\
\midrule
Loops & Uses looping blocks (e.g., Forever block) & forever, foreverIf, repeat, repeatUntil\\
\\
Parallelism & Parallel scripts with same ``hat'' block. & startHatTriggered, eventHatTriggered, keyHatTriggered, mouseHatTriggered\\
\\
Events & Uses ``when $<>$'' hat blocks & eventHatTriggered, keyHatTriggered, mouseHatTriggered, bounceOffEdge, turnAwayFromEdge, touching, touchingColor, colorSees, mousePressed, keyPressed, isLoud, sensor, sensorPressed, distanceTo\\
\\
Conditionals & Uses conditional blocks (e.g. ``if'' block) & waitUntil, foreverIf, if, ifElse, repeatUntil, bounceOffEdge, turnAwayFromEdge, touching, touchingColor, colorSees, mousePressed, keyPressed, isLoud, sensor, sensorPressed, lessThan, equalTo, greaterThan, and, or, not, listContains\\
\\
Operators & Uses operator blocks (e.g. ``+'' or ``or'' blocks) & lessThan, equalTo, greaterThan, and, or, not, add, subtract, multiply, divide, pickRandomFromTo, concatenateWith, letterOf, stringLength, mod, round, abs, sqrt, sin, cos, tan, asin, acos, atan, ln, log, e\textasciicircum, 10\textasciicircum\\
\\
Data & Uses data blocks (e.g. Variable block) & setVarTo, changeVarBy, showVariable, hideVariable, readVariable, addToList, deleteLineOfList, insertAtOfList, setLineOfListTo, contentsOfList, getLineOfList, lineCountOfList, listContains\\
\bottomrule
\end{tabular}
\caption{Mapping of CT concepts to Scratch blocks.}
\label{tab:concepts}
\end{table*}
The Scratch online community uses a database-driven web application that stores an extensive range of metadata on projects, users, and interactions on the website. This database also identifies, tracks, and presents data on whether projects are created through remixing. Additionally, the website stores each of the raw Scratch project files which can be further analyzed to reveal details such as projects' programming code and media elements. Our dataset is constructed by combining exported metadata about Scratch’s users, projects, and interactions with algorithmic analyses of each project.
Our unit of analysis is the Scratch project and our dataset includes every project shared on the Scratch online community from the moment the first project was shared in Scratch on marzo -d, 2007 through abril -d, 2012 -- a total of 2,426,894 projects.
Because our study aims to measure learning by looking at within-user changes, our analysis focuses on the 173,053 users who have shared at least two projects. We mark age data as missing for 4,354 users who report their age (at the time that they shared their first project) as less than 4 or more than 90 years. Finally, we omit projects for which we do not have data because of technical errors in our analytic tools or because of corruption in the project files. For analyzing overall repertoire size, we consider 1,625,988 de novo projects shared by all users within our restricted dataset, out of which we omit 4,059 projects due to missing data. While analyzing for specific concepts, we omit 4,950 projects due to missing data and analyze a total of 2,280,709 projects shared or remixed by the same set of users. For parallelism, we omit 11,229 projects due to technical errors in our analytic tools or missing or corrupt data and analyze a total of 2,274,435 projects.
To test hypothesis H1 about growth in programming repertoire, we operationalize a user's repertoire of programming concepts as the number of programming block types that they have ever used. To test hypothesis H2, we measure the presence or absence of particular computational thinking concepts. Toward this end, we adopt Brennan and Resnick's mapping of Scratch blocks to CT concepts \cite{brennan_new_2012} as described in Table \ref{tab:concepts}. For example, use of the ``repeat'' block in Scratch is one indicator of the use of the CT concept of \emph{loops}. To measure the CT concept of ``parallelism,'' we parse projects' code structure to detect multiple event-handling blocks that are listening for the same event. Figure \ref{fig:scratchcode} is a minimal example of Scratch code expressing parallelism with two instances of the same event-handler block with different blocks attached to each.
We deviate from Brennan and Resnick only in that we do not attempt to measure the CT concept of ``sequences'' as any block connected to another block is a sequence and nearly every Scratch project includes this.
\begin{table}[t!]
\centering
\begin{tabular}{rrrrl}
\toprule
& {M} & {$\bar{x}$} & {$\sigma$} & {Range} \\
\midrule
Cum. Repertoire & 23 & 28 & 21 & [0,142] \\
Remixes & 0 & 3 & 15 & [0,1008] \\
De Novo Projects & 4 & 13 & 46 & [1,6347] \\
Comments & 0 & 53 & 601 & [0,64374] \\
User Age (Years) & 14 & 18 & 10 & [4,90] \\
Experience (Days) & 13 & 99 & 206 & [0,1829] \\
Total Blocks & 146 & 767 & 3402 & [0,525964] \\
Downloads & 1 & 19 & 99 & [0,5591] \\
\bottomrule
\end{tabular}
\caption{Summary statistics for measures of users at the point of data collection. Columns are included for the median (M). mean ($\bar{x}$), standard deviation ($\sigma$) and range.}
\label{tab:sum}
\end{table}
Because our behavioral measures of learning may be influenced by a wide range of factors other than the amount of remixing a user has engaged in, we also include a range of control variables. Most critically, we include a count of the number of de novo projects a user has shared (\emph{De Novo Projects}) which provides the independent variable for modeling growth in repertoire associated with increased experience.
For some users of Scratch, interaction and sharing projects is primarily a social activity. Because these users' repertoires may grow more slowly, we include a control for the number of comments received by a user (\emph{Comments}); we expect a negative relationship between this measure and block repertoire. Because learning is related to development in general, we also include a self-reported measure of age in years at the moment that each project was shared (\emph{User Age}) and the age of each account in days (\emph{Experience}). Both of these variables may capture the sophistication of the user and we expect a positive relationship between these two variables and block repertoire.
In designing our study, we were concerned by several active sub-communities within Scratch that are characterized by large amounts of remixing and small amounts of project code. For example, many Scratch users engage in Scratch primarily through ``coloring-contests'' where users use remixing to modify image media in a given project \cite{nickerson_appropriation_2011}. Since these projects tend to have few programming blocks, we were concerned that this type of remixing might bias our estimate of the relationship between remixing through exposure to code. To account for this, we constructed a control for the aggregate corpus size across each user's de novo projects (\emph{Total Blocks}).
\begin{figure}[t]
\includegraphics[width=\columnwidth]{concept_survival_functions.pdf}
\caption{Kaplan-Meier ``survival'' curves for users in Scratch. The $y$ axes show the portion of users who have never uploaded a \emph{de novo} project using a given programming concept. The $x$ axes show ``time'' in terms of the number of \emph{de novo} projects. }
\label{fig:km}
\end{figure}
For our final control measure, we use the number of projects a user has downloaded (\emph{Downloads}). \emph{Downloads} is a complicated measure for several reasons. During the timespan our dataset covers, the only way to view the source code of a project shared on Scratch was to download it. If a user downloads a project, learns how to use a new block, but does not \emph{share} a modified version of the downloaded project publicly, it will not be counted in \emph{Remixes} in our dataset. That said, exposure to and reuse of a block through downloading would constitute ``remixing'' as defined by Manovich \cite{manovich_remix_2005} and as appropriation as defined by Papert \cite{papert_mindstorms:_1980}. In this sense, the relationship between downloads and measures of learning may itself reflect evidence of learning through remixing. Although our analysis focuses on the more conservative measure of remixing used within the Scratch community, we urge readers to interpret our results in this context.
Table \ref{tab:sum} includes summary statistics for each of our variables at the user level at the end of our period of data collection on abril -d, 2012. As is typical for data on participation in online communities, nearly all the variables are highly skewed. To assist in interpretation and the satisfaction of parametric assumptions in our models, we apply a log-transformation to each of our independent variables to ensure that their distribution of values more closely approximates normality.
\section{Analysis}
For H1, we analyze longitudinal data that represents the block repertoire of each Scratch user, use projects as our unit of analysis, and include an observation for every de novo project shared in Scratch during our window of data collection by a user with two or more projects. To model the effects of remixing within users, we use panel regression models with user level fixed effects that are equivalent to fitting a dummy variable for every user in our sample \cite{singer_applied_2003}. Our formal model to test H1 is shown below:
\begin{equation*}
\begin{split}
\mathrm{Cumulative~Repertoire} = \beta_{0} + \beta_{1}\log{\mathrm{Remixes}} \\
+ \beta_{2}\log{\mathrm{De~Novo~Projects}} + \beta_{3}\log{\mathrm{Comments}} \\
+ \beta_{4}\log{\mathrm{User~Age}} + \beta_{5}\log\mathrm{Experience} \\
+ \beta_{6}\log{\mathrm{Total~Blocks}} + \beta_{7}\log{\mathrm{Downloads}} \\
+ \boldsymbol{\beta_{u}}\mathbf{USER}\boldsymbol{_u} + \varepsilon
\end{split}
\end{equation*}
To test our theory about the effect of coloring-contests, we fit an exploratory model (M1) without controlling for \emph{Total Blocks} or \emph{Downloads}, a model (M2) with the \emph{Total Blocks} control, and a model (M3) which also adds the \emph{Downloads} control.
To test H2, we model the effect of remixing projects that contain particular CT concepts on the likelihood of the user employing those concepts in subsequent de novo projects. The demonstration of concepts can happen at any time during the tenure of a Scratch user and is, as a result, well suited to modeling using a continuous time-survival approach using the Cox proportional hazards model \cite{singer_applied_2003}. Originally created for epidemiology models of actual human survival against disease, the language of these models can seem strange when applied to positive events like learning. To support a Cox proportional hazards approach, we modified our dataset so that each observation represents a spell between projects.
Figure \ref{fig:km} includes non-parametric plots that show the proportion of users who have never used a particular block as a function of the number of projects they have shared. Users contribute data until they either use the concept, or they are censored because they did not contribute additional projects. Cox models estimate the association between predictor variables and the ``risk'' of experiencing an event as multipliers of a baseline ``hazard'' function that is itself a function of the ``survival'' functions shown in Figure \ref{fig:km}.
Rather than treating time as calendar time, which produced substantively similar estimates, we use the number of projects shared as our baseline time variable $t$. The independent variable in our Cox models is the number of remixes a user has shared that use the CT concept in question (\emph{Remixes w/ Concept}). We also include \emph{Remixes}, capturing total remixing activity as a control alongside the other control variables used in our repertoire models. We model the risk of a user demonstrating a new CT concept, having shared a given number of projects $\lambda(t)$, as a function of a baseline hazard function $\lambda_{0}(t)$ in the following model:
\begin{equation*}
\begin{split}
\lambda(t\vert X) = \lambda_{0}(t) \exp\{ \beta_{1}\log{\mathrm{Remixes~w/~Concept}} \\
+ \beta_{2}\log{\mathrm{Remixes}} + \beta_{3}\log{\mathrm{De~Novo~Projects}} \\
+ \beta_{4}\log{\mathrm{Comments}} + \beta_{5}\log{\mathrm{User~Age}} \\
+ \beta_{6}\log\mathrm{Experience} + \beta_{7}\log{\mathrm{Total~Blocks}} \\
+ \beta_{8}\log{\mathrm{Downloads}} \}
\end{split}
\end{equation*}
We estimate a separate model for each CT concept: conditionals, data, events, loops, operators, and parallelism.
\section{Results}
\begin{table}
\begin{center}
\begin{tabular}{l c c c }
\hline
& M1 & M2 & M3 \\
\hline
ln Remixes & $-0.304^{*}$ & $1.402^{*}$ & $0.253^{*}$ \\
& $(0.017)$ & $(0.016)$ & $(0.017)$ \\
ln De Novo Projects & $13.073^{*}$ & $4.779^{*}$ & $4.750^{*}$ \\
& $(0.023)$ & $(0.026)$ & $(0.026)$ \\
ln Comments & $0.413^{*}$ & $0.643^{*}$ & $0.080^{*}$ \\
& $(0.011)$ & $(0.010)$ & $(0.011)$ \\
User Age & $2.367^{*}$ & $2.912^{*}$ & $2.517^{*}$ \\
& $(0.019)$ & $(0.018)$ & $(0.018)$ \\
ln Experience & $0.862^{*}$ & $0.128^{*}$ & $-0.046^{*}$ \\
& $(0.008)$ & $(0.007)$ & $(0.007)$ \\
ln Total Blocks & & $5.965^{*}$ & $5.823^{*}$ \\
& & $(0.011)$ & $(0.011)$ \\
ln Downloads & & & $2.237^{*}$ \\
& & & $(0.012)$ \\
\hline
R$^2$ & 0.751 & 0.793 & 0.798 \\
Adj. R$^2$ & 0.673 & 0.711 & 0.715 \\
Num. obs. & 1578362 & 1578362 & 1578362 \\
\hline
\multicolumn{4}{l}{\scriptsize{$^*p<0.001$}}
\end{tabular}
\caption{Fitted regression models for Scratch users' cumulative block repertoires. All models use user-level fixed effects and reflect within-user estimates.}
\label{tab:rep}
\end{center}
\end{table}
Our tests of H1 and the relationship between remixing and blocks in a user's repertoire are shown in Table \ref{tab:rep}. In our first exploratory model M1, we find a negative association between the number of remixes a user has shared and the size of their repertoire. Controlling only for the total number of projects, comments, and users' age and experience, a one log-unit increase in the number of remixes a user shares is associated with a within-user decrease of 0.3 distinct blocks in her repertoire. Of course, this negative effect may simply reflect the fact that many users engaged in remixing are engaged in coloring-contests that involve little or no code. Model M2 attempts to address this concern by adding our control for the total number of blocks shared. Once we control for total block use, we find the parameter estimate for \emph{Remixes} is switched in sign and much larger. In this expanded model, we estimate that a one log-unit increase in number of remixes shared is associated with a 1.4 block increase in repertoire.
M3 adds a control variable for \emph{Downloads} and provides confirming evidence for our prediction that downloading -- a step toward remixing in Scratch -- moderates the effect of remixing. Model M3 estimates that a one log-unit increase in downloading is associated with a 2.24 block increase in a user's repertoire. This supports the idea that remixing is a pathway to learning, in the broader conceptual sense of remixing that is used by many theorists. However, even with this strong control for downloads, this model estimates that a one log-unit increase in the number of remixes shared on Scratch is associated with a 0.25 block marginal increase in repertoire. Although reflecting only a portion of a block, this effect is well estimated ($\sigma=0.02$) and statistically significant even at the conservative $\alpha=0.001$ level. Goodness of fit statistics shown in Table \ref{tab:rep} (e.g., $R^2=0.80$) suggest that these models fit the data well, and they explain a large majority of variation in Scratch users' cumulative repertoire of blocks.
The marginal effect sizes in these repertoire models are small, but can still be substantively meaningful. For example, M2 predicts that, holding all control variables at median values for similar users, a user who has shared 100 de novo projects -- but who had never remixed -- would be predicted to have a block repertoire of 81 blocks. If the user shared one remix for every three de novo projects, she would have a predicted repertoire of 86 blocks instead. When we control for downloads in M3, the estimated difference for this prototypical user is approximately one full block. Although small, this still reflects more than 1\% of the user's repertoire, and almost as large a proportion of all Scratch blocks.
\begin{table*}
\begin{center}
\begin{tabular}{l c c c c c c }
\hline
& Loops & Parallelism & Events & Conditionals & Operators & Data \\
\hline
ln Remixes w/ Concept & $0.388^{*}$ & $0.235^{*}$ & $0.282^{*}$ & $0.295^{*}$ & $0.158^{*}$
& $0.204^{*}$ \\
& $(0.009)$ & $(0.009)$ & $(0.009)$ & $(0.008)$ & $(0.007)$
& $(0.007)$ \\
ln Remixes & $-0.283^{*}$ & $-0.406^{*}$ & $-0.355^{*}$ & $-0.463^{*}$ & $-0.433^{*}$
& $-0.530^{*}$ \\
& $(0.012)$ & $(0.012)$ & $(0.012)$ & $(0.010)$ & $(0.009)$
& $(0.009)$ \\
ln Comments & $-0.586^{*}$ & $-0.286^{*}$ & $-0.574^{*}$ & $-0.728^{*}$ & $-0.702^{*}$
& $-0.657^{*}$ \\
& $(0.006)$ & $(0.006)$ & $(0.006)$ & $(0.005)$ & $(0.005)$
& $(0.005)$ \\
User Age & $-0.006^{*}$ & $-0.006^{*}$ & $-0.010^{*}$ & $-0.011^{*}$ & $-0.024^{*}$
& $-0.026^{*}$ \\
& $(0.000)$ & $(0.000)$ & $(0.000)$ & $(0.000)$ & $(0.000)$
& $(0.001)$ \\
ln Experience & $-0.152^{*}$ & $-0.286^{*}$ & $-0.191^{*}$ & $-0.100^{*}$ & $-0.061^{*}$
& $-0.045^{*}$ \\
& $(0.004)$ & $(0.004)$ & $(0.004)$ & $(0.003)$ & $(0.003)$
& $(0.003)$ \\
ln Total Blocks & $0.283^{*}$ & $0.431^{*}$ & $0.330^{*}$ & $0.287^{*}$ & $0.306^{*}$
& $0.323^{*}$ \\
& $(0.003)$ & $(0.004)$ & $(0.004)$ & $(0.004)$ & $(0.004)$
& $(0.004)$ \\
ln Downloads & $0.109^{*}$ & $0.033^{*}$ & $0.109^{*}$ & $0.068^{*}$ & $0.148^{*}$
& $0.139^{*}$ \\
& $(0.006)$ & $(0.006)$ & $(0.005)$ & $(0.005)$ & $(0.005)$
& $(0.005)$ \\
\hline
AIC & 963742 & 993339 & 908564 & 1193522 & 1231333
& 1152895 \\
R$^2$ & 0.142 & 0.143 & 0.150 & 0.140 & 0.081
& 0.073 \\
Max. R$^2$ & 0.996 & 0.995 & 0.995 & 0.966 & 0.888
& 0.829 \\
Num. events & 43615 & 44876 & 41122 & 51111 & 50658
& 46931 \\
Num. obs. & 179522 & 190933 & 179061 & 370524 & 584950
& 681062 \\
Missings & 4654 & 5415 & 5064 & 9148 & 15179
& 18393 \\
\hline
\multicolumn{7}{l}{\scriptsize{$^*p<0.001$}}
\end{tabular}
\caption{Fitted Cox proportional hazard models that estimate the ``risk'' that a Scratch user will share a de novo project that uses a block associated with a computational concept (e.g., loops) for the first time.}
\label{tab:surv}
\end{center}
\end{table*}
Our results for hypothesis H2 are shown in the six models described in Table \ref{tab:surv}. Although the relative size of the effects vary between concepts, the pattern of results are consistent. The marginal effect of the number of remixes with the CT concept is positive for all concepts. Parameter estimates can be interpreted as a ``magnifier'' of the hazard function. For example, a user who has remixed one log-unit more projects with conditionals would be estimated as being at 1.34 ($e^{0.29}$) times the risk of using conditionals in one of their de novo projects compared to an otherwise identical user.
To aid in interpretation, we include plots of model-derived estimates of the proportion of users who have used computational concepts for prototypical Scratch users in Figure \ref{fig:protosurv}. Estimates are shown for users from their first project up to 166 projects (the 99$\mathrm{th}$ percentile). Estimates are shown for two prototypical users: the first user has never shared a remix that uses the concept, while the second user has shared three remixes that do. All other variables are held at their overall sample median across all periods. The visualization shows that although the relative differences vary between concepts, Scratch users are systematically at higher ``risk'' of using blocks that demonstrate a given CT concept if they have remixed several projects that contain that concept.
\begin{figure*}[t]
\includegraphics[width=\textwidth]{concept_prototypical_values.pdf}
\caption{Plots of model-derived estimates of the proportion of users who have used blocks associated with CT concepts for prototypical Scratch users. Estimates are shown for two prototypical users: (a) a user who has never shared a remix that uses the concept, and (b) a user has shared three such projects. All other variables are held at their sample median.}
\label{fig:protosurv}
\end{figure*}
With Cox models, one concern is that the assumption of proportional hazards may be violated if a variable has a stronger or weaker association at different points in ``time.'' Unfortunately, the test for proportionality is a function of the size of one's dataset, and our dataset is far too large for the test to be useful. To address whether our inference is affected, we examined Schoenfeld residual plots and found that average values are largely flat, although in some cases they appear to decrease toward zero as more projects are shared. Although we believe that a violation of the proportional hazards assumption is not affecting our fundamental inference, we caution readers to remain skeptical about the specific parameter estimates in the models.
Overall, our controls correspond to our predictions. \emph{Total Blocks} has a positive association with learning across our models and acts to attenuate the effects of other variables. Although previous work offered mixed predictions for the relationship between \emph{Comments} and learning, the results -- a positive association with repertoire but a negative association with concept use across the models in Table \ref{tab:surv} -- was unexpected and suggests a particularly interesting area for future research. Results for our controls for \emph{Age} and \emph{Experience} were also contrary to our predictions in the survival models. Of particular note are the estimates associated with \emph{Remixes} in Table \ref{tab:surv}. Once we control for remixes including a particular concept, the effect of the raw number of remixes is consistently negative. This is not surprising, since remixing projects that do not involve a concept seems unlikely to increase the likelihood that a particular concept will be used. We believe that the negative effect is caused by social forms of remixing like coloring-contests that are imperfectly measured by \emph{Total Blocks}.
Also worth noting is that \emph{Downloads} is consistently positive across all the models that include it. This reveals an interesting secondary result, as it suggests that simply looking at projects has a strong positive association with our measure of learning, and that Scratch may help support learning through less active forms of engagement. Alternatively, as discussed above, this positive association can also be seen as evidence for learning through a broader definition of remixing.
The results from both sets of models provide evidence that supports our original hypotheses. Our regression models of block repertoire in Table \ref{tab:rep} provide support for H1 and suggest that, taking the raw amount of code produced into account, users who engage in more remixing tend to have larger repertoires of programming blocks than users who engage in less. The survival models shown in Table \ref{tab:surv} and Figure \ref{fig:protosurv} point to support for H2, and suggest that remixing more projects with particular computational thinking concepts is associated with using those concepts in new projects. Among the six concepts, those with weaker associations with remixing in our data also showed lower adoption in general. This may indicate a potential lacuna in the learnability of concepts like data and operators, and warrants further investigation.
\section{Limitations}
The validity of our findings may be affected by a number of threats and limitations. One of these threats arises from the fact that our measures of learning may increase for reasons unrelated to, but correlated with, levels of remixing. To show that remixing \emph{causes} learning, we would need exogenous sources of variation in remixing which we do not have. This is a common limitation in learning research and in research on informal learning environments in particular. We have tried to address this concern by adding control variables that address the effects of likely confounds. Although we have included variables that attempt to control for users' popularity, experience, age, and technical sophistication, there may be other important controls that we have omitted.
A limitation of our hazard models is that a Scratch user may copy a stack of blocks that includes a block we are looking for without understanding the underlying concept behind the block. In robustness checks, we attempted to address this by using an alternative threshold for demonstration of a CT concept in each of our hazard models that ignores any demonstration of a CT concept if it occurs in a script that is identical to one used in a previous remix. Results are included in the supplemental material. With this alternate threshold, we discarded 2-15\% of de novo projects that were previously associated with a given CT concept. Because this criterion omits all pure cut-and-pasting by remixing users as evidence of learning, this acts as a conservative test that can only reduce the strength of our findings. In all six cases, the results are still positive, statistically significant, and only moderately reduce the parameter estimates for \emph{Remixes w/ Concept}.
Another threat stems from the fact that users join Scratch with different skills and initial repertoires. For example, a user who joins with a deeper initial knowledge base may grow more slowly but also be more likely to remix. In a related sense, Yang et.\ al.\ \cite{yang_uncovering_2015} take initial block repertoire into account when comparing user learning trajectories. To mitigate this threat, we added a control variable to our repertoire models that captures users' initial block repertoire. The addition of this control did not substantively impact our results.
A final threat is that the design of Scratch may make the CT concept of parallelism emerge unintentionally. For example, as Scratch projects often have multiple sprites, and since the standard way to start a Scratch project is to have an event-handler for ``when green flag clicked,'' a large number of multi-sprite projects may exhibit parallelism even though the user may not be able to use the concept more generally. To address this, we extended our algorithm to discard instances of parallelism between sprites, and only take into account parallel scripts that are within the same sprite. Results are included in the supplemental material. In line with our expectations, this specification strengthens the size of our parameter estimate as we discard possible unintentional parallelism.
Several of these threats hint at a broader, unavoidable limitation: no behavioral measure of learning can see inside users' minds to know what they have learned. In this sense, measuring learning in informal learning environments like Scratch is itself an open area of research which we believe this work contributes to. Although we believe that our measures of learning build and improve on previous work in this area (e.g., \cite{maloney_programming_2008, scaffidi_skill_2011, yang_uncovering_2015}), other metrics may lead to different results and conclusions.
Finally, like other work that studies activity within a single online community, questions of generalizability are both important and difficult to answer. Without further research, we cannot know if these findings generalize beyond the users in our sample. A more nuanced version of this limitation involves the fact that the large majority of Scratch users never share more than a very small number of projects. Scaffidi and Chambers' \cite{scaffidi_skill_2011} point that learning is unlikely to happen when most users engage only for short periods of time reflects both a methodological and substantive limitation that is critical to keep in mind when considering the impact and application of our findings.
\section{Discussion}
Our work finds support for the theory that users can learn through remixing. At the outset, we find that a Scratch user's repertoire of programming blocks is positively associated with the amount of remixing they engage in, controlling for the total amount of code a user is publishing. Diving deeper, we also find that users' demonstration of six key computational thinking concepts is positively associated with users' exposure to these concepts through remixing. We find our results are robust to a series of potential threats to validity. Although limited in a number of ways, these results provide support for proponents of remixing as a pathway to learning.
As discussed, the estimated effect sizes for our predictors are relatively small. Treating the relationship between \emph{Downloads} and learning as part of remixing's effect is suggestive of more substantive effects, but does not add up to a large effect itself. This is not entirely unexpected. Although Scratch was designed to promote learning through remixing, remixing itself is very unconstrained on the platform and has become dominated by genres like coloring contests which were unanticipated by Scratch's designers. Given the informal and messy nature of observational data from Scratch, we believe that our marginal effects still reveal important evidence of learning, and it is easy to imagine stronger effects in a more structured context.
For example, teachers could frame and construct remixing experiences so that learners are exposed to code-intense projects or to particular concepts. Brennan's work on combining formal learning environments with Scratch is one useful guide for future work in this direction \cite{brennan_best_2013}. Larger effects might also be achieved through increased structure within informal environments. For example, research on other remixing platforms has suggested that large amounts of remixing activity can be encouraged and directed to particular source material by site administrators \cite{cheliotis_analysis_2009}. Even within Scratch, previous studies document efforts to promote ``high value'' forms of remixing through ``collaboration camps'' \cite{Roque:2012:TCD:2307096.2307130}.
Our work makes a series of contributions. Our most fundamental contribution is in testing the widespread and influential theory that remixing is positively associated with learning. Second, our paper builds on previous work to describe two novel methods for measuring computational thinking quantitatively. Because the Scratch dataset is becoming more widely available to researchers through a nascent public release, we hope our methods will provide a point of departure for future studies of Scratch. We also believe our approach can be adapted to other informal learning environments.
Of course, this work is only a first step. We hope to build on the current study by testing the relationship between our measures of learning and other predictors like levels and types of socialization. We hope that future work will critique and build on our theories, our measures, and our approach. As quantitative analysts of learning, we know that any attempt to measure learning is necessarily reductionist and potentially dangerous. We present our findings with the hope that others will build and improve upon our efforts.
Our findings have implications for designers interested in promoting learning, and our results support calls to incorporate remixing into the design of social computing systems. The differences between effect sizes in our hazard models suggest that remixing may be more effective at promoting engagement with some concepts, like loops, than others, like operators and data. This may be because the type of projects that are popular in the Scratch community tend to be those that leverage certain concepts over others. Alternatively, these differences may point toward certain concepts being less understandable or learnable, at least within the context of Scratch. Both of these reflect opportunities for designers to foster a wider range of engagement. Our technique for measuring learning has immediate implications and usefulness for the designers of informal learning systems. For example, we hope to work with the Scratch team at MIT to use the measures in this study to evaluate the design of new Scratch features.
As researchers and designers attracted to the promise of remixing, we feel our results are a welcome validation of the idea that remixing can act as a pathway to learning. That said, promoting remixing requires prolonged engagement, and remains extremely difficult in informal learning environments. Moreover, promoting remixing among engaged users remains both difficult and fraught with unpalatable trade-offs \cite{hill_remixing_2013, cheliotis_antecedents_2014}. The fact that remixing may be associated with learning does not make it any easier to promote or any better along other dimensions. Although our results provide evidence that the promise seen in remixing may not be misplaced, enormous work remains to realize its potential.
\section{Acknowledgments}
We would like to thank the Lifelong Kindergarten group at the MIT Media Lab for creating Scratch, as well as the millions of Scratch users who create and participate on the Scratch website. We would also like to acknowledge Mitchel Resnick, Natalie Rusk, Samantha Shorey, Samuel Woolley, and our anonymous reviewers for their thoughtful feedback. Financial support for this work came from the National Science Foundation (grants DRL-1417663 and DRL-1417952).
\balance{}
\bibliographystyle{SIGCHI-Reference-Format}
|
\section{Introduction}
\label{intro}
The following problem is well-known in the area of Probability. A biased coin is given, but the probability of heads is unknown. We flip it $n$ times and get $k$ heads. The problem is to estimate the probability of heads.
The most typical approach to solve this problem is the Maximum Likelihood method (see e.g. \cite{LC} or \cite{P}). Let $p=P(heads)$. Then
$$P(k \ {\rm heads\ out\ of}\ n\ {\rm tosses})= {n \choose k } p^k(1-p)^{n-k}.$$
Since we have absolutely no information about $p$, we choose an estimator $\widehat{p}\in [0,1]$ for which this expression is maximal, that is, $\widehat{p} = k/n$.
This approach has several shortcomings.
1.) For small $n$ we get unrealistic estimations. For example, if $n=1$ and we get a head, the method gives the estimation $p=1$, and if we get a tail, the method gives $p=0$.
2.) the method yields the most likely value of $p$, but does not take into account the error in the estimation. This can be seen in the following example: suppose we flip the coin $n=4$ times and get $k=2$ heads; of course, $\widehat{p} = .5$. However, while in the literature a coin is often considered ``reasonably fair'' when $.45 \leq p \leq .55$, in this case it is easy to check that $P\left((|p-.5|>.05)\ \vert \ (n=4, k=2)\right) = .8137...\,$, assuming $p$ is chosen following a uniform distribution, and, thus, the probability of dealing with a biased coin is high despite the Maximum Likelihood estimator.
Another approach for estimating the probability $p$ of a biased coin deals with Bayesian or Minimax Estimation. While in the former the goal is minimizing the average risk, in the latter the aim is to minimize the maximum risk; nevertheless, there are other well-known methods, such as the so-called Uniform Minimum Variance Unbiased (UMVU) estimator (for more information, see \cite{LC}). These methods require the use of a \textit{loss function}, which measures the difference between the parameter and its estimator. In the current paper, the loss function $|p-a_i|$ will be considered, where $a_i$ is the estimation of $p$ if $i$ heads are obtained after $n$ tosses, for $i=0,\ldots,n\,$. Then, the expected value of this loss function, commonly called \textit{risk} or \textit{penalty function}, is given by
\begin{equation}
\label{penalty}
D(a_0,...,a_n;p):=\sum_{k=0}^n {n \choose k} p^k(1-p)^{n-k} |p-a_k|
\end{equation}
and our goal will be to choose $a_i$'s that minimize the sup-norm of $D$. Figure \ref{FiveToss} below, corresponding to the case of $n=5$ tosses, illustrates the motivation for our analysis. It shows the penalty functions for the Maximum Likelihood choice (i.e. $a_k = k/n\,,\,k=0,\ldots,n$) and for the choice of the $a_i$'s we will prove to be optimal (see Section 2). Indeed, the sup-norm of the penalty function $D$ for the optimal choice is clearly smaller than the one corresponding to the Maximum Likelihood approach. Moreover, Figure \ref{FiveToss} shows that the optimal choice satisfies what we call hereafter the ``equimax'' property. This is very similar to the characterization of the set of interpolation nodes to reach the optimal Lebesgue function, as conjectured by Bernstein and Erd\H{o}s and proved forty years later by Kilgore \cite{K} and de Boor-Pinkus \cite{BP}. This similarity served as an important motivation to apply Approximation Theory type techniques for investigating this problem, and will be discussed in more detail in Section 2.
\begin{figure}[h]
\centering\includegraphics*[width=2.5in]{5tossplot.eps}
\caption{Graph of $D(a_0,...,a_n;p)$ in the case of $n=5$ tosses: Maximum Likelihood (green), optimal choice (red), and the maximum value of the penalty function for the optimal choice (black) }
\label{FiveToss}
\end{figure}
It is worth pointing out that in the Statistics literature one often prefers the use of squares instead of absolute values in \eqref{penalty}, that is, the minimization of
\begin{equation}\label{quadraticloss}
\widehat{D}(a_0,...,a_n;p):=\sum_{k=0}^n {n \choose k} p^k(1-p)^{n-k} (p-a_k)^2\,,\,p\in [0,1]\,,
\end{equation}
is considered, because of its analytical tractability and easier computations. Indeed, for the penalty function \eqref{quadraticloss} the optimal (minimax) strategy $\{a_0,\ldots,a_n\}$ is explicitly computed (see \cite{LC}):
\begin{equation}\label{quadraticoptimal}
a_k = \,\frac{1}{2}\,+\,\frac{\sqrt{n}}{1+\,\sqrt{n}}\,\left(\frac{k}{n}\,-\frac{1}{2}\right)\,,\,k=0,\ldots,n\,.
\end{equation}
Of course, this is the optimal strategy when measuring the loss using the least squares norm, but not in our ``uniform'' setting.
In Figure 2 below, we augment Figure 1 with the plot of the penalty function $D$ for the strategy \eqref{quadraticoptimal}.
\begin{figure}[h]
\centering\includegraphics*[width=3.0in]{fig1quad.eps}
\caption{Graph of $D(a_0,...,a_n;p)$ in the case of $n=5$ tosses: Maximum Likelihood (green), optimal choice (red), and the quadratic optimal choice (blue) }
\label{comparison}
\end{figure}
As it can be seen in Figure 2, the behavior of the Squared Error Minimax Estimator (hereafter, SEME) is similar, or even a bit better, towards the center of the interval, but clearly worse when we are close to the endpoints of the interval. Thus, for $n=5$, the sup-norm of the penalty function $D$ for the SEME is $0.1545$, while for our Absolute Error Minimax Estimator (AEME) is $0.131$. More generally, as mentioned above, along with the Minimax Estimators, the so-called Bayes Estimators are also often employed (see \cite[Ch. 4]{LC}). In this setting, given a loss function $R(\theta, \delta)$, some ``prior'' distribution $\Lambda$ for the parameter $\theta$ to be determined (in our case, $\theta = p$) is selected, and the estimator $\delta_{\Lambda}$ is chosen in order to minimize the weighted average risk
\begin{equation}\label{Bayesrisk}
r(\lambda, \delta) = \int \, R(\theta, \delta)\,d \Lambda (\theta),
\end{equation}
\noindent also called Bayes risk. As established in \cite[Theorem 5.1.4]{LC} and some corollaries, in certain situations a Bayes estimator is also a Minimax estimator. This connection will be used below (Section 4) when discussing the interpretation of our method within the framework of Game Theory. Thus, roughly speaking, we can say that in this paper we are dealing with a ``multi-faceted'' topic, which we investigate from the points of view of Point Estimation Theory, Approximation Theory, and Game Theory.
On the other hand, the problem of estimating the probability of a coin (more generally, the parameter of a Bernoulli distribution) from a few tosses has been often considered as a toy-model for randomization processes and for checking the effectiveness of different methods in statistical inference. In this sense, it is noteworthy that some recent papers have dealt with the problem of simulating a coin with a certain prescribed probability $f(p)$, by using a coin whose probability is actually $p$ (see \cite{Holtz} and \cite{Nacu}). Actually, this idea comes from a seminal paper by von Neumann \cite{Neumann}. In \cite{Holtz} and \cite{Nacu} the authors also show, just as we do in the current paper, the utility of techniques from Approximation Theory to solving problems from Probability.
The paper is structured as follows. In Section 2, our minimax estimation is thoroughly studied and the optimal choice is established by Theorem \ref{thm:equimax}, which represents the main result of this paper. Some computational results are included. The asymptotic distribution of the set of nodes corresponding to such optimal strategies, when the number of tosses approaches infinity, is established in Section 3. In Section 4 we discuss the problem from the Game Theory standpoint, as a non-cooperative two-player game, which is described in detail. Also, the existence of a Nash-equilibrium is established. Furthermore, in Section 5, this Nash-equilibrium is explicitly solved for the case of $n=2$ tosses.
In both Sections 4 and 5, the solution of this Nash-equilibrium problem is related to the connection between Minimax and Bayes estimators.
Finally, the last section contains some further remarks and conclusions.
\setcounter{equation}{0}
\setcounter{theorem}{0}
\section{The minimax estimation}
\label{Main}
\vskip 3mm
As it was discussed above, the optimal strategy we consider is to choose $a_0,\ldots,a_n \in [0,1]$ in order to minimize the sup-norm of the penalty function $D(a_0,\ldots,a_n;p)$. Thus, our minimax problem has a striking resemblance with the well-known problem of the optimization of the Lebesgue function in polynomial interpolation. Indeed, let us consider the polynomial interpolation of a function $f$ over a set of nodes $x_0,\ldots,x_n \in [0,1]$. By the classical Lagrange formula, we know the expression of such interpolating polynomial is: $$L_n (f;x_0,\ldots,x_n;x) = \sum_{k=0}^n\,f(x_k)\,l_k(x_0,\ldots,x_n;x)\,,$$ where $l_k(x_0,\ldots,x_n;x),\,k=0,\ldots,n,\,$ are the well-known Lagrange interpolation polynomials, and they form a basis for $\mathcal{P}_n$, the space of polynomials of degree less than or equal to $n$. Since the norm of the projection operator from $C[0,1]$, the space of all continuous functions on $[0,1]$, onto $\mathcal{P}_n$ is given by the sup-norm of the Lebesgue function
\begin{equation}\label{Lebesgue}
\Lambda(x_0,\ldots,x_n;x) = \sum_{k=0}^n\,|l_k(x_0,\ldots,x_n;x)|\,,
\end{equation}
\noindent the problem of finding optimal choices of nodes $x_0,\ldots,x_n \in [0,1]$ minimizing the sup-norm of \eqref{Lebesgue} arises in a natural way. It is well known that if the endpoints of the interval belong to the set of nodes, then the solution is unique. As for the characterization of the solution, the famous Bernstein-Erd\H{o}s conjecture asserted that for an optimal choice, the corresponding Lebesgue function \eqref{Lebesgue} must exhibit the following ``equimax'' property: If the absolute maximum of $\Lambda(x_0,\ldots,x_n;x)$ on each subinterval $[x_{i-1},x_i]$ is denoted by $\lambda_i = \lambda_i(x_0,\ldots,x_n)\,,i=1,\ldots,n$\, then we have:
\begin{equation*}\label{equimaxLeb}
\lambda_1 = \ldots = \lambda_n.
\end{equation*}
This conjecture was finally proved by Kilgore \cite{K} (see also \cite{BP}).
Now, the following result shows that the above mentioned resemblance between both minimax problems can be extended to the characterization of the optimal solutions. Let $f(p):=D(a_0,\ldots,a_n;p)$ be an optimal penalty function in the sense of minimizing the sup-norm of \eqref{penalty}. Then, the following result, which gathers some necessary conditions to be satisfied for an optimal choice $\{a_0,\ldots,a_n\}\,,$ will be useful. It will be stated without assuming that the points are ``well-ordered'', i.e. that $a_0 <a_1 <\ldots <a_n$. Although this fact may seem obvious, it does require a proof (see Theorem \ref{thm:equimax} below).
\begin{lemma}\label{lem:maxima}
Let
\begin{equation}M(f):=\{ x\in [0,1] \ : \ f(x)=\| f \|_\infty\}\end{equation}
be the set of absolute maxima of an optimal penalty function $f$. Then,
\begin{itemize}
\item [(i)] $M(f) \cap \{a_0,...,a_n\}\,=\,\emptyset$.
\item [(ii)] $M(f) \cap [0,\min \{a_i\})\,\neq\,\emptyset\,,\;M(f) \cap (\max \{a_i\},1]\,\neq\,\emptyset\,.$
\item [(iii)] $a_0 \leq 1/2 \leq a_n$ and $M(f) \cap (a_0, a_n)\,\neq\,\emptyset\,.$
\end{itemize}
\end{lemma}
\begin{proof}
The proof of part (i) easily follows from the fact that, from \eqref{penalty}, the derivative $f^\prime (p)$ has a positive jump as $p$ passes through $a_i$ and, thus, $f(p)$ cannot be increasing/decreasing as we pass through $a_i$.
\noindent As for part (ii), suppose that $M(f)\cap [0,\min \{a_i\}) = \emptyset$. Then, $\displaystyle \max_{[0,\min \{a_i\})} f(p)<\|f \|_\infty$. Since by (i), $\min \{a_i\} \not\in M(f)$, we have that there is a $\delta>0$ s.t. $\displaystyle \max_{[0,\min \{a_i\}+\delta]}f(p)<\|f \|_\infty$. But then, $\| D(a_0,a_1,\ldots,\min \{a_i\}+\delta,\ldots, a_n;p)\|_\infty<\|f\|_\infty$, which contradicts the optimality of $f(p)$. The argument that $M(f)\cap [\max \{a_i\},1)\not= \emptyset$ is similar.
\noindent To prove (iii) we first notice that $a_0\leq1/2\leq a_n$. Indeed, it is easily seen that \linebreak{$\|D(1/2,\dots,1/2;p)\|_\infty \leq 1/2$}, which implies that $\|f\|_\infty \leq 1/2$. Therefore, {$a_0=f(0)\leq \| f\|_\infty\leq 1/2$}. Similarly, $1-a_n = f(1)\leq 1/2$, and thus $a_0\leq 1/2\leq a_n$. We note that, as a consequence, we also have that $\min\{ a_i \}\leq a_0\leq 1/2\leq a_n\leq \max\{ a_i \}$.
\noindent Suppose now that $M(f)\cap[a_0, a_n]=\emptyset$. Since $a_i\not\in M(f)$, there is $\delta>0$, such that $[a_0-\delta,a_n+\delta]\cap M(f)=\emptyset$. Using a similar argument as in the proof of (ii), we have that for $\epsilon>0$ small enough and $0\leq p \leq a_0 - \delta<1/2$,
\[
\begin{split}
D(a_0-\epsilon,\dots,a_n + \epsilon,\dots,a_n ; p)&=D(a_0,\dots,a_n;p)-\epsilon\left[ (1-p)^n-p^n \right] \\ &<D(a_0,\dots,a_n;p)
\end{split}\]
\noindent We can also prove the same inequality for $a_n+\delta < p\leq 1$ from which we conclude that
\[
\|D(a_0-\epsilon,a_1,\dots, a_{n-1}, a_n + \epsilon ; p)\|_\infty<\| D(a_0,\dots,a_n;p)\|_\infty=\|f \|_\infty,
\]
\noindent which is a contradiction with the optimality of $f$.
\end{proof}
The following theorem establishes the {\em equimax property} for our minimax estimation and represents one of the two main results of this paper,
the other being Theorem \ref{LimDistr}. In addition, the ``well-ordering" of such optimal choice is also proved.
\vskip 3mm
\begin{theorem}\label{thm:equimax}
Suppose that
\begin{equation}f(p):=D(a_0, \dots , a_n;p)=D(T;p)\end{equation} is an optimal penalty function. Then the node set $T$ satisfies $a_0<a_1<\cdots <a_n $ and the {\em equimax property} holds, that is: $M(f)\cap (a_i ,a_{i+1} )\not= \emptyset$, $i=0,\dots,n-1$.
\end{theorem}
\begin{proof} First, we will prove that for an optimal penalty function $D(T;p)$ the node set $T$ is well-ordered, i.e. $a_i\leq a_{i+1}$ for all $i=0,\dots, n-1$. Indeed, suppose it is not. Then there is an index $i<n$ such that $a_i>a_{i+1}$. We will perturb the node set $T$ to obtain a penalty function with smaller norm.
\noindent Select $\epsilon>0$ small enough so that $\|f \|>f(p)$ for $p\in(a_i -\psi, a_i +\psi) \cup (a_{i+1}-\delta, a_{i+1}+\delta)$, where
\begin{equation}\label{psidelta}
\psi:=\epsilon\,(a_i+a_{i+1})/(2-a_i -a_{i+1})\,,\;\delta:=\epsilon \, (i+1)/(n-i)
\end{equation}
and $\max(\psi,\delta)<(a_i-a_{i+1})/2$. Denote by $T_{\epsilon}$ the node set obtained by perturbing the nodes $a_i$ and $a_{i+1}$ to $a_i-\psi$ and $a_{i+1}+\delta$, respectively. Then, from \eqref{penalty}, we have
\begin{equation*}
D(T_\epsilon;p)-D(T;p) =
\epsilon\, {n \choose i}p^i (1-p)^{n-i} g(p),
\end{equation*}
where
\begin{equation*}
g(p):=\left \{ \begin{array}{lc}
\displaystyle{- \frac{a_i+a_{i+1}}{2-a_i-a_{i+1}}+\frac{p}{1-p}},\quad&p\leq a_{i+1}\\
&\\
\displaystyle{ -\frac{a_i+a_{i+1}}{2-a_i-a_{i+1}}-\frac{p}{1-p}},\quad&a_{i+1}+\delta \leq p \leq a_{i}-\psi\\
&\\
\displaystyle{\frac{a_i+a_{i+1}}{2-a_i-a_{i+1}}-\frac{p}{1-p}},\quad&p\geq a_{i}\,,\\
\end{array}\right.
\end{equation*}
\noindent and using the fact that $x/(1-x)$ is an increasing function, it is easy to see that $g(p)<0$ for all $p\in [0,1]\setminus \{(a_i -\psi, a_i) \cup (a_{i+1}, a_{i+1}+\delta) \}$. Additionally, as $f(p)<\|f\|$ on $(a_i -\psi, a_i +\psi) \cup (a_{i+1}-\delta, a_{i+1}+\delta)$, by selecting $\epsilon>0$ smaller if needed, we can guarantee $\| D(T_\epsilon;p)\|<\| f \|$, a contradiction with the optimality of $f$. This implies that for optimal penalty function the node set $T$ is well-ordered, i.e. $a_0<a_1<\cdots <a_n $.
\noindent Next, we prove the {\em equimax property}. Denote the global maxima on the consecutive subintervals by $\mu_{-1}:= \max_{[0,a_{0}]} f(p)$, $\mu_j:=\max_{[a_j,a_{j+1}]} f(p)$, $j=0,2,\dots, n-1$, and $\mu_n:= \max_{[a_{n},1]} f(p)$. We want to show that
$\mu_0=\dots=\mu_{n-1}= \|f\|_{\infty}$ (we already know from Lemma \ref{lem:maxima} that $\mu_{-1}=\mu_{n}= \|f\|_{\infty}$).
\noindent By contradiction, assume that for some $i\in \{0,\ldots,n-1\}$ we have $\mu_i< \|f\|_{\infty}$. Then, as in the first part of the proof, we will construct a perturbation of the initial set of nodes, $T_{\epsilon}$, for which the corresponding penalty function has a smaller norm.
\noindent Fix $\epsilon>0$ small enough so that $\mu_i (T_\epsilon)<\|f\|_{\infty}$, where $T_\epsilon := \{a_0,\dots,a_i -\psi, a_{i+1} + \delta ,\dots, a_n \}$, with $\delta$ and $\psi$ given in \eqref{psidelta} (notice that now we are enlarging the original interval, while above it was shortened).
We will show that $\mu_k (T_\epsilon)<\mu_k (T)$ for all $k$, from which we will obtain $\|D(T_\epsilon;p \|< \|f\|_{\infty}$, a contradiction with the optimality of $f$.
\noindent Indeed, let $q \in [a_k,a_{k+1}]$ be such that $D(T_\epsilon;q)=\mu_k (T_\epsilon)$. Then
$$D(T_\epsilon;q)-D(T;q) =
\epsilon\, {n \choose i}q^i (1-q)^{n-i} h(q),$$ where
\begin{equation*}
h(q):=\left \{ \begin{array}{lc}
\displaystyle{- \frac{a_i+a_{i+1}}{2-a_i-a_{i+1}}+\frac{q}{1-q}},\quad&q<a_i\\
&\\
\displaystyle{\frac{a_i+a_{i+1}}{2-a_i-a_{i+1}}-\frac{q}{1-q}},\quad&q>a_{i+1}\,,\\
\end{array}\right.
\end{equation*}
\noindent and, thus, the fact that $x/(1-x)$ is an increasing function implies that $\mu_k (T_\epsilon)<\mu_k\leq \|f\|_{\infty}$ for all $k\not= i$. Since the choice of $\epsilon$ implies that $\mu_i (T_\epsilon)< \|f\|_{\infty}$, then we obtain the desired contradiction.
\end{proof}
\begin{example} {\bf The Optimal Penalty Function for some values of $n$.} {\rm For illustration, in Figures \ref{3tossplot} - \ref{7tossplot} we present the computational results we obtained for the cases of $n=3, 4, 6$, and $7$ tosses (the plot for $n=5$ was already shown in Figure \ref{FiveToss}). They all confirm the conclusions of Theorem \ref{thm:equimax}. In all of these figures, we plot the optimal choice for the penalty function $D(a_0,...,a_n;p)$, that is, AEME, (red), the Maximum Likelihood function (green), and the maximum value of the optimal penalty function (black).}
\end{example}
\begin{figure}
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=.7\linewidth]{3tossplot}
\caption{The case of $n=3$ tosses}
\label{3tossplot}
\end{minipage
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=.7\linewidth]{4tossplot}
\caption{The case of $n=4$ tosses}
\label{4tossplot}
\end{minipage}
\end{figure}
\begin{figure}
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=.7\linewidth]{6tossplot}
\caption{The case of $n=6$ tosses}
\label{6tossplot}
\end{minipage
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=.7\linewidth]{7tossplot}
\caption{The case of $n=7$ tosses}
\label{7tossplot}
\end{minipage}
\end{figure}
\begin{remark}\label{constantrisk}
It is relevant to point out here that when using the penalty function $\widehat{D}$ corresponding to the squared error loss function (see \eqref{quadraticloss}), it is shown in \cite[Ch. 5]{LC} that the related minimax estimation (SEME), given by \eqref{quadraticoptimal}, has a constant risk function, as shown in Figure \ref{fig:Dhat}, where the risk function for SEME (blue) is compared with those for AEME (red) and for the Maximum Likelihood (hereafter, MLE).
Since a constant function obviously satisfies the equimax property, this supports our conjecture that this equimax characterization should hold for Minimax Estimators corresponding to any convex loss function.
\begin{figure}[h]
\centering\includegraphics*[width=2.0in]{Figure7.eps}
\caption{Graph of $\widehat{D}(a_0,...,a_n;p)$ in the case of $n=5$ tosses: Maximum Likelihood (green), Absolute Error Minimax Estimator (red), and the Squared Error Minimax Estimator (blue) }
\label{fig:Dhat}
\end{figure}
\end{remark}
\vspace{.5cm}
\setcounter{equation}{0}
\setcounter{theorem}{0}
\section{The asymptotic behavior of the minimax estimations}
\label{Limit}
In this section we shall consider the limiting distribution of the minimax estimations as the number of tosses $n$ approaches infinity.
\begin{definition} For every positive integer $n$ let $A_n:=\{ a_{0n}, a_{1n}, \dots, a_{nn} \}$ denote a {\em minimax estimations set} or, simply, a {\em minimax node set}, namely
\[\|D(A_n,p)\|:=\|D(a_{0n}, \dots, a_{nn};p) \|=\,\min_{x_0,...,x_n \in [0,1]} \| D(x_0,...,x_n;\cdot)\|_\infty .\]
\end{definition}
Observe, that by Theorem \ref{thm:equimax} the node set $A_n$ is ordered. Moreover, if $B_n:=\{k/n\}_{k=0}^n$ denotes the {\em uniform node set} corresponding to the Maximum Likelihood estimation, then
\begin{equation}\label{Cond1}\|D(A_n, p)\|\leq \|D(B_n, p)\|=\mathcal{O} \left(\frac{1}{\sqrt{n}} \right)\longrightarrow 0\quad {\rm as}\quad n\to \infty.\end{equation}
Indeed, the $\mathcal{O}(1/\sqrt{n})$ estimate follows as an application of the Jensen's inequality to the convex function $f(x)=x^2$
\begin{equation*}
\begin{split}
D(B_n,p)^2&= \left( \sum_{k=0}^n {n \choose k} p^k(1-p)^{n-k} \left| p-\frac{k}{n}\right| \right)^2 \\
&\leq \sum_{k=0}^n {n \choose k} p^k(1-p)^{n-k} \left( p-\frac{k}{n} \right)^2 = \frac{p(1-p)}{n}\leq \frac{1}{4n},
\end{split}
\end{equation*}
where we used the fact that the mean of the binomial distribution is $\mu=np$ and its variance is $\sigma^2=np(1-p)$.
\vspace{.25cm}
We will prove (see Theorem \ref{LimDistr} below) that the ordering of $A_n$, established in Theorem 3.1, and equation \eqref{Cond1} yield that the limiting distribution is uniform. For this purpose let us remind the reader the definition of weak$^*$ convergence of a sequence of measures.
\begin{definition} \label{Weak*Conv} Let $\{\mu_n \}$ be a sequence of measures supported on $[0,1]$. We say that it converges weakly (or weak$^*$) to a measure $\mu$ if
\begin{equation}\label{weak*1}
\lim_{n\to\infty} \int_0^1 f(t)\, d \mu_n (t) =\int_0^1 f(t)\, d \mu (t) \quad {\rm for\ \ all}\quad f\in C[0,1],
\end{equation}
where $C[0,1]$ denotes all continuous functions on $[0,1]$, or equivalently
\begin{equation} \label{weak*2}
\lim_{n\to \infty} \mu_n ([a,b]) = \mu([a,b]) \quad {\rm for\ all} \quad [a,b]\subset [0,1].
\end{equation}
We denote this as
\[ \mu_n \stackrel{*}{\longrightarrow} \mu, \quad {\rm as}\quad n\to \infty.\]
\end{definition}
\begin{definition} \label{CountingMeasDef}
Given a finite set $K_n := \{ \alpha_{0n},\alpha_{1n}, \dots, \alpha_{nn} \}$ we call the measure
\begin{equation}
\delta_{K_n} := \frac{1}{n+1}\sum_{j=0}^n \delta_{\alpha_{jn}},
\end{equation}
a {\em normalized counting measure} of $K_n$. Here $\delta_x$ denotes the Dirac-delta measure at the point $x$.
\end{definition}
\begin{theorem}\label{LimDistr}
Suppose that the node sets $K_n := \{ \alpha_{0n},\alpha_{1n}, \dots, \alpha_{nn} \}\,\subset\,[0,1]$, $n=1,\dots,\infty$, are ordered and that
\begin{equation}\label{Cond2}\lim_{n\to \infty} \|D(K_n, p)\|= 0.\end{equation}
Then the asymptotic distribution of $K_n$, as $n$ tends to infinity, is uniform, namely
\begin{equation}\label{ConvThm} \delta_{K_n} \stackrel{*}{\longrightarrow} dx, \quad {\rm as}\quad n\to \infty,\end{equation}
where $dx$ denotes the Lebesgue measure on the interval $[0,1]$.
\end{theorem}
\begin{proof}
To prove \eqref{ConvThm} it is sufficient to establish that for all $0<p<1$ we have
\[p\leq \liminf_{n\to \infty} \frac{|K_n \cap [0,p]|}{n+1} \leq \limsup_{n\to \infty} \frac{| K_n \cap [0,p] |}{n+1} \leq p .\]
We shall prove the third inequality, the first being similar and the second being obvious.
Suppose that there is a $0<p<1$ for which it fails. Then there is an $\epsilon >0$ such that
\[\limsup_{n\to \infty} \frac{| K_n \cap [0,p] |}{n+1} > p +\epsilon.\]
This implies that there is a subsequence $\{n_i\}$ and a number $M$, such that
\begin{equation}\label{inequality}
\frac{| K_{n_i} \cap [0,p] |}{n_i +1} > p +\frac{\epsilon}{2}=:q \quad {\rm for \ all} \quad i \geq M .
\end{equation}
For every $i$ denote $k_i :=| K_{n_i} \cap [0,p] |$. Then the ordering of $K_{n_i}$ implies that $|q-\alpha_{\ell,n_i}|>\epsilon/2$ for all $\ell\leq k_i$ and hence
\begin{equation}
\begin{split}
D(K_{n_i},q)&=\sum_{\ell=0}^{n_i} {n_i \choose \ell} q^\ell(1-q)^{n_i-\ell} |q-\alpha_{\ell,n_i}| \\
&\geq \sum_{\ell=0}^{k_i} {n_i \choose \ell} q^\ell(1-q)^{n_i-\ell} |q-\alpha_{\ell,n_i}| \\
& > \frac{\epsilon}{2}\sum_{\ell=0}^{k_i} {n_i \choose \ell} q^\ell(1-q)^{n_i-\ell}>\frac{\epsilon}{4},
\end{split}
\end{equation}
where in the last inequality we apply \cite[Theorem 1]{KB}, using the fact that \eqref{inequality} yields that $k_i-qn_i>q$. This contradicts \eqref{Cond2}, which proves the theorem.
\end{proof}
As the minimax node sets $A_n$ are ordered, \eqref{Cond1} allows us to establish the following.
\begin{corollary} \label{cor:weakconv}
The limiting distribution of the minimax node sets $A_n$ is the uniform distribution $dx$ on $[0,1]$.
\end{corollary}
\begin{remark}\label{rem:universality}
Observe that Theorem \ref{LimDistr} establishes that the limiting distribution is the uniform one not just for the sequence of optimal choices, but for every sequence of ``acceptable'' strategies, in the sense that they are well ordered and the sup-norm of their corresponding penalty functions approaches zero, as $n\rightarrow \infty$. Thus, we see that all these acceptable estimators are asymptotically unbiased, in the sense that their limit distribution as $n\rightarrow \infty$, is the uniform distribution, the same as for the Maximum Likelihood estimators, which are the unique unbiased estimators for the parameter $p$.
It is also remarkable that this conclusion also holds for the sequence of Squared Error Minimax Estimators given by \eqref{quadraticoptimal}, which is easy to check.
\end{remark}
\setcounter{equation}{0}
\setcounter{theorem}{0}
\section{A problem in Game Theory}
\label{Description}
As it was said above, now we are going to see our estimation problem from the viewpoint of the Game Theory (see e.g. \cite{BG}, \cite{Fudenberg} or \cite{Osborne}). In particular, a non-cooperative, two-player, zero-sum and mixed-strategy game will be posed, as we explain below.
Indeed, we are dealing with a simple two-player game, where Player I selects a probability $p\in [0,1]$ and creates a
coin such that $P(heads)=p$. He tosses the coin $n$ times and provides the number $i\in \{0,1,...,n\}$ of heads observed to Player II. Then, based on this value,
Player II makes a guess $a_i\in [0,1]$ for the value of $p$ and he will pay a loss of $|p-a_i|$ to Player I. Obviously, the goal of Player I is to maximize this loss, while
Player II wants to minimize it.
More generally, let us assume that both players are allowed to follow what is commonly referred to as a ``mixed strategy'' in the Game Theory framework: that is, the choices of the players are not deterministic (``pure strategy''), but the available actions are selected according to certain probability distributions. Thus, let $\Omega$ denote the set of all probability distributions on the interval $[0,1]$. Suppose that when
making their decisions, Player I is allowed to choose $\mu \in \Omega$ and he picks $p$ to
follow the distribution $d\mu$, and Player II picks $x_i$ to follow his choice of $d\sigma_i$
distributions, where $\sigma_i \in \Omega, \ i=0,1,...,n$.
Therefore, the expected penalty of Player II is
\begin{equation}
E(\sigma_0,...,\sigma_n;\mu):=\int \cdots \int D(x_0,...,x_n;p) d\sigma_0(x_0) ... d\sigma_n(x_n) d\mu(p),
\end{equation}
where $D(x_0,...,x_n;p)$ is the penalty function given in \eqref{penalty}. In these terms, the goal of Player I will be to find
\begin{equation} \label{maximin} \max_{\mu\in\Omega} \min_{\sigma_0,...,\sigma_n \in\Omega} E(\sigma_0,...,\sigma_n;\mu)\,,\end{equation}
while the second player will try to get
\begin{equation} \label{minimax} \min_{\sigma_0,...,\sigma_n \in\Omega} \max_{\mu\in\Omega} E(\sigma_0,...,\sigma_n;\mu\,).\end{equation}
Using again the terminology from Game Theory, this is a ``zero-sum'' game (that is, the total gains of players minus the total losses add up to zero). Now, we are looking for the so-called mixed-strategy Nash-equilibrium. The basic notion of equilibrium in Game Theory finds its roots in the work by Cournot (1838), but was formalized in the celebrated papers by Nash, \cite{N1}-\cite{N2}, where the (now called) Nash-equilibrium was established for finite games by using Kakutani's Fixed Point Theorem \cite{Ka}. The problem for continuous games, where the players may choose their strategies in continuous sets (as in the present case), is more involved.
The following result, establishing the mixed-strategy Nash-equilibrium for our problem, is a direct application of the Glicksberg's Theorem \cite{Glicksberg}, who made use of an extension of Kakutani's Theorem to convex linear topological spaces. An alternate method of proof consists in taking finer and finer discrete approximations of our continuous game, for which the existence of a Nash-equilibrium was established, and then using the continuity of \eqref{penalty} and standard arguments of weak convergence (for more information about the successive extensions of the Nash-equilibrium problem, see for example the monographs \cite{Fudenberg} and \cite{Osborne}; there are also many papers about such extensions from the point of view of the applications to Business, see e.g. \cite{McKe} and \cite{COP}, to only cite a few).
\begin{theorem}
\begin{equation} \label{Nash generalized}
\min_{\sigma_0,...,\sigma_n \in\Omega} \max_{\mu\in\Omega} E(\sigma_0,...,\sigma_n;\mu) =
\max_{\mu\in\Omega} \min_{\sigma_0,...,\sigma_n \in\Omega} E(\sigma_0,...,\sigma_n;\mu).
\end{equation}
\end{theorem}
\vskip 3mm
For our analysis it is important to make use of the following discretization of the probability distribution setting in \eqref{Nash generalized} related to the second player's strategy.
\begin{theorem}\label{thm:discret} The minimax problem for distributions \eqref{Nash generalized} admits the following discretization
\begin{equation}\label{discret}
\begin{split}
\min_{\sigma_0,...,\sigma_n \in\Omega} \max_{\mu\in\Omega} E(\sigma_0,...,\sigma_n;\mu) & = \min_{a_0,...,a_n \in [0,1]} \max_{p\in [0,1]} D(a_0,...,a_n;p) \\
& \\
& =\min_{a_0,...,a_n \in [0,1]} \| D(a_0,...,a_n;\cdot)\|_\infty\,.
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
Let $a_i:=\int_0^1 \theta \, d\sigma_i (\theta)$, $i=0,1,\dots,n$. Since $\displaystyle \int_0^1\,|p-\theta|\,d\sigma_i(\theta)\,\geq\,|p-a_i|\,,$ we have that
$$E(\sigma_0,...,\sigma_n;\mu)\,\geq\,E(a_0,...,a_n;\mu)\,.$$
Then, by the continuity of $D(a_0,\dots,a_n;p)$, we get
$$ \max_{\mu\in\Omega} E(a_0,...,a_n;\mu)= \| D(a_0,...,a_n;\cdot)\|_\infty.$$
Hence,
\begin{equation}\label{relation1}
\min_{\sigma_0,...,\sigma_n \in\Omega} \max_{\mu\in\Omega} E(\sigma_0,...,\sigma_n;\mu)\,\geq\,\min_{a_0,...,a_n} \max_{p\in [0,1]} D(a_0,...,a_n;p)
\end{equation}
On the other hand, if for fixed points $a_0,\ldots,a_n\in [0,1]$, we take $\displaystyle \sigma_i = \delta_{a_i}\,,$ it is clear that
$E(\sigma_0,...,\sigma_n;\mu) = E(a_0,...,a_n;\mu)$ and, thus
\begin{equation}\label{relation2}
\min_{\sigma_0,...,\sigma_n \in\Omega} \max_{\mu\in\Omega} E(\sigma_0,...,\sigma_n;\mu)\,=\,\min_{a_0,...,a_n \in [0,1]} \max_{\mu\in\Omega} E(a_0,...,a_n;\mu)\,,
\end{equation}
but, for some $\mu^* \in \Omega$,
\begin{equation}
\begin{split}\label{relation3}
\max_{\mu\in\Omega} E(a_0,...,a_n;\mu) &= \int_0^1\,D(a_0,...,a_n;p)\,d\mu^*(p)\\
&\leq\,\max_{p\in [0,1]} D(a_0,...,a_n;p)
=\,\|D(a_0,...,a_n;\cdot)\|_{\infty}\,,
\end{split}
\end{equation}
and this settles the proof of \eqref{discret}.
\end{proof}
\vskip 3mm
\begin{remark}\label{doneinsection2}
The above discretization \eqref{discret} shows that the optimal strategy for Player II described in the previous section (Theorem \ref{thm:equimax}) agrees with the set of the Absolute Error Minimax Estimators (AEME), using the language of Point Estimation Theory.
\end{remark}
Therefore, our main concern now is the Optimal Strategy for the Player I. But in this sense, the arguments used in the proof of Theorem \ref{thm:discret} also have an important consequence for the Player I's strategy. Indeed, if we denote by $\mu^*$ an optimal distribution for the first player and by $\supp \mu^* \subset [0,1]$, its support, then we have
\begin{lemma}\label{lem:support}
\begin{equation}\label{supp}
\supp \mu^* \subset M(f)
\end{equation}
\end{lemma}
\begin{proof}It is enough to realize that in the proof of Theorem \ref{thm:discret}, equations \eqref{relation1}-\eqref{relation2} show that for an extremal measure $\mu^*$ (for which the Nash-equilibrium \eqref{Nash generalized} is attained), \eqref{relation3} is actually an equality and, therefore, $\supp \mu^*$ must be contained in the set where the equality
$f(p)=\|f\|_\infty = \|D(a_0,\ldots,a_n;.)\|_{\infty}\,$ is attained (with $(a_0,\ldots,a_n)$ being an optimal choice for Player II). This establishes \eqref{supp}.
\end{proof}
Theorem \ref{thm:discret} and Lemma \ref{lem:support} show that, from this point on, we may assume that both players follow strategies based on discrete distributions. Indeed, while a Player II's optimal (pure) strategy will be a choice $\{a_0,\ldots,a_n\} \subset [0,1]\,,$ an optimal (mixed) strategy for the Player I will be based on an atomic measure $\displaystyle \mu^* = \sum_{j=1}^k\,m_j\,\delta_{p_j}\,,$ where $\displaystyle \{p_j\}_{j=1}^k \subset M(f)$ and $\sum_{j=1}^k\,m_j = 1\,.$
\begin{example}{The case of $n=1$ toss.{\rm }}
\label{1 toss}
\vskip 2mm
\noindent {\rm Because of the simplicity and the symmetry of the problem, the method to find the strategies satisfying the Nash-equilibrium (\ref{Nash generalized}) can be carried out easily in this simple case by using Lemmas \ref{lem:maxima} and \ref{lem:support} above. Therefore, we skip the details.
\noindent The optimal discrete strategy $\mu$ of Player I is the following: choose the atomic measure $\displaystyle \mu = \sum_{k=0}^2\,m_i\,\delta_{p_i}$, with $p_0=0, p_1=0.5, p_2=1$ and
the corresponding weights given by $m_0=0.25, m_1=0.5, m_2=0.25$.
Further, for $a_0\in [0,0.5], a_1\in [0.5,1]$, we have $E(a_0,a_1;\mu)=0.25$ and for any $a_0,a_1\in [0,1]$ we have $0.25\le E(a_0,a_1;\mu)$.
Thus, for any $\sigma_0,\sigma_1\in $ we have $0.25\le E(\sigma_0,\sigma_1;\mu)$.
\noindent The optimal discrete strategy of Player II is the following: $\sigma_0$ is the unit point mass at $a_0=0.25$, and $\sigma_1$ is the unit point mass at $a_1=0.75$.
The graph of $f(p)=D(0.25,0.75;p)=(1-p)|p-0.25|+p|p-0.75|$ is shown in Figure \ref{1tossplot}.
Since for all $p\in[0,1]$ we have $f(p)\le 0.25$, we conclude that for any $\mu\in\Omega$, $E(\sigma_0,\sigma_1;\mu)\le 0.25$.
So the Nash-equilibrium is established, and if both players follow the outlined strategies, then Player II pays $E(a_0,a_1;\mu)=0.25$ dollars to Player I.}
\begin{figure}[h]
\centering\includegraphics*[width=3.0in]{1tossplot.eps}
\caption{Graph of $f(p)$ in the case of $n=1$ toss.}
\label{1tossplot}
\end{figure}
\end{example}\
\begin{remark}\label{rem:leastfav}
Going back to the Point Estimation Theory approach, the results in previous Theorem 4.2 and Lemma 4.4 admit an interpretation in terms of the connection between Bayes and Minimax estimators, as
mentioned in the Introduction. Indeed, Theorem 5.1.4 and especially Corollary 5.1.5 in \cite{LC} establish sufficient conditions to ensure that a Bayes estimator is also a Minimax one, namely, that the average risk (or penalty) of the Bayes estimator $\delta_{\Lambda}$ for a certain prior distribution $\Lambda$ (see \eqref{Bayesrisk}) agrees with the value of the maximum of that risk and, hence, that for such distribution $\Lambda$, the risk function must be constant. If this is the case, the prior distribution $\Lambda$ is said to be a \emph{least favorable} one. In this sense, our results are a sort of converse of the ones in \cite{LC}.
When the squared error loss function is used, one could see in Figure \ref{fig:Dhat} that the risk function for the corresponding Minimax estimator is constant throughout the interval $[0,1]$. In this case, it is proven in \cite[Ex. 5.1.7]{LC} that this Minimax estimator is indeed a Bayes estimator with respect to a continuous distribution supported in $[0,1]$, namely the Beta distribution $\displaystyle \mathcal{B} \left(\frac{\sqrt{n}}{2}, \frac{\sqrt{n}}{2}\right)$. Therefore, this last distribution plays the role of a least favorable one and its corresponding Bayes estimator is proven to be unique, hence, by \cite[Th. 5.1.4]{LC}, it is also unique as Minimax estimator.
Taking into account our previous results, this least favorable distribution would represent the optimal strategy for Player I in this case.
The above connection for our absolute error loss function is more involved and it will be discussed in the next section, where the Nash-equilibrium for the case of $n=2$ tosses is thoroughly analyzed.
\end{remark}
\setcounter{equation}{0}
\setcounter{theorem}{0}
\section{A constructive proof of the Nash-equilibrium for the case of $n=2$ tosses}
\label{Main1}
Now, we are concerned with the existence and uniqueness of a strategy pair solving the Nash equilibrium \eqref{Nash generalized} in the case of $n=2$ tosses. Our method will be based on previous Theorem \ref{thm:equimax} and Lemmas \ref{lem:maxima}-\ref{lem:support}.
Recall that the strategy of Player II is to minimize the maximum of the penalty function, namely determine optimal outcomes $\{a_0^*, a_1^*, a_2^* \}$ defining an optimal penalty function
$f(p):=D(a_0^*,a_1^*,a_2^*;p)$ such that
\begin{equation}
\label{PlayII}
\min_{\{a_0, a_1, a_2 \}\subset [0,1]} \max_p D(a_0, a_1, a_2;p)=\min_{\{a_0, a_1, a_2 \}\subset [0,1]} \| D(a_0,a_1,a_2;p) \|_\infty= \|f\|_\infty\,.
\end{equation}
\noindent That such an $f$ exists follows easily by a compactness argument.
The strategy of Player I is to find a probability measure $d\mu^*(p)$, supported on $[0,1]$, that maximizes the expected penalty no matter what the choice of Player II is, i.e. determine
\begin{equation}
\label{PlayI}
\mathcal{F}:=\max_{\mu} \min_{\{a_0, a_1, a_2 \} \subset [0,1]}\int_0^1 D(a_0, a_1, a_2; p) d\mu (p).
\end{equation}
Clearly, for any $\mu \in \Omega$,
\[\int_0^1 D(a_0,a_1,a_2;p) d\mu (p) \leq \|D(a_0,a_1,a_2;p)\|_\infty ,\]
so,
\[ \min_{\{a_0, a_1, a_2 \} \subset [0,1] }\int_0^1 D(a_0,a_1,a_2;p) d\mu (p)\leq \|f\|_\infty ,\]
\noindent which implies, after taking the $\max$ over all $\mu$, that $\mathcal{F}\leq \|f\|_\infty$. Then, our goal in this section is to find an optimal strategy pair
$\{a_0^*, a_1^*, a_2^* \}\subset [0,1],\,\mu^* \in \Omega,$ for which the Nash-equilibrium, $\mathcal{F} = \|f\|_\infty$, is uniquely reached.
\vskip 1mm
The main result in this section is stated as follows.
\begin{theorem}\label{thm:2tosses}
The Nash-equilibrium is reached if the players use the following strategies:
\begin{itemize}
\item [\textbf{Player I:}] Choose $p$ according to the distribution $\displaystyle \mu = \sum_{i=0}^3\,m_i\,\delta_{p_i}\,,$ where
\begin{equation}\label{p1}
p_1 = \frac{1}{3} \Big( 1+ \sqrt[3]{1+3\sqrt{57}} - \frac{8}{\sqrt[3]{1+3\sqrt{57}}} \Big) \approx 0.3611
\end{equation}
is the unique real root of the polynomial $x^3-x^2+3x-1$, and
\begin{equation}p_2 = 1-p_1 \approx 0.6389\,,\,p_0 = 0, p_3 = 1\,.
\end{equation}
The weights $m_i$ are given by:
\begin{equation}\label{weights}
\begin{split}
m_1&=\frac{0.5}{p_1^2+(1-p_1)^2+1} \approx 0.325 \\
m_0 &= 0.5 - m_1 \approx 0.175\,,\;m_2 = m_1\,,\;m_3 = m_0\,.
\end{split}
\end{equation}
\item [\textbf{Player II:}] Choose the following values $a_i$
\begin{equation}\label{ai}
a_0 = \frac{2p_1(1-p_1)^2}{p_1^2+(1-p_1)^2+1} \approx 0.1916\,,\;a_2 = 1- a_0 \approx 0.8084\,,\; a_1 = 0.5\,.
\end{equation}
\end{itemize}
Furthermore, the above pair of strategies is unique in the following sense: if the choice of distributions $\{\sigma_0, \sigma_1, \sigma_2,\mu \}$ satisfies the Nash-equilibrium \eqref{Nash generalized}, then $E(\sigma_i)=a_i, i=0,1,2$, and $\displaystyle \mu= \sum_{i=0}^3 m_i \delta_{p_i}$, where $a_i, m_i$ and $p_i$ are as above.
\end{theorem}
The graph of the optimal penalty function given in Theorem \ref{thm:2tosses}, $f(p)=D(a_0,a_1,a_2;p)$, is shown in Figure 8. Observe that the interlacing property $p_0 < a_0 < p_1 < a_1 < p_2 < a_2 < p_3$ holds.
For the proof of Theorem \ref{thm:2tosses} two technical lemmas are needed.
\begin{lemma} \label{lem:bounds}
The optimal choice $\{a_0, a_1, a_2 \}$ satisfies $0<a_0<0.2<0.4<a_1<0.6<0.8<a_2<1$. In addition, $a_1-a_0 < 0.4$ and $a_2-a_1 < 0.4$.
\end{lemma}
\begin{proof} It is easy to check that for the symmetric choice $\{a_0,a_1,a_2\}$, where $a_0 = .195\,,\,a_1 = .5\,,\,a_2 = 1-a_0 = .805$, we have that $\|D(a_0,a_1,a_2;p)\|_\infty=0.195\,<\,0.2$. Therefore, $\| f \|_\infty <0.2$. Since $f(0)=a_0$ and $f(1)=1-a_2$ we immediately obtain that $a_0<0.2$ and $0.8<a_2$.
On the other hand, if there exists $p\in[0,1]$ such that $|p-a_i | \geq 0.2$, $i=0,1,2$, then $\|f \|_\infty \geq \min_i |p-a_i | \geq 0.2$, which is a contradiction. Thus, $\min_i |p-a_i |<0.2 $ holds for all $p\in[0,1]$. This implies that $0<a_0$ and $a_2<1$ (otherwise, the Dirichlet Pigeonhole Principle implies there exists a $p$ such that $\min_i |p-a_i |\geq 0.2$). The fact that $\min_i |p-a_i |<0.2\,,\, p\in[0,1]\,,$ also implies that $a_1-a_0 < 0.4$ and $a_2-a_1 < 0.4$.
\noindent To derive that $0.4<a_1<0.6$ we need only use the values $p=0.4$ and $p=0.6$ to determine that from $\min_i |0.4-a_i |<0.2$ we must have $|0.4-a_1|<0.2$, or $a_1<0.6$, and from $\min_i |0.6-a_i |<0.2$ we must have $|0.6-a_1|<0.2$, or $0.4<a_1$.
\end{proof}
\begin{remark}\label{symchoice}
While the ``test''values used in the above proof might seem, at first glance, quite arbitrary, their usefulness can be easily shown by a simple numerical experimentation. In particular, for such a symmetric choice with $0<a_0\,,\,a_1 = .5$ and $a_2 = 1-a_0 < 1$, we see that the restrictions of the penalty function $f$ to the ``end'' subintervals $[0,a_0]$ and $[1-a_0,1]$ are straight lines whose respective maximum values (attained at the endpoints $0$ and $1$) are given by $a_0$, and the restrictions to the ``central'' intervals $[a_0,.5]$ and $[.5,1-a_0]$, are concave functions. This fact will be used again in the proof of Theorem \ref{thm:2tosses} below.
\end{remark}
Now, as a consequence of Lemma \ref{lem:bounds} we have the following result, which completes the previous Lemma \ref{lem:maxima}
\begin{lemma}\label{lem:01}
For an optimal choice $\{a_0, a_1, a_2 \}\,,$ we have that $\{0,1\} \subset M(f)\,.$
\end{lemma}
\begin{proof} Indeed, consider the restriction of $f(p)$ to $[0,a_0]$. Then,
$$f(p)=(a_0-2a_1+a_2)p^2 -(1+2a_0-2a_1)p+a_0.$$
Since the global maximum is attained on $[0,a_0]$, and it is not at $a_0$,
the only possibility of it not being at $p=0$ is when $a_0-2a_1+a_2<0$ and the $x$-coordinate of the parabola's vertex $(1/2+a_0-a_1)/( a_0-2a_1+a_2)>0$, or $1/2+a_0-a_1<0$.
This implies that $a_1-a_0>1/2$, which, as shown in Lemma \ref{lem:bounds}, is impossible if $f$ is the optimal solution of \eqref{PlayII}. Therefore, we derive that $0\in M(f)$.
In a similar fashion, one gets $1\in M(f)$.
\end{proof}
\vspace{3mm}
\noindent {\bf Proof of Theorem \ref{thm:2tosses}}. From Lemmas 2.3-2.4, Theorem 3.1 and Lemma 4.4, we already know that
\[ d\mu(p)=\beta_0\delta_0+\beta_1\delta_{p}+\beta_2 \delta_{q} +\beta_3 \delta_1 , \]
for some $\beta_i\geq 0$, $\beta_0+\beta_1+\beta_2+\beta_3=1$, and $p \in (a_0,a_1)$, $q\in (a_1,a_2)$. Since we are in the case of Nash-equilibrium, the quantity
\begin{eqnarray*} &&E(a_0,a_1, a_2;\mu) = \mathcal{E}(a_0,a_1, a_2; p,q):=\int_0^1 f(r)\, d\mu(r)\\
&& \quad =\beta_0 f(0)+\beta_1 f(p)+\beta_2 f(q)+\beta_3 f(1) \\
&& \quad =
a_0\left[ \beta_0 -\beta_1(1-p)^2-\beta_2(1-q)^2\right] +a_1\left[2\beta_1 p(1-p)-2\beta_2 q(1-q)\right]\\
&& \quad \ +a_2\left[ \beta_1 p^2+\beta_2 q^2 -\beta_3 \right] +\beta_3-\beta_1 p+\beta_2 q +2\beta_1 p(1-p)^2-2\beta_2 q^3.
\end{eqnarray*}
is a global minimum (w.r.t. $\{a_0, a_1, a_2\} \in (0,1)$), which implies that \[\frac{\partial \mathcal{E}}{\partial a_i}=0. \quad i=0,1,2.\]
Therefore, the coefficients in front of $a_0, a_1$, and $a_2$ vanish for the given choice of $p,\ q$, and $\beta_i$, $i=0,1,2,3$. Hence, the optimization problem \eqref{PlayII} becomes a constrained minimization problem
\begin{equation}\label{CMP}
\begin{split}
& Maximize \quad \beta_3-\beta_1 p+\beta_2 q +2\beta_1 p(1-p)^2-2\beta_2 q^3\\
& \\
&Subject \ to\ \left \{ \begin{array}{rcrcrcrcr}
\beta_0&+&\beta_1&+&\beta_2&+&\beta_3&=&1\\
\beta_0&-&(1-p)^2\beta_1&-&(1-q)^2\beta_2&\ &\ &=&0\\
\ &\ &2p(1-p)\beta_1& -&2 q(1-q)\beta_2 &\ &\ &=&0\\
\ &\ &p^2 \beta_1&+&q^2 \beta_2& -&\beta_3&=&0\\
\ & \ &p,q\in [0,1], &\ &\beta_i \geq 0.
\end{array}\right.
\end{split}
\end{equation}
Eliminating $\beta_0$ and $\beta_3$ we reduce \eqref{CMP} to
\begin{equation}\label{CMP2}
\begin{split}
& Maximize \quad \beta_1 p(1-p)(1-2p)+\beta_2 q(1-q)(1+2q)\\
& \\
&Subject \ to\ \left \{ \begin{array}{rcrcr}
2[1-p(1-p)]\beta_1&+&2[1-q(1-q)]\beta_2&=&1\\
p(1-p)\beta_1& -& q(1-q)\beta_2 &=&0\\
p,q\in [0,1], \beta_i \geq 0.
\end{array}\right.
\end{split}
\end{equation}
Further, eliminating $\beta_2$ from \eqref{CMP2} we derive
\begin{equation}\label{CMP3}
\begin{split}
& Maximize \quad 2\beta_1 p(1-p)(1-p+q)\\
& \\
&Subject \ to \quad
2\beta_1 p(1-p) \left[ \frac{1}{p(1-p)}+\frac{1}{q(1-q)}-2 \right]=1, \quad p,q,\beta_1 \in [0,1].
\end{split}
\end{equation}
Substituting $2\beta_1 p(1-p)$ and denoting $x=1-p$, $y=q$, we obtain the minimization problem
\begin{equation}\label{CMP4} Minimize \quad \frac{\displaystyle{\frac{1}{x(1-x)}+\frac{1}{y(1-y)}-2}}{x+y},\quad x,y \in [0,1].\end{equation}
\noindent Since $1/[x(1-x)]$ is a convex function, it is easy to see that if $x+y$ is kept constant, the minimum in \eqref{CMP4} is attained when $x=y$, which implies that $q=1-p$ (and subsequently $\beta_3=\beta_0$ and $\beta_2=\beta_1$) is a necessary condition for a Nash-equilibrium selection. Moreover, assuming $x=y$, we obtain that $x$ must minimize the function
\[ g(x)=\frac{1-x+x^2}{x^2(1-x)}=\frac{1}{x^2}+\frac{1}{1-x}.\]
Differentiating, we see that $g'(x)=0$ has only one solution in $[0,1]$, which satisfies $x^3=2(1-x)^2$, or $(1-p)^3=2p^2$, so $p=p_1$, where $p_1$ is given in \eqref{p1}. It now follows easily that $\beta_0=m_0$ and $\beta_1=m_1\,,$ where $m_0$ and $m_1$ are given in \eqref{weights}.
Therefore, we have proven that if $p = p_1$ and $q = 1-p_1$, with $0 < a_0 < p_1 < a_1 < q=1-p_1 < a_2 < 1$ as above, then $E(a_0,a_1, a_2;\mu)$ does not depend on $a_0, a_1, a_2$, or,
$$E(a_0,a_1, a_2;\mu) \equiv E(p_1) = \frac{2p_1(1-p_1)^2}{p_1^2+(1-p_1)^2+1} \approx 0.1916\,.$$
\noindent Keeping in mind the results of Theorem \ref{thm:equimax}, it follows that the optimal strategy for Player II, is given by: $a_0= E(p_1) \approx .1916...$, $a_1=0.5, a_2=1-a_0 \approx .8084$.
The graph of $f(p)=D(a_0,a_1,a_2;p)$ is shown in Figure \ref{2tossplot}.
\begin{figure}[h]
\centering\includegraphics*[width=3.0in]{2tossplot.eps}
\caption{Graph of $f(p)$ for $n=2$ tosses}
\label{2tossplot}
\end{figure}
\noindent Indeed, $f(0)=f(1)=a_0$, and direct but cumbersome calculations show that $f(p_1)=a_0$ and $f'(p_1)=0$, and by symmetry $f(1-p_1)=a_0$, $f'(1-p_1)=0$.
One has $f'(0)=-f'(1) \approx -0.38$ and $f''(p_1)=f''(1-p_1) \approx -4.43$. Since $f$ is a continuous piecewise-polynomial function whose restrictions to $[0,a_0]$ and to $[a_n,1]$ are straight lines, while the restrictions to $[a_0,a_1]$ and $[a_1,a_2]$ are concave functions, it follows that the set of the absolute maxima of $f$, $M(f)$, cannot contain more than $4$ points, which are precisely $0, p_1, 1-p_1, 1$. Thus, $\|f\|_{\infty} = a_0 = E(p_1) = 0.1916...$.
\noindent Therefore, the Nash-equilibrium (\ref{Nash generalized}) is established. $\quad\Box $
\begin{remark}\label{rem:bilingual}
The proof of Theorem \ref{thm:2tosses}, as well as Example 4.5 (the one-toss case), show that the Nash strategy pair for Players II and I may be seen as the pair consisting of, the set of absolute error minimax estimators, and the least favorable prior distribution for which the former become Bayes estimators. However, unlike the case when the squared error loss function is used (see Remark 4.6), in the above proof of Theorem \ref{thm:2tosses}, as well as in the discussion of Example 4.5, it was shown that for the current absolute error loss function, given the least favorable distribution $\mu$ there is no unique corresponding Bayes estimator. Indeed, we have seen that for $n=1$ or $2$ tosses and for $\mu = \sum_{k=0}^{n+1}\,m_{i,n}\,p_{i,n}$ given in Example 4.5 and Theorem \ref{thm:2tosses}, respectively, every configuration $\{a_0,\ldots,a_n\}$ satisfying the interlacing property $0 = p_0 \leq a_0 \leq p_1 \leq \ldots \leq a_n \leq p_{n+1} = 1$ is a Bayes estimator for $\mu$. Does this hold for $n>2$? And, furthermore, in spite of the lack of uniqueness of the Bayes estimator, is our Minimax estimator unique? These issues will be taken up again in the next section.
\end{remark}
\setcounter{equation}{0}
\setcounter{theorem}{0}
\section{Conclusions and further remarks}
\label{Conclusions}
In this paper, minimax type techniques from Approximation Theory have been applied to a classical problem in Probability: estimating the probability of a biased coin after a few tosses. We used the Minimax Estimation with absolute error loss function to solve the problem, characterizing the optimal solution and studying the asymptotics of the optimal estimators as the number of tosses tend to infinity. In addition, the method employed has been described within the framework of a non-cooperative (in particular, zero-sum) two-player game, where both players are allowed to make use of mixed strategies which, in turn, is closely related to the connection between Minimax and Bayes estimators in Point Estimation Theory. Our main results are Theorem \ref{thm:equimax}, where the optimal strategy choice for the second player is characterized by means of a property with a striking resemblance to a well-known problem in Polynomial Interpolation, and Theorem \ref{LimDistr}, with its Corollary \ref{cor:weakconv}, where the uniform limiting distribution of the optimal choices is established. Likewise, the result of Theorem \ref{thm:2tosses}, where the Nash-equilibrium for the case of $n=2$ tosses is uniquely solved, is also remarkable.
In view of the results obtained, some further remarks and questions are noteworthy and will be subject of further research.
\begin{itemize}
\item We have shown that the sup-norm of the penalty function for the optimal choice is clearly smaller than the one corresponding to the Maximum Likelihood (MLE) choice (i.e. $a_k = k/n\,,\,k=0,\ldots,n\,,$ see Figure \ref{FiveToss}), especially for small values of $n$. But it is interesting to observe that for a slight modification of the MLE, namely taking $a_k = (k+1)/(n+2)\,,\,k=0,\ldots,n$, the results are much closer to those corresponding to the optimal choice (see Figure \ref{FiveTossmod}, where the initial Figure \ref{FiveToss} has been augmented by adding the graph of the penalty function for the modified MLE). Indeed, this modified MLE (MMLE) is much easier to compute than the optimal choice (especially for big values of $n$) and seems to provide near optimal results. In other words, if we replace the optimal selection by this MMLE, a {\em near Nash-equilibrium} arises (also commonly referred to as an $\varepsilon$-Nash equilibrium). This is a well-known problem in Game Theory: since the optimal strategies are often difficult to compute, it is customary to look for easily computable approximations for which the deviation from equality in the ``pure'' Nash-equilibrium \eqref{Nash generalized} is small enough. From a Game Theory standpoint, the difference between pure and near (or $\varepsilon$-) Nash-equilibria consists in the fact that while in the "pure" setting no player has motivation to modify its strategy (corresponding to the optimal strategy pair), in the near equilibrium setting there exists a small incentive to do it. Of course, our version of the near Nash-equilibrium using the MMLE only deals with the second player's strategy (for more information about near Nash-equilibrium, see e.g. \cite{nearNash}).
This difference between ``pure'' and near Nash-equilibrium also has a counterpart regarding the optimality of the nodes in the sense of the Lebesgue constant in the context of polynomial interpolation. Indeed, if the interval $[-1,1]$ is considered, it is well known that the so-called Extended Chebyshev nodes (that is, the zeros of the Chebyshev polynomial $T_{n+1}$ adjusted in such a way that the first and last zero fall on the endpoints of the interval) provide a near optimal choice of interpolating nodes (see \cite{deBoorH}--\cite{Brutman}; see also \cite{Smith} for a deeper discussion on near optimal choices of nodes).
\begin{figure}[h]
\centering\includegraphics*[width=3.0in]{5tossplot-mod.eps}
\caption{Graph of $D(a_0,...,a_n;p)$ in the case of $n=5$ tosses: MLE (green), MMLE (blue), optimal choice (AEME), (red), and the maximum value of the penalty function for the optimal choice (black) }
\label{FiveTossmod}
\end{figure}
\item However, Theorem \ref{LimDistr} shows that, as $n$ increases, the optimal choice, as well as the MLE and MMLE ones, or any other ``acceptable'' choice (such as, for instance, the SEME, given by (1.3)), all approach the same limiting distribution, in light of Remark \ref{rem:universality}. In other words, the advantage of using the optimal strategy over the MLE is worthwhile only for a small, or not very large, number of tosses.
\item In the solution of the Nash-equilibrium for the case of $n=2$ tosses we found that the interlacing property $0=p_0<a_0<p_1<a_1<p_2<a_2<p_3=1$ was satisfied. But what about the asymptotic distribution of the atomic measures $\displaystyle \sum_{j=0}^{n+1}\,m_{j,n}\,\delta_{p_{j,n}}\,$, with $\displaystyle \sum_{j=0}^{n+1}\,m_{j,n} = 1\,,$ which give the optimal strategy of the first player for each $n$ (provided they exist)? Of course, the same question may be posed using the language of Point Estimation Theory, regarding the pair formed by the Minimax estimator and its corresponding least favorable prior distribution.
\item As for the equimax property for the optimal choice of Player II (Minimax Estimation) given in Theorem \ref{thm:equimax}, it is necessary to point out a couple of pending questions about the uniqueness of the solution. First, is there a unique optimal configuration $\{a_0,\ldots,a_n\}$ for each $n$? And secondly, Theorem \ref{thm:equimax} proved that the intersection of the set of maxima $M(f)$ with each subinterval $(a_i,a_{i+1})$ corresponding to an optimal choice is nonempty. But, is there just a single absolute maximum on each subinterval? It was only established for the case of $n=2$ tosses and numerical results seem to confirm that it also holds for larger values of $n$.
\section*{Acknowledgement} The proof of Theorem \ref{LimDistr} is inspired from an argument presented by Vilmos Totik to the third author, for which the authors are very grateful. In addition, the authors wish to thank Prof. Jos\'{e} Luis Fern\'{a}ndez P\'{e}rez (Universidad Aut\'{o}noma, Madrid) for helping us to suitably place the present work within the
Point Estimation Theory literature.
\end{itemize}
|
\section*{Conclusions}
\label{sect:Conclusions}
The main conclusion from the detailed analysis of census tracts is that they do not provide a good basis on which to build a licensing framework. The FCC rightly recognizes that smaller license areas provide the potential for unlocking efficient sharing of spectrum in 3.5GHz band using small cell deployments. But as our results have shown, even under propagation conditions that are favorable, many unit license areas are rendered unusable and that potential is not unlocked.
The problem of spectrum waste over the license area, reflected through the percentage of area loss and number of people precluded from spectrum use points to the fact that the boundaries of census tracts are not appropriate boundaries for a license area in a geographic licensing scheme envisioned for 3.5 GHz band. Not only that they mismatch with the spectrum propagation of small cell deployments, the main deployments that could put 3.5 GHz spectrum to more efficient use, but also finding the CBRS-allowed remains a problem with no solution even if the boundary limits are less stringent. In highly dense urban areas, census tracts are so small that it is not possible to form the CBRS-allowed area at all to fit the diverse spectrum usage so the interference between adjacent tracts is reduced. As we have shown, the smaller the census tracts are, more spectrum is wasted over the area.
Getting the licensing framework right is essential. The landscape in spectrum management has changed, in that it is accepted that some form of spectrum sharing will be part of the future. However it is still an open question as to which approaches will succeed. The 3.5GHz SAS-based sharing promises much and may be hampered in delivering if census tracts are at the core of its licensing framework.
Therefore we are pointing out that it is worth re-considering the licensing scheme, because the size and the shape of the license area matters. We propose the following geographic licensing models to take into consideration towards building an effective licensing framework that can fully support the SAS potential. Geographic area of deployment calculated based on the predicted actual spectrum usage should be the base area unit of the licensing scheme, i.e. area whose size is not fixed nor created for purposes other than spectrum usage. Describing the actual spectrum usage of an operator's network is a research question we plan to address in our future work, in order to investigate how different network deployments need to be so that they can be allocated in the same license area and even in the same spectrum chunk.
Furthermore, the currently existing geographic licensing models for spectrum sharing are implemented on the borders of the U.S. with Canada and Mexico in 800MHz band. These models are built on the basis of geographic sharing while giving assurance that spectrum is assigned in a fair manner to adjacent operators, whose edge users will get the service under these conditions. Defined sharing and protection zones reduce the need for grey space non-allowed area and there is always someone using the frequencies in time and space, showing in a way that the operator networks can overlap and still avail from spectrum use.
\section*{Acknowledgement}
This work is supported by the Seventh Framework Programme for Research of the European Commission under grant ADEL-619647 and the Science Foundation Ireland under grant CONNECT 13/RC/2077. We would like to thank Dr. Tim Forde for his insightful comments and valuable feedback which improved the quality of this paper.
\section{Census Tracts Study}
\label{sect:study}
In this section, we describe the study conducted and the approach to the problem. First, we introduce the selection criteria for the census tract areas. Methods of collecting and analysing data for the study are described in brief, followed by the propagation analysis and the method for computing the \textit{CBRS-allowed area}, i.e. the portion of the area of a census tract in which CBSD deployment is allowed. To evaluate census tracts-based licensing scheme, we introduce two efficiency metrics, which take into account two main census tracts characteristics: area and their population.
\subsection{Census tracts datasets}
\label{sect:data}
Traffic profile of metropolitan areas (particularly highly dense, urban areas) consists mostly of indoor residential and commercial traffic, but also outdoor WiFi traffic. The rules on technical operations in 3.5GHz band describe the PA and GAA spectrum usage corresponding to such traffic profiles, with the small cell deployments for indoor and outdoor categories of CBSDs. Based on their strategic importance and census statistics data, we selected 3 sample cities. The subjects of our study are: (1) New York City (Manhattan Borough), NY, (2) Washington, DC and (3) San Francisco, CA.
The area of Manhattan, as an example of highly dense and urban area, is divided into 288 very small census tracts, with an extremely high population density of 69,467 people per square mile. The population density in urban area profiles of San Francisco and Washington DC is much lower (17,179 and 9,856 population per square mile, respectively), but census tracts areas are still small.
The data for this study comes from the official U.S. census databases\footnote[6]{http://www.census.gov/geo/maps-data/index.html}.
Census tracts are presented in the form of cartographic boundary files, given in .kml and .shp format\footnote[7]{We used .kml files, which are based on WGS84 geodetic system. World Geodetic system 1984 is the current geodetic system being used by GPS and U.S. DoD to satisfy mapping, charting and geodetic requirements. It is geocentric and globally consistent within 1m.}. To apply Cartesian geometry which is required by our study, we used the UTM (Universal Transverse Mercator) projection maps to convert geodetic coordinates of census tracts to Cartesian friendly, UTM coordinates (WGS84).
\subsection{Propagation Analysis}
\label{sec:propagation}
\begin{table}[!t]
\centering
\caption{Propagation Analysis - parameters}
\label{tab:study}
\begin{tabular}{|| c || c || c || c || c || p{5cm} |}
\hline
\bfseries Deployment & \bfseries Building Loss & \bfseries Path Loss & \bfseries Set-back \\ \hline
Outdoor & - & 110 dB & 2.1 km \\ \hline
Indoor residential & 10 dB (wood) & 100 dB & 663 m \\ \hline
Indoor commercial & 20 dB (metal) & 90 dB & 210 m \\ \hline
\end{tabular}
\end{table}
Being above the 3GHz threshold up to which mobile cellular spectrum usage is ideal, the 3.5 GHz\footnote[8]{Refering to the 3.5GHz band means the 150 MHz of spectrum between 3550-3700 MHz.} band is not ideal for exclusive, licensed and commercial mobile broadband usage. The core advantage of the band is a huge potential for geographic sharing. Having low powered deployments with the limited propagation range of the band will allow more users to operate in closer proximity\footnote[9]{The limited propagation of 3.5 GHz needs a technology that requires less range than macrocell networks to meet users demand. Propagation characteristics of 3.5GHz make the signal decay faster, which is why low-powered and dense small cell deployments empower this band. Small cells bring higher spatial and spectral reuse with its applications, since their larger density reduces the risk of interference in geography and spectrally which results in increased frequency reuse and network capacity. In addition, signal propagation of the band still allows flexible topologies, appropriate for non-line-of-site use.}.
Geographic-based licensing in this band is on the level of a census tract. Therefore, the power limits on a license boundary are proposed by the FCC in the rules for technical operation in 3.5GHz \cite{RO}. The rules specify the conditions in which neighboring CBRS deployments should operate to reduce the risk of interference due to tier-interactions.
Anywhere along PA service area boundaries between different CBRS users, a signal strength level limit of -80 dBm, measured by a 0 dBi isotropic antenna in a 10 MHz bandwidth is proposed. The path loss amount required to meet the boundary limit is the difference of: allowed EIRP (30 dBm/10MHz), the boundary limit and building loss where existing, depending on the type of deployments. The distance, calculated from path loss free space model formula, for frequency of 3.6GHz (the frequency range is 3550-3700 MHz) and antenna gains of 0 dBi is the required set-back distance under different building losses. We summarise in Table \ref{tab:study} the path loss for different cases and the corresponding set-back distances.
In the following section, we use this distance to set the constraint on a boundary of the license area in order to compute the portion of the area that would be off-limits to CBSDs deployments, i.e. area in which CBSDs cannot meet the adopted boundary signal strength limit. As summarised in Table \ref{tab:study}, CBSD deployments considered are indoor residential and indoor commercial and we used full-power CBSDs, Category A.
\subsection{Computing the CBRS-allowed area}
\label{sec:model}
Census tracts polygons are \textit{non-convex polygons}, irregular in shape and in size. In general, they are layed out as sets of multiple polygons, each of them given as a set of points in the geodetic coordinate system\footnote[10]{To obtain the list of points, coordinates conversion is applied so that the boundary constraint in optimisation model could be set for Euclidan distance (distance between geodetic coordinates is Haversine distance).}.
The problem of computing the \textit{CBRS-allowed area} is formulated as a collection of quadratic programming problems with quadratic constraints. The assumption underlying the model is that the adjacent census tracts are allocated to different operators. This case can be interpreted as a worst case scenario, but extremely important one to look at since one of the core ideas of 3.5GHz framework is to allow operators to operate locally. For each of the original polygons we compute the largest polygon (if exists) whose distance from the boundary of the original one is at least the set-back distance determined by propagation analysis, based on FCC technical requirements for CBRS devices.
Let us denote by $\{P_1,...,P_N\}$ the set of points in UTM that describe a polygon. For each point $P_i$, we determine a point $P_i^{*}$ such that the distance between $P_i$ and $P_i^{*}$ is minimized, whilst a minimum distance $d$ (which is the set-back distance) between $P_i^{*}$ and every $P_j$ $\in\{P_1, P_2...,P_N\}$ is guaranteed, as formulated in (\ref{problem}). The set of points $\{P_1^{*}, P_2^{*}, ...,P_N^{*}\}$ defines the new polygon.
\begin{equation}
\label{problem}
\begin{aligned}
& \underset{P_i^{*}}{\min} & & (P_i^{*}-P_i)^2\\
&\text{s.t.} & & (P_i^{*}-P_j)^2\geq d^2, \forall j \in \{1,...,N\}.
\end{aligned}
\end{equation}
It is important to note that in case the length of the segments that define the original polygon is greater than $2\times d$, the solution to (\ref{problem}) could include points in boundary of the original polygon. To avoid this problem, we 'densify' the original polygon by adding additional points any time the length of a segment is larger than $d$. This also improves accuracy in the proximity of corners.
\subsection{Metrics}
\label{sec:metrics}
Here, we introduce two efficiency metrics to evaluate the census tract-based licensing. They are selected so that they encompass the main characteristics of census tracts: their boundary and their population.
The area of each census tract polygon and the CBRS-allowed area\footnote[11]{As noted in \ref{sec:model}, CBRS-allowed area defines the region where CBRS nodes can be deployed, under the rules on boundary signal strength limits adopted by FCC for CBRS operations.} are computed in order to evaluate the licensing scheme through the efficiency metrics of \textit{Area Loss Percentage} and \textit{Population of Census Tract with Access to Spectrum}.
The area of each polygon is computed as the area of an irregular polygon with known coordinates, based on Green's theorem.
The ALP metric, as defined in (\ref{eqn:arealoss})
takes into account the ratio between the CBRS-allowed area, denoted as $A_{\rm CBRS}$, and the area of the original census tract polygon, denoted as $A_{\rm CT}$. Under the propagation analysis applied, ALP gives the percentage of the area which cannot be used for sharing under technical rules for 3.5 GHz band.
\begin{equation}\label{eqn:arealoss}
\begin{aligned}
&{\rm ALP}=1-\frac{A_{\rm CBRS}}{A_{\rm CT}}, 0<{\rm ALP}\leqslant 1.
\end{aligned}
\end{equation}
To account for the basic characteristic of census tracts, i.e. their population - we introduce the metric PCTAS, defined in (\ref{eqn:populationloss}) as the percentage of the area of a census tract in which spectrum is not wasted, weighted by the population count $P$ of the census tract.
\begin{equation}\label{eqn:populationloss}
\begin{aligned}
&{\rm PCTAS}=(1-{\rm ALP})\times P.
\end{aligned}
\end{equation}
Under the propagation analysis applied and under the assumption of uniformely distributed population, PCTAS gives the population in census tract that can consume network capacity under the 3.5GHz spectrum sharing rules in the cities selected for the study.
\section{Conclusions}
\label{sect:Conclusions}
The main conclusion from the detailed analysis of census tracts is that they do not provide a good basis on which to build a licensing framework. The FCC rightly recognizes that smaller license areas provide the potential for unlocking efficient sharing of spectrum in 3.5GHz band using small cell deployments. But as out results have shown, even under propagation conditions that are favorable, many unit license areas are rendered unusable and that potential is not unlocked.
The problem of spectrum waste over the license area, reflected through the percentage of area loss and number of people precluded from spectrum use points to the fact that the boundaries of census tracts are not appropriate boundaries for a license area in a geographic licensing scheme envisioned for 3.5 GHz band. Not only that they mismatch with the spectrum propagation of small cell deployments, the main deployments that could put 3.5 GHz spectrum to more efficient use, but also finding the CBRS-allowed remains a problem with no solution even if the boundary limits are less stringent. In highly dense urban areas, census tracts are so small that it is not possible to form the CBRS-allowed area at all to fit the diverse spectrum usage so the interference between adjacent tracts is reduced. As we have shown, the smaller the census tracts are, more spectrum is wasted over the area.
Size of the license territory is one aspect of the license design. However, to design a flexible license based on geographic area, it is necessary to embed efficient interference protection mechanism in terms of limits on the license boundary, as well.
Getting the licensing framework right is essential. The landscape in spectrum management has changed, in that it is accepted that some form of spectrum sharing will be part of the future. However it is still an open question as to which approaches will succeed. The 3.5GHz SAS-based sharing promises much and may be hampered in delivering if census tracts are at the core of its licensing framework.
Therefore we propose the following geographic licensing models to take into consideration towards building an effective licensing framework that can fully support the SAS potential. Geographic area of deployment calculated based on the predicted actual spectrum usage should be the base area unit of the licensing scheme, i.e. area whose size is not fixed nor created for purposes other than spectrum usage. Describing the actual spectrum usage of an operator's network is a valid research question we plan to address in our future work, in order to investigate a particular challenge of how different network deployments need to be to be allocated in the same license area.
The model of spatial sharing that is adopted for the borders of the U.S and Canada and the U.S. and Mexico is an sharing arrangement in 800MHz band that is based on the geographic area so that the edge users on the borders can be served.
We are pointing out that it is worth re-considering the licensing scheme, because the size and the shape of the license area matters. As we have shown, it affects the spectrum utilisation in different ways. To grant licenses that are geographically based in a meanigful way, license design needs to be based on these recommendations:(1) the license area should not be of fixed size, and (2) license should be issued based on actual spectrum use.
\section{Results}
\label{sec:results}
\begin{figure*}
\centering
\subfloat[Case I, set-back distance of 210m ]{\includegraphics[width=6.5cm]{DC2.jpg}
\label{DCd1}}
\hfil
\subfloat[Case II, set-back distance of 663m]{\includegraphics[width=6.5cm]{DC1.jpg}
\label{DCd2}}
\caption{Washington DC area}
\label{Washington}
\end{figure*}
\begin{figure*}
\centering
\subfloat[Case I, set-back distance of 210m ]{\includegraphics[width=8.5cm]{lastCDF_ALP210.jpg}
\label{CDFAREALOSS}}
\hfil
\subfloat[Case II, set-back distance of 663m]{\includegraphics[width=8.5cm]{lastCDF_ALP663.jpg}
\label{CDFAREALOSS2}}
\caption{CDF of ALP - Area Loss Percentage}
\label{CDF_AREALOSS2}
\end{figure*}
\begin{figure*}
\centering
\subfloat[Case I, set-back distance of 210m ]{\includegraphics[width=8.5cm]{lastCDFPCTAS210.jpg}
\label{CDFPOPLOSS}}
\hfil
\subfloat[Case II, set-back distance of 663m]{\includegraphics[width=8.5cm]{lastCDFPCTAS663.jpg}
\label{CDFPOPLOSS2}}
\caption{CDF of PCTAS - Population of Census Tract with Access to Spectrum}
\label{CDF_POPLOSS}
\end{figure*}
The results of our analysis are presented in Fig. \ref{Washington}, Fig. \ref{CDF_AREALOSS2} and Fig. \ref{CDF_POPLOSS}. To give a visual insight into the problem, Fig.\ref{DCd1} (210m set-back distance) and Fig.\ref{DCd2} (663m set-back distance) show the results of the area study for Washington DC. Green areas correspond to the portion of census tracts where CBRS deployment is allowed. Red areas correspond to areas off-limits to CBRS deployment. Compared to Manhattan and San Francisco, Washington DC results show a good outcome in terms of spectrum waste over the area for Washington. In fact, in Case I, the red areas cover almost entirely Manhattan and 35\% of San Francisco. Only 18\% of census tracts in Manhattan could accommodate CBRS deployments to satisfy the boundary constraint on a minimum distance dictated by propagation analysis (i.e., the best propagation conditions case). The impact of 663m set-back distance is severe in all three cities. As seen on Fig.\ref{DCd2}, only 9 tracts, out of 179 tracts of the District of Columbia - will be available to accommodate CBRS deployments under these conditions. For 2.1km set-back distance, no census tract will be available for use in all the considered cities. \\
If we look at the cumulative distribution function (CDF) for the two efficiency metrics described in Section \ref{sec:metrics}, we can get an insight into spectrum utilisation for the cities analyzed. In the Case I, as shown in Fig.\ref{CDFAREALOSS}, 82\% of Manhattan census tracts will not be available for CBRS deployment. For the rest of the Manhattan areas, where deployment is possible, the area that is available for deployment is limited. For example, the probablity that ALP is less than 50\% is 0.0035. It means that even in the case the boundary limit can be satisfied, the potential for spectrum sharing could hardly be unlocked. In Case I, San Francisco area will be unavailable for deployments in about 35\% of census tracts, but the probablity that ALP is less than 50\% is still very small, namely 0.025. In the second case of a larger set-back distance, the Fig.\ref{CDFAREALOSS2} shows that Manhattan area will not be available for deployments at all, and only 3-4\% of Washington DC and San Francisco census tracts could be used as license areas.
\\
We now look at the unavailability of spectrum for network capacity consumers within census tract boundaries. Fig.\ref{CDF_POPLOSS} shows the CDF of PCTAS for 210m and 663m cases of propagation conditions. In the Case I, Fig.\ref{CDFPOPLOSS}, we can see that the probability that no one in Manhattan can consume the network service is 82\%. Further, it is certain that less than half of the average number of people per census tract in Manhattan (5500) can be served. Also, with probability $1$, less than three quarters of the average number of people per census tract (4087) in San Francisco will be served. Finally, in Washington DC the probability that less than average number of people per census tract (3360) will have a network service is 0.977. In Washington tracts with size of the population larger than 4000 people per tract, no one will take advantage of spectrum use. In the Case II, Fig.\ref{CDFPOPLOSS2}, no one in Manhattan can be served, almost no one in San Francisco and only less than 2000 people per census tract in Washington DC can consume network capacity.
\section{Introduction}
\label{sect:intro}
The licensing framework for sharing the 3.5 GHz band is based on the notion of \textit{census tracts}. Census tracts are geographical areas defined on the basis of population statistics with their area boundaries not expected to change much over time. The Federal Communications Commission has adopted the census tract demographic areas as the \textit{licensing area units} - their cartographic boundaries will serve as boundaries of allowed licensed operation in the 3.5GHz band\cite{RO}. In this paper, we explore whether this choice makes sense and what implications this choice has for efficient spectrum usage.
Attention from the academia and spectrum regulation sector in the past few years has been directed towards the Spectrum Access System (SAS) - a technologically sophisticated system of multiple databases that will manage the tiers of diverse users, set in a complex spectrum sharing environment of the 3.5GHz band. Specially three areas of SAS-based sharing have been discussed in the literature: (1) architectural implementation and feasibility of the SAS system \cite{SASarch}, \cite{GoogleSAS} , (2) regulatory and policy issues around license design, spectrum rights and enforcement classes for the future SAS system \cite{LEHR}, \cite{WEISS}, and (3) incumbent tier coexistence with small cell operators in the band \cite{Incumbents}, \cite{Radars}.
Despite being the basic unit on which the licensing framework for 3.5 GHz band is based, the impact of census tracts on the effectiveness of the framework has not been analyzed in detail.
A \textit{license} to use spectrum is granted with the purpose of avoiding harmful interference that the licensee may experience from the neighboring spectrum users. Therefore to efficiently manage the spectrum use, spectral and geographical neighbors have to be issued with compatible licenses. For the 3.5 GHz band, the spectrum users are Citizens Broadband Radio Service (CBRS) users, a new class of service established by the FCC in order to promote innovative sharing for 3.5 GHz band \cite{NPRM}. The\textit{ rights to transmit} (specified in the license content) for CBRS users will define the area of operation, the frequencies, and the power levels of operation. One license for CBRS users will be issued per census tract. Therefore to limit the risk of interference to neighbouring census tracts, CBRS nodes cannot be deployed too close to the boundary of the census tracts. More specifically, if licenses for two adjacent census tracts are issued to different CBRS networks - the nodes of these networks can only be deployed within the \textit{set-back distance} from the border of the census tract. The distance is determined by the propagation characteristics of the band and FCC technical rules on CBRS operation \cite{FNPRM}. As illustrated in Fig.\ref{fig:censustracts}, under these conditions, the existence of a grey space area where CBRS networks cannot be deployed points to the problem of \textit{spectrum waste over the area}, which we address in this paper.
Google Inc. has raised the problem of the census tract within the stakeholders' filings to the FCC\footnote[1]{Stakeholders filings to the Commission are available at ECFS home page: http://apps.fcc.gov/ecfs/).}, under the 3.5 GHz docket 12-354 \cite{Comments_RF}. To substantiate their concerns, Google carried out a fairly simplified geometry analysis of the census tract area boundaries in which they are represented as equivalent circular areas. Within these circles, the set-back area in which no CBRS devices can be deployed was shown to be substantial.
In the same filing, other stakeholders expressed their support for the Commission's proposal on census tracts. The arguments in favor included: (1) population characteristics of census tracts will give operators an option to plan their network deployments to target certain groups of customers for the service they aim to provide in that area and (2) the official census database could be incorporated into SAS to more effectively support the accuracy of spectrum usage information, i.e. geographic data about spectrum users.
\begin{figure}
\centering
\includegraphics[width=9cm]{tractspaper.jpg}
\caption{Adjacent census tracts allocated to different operators: the set-back distance and the sterile, grey space area in which the red nodes cannot be used.}
\label{fig:censustracts}
\end{figure}
\label{sect:introduction}
The purpose of our paper is to conduct a wider analysis of census tracts and further explore such geographic licensing scheme. It is not clear whether the existence of intersecting\footnote[2]{Intersection of census tracts is illustrated in \cite{Comments_RF} as a location where several census tracts meet, e.g. Washington Convention Center. If an operator wishes to serve the users at this location, it would need to aggregate several licenses for all census tracts that meet there. Details about the rules in Licensing Framework for 3.5 GHz are elaborated in Section \ref{sect:framework}.} census tracts or tracts of inappropriate size across the country would severely jeopardize the flexibility of the licensing framework for 3.5 GHz band or whether it is an issue only for a number of specific locations.
Therefore we use official census data (cartographic files and census statistics) in order to analyze boundaries of real census tracts. We conduct studies on entire cities, classified in the U.S. census databases as Major Economic Areas (MEAs)\footnote[3]{http://transition.fcc.gov/oet/info/maps/areas/}. Sample cities for the study are: New York City (Manhattan Burough), Washington DC and San Francisco, CA - selected as urban area types based on their diverse population densities.
As a proxy of spectrum utilisation, we use two efficiency metrics to analyze the effectiveness of census tracts-based licensing.
The first metric of Area Loss Percentage (ALP) as introduced in \cite{Comments_RF}, takes into account the boundaries of the license areas defined through the ratio of CBRS-allowed area and the corresponding census tract area.
The second takes account of the essential characteristic of a census tract, namely the demographics. Hence we introduce a metric we term Population of Census Tract with Access to Spectrum (PCTAS), which gives us an opportunity to evaluate spectrum utilisation in terms of real consumers of network capacity who will be affected by the sharing rules for 3.5 GHz band. \\
Our results show that even in the best propagation conditions in the band, for which the area that would be off-limits to CBRS nodes is minimised - the implications can be stark. For example, in the case of Manhattan, 82\% of census tracts license areas will not be available for deployment if one license is issued per one census tract. For the less favourable propagation conditions, the Manhattan area cannot be used for spectrum sharing in 3.5 GHz at all, and results for Washington DC and San Francisco show that less that 10\% of census tracts in both cities is actually available for deployments.
To begin with the framework itself, in Section \ref{sect:framework} we introduce the aspects of SAS-based sharing and potential issues with the licensing scheme. New classes of users created for this band are explained with the specific technical rules that will affect the licensing framework whose area unit is a census tract and we overview the details of census tracts. In Section \ref{sect:study} we describe the study conducted followed by the results presentation and discussion in Section \ref{sec:results}. The last section concludes the paper.
\section{Sharing Framework and SAS}
\label{sect:framework}
\begin{figure*}
\centering
\includegraphics[scale=0.3]{newSASedited.jpg}
\caption{Tiers of SAS and SAS functionalities: The 3.5GHz band becomes open to licensed and opportunistic use, accommodating critical operations facilities (hospitals, public safety networks), business and residential users alongside with federal users and fixed satellite stations which encumber the band.}
\label{fig:SAS}
\end{figure*}
The President’s Council of Advisors on Science and Technology (PCAST) Report \cite{PCAST} has called for innovation in spectrum sharing by mapping the technology and regulatory enablers and advancing them both so that sharing becomes a norm. The sharing framework for 3.5 GHz band accommodates diversity of users, envisions a range of different technologies to be used and supports many different applications to enable efficient sharing.
One of the many inspiring messages of the PCAST report was that wireless users will get to share military spectrum. To embed this promising sharing diversity into the set of effective and clear regulatory rules of operation, the FCC has put a stamp of approval on such message by first - creating an entirely new class of service for the band \cite{NPRM}, and second - adopting the technical rules of CBRS operation \cite{FNPRM}. The broad range of future usage is suggested in the title itself, Citizens Broadband Radio Service. The type of usage and the classes of services under the CBRS name, as depicted in Fig.\ref{fig:SAS}, range from licensed carrier cells, fixed wireless broadband to advanced home networking and any other uses.
Incumbent Access (IA) users, as primary holders of spectrum rights have the highest priority to access the spectrum which is reflected in exclusion and protection zones in which Priority Access (PA) and General Authorized Access (GAA) users cannot operate. For the 3.5 GHz band, the IA users are federal users (military high-powered radars on ship platforms across the coastline) and grandfathered fixed satellite services (FSS). Federal incumbent use is a matter of national security and therefore such users need continuous protection from interference. However, national defence missions are usually executed in shorter times and in limited geographic areas, therefore unused portions of spectrum in space, time and frequency can be reallocated to other users. These users will receive protection from harmful interference that commercial users in lower tiers would generate; moreover they are not required to mitigate interference they generate to lower tiers.
\subsection{The Licensing scheme and Frequency Assignments}
\textit{PA users} need to be issued with a license to access spectrum through a geographic licensing scheme - to operate within the boundaries of one census tract area.
The U.S. is divided into 74,000 census tracts, whose boundaries follow political (e.g., subdivision of counties) or geographic boundaries like rivers and roads. They are constructed to encompass a population of 2,000-4,000 on average.
A PA network operator will be issued a license to operate within one census tract in one 10 MHz channel. The license duration is 3 years and is non-renewable. However, aggregation of spectrum is allowed in all three dimensions(space, frequency and time). Spatially, one PA user is allowed to aggregate an unlimited number of census tract areas in order to serve more users. In frequency, a maximum of four 10 MHz channels within one census tract area can be aggregated. In time, PA license (PAL) applicants can apply for 2 consecutive license terms within the first applicant window, gaining in such way the license for 6 years.
The size of the census tracts varies, from areas less then a square mile in high dense urban regions, to 85,000 square miles in less inhabited rural regions. They often intersect roads, strategic touristic attractions or institutions of high commercial interest. Their intersection combined with the technical rules for a licensed operation means that an operator may have to acquire several adjacent census tracts licenses to deploy a small cell network and serve one particular building. On the other hand, when census tracts are too large, spectrum is wasted per area because of exclusivity of only one PAL operating in that tract per one 10MHz chunk. \\
\textit{GAA users} do not need to get a license and access spectrum opportunistically, similarly to an unlicensed mode of operation with the difference of being licensed-by-rule in the framework. This means that SAS assigns GAA users dynamically based on the demand, whenever and wherever spectrum is free from PA use. It also means that GAA devices have to be FCC technically certified (ability to tune to given frequencies or having embedded sensing capabilities in the device). \\
Despite the fixed bandwidth that SAS will assign to PA users (multiple of 10MHz), according to the demand and interference conditions SAS may move them to other 10 MHz spectrum chunks, if needed. To GAA users SAS will allocate from the pool of frequencies instead, i.e. any frequency free from PA use in the portion of 150 MHz, where interference constraints are satisfied. The dynamic assignments of 10MHz chunks to PAs and free frequencies from GAA pool of 150MHz, performed by a SAS system will follow \textit{the demand and supply} principle whilst providing assurance that some level of assignment convention is met. The directive on more efficient spectrum use also means that spectrum is assigned contiguously and continuously\footnote[4]{Initially, FCC proposed frequency assignment for GAA users of minimum 50 percent of the 150 MHz to be reserved for them in any census tract. After commenters questioned the proportional approach as a potential cause of uncertainty in the marketplace, FCC revised the proposal and concluded that a maximum of 70 MHz can be reserved for PALs in any given license area. This means that GAA users may access all of 150 MHz in areas where spectrum is free (no PALs issued or in use) or up to 80 MHz where all PALs are in use.}. For this reason, to assure fairness among users in the framework - GAA users are allowed in the enitre band of 150 MHz as long as incumbent and PA users tiers are not utilising the specific channels.
While the GAA users do not need licenses, since they get authorized and allocated by the SAS once they register, the concept of census tracts has implications on them. This is because the rules of operation adopt the same boundary limits for all CBSDs in the band, which means that neither PA nor GAA users can access the grey space strip area as shown in Fig.\ref{fig:censustracts}, in order to reduce the effect of interference leakage into adjacent census tracts.
As depicted in Fig.\ref{fig:SAS}, the SAS is in charge of PA and GAA assignments and authorization, tiers coordination and management of interference due to tier-interactions. The types of interference that may occur in tier interactions are: (1) \textit{co-channel interference} (from PA and GAA to IA, between two adjacent PA users, between nearby GAAs, between PA and GAA users) and (2) \textit{adjacent channel interference} (from PA and GAA to IA, between PA users across adjacent license areas, between PA users within a license area, between PA and GAA users and between nearby GAAs). The SAS needs to coordinate the users in the band dynamically and potentially across multiple bands, so it also needs an embedded mechanism for monitoring and reporting spectrum usage to be able to manage the spectrum usage automatically.
Finally, CBRS devices (CBSDs) are divided into two groups, group A is small cell indoor/outdoor low power use and group B is point-to-point use (higher power devices)\footnote[5]{All CBSDs are required to register with a SAS and provide their location, antenna height above the ground, authorization status they request for (PA or GAA), FCC ID number, serial number of the device. Optionally, CBSDs should report on their sensing capabilities. Since frequency and power assignments are established by the SAS, all CBSDs must execute SAS instructions within 60 seconds.} Technical parameters around CBSD specifications can be found in \cite{RO} and we later summarise the specific parameters used in our study are in Table \ref{tab:study}.
|
\section{Introduction}
A very recent paper by Oka et al. (2013) demonstrates that
profiles of the still unidentified diffuse interstellar bands
(DIBs), in particular the very well--correlated ones near 6196 and
6614~\AA, may not change in unison (see their Figs. 1 and 2).
The rotational temperatures of the polar molecules CH and CH$^+$,
found in Herschel 36, are as high as about 15K. In this special
object the 6614 profile shows a very pronounced extended red wing
while the latter is barely seen in 6196.
There are several possible explanations for
the origin of red wing seen in DIB 6614: Bernstein et al. (2015)
offer the hypothesis of accidentally overlapping diffuse bands,
Oka et al. (2013) consider a single rotational contour where
high rotational transitions being pumped by the nearby irradiation from Her 36.
Lastly, Marshall et al (2015) suggest that "...the extended
red tails would arise principally from vibrational hot bands that
are red-shifted with respect to the origin band..."
The possible influence of
the rotational temperatures of centrosymmetric species (C$_2$) on
some DIB profiles was already suggested by Ka{\'z}mierczak et al.
(2010). This result received a support very recently
(Gnaci{\'n}ski \& Kre{\l}owski (2014); some DIB profile widths
have been shown as related to the rotational temperature of H$_2$
-- the homonuclear molecule.
\begin{table*}
\caption{Equivalent widths of the considered DIBs in spectra of
our targets. Origin of spectra is marked as follows: b - BOES, f - FEROS, h - HARPS, M - McDonald, t - Terskol, u - UVES.
Weighted averages of the same target data are boldfaced.
}
\label{ew}
\begin{tabular}{lrr lrr lrr}
\hline
Star & EW(6196) & EW(6614) & Star & EW(6196) & EW(6614) & Star & EW(6196) & EW(6614) \\
\hline
bd59456b & 52.8$\pm$2.3 & 224.2$\pm$6.0 & 144470f & 15.8$\pm$1.3 & 60.4$\pm$2.7 & 154445u & 24.0$\pm$0.7 & 103.9$\pm$1.3 \\
bd60594b & 40.2$\pm$1.7 & 171.1$\pm$5.1 & 144470h & 18.6$\pm$3.1 & 57.8$\pm$5.2 & 157038u & 46.8$\pm$0.6 & 179.0$\pm$2.7 \\
bd404220b & 89.5$\pm$2.5 & 300.7$\pm$5.2 & &{\bf 16.2$\pm$1.0}&{\bf 59.8$\pm$1.1}& 161056u & 38.0$\pm$1.5 & 161.0$\pm$3.6 \\
bd404227t & 84.6$\pm$4.6 & 327.4$\pm$9.9 & 145037u & 72.6$\pm$2.1 & 244.0$\pm$6.0 & 163800h & 27.8$\pm$2.5 & 118.2$\pm$2.5 \\
CD-324348u & 82.4$\pm$0.8 & 344.6$\pm$2.5 & 145502f & 15.0$\pm$0.6 & 58.0$\pm$1.9 & 166734u & 97.4$\pm$1.1 & 417.1$\pm$3.7 \\
CygOB27b & 90.0$\pm$4.1 & 344.8$\pm$9.9 & 147165h & 18.0$\pm$0.7 & 60.1$\pm$1.8 & 167971u & 62.5$\pm$1.0 & 250.9$\pm$3.5 \\
CygOB28ab & 94.9$\pm$4.0 & 348.7$\pm$7.0 & 147165u & 17.6$\pm$0.6 & 58.0$\pm$1.5 & 168607u & 97.3$\pm$1.2 & 387.4$\pm$3.8 \\
CygOB211b &112.0$\pm$5.7 & 380.0$\pm$7.8 & &{\bf 17.8$\pm$0.2}&{\bf 58.9$\pm$1.0}& 168625u &100.5$\pm$1.0 & 434.2$\pm$2.2 \\
CygOB212b &116.8$\pm$9.8 & 386.4$\pm$9.9 & 147889h & 46.2$\pm$2.0 & 179.5$\pm$3.4 & 169454u & 59.9$\pm$2.8 & 196.6$\pm$3.3 \\
15785b & 46.7$\pm$1.2 & 221.3$\pm$4.8 & 147889u & 45.4$\pm$0.9 & 181.6$\pm$1.5 & 169454M & 59.3$\pm$1.2 & 194.8$\pm$2.0 \\
23180t & 15.3$\pm$1.6 & 61.8$\pm$5.1 & &{\bf 45.5$\pm$0.3}&{\bf 181.3$\pm$0.8}& &{\bf 59.4$\pm$0.2}&{\bf 195.3$\pm$0.8}\\
24398b & 13.9$\pm$1.0 & 58.1$\pm$4.6 & 147933h & 15.9$\pm$1.0 & 62.7$\pm$2.7 & 170740u & 28.8$\pm$1.1 & 116.9$\pm$1.7 \\
24912t & 22.7$\pm1.3$ & 95.3$\pm$3.2 & 148184h & 12.5$\pm$0.7 & 42.1$\pm$2.0 & 179406h & 20.0$\pm$0.9 & 94.2$\pm$2.2 \\
27778M & 10.6$\pm$0.8 & 45.5$\pm$1.9 & 148184u & 13.2$\pm$0.6 & 40.5$\pm$1.7 & 183143m & 96.0$\pm$1.3 & 358.5$\pm$2.8 \\
34078M & 24.9$\pm$1.2 & 66.0$\pm$3.1 & &{\bf 12.9$\pm$0.3}&{\bf 41.2$\pm$0.8}& 184915m & 20.4$\pm$1.2 & 77.3$\pm$2.1 \\
34078t & 22.0$\pm$1.2 & 54.0$\pm$3.1 & 148379u & 42.8$\pm$0.8 & 159.0$\pm$2.5 & 184915t & 17.8$\pm$2.7 & 73.2$\pm$4.2 \\
34078u & 25.4$\pm$2.1 & 65.9$\pm$4.0 & 149408u & 45.9$\pm$1.4 & 177.0$\pm$2.6 & &{\bf 20.0$\pm$1.0}&{\bf 76.5$\pm$1.6}\\
&{\bf 23.7$\pm$1.1}&{\bf 61.4$\pm$4.1}& 149757h & 10.5$\pm$0.7 & 41.8$\pm$1.9 & 194839t & 66.1$\pm$3.9 & 197.5$\pm$9.0 \\
63804u & 87.6$\pm$1.7 & 332.0$\pm$3.5 & 149757u & 11.2$\pm$0.8 & 40.7$\pm$1.5 & 203532M & 14.0$\pm$0.8 & 58.4$\pm$1.1 \\
73882u & 17.1$\pm$0.9 & 48.1$\pm$1.8 & &{\bf 10.8$\pm$0.3}&{\bf 41.1$\pm$0.5}& 204827b & 41.2$\pm$2.6 & 173.1$\pm$4.0 \\
78344u & 76.6$\pm$2.1 & 281.4$\pm$4.6 & 151932u & 31.6$\pm$0.9 & 121.8$\pm$2.1 & 207198t & 33.0$\pm$2.0 & 125.2$\pm$3.7 \\
80077u & 80.2$\pm$1.2 & 291.3$\pm$1.9 & 152003u & 32.9$\pm$2.0 & 105.8$\pm$2.9 & 210839t & 32.9$\pm$0.8 & 145.5$\pm$3.2 \\
110432u & 17.0$\pm$0.3 & 74.4$\pm$1.6 & 152233h & 22.6$\pm$1.2 & 78.3$\pm$2.4 & 226868b & 49.0$\pm$3.8 & 180.2$\pm$4.9 \\
114213u & 38.9$\pm$0.9 & 145.2$\pm$3.7 & 152233M & 22.7$\pm$1.1 & 79.3$\pm$2.1 & 228712b & 57.4$\pm$2.7 & 181.4$\pm$3.6 \\
115842u & 36.8$\pm$1.1 & 151.0$\pm$2.0 & &{\bf 22.7$\pm$0.1}&{\bf 78.9$\pm$0.5}& 228779b & 71.8$\pm$2.1 & 285.9$\pm$6.0 \\
125241u & 57.1$\pm$0.8 & 224.3$\pm$3.4 & 152235h & 36.3$\pm$1.1 & 132.5$\pm$2.6 & 229059t & 63.8$\pm$4.8 & 231.4$\pm$7.5 \\
142468u & 56.4$\pm$0.8 & 220.2$\pm$3.4 & 152236u & 39.3$\pm$1.2 & 129.0$\pm$2.1 & 254577b & 51.7$\pm$2.2 & 206.0$\pm$4.5 \\
144217h & 13.2$\pm$0.8 & 40.9$\pm$2.0 & 152249h & 20.0$\pm$1.0 & 86.4$\pm$2.7 & 278942b & 32.0$\pm$2.1 & 166.0$\pm$4.6 \\
144217u & 11.1$\pm$0.3 & 45.6$\pm$1.5 & 152249u & 19.2$\pm$0.6 & 81.5$\pm$1.9 & Hersch36 & 40.6$\pm$4.1 & 151.3$\pm$9.9 \\
&{\bf 11.4$\pm$0.7}&{\bf 43.9$\pm$2.3}& &{\bf 19.4$\pm$0.4}&{\bf 83.1$\pm$2.3}& & & \\
144218h & 14.1$\pm$3.1 & 57.4$\pm$8.6 & 154368u & 33.8$\pm$1.9 & 150.0$\pm$3.3 & & & \\
\hline
\end{tabular}
\end{table*}
The pair of the diffuse bands near 6196 and 6614~\AA\ was found as
very well correlated by Moutou et al. (1999). The intensities of
the two DIBs were found as correlated with the Pearson's
coefficient as high as 0.97. The first guess could thus be that
6196 and 6614 may form a ``family'' sharing a common carrier.
However, the very high resolution analysis of the profiles of the
two DIBs created some doubts (Galazutdinov et al., 2002). Their
Fig. 4 demonstrated that while EW's of 6196 and 6614 do correlate
quite tightly, their FWHM's do not. Apparently the profiles of
both DIBs do not react in the same fashion to the varying physical
conditions inside interstellar clouds. This fact is also confirmed
by differing shapes of both DIB profiles (Galazutdinov, Lo Curto
\& Kre{\l}owski 2008).
The nearly perfect correlation of the two bands was confirmed by
McCall et al. (2010) which is better than any other pair of DIBs.
Nevertheless, the authors emphasized the need to explain the very different profiles of
the 6196 and 6614 diffuse bands if they are of the same origin. Indeed, profile of the former one
is much more symmetric and narrow than that of the latter.
Naturally, it was the very first guess that trying to divide
diffuse bands into sets of features carried by the same molecules
we should put into one ``family'' the features which are really
well correlated (e.g. Kre{\l}owski, Schmidt \& Snow (1997)).
However, the above mentioned papers expressed some doubts on
whether the two nearly perfectly correlated DIBs share the same
carrier. The profiles are of very different widths and these
widths vary but not in the same way. It is thus important to check
whether the observed scatter, seen despite a very high correlation
coefficient, follows just measurement errors or has some physical
background.
It is commonly believed that the DIB carriers are some complex
molecules. The most popularly considered molecular species, being
candidates for DIB carriers, are chain molecules, based on a
carbon skeleton. They have been observed in star forming regions
but the attempts to find them in translucent clouds failed
(Motylewski et al. 2000). Another candidature is that of
polycyclic aromatic hydrocarbons (PAHs). A kind of their mixture
is believed as the carrier of a few emission infrared bands.
However, the attempts to identify any specific PAH -- this is
possible only in the near UV range -- failed as well (Salama et
al. 2011; Gredel et al. 2011). Another possibility -- fullerenes
-- seems to be restricted to a few DIBs only, despite this species
were discovered in star forming regions (Cami et al. 2010).
Recently Campbell et al. (2016) reported
the identification of C$_60^+$ near infrared features in spectra of relatively cool object (B7Ia) HD\,183143.
Earlier attempts to identify C$_60^+$, e.g. by Jenniskens et al. (1997), Galazutdinov et al. (2000) pointed out
that the lack of analysis of spectra of differentiated spectral types makes the identification of C$_60^+$
uncertain because the interstellar features (especially 9633) may be severely contaminated with stellar lines,
being specific in each case. Unfortunately, this critical issue have not been addressed by Campbell et al.
\section{Observations}
The UVES spectra of our targets were collected at the 8-m UT2
telescope (Paranal Observatory, ESO, Chile). We used our own
observations as well as those available from the public ESO archive.
The spectra were carried out in two standard modes DIC1(346+580)
and DIC2(437-860) covering the whole wavelength range $\sim$3050 -
$\sim$10400 \AA\AA\, some of them with a gap between $\sim$5770 -
$\sim$5830 \AA\AA\, due to the inaccurate set-up of the optical
elements of spectrograph.
The applied slit width was 0.4 and 0.3 arcsecond for the blue and
red branches of the spectrograph respectively. These widths
satisfy to the 2 pixel criteria for the slit image projection to
the corresponding CCD cameras, providing the highest possible
resolving power $\sim$80,000 and $\sim$110,000 for the blue and
red spectrograph branches respectively.
Several spectra used in this investigation were acquired using the
HARPS spectrograph fed with the 3.6\textendash m LaSilla
telescope. This instrument offers a very high resolution
(R=115,000); the wavelength range is 3,800 -- 6,900~\AA\ divided
into 72 orders.
Additional spectra were collected with the aid of the
Magellan/Clay telescope at the Las Campanas Observatory (Chile)
using the MIKE spectrograph (Bernstein et al. 2003) with a
0.35$\times$5 arcs slit. The resulting high resolution spectrum
(R$\sim$65,000) is an average of several individual exposures,
achieves the high S/N ratio ($\sim$ 1000) and covers the broad
wavelength range $\sim$3600 - $\sim$ 9400 \AA\AA) with the lack of
gaps between spectral orders.
A big portion of the data was collected at the ICAMER (North
Caucasus, Russia) observatory, with the aid of the coud\'e echelle
spectrograph MAESTRO (Musaev 1999)
fed by a 2\textendash m telescope. These spectra may be obtained
in two operating modes of the instrument, with resolving powers
respectively of 45,000 and 120,000. In this paper we used only the
high (R=120,000) resolution.
We have measured our selected diffuse bands' equivalent widths
also in the spectra collected at the McDonald Observatory in 1993
by one of us (JK). The Sandiford echelle spectrograph allows the
resolution as high as R=60,000 and the wavelength range 5,650 --
7,000~\AA\ is divided into 27 orders.
Several targets have been observed using the Feros spectrograph,
fed with the 2.2\textendash m telescope. The resolution of this
instrument is R=48,000 and the spectral range covers 3,700 --
9,000~\AA. The latter is divided into 37 orders.
Lastly, our sample includes seven spectra obtained through the
fibre\textendash fed echelle spectrograph BOES (Kim et al., 2007)
installed on the 1.8-m telescope at the Bohyunsan Observatory
(South Korea). In this case the resolution may be either 30,000;
45,000 or 90,000; the wavelength range is always 3,800 --
10,000~\AA\ divided into 76 orders.
All spectra were processed and measured in a standard way using
both {\sevensize IRAF} (Tody 1986) and our own
{\sevensize DECH}\footnote[1]{http://gazinur.com/DECH-software.html} codes.
\section{Results}
Oka (2013) demonstrated that profiles of both 6196 and 6614
(especially the latter) may change seriously while the rotational
temperatures of CH or CH$^+$ are extraordinarily high, forming the
extended red wings. Our sample of high S/N ratio targets shows
that similar extended red wings may be observed in relatively many
objects though not as evident as in Herschel 36. We have selected
three spectra of HD179406, HD147889 (mentioned as differing
seriously in T$_{rot}$ of C$_2$ by Ka\'zmierczak et al. 2010) and
Herschel 36 to compare the DIB profiles. The result is shown in
Fig. \ref{erw03}. Apparently HD179406 is the case of low
rotational temperatures of molecular species and HD147889 is
somewhere in between of HD179406 and Herschel 36. We would like to
draw attention to a comparison of two spectra of HD147889 and
HD179406: one from HARPS and one from UVES -- both of very high
S/N ratio. The DIB profiles are not identical in both these
sources. Apparently the extended red wings develop with the
growing rotational temperatures of molecular species. In the case
of 6196 the DIB core is broadened when T$_{rot}$ grows but the
extended red wing is relatively shallow while in the case of 6614
it is very pronounced.
\begin{figure}
\includegraphics[angle=270,width=8.5cm]{erw03.eps}
\caption{The extended red wings of the considered DIBs growing in
unison with T$_{rot}$ of identified species.}
\label{erw03}
\end{figure}
Our measurements of the equivalent widths of both DIBs in our
sample of high S/N ratio spectra are given in Table 1.
Equivalent width error estimations are based on use of equation 7
from Vollmann \& Eversberg (2006). The method neglects possible continuum
placement errors, though hard to expect sufficient incorrectness in case of relatively narrow DIBs like 6196.
It is important to emphasize that the extended red wings are incorporated into the
integrated profiles if present. This set of data is especially
reliable as we have selected from our database only the best
spectra, i.e. with S/N ratio exceeding 250 and where the DIB
profiles are not contaminated with stellar lines. Our sample
contains thus 80 spectra; some of the targets have been observed
using more than one instrument. We consider the fact that
equivalent widths, measured using different instruments, coincide
inside the one sigma errors, as very important. They prove that
all measurements can be repeated with the same result. The
correlation plot is demonstrated in Fig. \ref{corr}.
\begin{figure}
\includegraphics[angle=270,width=8.5cm]{corr.eps}
\caption{The tight relation between 6196 and 6614 diffuse bands. A
few departing points are labeled with the stellar HD or BD
numbers. For HD34078 we plotted measurements from UVES, MIKE and
Maestro. Linear fit done with including errors for both coordinates, see Fasano \& Vio(1988).
} \label{corr}
\end{figure}
Fig. \ref{corr} clearly supports the conclusions of the above
mentioned papers -- that the correlation between both DIBs is
nearly perfect. However, it seems interesting to compare the
departing points and so -- to check whether they really do not
follow the main stream because of the measurements' errors or the
strength ratios of 6196 and 6614 in these targets are really
different.
There are some evidently departing points which represent
HD34078 (weak DIBs but three different spectra) and HD15785 where
both DIBs are strong and the BOES spectrum is of very high
quality. The profiles of both diffuse bands are shown in the
radial velocity scale Fig. \ref{ratio}. Another departing star - HD169454, is
represented by two points: one based on UVES and one -- on MIKE
spectra, both of very high quality.
\begin{figure}
\includegraphics[angle=0,width=8.5cm]{ratio.eps}
\caption{The profiles of both DIBs plotted in the radial velocity
scale for the two special targets: HD34078 (AE Aur) and HD15785.
The different strength ratio follows the scatter seen in Fig.
\ref{corr}.}
\label{ratio}
\end{figure}
Fig. \ref{ratio} demonstrates that the strength ratio of the two
DIBs under consideration really varies from target to target.
Apparently there are two different carriers causing these DIBs
though they must be related in a way -- their abundance ratios are
similar but not identical. This seems to support the conclusions
of Galazutdinov et al. (2002) that the FWHM's of the DIBs do not correlate
and the recent result of Oka et al. (2013) which demonstrates
evident extended red wing in 6614 while the latter is barely seen
in 6196. This extended red wing seems to appear in HD34078 and
thus the object may be a transition one between typical objects
with low rotational temperatures of molecules and the extreme
object -- Herschel 36. It is worth to mention that the T$_{rot}$ of the
CN molecule in this object (HD34078) is as high as 4.0K
(Kre{\l}owski et al., 2012). It is interesting that the point, representing Herschel 36 in Fig.
\ref{corr} is situated exactly at the average relation. The
observed 6196/6614 strength ratios seemingly do not depend on
rotational temperatures of identified molecular species.
\begin{figure}
\includegraphics[angle=0,width=8.5cm]{ratio_01.eps}
\caption{The same as in Fig. \ref{ratio} but for Cyg OB2 11 and
HD168625. The different strength ratio in both spectra is shown
beyond a doubt. Apparently the scatter, seen in Fig. \ref{corr}
does not follow measurements' errors -- it reflects different
physical conditions.}
\label{ratio_01}
\end{figure}
Fig. \ref{corr} allows to select other pairs of targets situated
at both sides of the average relation. We also demonstrate the same
as in Fig. \ref{ratio} effect for Cyg OB2 11 and HD168625 and for
HD278942 and HD152003 -- Fig. \ref{ratio_01} and Fig.
\ref{ratio_pl}. The plots clearly support the conclusion that the
two DIBs' ratios really differ from object to object. Apparently
the physical scatter is not related to the changes of DIB
profiles. The profiles apparently depend on rotational
temperatures of the carriers while the strength ratios -- on their
abundances.
\begin{figure}
\includegraphics[angle=0,width=8.5cm]{ratio_pl.eps}
\caption{The same as in Fig. \ref{ratio} but for HD278942 and
HD152003. The different strength ratio in both spectra is once
again shown beyond a doubt. } \label{ratio_pl}
\end{figure}
\section{Conclusions}
Our analysis of the 6196 and 6614 \AA\ diffuse bands leads to the
following conclusions:
\begin{itemize}
\item
profiles of both diffuse bands change
with the rotational temperature of identified species; the growing
T$_{rot}$ results with the extended red wings
\item
both considered DIBs correlate almost perfectly (with the Pearsons
correlation coefficient 0.98); the average 6614/6196 strength
ratio being 3.88$\pm$0.13
\item
the observed scatter is not a result of measurement's errors, i.e. the difference is higher than uncertainty of individual measurements; the
6614/6196 strength ratios are proved to be variable
\item
the variable strength ratio is not related to the observed profile
changes; apparently both DIBs are caused by different carriers but
their abundances are pretty closely related.
\end{itemize}
Apparently the very first idea, expressed by Kre{\l}owski \&
Walker (1987) and by Kre{\l}owski \& Westerlund (1988), that all
DIBs do not share the same carriers but can be divided into
several ``families'' cannot be helpful while trying to identify
DIB carriers. An analysis of DIB profiles, pioneered by Westerlund
\& Kre{\l}owski (1988) is necessary as well. As yet it was not
possible to find any pair of DIBs being certainly carried by the
same species. Seemingly all stronger DIBs which can be reliably
measured in statistically meaningful samples of objects are
carried by a different molecule each.
\section*{Acknowledgments}
JK acknowledges the financial support of the Polish National
Center for Science during the period 2016 - 2018 (grant
UMO-2015/17/B/ST9/03397). GAG thanks the Russian Science Foundation (project 14-50-00043, the Exoplanets program) for
support of observational and interpretational parts of this study.
|
\section{Introduction}
In asymptotic geometric analysis, the hyperplane conjecture is one of the outstanding open problems that first appeared explicitly in a work of Bourgain \cite{Bou}. It asks about the existence of an absolute constant $c>0$ such that every convex body of unit volume has a hyperplanar section of volume bounded from below by $c$, independently of the space dimension. An equivalent formulation, following from a result of Hensley \cite{He}, is that the isotropic constant of every convex body is bounded from above by an absolute constant (a formal definition is provided in Section \ref{sec:prelim} below). Although the hyperplane conjecture has an affirmative answer for several classes of convex bodies such as unconditional convex bodies \cite{Bou,MP}, zonoids, duals of zonoids \cite{B}, bodies with a bounded outer volume ratio \cite{MP}, or unit balls of Schatten norms \cite{KMP}, the general case still remains one of the central open problems in this area. The best general upper bound for the isotropic constant known up to now is due to Klartag \cite{K06} and gives an upper bound of order $n^{1/4}$ with $n$ being the space dimension. This improves by a logarithmic factor the previous bound of Bourgain \cite{B91}.
Milman and Pajor \cite{MP} discovered an interesting connection between the hyperplane conjecture and geometric properties associated with random polytopes. More precisely, they proved that the second moment of the volume of a random simplex in an isotropic convex body is closely related to the value of its isotropic constant, see \cite[Theorem 3.5.7]{IsotropicConvexBodies}. Furthermore, since the pioneering work of Gluskin \cite{G}, random polytopes are a major source for extremizers and counterexamples in high-dimensional convex geometry. As a consequence, they are natural candidates for a potential counterexample to the hyperplane conjecture stated above. Following this philosophy, the isotropic constant of several classes of random polytopes has been studied in recent years. More exactly, it has been shown in \cite{A,DGG,HHRT,KK} that the isotropic constant of these random polytopes is bounded by an absolute constant with probability tending to one, as the space dimension tends to infinity. The models studied so far are Gaussian random polytopes \cite{KK}, random convex hulls of points from the Euclidean unit sphere \cite{A}, random polytopes that arise from uniform random points chosen in the interior of an isotropic convex body \cite{DGG} and random polyhedra associated with a parametric class of Poisson hyperplane tessellations \cite{HHRT}. Following along the lines of \cite{KK}, it was recently proved independently in \cite{ALT} and \cite{GHT} that if $K_N$ is the symmetric convex hull of $N\geq n$ independent random vectors uniformly distributed in the interior of an $n$-dimensional isotropic convex body $K$, then the isotropic constant $L_{K_N}$ of $K_N$ is bounded by a constant multiple of $\sqrt{\log (2N/n)}$ with overwhelming probability (see also \cite{A11,ABBW,J94} for earlier results).
The principal aim of the present paper is to investigate the isotropic constant of random polytopes that arise by taking the convex hull of random points chosen with respect to the cone measure from the boundary of an $\ell_p$-sphere in $\R^n$ with $1\leq p<\infty$. With the exception of \cite{A}, where the special case $p=2$ of the Euclidean sphere has been investigated (in that case, cone and surface measure coincide), the isotropic constant of such a class of random polytopes has not been studied yet to the best of our knowledge. On the other hand, random polytopes with vertices on the boundary of a prescribed convex body have previously been investigated from a different angle by B{\"o}r{\"o}czky, Fodor and Hug \cite{BoeroeEtAl}, Reitzner \cite{Reitzner2002,Reitzner2003}, Richardson, Vu and Wu \cite{RichardsonVuWu} as well as Sch\"utt and Werner \cite{SchuettWernerRPBoundary}. These papers concentrate on expectation and variance asymptotic for several key geometric functionals, such as the volume and the vertex number of these random polytopes, as the number of vertices tends to infinity. In contrast, we essentially analyse the geometric behaviour of random polytopes with vertices on the boundary in high dimensions, that is, as the number of vertices and the space dimension tend to infinity simultaneously.
As in most of the previous results that verified the boundedness of the isotropic constant of random polytopes, our proof follows the approach invented by Klartag and Kozma \cite{KK} (see also \cite{A,DGG} and we refer to \cite{HHRT} for a notable exception based on \cite{HH14}). On the other hand, we would like to emphasize that in the case we study several additional tools are required. These include
\renewcommand\labelitemi{\tiny$\bullet$}
\begin{itemize}
\item a coupling argument to estimate the volume radius from below,
\item a probabilistic representation of the cone measure of an $\ell_p$-sphere due to Schechtman and Zinn \cite{SZ} (see also \cite{RR91}),
\item moment estimates for sums of independent random variables with log-concave tails proved by Barthe, Gu\'edon, Mendelson and Naor \cite{BGMN}, which are based on earlier work of Gluskin and Kwapie\'n \cite{GK}, and
\item a $\psi_2$-estimate for the cone measure on an $\ell_p$-sphere in the flavour of Bobkov and Nazarov (see \cite{BobkovNazarov}).
\end{itemize}
\medskip
The rest of this paper is structured as follows. Our main result, Theorem \ref{thm:Main} below, together with its mathematical framework, is presented in the next section. In Section \ref{sec:prelim} we collect some preliminary material that is needed in our further arguments. All proofs are presented in Section \ref{sec:ProofThm} at the end of this paper.
\section{Presentation of the result}
By $\K^n$ we denote the class of convex bodies in $\R^n$, that is, the class of compact convex subsets of $\R^n$ having non-empty interior. We denote the $n$-dimensional Lebesgue measure of $K\in\K^n$ by $\vol_n(K)$.
One says that $K\in\K^n$ is isotropic or in isotropic position, if $\vol_n(K)=1$, $K$ has its barycentre at the origin and if
\[
\int_K\langle x,\theta\rangle^2 \, \dint x=L_K^2\,,\qquad \theta\in \SSS^{n-1}\,.
\]
Here, $\langle \,\cdot\,,\,\cdot\,\rangle$ stands for the standard scalar product on $\R^n$ and $L_K$ is a constant independent of $\theta$, the so-called {isotropic constant} of $K$.
For $1\leq p <\infty$, we denote by $\ell_p^n$ the space $\R^n$ supplied with the norm
\[
\|x\|_p=\bigg(\sum_{i=1}^n |x_i|^p\bigg)^{1/p},\qquad x=(x_1,\ldots,x_n)\in\R^n.
\]
The unit ball in $\ell^n_p$ is denoted by $\B_p^n=\{x\in\R^n \,:\, \|x\|_p \leq 1 \}$ and we write $\SSS_p^{n-1}=\{x\in\R^n \,:\, \|x\|_p=1 \}$ for the corresponding unit sphere. If $p=2$, we just write $\SSS^{n-1}$ instead of $\SSS_2^{n-1}$.
The {cone (probability) measure} $\bM_K$ of a convex body $K\in\K^n$ is defined as
$$
\bM_K(B) = \frac{\vol_n\big(\{rx:x\in B\,,0\leq r\leq 1\}\big)}{\vol_n(K)}\,,
$$
where $B\subseteq\bd\,K$ is a Borel subset of the boundary $\bd\,K$ of $K$. We remark that the cone measure $\bM_{\B_p^n}$ of $\B_p^n$ coincides with the normalized surface measure on $\SSS_p^{n-1}$ if and only if $p=1$ or $p=2$ (and additionally if $p=\infty$, but we exclude this case in our text). For a more detailed account to the relationship between the cone and the surface measure on $\ell_p$-balls we refer the reader to \cite{N,NR2002}.
Given an isotropic convex body $K\in\K^n$ and $N\geq n+1$, we let $X_1,\ldots,X_N$ be independent random vectors that are distributed on the boundary $\bd\,K$ of $K$ according to the probability law $\bM_K$. By
$$
K_N=\conv\big(\{\pm X_1,\ldots,\pm X_N\}\big)
$$
we denote the symmetric convex hull of these points. The random polytopes $K_N$ are in the focus of our attention in this paper. Our main theorem shows that for $K=\B_p^n$, the $\ell_p$-ball in $\R^n$, the isotropic constant of $K_N$ is absolutely bounded with overwhelming probability. (To simplify the presentation we shall retain the general notation $K_N$ also for the particular choice $K=\B_p^n$, the meaning will always be clear from the context.)
\begin{theorem}\label{thm:Main}
There exist absolute constants $c,c_1,c_2\in(0,\infty)$ such that if $1\leq p<\infty$, $n+1\leq N\leq e^{\sqrt{n}}$, and $X_1,\dots,X_N\in\SSS_p^{n-1}$ are independent random vectors with distribution $\bM_{\B_p^n}$, then
\[
\bP\big(L_{K_N} \leq c\big) \geq 1-e^{-c_1\sqrt{N}}-2e^{-c_2n\log(2N/n)}\,.
\]
\end{theorem}
The result of Theorem \ref{thm:Main} in particular implies that the probability for the event that the isotropic constant of random polytopes with vertices on any $\ell_p$-sphere $\SSS_p^{n-1}$ stays absolutely bounded (independently of $p$) tends to one exponentially fast, as the space dimension (and simultaneously the number of vertices) tends to infinity.
\begin{remark}
We remark that for $p=2$, the result reduces to the main theorem of \cite{A}. In fact, in \cite{A} the number $N$ is allowed to vary in the range $N\geq n+1$. It might also be possible to extend Theorem \ref{thm:Main} to this range, but we have made no attempt in this direction. We also remark that if the number of points $N$ is proportional to the space dimension $n$, the main result from \cite{ABBW} already implies boundedness of the isotropic constant $L_{K_N}$ with probability one.
\end{remark}
In principle, our proof follows along the lines of \cite{KK} (see also \cite{A,DGG}), but requires several additional tools that have not found attention in this context before. On the one hand, we need a high probability lower bound on the volume radius of our random polytope. This is achieved by a coupling argument that allows us to use the known results from \cite{DGT1} when the random points are uniformly distributed inside an isotropic convex body. On the other hand, we need to apply the $\psi_2$-version of Bernstein's inequality and to this end, we need to study the $\psi_2$-behaviour of $\langle X,\theta \rangle$, where $X$ is distributed according to the cone measure of $\B_p^n$ and $\theta\in \SSS^{n-1}$. Since $\B_p^n$ is a $\psi_2$-body iff $p\geq 2$, the $\psi_2$-constant of $n^{1/p}\langle X,\theta \rangle$ is uniformly bounded in that case and we can conclude along the lines of \cite{A,KK}. In contrast, if $1\leq p<2$, we will have to take a different route, which first consists in establishing in this special situation an analogue for the cone measure of a result of Bobkov and Nazarov \cite{BobkovNazarov}. To obtain such an estimate we will carefully combine a probabilistic representation of the cone measure of $\B_p^n$ due to Schechtman and Zinn (see \cite{SZ} as well as \cite{RR91} and also \cite{BGMN,SZ2} for extensions) with moment estimates for sums of independent random variables having logarithmically concave tails due to Gluskin and Kwapie\'n (see \cite{GK} and also \cite{BGMN}). We then build on the methods from a paper of Dafnis, Giannopoulos and Gu\'edon \cite{DGG} to prove boundedness of the isotropic constant in the regime $1\leq p<2$.
\section{Some Preliminaries} \label{sec:prelim}
\subsection{General notation}
Fix a space dimension $n\geq 1$ and, as in the previous section, denote by $\K^n$ the space of convex bodies in $\R^n$. A convex body $K\in\K^n$ is symmetric (with respect to the origin) if $x\in K$ implies that also $-x$ belongs to $K$.
The {support function} of a convex body $K\in\K^n$ is defined as
$$
h_K(u) = \max\big\{\langle x,u\rangle:x\in K\big\}\,,\qquad u\in\R^n\,.
$$
It is well known that a convex body $K$ is uniquely determined by its support function $h_K$.
\smallskip
For general $K\in\K^n$ we recall that the {isotropic constant} $L_K$ of $K$ satisfies
\begin{equation}\label{eq:IsotropicConstantDef}
nL_K^2 = \min \bigg\{ \frac{1}{\vol_n(AK)^{1+{2\over n}}}\int_{z+AK} \|x\|_2^2\,\dint x \,:\, z\in\R^n, \, A\in {\rm GL}(n) \bigg\}\,,
\end{equation}
where ${\rm GL}(n)$ is the group of invertible linear transformations of $\R^n$ (see, e.g., \cite[Definition 10.1.6]{AsymGeomAnalyBook}). Moreover, according to \cite[Lemma 11.5.2]{IsotropicConvexBodies} there exists an absolute constant $c\in(0,\infty)$ such that
\begin{equation}\label{eq:IsotropicConstantL1Bound}
nL_K \leq {c\over\vol_n(K)^{1+{1/ n}}}\int_K\|x\|_1\,\dint x
\end{equation}
for any symmetric convex body $K\in\K^n$.
\smallskip
For a symmetric $K\in\K^n$, its (Minkowski) gauge, $\|\cdot\|_K$, is given by
\[
\|x\|_K = \inf\{\lambda>0 \,:\, x\in \lambda K \}\,,\qquad x\in\R^n\,.
\]
Obviously this is a norm on $\R^n$, which has $K$ as its unit ball.
We define the {Minkowski map} $\mu_K:K\to\bd\,K$ by
\[
x\mapsto \mu_K(x)={x\over\|x\|_K}\,.
\]
\smallskip
Finally, for two sequences $(a(n))_{n\in\N}$ and $(b(n))_{n\in\N}$ of real numbers, we write $a(n)\gtrsim b(n)$ (or $a(n)\lesssim b(n)$) provided that there is a constant $c\in(0,\infty)$ such that $a(n)\geq cb(n)$ (or $a(n)\leq cb(n)$) for all $n\in\N$. Moreover, we write $a(n)\approx b(n)$ if $a(n) \lesssim b(n)$ and $a(n) \gtrsim b(n)$.
\subsection{$\psi_2$-estimates and Bernstein's inequality}
Let $(\Omega,\mathcal F, \bP)$ be a probability space and recall that a convex function $M:[0,\infty)\to[0,\infty)$ with $M(0)=0$ is called an {Orlicz function}. The set of all (equivalence classes of) measurable functions $f:\Omega\to\R$ such that for some $\lambda>0$,
\[
\int_{\Omega}M\left({|f|\over \lambda}\right)\,\dint \bP < \infty,
\]
is called {Orlicz space} associated with $M$ and is denoted by $L_M(\Omega,\bP)$. This space becomes a Banach space when it is supplied with the Luxemburg norm
\[
\|f\|_{M} = \inf\bigg\{ \lambda>0 \,:\, \int_{\Omega}M\left({|f|\over \lambda}\right)\,\dint \bP \leq 1 \bigg\}\,,
\]
which is equivalent to the Orlicz norm on that space.
As already discussed in the previous two sections, we are interested in the case in which the Orlicz function is $\psi_2(x)= e^{x^2}-1$ and in which $\Omega=\SSS_p^{n-1}$ for some $1\leq p<\infty$, $\mathcal F$ is the Borel $\sigma$-field on $\SSS_p^{n-1}$ and $\bP=\bM_{\B_p^n}$ is the cone probability measure of $\B_p^n$.
We will use the following equivalent expression for the $\psi_2$-norm (see, e.g., \cite[Lemma 2.4.2]{IsotropicConvexBodies}). To formulate it and to simplify notation, instead of $L_M(\Omega)$ for $M(x)=x^q$, $q\geq 1$, let us write $L_q(\Omega)$, and $\|\,\cdot\,\|_{L^q(\Omega)}$ for the corresponding norm.
\begin{proposition}\label{prop:psi 2 norm equivalence}
Let $(\Omega,\mathcal F, \bP)$ be a probability space and $f:\Omega\to\R$ be measurable. Then,
\[
\|f\|_{\psi_2}\approx \sup_{q\geq 2} \frac{\|f\|_{L_q(\Omega)}}{\sqrt{q}}\,.
\]
\end{proposition}
The latter will be used for the space $(\SSS_p^{n-1},\bM_{\B_p^n})$ and for the functionals $\langle\,\cdot\,,\theta \rangle:\SSS_p^{n-1}\to\R$ with $\theta\in\R^n$.
One of the main tools in our proof of Theorem \ref{thm:Main} will be the following Bernstein type inequality obtained in \cite[Proposition 1]{BLM88} (see also \cite[Lemma 11.4.6]{IsotropicConvexBodies}).
\begin{proposition}\label{prop:bernstein}
Let $\xi_1,\dots,\xi_N \in L_{\psi_2}(\Omega,\bP)$ be independent, mean zero random variables defined on a common probability space $(\Omega,\mathcal{F},\bP)$. Assume that $\|\xi_i\|_{\psi_2} \leq c$ for all $i=1,\dots,N$ and some $c\in(0,\infty)$. Then, for all $\varepsilon>0$,
\[
\bP \bigg( \Big| \sum_{i=1}^N \xi_i\Big| > \varepsilon N \bigg) \leq 2\exp\Big({-\frac{\varepsilon^2N}{8c^2}}\Big)\,.
\]
\end{proposition}
\subsection{$L_q$-centroid bodies}
Let $K\in\K^n$ be such that $\vol_n(K)=1$. For $q\geq 1$ the $L_q$-centroid body $Z_q(K)$ is the unique convex body that has support function given by
$$
h_{Z_q(K)}(u) = \bigg(\int_K|\langle x,u\rangle|^q\,\dint x\bigg)^{1/q}\,,\qquad u\in\R^n\,.
$$
We remark that, using the language of centroid bodies, the condition that a convex body $K\in\K^n$ is isotropic can be rephrased by saying that $Z_2(K)$ is a Euclidean ball. The study of $L_q$-centroid bodies, considered from an asymptotic point of view, was initiated by Paouris in \cite{Paouris2,Paouris}.
We recall from the work of Paouris \cite{P06} and of Klartag and Milman \cite{KM12} that the volume radius $\vol_n(Z_q(K))^{1/n}$ of the $L_q$-centroid body of $K$ satisfies the estimate
\begin{equation}\label{eq:volume centroid body}
\vol_n\big(Z_q(K)\big)^{1/n}\approx \sqrt{\frac{q}{n}}\,L_K\,,
\end{equation}
as long as $1\leq q\leq \sqrt{n}$ (compare with \cite[Theorem 2.2]{GHT}). Note that we are working with isotropic convex bodies and not with isotropic log-concave measures, whence the constant $L_K$ in Relation \eqref{eq:volume centroid body}. We remark that this improves upon the previous bound of Lutwak, Yang and Zhang (see \cite[Proposition 5.1.16]{AsymGeomAnalyBook}). We will see below that \eqref{eq:volume centroid body} will allow us to derive a lower bound for the volume radius of our random polytopes $K_N$ that holds with overwhelming probability.
\subsection{Geometry of $\ell_p$-balls}
Recall that for $1 \leq p < \infty$ and $n\in\N$, we denote by $\ell_p^n$ the space $\R^n$ equipped with the norm $\|x\|_p = (\sum_{i=1}^n |x_i|^p)^{1/p}$. The unit ball of $\ell_p^n$ is denoted by $\B_p^n$ and we let $\SSS_p^{n-1}$ be the unit sphere in $\ell_p^n$. It is convenient for us to write $\B^n$ and $\SSS^{n-1}$ instead of $\B^n_2$ and $\SSS_2^{n-1}$, respectively. The volume of $\B_p^n$ is given by
\[
\vol_n(\B_p^n) = \frac{\big(2\Gamma(1+\frac{1}{p})\big)^n}{\Gamma(1+\frac{n}{p})}\,,
\]
see \cite[page 180]{AsymGeomAnalyBook}. It follows directly from Stirling's formula that asymptotically, as $n\to\infty$, $\vol_n(\B_p^n)^{1/n} \approx n^{-1/p}$.
\smallskip
We rephrase the following result of Schechtman and Zinn \cite[Lemma 1]{SZ} (independently obtained by Rachev and R\"uschendorf in \cite{RR91}) that provides a probabilistic representation of the cone measure $\bM_{\B_p^n}$ of the unit ball of $\ell_p^n$ (see also \cite{BGMN} for an extension).
\begin{proposition}\label{thm:SZ}
Let $n\in\N$, $1\leq p < \infty$, and $g_1,\dots,g_n$ be independent real-valued random variables that are distributed according to the density
\[
f(t) = \frac{e^{-|t|^p}}{2\Gamma\big(1+{1\over p}\big)}, \qquad t\in\R\,.
\]
Consider the random vector $G=(g_1,\dots,g_n)\in\R^n$ and put $Y:=G/\|G\|_p$. Then $Y$ is independent of $\|G\|_p$ and has distribution $\bM_{\B^n_p}$.
\end{proposition}
Let us also recall the following result on moments of sums of independent random variables with log-concave tails that was originally obtained by Gluskin and Kwapie\'n in \cite[Theorem 1]{GK} and is stated here as in \cite[Proposition 7]{BGMN}, reflecting our particular set-up. In what follows, $\bE$ will denote expectation (integration) with respect to the probability measure $\bP$.
\begin{proposition}\label{thm:bgmn}
Let $n\in\N$, $1\leq p,q<\infty$ and $a=(a_i)_{i=1}^n\in\R^n$. Further, let the random vector $G$ be as in Proposition \ref{thm:SZ}. Then,
\begin{align}\label{eq:bgmn expression}
\big(\bE\,\big|\langle a,G \rangle\big|^q\big)^{1/q} \approx q^{1/p}\, \big\|(a_i^*)_{i=1}^q\big\|_{p^*} + \sqrt{q}\,\big\|(a_i^*)_{i=q+1}^n\big\|_2\,,
\end{align}
where $(a_i^*)_{i=1}^n$ is the non-increasing rearrangement of $(|a_i|)_{i=1}^n$ and $p^*$ is the conjugate of $p$ that is defined via the relation $\frac{1}{p}+\frac{1}{p^*}=1$.
\end{proposition}
\begin{remark}
The formulation of Proposition \ref{thm:bgmn} is taken from \cite{BGMN} and follows essentially from \cite[Theorem 1]{GK}, but requires computing and estimating the parameters that appear in \cite[Theorem 1]{GK}. The original formulation of the result contains a constant
$$
c_p=\bigg({1\over p}\bigg)^{1/p}\bigg({1\over p^*}\bigg)^{1/p^*}\,,
$$
see \cite[Remark 1]{GK}, and parameters $\kappa_1,\kappa_2$ depending on a convex function $N$ that determines the tail behaviour of the random variables involved. However, as can be seen in the proof of \cite[Proposition 7]{BGMN}, the equivalence in \eqref{eq:bgmn expression} holds with absolute constants. In the case that $q\geq n$, the right-hand side of \eqref{eq:bgmn expression} reduces to $q^{1/p}\|(a_i)_{i=1}^n \|_{p^*}$.
\end{remark}
\subsection{Two bounds for the average norm}
In order to bound the isotro\-pic constant of our random polytopes from above in the regime $2\leq p<\infty$, the following result is used. It was already applied in earlier papers concerning the isotropic constant of random polytopes and is explicitly written in \cite[Lemma 3.2]{ALT}.
\begin{proposition}\label{prop:IntegralBoundedBySup}
Let $1\leq n \leq N$ and $P=\conv\{ P_1,\dots,P_N\}$ be a non-degenerated symmetric polytope in $\R^n$. Then,
\[
\frac{1}{\vol_n(P)}\int_P \|x\|_2^2\,\dint x \leq \frac{1}{(n+1)(n+2)}\sup_{\substack{E\subseteq \{1,\dots,N\}\\|E|=n}}\Bigg[ \sum_{i\in E}\|P_i\|_2^2 + \Big\| \sum_{i\in E} P_i \Big\|_2^2\Bigg]\,.
\]
\end{proposition}
For other closely related versions of Proposition \ref{prop:IntegralBoundedBySup} we refer the reader, for example, to \cite{A,IsotropicConvexBodies,KK}.
\medskip
Finally, in the case that $1\leq p<2$ we shall use instead the following bound for the integral of the $1$-norm. We directly formulate the result for the symmetric random polytopes $K_N$.
\begin{proposition}\label{prop:IntegralBoundL1Version}
Let $K_N$ be the symmetric random polytope generated by $n+1\leq N\leq e^{\sqrt{n}}$ independent random points $X_1,\ldots,X_N$ that are distributed on $\SSS_p^{n-1}$ according to ${\bf m}_{\mathbb{B}_p^n}$. Then,
\begin{align*}
{1\over\vol_n(K_N)}\int_{K_N}\|x\|_1\,\dint x \leq \max_{F=\conv(\{X_{i_1},\ldots,X_{i_n}\})\atop\varepsilon_{i_1},\ldots,\varepsilon_{i_n}=\pm 1}\!\!{1+\sqrt{2}\over n}\,\|\varepsilon_{i_1}X_{i_1}+\ldots+\varepsilon_{i_n}X_{i_n}\|_1\,,
\end{align*}
where the maximum is taken over all facets $F=\conv(\{X_{i_1},\ldots,X_{i_n}\})$ of $K_N$ with vertices $\{X_{i_1},\ldots,X_{i_n}\}\subseteq\{X_1,\ldots,X_N\}$ and over all choices of signs $\varepsilon_{i_1},\ldots,\varepsilon_{i_n}\in\{-1,+1\}$.
\end{proposition}
\begin{proof}
This is a direct consequence of \cite[Lemma 11.5.4]{IsotropicConvexBodies} and \cite[Identity (2.26)]{DGG} together with the fact that the $\ell_p$-balls $\B_p^n$ are symmetric with respect to all coordinate hyperplanes.
\end{proof}
\section{Proof of Theorem \ref{thm:Main}}\label{sec:ProofThm}
\subsection{The starting point}
Assume that $K\in\K^n$. It follows from the definition \eqref{eq:IsotropicConstantDef} of the isotropic constant $L_K$ of $K$ that
\[
nL_K^2 \leq \frac{1}{\vol_n(K)^{2/ n}}{1\over\vol_n(K)}\int_K \|x\|_2^2\,\dint x\,,
\]
and, as we stated in \eqref{eq:IsotropicConstantL1Bound},
\[
nL_K \leq {c\over\vol_n(K)^{1/ n}}{1\over\vol_n(K)}\int_K\|x\|_1\,\dint x
\]
for any symmetric $K$, $c\in(0,\infty)$ being an absolute constant.
Therefore, to prove boundedness of the isotropic constant of our random polytopes $K_N = \conv\{\pm X_1,\dots,\pm X_N \}$ with vertices $\pm X_1,\ldots,\pm X_N\in\SSS_p^{n-1}$ chosen with respect to the cone measure $\bM_{\B_p^n}$ of $\B_p^n$ with overwhelming probability, we proceed in two steps:
\begin{itemize}
\item Step 1: derive a high probability lower bound for $$\vol_n(K_N)^{1/n}\,.$$
\item Step 2: derive a high probability upper bound for $$\frac{1}{\vol_n(K_N)}\int_{K_N} \|x\|_2^2\,\dint x\,$$
if $2\leq p < \infty$, and for
\[
{1\over\vol_n(K_N)}\int_{K_N}\|x\|_1\,\dint x
\]
if $1\leq p < 2$.
\end{itemize}
Carrying out Step 1 and Step 2 is the content of the next three subsections. We subdivide Step 2 into the cases $1\leq p < 2$ (see Subsection \ref{sec:Step2p<2}) and $2\leq p <\infty$ (see Subsection \ref{sec:Step2}) and combine everything in Subsection \ref{sec:EndOfProof} to complete the proof. It will in fact become evident from the proof of the case $2\leq p <\infty$ that the case $1\leq p <2$ has to be treated differently.
This method of bounding the isotropic constant of random polytopes has been introduced in \cite{KK} and then been further developed in \cite{A,DGG}. As already explained earlier, we shall adapt this route to our model, which in particular means that in the regime $1\leq p<2$ we have to establish a new bound on the $\psi_2$-behaviour of the functional $\langle \,\cdot\,,\theta\rangle$, $\theta\in\R^n$, with respect to the cone measure on an $\ell_p$-sphere.
\subsection{Step 1}\label{sec:Step1}
We first introduce another random polytope model. We do this in a general framework in order to distinguish arguments working for general underlying convex bodies from those valid only for $\ell_p$-spheres. For that purpose, let $K\in\K^n$ be isotropic and symmetric and assume that $n+1\leq N\leq e^{\sqrt{n}}$. Now, let $Y_1,\ldots,Y_N$ be independent random points that are selected according to the uniform distribution on $K$, that is, the restriction of the Lebesgue measure to $K$, and define the random polytope
\[
\widetilde{K}_N=\conv\big(\{\pm Y_1,\ldots,\pm Y_N\}\big)\,.
\]
It should be noted that, in contrast to the random polytopes $K_N$, not all of the points $\pm Y_1,\ldots,\pm Y_N$ are necessarily vertices of $\widetilde{K}_N$. It is one of the main results of the paper \cite{DGT1} that
\begin{equation}\label{eq:RandomPolyCentroid}
\widetilde{K}_N \supseteq c_1 Z_{\log(2N/n)}(K)
\end{equation}
with probability at least $1-e^{-c_2\sqrt{N}}$, where $c_1,c_2\in(0,\infty)$ are absolute constants and $N\geq c_0n$ for another absolute constant $c_0\in(0,\infty)$. Together with \eqref{eq:volume centroid body} this immediately implies that
\begin{equation}\label{eq:volume random polytope}
\vol_n\big(\widetilde{K}_N\big)^{1/n}\gtrsim \sqrt{\frac{\log(2N/n)}{n}}\, L_K
\end{equation}
with probability bounded from below by $1-e^{-c_2\sqrt{N}}$, whenever $N\geq c_0 n$. If we work with a linear number of points, that is, with $n\leq N \leq \gamma n$ for some $\gamma>1$, then this goes back to Pivovarov \cite[Proposition 2]{Piv} (see also \cite[Theorem 11.3.7]{IsotropicConvexBodies} for a version combining both regimes).
In the remaining part of this section, let $K_N$ be a random polytope in $K$ that arises as the symmetric convex hull of $N$ random points distributed according to the cone probability measure $\bM_K$ on $K$. Our goal is to construct a coupling of these random polytopes $K_N$ and the random polytopes $\widetilde{K}_N$ introduced above such that pathwise the set inclusion $K_N\supseteq\widetilde{K}_N$ holds. To construct such a coupling, we let, as above, $Y_1,\ldots,Y_N$ be independent random points selected according to the uniform distribution on $K$ and let $(\Omega,\mathcal{F},\bP)$ be the underlying probability space on which these random points are defined. Now, we apply to each point $Y_i$ the Minkowski map $\mu_K$ and define $X_1(\omega):=\mu_K(Y_1(\omega)),\ldots,X_N(\omega):=\mu_K(Y_N(\omega))$ for all $\omega\in\Omega$ for which $\|Y_i\|_K\neq 0$ for all $i\in\{1,\ldots,N\}$ (in the other case, we put $X_1(\omega)=\ldots=X_N(\omega)$ to be some arbitrary boundary point of $K$). The random points $X_1,\ldots,X_N$ are clearly independent and are located on the boundary of $K$. Next, we notice that the push-forward measure of the uniform distribution on $K$ is -- by its very definition -- precisely the cone measure $\bM_K$ of $K$, i.e., for every Borel set $B\subseteq\bd\,K$ we have that
\begin{align*}
\bP(X_i\in B) &= \bP(\mu_K(Y_i)\in B) = \bP(Y_i\in\mu_K^{-1}(B))=\bM_K(B)\,.
\end{align*}
In other words this means that the independent random points $X_1,\ldots,X_N$ are distributed according to the cone measure $\bM_K$, implying that the random convex hull $\conv\big(\{\pm X_1,\ldots,\pm X_N\}\big)$ has the same distribution as our random polytope $K_N$. From now on we understand by $K_N$ the random polytope that arises in the way just described. Moreover, by construction, we have that $K_N(\omega)\supseteq\widetilde{K}_N(\omega)$ for all $\omega\in\Omega_0$, where $\Omega_0\subseteq\Omega$ is a subset satisfying $\bP(\Omega_0)=1$. Using this coupling, we conclude from \eqref{eq:volume random polytope} that
\begin{equation}\label{eq:volume random polytopeFINAL}
\vol_n(K_N)^{1/n}\geq\vol_n(\widetilde{K}_N)^{1/n} \gtrsim\sqrt{\frac{\log(2N/n)}{n}}\, L_K
\end{equation}
on a subset of $\Omega$ with $\bP$-measure at least $1-e^{-c\sqrt{N}}$, where $c\in(0,\infty)$ is an absolute constant. This completes the proof of Step 1.
In the special case of the unit balls $\B_p^n$, we summarize the result in the following proposition. We emphasize that the additional factor $n^{-1/p}$ there just arises from the different volume normalization.
\begin{proposition}\label{prop:lower volume bound ell_p}
Let $1\leq p <\infty$ and $n+1\leq N \leq e^{\sqrt{n}}$. Assume that $X_1,\dots,X_N\in\SSS_p^{n-1}$ are independent random points with distribution $\bM_{\B_p^n}$ and write $K_N={\rm conv}(\{\pm X_1,\ldots,\pm X_N\})$. Then,
\[
\bP \Bigg( \vol_n(K_N)^{1/n} \geq c_1\frac{\sqrt{\log(2N/n)}}{n^{1/2+1/p}}\Bigg) \geq 1-e^{-c_2\sqrt{N}},
\]
where $c_1,c_2\in(0,\infty)$ are absolute constants.
\end{proposition}
\begin{remark}
In principle, the coupling just described also works for $p=\infty$, in which case $\B_\infty^n$ is nothing else than the cube $[-1,1]^n$ in $\R^n$. However, since in the second step below we can only treat the case $p<\infty$, we decided to formulate Proposition \ref{prop:lower volume bound ell_p} merely in this set-up.
\end{remark}
\subsection{Step 2 for $2\leq p<\infty$}\label{sec:Step2}
In view of the lower bound on the volume radius we obtained, to prove Theorem \ref{thm:Main} for $2\leq p<\infty$, it is enough to show that the inequality
\[
\frac{1}{\vol_n(K_N)}\int_{K_N} \|x\|_2^2\,\dint x \lesssim \frac{\log(2N/n)}{n^{2/p}}
\]
holds with high probability. To prove this, we will apply Bernstein's inequality. Thus, we first have to make sure that the functionals $n^{1/p}\langle X,\theta\rangle$, where $X$ is distributed according to $\bM_{\B_p^n}$ and $\theta\in \SSS^{n-1}$, satisfy the required $\psi_2$-bound. Here we make use of the Schechtman-Zinn and Gluskin-Kwapie\'n results, see Propositions \ref{thm:SZ} and \ref{thm:bgmn}, respectively. We will show the following.
\begin{proposition}\label{thm:psi_2 bound}
There exists an absolute constant $c_{\psi_2}\in(0,\infty)$ such that
\[
\| n^{1/p}\langle\, \cdot\,,\theta \rangle \|_{\psi_2} \leq c_{\psi_2}
\]
for all $2\leq p<\infty$, $n\in\N$ and $\theta\in\SSS^{n-1}$, where $\|\,\cdot\,\|_{\psi_2}$ is the norm on the Orlicz space $L_{\psi_2}(\SSS_p^{n-1},\bM_{\B^n_p})$.
\end{proposition}
Before we turn to the proof of Proposition \ref{thm:psi_2 bound}, we first prove a lemma that provides an equivalent expression for the $q$th moment of $\langle X,\theta \rangle$ that will be more convenient to work with and that is also used in the next subsection.
\begin{lemma}\label{lem:qth moment estimate}
For all $1\leq p,q<\infty$ and all $\theta=(\theta_1,\ldots,\theta_n)\in \R^n$, we have
\begin{align*}
&\bigg(\int_{\SSS_p^{n-1}} |\langle x,\theta \rangle|^q \,\dint\bM_{\B^n_p}(x)\bigg)^{1/q} \\
&\qquad\approx \left(\frac{\Gamma(\frac{n}{p})}{\Gamma(\frac{n+q}{p})}\right)^{1/q}\Big[ q^{1/p}\, \big\|(\theta_i^*)_{i=1}^q\big\|_{p^*} + \sqrt{q}\,\big\|(\theta_i^*)_{i=q+1}^n\big\|_2 \Big]\,,
\end{align*}
where $(\theta_i^*)_{i=1}^n$ is the non-increasing rearrangement of $(\theta_i)_{i=1}^n$ and $p^*$ is determined by the equation ${1\over p}+{1\over p^*}=1$.
\end{lemma}
\begin{proof}
Let $\theta \in \SSS^{n-1}$. Then, by Proposition \ref{thm:SZ} and with $G$ as given there,
\begin{align*}
\bE \,\big|\big\langle G,\theta \big\rangle\big|^q & = \bE\,\bigg[\|G\|_p^q \cdot \Big|\Big\langle \frac{G}{\|G\|_p},\theta \Big\rangle \Big|^q\bigg] \cr
& = \bE\, \|G\|_p^q \cdot \int_{\SSS_p^{n-1}}|\langle y,\theta \rangle|^q\,\dint\bM_{\B_p^n}(y).
\end{align*}
We now use the following polar integration formula to compute the $q$th moment of $\|G\|_p$. For $K\in\K^n$ one has that
\begin{align*}
\int_{\R^n}f(x)\,\dint x = n\,\vol_n(K)\int_0^\infty\int_{\bd\,K}f(ry)\,r^{n-1}\,\dint\bM_K(y)\dint r
\end{align*}
for all non-negative measurable functions $f:\R^n\to\R$ (in fact, this may alternatively be used as a definition for the cone measure $\bM_K$ of $K$). We apply this to the function
\[
f(x) = \frac{\|x\|_p^q\cdot e^{-\|x\|_p^p}}{\big(2\Gamma(1+1/p)\big)^n}\,,\qquad x\in\R^n,
\]
and the unit ball $\B^n_p$ of $\ell_p^n$, and obtain
\begin{equation*}
\begin{split}
\bE\, \|G\|_p^q &= \int_{\R^n} f(x) \, \dint x = \frac{n\,\vol_n(\B_p^n)}{(2\Gamma(1+\frac{1}{p}))^n} \, \int_0^\infty r^{n+q-1}e^{-r^p} \, \dint r \\
& = \frac{n\, \vol_n(\B_p^n) \, \Gamma(\frac{n+q}{p})}{p \, (2\Gamma(1+\frac{1}{p}))^n} = \frac{\Gamma(\frac{n+q}{p})}{\Gamma(\frac{n}{p})} \,,
\end{split}
\end{equation*}
where we also used the definition of the gamma function as well as the formula for $\vol_n(\B_p^n)$ stated in Section \ref{sec:prelim} above.
What is left to bound is the term $\bE\big[\big|\big\langle G,\theta \big\rangle\big|^q\big]$. This is done by means of Proposition \ref{thm:bgmn}, which yields that
\[
\bE \Big[\big|\big\langle G,\theta \big\rangle\big|^q\Big] \approx \Big[q^{1/p}\, \big\|(\theta_i^*)_{i=1}^q\big\|_{p^*} + \sqrt{q}\,\big\|(\theta_i^*)_{i=q+1}^n\big\|_2\Big]^q\,.
\]
This completes the proof.
\end{proof}
We can now prove the desired $\psi_2$-estimate for the linear functionals $\langle\,\cdot\,,\theta\rangle$, $\theta\in\SSS^{n-1}$, on the boundary of $\B_p^n$.
\begin{proof}[Proof of Proposition \ref{thm:psi_2 bound}]
By definition of an Orlicz norm, it is enough to show that
\[
\int_{\SSS_p^{n-1}} \exp\bigg(\frac{n^{2/p}|\langle x,\theta \rangle|^2}{c_{\psi_2}^2}\bigg) \,\dint\bM_{\B^n_p}(x) \leq 2\,.
\]
Using the series expansion of the exponential function and Lemma \ref{lem:qth moment estimate} applied with $2q$ instead of $q$ there, we obtain that
\begin{align*}
&\int_{\SSS_p^{n-1}} \exp\bigg(\frac{n^{2/p}|\langle x,\theta \rangle|^2}{c_{\psi_2}^2}\bigg) \,\dint\bM_{\B^n_p}(x)\\
& = \sum_{q=0}^\infty \frac{n^{2q/p}}{q!c_{\psi_2}^{2q}}\int_{\SSS_p^{n-1}} |\langle x,\theta \rangle|^{2q}\,\dint\bM_{\B^n_p}(x) \\
& \leq 1+ \sum_{q=1}^\infty c_1 \frac{n^{2q/p}}{q!c_{\psi_2}^{2q}}\,\frac{\Gamma\big(\frac{n}{p}\big)}{\Gamma\big(\frac{n+2q}{p}\big)} \, \Big[ (2q)^{1/p}\, \big\|(\theta_i^*)_{i=1}^{2q}\big\|_{p^*} + \sqrt{2q}\,\big\|(\theta_i^*)_{i=2q+1}^n\big\|_2 \Big]^{2q}\,,
\end{align*}
where $(\theta_i^*)_{i=1}^n$ is the non-increasing rearrangement of $(\theta_i)_{i=1}^n$ and $c_1\in(0,\infty)$ is an absolute constant.
Observe that, since $\theta\in \SSS^{n-1}$,
$$
\big\|(\theta_i^*)_{i=1}^{2q}\big\|_{p*}\leq(2q)^{{1\over p^*}-{1\over 2}}\big\|(\theta_i^*)_{i=1}^{2q}\big\|_2 \leq (2q)^{{1\over p^*}-{1\over 2}}
$$
if $1\leq p^*\leq 2$. In view of $1/p+1/p^*=1$, this implies that
$$
(2q)^{1/p}\, \big\|(\theta_i^*)_{i=1}^{2q}\big\|_{p^*} + \sqrt{2q}\,\big\|(\theta_i^*)_{i=2q+1}^n\big\|_2 \leq 2\sqrt{2q}\,,
$$
since $2\leq p<\infty$ has been assumed. Using Stirling's formula repeatedly, this yields
\begin{align*}
&\int_{\SSS_p^{n-1}} \exp\bigg(\frac{n^{2/p}|\langle x,\theta \rangle|^2}{c_{\psi_2}^2}\bigg) \,\dint\bM_{\B^n_p}(x)\\
&\qquad\leq 1+c_1\sum_{q=1}^\infty {n^{2q/p}\over q!c_{\psi_2}^{2q}}\big(2\sqrt{2q}\big)^{2q}\frac{ \Gamma\big(\frac{n}{p}\big)}{\Gamma\big(\frac{n+2q}{p}\big)}\\
&\qquad\leq 1+c_2\sum_{q=1}^\infty {n^{2q/p}\over q!c_{\psi_2}^{2q}}\big(2\sqrt{2q}\big)^{2q}\sqrt{\frac{n+2q}{n}}\Big(\frac{n}{n+2q}\Big)^{n/p}\Big(\frac{pe}{n+2q}\Big)^{2q/p}\\
&\qquad \leq 1+c_2\sum_{q=1}^\infty {(pe)^{2q/p}\over q!c_{\psi_2}^{2q}}\big(2\sqrt{2q}\big)^{2q}2\sqrt{q}\\
&\qquad \leq 1+c_3\sum_{q=1}^\infty {(2e)^{2q}\over c_{\psi_2}^{2q}}\big(2\sqrt{2q}\big)^{2q}\sqrt{q}\frac{1}{\sqrt{2\pi q}}\Big(\frac{e}{q}\Big)^q\\
&\qquad = 1+c_4\sum_{q=1}^\infty \bigg(\frac{32e^3}{c_{\psi_2}^2}\bigg)^q \,,
\end{align*}
where $c_2,c_3,c_4\in(0,\infty)$ are again absolute constants. We can achieve that the last expression is bounded from above by $2$, provided that the constant $c_{\psi_2}\in(0,\infty)$ is chosen large enough. This completes the proof.
\end{proof}
Let us proceed with the proof of Step 2 by establishing the following result.
\begin{proposition}\label{prop:upperbound}
Let $2\leq p<\infty$ and $n+1\leq N \leq e^{\sqrt{n}}$. Then,
\begin{align*}
&\bP\Bigg(\frac{1}{\vol_n(K_N)}\int_{K_N} \|x\|_2^2\,\dint x \leq c_1\,\frac{\log(2N/n)}{n^{2/p}}\Bigg) \\
&\hspace{4cm}\geq 1-3\exp\Big(-c_2n\log(2N/n)\Big),
\end{align*}
where $c_1,c_2\in(0,\infty)$ are absolute constants.
\end{proposition}
\begin{proof}
Since $2\leq p<\infty$, the result of Proposition \ref{thm:psi_2 bound} ensures that we can apply Bernstein's inequality to the independent and centered random variables $n^{1/p}\langle X_i,\theta\rangle$, $i\in\{1,\ldots,N\}$, where $\theta\in\SSS^{n-1}$ is a fixed direction. Let $E\subseteq \{1,\dots,N\}$ with $|E|=n$. Bernstein's inequality (see Proposition \ref{prop:bernstein}) implies that
\begin{align*}
\bP\bigg(\bigg|\sum_{i\in E}\langle X_i,\theta\rangle\bigg|\geq \varepsilon n\bigg)\leq 2\exp\bigg(-{\varepsilon^2n^{2/p+1}\over 8c_{\psi_2}^2}\bigg)
\end{align*}
for all $\varepsilon>0$. Now, let $\mathcal N$ be a $1\over 2$-net of $\SSS^{n-1}$ with cardinality $|{\mathcal N}|\leq 5^n$ (the existence of such a net is ensured by \cite[Lemma 5.2.5]{AsymGeomAnalyBook}, for example). Then the union bound yields that
\begin{align*}
\bP\bigg(\exists\theta\in\mathcal{N}:\bigg|\sum_{i\in E}\langle X_i,\theta\rangle\bigg|\geq \varepsilon n\bigg) &\leq 2\cdot 5^n\exp\bigg(-{\varepsilon^2n^{2/p+1}\over 8c_{\psi_2}^2}\bigg)\\
&=2\exp\bigg(-{\varepsilon^2n^{2/p+1}\over 8c_{\psi_2}^2}+n\log 5\bigg)
\end{align*}
or, equivalently,
\begin{align*}
\bP\bigg(\forall\theta\in\mathcal{N}:\bigg|\sum_{i\in E}\langle X_i,\theta\rangle\bigg|< \varepsilon n\bigg) \geq 1-2\exp\bigg(-{\varepsilon^2n^{2/p+1}\over 8c_{\psi_2}^2}+n\log 5\bigg)\,.
\end{align*}
Now we want to show that the length of $\sum_{i\in E} X_i$ is bounded from above by $c\varepsilon n$ with overwhelming probability, where $c\in(0,\infty)$ is an absolute constant. To this end, note that each $\theta\in\SSS^{n-1}$ has a representation of the form $\theta = \sum_{j=1}^\infty \delta_j \theta_j$ with $\theta_j\in\mathcal N$ and $0 \leq \delta_j \leq (\frac{1}{2})^{j-1}$, see \cite{A}. Observe that for any $\omega\in\Omega$ satisfying $|\sum_{i\in E} \langle X_i(\omega),\theta\rangle| \leq \varepsilon n$ for every $\theta\in\mathcal N$, we have that
\begin{align*}
\Big| \Big\langle \sum_{i\in E} X_i(\omega),\theta \Big\rangle \Big|
&= \Big| \Big\langle \sum_{i\in E} X_i(\omega),\sum_{j=1}^\infty \delta_j\theta_j \Big\rangle \Big| \\
&\leq \sum_{j=1}^\infty \delta_j \Big|\Big\langle \sum_{i\in E} X_i(\omega),\theta_j \Big\rangle \Big| \leq 2\varepsilon n
\end{align*}
for all $\theta\in\SSS^{n-1}$. Hence,
\begin{equation}\label{eq:small ball estimate euclidean norm of sum}
\begin{split}
\bP\bigg( \Big\|\sum_{i\in E} X_i \Big\|_2 \leq 2 \varepsilon n\bigg)
& = \bP \bigg( \sup_{\theta\in \SSS^{n-1}}\Big\langle\sum_{i\in E} X_i, \theta\Big\rangle \leq 2 \varepsilon n\bigg) \\
& \geq 1- 2\exp\bigg(-{\varepsilon^2n^{2/p+1}\over 8c_{\psi_2}^2}+n\log 5\bigg)\,.
\end{split}
\end{equation}
Next, we notice that, according to Proposition \ref{prop:IntegralBoundedBySup},
\begin{align*}
& \bP\Bigg(\frac{1}{\vol_n(K_N)}\int_{K_N} \|x\|_2^2\,\dint x \leq c\frac{\log(2N/n)}{n^{2/p}}\Bigg) \\
& \geq \bP\Bigg(\sup_{\substack{E\subseteq \{1,\dots,N\}\\|E|=n}}\Bigg[ \sum_{i\in E}\|X_i\|_2^2 + \Big\| \sum_{i\in E} X_i \Big\|_2^2\Bigg] \leq cn^{2-2/p}\log(2N/ n)\Bigg)\,,
\end{align*}
where $c\in(0,\infty)$ is a sufficiently large absolute constant. Now, for every $x\in \R^n$, we have that $\| x\|_2 \leq n^{1/2-1/p} \|x\|_p$ by H\"older's inequality and the fact that $2\leq p<\infty$. Thus,
\begin{align*}
& \bP\Bigg(\sup_{\substack{E\subseteq \{1,\dots,N\}\\|E|=n}}\Bigg[ \sum_{i\in E}\|X_i\|_2^2 + \Big\| \sum_{i\in E} X_i \Big\|_2^2\Bigg] \leq cn^{2-2/p}\log(2N/ n)\Bigg) \\
& \geq \bP\Bigg(\sup_{\substack{E\subseteq \{1,\dots,N\}\\|E|=n}} \Big\| \sum_{i\in E} X_i \Big\|_2^2 \leq cn^{2-2/p}\log(2N/ n)-n^{2-2/p}\Bigg) \\
& \geq \bP\Bigg(\sup_{\substack{E\subseteq \{1,\dots,N\}\\|E|=n}} \Big\| \sum_{i\in E} X_i \Big\|_2^2 \leq c_1n^{2-2/p}\log(2N/ n)\Bigg) \\
& = \bP\Bigg(\sup_{\substack{E\subseteq \{1,\dots,N\}\\|E|=n}} \Big\| \sum_{i\in E} X_i \Big\|_2 \leq \sqrt{c_1}n^{1-1/p}\sqrt{\log(2N/ n)}\Bigg)\,,
\end{align*}
where $c_1\in(0,\infty)$ is an absolute constant. Therefore, changing to the complementary event, using the union bound and applying \eqref{eq:small ball estimate euclidean norm of sum} with $\varepsilon=\sqrt{c_1}n^{-1/p}\sqrt{\log(2N/n)}/2$, together with Stirling's formula, we obtain
\begin{align*}
& \bP\Bigg(\frac{1}{\vol_n(K_N)}\int_{K_N} \|x\|_2^2\,\dint x \leq c\frac{\log(2N/n)}{n^{2/p}}\Bigg) \\
& \geq 1-2{{N}\choose{n}}\exp\Bigg(-\frac{c_1}{32c_{\psi_2}^2}n\log(2N/n)+n \log 5\Bigg) \\
& \geq 1-2\exp\Bigg(-\frac{c_1}{32c_{\psi_2}^2}n\log(2N/n)+ n\log(2eN/n) +n \log 5\Bigg) \\
& \geq 1-2 \exp\Big(-c_2n\log(2N/n)\Big)\,,
\end{align*}
where we choose the constant $c$ above so large that $c_1/(32c_{\psi_2}^2)\geq 1$ and put $c_2:=(1+2+\log 5)c_1/(32c_{\psi_2}^2)=c_1/(8c_{\psi_2}^2)$, for example.
\end{proof}
\subsection{Step 2 for $1\leq p<2$}\label{sec:Step2p<2}
The purpose of this section is to establish a high-probability upper bound for the expression
$$
{1\over\vol_n(K_N)}\int_{K_N}\|x\|_1\,\dint x
$$
and to prove Step 2 in the case that $K_N$ is a symmetric random polytope with vertices from $\SSS_p^{n-1}$ with $1\leq p<2$. To this end, we first establish a suitable substitute for Proposition \ref{thm:psi_2 bound}, which can be regarded as the analogue for the cone measure of $\B_p^n$ of a theorem of Bobkov and Nazarov \cite{BobkovNazarov}. In what follows, we shall write $\SSS_\infty^{n-1}$ for the unit sphere in $\R^n$ supplied with the $\ell_\infty$-norm $\|x\|_\infty=\max\{|x_i|:i=1,\ldots,n\}$, $x=(x_1,\ldots,x_n)\in\R^n$.
\begin{proposition}\label{prop:Psi_2boundForp<2}
There exists an absolute constant $c_{\psi_2}\in(0,\infty)$ such that, for every $1\leq p < 2$ and all $\theta\in\SSS_{\infty}^{n-1}$,
\[
\Big\|{1\over n^{1/2-1/p}}\langle\,\cdot\,,\theta \rangle\Big\|_{\psi_2} \leq c_{\psi_2}\,,
\]
where $\| \cdot\|_{\psi_2}$ is the norm on the Orlicz space $L_{\psi_2}(\SSS_p^{n-1},\bM_{\B_p^n})$.
\end{proposition}
\begin{proof}
Fix $1\leq p < 2$ and $\theta=(\theta_1,\dots,\theta_n)\in\SSS_{\infty}^{n-1}$. By Proposition \ref{prop:psi 2 norm equivalence}, we have that
\[
\|\langle\,\cdot\,,\theta \rangle\|_{\psi_2} \lesssim \sup_{q\geq 2}\frac{\|\langle\,\cdot\,,\theta \rangle\|_{q}}{\sqrt{q}}\,,
\]
where $\|\,\cdot\,\|_{q}$ is the $L_q$-norm on $\SSS_p^{n-1}$. Let us now consider the two cases $q<n$ and $q\geq n$ separately. From Lemma \ref{lem:qth moment estimate} we obtain that, for any $1\leq q <\infty$,
\begin{align*}
\|\langle\,\cdot\,,\theta \rangle\|_{q} & \lesssim \left(\frac{\Gamma(\frac{n}{p})}{\Gamma(\frac{n+q}{p})}\right)^{1/q}\Big[ q^{1/p}\, \big\|(\theta_i^*)_{i=1}^q\big\|_{p^*} + \sqrt{q}\,\big\|(\theta_i^*)_{i=q+1}^n\big\|_2 \Big]\cr
& \lesssim
\begin{cases}
\big[q+\sqrt{q}\,\sqrt{n-q}\,\big]\left(\frac{\Gamma(\frac{n}{p})}{\Gamma(\frac{n+q}{p})}\right)^{1/q} &: 2\leq q< n \\
q \left(\frac{\Gamma(\frac{n}{p})}{\Gamma(\frac{n+q}{p})}\right)^{1/q}&: q\geq n\,.
\end{cases}
\end{align*}
Here, we have estimated in the second step the two norms against the $\ell_\infty$-norm and used the fact that ${1\over p}+{1\over p^*}=1$ and that $\|\theta\|_\infty=1$. Applying Stirling's formula, we see that
\begin{align*}
\left(\frac{\Gamma(\frac{n}{p})}{\Gamma(\frac{n+q}{p})}\right)^{1/q} &\lesssim
n^{(\frac{n}{p}-\frac{1}{2})\frac{1}{q}}(n+q)^{(\frac{1}{2}-\frac{n+q}{p})\frac{1}{q}} \cr
& \lesssim (n+q)^{(\frac{n}{p}-\frac{1}{2})\frac{1}{q}}(n+q)^{(\frac{1}{2}-\frac{n+q}{p})\frac{1}{q}} = \frac{1}{(n+q)^{1/p}}\,.
\end{align*}
Now, we note that if $2\leq q<n$, $\frac{\|\langle\,\cdot\,,\theta \rangle\|_{q}}{\sqrt{q}}\lesssim{\sqrt{q}+\sqrt{n-q}\over(n+q)^{1/p}}$ is maximal for $q=2$, where it is $c_1n^{1/2-1/p}$ with some absolute constant $c_1\in(0,\infty)$. Similarly, if $q\geq n$, $\frac{\|\langle\,\cdot\,,\theta \rangle\|_{q}}{\sqrt{q}}\lesssim{\sqrt{q}\over (n+q)^{1/p}}$ is maximal for the choice $q=n$, where the expression reduces to $c_2n^{1/2-1/p}$ with another absolute constant $c_2\in(0,\infty)$. As a consequence, we find that
$$
\|\langle\,\cdot\,,\theta \rangle\|_{\psi_2} \lesssim n^{1/2-1/p}
$$
and the proof is complete.
\end{proof}
We can now proceed and derive the announced high-probability upper bound by partially following the strategy in \cite{DGG} (see also \cite[Chapter 11.5.1]{IsotropicConvexBodies}).
\begin{proposition}
Let $1\leq p<2$ and $n+1\leq N\leq e^{\sqrt{n}}$. Then,
\begin{align*}
&\bP\Bigg(\frac{1}{\vol_n(K_N)}\int_{K_N} \|x\|_1\,\dint x \leq c_1\,n^{{1\over 2}-{1\over p}}\sqrt{\log\Big({2N\over n}}\Big)\Bigg) \\
&\hspace{4cm}\geq 1-2\exp\Big(-c_2n\log(2N/n)\Big),
\end{align*}
where $c_1,c_2\in(0,\infty)$ are absolute constants.
\end{proposition}
\begin{proof}
Let $X_1,\ldots,X_n$ be independent and identically distributed on $\SSS_p^{n-1}$ according to the cone measure of $\B_p^n$, and fix $\theta\in\SSS_\infty^{n-1}$ as well as $\varepsilon_1,\ldots,\varepsilon_n\in\{-1,+1\}$. We define $g_j(X_1,\ldots,X_n):=\langle\varepsilon_jX_j,\theta\rangle$ for $j\leq n$ and notice that $\|g_j\|_{\psi_2}\lesssim n^{1/2-1/p}$ according to Proposition \ref{prop:Psi_2boundForp<2} and since $\B_p^n$ is symmetric with respect to all coordinate hyperplanes. Now, using Bernstein's inequality (see Proposition \ref{prop:bernstein}), we conclude that
\begin{align*}
\bP\bigg(\Big|\sum_{j=1}^ng_j(X_1,\ldots,X_n)\Big|>t\,n\bigg) &=\bP\big(|\langle\varepsilon_1X_1+\ldots+\varepsilon_nX_n,\theta\rangle|>t\,n\big)\\
&\leq 2\exp\big(-c\,t^2\,n^{2/p}\big)
\end{align*}
for all $t>0$ and with an absolute constant $c\in(0,\infty)$.
Thus, choosing $t=c_1n^{1/2-1/p}\sqrt{\log\big({2N\over n}\big)}$ with some absolute constant $c_1\in(0,\infty)$, we have that
\begin{equation}\label{eq:ProofL1Gleichung1}
\begin{split}
&\bP\Bigg(|\langle\varepsilon_1X_1+\ldots+\varepsilon_nX_n,\theta\rangle|>c_1n^{{3\over 2}-{1\over p}}\sqrt{\log\Big({2N\over n}\Big)}\,\Bigg)\\
&\qquad\qquad\qquad\leq 2\exp\Big(-c_2n\log\Big({2N\over n}\Big)\Big)
\end{split}
\end{equation}
with $c_2=c\cdot c_1$. Next, we notice that
$$
\sup_{\theta\in\SSS_\infty^{n-1}}|\langle\varepsilon_1X_1+\ldots+\varepsilon_nX_n,\theta\rangle| = \sup_{\theta\in{\rm extr}(\SSS_\infty^{n-1})}|\langle\varepsilon_1X_1+\ldots+\varepsilon_nX_n,\theta\rangle|\,,
$$
where ${\rm extr}(\SSS_\infty^{n-1})$ is the set of extreme points of $\SSS_\infty^{n-1}$. Clearly, any such extreme point has the representation $\theta=(\theta_1,\ldots,\theta_n)$ with $\theta_1,\ldots,\theta_n\in\{-1,+1\}$. Thus, using \eqref{eq:ProofL1Gleichung1} and the union bound, we find that
\begin{align*}
&\bP\Bigg(\max_{\varepsilon_j=\pm 1}\|\varepsilon_1X_1+\ldots+\varepsilon_nX_n\|_1 > c_3n^{{3\over 2}-{1\over p}}\sqrt{\log\Big({2N\over n}\Big)}\Bigg)\\
&=\bP\Bigg(\max_{\varepsilon_j=\pm 1}\sup_{\theta\in\SSS_\infty^{n-1}}|\langle\varepsilon_1X_1+\ldots+\varepsilon_nX_n,\theta\rangle| > c_3n^{{3\over 2}-{1\over p}}\sqrt{\log\Big({2N\over n}\Big)}\Bigg)\\
&=\bP\Bigg(\max_{\varepsilon_j=\pm 1}\sup_{\theta\in\{-1,+1\}^n}|\langle\varepsilon_1X_1+\ldots+\varepsilon_nX_n,\theta\rangle| > c_3n^{{3\over 2}-{1\over p}}\sqrt{\log\Big({2N\over n}\Big)}\Bigg)\\
&\leq\sum_{\varepsilon_1,\ldots,\varepsilon_n\in\{-1,+1\}\atop \theta\in\{-1,+1\}^n}\bP\Bigg(|\langle\varepsilon_1X_1+\ldots+\varepsilon_nX_n,\theta\rangle| > c_3n^{{3\over 2}-{1\over p}}\sqrt{\log\Big({2N\over n}\Big)}\Bigg)\\
&\leq 2^n\cdot 2^n\cdot 2\exp\Big(-c_2n\log\Big({2N\over n}\Big)\Big)\\
&= 2\exp\Big(-c_2n\log\Big({2N\over n}\Big)+n\log 4\Big)\\
&\leq 2\exp\Big(-c_4n\log\Big({2N\over n}\Big)\Big)
\end{align*}
with an absolute constant $c_4\in(0,\infty)$, whenever the constant $c_1$ above is chosen sufficiently large. Equivalently, this can be re-phrased by saying that
\begin{align*}
&\bP\Bigg(\max_{\varepsilon_j=\pm 1}\|\varepsilon_1X_1+\ldots+\varepsilon_nX_n\|_1 \leq c_3n^{{3\over 2}-{1\over p}}\sqrt{\log\Big({2N\over n}\Big)}\Bigg)\\
&\qquad\qquad\qquad\geq 1-2\exp\Big(-c_4n\log\Big({2N\over n}\Big)\Big)\,.
\end{align*}
The result follows now from Proposition \ref{prop:IntegralBoundL1Version}, since each facet of $K_N$ has the same distribution as the convex hull $\conv(\{X_1,\ldots,X_n\})$ of $X_1,\ldots,X_n$.
\end{proof}
\subsection{Completion of the proof}\label{sec:EndOfProof}
Using the estimates we obtained in Step 1 and Step 2, we will now complete the proof of Theorem \ref{thm:Main}. Recall that for every $K\in\K^n$,
\[
nL_K^2 \leq \frac{1}{\vol_n(K)^{2/ n}}{1\over\vol_n(K)}\int_K \|x\|_2^2\,\dint x\,.
\]
Thus, by combining this with Proposition \ref{prop:lower volume bound ell_p} and Proposition \ref{prop:upperbound}, we can find absolute constants $c_1,c_2,c_3\in(0,\infty)$ such that for all $2\leq p<\infty$, $n+1\leq N\leq e^{\sqrt{n}}$ and all random polytopes $K_N = \conv\{\pm X_1,\dots,\pm X_N \}$ with vertices $\pm X_1,\ldots,\pm X_N$ chosen with respect to the cone measure $\bM_{\B_p^n}$ from $\SSS_p^{n-1}$,
\begin{align*}
\bP\big(L_{K_N} \leq c_1\big) & \geq \bP\left(\frac{1}{\vol_n(K_N)^{2/ n}}{1\over\vol_n(K_N)}\int_{K_N} \|x\|_2^2\,\dint x \leq n\,c_1^2 \right)\\
& \geq 1- e^{-c_2\sqrt{N}} - 2e^{-c_3n\log(2N/n)}\,.
\end{align*}
Now, recall that
\[
nL_K \leq {c\over\vol_n(K)^{1+{1/ n}}}\int_K\|x\|_1\,\dint x
\]
for any symmetric $K$, where $c\in(0,\infty)$ is an absolute constant. Hence, in the situation $1\leq p<2$, combining Proposition \ref{prop:lower volume bound ell_p} with Proposition \ref{prop:Psi_2boundForp<2}, we conclude that there are absolute constants $c_4,c_5,c_6\in(0,\infty)$ such that for all $1\leq p<2$, $n+1\leq N\leq e^{\sqrt{n}}$ and all random polytopes $K_N = \conv\{\pm X_1,\dots,\pm X_N \}$ with vertices $\pm X_1,\ldots,\pm X_N$ chosen with respect to the cone measure $\bM_{\B_p^n}$ from $\SSS_p^{n-1}$,
\begin{align*}
\bP\big(L_{K_N} \leq c_4\big) & \geq \bP\left(\frac{1}{\vol_n(K_N)^{1/ n}}{1\over\vol_n(K_N)}\int_{K_N} \|x\|_1\,\dint x \leq n\,c_4 \right)\\
& \geq 1- e^{-c_5\sqrt{N}} - 2e^{-c_6n\log(2N/n)}\,.
\end{align*}
This completes the proof of Theorem \ref{thm:Main}.\hfill $\Box$
\section*{Acknowledgement}
We would like to thank David Alonso-Guti\'errez and Apostolos Giannopoulos for useful conversations and interesting hints and remarks. We would also like to thank an anonymous referee for many helpful suggestions and especially for pointing us to an error in an earlier version of this manuscript. The financial support of the Mercator Research Center Ruhr has made possible a research stay of the second author at Ruhr University Bochum.
\bibliographystyle{plain}
|
\section{Introduction}
Recently, the ATLAS and CMS collaborations have reported
the results of the LHC Run-2 at
center-of-mass energy $\sqrt{s}=13$~TeV,
based on the integrated luminosity of
3.2~fb$^{-1}$ and 3.3~fb$^{-1}$, %
respectively
\cite{LHC-diphoton}.
Both the collaborations have shown
a possible excess in the events containing two photons,
suggesting the existence of a new $s$-channel resonance particle $\phi$.
The distribution of the observed events at ATLAS favours
a mass of the resonance $M\approx 750$~GeV, and
a width-to-mass ratio $\Gamma/M\approx 0.06$ with
a local (global) significance of 3.9~$\sigma$ (2.3~$\sigma$).
In the assumption of a narrow width,
the corresponding local (global) significance is 3.6~$\sigma$ (2.0~$\sigma$).
The CMS collaboration has also reported
a similar excess at $M \approx 760$~GeV with
a local (global) significance of $2.9\sigma$ ($<1 \sigma$),
and the event distribution slightly favours a narrow width.
A combined analysis of the CMS Run-1 (8~TeV) and Run-2 data showed that
the local (global) significance of the diphoton excess increases to
3.4~$\sigma$ (1.6~$\sigma$) with
the best-fit diphoton invariant mass close to $750$~GeV
\cite{LHC-diphoton-CMS-1603}.
If the two photons arise directly from the decay of the resonance $\phi$,
the resonance must be electrically neutral, and
its spin can be 0 or 2 due to the Landau-Yang theorem
\cite{LangdauYang}.
Assuming a large width,
the ATLAS (CMS) data favour a diphoton production cross section
$10\pm 3~\mbox{fb}$ ($6\pm 3~\mbox{fb}$)
\cite{
Franceschini:2015kwy%
}.
Other analyses assuming narrow width give
$\sim 6.2 \ (5.6)$~fb %
for ATLAS (CMS)
\cite{
Buttazzo:2015txu,%
Falkowski:2015swt%
}.
The LHC diphoton excess, if confirmed,
is a clear indication of new physics beyond the standard model (SM).
Furthermore, $\phi$ is unlikely to be the only new particle.
Since $\phi$ is electrically neutral,
it can only couple to photons through loop processes.
If the loops involve only the SM charged particles,
$\phi$ should decay into these SM particles with large rates,
as $\phi$ is much heavier than all the SM particles.
The corresponding production cross sections can easily reach $\mathcal{O}(\mbox{pb})$
which are too large to escape the detection at LHC Run-1
(see e.g.
\cite{
Low:2015qep,
Agrawal:2015dbf%
}).
If the large width $\Gamma/M\approx 0.06$ favoured by ATLAS is confirmed,
the resonance $\phi$ is likely to have additional tree-level invisible decays.
An intriguing possibility is that
$\phi$ also couples to the dark matter (DM) particles
which contribute to $\sim 26.8\%$ of the energy budget of our Universe.
In this scenario,
$\phi$ plays the role of a portal connecting
the invisible and visible world.
The excess of diphoton events suggests that
the DM particle should at least couple to photons,
and also couple to gluons or quarks depending on
the production mechanism of $\phi$ at the LHC.
The phenomenological implications such as the DM relic density,
DM direct and indirect detections have been extensive studied
\cite{
Mambrini:2015wyu,
Backovic:2015fnp,
Franceschini:2015kwy,%
Bi:2015uqd,
Barducci:2015gtd,
Bai:2015nbs,
Han:2015cty,
Bauer:2015boy,
Dev:2015isx,
Davoudiasl:2015cuo,%
D'Eramo:2016mgv,%
Chu:2012qy,%
Park:2015ysf%
}.
If the DM particles can couple to the SM particles indirectly,
the annihilation of the DM particles in the Galactic halo can generate
extra flux of cosmic-ray particles and photons.
Compared with the cosmic ray charged particles,
the photons are not deflected by the Galactic magnetic fields and
do not loss energy during the propagation in the Galactic halo.
Thus they are of crucial importance in searching for
the signals of halo DM annihilation.
The Galactic Center (GC) is expected to harbour high densities of DM,
as suggested by N-body simulations,
which makes it a promising place to
look for photon signals of DM annihilation or decay.
Recently,
a number of groups including Ferm-LAT collaboration have independently
found statistically strong evidence for an excess in cosmic gamma-ray fluxes at
energy $\sim2$~GeV towards
the inner regions around the Galactic center (GC) from the data of Fermi-LAT
\cite{Goodenough:2009gk,
Hooper:2010mq,
Boyarsky:2010dr,
Abazajian:2010zy,
Hooper:2011ti,
Abazajian:2012pn,
Gordon:2013vta,
Macias:2013vya,
Abazajian:2014fta,%
Hooper:2013rwa,
Huang:2013pda,
Daylan:2014rsa,
Zhou:2014lva,%
Calore:2014xka,%
TheFermi-LAT:2015kwa%
}.
The morphology of this GC excess (GCE) emission is consistent with
a spherical emission profile expected from DM annihilation.
The origin of the GCE is still under debate.
There exists plausible astrophysical explanations
such as the unresolved point sources of mili-second pulsars
\cite{
Abazajian:2010zy,%
Hooper:2011ti,%
Abazajian:2012pn,%
Gordon:2013vta,%
Bartels:2015aea,%
Lee:2015fea%
}
and
the interactions between the cosmic rays and the molecular gas
\cite{Macias:2013vya,%
Abazajian:2014fta,%
Gaggero:2015nsa%
}.
Halo DM annihilation can also provide a reasonable explanation.
The determined energy spectrum of the excess emission
although depending on the choices of diffuse gamma-ray background templates,
is in general compatible with the scenario of
$\sim40$~GeV DM particles self-annihilating into $b \bar b $ final states with
a cross section
$\langle \sigma v\rangle \approx (1-2)\times 10^{-26}\text{ cm}^{3}\text{s}^{-1}$
close to the typical thermal cross section for the observed DM relic aboundence
\cite{
Hooper:2013rwa,
Daylan:2014rsa%
}
(other possible final states were considered in Refs.
\cite{
Calore:2014xka,
Agrawal:2014oha,
Chu:2012qy%
}).
The possible connection between the LHC diphoton excess and
the GCE was first explored in
\cite{
Huang:2015svl%
}.
Assuming a pesudoscalar $\phi$ which couples dominantly to
$gg$, $\gamma\gamma$ and scalar DM particles,
it was shown that the two reported photon excesses can be
simultaneously explained
if the total width of $\phi$ is large enough $\Gamma/M\gtrsim\mathcal{O}(10^{-2})$ which is favoured by the current ATLAS data.
The phenomenological consequences of such a connection was further discussed in
Refs.~\cite{
Hektor:2016uth%
} and
\cite{
Krauss:2016cdi%
}.
A large total width of $\phi$, if confirmed, implies that
the new physics sector is strongly coupled, or
the resonance $\phi$ has large number of decay channels.
In this work, we investigate the generic conditions for
a consistent explanation for possible the LHC diphoton excess and GCE,
especially the requirement on total width of $\phi$ in
a wide range of DM models where
the DM particle can be scalar, fermionionic and vector, and
$\phi$ can be generated by
$s$-channel gluon fusion or quark-antiquark annihilation at parton level.
We show that the minimally required $\phi$ width is determined by
a single parameter proportional to $(m_{\chi}/M)^{n}$,
where the integer $n$ depends on
the spins of the DM particle and its decay final states.
We find that
for scalar DM model with $\phi$ generated from $q\bar q$ annihilation,
the minimally required $\Gamma/M$ can be as low as $\mathcal{O}(10^{-3})$.
For scalar DM model with $\phi$ generated from $gg$ fusion and
fermionic DM model with $\phi$ from $q\bar q$ annihilation,
the required $\Gamma/M$ reaches $\mathcal{O}(10^{-2})$.
Other models such as the vector DM model requires larger $\Gamma/M$ of order one
which is already disfavoured by the current data.
For the same DM model, the required width of $\phi$ is always smaller
in $q\bar q$ channel than that in the $gg$ channel.
For the DM models which can simultaneously account for the diphoton excess and the GCE,
the predicted cross sections for gamma-ray line are typically of
$\mathcal{O}(10^{-30})~\text{cm}^{3}\text{s}^{-1}$,
which is close to the current limits imposed by the Fermi-LAT data.
These models can be distinguish by LHC and Fermi-LAT in the near future.
The rest of this paper is organized as follows.
In section 2,
we overview the interpretation of the diphoton excess, and
derive model-independent conditions for a consistent explanation for
the diphoton excess and the GCE.
In section 3,
we discuss model-independently the implications of the GCE for the DM properties.
In section 4, we determine the allowed parameters in various DM models in which
the DM particles can be scalar, fermionic and vector with $\phi$ generated by $gg$ fusion
and $q\bar q$ annihilation.
The conclusion is given in section 5.
\section{The LHC diphton excess}
We consider the simplest scenario where
the diphoton events are produced from
the decay of the $s$-wave resonance $\phi$ which
is generated through $X\bar{X}$ fusion or annihilation process,
where $X\bar X=gg$, $\gamma\gamma$ and $q\bar q$ ($q=u,d,c,s,t,b$).
The production cross section for
the process $pp\to\phi\to\gamma\gamma$ in
the narrow-width approximation is given by
\begin{align}\label{eq:diphoton}
\sigma_{\gamma\gamma}=\frac{2J+1}{(\Gamma/M)s}
\left(\sum_{X}C_{X\bar{X}}\frac{\Gamma_{X\bar{X}}}{M}\right)
\left(\frac{\Gamma_{\gamma\gamma}}{M}\right),
\end{align}
where
$J$ is the spin of $\phi$, and
the coefficients $C_{X\bar{X}}$ incorporate the integration over
the parton distribution functions of the protons.
For instance,
at the center-of-mass energy $\sqrt{s}=13\ (8)$ TeV,
$C_{gg}\approx2137\ (174)$,
$C_{b\bar{b}}\approx15.3\ (1.07)$ and
$C_{c\bar{c}}\approx36\ (2.7)$
\cite{Franceschini:2015kwy}.
Higher order QCD corrections can be taken into account by including
the $K$-factors with typical values $K_{gg\ (q\bar{q})}\approx1.48\ (1.20)$.
For the sake of simplicity,
we consider the case where $\phi$ is spin zero, and
one channel of $X\bar X$ dominates the $\phi$ production at a time.
The process of $\gamma\gamma$ fusion is always included,
as it is irreducible.
In the limit of $\Gamma_{X\bar X}\gg \Gamma_{\gamma\gamma}$,
the values of the partial decay widths required to account for
the diphoton excess at Run-2 are estimated as
\begin{align}
\left(\frac{\Gamma_{X\bar{X}}}{M}\right)
\left(\frac{\Gamma_{\gamma\gamma}}{M}\right)
\approx \frac{2.1\times 10^{-4}}{C_{X\bar{X}}}
\left(\frac{\sigma_{\gamma\gamma}}{8~\mbox{fb}}\right)
\left(\frac{\Gamma/M}{0.06}\right) .
\end{align}
The non-observation of any excess at Run-1 (8~TeV)
already imposes stringent limits on the cross sections for
a number of final states generated from the decay of a generic resonance
\begin{equation}\label{eq:run-1-limits}
\begin{tabular}{ccc}
$\sigma_{Z\gamma} \leq 4.0~\text{fb}$\cite{Aad:2014fha},
$\sigma_{ZZ} \leq 12~\text{fb}$\cite{Aad:2015kna},
$\sigma_{WW} \leq 40~\text{fb}$\cite{Khachatryan:2015cwa,Aad:2015agg},
\\
$\sigma_{\gamma\gamma} \leq 1.5~\text{fb}$
\cite{
Aad:2015mna%
},
$\sigma_{jj} \leq 2.5~\text{pb}$\cite{Aad:2014aqa},
$\sigma_{b\bar b} \leq 1.0~\text{pb}$
\cite{
Khachatryan:2015tra%
},
\end{tabular}
\end{equation}
The enhancement of the production cross section at Run-2 relative to
that at Run-1 can be described by the gain factor
$r=\sigma_{\text{13TeV}}/\sigma_{\text{8TeV}}
\approx 0.38 C_{X\bar X}(13\text{TeV})/C_{X\bar X}(8\text{TeV})$.
In order to account for an excess seen at Run-2 but not Run-1,
a large value of $r$ is favoured.
The production channels with leading $r$ factors are
$r_{b\bar{b}}\approx 5.4$, $r_{gg}\approx 4.7$, and
$r_{c\bar{c}}\approx 5.1$.
Other channels have smaller gain factors,
for instance,
$r_{ss}\approx 4.3$,
$r_{dd} \approx 2.7$,
$r_{uu}\approx 2.5$, and
$r_{\gamma\gamma} \approx 1.9$.
Thus they are not considered further in this work.
In the case where $\phi$ also couples to DM particles,
the total width of $\phi$ is given by
\begin{align}\label{eq:total-width}
\Gamma & =\Gamma_{gg(q\bar{q})}+\kappa\Gamma_{\gamma\gamma}+\Gamma_{\chi\chi},
\end{align}
where the factor $\kappa=(1+\Gamma_{ZZ}/\Gamma_{\gamma\gamma}+\Gamma_{Z\gamma}/\Gamma_{\gamma\gamma}+\Gamma_{WW}/\Gamma_{\gamma\gamma})$
absorbs the contributions from
$ZZ$, $Z\gamma$ and $WW$ final states,
which depends on the couplings
between $\phi$ and the SM weak gauge bosons in a given model.
If the total width $\Gamma$ can be determined by the experiment,
\eq{eq:total-width} can place an important constraint on
the properties of the DM particle.
If the diphoton events are generated dominantly by
the process of gluon fusion (quark-antiquark annihilation)
$gg\ (q\bar q) \to \phi \to \gamma\gamma$,
the cross sections of diphoton production and DM annihilation are
strongly correlated,
as the DM particles inevitably annihilate into these states through the
same intermediate state,
$\chi\bar\chi \to \phi\to gg, (q\bar q), \gamma \gamma$.
The same DM annihilation process determines
both the DM relic density and the DM indirect detection signals.
For the $s$-channel DM annihilation process $\chi\chi\to\phi\to X\bar{X}$,
the corresponding thermally-averaged
product of the DM annihilation cross section and the DM relative velocity can be
written in a generic form
\begin{align}\label{eq:sigmav}
\langle\sigma v\rangle_{X\bar{X}} & =\frac{8\pi\eta_{\chi}R_{X\bar{X}}}{s_{\chi}^{2}\left[\left(1-4m_{\chi}^{2}/M^{2}\right)^{2}+(\Gamma/M)^{2}\right]m_{\chi}^{2}\beta_{\chi}(M^{2})}\left(\frac{\Gamma_{\chi\chi}}{M}\right)\left(\frac{\Gamma_{X\bar X}}{M}\right),
\end{align}
where
$\eta_{\chi}=2\ (1)$ for the DM particle (not) being its own antiparticle,
$s_{\chi}$ is the spin degrees-of-freedom of the DM particle with
$s_{\chi}=$1, 2 and 3 for the DM being a scalar, fermion and vector,
respectively.
The quantity $\beta_{X}(s)\equiv(1-4m_{X}^{2}/s)^{1/2}$
is the velocity of the particle $X$ from the decay $\phi^{(*)}\to X\bar{X}$
with a squared center-of-mass energy $s$.
The function $R_{X\bar{X}}$ is essentially
the ratio of $\phi$ decay squared amplitudes at $s\approx4m_{\chi}^{2}$ and $M^{2}$
\begin{align}
R_{X\bar{X}}(m_{\chi}^{2}/M^{2}) & =\frac{\sum|M_{\phi\to\chi\chi}(s=4m_{\chi}^{2})|^{2}\sum|M_{\phi\to X\bar{X}}(s=4m_{\chi}^{2})|^{2}\beta_{X}(4m_{\chi}^{2})}{\sum|M_{\phi\to\chi\chi}(s=M^{2})|^{2}\sum|M_{\phi\to X\bar{X}}(s=M^{2})|^{2}\beta_{X}(M^{2})}.
\end{align}
For a consistent explanation of the diphoton excess and the GCE,
\eqs{eq:diphoton}, (\ref{eq:total-width}) and (\ref{eq:sigmav}) must be satisfied simultaneously.
The corresponding solutions for the $\phi$ partial decay widths
in the limits $m_{\chi}/M\ll 1$ and $\Gamma/M\ll 1$
are given by
\begin{align}\label{eq:twoSolutions}
\left(\frac{\Gamma_{X\bar X}}{M}\right)=\frac{1}{2}\left[\left(\frac{\Gamma}{M}\right)\pm\Delta^{1/2}\right], & \left(\frac{\Gamma_{\text{\ensuremath{\gamma\gamma}}}}{M}\right)=\frac{\sigma_{\gamma\gamma}s(\Gamma/M)}{C_{X\bar X}(\Gamma_{X\bar X}/M)} ,
\end{align}
where
\begin{align}\label{eq:delta}
\Delta
\equiv
\left(\frac{\Gamma}{M}\right)^{2}
-4\left[
\frac{s_{\chi}^{2}m_{\chi}^{2}\beta_{\chi}(M^{2})\langle\sigma v\rangle_{X\bar X}}{8\pi\eta_{\chi}R_{X\bar X}}
+\frac{\kappa\sigma_{\gamma\gamma}s}{C_{X\bar X}}\frac{\Gamma}{M}
\right] .
\end{align}
The necessary condition for the existence of the solutions is $\Delta \geq 0$.
As it can be seen in the following sections,
in most DM models $R_{X\bar X}\propto (m_{\chi}/M)^{2n},\ (n=1,2,3,\dots)$.
Since we are interested in the case of GCE where
$m_{\chi}\ll M$,
the second term in the square brackets in \eq{eq:delta} can be
safely neglected.
In a good approximation, the condition can be written as
\begin{align}\label{eq:totalWidthLimit}
\frac{\Gamma}{M}
\gtrsim
\left(
\frac{s_{\chi}^{2}m_{\chi}^{2}\beta_{\chi}\langle\sigma v\rangle_{X\bar X}}{2\pi\eta_{\chi}R_{X\bar X}}
\right)^{1/2} .
\end{align}
For a given DM model,
the factor $R_{X\bar X}$ is fixed.
If the diphoton excess is consistent with the DM thermal relic density
which is set by DM annihilation into $X\bar X$,
then the annihilation cross section must be close to
the typical thermal cross section
$\langle\sigma v\rangle_{X\bar{X}}
\approx
\langle\sigma v\rangle_{F}
=
3\times 10^{-26}\text{ cm}^{3}\text{s}^{-1}$.
From the value of $\Gamma/M$ determined by the experiment,
one can derived an upper limit on $\langle\sigma v\rangle_{X\bar{X}}$
as a function of $m_{\chi}$ from \eq{eq:totalWidthLimit},
which depends on the nature of the DM particle and
the final state $X\bar X$.
On the other hand,
if the diphoton excess is required to be consistent with the GCE,
since both $\langle\sigma v\rangle_{X\bar{X}}$ and
$m_{\chi}$ can be determined by the GCE data,
\eq{eq:totalWidthLimit} can lead to a minimal requirement on
the total width $\Gamma/M$.
The diphoton excess suggests that
the DM particles inevitably annihilate into two-photon final states through
$s$-channel $\phi$ exchange,
which results in a spectral line in the generated gamma-ray flux with
photon energy centered at $E_{\gamma}=m_{\chi}$.
The spectral line is difficult to be mimicked by conventional astrophysical
contributions, if observed, can be a strong evidence for halo DM annihilation or decay.
If the diphoton excess is generated from $X\bar X$ initial states,
from \eq{eq:diphoton} and (\ref{eq:sigmav}), it follows that
\begin{align}
\langle\sigma v\rangle_{\gamma\gamma}
=\frac{\sigma_{\gamma\gamma}}{\sigma_{X\bar X}}
\frac{R_{\gamma\gamma}}{R_{X\bar X}}
\langle\sigma v\rangle_{X\bar X} ,
\end{align}
where $\sigma_{X\bar X}$ is the cross section for
the production of $X\bar X$ final states through
intermediate state $\phi$ from $X\bar X$ fusion or annihilation at the LHC,
i.e. $X\bar X\to \phi \to X\bar X$.
For a given DM model,
the values of $R_{\gamma\gamma}/R_{X\bar X}$ is fixed.
Thus, from the Run-1 upper limit on $\sigma_{X\bar X}$,
one can obtain a $lower$ limit on $\langle\sigma v\rangle_{\gamma\gamma}$.
It was shown in Ref.~\cite{Huang:2015svl} that
a lower limit of $\langle \sigma v \rangle_{\gamma\gamma}\gtrsim ~4.8\times 10^{-30}~\mbox{cm}^{3}\mbox{s}^{-1} $ can be obtained in a scalar DM model with $\phi$ generated through $gg$
fusion.
If $\phi$ is allowed to couple to $Z\gamma$,
the DM annihilation can generate a gamma-ray line with
photon energy at
$E_{\gamma} = m_{\chi} (1-m_{Z}^{2}/4m_{\chi}^{2})$.
The annihilation cross section for the process $\chi\chi\to\phi\to Z\gamma$ is related to
that for $\chi\chi\to\phi\to \gamma\gamma$ as follows
\begin{align}
\langle\sigma v\rangle_{Z\gamma}
=
\frac{\sigma_{Z\gamma}}{\sigma_{\gamma\gamma}}
\frac{\tilde\beta^{6}_{Z}(4m^{2}_{\chi})}{\tilde\beta^{6}_{Z}(M^{2})}
\langle\sigma v\rangle_{\gamma\gamma} ,
\end{align}
where $\tilde\beta_{X}(s)=(1-m_{X}^{2}/s)^{1/2}$.
Since $\sigma_{Z\gamma}/\sigma_{\gamma\gamma}$ is a known in a give model,
a lower limit on $\langle \sigma v \rangle_{Z\gamma}$ can be obtained
in a similar way.
\section{The Galactic Center Excess}
The annihilation of DM particles into $X\bar X$ final states generates
diffuse gamma rays with a broad energy spectrum due to hadronization,
while the annihilation into $\gamma\gamma$ generates
a line-shape spectrum with energy centered at the DM particle mass.
Both the signatures are under active searches by
the current DM indirect detection experiments.
The differential gamma-ray flux,
averaged over a solid angle $\Delta\Omega$ is given by
\begin{align}\label{eq:fluxAvg}
\frac{d\Phi}{dE}=
\frac{\eta_{\chi} \rho_{0}^{2} r_{\odot} }{16\pi }
\frac{\langle\sigma v\rangle}{m_{\chi}^{2}}
\frac{dN_{\gamma}}{dE} J ,
\end{align}
where
$r_{\odot} \approx 8.5$~kpc is the distance from the Sun to the GC,
$\rho_{0} \approx 0.4\mbox{ GeV}/\mbox{cm}^{3}$ is
the local DM density in the solar neighbourhood, and
$dN_{\gamma}/dE$ is the gamma-ray spectrum per DM annihilation.
The dimensionless $J$-factor which contains the information of
DM density distribution is given by
\begin{align}
J=\int\frac{d\Omega}{\Delta\Omega}\int_{\text{l.o.s}}
\left( \frac{\rho(r)}{\rho_0} \right)^{2}
\frac{ds}{r_{\odot}} ,
\end{align}
where
$\rho(r)$ is the spatial distribution of halo DM energy density,
with $r$ the distance to the GC.
The integration is to be performed over the distance $s$
along the light-of-sight which is related to $r$ through the relation
$r^{2}=r_{\odot}^{2}+s^{2}-2sr_{\odot}\cos\psi$, where $\psi$ is
the angle of the direction away from the GC.
N-body simulations suggest an universal DM density profile of
the Navarro-Frenk-White (NFW) form \cite{Navarro:1996gj}
\begin{align}
\rho(r)=\rho_{s}\left(\frac{r}{r_{s}}\right)^{-\gamma}\left[1+\left(\frac{r}{r_{s}}\right)^{\alpha}\right]^{\frac{\gamma-\beta}{\alpha}} ,
\end{align}
which is characterized by the parameters $\alpha$, $\beta$, $\gamma$,
and a reference scale $r_{s}\simeq20$~kpc. For the standard NFW
profile, $\alpha=\gamma=1$ and $\beta=3$.
The normalization factor $\rho_{s}$ is determined by
the local DM density
$\rho(r_{\odot})=\rho_{0}$.
We determine the favoured values of
$m_{\chi}$ and $\langle\sigma v\rangle_{X\bar X}$ for
a number of annihilation final states such as
$gg$, $b\bar b$, $c \bar c$ and $u\bar u$,
from fitting to the GCE data derived in Ref.~\cite{Calore:2014xka}.
In total there are 24 data points.
The spectra of the prompt gamma rays $dN_{\gamma}/dE$ for
DM annihilating into $X\bar X$ are generated
by the Monte-Carlo simulation package Pythia 8.201~\cite{Sjostrand:2007gs}.
For the considered final states,
the contributions from the inverse Compton scatterings can be safely neglected.
We choose a modified NFW profile with an inner slope $\gamma=1.26$,
as suggested by the observed morphology of the gamma-ray emission
\cite{Abazajian:2012pn,Gordon:2013vta,Hooper:2013rwa,Calore:2014xka}.
Making use of \eq{eq:fluxAvg},
the calculated diffuse gamma-ray fluxes are averaged over
a square region of interest (ROI) $20^{\circ}\times20^{\circ}$ in
the sky with latitude $|b|<2^{\circ}$ masked out.
The corresponding $J$-factor is $J=57.6$.
The best-fit DM particle masses and annihilation cross sections,
and the corresponding $\chi^{2}$ and $p$-values are summarized in \tab{tab:GCEdm}.
\begin{table}\begin{center}
\begin{tabular}{ccccc}
\hline\hline
Channel & $m_\chi$ (GeV) & $\langle\sigma v\rangle_{bb}$ $(10^{-26}\text{cm}^3\text{s}^{-1}) $ & $\chi^2_{\text{min}}/\text{d.o.f.} $ & $p$-value \\%\tabincell{c}{$m_\chi$\\[GeV]} \tabincell{c}{$\langle\sigma v\rangle_{bb}$ \\ $ [10^{-26}\text{cm}^3\text{/s}] $}
\hline
$b\bar{b}$ & $46.15^{+5.81}_{-3.53}$ & $1.42^{+0.18}_{-0.17}$ & 24.572/22 & 0.32 \\
$c\bar{c}$ & $35.54^{+3.10}_{-4.12}$ & $0.95^{+0.12}_{-0.12}$ & 25.626/22 & 0.27 \\
$u\bar{u}$ & $22.26^{+2.83}_{-1.91}$ & $0.62^{+0.10}_{-0.08}$ & 28.495/22 & 0.16 \\
$gg$ & $ 62.01^{+6.56}_{-6.35}$ & $1.96^{+0.26}_{-0.24}$ & 24.665/22 & 0.31 \\
\hline\hline
\end{tabular}
\caption{
Values of DM mass and annihilation cross sections determined from
fitting to the GCE data.
The DM particle is assumed to be its own antiparticle.
}\label{tab:GCEdm}
\end{center}\end{table}
In \fig{fig:GCEfit}, we show the contours of the allowed regions for
the parameters $m_{\chi}$ and $\langle \sigma v\rangle_{X\bar X}$ at
$68\%$ and $95\%$ C.L. for two parameters,
corresponding to $\Delta \chi^{2}=2.3$ and 6.0, respectively.
As can be seen from the table,
in the DM interpretation of the GCE,
the required DM particle mass is in the range $\sim(20-70)$~GeV with
a cross section $(0.5-2)\times 10^{-26}\text{cm}^{3}\text{s}^{-1}$.
The most favoured channel is $b\bar b$.
We emphasize that the $gg$ channel also gives reasonably good fit with
a larger DM mass $\sim 60$~GeV,
which is crucial for a consistent explanation with the diphoton excess,
as $gg$ fusion is also the favoured channel for
the production of $\phi$ at the LHC Run-2.
These results are in good agreement with
the previous analysis in Ref.~\cite{Calore:2014nla}.
At present,
the most stringent constraints on the DM annihilation cross sections
are provide by the Fermi-LAT data on the diffuse gamma rays
of the dwarf spheroidal satellite galaxies (dSphs)
\cite{
Ackermann:2015zua%
}.
These limits are also shown in \fig{fig:GCEfit} for comparison purpose,
where the limits on $gg$ channel was derived using
a conservative rescaling approach detailed in Ref.
\cite{
Huang:2015svl%
}.
It is known that there is an apparent tension between
the GCE favoured regions and the Fermi-LAT limits.
Note that the DM velocity dispersion in the Galactic halo is
quite different from that in the dSphs.
The DM annihilation cross section favoured by the GCE data and
constrained by the gamma rays of dSphs can only be compared under
the assumption that the cross section is velocity independent,
which is in general not the case.
In the analysis of the Fermi-LAT collaboration,
the uncertainties in the $J$-factors were taken into account assuming
a NFW type parametrization of the DM density profile.
A recent analysis directly using the spherical Jeans equations rather than taking a
parametric DM density profile as input showed that
the $J$-factor can be smaller by a factor about $2-4$ for
the case of Ursa Minor,
which relaxes the constraints on the DM annihilation cross section to
the same amount
\cite{
Ullio:2016kvy%
}.
The annihilation of halo DM also generates cosmic-rayparticles
such as protons/antiprotons, electrons/positrons and neutrinos.
Compared with gamma-rays,
the predictions for the flux of cosmic-ray charged particles from DM annihilation suffer from
large uncertainties in the cosmic-ray propagation models.
For a DM particle mass below $\sim 100$~GeV,
the predicted $\bar p/p$ ratio peaks at lower energies below $\sim10$~GeV,
which suffer from additional uncertainties due to the solar activities.
The upper limits on the DM annihilation cross section
from the AMS-02 and PAMELA data on $\bar p/p$ ratio for various channels
have been studied for variuos propagation models
and DM density profiles ( see e.g.
\cite{
Jin:2015sqa,%
Jin:2015mka,%
Hooper:2014ysa,
Jin:2012jn,%
Jin:2014ica,%
Jin:2013nta%
}).
In general, the obtained limits are weaker than that derived from
the gamma rays of dSphs.
The constraints from the cosmic-ray positrons depends strongly on
the annihilation final states.
For leptonic final states such as $e^{+}e^{-}$ and $\mu^{+}\mu^{-}$,
the derived upper limits from the AMS-02 positron flux can reach the typical thermal cross section
for DM particle mass below 50--100~GeV
\cite{
Ibarra:2013zia%
}.
But for hadronic final states such as $b\bar b$, the corresponding limits are rather weak,
typically at $\mathcal{O}(10^{-24})~\text{cm}^{3}\text{s}^{-1}$.
The $gg$ final state generates a softer positron spectrum in comparison with
the $b\bar b$ final states.
Thus the corresponding limits are expected to be even weaker.
\begin{figure}[!htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{GCESigmav_Scalar_gg.eps}
\includegraphics[width=0.45\textwidth]{GCESigmav_Scalar_qq.eps}
\\
\includegraphics[width=0.45\textwidth]{GCESigmav_Majorana_gg.eps}
\includegraphics[width=0.45\textwidth]{GCESigmav_Majorana_qq.eps}
\caption{
Upper panels)
Left:
regions of DM mass and annihilation cross section allowed by the GCE
data at $68\%$ and $95\%$ C.Ls.
for the scalar DM particle annihilating into
$gg$ final states through $\phi$ exchange.
Upper limits on the annihilation cross section as a function of DM particle mass
for fixed values of $\Gamma/M$ are shown.
See text for detailed explanations.
The horizontal line indicates the typical thermal annihilation cross section of
$3\times 10^{-26}~\text{cm}^{3}\text{s}^{-1}$.
The $95\%$ C.L. upper limits from the Fermi-LAT data on the gamma rays from dSphs
\cite{
Ackermann:2015zua%
}
are also shown.
The limits on $gg$ channel was derived using
a conservative rescaling approach detailed in
\cite{
Huang:2015svl%
}.
Right: the same as left but for DM annihilation into $b\bar b$, $c\bar c$ and $u\bar u$ final states.
Lower pannels)
The same as upper panels but for Majorana fermionic DM model.
}
\label{fig:GCEfit}
\end{center}
\end{figure}
\section{DM Models}
We focus on the scenario where
the 750 GeV resonance $\phi$ is a pseudo-scalar particle.
For a scalar resonance, the related couplings are severely constrained by
the null results of the DM direct detection experiments.
The UV origins of the pseudoscalar $\phi$ can be
axion-like particles from the breaking of the Peccei-Quinn symmetry
\cite{PQsym},
pesudo-Goldstone boson from composite Higgs models
\cite{
Gripaios:2009pe%
},
or from the extended Higgs sectors of the SM
\cite{
Angelescu:2015uiz,
Han:2015qqj,
Huang:2015rkj,
Moretti:2015pbj,
Badziak:2015zez,
Han:2016bus,
Han:2016bvl,
Gopalakrishna:2016tku,
Ge:2016xcq%
},
or left-right symmetric models
\cite{
Dey:2015bur,
Dasgupta:2015pbr,
Borah:2016uoi,
Huong:2016kpa,
Dev:2015vjd,%
Guo:2011zze,Guo:2010sy,Guo:2010vy,Guo:2008si,Wu:2007kt
}.
If $\phi$ is a SM singlet and couples to the SM gauge bosons through
vector-like heavy fermions which have
small mixings with the SM fermions,
the constraints from the oblique parameters $S$ and $T$,
the EW precision test,
and the flavor physics can be relaxed
\cite{
Ellis:2015oso%
}.
A pseudo-scalar does not mix with the SM Higgs boson,
and is less constrained by the measured properties of the Higgs boson.
For fermionic DM, its annihilation into SM particles through $s$-channel
pseudoscalar exchange is not suppressed by the low velocity dispersion of
the halo DM.
Since $\phi$ is a pseudo-scalar,
the cross sections for DM-nucleus scattering through
quarks or gluons within the nucleus are
either velocity suppressed or vanishing,
which easily relaxes the stringent upper limits from
various DM direct detection experiments.
We assume that $\phi$ can couple to the SM quarks directly and
the SM gauge fields indirectly through loop processes (see e.g. Refs.~\cite{
others%
}).
Since $\phi$ is much heavier than the electroweak (EW) scale,
we start with EW gauge-invariant effective interactions up to dimension-five
\begin{align}
\mathcal{L}\supset &
\frac{1}{2}\partial_{\mu} \phi \partial^{\mu} \phi
-\frac{1}{2}M^{2}\phi^{2}
-i y_{q}\bar q \gamma^{5} q \phi
-\frac{g^{2}_{1}}{2\Lambda}\phi B_{\mu\nu}\tilde{B}^{\mu\nu}
\nonumber\\
& -\frac{g^{2}_{2}}{2\Lambda}\phi W_{\mu\nu}^{}\tilde{W}^{\mu\nu}
-\frac{g^{2}_{g}}{2\Lambda}\phi G_{\mu\nu}^{}\tilde{G}^{\mu\nu},
\end{align}
where for the gauge fields
$\tilde{F}_{\mu\nu}=\frac{1}{2}\epsilon_{\mu\nu\alpha\beta}F^{\alpha\beta}$,
$g_{1,2,g}$ are the dimensionless effective coupling strengths,
$y_{q}$ is the Yukawa coupling strength, and
$\Lambda$ is a common energy scale.
After the EW symmetry breaking,
the interaction terms involving physical EW gauge bosons
$A$, $Z$ and $W$ are given by
\begin{align}
\mathcal{L}_{\text{gauge}}
\supset &
-\frac{g^{2}_{A}}{2\Lambda}\phi A_{\mu\nu}\tilde{A}^{\mu\nu}
-\frac{g^{2}_{Z}}{2\Lambda}\phi Z_{\mu\nu}\tilde{Z}^{\mu\nu}
-\frac{g^{2}_{AZ}}{2\Lambda}\phi A_{\mu\nu}\tilde{Z}^{\mu\nu}
\nonumber\\
&-\frac{g^{2}_{W}}{2\Lambda}\phi W_{\mu\nu}\tilde{W}^{\mu\nu}
-\frac{g^{2}_{g}}{2\Lambda}\phi G_{\mu\nu}\tilde{G}^{\mu\nu},
\end{align}
where the physical gauge couplings $g_{A}$, $g_{Z}$, $g_{ZA}$ and
$g_{W}$ are related to that in the gauge basis as
\begin{align}
g^{2}_{A} &=g^{2}_{1}c_{W}^{2}+g^{2}_{2}s_{W}^{2},
\quad
g^{2}_{Z}=g^{2}_{1}s_{W}^{2}+g^{2}_{2}c_{W}^{2},
\nonumber\\
g^{2}_{ZA} &=2s_{W}c_{W}(g^{2}_{2}-g^{2}_{1}),
\quad
g^{2}_{W}=g^{2}_{2}
\end{align}
with $s_{W}^{2}=1-c_{W}^2=\sin^{2}\theta_{W}\approx 0.23$.
For the three extreme cases: $g_{1}=0$,
$g_{1}=g_{2}$ and $g_{2}=0$, the partial widths of $ZZ$, $Z\gamma$
and $WW$ relative to that of $\gamma\gamma$ and the values of $\kappa$
are listed in \tab{tab:couplings}.
We should focus on the case of $g_{2}=0$,
namely $\phi$ is not charged under the $SU(2)_{L}$ gauge group.
Note that the case of $g_{1}=0$ is severely constrained by the Run-I data on
the $Z\gamma$ and $ZZ$ production rates,
as it can be seen from \eq{eq:run-1-limits} and \tab{tab:couplings}.
The related phenomenology in the case of $g_{1}=g_{2}$ is similar to
that in the case of $g_{2}=0$,
except that the DM annihilation into $Z\gamma$ is forbidden.
Thus there is no gamma-ray line generated from $Z\gamma$ final states.
The partial decay widths for $\phi$ decaying into the SM gauge bosons and
the fermions are given by
\begin{align}\label{eq:widths}
\frac{\Gamma_{\gamma\gamma}}{M}
&=
\pi\alpha_{A}^{2}\left(\frac{M}{\Lambda}\right)^{2}, \
\frac{\Gamma_{gg}}{M}=8\pi\alpha^{2}_{g}\left(\frac{M}{\Lambda}\right)^{2},
\
\frac{\Gamma_{q\bar q}}{M}=\frac{3y_{q}^{2}\beta_{q}(M^{2})}{8\pi} ,
\end{align}
respectively,
where $\alpha_{A,g}=g^{2}_{A,g}/4\pi$.
For the spin nature of DM particles,
we consider three classes of models
where the DM particles can be scalar, fermionic and vector.
\begin{table}\begin{center}
\begin{tabular}{ccccc}
\hline\hline
models & $\Gamma_{ZZ}/\Gamma_{\gamma\gamma}$ & $\Gamma_{Z\gamma}/\Gamma_{\gamma\gamma}$ & $\Gamma_{WW}/\Gamma_{\gamma\gamma}$ & $\kappa$\tabularnewline
\hline
$g_{1}=0$ & 10 & 6.4 & 35 & 53\tabularnewline
\hline
$g_{1}=g_{2}$ & 0.9 & 0 & 1.9 & 3.8\tabularnewline
\hline
$g_{2}=0$ & 0.081 & 0.57 & 0 & 1.7 \tabularnewline
\hline\hline
\end{tabular}
\caption{
Ratios of $\phi$ decay widths
$\Gamma_{ZZ}/\Gamma_{\gamma\gamma}$,
$\Gamma_{Z\gamma}/\Gamma_{\gamma\gamma}$,
$\Gamma_{WW}/\Gamma_{\gamma\gamma}$
and the value of $\kappa$ defined in \eq{eq:total-width} for
three cases of $\phi$ couplings with the SM gauge bosons,
$g_{1}=0$, $g_{1}=g_{2}$ and $g_{2}=0$, respectively.
}\label{tab:couplings}
\end{center}\end{table}
\subsection{Real scalar DM}
In the real scalar DM model,
the Lagrangian for the DM particle $\chi$ and its interaction with $\phi$ is given by
\begin{align}
\mathcal{L}
\supset
\frac{1}{2} \partial_{\mu} \chi \partial^{\mu} \chi
-\frac{1}{2}m_{\chi}^{2}\chi^{2}
-\frac{1}{2}g_{\chi}\phi\chi^{2} ,
\end{align}
where $g_{\chi}$ is a dimensionful coupling strength, and
we have only included the most relevant interaction term.
Other possible interaction terms such as $\lambda\phi^{2}\chi^{2}/4$
are neglected by assuming small couplings.
For DM annihilation into $gg$ final states,
the corresponding factor $R_{gg}$ defined in \eq{eq:sigmav} is given by
\begin{align}
R_{gg}=16\left(\frac{m_{\chi}}{M}\right)^{4} ,
\end{align}
which is typically of $\mathcal{O}(10)^{-4}$ for $m_{\chi}\approx 60$~GeV.
In this case,
\eq{eq:totalWidthLimit} can be rewritten as
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{scalar},gg}
\geq
\frac{\beta_{\chi}^{1/2}(M^{2})\langle \sigma v\rangle_{gg}^{1/2} M^{2}}{8\pi^{1/2}m_{\chi}} .
\end{align}
For a given value of $\Gamma/M$,
the above inequality can be interpreted as
the upper limit on $\langle \sigma v\rangle_{gg}$ as a function of $m_{\chi}$.
In the upper-left panel of \fig{fig:GCEfit},
we show this relation for three choices of
$\Gamma/M=0.06$, 0.03 and 0.02, respectively.
If $\langle \sigma v\rangle_{gg}$ is required to meet the thermal value $\langle \sigma v\rangle_{F}$,
it can be seen that the DM particle mass has to be
larger than $\sim 65$ GeV (90 GeV) for $\Gamma/M=0.03$ (0.02).
While for $\Gamma/M=0.06$, the constraint on the DM particle mass is rather weak.
If $\Gamma/M$ is not fixed,
using the best-fit values of $m_{\chi}$ and $\langle \sigma v\rangle_{gg}$ for
$gg$ channel from \tab{tab:GCEdm},
a lower limit on the required total width of $\phi$ can be obtained as follows
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{scalar},gg}
\gtrsim
0.026
\left( \frac{M}{750~\text{GeV}}\right)^{2}
\left( \frac{62~\text{GeV}}{m_{\chi}}\right)
\left( \frac{\langle \sigma v\rangle_{gg}}{2.0\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2} .
\end{align}
Thus the GCE required typical minimal width-to-mass ratio is
quite large of $\mathcal{O}(10^{-2})$,
which is currently favoured by ATLAS, and
can be confirmed or ruled out soon by the upcoming LHC updated results.
Assuming $\phi$ is generated dominantly by $gg$ fusion,
a combined fit to both the data of diphton excess and GCE in
this model has been carried out in Ref.
\cite{
Huang:2015svl%
},
which showed that
the total width of $\phi$ is dominated by $\Gamma_{\chi\chi}$, and
the favoured partial widths $\Gamma_{gg}/M$ and $\Gamma_{\gamma\gamma}/M$ are of $\mathcal{O}(10^{-3})$ and $\mathcal{O}(10^{-5})$, respectively.
According to \eq{eq:twoSolutions},
there are actually two solutions for $\Gamma_{gg}/M$.
We update the previous analysis by considering wider ranges of parameters
and including the contribution from photon fusion
which is non-negligible when $\Gamma_{gg}/M$ is below $\mathcal{O}(10^{-4})$.
In \fig{fig:contour-scalar-gg},
we show the regions of the partial decay widths allowed by
the diphoton excess and GCE in wide ranges of parameter space in
$(\Gamma_{gg}/M,\Gamma_{\gamma\gamma}/M)$ and
$(\Gamma_{\chi\chi}/M,\Gamma_{\gamma\gamma}/M)$ planes.
The allowed regions are at $68\%$ and $95\%$ C.Ls. for two parameters,
corresponding to $\Delta \chi^{2}=2.3$ and 6.0, respectively,
together with the allowed regions by each individual experiment,
for the case of $\Gamma/M=0.06$ and 0.03.
It can be clearly seen that there is another solution located at
$\Gamma_{gg}/M\approx \Gamma/M$ which corresponds to the case where
the total width is dominated by gluon final states.
However, this solution is ruled out by the limit on the dijet production at Run-1,
as can be seen from the figure.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_gg_006_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_gg_006_xy.eps}
\\
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_gg_003_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_gg_003_xy.eps}
\caption{
Upper panels)
Left: Allowed regions at
$60\%$ and $95\%$ C.L. in
$(\Gamma_{gg}/M, \Gamma_{\gamma\gamma}/M)$ plane from
a combined fit to both the LHC diphoton excess and the GCE in
a scalar DM model with $gg$ fusion,
together with the regions allowed by each individual experiment.
The upper limits from Run-1 on the dijet and diphoton production
cross sections are also shown.
The total width is fixed at $\Gamma/M=0.06$.
Right: The same as upper-left,
but in $(\Gamma_{\chi\chi}/M, \Gamma_{\gamma\gamma}/M)$ plane.
Lower panels) The same as upper pannels but for $\Gamma/M=0.03$.
}
\label{fig:contour-scalar-gg}
\end{center}
\end{figure}
If the diphoton events are generated from parton level $q\bar q$ annihilation,
the situation is quite different.
For $q\bar q$ annihilation channel,
the function $R_{q\bar q}$ is given by
\begin{align}
R_{q\bar{q}}=4\frac{\beta_{q}(4m_{\chi}^{2})}{\beta_{q}(M^{2})} \left(\frac{m_{\chi}}{M}\right)^{2} ,
\end{align}
Since $R_{q \bar q}$ is proportional to $(m_{\chi}/M)^{2}$ instead of
$(m_{\chi}/M)^{4}$,
the lower limit on $\Gamma/M$ can be much smaller.
For DM annihilation dominantly into $b\bar b$ final states,
\eq{eq:totalWidthLimit} can be rewritten as
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{scalar},b\bar b}
\gtrsim
\frac{\beta_{\chi}^{1/2}(M^{2})\langle \sigma v\rangle_{b\bar b}^{1/2} M }{4\pi^{1/2}} .
\end{align}
Note that for $b\bar b$ final states, it is determined by
$\langle \sigma v\rangle_{b\bar b}$ alone,
as the leading $m_{\chi}$ dependence cancels out in \eq{eq:totalWidthLimit}.
This observation holds for all the $q\bar q$ final states.
In the upper-right panel of \fig{fig:GCEfit},
we show the maximally allowed value of
$\langle \sigma v\rangle$ as a function of DM particle mass
for three choices of $\Gamma/M=0.006$, 0.003 and 0.002, respectively.
If $\langle \sigma v\rangle_{b\bar b}$ is required to be equal to $\langle \sigma v\rangle_{F}$,
we find that the required $\Gamma/M$ should be above $\sim$ 0.006,
which is insensitive to the DM particle mass.
Using the best-fit values of $m_{\chi}$ and
$\langle \sigma v\rangle_{b\bar b}$ for $b\bar b$ channel in \tab{tab:GCEdm},
the corresponding minimal value of the width-to-mass ratio is found to be
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{scalar}, b\bar b}
\gtrsim
3.6\times 10^{-3}
\left( \frac{M}{750~\text{GeV}}\right)
\left( \frac{\langle \sigma v\rangle_{b\bar b}}{1.4\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2} .
\end{align}
Similarly, for the $c\bar c$ channel, the minimal width is given by
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{scalar}, c\bar c}
\gtrsim
3.0\times 10^{-3}
\left( \frac{M}{750~\text{GeV}}\right)
\left( \frac{\langle \sigma v\rangle_{c\bar c}}{0.95\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2} .
\end{align}
Thus for $q\bar q $ annihilation, the required minimal width-to-mass ratio
can be reduced to $\mathcal{O}(10^{-3})$, an order of magnitude lower
than that in the case of $gg$ fusion.
We perform analogous $\chi^{2}$ fits to
the data of diphoton excess and the GCE in
$b\bar b$ and $c\bar c$ channels
to determine the allowed values of the the parameters
$m_{\chi}$, $\Gamma_{\chi\chi}/M$, $\Gamma_{q\bar q}/M$ and
$\Gamma_{\gamma\gamma}/M$ for two typical values of total width
$\Gamma/M=0.06$ and 0.006, respectively.
For the diphoton excess,
we take a naively weighted average of ATLAS and CMS results
$\sigma_{\gamma\gamma}=8\pm2.1~\text{fb}$.
The Run-1 limits on dijet and diphoton productions are taken into account.
For the fit to the GCE, the data and the selection of the region of interest
in the sky are the same as the fit in Sec.~3.
The results of the best-fit values and uncertainties of these parameters are
summarized in \tab{tab:fitCombined}.
Compared with the fits to the GCE data alone,
there are no significant changes in the determined DM particle mass.
The values of $\chi^{2}/\text{d.o.f}$ are also comparable,
which indicates that the diphoton excess and GCE can be consistently
explained in this model.
The allowed regions for the partial decay widths at
$68\%$ and $95\%$ C.Ls.,
together with the allowed regions by each individual experiment,
for the case of $\Gamma/M=0.06\ (0.006)$ are shown in \fig{fig:contour-scalar-qq-0.06} (\fig{fig:contour-scalar-qq-0.006}).
For the $q\bar q$ annihilation channels,
the two solutions of \eq{eq:twoSolutions} can be seen
as the two well-separated regions characterized by
\begin{align}
\frac{\Gamma_{\chi\chi}}{M} \approx \frac{\Gamma}{M}, \quad
\frac{\Gamma_{q\bar q}}{M} \ll \frac{\Gamma_{\chi\chi}}{M} ,
& \qquad\qquad \text{(i)}
\nonumber\\
\frac{\Gamma_{q\bar q}}{M} \approx \frac{\Gamma}{M}, \quad
\frac{\Gamma_{\chi\chi}}{M} \ll \frac{\Gamma_{q\bar q}}{M} .
& \qquad\qquad \text{(ii)}
\end{align}
The solution (i) corresponds to case of DM dominance while
the solution (ii) corresponds to the quark dominance in the total width.
In $q\bar q$ channels, the Run-1 dijet constraint does not apply.
However, the Run-1 constraint on the diphoton production cross section
$\sigma_{\gamma\gamma}$ becomes relevant.
In the large width case with $\Gamma/M=0.06$,
for both the $b\bar b$ and $c\bar c$ channels,
the solution (i) is ruled out by
the Run-1 limit on the diphoton production,
as the required $\Gamma_{\gamma\gamma}$ is above $\mathcal{O}(10^{-3})$.
The solution (ii) is consistent with the data, and
the favoured $\Gamma_{\chi\chi}/M$ are of $\mathcal{O}(10^{-5})$, and
$\Gamma_{\gamma\gamma}/M$ are of $\mathcal{O}(10^{-4})$.
In the solution (ii), from $\Gamma_{q\bar q}/M\approx \Gamma/M=0.06$,
the size of the Yukawa coupling is found to be $y_{q}^{2}\approx 0.5$
which is marginally within the perturbative regime.
In the small width case with $\Gamma/M=0.006$,
for $b\bar b$ channel, both the solutions are close to
the Run-1 diphoton limit.
But the solution (ii) is favoured against solution (i).
For the $c\bar c$ channel, the situation is similar.
In the small width case $\Gamma/M=0.006$,
the favoured $\Gamma_{\chi\chi}/M$ is comparable with
$\Gamma_{\gamma\gamma}$, both are of $\mathcal{O}(10^{-4})$.
The determined values of $m_{\chi}$ and
$\langle \sigma v\rangle_{b\bar b, c\bar c}$ for the solution (ii)
are listed in \tab{tab:fitCombined}.
Since in the $q\bar q$ channel, the total width is not DM dominated.
The predicted cross section for gamma-ray lines which is proportional to
$\Gamma_{\chi\chi}\Gamma_{\gamma\gamma}$
can be smaller.
In \fig{fig:gammaline-scalar-qq}, we give the predicted cross sections for DM annihilation into $\gamma\gamma$ which gives rise to the gamma-ray spectral lines,
based on the parameters determined from the fit results listed in \tab{tab:fitCombined}
for $b\bar b$ and $c\bar c$ channels with
two different values of $\Gamma/M=0.06$ and 0.006, respectively.
In all the cases the predictions are well below the current upper limits set by Fermi-LAT.
For $\Gamma/M=0.06$, the predicted cross section
$\langle \sigma v \rangle_{\gamma\gamma}$ is
$\sim 5\times 10^{-31}~\text{cm}^{3}\text{s}^{-1}$ for $b\bar b$ channel,
and $\langle \sigma v \rangle_{Z\gamma}$ is below
$\sim 10^{-31}~\text{cm}^{3}\text{s}^{-1}$.
For $c\bar c$ channel, the predicted cross section
$\langle \sigma v \rangle_{\gamma\gamma}$ is
$\sim 1\times 10^{-31}~\text{cm}^{3}\text{s}^{-1}$.
The $Z\gamma$ final state is kinematically forbidden due to
th low mass of the DM particle.
For $\Gamma/M=0.06$, the predictions are relatively higher,
which is due to the fact that a larger $\Gamma_{\chi\chi}/M$ of $\mathcal{O}(10^{-4})$
is favoured.
\begin{table}\begin{center}
\begin{tabular}{llccccc}
\hline\hline
Channel & $\Gamma/M$ & $m_\chi$ (GeV) & $\Gamma_{\gamma\gamma}/M (\times 10^{-4}) $ & $\Gamma_{\chi\chi}/M $ & $\chi^2_{\text{min}}/\text{d.o.f.} $ & $p$-value \\%& $\Gamma_{qq}/M $
\hline
Scalar DM, $b\bar{b}$ & 0.06 & $46.15^{+5.81}_{-3.53}$ & $1.90^{+0.49}_{-0.50}$ & $5.52^{+0.68}_{-0.68}\times10^{-5}$ & 25.418/23 & 0.33 \\% & $0.05962^{+0.00008}_{-0.00008}$
& 0.006 & $46.15^{+5.81}_{-3.53}$ & $2.04^{+0.46}_{-0.48}$ & $6.56^{+0.97}_{-0.92}\times10^{-4}$ & 24.578/23 & 0.37 \\ %
Scalar DM, $c\bar{c}$ & 0.06 & $35.54^{+3.10}_{-4.12}$ & $0.81^{+0.21}_{-0.21}$ & $3.80^{+0.47}_{-0.47}\times10^{-5}$ & 26.664/23 & 0.27 \\ %
& 0.006 & $35.54^{+3.10}_{-4.12}$ & $0.89^{+0.23}_{-0.23}$ & $4.11^{+0.56}_{-0.54}\times10^{-4}$ & 26.311/23 & 0.29 \\ %
\hline
Fermionic DM, $b\bar{b}$ & 0.06 & $46.20^{+6.37}_{-2.68}$ & $4.55^{+2.39}_{-2.11}$ & $ 3.49^{+0.85}_{-1.75}\times10^{-2}$ & 24.906/23 & 0.36 \\ %
Fermionic DM, $c\bar{c}$ & 0.06 & $36.48^{+3.13}_{-2.11}$ & $1.62^{+0.82}_{-0.55}$ & $3.01^{+0.81}_{-0.85}\times10^{-2}$ & 27.048/23 & 0.25 \\ %
\hline\hline
\end{tabular}%
\caption{
Values of DM mass $m_{\chi}$, partial width-to-mass ratios
$\Gamma_{\gamma\gamma}/M$ and $\Gamma_{\chi\chi}/M$
determined from
combined fits to both the LHC diphoton excess and the GCE
in scalar and fermoinic DM models with constraints from Run-1 data
on the dijet and diphoton searches included.
For scalar DM models,
the results for the cases of $\phi$ coupling dominantly to
$b\bar b$ or $c\bar c$ with total width $\Gamma/M=0.06$ and 0.006 are given.
For fermionic DM models, the results are for $\Gamma/M=0.06$.
The corresponding $\chi^{2}/\text{d.o.f}$ and $p$-values for each fit are also shown.
}
\label{tab:fitCombined}
\end{center}\end{table}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_bb_006_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_bb_006_xy.eps}
\\
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_cc_006_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_cc_006_xy.eps}
\caption{
The same as \fig{fig:contour-scalar-gg}, but for scalar DM models with
$\phi$ generated from $b\bar b$ (upper panels) and
$c\bar c$ (lower panels) annihilation at the LHC for
$\Gamma/M=0.06$.
}
\label{fig:contour-scalar-qq-0.06}
\end{center}
\end{figure}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_bb_0006_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_bb_0006_xy.eps}
\\
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_cc_0006_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Scalar_cc_0006_xy.eps}
\caption{
The same as \fig{fig:contour-scalar-qq-0.06}, but for $\Gamma/M=0.006$.
}
\label{fig:contour-scalar-qq-0.006}
\end{center}
\end{figure}
\subsection{Fermionic DM}
For fermionic DM,
we shall focus on the case where $\chi$ is a Majorana fermion.
The results for the Dirac DM particle can be obtained
in a straight forward way.
The Lagrangian for the Majorana DM particle and
its interaction with $\phi$ is given by
\begin{align}
\mathcal{L} & \supset
\frac{1}{2}\bar\chi (i\gamma^{\mu}\partial_{\mu}-m_{\chi})\chi
-\frac{1}{2} y_{\chi}\bar\chi i\gamma^{5} \chi \phi .
\end{align}
For the DM annihilation into $gg$ and $q\bar q$ final states,
the corresponding $R_{X\bar{X}}$ factors are
\begin{align}
R_{gg}=64 \left( \frac{m_{\chi}}{M}\right)^{6}
\quad \text{and} \quad
R_{q\bar{q}}=16\frac{\beta_{q}(4m_{\chi}^{2})}{\beta_{q}(M^{2})}
\left( \frac{m_{\chi}}{M}\right)^{4} .
\end{align}
For $gg$ annihilation final states in this model,
the \eq{eq:totalWidthLimit} can be written as
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{fermion},gg}
\geq
\frac{\beta_{\chi}^{1/2}(M^{2})\langle \sigma v\rangle_{gg}^{1/2}M^{3}}{8\pi^{1/2}m_{\chi}^{2}} .
\end{align}
Since in this model the $R_{gg}$ factor is proportional to $(m_{\chi}/M)^{6}$,
the required total width is quite large.
In the lower-left panel of \fig{fig:GCEfit},
we show the upper limit on $\langle \sigma v\rangle_{gg}$ as a function of $m_{\chi}$
for three choices of $\Gamma/M=0.5$, 0.2 and 0.06, respectively.
For DM particle mass below $\sim100$ GeV,
the value of $\langle \sigma v\rangle_{gg}$ is
far below the typical thermal cross section.
For a consistent explanation to the DM relic density,
the required DM particle mass should be above $\sim 150$ GeV,
for $\Gamma/M=0.06$.
In fermionic DM model,
the factor $R_{q\bar q}$ is the same as $R_{gg}$ in the scalar DM model.
Thus the upper limit on the cross sections can be obtained from
that in the scalar DM model by a rescaling factor $1/s_{\chi}=1/4$.
For $gg$ channel,
using the best-fit values of $m_{\chi}$ and
$\langle \sigma v\rangle_{gg}$ in \tab{tab:GCEdm},
the corresponding minimal value of the width-to-mass ratio is found to be
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{fermion}, gg}
\gtrsim
0.31
\left( \frac{M}{750~\text{GeV}}\right)^{3}
\left( \frac{62~\text{GeV}}{m_{\chi}}\right)^{2}
\left( \frac{\langle \sigma v\rangle_{gg}}{1.96\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2} .
\end{align}
Such a large with is not favoured by the
current experimental data and is theoretically unnatural.
For $q\bar q$-channel, since $R_{q\bar q}$ is proportional to $(m_{\chi}/M)^{4}$,
the required total width is similar to the case of $gg$-channel of scalar DM.
For $b\bar b$ channel it is found that
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{fermion}, b\bar b}
\gtrsim
0.058
\left( \frac{M}{750~\text{GeV}}\right)^{2}
\left( \frac{46~\text{GeV}}{m_{\chi}}\right)
\left( \frac{\langle \sigma v\rangle_{gg}}{1.42\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2} ,
\end{align}
and the result is similar for the $c\bar c$ channel
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{fermion}, c\bar c}
\gtrsim
0.062
\left( \frac{M}{750~\text{GeV}}\right)^{2}
\left( \frac{35.5~\text{GeV}}{m_{\chi}}\right)
\left( \frac{\langle \sigma v\rangle_{gg}}{0.95\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2} .
\end{align}
We perform $\chi^{2}$-fit to the diphoton and GCE data in
the fermionic DM model for $b\bar b$ and $c\bar c$ channels with
$\Gamma/M=0.06$.
The determined parameters are shown in \tab{tab:fitCombined}, and
the allowed regions of the parameters in
$(\Gamma_{b\bar b}/M, \Gamma_{\gamma\gamma}/M)$ and
$(\Gamma_{\chi\chi}/M, \Gamma_{\gamma\gamma}/M)$ planes
are shown in \fig{fig:contour-Majorana-qq-0.06}.
Compared with the same channel in the scalar DM model,
a visible change in the allowed regions is that
the regions corresponding to the two solutions merge together,
which is due to the fact that in fermionic DM models,
the value of $\Delta$ is quite small as
the minimally required width is close to 0.06.
The determined values of $\Gamma_{q\bar q}/M$ and $\Gamma_{\chi\chi}/M$
are roughly the same order of magnitude about $\mathcal{O}(10^{-2})$.
The allowed regions are consistent with
the Run-1 limit on cross section of the diphoton production .
Since in the fermionic DM model,
$\Gamma_{\chi\chi}/M$ can reach $\mathcal{O}(10^{-2})$,
it is expected that the predicted cross sections for the gamma-ray line
are significantly larger than that in the scalar DM model.
In \fig{fig:gammaline-fermionic-qq},
we show the predicted cross sections in this model for $b\bar b$ and
$c\bar c$ channel with $\Gamma/M=0.06$.
The cross section can reach $\mathcal{O}(10^{-29}) \text{ cm}^{3}\text{s}^{-1}$,
which is very close to the current Fermi-LAT limit, and
can be tested in the future by Fermi-LAT, HESS and CTA.
In the case where $\chi$ is Dirac,
the corresponding values of $R_{X\bar X}$ are the same.
However, the required product of $(\Gamma_{\chi\bar\chi}/M)(\Gamma_{X\bar X})/M$
increases by a factor of four from Eqs. (\ref{eq:sigmav}) and (\ref{eq:fluxAvg}).
Thus the required total width $\Gamma/M$ is expected to be larger compared
with all the cases of Majorana DM.
\begin{figure}[htb]
\begin{center}
\includegraphics[width=0.45\textwidth]{Gamma_Majorana_bb_006_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Majorana_bb_006_xy.eps}
\\
\includegraphics[width=0.45\textwidth]{Gamma_Majorana_cc_006_fy.eps}
\includegraphics[width=0.45\textwidth]{Gamma_Majorana_cc_006_xy.eps}
\caption{
The same as \fig{fig:contour-scalar-qq-0.06} but for
the Majoranna fermionic DM model.
}
\label{fig:contour-Majorana-qq-0.06}
\end{center}
\end{figure}
\subsection{Vector DM}
In the case where the DM particle is a Majorana fermion,
the Lagrangian for DM and its interaction with $\phi$ is given by
\begin{align}
\mathcal{L} & \supset
\frac{1}{4 } \chi_{\mu\nu}\chi^{\mu\nu}
-\frac{1}{2} m_{\chi}^{2} \chi_{\mu}\chi^{\mu}
-\frac{1}{2} g_{\chi} \phi \chi_{\mu}\chi^{\mu}.
\end{align}
For $\text{\ensuremath{\chi\chi\to\phi\to}}gg, q\bar{q}$,
the corresponding $R_{X\bar{X}}$ factors are given by
\begin{align}
R_{gg}=\frac{192m_{\chi}^{8}}{M^{8}T(m_{\chi}/M)}, \quad
R_{q\bar{q}}=\frac{64m_{\chi}^{6}\beta_{q}(4m_{\chi}^{2})}{M^{6}\beta_{q}(M^{2})T(m_{\chi}/M)} ,
\end{align}
where $T(x)=1-4x^{2}+12x^{4}$.
Since in vector DM model $R_{gg}\propto (m_{\chi}/M)^{8}$ and
$R_{q\bar q}\propto (m_{\chi}/M)^{6}$,
it is expected that a very large $\Gamma/M$ is required.
For $gg$ channel, the minimally required width is given by
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{vector}, gg}
\gtrsim
3.3
\left( \frac{M}{750~\text{GeV}}\right)^{4}
\left( \frac{62~\text{GeV}}{m_{\chi}}\right)^{3}
\left( \frac{\langle \sigma v\rangle_{gg}}{1.96\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2},
\end{align}
and for $q\bar q$ channel
\begin{align}
\left(\frac{\Gamma}{M} \right)_{\text{fermion}, b\bar b}
\gtrsim
0.73
\left( \frac{M}{750~\text{GeV}}\right)^{3}
\left( \frac{46~\text{GeV}}{m_{\chi}}\right)^{2}
\left( \frac{\langle \sigma v\rangle_{gg}}{1.42\times 10^{-26}~\mbox{cm}^{3}\mbox{s}^{-1}}\right)^{1/2} .
\end{align}
The results for the $c\bar c$ channel is similar to that in the $b\bar b$ channel.
Thus in all the cases, the required $\phi$ width are too large and
already ruled out by the current Run-2 data,
which indicates that the vector DM model can not provide a
consistent explanation to the LHC Run-2 diphoton excess and the GCE.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{Limit_Scalar_qq_006.eps}
\includegraphics[width=0.45\textwidth]{Limit_Scalar_qq_0006.eps}
\caption{
Left)
Predictions for $\langle \sigma v \rangle_{\gamma\gamma}$ and
$\langle \sigma v \rangle_{Z\gamma}$ as a function of photon energy
in the scalar DM model with $\phi$ coupling dominantly with $b\bar b$ and $c\bar c$,
using the allowed parameters from the fit to the data of diphoton exces and GCE,
for $\Gamma/M=0.06$. The exclusion limits at $95\%$ C.L. of Fermi-LAT
\cite{Ackermann:2015lka} for region R16 are also shown.}
\label{fig:gammaline-scalar-qq}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\textwidth]{Limit_Majorana_qq_006.eps}
\caption{
The same as \fig{fig:gammaline-scalar-qq}, but for Majorana DM model and
$\Gamma/M=0.06$.
}
\label{fig:gammaline-fermionic-qq}
\end{center}
\end{figure}
\section{Conclusions}
In summary, we have investigate the conditions for
a consistent explanation for possible the LHC diphoton excess and GCE,
especially the requirement on total width of $\phi$ in
a wide range of DM models where
the DM particle can be scalar, fermionionic and vector, and
$\phi$ can be generated by
$s$-channel gluon fusion or quark-antiquark annihilation ($b\bar b $ and $c\bar c$)
at parton level.
We have shown that the required $\Gamma/M$ is determined by
a single parameter proportional to $(m_{\chi}/M)^{n}$.
We have found that three models can explain the two excesses successfully:
i) scalar DM model with $\phi$ coupling dominantly with $q\bar q$.
the minimally required $\Gamma/M$ can be as low as $\mathcal{O}(10^{-3})$;
ii) scalar DM model with $\phi$ coupling dominantly with $gg$, the required $\Gamma/M$ is about $\mathcal{O}(10^{-2})$;
iii) fermionic DM model with coupling dominantly with $q\bar q$, the required $\Gamma/M$ reaches $\mathcal{O}(10^{-2})$.
Other models such as the vector DM model requires larger $\Gamma/M$ of order one
which is already disfavoured by the current data.
For the same DM model, the required width of $\phi$ is always smaller
in $q\bar q$ channel than that in the $gg$ channel.
For the DM models which can simultaneously account for the diphoton excess and the GCE,
the predicted cross sections for gamma-ray line are typically of
$\mathcal{O}(10^{-30})~\text{cm}^{3}\text{s}^{-1}$,
which is close to the current limits imposed by the Fermi-LAT data.
These models can be distinguish soon by the updated LHC data,
through the measurement of the total width,
and Fermi-LAT data on the gamma-ray line searches in the near future.
{\bf Acknowledgements.}
This work is supported
by
the NSFC under Grants
No. 11335012 and
No. 11475237.
\bibliographystyle{JHEP} %
|
\section{Introduction}
Consider the setting in which we want to recover a vector $\vect{x} \in \RR{d}$ from linear measurements
\begin{eqnarray}
\label{eq:meas}
\vect{y} = \matr{M}\vect{x} + \vect{e},
\end{eqnarray}
where $\matr{M} \in \RR{m \times d}$ is the measurement matrix and $\vect{e} \in \RR{d}$ is additive noise.
This setup appears in many fields including statistics (e.g., regression), image processing (e.g., deblurring and super-resolution), and medical imaging (e.g., CT and MRI), to name just a few.
Often the recovery of $\vect{x}$ from $\vect{y}$ is an ill-posed problem. For example, when $\matr{M}$ has fewer rows than columns ($m < n$), rendering \eqref{eq:meas} an underdetermined linear system of equations. In this case, it is impossible to recover $\vect{x}$ without introducing additional assumptions on its structure. A popular strategy is to assume that $\vect{x}$ resides in a low dimensional set $\mathcal K$, e.g., sparse vectors \cite{Bruckstein09From, Candes05Decoding, Elad10Sparse, Eldar15Sampling} or a Gaussian Mixture Model (GMM) \cite{Yu12Solving}. The natural by-product minimization problem then becomes
\begin{eqnarray}
\label{eq:min_const}
\min_{\vect{x}} \norm{\vect{y} - \matr{M}\vect{x}}_2^2 & \text{s.t.} & \vect{x} \in \mathcal K.
\end{eqnarray}
This can be reformulated in an unconstrained form as
\begin{eqnarray}
\label{eq:min_unconst}
\min_{\vect{x}} \norm{\vect{y} - \matr{M}\vect{x}}_2^2 + \lambda f(\vect{x}),
\end{eqnarray}
where $\lambda$ is a regularization parameter and $f(\cdot)$ is a cost function related to the set $\mathcal K$. For example, if $\mathcal K = \left\{ \vect{x} \in \RR{d} : \norm{\vect{x}}_0 \le k\right\}$ is the set of $k$-sparse vectors, then a natural choice is $f(\cdot) = \norm{\cdot}_0$ or its convex relaxation $f(\cdot) = \norm{\cdot}_1$.
A popular technique for solving $\eqref{eq:min_const}$ and $\eqref{eq:min_unconst}$ is using iterative programs such as proximal methods \cite{Combettes11Proximal, Bottou16Optimization} that include the iterative shrinkage-thresholding algorithm (ISTA) \cite{Beck09Fast, Elad07Coordinate, Daubechies04iterative} and the alternating direction method of multipliers (ADMM) \cite{Boyd11Distributed, Gabay76dual}. This strategy is particularly useful for large dimensions $d$.
Many applications impose time constraints, which limit the number of computations that can be performed to recover $\vect{x}$ from the measurements. One way to minimize time and computations is to reduce the number of iterations without increasing the computational cost of each iteration. A different approach is to use momentum methods \cite{Wilson16Lyapunov} or random projections \cite{Pilanci15Iterative, Pilanci16Iterative, Pilanci15Newton, Tu17Breaking} to accelerate convergence.
Another alternative is to keep the number of iterations fixed while reducing the cost of each iteration.
For example, since the complexity of iterative methods rely, among other things, on $m$, a common technique to save computations is to sub-sample the measurements $\vect{y}$, removing ``redundant information,'' to an amount that still allows reconstruction of $\vect{x}$. A series of recent works \cite{Bottou08Tradeoffs, Shalev08SVM, Daniely13More, Bruer15Designing, Chandrasekaran13Computational, Oymak15Sharp} suggest that by obtaining more measurements one can benefit from simple efficient methods that cannot be applied with a smaller number of measurements.
In \cite{Bottou08Tradeoffs} the generalization properties of large-scale learning systems have been studied showing a tradeoff between the number of measurements and the target approximation. The work in \cite{Shalev08SVM} showed how it is possible to make the run-time of SVM optimization decrease as the size of the training data increases. In \cite{Daniely13More}, it is shown that the problem of supervised learning of halfspaces over 3-sparse vectors with trinary values $\{-1, 1, 0\}$ may be solved with efficient algorithms only if the number of training examples exceeds a certain limit. Similar phenomena are encountered in the context of sparse recovery, where efficient algorithms are guaranteed to reconstruct the sparsest vector only if the number of samples is larger than a certain quantity \cite{Eldar12Compressed, Elad10Sparse, Foucart13Mathematical}.
In \cite{Chandrasekaran13Computational} it was shown that by having a larger number of training examples it is possible to design more efficient optimization problems by projecting onto simpler sets. This idea is further studied in \cite{Bruer15Designing} by changing the amount of smoothing applied in convex optimization. In \cite{Oymak15Sharp} the authors show that more measurements may allow increasing the step-size in the projected gradient algorithm (PGD) and thus accelerating its convergence.
While these works studied a tradeoff between convergence speed and the number of available measurements, this paper takes a different route.
Consider the case in which due to time constraints we need to stop the iterations before we achieve the desired reconstruction accuracy. For the original algorithm, this can result in the recovery being very far from the optimum.
An important question is whether we can modify the original iterations (e.g., those dictated by the shrinkage or ADMM techniques), such that the method convergences to an improved solution with fewer iterations without adding complexity to them.
This introduces a tradeoff between the recovery error we are willing to tolerate and the computational cost. As we demonstrate, this goes beyond the trivial relationship between the approximation error and the number of iterations that exists for various iterative methods \cite{Beck09Fast}.
Such a tradeoff is experimentally demonstrated by the success of {\it learned} ISTA (LISTA) \cite{Gregor10Learning} for sparse recovery with $f(\cdot) = \norm{\cdot}_1$. This technique learns a neural network with only several layers, where each layer is a modified version of the ISTA iteration.\footnote{ISTA and its variants is one of the most powerful optimization techniques for sparse coding.} It achieves virtually the same accuracy as the original ISTA using one to two orders of magnitude less iterations. The acceleration of iterative algorithms with neural networks is not unique only to the sparse recovery problem and $f(\cdot) = \norm{\cdot}_1$. This behavior was demonstrated for other models such as the analysis cosparse and low-rank matrix models \cite{Sprechmann15Learning}, Poisson noise \cite{Remez15Picture}, acceleration of Eulerian fluid simulation \cite{Tompson16Accelerating}, and feature learning \cite{Andrychowicz16Learning}.
However, a proper theoretical justification to this phenomena is still lacking.
{\bf Contribution.} In this work, we provide theoretical foundations elucidating the tradeoff between the allowed minimization error and the number of simple iterations used for solving inverse problems. We formally show that if we allow a certain reconstruction error in the solution, then it is possible to change iterative methods by modifying the linear operations applied in them such that each iteration has the same complexity as before but the number of steps required to attain a certain error is reduced.
Such a tradeoff seems natural when working with real data, where both the data and the assumed models are noisy or approximate; searching for the {\it exact} solution of an optimization problem, where all the variables are affected by measurement or model noise may be an unnecessary use of valuable computational resources.
We formally prove this relation for iterative projection algorithms.
Interestingly, a related tradeoff exists also in the context of sampling theory, where by allowing some error in the reconstruction we may use fewer samples and/or quantization levels \cite{Kipnis16Sampling}.
We argue that the tradeoff we analyze may explain the smaller number of iterations required in LISTA compared to ISTA.
Parallel efforts to our work also provide justification for the success of LISTA. In \cite{Moreau16Adaptive}, the fast convergence of LISTA is justified by connecting between the convergence speed and the factorization of the Gram matrix of $\matr{M}$. In \cite{Xin16Maximal}, the convergence speed of ISTA and LISTA is analyzed using the restricted isometry property (RIP) \cite{Candes05Decoding}, showing that LISTA may reduce the RIP, which leads to faster convergence. A relation between LISTA and approximate message passing (AMP) strategies is drawn in \cite{Borgerding16Onsager}.
Our paper differs from previous contributions in three main points: (i) it goes beyond the case of standard LISTA with sparse signals and considers variants that apply to general low-dimensional models; (ii) our theory relies on the concept of inexact projections and their relation to the tradeoff between convergence-speed and recovery accuracy, which differs significantly from other attempts to explain the success of LISTA; and (iii) besides exploring LISTA, we provide acceleration strategies to other programs such as model-based compressed sensing \cite{Eldar12Compressed} and sparse recovery with side-information.
{\bf Organization.} This paper is organized as follows. In Section~\ref{sec:prelim} we present preliminary notation and definitions, and describe the ISTA, LISTA and PGD techniques. Section~\ref{sec:PGD_new_theory} introduces a new theory for PGD for non-convex cones.
Section~\ref{sec:iter_proj_tradeoofs} shows how it is possible to tradeoff between convergence speed and reconstruction accuracy by introducing the inexact projected gradient descent (IPGD) method using spectral compressed sensing \cite{Duarte13Spectral} as a motivating example. The reconstruction error of IPGD is analyzed as a function of the iterations in Section~\ref{sec:IPGD_theory}.
Section~\ref{sec:examples} discusses the relation between our theory and model-based compressed sensing \cite{Baraniuk10Model} and sparse recovery with side information \cite{Friedlander12Recovering, Khajehnejad09Weighted, Vaswani10Modified, Weizman15Compressed}. Section~\ref{sec:LIPGD} proposes a LISTA version of IPGD, the learned IPGD (LIPGD), and demonstrates its usage in the task of image super-resolution.
Section~\ref{sec:cs_DL} relates the approximation of minimization problems studied here with neural networks and deep learning, providing a theoretical foundation for the success of LISTA and suggesting a ``mixture-model'' extension of this technique. Section~\ref{sec:conc} concludes the paper.
\section{Preliminaries and Background}
\label{sec:prelim}
Throughout this paper, we use the following notation. We write
$\norm{\cdot}$ for the Euclidian norm for vectors and the spectral norm for matrices, $\norm{\cdot}_1$ for the $\ell_1$
norm that sums the absolute values of a vector and $\norm{\cdot}_0$
for the $\ell_0$ pseudo-norm, which counts the number of non-zero elements in a vector. The conjugate transpose of $\matr{M}$ is denoted by
$\matr{M}^*$ and the orthogonal projection onto the set $\mathcal K$ by $\mathcal P_{\mathcal K}$. The original unknown vector is denoted by $\vect{x}$, the given measurements by $\vect{y}$, the measurement matrix by $\matr{M}$ and the system noise by $\vect{e}$. The $i$th entry of a vector $\vect{v}$ is denoted by $\vect{v}[i]$. The sign function $\sgn(\cdot)$ equals $1$, $-1$ or $0$ if its input is positive, negative or zero respectively. The $d$-dimensional $\ell_2$-ball of radius $r$ is denoted by $\mathbb{B}^d_r$. For balls of radius $1$, we omit the subscript and just write $\mathbb{B}^d$.
\subsection{Iterative shrinkage-thresholding algorithm (ISTA)}
A popular iterative technique for minimizing \eqref{eq:min_unconst} is ISTA. Each of its iterations is composed of a gradient step with step size $\mu$, obeying $\frac{1}{\mu} \ge \norm{\matr{M}}$ to ensure convergence \cite{Beck09Fast}, followed by a proximal mapping $\mathcal S_{f,\lambda}(\cdot)$ of the function $f$, defined as
\begin{eqnarray}
\mathcal S_{f,\lambda }(\vect{v}) = \argmin_{\vect{z}} \frac{1}{2}\norm{\vect{z} - \vect{v}} + \lambda f(\vect{z}),
\end{eqnarray}
where $\lambda$ is a parameter of the mapping.
The resulting ISTA iteration can be written as
\begin{eqnarray}
\label{eq:ISTA}
\vect{z}_{t+1} = \mathcal S_{f,\mu\lambda}\left(\vect{z}_t + \mu \matr{M}^*(\vect{y} - \matr{M}\vect{z}_t)\right),
\end{eqnarray}
where $\vect{z}_t$ is an estimate of $\vect{x}$ at iteration $t$. Note that the step size $\mu$ multiplies the parameter of the proximal mapping.
The proximal mapping has a simple form for many functions $f$. For example, when $f(\cdot) = \norm{\cdot}_1$, it is an element-wise shrinkage function,
\begin{eqnarray}
\label{eq:soft_thresh}
\mathcal S_{\ell_1,\lambda}(\vect{v})[i] = \sgn\left(\vect{v}[i]\right)\max(0, \abs{\vect{v}[i]} - \lambda).
\end{eqnarray}
Therefore, the advantage of ISTA is that its iterations require only the application of matrix multiplications and then a simple non-linear function.
Nonetheless, the main drawback of ISTA is the large number of iterations that is typically required for convergence.
\begin{figure}[t]
\begin{center}
{\includegraphics[width=0.45\textwidth]{LISTA}
}
\end{center}
\vspace{-0.15in}
\caption{The LISTA scheme.}
\label{fig:LISTA_scheme}
\end{figure}
Many acceleration techniques have been proposed to speed up convergence of ISTA (see \cite{Beck09Fast, Beck09FastGradient, Elad07Coordinate, Nesterov83Method, Treister12Multilevel, Goldstein09Split, Su14Differential, Machart12Optimal, Aujol15Stability, Wibisono16variational} as a partial list of such works).
A prominent strategy is LISTA, which has the same structure as ISTA but with different linear operations in \eqref{eq:ISTA}. Empirically, it is observed that it is able to attain a solution very close to that of ISTA with a significantly smaller fixed number of iterations $T$. The LISTA iterations are given by\footnote{We present the more general version \cite{Sprechmann15Learning} that can be used for any signal model and not only for sparsity.}
\begin{eqnarray}
\label{eq:LISTA}
\vect{z}_{t+1} = \mathcal S_{f,\lambda}\left(\matr{A}\vect{y} + \matr{U}\vect{z}_t\right),
\end{eqnarray}
where $\matr{A}$, $\matr{U}$ and $\lambda$ are learned from a set of training examples by back-propagation with the objective being the $\ell_2$-distance between the final ISTA solution and the LISTA one (after $T$ iterations) \cite{Gregor10Learning}.
Other minimization objectives may be used, e.g., training LISTA to minimize \eqref{eq:min_unconst} directly \cite{Sprechmann15Learning}. Notice that LISTA has a structure of a recurrent neural network as can be seen in Fig.~\ref{fig:LISTA_scheme}.
While other acceleration techniques for ISTA have been proposed together with a thorough theoretical analysis, the powerful LISTA method has been introduced without mathematical justification for its success.
In this work, we focus on the PGD algorithm, whose iterations are almost identical to the ones of ISTA but with an orthogonal projection instead of a proximal mapping. We propose an acceleration technique for it, which is very similar to the one of LISTA, accompanied by a theoretical analysis.
\subsection{Projected gradient descent (PGD)}
The PGD iteration is given by
\begin{eqnarray}
\label{eq:PGD}
\vect{z}_{t+1} = \mathcal P_{\mathcal K}\left(\vect{z}_t + \mu \matr{M}^*(\vect{y} - \matr{M}\vect{z}_t)\right),
\end{eqnarray}
where $\mathcal P_{\mathcal K}$ is an orthogonal projection onto a given set $\mathcal K$.
For example, if $\mathcal K$ is the $\ell_1$-ball then $\mathcal P_{\mathcal K}$ is simply soft thresholding with a value that varies depending on the projected vector \cite{Duchi08Efficient}. Note the similarity to the proximal mapping in ISTA with $f$ as the $\ell_1$-norm, which is also the soft thresholding operation but with a fixed threshold \eqref{eq:soft_thresh}. This similarity is not unique to the $\ell_1$-norm case but happens also for other types of $f$ such as the $\ell_0$ pseudo-norm and the nuclear norm.
The step size $\mu$ is assumed to be constant for the sake of simplicity, as in \eqref{eq:ISTA}. In both methods it may vary between iterations.
PGD is a generalization of the iterative hard thresholding (IHT) algorithm, which was developed for $\mathcal K$ being the set of sparse vectors \cite{Blumensath09Iterative}.
This important method has been analyzed in various works. For example, for standard sparsity in \cite{Blumensath09Iterative}, for sparsity patterns that belong to a certain model in \cite{Baraniuk10Model}, for a general union of subspaces in \cite{Blumensath11Sampling}, for nonlinear measurements in \cite{Beck13Sparsity},
and more recently in \cite{Oymak15Sharp} for a set of the form $$\mathcal K = \left\{ \vect{z} \in \RR{d} : f(\vect{z}) \le R \right\}.$$ The formulation \eqref{eq:PGD} generalizes the special cases above. For example, if $f(\cdot) = \norm{\cdot}_0$ and $R$ is the sparsity level then we have the IHT method from \cite{Blumensath09Iterative}; when $f$ counts the number of non-zeros of only certain sparsity patterns, which are bounded by $R$, we have the model-based IHT of \cite{Baraniuk10Model}. PGD may also be applied to non-linear inverse problems \cite{Yang16Sparse, Oymak15Sharp}.
Theorem~\ref{thm:PGD_error_t} below provides convergence guarantees on PGD (it is the noiseless version of Theorem 1.2 in \cite{Oymak15Sharp}).
Before presenting the result, we introduce several properties of the set $\mathcal K$ and some basic lemmas.
\begin{defn}[Descent set and tangent cone]
The descent set of the function $f$ at a point $\vect{x}$ is defined as
\begin{eqnarray}
D_f(\vect{x}) = \left\{ \vect{h} \in \RR{d} : f(\vect{x} + \vect{h}) \le f(\vect{x}) \right\}.
\end{eqnarray}
The tangent cone $C_f(\vect{x})$ at a point $\vect{x}$ is the conic hull of $D_f(\vect{x})$, i.e., the smallest closed cone $C_f(\vect{x})$ satisfying $D_f(\vect{x}) \subseteq C_f(\vect{x})$.
\end{defn}
For concise writing, below we denote $D_f(\vect{x})$ and $C_f(\vect{x})$ as $D$ and $C$, respectively.
\begin{lem}[Lemma 6.2 in \cite{Oymak15Sharp}]
\label{lem:P_c_sup}
Let $\vect{v}\in \mathbb R^d$ and $C \subset \mathbb R^d$ be a closed cone. Then
\begin{eqnarray}
\norm{\mathcal P_C(\vect{v})} = \sup_{\vect{u} \in C \cap \mathbb{B}^d} \vect{u}^* \vect{v}.
\end{eqnarray}
\end{lem}
\begin{lem}
\label{lem:Kappa_D_rel}
If for $\vect{x} \in \RR{d}$, $\mathcal K = \{ \vect{z} \in \RR{d} : f(\vect{z}) \le f(\vect{x}) \} \subset \RR{d}$ is a closed set, then for all $\vect{v} \in \RR{d}$,
\begin{eqnarray}
\mathcal P_\mathcal K(\vect{x} + \vect{v} ) -\vect{x} = \mathcal P_{\mathcal K - \left\{\vect{x} \right\}}(\vect{v}) = \mathcal P_{D} (\vect{v} ).
\end{eqnarray}
\end{lem}
{\it Proof:} From the definition of the descent cone we have
$D = \left\{\vect{h} \in \RR{d} : f(\vect{h} + \vect{x}) \le f(\vect{x}) \right\} = \left\{\vect{z} - \vect{x} : f(\vect{z}) \le f(\vect{x}) \right\} = \left\{\vect{z} - \vect{x} : \vect{z} \in \mathcal K \right\} = \mathcal K - \left\{\vect{x} \right\}$,
where the second equality follows from a simple change of variables, and the last ones from the definitions of the set $\mathcal K$ and the Minkowski difference. Therefore, projecting onto $D$ is equivalent to a projection onto $\mathcal K - \left\{\vect{x} \right\}$.
\hfill $\Box$
\begin{lem}[Lemma~6.4 in \cite{Oymak15Sharp}]
\label{lem:C_D_rel}
Let ${D}$ and $C$ be a nonempty and closed set and cone, respectively, such that $0 \in D$ and $D \subseteq C$. Then for all $\vect{v} \in \RR{d}$
\begin{equation}
\norm{\mathcal P_D(\vect{v})} \le \kappa_f \norm{\mathcal P_C(\vect{v})},
\end{equation}
where $\kappa_f = 1$ if $D$ is a convex set and $\kappa_f = 2$ otherwise.
\end{lem}
We now introduce the convergence rate provided in \cite{Oymak15Sharp} for PGD. For brevity, we present only its noiseless version.
\begin{thm}[Noiseless version of Theorem~1.2 in \cite{Oymak15Sharp}]
\label{thm:PGD_error_t}
Let $\vect{x} \in \RR{d}$, $f: \RR{d} \rightarrow \mathbb R$ be a proper function, $\mathcal K = \left\{ \vect{z} \in \RR{d} : f(\vect{z}) \le f(\vect{x}) \right\}$,
$C = C_f(\vect{x})$ the tangent cone of the function $f$ at point $\vect{x}$, $\matr{M} \in \RR{m \times d}$ and $\vect{y} = \matr{M}\vect{x}$ a vector containing $m$ linear measurements. Assume we are using PGD with $\mathcal K$ to recover $\vect{x}$ from $\vect{y}$. Then the estimate $\vect{z}_t$ at the $t$th iteration (initialized with $\vect{z}_0 = 0$) obeys
\begin{eqnarray}
\norm{\vect{z}_t - \vect{x}} \le (\kappa_f \rho(C))^t\norm{\vect{x}},
\end{eqnarray}
where $\kappa_f$ is defined in Lemma~\ref{lem:C_D_rel}, and
\begin{eqnarray}
\label{eq:rho_C}
\rho(C) = \rho(\mu, \matr{M}, f, \vect{x}) = \sup_{\vect{u}, \vect{v} \in C \cap \mathbb{B}^d} \vect{u}^*\left(\matr{I} - \mu\matr{M}^*\matr{M} \right)\vect{v},
\end{eqnarray}
is the convergence rate of PGD.
\end{thm}
\subsection{Gaussian mean width}
\label{sec:mean_width_def}
When $\matr{M}$ is a random matrix with i.i.d. Gaussian distributed entries $\mathcal{N}(0,1)$, it has been shown in \cite{Oymak15Sharp} that the convergence rate
$\rho(C)$ is tightly related to the dimensionality of the set (model) $\vect{x}$ resides in. A very useful expression for measuring the ``intrinsic dimensionality'' of sets is the (Gaussian) mean width.
\begin{defn}[Gaussian mean width] The Gaussian mean width of a set $\Upsilon$ is defined as
\begin{eqnarray}
\label{eq:omega}
\omega(\Upsilon) = E[\sup_{\vect{v} \in \Upsilon \cap \mathbb{B}^d} \langle \vect{g}, \vect{v} \rangle ], & \vect{g} \sim \mathcal{N}(0,\matr{I}).
\end{eqnarray}
\end{defn}
Two variants of this measure are generally used. The {\em cone Gaussian mean width},
$\omega_C = \omega(C)$,
which measures the dimensionality of the tangent cone $C = C_f(\vect{x})$; and the {\em set Gaussian mean width}, $\omega_\mathcal K = \omega(\mathcal K - \mathcal K)$, which is related directly to the set $\mathcal K$ through its Minkowski difference $\mathcal K - \mathcal K = \left\{ \vect{z} - \vect{v} : \vect{z},\vect{v} \in \mathcal K \right\}$.
The cone Gaussian mean width relies on both the set $\mathcal K$ (through $f$) and a specific target point $\vect{x}$, while the set Gaussian mean width considers only $\mathcal K$. On the other hand, the dependence of $\omega_C$ on $\mathcal K$ is indirect via the descent set at the point $\vect{x}$.
There is a series of works, which developed convergence and reconstruction guarantees for various methods based on $\omega_C$ \cite{Chandrasekaran12Convex, Amelunxen14Living, Oymak15Sharp}, and others that rely on $\omega_\mathcal K$ \cite{Plan13Robust, Tirer17Generalizing}. The first ($\omega_C$) is mainly employed in the case of convex functions $f$, which are used to relax the non-convex set in which $\vect{x}$ resides. In this setting, often $D$ is convex and $\vect{x} \in \mathcal K$.
As an example of $\omega_C$ consider the case in which $\mathcal K$ is the $\ell_1$-ball and $\vect{x}$ is a $k$-sparse vector. Then $\omega_C \simeq \sqrt{2 k \log(d/k)}$.
If we add constraints on $\vect{x}$ such as having a tree structure, i.e., belonging to the set
\begin{eqnarray}
\label{eq:Kappa_hat_tree}
\hat{\mathcal K} = \{ \vect{z}\in \RR{d} : \norm{\vect{z}}_0 \le \norm{\vect{x}}_0 \text{\&} ~ \vect{z} ~ \text{obeys a tree structure} \},
\end{eqnarray}
where an entry may be non-zero only if its parent node is non-zero,
then the value of $\omega_C$ does not change. Although the definition of $\omega_\mathcal K$ is very similar to $\omega_C$, it yields different results. For $\mathcal K$ the set of $k$-sparse vectors $\omega_\mathcal K = O(\sqrt{k\log(d/k)})$, while for \eqref{eq:Kappa_hat_tree}, $\omega_{\hat{\mathcal K}} = O(\sqrt{k})$. The first result is similar to the expression of $\omega_C$ for the $\ell_1$ ball with a $k$-sparse vector $\vect{x}$, yet, the second provides a better measure of the set $\hat{\mathcal K}$ in \eqref{eq:Kappa_hat_tree}.
\subsection{PGD convergence rate and the cone Gaussian mean width}
\label{sec:conv_Gaussian_width}
In \cite{Oymak15Sharp}, it has been shown that the smaller $\omega_C$, the faster the convergence. More specifically, if $m$ is very close to $\omega_C$, then we may apply PGD with a step-size $\mu =\frac{1}{(\sqrt{d}+\sqrt{m})^2} \simeq \frac{1}{d}$ and have a convergence rate of (Theorem~2.4 in \cite{Oymak15Sharp})
\begin{eqnarray}
\label{eq:rho_omegaC1}
\rho(C) = 1- O\left(\frac{\sqrt{m} - \omega_C }{m+d}\right).
\end{eqnarray}
If $\omega_C$ is smaller than $\sqrt{m}$ by a certain constant factor, then we may apply PGD with a larger step size $\mu \simeq \frac{1}{m}$, which leads to improved convergence (Theorem~2.2 in \cite{Oymak15Sharp})
\begin{eqnarray}
\label{eq:rho_omegaC2}
\rho(\mathcal K) = O\left(\frac{\omega_C}{\sqrt{m}}\right).
\end{eqnarray}
These relationships rely on the fact that with larger $m$ the eigenvalues of $\matr{I} - \mu \matr{M}^T\matr{M}$ (after projection onto $C$, see \eqref{eq:rho_C}) are better positioned such that it is possible to improve convergence by increasing $\mu$.
Both \eqref{eq:rho_omegaC1} and \eqref{eq:rho_omegaC2} set a limit on the minimal value $m$ for which PGD iterations converge to $\vect{x}$, namely $m=O(\omega_C^2)$. This implies that $m \gtrapprox 2 k \log(d/k)$ (bigger than approximately $2 k \log(d/k)$)
for $\mathcal K$ as the $\ell_1$-ball and a $k$-sparse vector $\vect{x}$. This is known to be a tight condition. See more examples for this relationship between $\omega_C$ and $m$ in \cite{Chandrasekaran12Convex}.
The connections in \eqref{eq:rho_omegaC1} and \eqref{eq:rho_omegaC2} between $\rho(C)$ and $\omega_C$ are not unique only to the case that $\matr{M}$ is a random Gaussian matrix. Similar relationships hold for many other types of matrices \cite{Oymak15Sharp}.
\section{PGD Theory based on the Projection Set}
\label{sec:PGD_new_theory}
While Theorem~\ref{thm:PGD_error_t} covers many sets $\mathcal K$, there are interesting examples that are not included in it such as the set of $k$-sparse vectors corresponding to $f$ being the $\ell_0$ pseudo-norm, which is not a proper function. Even if we ignore this condition and try to use the result of Theorem~\ref{thm:PGD_error_t} in the case that $\matr{M}$ is a random Gaussian matrix we face a problem. Using the relationship between $\rho(C)$ and the Gaussian mean width $\omega_C$ in \eqref{eq:rho_omegaC1} and \eqref{eq:rho_omegaC2}, and the fact that in this case $\omega_C= \sqrt{d}$,
we get the condition $m >d$. This demand on $m$ is inferior to existing theory that in this scenario guarantees convergence with $m = O(k \log(d/k)$ \cite{Blumensath09Iterative}.
One way to overcome this problem is by considering the convex-hull of the set of $k$-sparse vectors with bounded $\ell_2$ norm. In this case $\omega_C = O(\sqrt{k\log(d/k)})$ \cite{Plan13Robust}. However, as mentioned above, guarantees for PGD exist for the $k$-sparse case without a bound on the $\ell_2$-norm.
A similar phenomenon also occurs with the set of sparse vectors with a tree structure $\hat{\mathcal K}$ (see \eqref{eq:Kappa_hat_tree}),
where again $\omega_C = \sqrt{d}$ implying $m = O(d)$. Yet, from the work in \cite{Baraniuk10Model}, we know that in this setting it is sufficient to choose $m = O(k)$. Note that for $\hat{\mathcal K}$, the set Gaussian mean width is $\omega_{\hat{\mathcal K}} = \sqrt{k}$. If we would have relied on it instead of on $\omega_C$ in the bound for the required size of $m$, it would have coincided with \cite{Baraniuk10Model}.
In order to address these deficiencies in the convergence rate, we provide a variant of Theorem~\ref{thm:PGD_error_t} that relies on the set $\mathcal K$ directly through the Minkowski difference $\mathcal K - \mathcal K$ in lieu of $C$.
For simplicity we present only the noiseless case but the extension to the noisy setting can be performed using the strategy in \cite{Oymak15Sharp}.
\begin{thm}
\label{thm:PGD_error_K}
Let $\vect{x} \in \mathcal K$,
$\mathcal K \subset \RR{d}$ be a closed cone,
$\matr{M} \in \RR{m \times d}$ and $\vect{y} = \matr{M}\vect{x}$ a vector containing $m$ linear measurements. Assume we are using PGD with $\mathcal K$ to recover $\vect{x}$ from $\vect{y}$. Then the estimate $\vect{z}_t$ at the $t$th iteration (initialized with $\vect{z}_0 = 0$) obeys
\begin{eqnarray}
\norm{\vect{z}_t - \vect{x}} \le (\kappa_\mathcal K \rho(\mathcal K))^t\norm{\vect{x}},
\end{eqnarray}
where $\kappa_\mathcal K=1$ if $\mathcal K$ is convex and $\kappa_\mathcal K = 2$ otherwise, and
\begin{eqnarray}
\rho(\mathcal K) = \rho(\mu, \matr{M}, \mathcal K) = \hspace{-0.1in} \sup_{\vect{u}, \vect{v} \in (\mathcal K - \mathcal K) \cap \mathbb{B}^d} \hspace{-0.1in} \vect{u}^*\left(\matr{I} - \mu\matr{M}^*\matr{M} \right)\vect{v},
\end{eqnarray}
is the convergence rate of PGD.
\end{thm}
{\it Proof:}
We repeat similar steps to the ones in the proof of Theorem~1.2 in \cite{Oymak15Sharp}.
We start by noting that the PGD error at iteration $t+1$ is,
\begin{eqnarray}
\label{eq:PGD_error_def}
\hspace{-0.27in} \norm{\vect{z}_{t+1} - \vect{x}} &=& \norm{\mathcal P_{\mathcal K}\left(\vect{z}_t + \mu \matr{M}^*(\vect{y} - \matr{M}\vect{z}_t)\right) - \vect{x}}
\\ \nonumber \hspace{-0.22in} &=& \norm{\mathcal P_{D}\left(\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x}) \right) },
\end{eqnarray}
where the last inequality is due to Lemma~\ref{lem:Kappa_D_rel} and the fact that $\vect{y} = \matr{M}\vect{x}$. Since $\mathcal K$ is a closed cone, also the Minkowski difference $\mathcal K - \mathcal K$ is a closed cone. Moreover, $D \subset \mathcal K-\mathcal K$ as $\vect{x} \in \mathcal K$. Thus, following Lemma~\ref{lem:C_D_rel} we have
\begin{eqnarray}
\label{eq:PGD_step_error_proof_1}
&& \hspace{-0.27in} \norm{\vect{z}_{t+1} - \vect{x}}
\le \kappa_\mathcal K\norm{\mathcal P_{\mathcal K-\mathcal K}\left(\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x}) \right) }
\\ \nonumber && \hspace{-0.27in}
\le \hspace{-0.1in} \sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_{\mathcal K}(\vect{v})- \vect{x}}
\le \norm{\vect{z}_t - \vect{x}}}} \hspace{-0.1in}
\kappa_\mathcal K\norm{\mathcal P_{\mathcal K-\mathcal K}\left(\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\mathcal P_\mathcal K(\vect{v}) - \vect{x}) \right) }
\\ \nonumber && \hspace{-0.27in}
\le \hspace{-0.1in} \sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(\vect{v}- \vect{x})}
\le \norm{\vect{z}_t - \vect{x}}}} \hspace{-0.1in}
\kappa_\mathcal K\norm{\mathcal P_{\mathcal K-\mathcal K}\left(\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\mathcal P_D(\vect{v} - \vect{x})) \right) },
\end{eqnarray}
where the second inequality is due to the fact that $\vect{z}_t$ is of the form $\mathcal P_\mathcal K(\vect{v})$ for some vector $\vect{v}$, and the last inequality follows from Lemma~\ref{lem:Kappa_D_rel}.
Noticing that the constraint $\norm{\mathcal P_D(\vect{v}- \vect{x})}
\le \norm{\vect{z}_t - \vect{x}}$ is equivalent to $\vect{v}- \vect{x} \in D \cap \mathbb{B}^d_{\norm{\vect{z}_t - \vect{x}}}$ (where $ \mathbb{B}^d_{\norm{\vect{z}_t - \vect{x}}}$ is the $\ell_2$-ball of radius $\norm{\vect{z}_t - \vect{x}}$) and
using the relation $D \subset \mathcal K-\mathcal K$ leads to
\begin{eqnarray}
\label{eq:PGD_step_error_proof_2}
&& \hspace{-0.27in} \norm{\vect{z}_{t+1} - \vect{x}}
\\ \nonumber && \hspace{-0.27in}
\le \sup_{\vect{v} \in (\mathcal K-\mathcal K) \cap \mathbb{B}^d}
\kappa_\mathcal K\norm{\mathcal P_{\mathcal K-\mathcal K}\left(\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\vect{v} \right) }\norm{\vect{z}_t - \vect{x}}
\\ \nonumber && \hspace{-0.27in}
\le \sup_{\vect{v}, \vect{u} \in (\mathcal K-\mathcal K) \cap \mathbb{B}^d}
\kappa_\mathcal K\norm{\vect{u}^*\left(\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\vect{v} \right) }\norm{\vect{z}_t - \vect{x}},
\end{eqnarray}
where the last inequality follows from Lemma~\ref{lem:P_c_sup}.
Using the definition of $\rho(\mathcal K)$ and applying the inequality in \eqref{eq:PGD_step_error_proof_2} recursively leads to the desired result.
\hfill $\Box$
When $\matr{M}$ is a random Gaussian matrix, the relationships in \eqref{eq:rho_omegaC1} and \eqref{eq:rho_omegaC2} hold with $\rho(\mathcal K)$ and $\omega_{\mathcal K}$ replacing $\rho(C)$ and $\omega_C$ respectively. This implies that we need $m = O(\omega_\mathcal K^2)$ for convergence. This result is in line with the conditions on $m$ that appear in previous works for $k$-sparse vectors \cite{Blumensath09Iterative}, for which $\omega_\mathcal K = O(\sqrt{k\log(d/k)})$, and for sparse vectors with tree structure \cite{Baraniuk10Model}, where $\omega_\mathcal K = O(\sqrt{k})$.
As discussed in Section~\ref{sec:mean_width_def}, the measure $\omega_\mathcal K$ is related directly to the set $\mathcal K$ (may be non-convex) in which $\vect{x}$ resides.
Thus, it provides a better measure for the complexity of $\mathcal K$ when it is unbounded or has some specific structure as is the case for sparsity with tree structure \cite{Tirer17Generalizing}. In such settings, Theorem~\ref{thm:PGD_error_K} should be favored over Theorem~\ref{thm:PGD_error_t}.
Notice that if $\vect{x} \in \mathcal K$, then we have $D \subset \mathcal K - \mathcal K$. Thus, in the settings that $D$ is convex and $\vect{x} \in \mathcal K$, we have $C = D \subset \mathcal K - \mathcal K$ implying that $\omega_C \le \omega_\mathcal K$; when $\matr{M}$ is random Gaussian, this also implies $\rho(C) \le \rho(\mathcal K)$. Therefore, in this scenario Theorem~\ref{thm:PGD_error_t} has an advantage over Theorem~\ref{thm:PGD_error_K}.
\section{Inexact projected gradient descent (IPGD)}
\label{sec:iter_proj_tradeoofs}
It may happen that the function $f$ or the set $\mathcal K$ are too loose for describing $\vect{x}$. Instead, we may select a set $\hat{\mathcal K}$ that better characterizes $\vect{x}$ and therefore leads to a smaller $\omega$, resulting in faster convergence. This improvement can be very significant; smaller $\omega$ both improves the convergence rate and allows using a larger step-size (see Section~\ref{sec:conv_Gaussian_width}).
For example, consider the case of a $k$-sparse vector $\vect{x}$, whose sparsity pattern obeys a tree structure.
If we ignore the structure in $\vect{x}$ and choose $f$ as the $\ell_0$ or $\ell_1$ norms, then the mean widths are $\omega_{\mathcal K} = O(k \log(d/k))$ \cite{Plan13Robust} and $\omega_{C} \simeq 2(k \log(d/k))$ \cite{Chandrasekaran12Convex} respectively. However, if we take this structure into account and use the set of $k$-sparse vectors with tree structure (see \eqref{eq:Kappa_hat_tree}),
then $\omega_{\hat{\mathcal K}} = O(k)$ \cite{Tirer17Generalizing}. As mentioned above, this improvement may be significant especially when $m$ is very close to $\omega_{\mathcal K}$.
Such an approach was taken in the context of model-based compressed sensing \cite{Baraniuk10Model}, where it is shown that faster convergence is achieved by projecting onto the set of $k$-sparse vectors with tree structure instead of the standard $k$-sparse set.
A related study \cite{Yu12Solving} showed that it is enough to use a small number of Gaussians to represent all the patches in natural images instead of using a dictionary that spans a much larger union of subspaces. This work relied on Gaussian Mixture Models (GMM), whose mean width scales proportionally to the number of Gaussians used, which is significantly smaller than the mean width of the sparse model.
\subsection{Inexact projection}
A difficulty often encountered is that the projection onto $\hat\mathcal K$, which may even be unknown, is more complex to implement than the projection onto $\mathcal K$. The latter can be easier to project onto but provides a lower convergence rate.
Thus, in this work we introduce a technique that compromises between the reconstruction error and convergence speed by using
PGD with an inexact ``projection'' that projects onto a set that is approximately as small as $\hat{\mathcal K}$ but yet is as computationally efficient as the projection onto $\mathcal K$. In this way, the computational complexity of each projected gradient descent iteration remains the same while the convergence rate becomes closer to that of the more complex PGD with a projection onto $\hat{\mathcal K}$.
The ``projection'' we propose is composed of a simple operator $p$ (e.g., a linear or an element-wise function) and the projection onto $\mathcal K$, $\mathcal P_{\mathcal K}$, such that it introduces only a slight distortion into $\vect{x}$. In particular, we require the following:
\subsubsection{The projection condition for convex \texorpdfstring{$\mathcal K$}{K}}
If $\mathcal K$ is convex, then we require
\begin{eqnarray}
\label{eq:Kp_eps_ineq}
\norm{\vect{x} - \mathcal P_{\mathcal K}\left( p(\vect{x})\right)} \le \epsilon\norm{\vect{x}}.
\end{eqnarray}
Due to Lemma~\ref{lem:Kappa_D_rel}, this is equivalent to
\begin{eqnarray}
\label{eq:Dp_eps_ineq}
\norm{\mathcal P_{D}\left(\vect{x} - p(\vect{x})\right)} \le \epsilon\norm{\vect{x}}.
\end{eqnarray}
From the fact that $\norm{\mathcal P_{D}\left(\vect{x} - p(\vect{x})\right)} \le \norm{\vect{x} - p(\vect{x})}$, it is sufficient that
\begin{eqnarray}
\label{eq:p_eps_ineq}
\norm{\vect{x} - p(\vect{x})} \le \epsilon\norm{\vect{x}},
\end{eqnarray}
to ensure \eqref{eq:Kp_eps_ineq}.
Examples for projections that satisfy condition \eqref{eq:Kp_eps_ineq} are given hereafter in sections~\ref{sec:IPGD_desc} and \ref{sec:sparse_side}.
\subsubsection{The projection condition for non-convex \texorpdfstring{$\mathcal K$}{K}}
In the case that $\mathcal K$ is non-convex, we require
\begin{equation}
\label{eq:P_K_pv_x_px_epsilon}
\norm{\mathcal P_\mathcal K\left( p \vect{v} - p\vect{x} \right) - \mathcal P_\mathcal K\left( p \vect{v} - \vect{x} \right)} \le \epsilon\norm{\vect{x}}, \forall \vect{v} \in \mathbb R^d.
\end{equation}
Due to Lemma~\ref{lem:Kappa_D_rel} and a simple change of variables, \eqref{eq:P_K_pv_x_px_epsilon} is equivalent to
\begin{equation}
\label{eq:P_D_pv_x_px_epsilon}
\norm{\mathcal P_D\left( p \vect{v} - \vect{x} + p\vect{x} \right) - \mathcal P_D\left( p \vect{v} \right)} \le \epsilon\norm{\vect{x}}, \forall \vect{v} \in \mathbb R^d,
\end{equation}
which by another simple change of variables is the same as
\begin{equation}
\label{eq:P_D_pv_x_pv_px_epsilon}
\norm{\mathcal P_D\left( p \vect{v} - \vect{x} \right) - \mathcal P_D\left( p \vect{v} - p\vect{x}\right)} \le \epsilon\norm{\vect{x}}, \forall \vect{v} \in \mathbb R^d.
\end{equation}
An example for a projection that satisfies condition \eqref{eq:P_K_pv_x_px_epsilon} is provided in Section~\ref{sec:tree}.
\begin{figure*}[t]
\begin{center}
{
{
\subfigure[x2 redundant DCT dictionary with sparsity $k=2$.]{\includegraphics[width=0.48\textwidth]{spect_error_2redundant}
\label{fig:spect_error_2redundant}}
\hspace{0.05in}
\subfigure[x4 redundant DCT dictionary with sparsity $k=4$.]{\includegraphics[width=0.48\textwidth]{spect_error}
\label{fig:spect_error_4redundant}}
}
}
\end{center}
\caption{Reconstruction error as a function of iterations for sparse recovery with a dictionary with high coherence between neighboring atoms. The sparse representation in the dictionary is generated such that there are three correlated neighboring atoms close to each other with location distance 1 or 4.
PGD is applied with $\mathcal K$ being the $\ell_1$ ball. IPGD is used with the same $\mathcal K$ and $p$ being a non-linear function that for a given vector keeps at most only one dominant entry in every neighborhood of fixed size (zeroing the smaller values).
This shows that IPGD may accelerate convergence compared to PGD and in some cases (right figure) even achieve lower recovery error.
}
\label{fig:spect_error}
\end{figure*}
\subsection{Inexact PGD}
\label{sec:IPGD_desc}
Plugging the inexact projection into the PGD step results in the proposed {\it inexact} PGD (IPGD) iteration (compare to \eqref{eq:PGD})
\begin{eqnarray}
\label{eq:IPGD}
\vect{z}_{t+1} = \mathcal P_{\mathcal K}\left(p\left(\vect{z}_t\right) + \mu p\left( \matr{M}^*(\vect{y} - \matr{M}\vect{z}_t)\right)\right).
\end{eqnarray}
To motivate this algorithm consider the problem of spectral compressed sensing \cite{Duarte13Spectral}, in which one wants to recover a sparse representation in a dictionary that has high local coherence. It has been shown that if the non-zeros in the representation are far from each other then it is easier to obtain good recovery \cite{Candes14Towards}.
Let $\matr{M}$ be a two times redundant DCT dictionary and $\tilde{\vect{x}}$ be a $k$-sparse vector, with sparsity $k=2$, of dimension $d =128$, such that the minimal distance (with respect to the location in the vector) between non-zero neighboring coefficients in it is greater than $5$ (indices) and the value in each non-zero coefficient is generated from the normal distribution. We construct the vector $\vect{x}$ by adding to $\tilde{\vect{x}}$ random Gaussian values with zero mean and variance $\sigma^2 = 0.05$ at the neighboring coefficients of each non-zero entry in $\tilde{\vect{x}}$ with (location) distance $1$ or $4$ (two different experiments).
As mentioned above, a better reconstruction is achieved by estimating $\tilde{\vect{x}}$ from $\matr{M}\tilde{\vect{x}}$ than by estimating $\vect{x}$ from $\matr{M}\vect{x}$ due to the highly correlated columns in $\matr{M}$. A common practice to improve the recovery in such a case is to force the recovery algorithm to select a solution with separated coefficients. In our context it is simply using the IPGD with a projection onto the $\ell_1$ ball and $p(\cdot)$ that keeps at most only one dominant entry (in absolute value) in every neighborhood of size $5$ in a given representation by zeroing out the other values.
The operator $p$ causes an error in the model (with $\epsilon\simeq 0.05\sqrt{2} \simeq 0.1$) and therefore reaches a slightly higher final error than PGD with projection onto the $\ell_1$ ball. Compared to PGD, IPGD projects onto a simpler set with a smaller Gaussian mean width, thus, attaining faster convergence at the first iterations, where the approximation error is still significantly larger than $\epsilon$ as can be seen in Fig.~\ref{fig:spect_error_2redundant}. When the coherence is larger (as in the case of added coefficients at distance $1$), the advantage of IPGD over PGD is more significant.
In some cases IPGD may even attain a lower final recovery error compared to PGD. For example, consider the case of $\matr{M}$ being a four times redundant DCT dictionary and $\vect{x}$ generated as above but with $k=4$.
Due to the larger redundancy in the dictionary, the coherence is larger in this case. Thus, the recovery of $\vect{x}$ is harder.
Here, PGD with a projection onto the $\ell_1$ ball converges slower and reaches a large error due to the high correlations between the atoms. Using IPGD with the $\ell_1$ ball and a projection $p(\cdot)$ that keeps at most only one dominant entry (in absolute value) in every neighborhood of size $5$, leads both to faster convergence and better final accuracy as can be seen in Fig.~\ref{fig:spect_error_4redundant}.
\section{IPGD Convergence Analysis}
\label{sec:IPGD_theory}
We turn to analyze the performance of IPGD.
For simplicity of the discussion, we analyze the convergence of this technique only for a linear operator $p$ and the noiseless setting, i.e., $\vect{e} = 0$. The extension to other types of operators and the noisy case is straightforward by arguments similar to those used in \cite{Oymak15Sharp} for treating the noise term and other classes of matrices.
We present two theorems on the convergence of IPGD. The first result provides a bound in terms of $\rho(C)$ (i.e., depends on $\omega_C$ if $\matr{M}$ is a random Gaussian matrix) for the case that $D$ is convex corresponding to $\kappa_f=1$ in Theorem~\ref{thm:PGD_error_t}; the second provides a bound in terms of $\rho(K)$ (i.e., depends on $\omega_\mathcal K$ if $\matr{M}$ is a random Gaussian matrix) when $\mathcal K$ is a closed cone but not necessarily convex.
The proofs of both theorems are deferred to appendices~\ref{sec:IPGD_error_t_proof} and \ref{sec:IPGD_error_K_proof}.
\begin{thm}
\label{thm:IPGD_error_t}
Let $\vect{x} \in \RR{d}$, $f: \RR{d} \rightarrow \mathbb R$ be a proper function, $\mathcal K = \left\{ \vect{z} \in \RR{d} : f(\vect{z}) \le f(\vect{x}) \right\}$,
$D = D_{f}(\vect{x})$ and $C = C_{f}(\vect{x})$ the descent set and the tangent cone of the function $f$ at point $\vect{x}$ respectively, $p(\cdot)$ a linear operator satisfying \eqref{eq:Dp_eps_ineq},
$\matr{M} \in \RR{m \times d}$ and $\vect{y} = \matr{M}\vect{x}$ a vector containing $m$ linear measurements. Assume we are using IPGD with $\mathcal K$ and $p$ to recover $\vect{x}$ from $\vect{y}$ and that $D$ is convex. Then the estimate $\vect{z}_t$ at the $t$th iteration (initialized with $\vect{z}_0 = 0$) obeys
\begin{eqnarray*}
\norm{\vect{z}_t - \vect{x}}
\le \left(\left(\rho_p(C)\right)^t + \frac{1-\left( \rho_p({C}) \right)^t}{1- \rho_p({C}) }(2+\rho_p(C))\epsilon \right)\norm{\vect{x}},
\end{eqnarray*}
where
\begin{eqnarray*}
\rho_p(C) &=& \rho(\mu, \matr{M}, f, p, \vect{x}) \\ \nonumber
&=& \hspace{-0.15in} \sup_{\vect{u},\vect{v} \in C\cap \mathbb{B}^d}
p(\vect{u})^*\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)p(\vect{v})
\end{eqnarray*}
is the ``effective convergence rate'' of IPGD for small $\epsilon$.
\end{thm}
\begin{thm}
\label{thm:IPGD_error_K}
Let $\vect{x} \in \mathcal K$,
$\mathcal K \subset \RR{d}$ be a closed cone, $p(\cdot)$ a linear operator satisfying \eqref{eq:P_K_pv_x_px_epsilon},
$\matr{M} \in \RR{m \times d}$ and $\vect{y} = \matr{M}\vect{x}$ a vector containing $m$ linear measurements. Assume we are using IPGD with $\mathcal K$ and $p$ to recover $\vect{x}$ from $\vect{y}$. Then the estimate $\vect{z}_t$ at the $t$th iteration (initialized with $\vect{z}_0 = 0$) obeys
\begin{eqnarray}
&& \hspace{-0.5in} \norm{\vect{z}_t - \vect{x}}
\le \left(\left(\kappa_\mathcal K\rho_p(\mathcal K)\right)^t + \frac{1-\left( \kappa_\mathcal K\rho_p({\mathcal K}) \right)^t}{1- \kappa_\mathcal K\rho_p({\mathcal K}) }\gamma \right)\norm{\vect{x}},
\end{eqnarray}
where $\kappa_\mathcal K$ and $\rho(K)$ are defined in Theorem~\ref{thm:PGD_error_K},
\begin{eqnarray}
\gamma \triangleq (2\rho(\mathcal K)\kappa_\mathcal K+\rho_p(\mathcal K)\kappa_\mathcal K+1)\epsilon,
\end{eqnarray}
and
\begin{eqnarray}
\label{eq:rho_p_K}
\rho_p(\mathcal K) &=& \rho(\mu, \matr{M}, \mathcal K, p) \\ \nonumber &=& \hspace{-0.15in} \sup_{\vect{u}, \vect{v} \in (\mathcal K - \mathcal K) \cap \mathbb{B}^d} p(\vect{u})^*\left(\matr{I} - \mu\matr{M}^*\matr{M} \right)p(\vect{v})
\end{eqnarray}
is the ``effective convergence rate'' of IPGD for small $\epsilon$.
\end{thm}
Theorems~\ref{thm:IPGD_error_t} and \ref{thm:IPGD_error_K} imply that if $\epsilon$ is small enough (compared to $\rho_p^t$, where $t$ is the iteration number and $\rho_p$ is defined in \eqref{eq:rho_p_K}) then IPGD has an effective convergence rate of $\rho_p=\rho_p(C)$ when $D$ is convex, and $\rho_p=\kappa_\mathcal K\rho_p(\mathcal K)$ in the case that $\mathcal K$ is a closed cone but not necessarily convex. Note that if $p = \matr{I}$ then $\epsilon =0$ and our results coincide with theorems~\ref{thm:PGD_error_t} and \ref{thm:PGD_error_K}.
As we shall see hereafter, for some operators $p$ the rate $\rho_p$ may be significantly smaller than $\rho(C)$ and $\rho(\mathcal K)$. The smaller the set that $p$ maps to, the smaller $\rho_p$ becomes. At the same time, when $p$ maps to smaller sets it usually provides a ``coarser estimate'' and thus the approximation error $\epsilon$ in \eqref{eq:Dp_eps_ineq} and \eqref{eq:P_K_pv_x_px_epsilon} increases.
Thus, IPGD allows us to tradeoff approximation error $\epsilon$ and improved convergence $\rho_p$.
\begin{figure*}[t]
\begin{center}
{
{
\subfigure[Recovery error as a function of iteration number]{\includegraphics[width=0.48\textwidth]{tree_error}
\label{fig:tree_error}}
\hspace{0.05in}
\subfigure[Recovery error as a function of running time (in sec)]{\includegraphics[width=0.48\textwidth]{tree_error_run_time}
\label{fig:tree_error_run_time}}
\subfigure[Recovery error as a function of iteration number zoomed]{\includegraphics[width=0.48\textwidth]{tree_error_zoom}
\label{fig:tree_error_zoom}}
\hspace{0.05in}
\subfigure[Recovery error as a function of running time (in sec) zoomed]{\includegraphics[width=0.48\textwidth]{tree_error_run_time_zoom}
\label{fig:tree_error_run_time_zoom}}
}
}
\end{center}
\caption{Reconstruction error as a function of the iterations (left) and the running time (right) for recovering a sparse vector with tree structure.
Since we initialize all algorithms with the zero vector, the error at iteration/time zero is $\norm{\vect{x}}$. Zoomed version of the first $10$ iterations and first $1ms$ appears in the bottom row.
This figure demonstrates the convergence rate of PGD with projections onto the sparse set and sparse tree set compared to IPGD with $p$ that projects onto a certain number of levels of the tree and IPGD with changing $p$ that projects onto an increasing number of levels as the iterations proceed. Note that while PGD with a projection onto a tree structure converges faster than IPGD as a function of the number of iterations (left figure), it converges slower than IPGD if we take into account the actual run time of each iteration, as shown in the right figure, due to the higher complexity of the PGD projections.
}
\label{fig:tree}
\end{figure*}
The error term in theorems~\ref{thm:IPGD_error_t} and \ref{thm:IPGD_error_K} at iteration $t$ is comprised of two components. The first goes to zero as $t$ increases while the second increases with iterations and is on the order of $\epsilon$. The fewer iterations we perform the larger $\epsilon$ we may allow. An alternative perspective is that the larger the reconstruction error we can tolerate, the larger $\epsilon$ may be and thus we require fewer iterations.
Therefore, the projection $p$ introduces a tradeoff. On the one hand, it leads to an increase in the reconstruction error. On the other hand, it simplifies the projected set, which leads to faster convergence (to a solution with larger error).
The works in \cite{Giryes14Greedy, Giryes15GreedySignal, Hegde14Approximation} use a similar concept of near-optimal projection (compared to \cite{Baraniuk10Model} that assumes only exact projections). The main difference between these contributions and ours is that these papers focus on specific models, while we present a general framework that is not specific to a certain low-dimensional prior.
In addition, in these papers the projection is performed to make it possible to recover a vector from a certain low-dimensional set, while in this work the main purpose of our inexact projections is to accelerate the convergence within a limited number of iterations. For a larger number of iterations these projections may not lead to a good reconstruction error.
\section{Examples}
\label{sec:examples}
This section presents examples of IPGD with an operator $p$ that accelerates the convergence of PGD for a given set $\mathcal K$.
\subsection{Sparse recovery with tree structure}
\label{sec:tree}
To demonstrate our theory we consider a variant of the $k$-sparse set with tree structure in \eqref{eq:Kappa_hat_tree} that has smaller weights in the lower nodes of the tree.
We generate a $k$-sparse vector $\vect{x} \in \RR{127}$ with $k = 13$ and a sparsity pattern that obeys a tree structure. Moreover, we generate the non-zero entries in $\vect{x}$ independently from a Gaussian distribution with zero mean and variance $\sigma^2=1$ if they are at the first two levels of the tree
and $\sigma^2 = 0.2^2$ for the rest.
The best way to recover $\vect{x}$ is by using a projection onto the set $\hat\mathcal K$ in \eqref{eq:Kappa_hat_tree}, which is the strategy proposed in the context of model-based compressed sensing \cite{Baraniuk10Model}. Yet, this projection requires some additional computations at each iteration \cite{Baraniuk10Model}. Our technique suggests to approximate it by a linear projection onto the first levels of the tree (a simple operation) followed by a projection onto $\mathcal K = \{\vect{z} : \norm{\vect{z}}_0 \le k \}$.
The more levels we add in the projection $p$, the smaller the approximation error $\epsilon$ turns out to be. More specifically, it is easy to show that $\epsilon$ in \eqref{eq:P_K_pv_x_px_epsilon} is bounded by two times the energy of the entries eliminated from $\vect{x}$ divided by the total energy of $\vect{x}$, i.e., by $2\frac{\norm{p(\vect{x}) - \vect{x}}}{\norm{\vect{x}}}$. Clearly, the more layers we add the smaller $\epsilon$ becomes. Yet, assuming that all nodes in each layer are selected with equal probability, the probability of selecting a node at layer $l$ is equal to $\prod_{i=1}^l 0.5^{i-1}$, where we take into account the fact that a node can be selected only if all its forefathers have been chosen. Thus, the upper layers have more significant impact on the values of $\epsilon$.
On the other hand, the convergence rate $\rho_p(\mathcal K)$ for a projection with $l$ layers is equivalent to the convergence rate for the set of vectors of size $2^{l}$ (denoted by $\mathcal K_l$). Thus, we get that $\rho_p(\mathcal K) = \rho(\mathcal K_l)$, which is dependent on the Gaussian mean width $\omega_{\mathcal K_l}$ that scales as $\max(k l, k\log(2^l/k))$. Clearly, when we take all the layers $l = \log(d)$ and we have $\omega_{\mathcal K_l} = \omega_\mathcal K = O(k\log(d/k))$.
Figure~\ref{fig:tree_error} presents the signal reconstruction error ($\norm{\vect{x} - \vect{z}_t}_2$) as a function of the number of iterations for PGD with the sets $\mathcal K$ (IHT \cite{Blumensath09Iterative}) and $\hat\mathcal K$ (model-based IHT \cite{Baraniuk10Model})\footnote{For demonstration purposes we plot only the cases where model-based IHT converges to zero.} and for the proposed IPGD with $p$ that projects onto a different number of levels (1-5) of the tree. All algorithms use step size $\mu =\frac{1}{(\sqrt{d}+\sqrt{m})^2}$.
It is interesting to note that if $p$ projects only onto the first layer, then the algorithm does not converge as the resulting approximation error $\epsilon$ is too large.
However, starting from the second layer, we get a faster convergence at the first iterations with $p$ that projects onto a smaller set, which yields a smaller $\rho$. As the number of iterations increases, the more accurate projections achieve a lower reconstruction error, where the plateau attained is proportional to the approximation error of $p$ as predicted by our theory.
This tradeoff can be used to further accelerate the convergence by changing the projection in IPGD over the iterations. Thus, in the first iterations we enjoy the fast convergence of the coarser projections and in the later ones we use more accurate projections that allow achieving a lower plateau. The last line in Fig.~\ref{fig:tree} demonstrates this strategy, where at the first iteration $p$ is set to be a projection onto the first two levels, and then every four iterations another tree level is added to the projection until it becomes a projection onto all the tree levels (in this case IPGD coincides with PGD). Note that IPGD converges faster than PGD also when the projection in it becomes onto all the tree levels. This can be explained by the fact that typically convergence of non-linear optimization techniques depends on the initialization point \cite{Bertsekas99Nonlinear}.
While here we arbitrarily chose to add another level every fixed number of iterations, in general, a control set can be used for setting the number of iterations to be performed in each training level. We demonstrate this strategy in Section~\ref{sec:LIPGD}.
Since PGD with $\hat{\mathcal K}$ does not introduce an error in its projection and projects onto a precise set, it achieves the smallest recovery error throughout all iterations. Yet, as its projection is computationally demanding, it converges slower than IPGD if we take into account the run time of each iteration, as can been seen in Fig.~\ref{fig:tree_error_run_time}. This clearly demonstrates the advantage of using simple projections with IPGD compared to accurate but more complex projections with PGD.
\begin{figure}[t]
\begin{center}
{
{\subfigure[House image.]{\includegraphics[width=0.23\textwidth]{house}}
\label{fig:house_image}
\hspace{0.1in}
\subfigure[Patch from house image.]{\includegraphics[width=0.23\textwidth]{house_patch}}
\label{fig:house_patch} \\ \vspace{0.2in}
\subfigure[Patch representation magnitudes with DCT.]{\includegraphics[width=0.23\textwidth]{house_dct}}
\label{fig:house_dct}
\hspace{0.05in}
\subfigure[Patch representation magnitudes with Haar.]{
\includegraphics[width=0.23\textwidth]{house_haar}}
\label{fig:house_haar}
}
}
\end{center}
\caption{House Image (top left) and a random patch selected from it (top right) with the sorted magnitude (in log scale) of the representation of this patch in the DCT (bottom left) and Haar (bottom right) bases.}
\label{fig:house}
\end{figure}
\begin{figure}[t]
\begin{center}
{
{\includegraphics[width=0.48\textwidth]{side_inf_error}
\label{fig:side_inf_error}}
}
\end{center}
\caption{Reconstruction error as a function of the iterations for sparse recovery with side information. This demonstrates the convergence rate of (i) PGD with a projection onto the $\ell_1$ ball compared to (ii) IPGD with oracle side information on the columns of the representation of $\vect{x}$ in the Haar basis; (iii) IPGD with oracle side information that projects onto an increasing number of columns from the Haar basis ordered according to their significance in representing $\vect{x}$; (iv) IPGD with a projection onto the first $512$ columns of that Haar basis; and (v) IPGD with a changing $p$ that projects onto an increasing number of columns from the Haar basis.}
\label{fig:side_inf}
\end{figure}
\subsection{Sparse recovery with side information}
\label{sec:sparse_side}
Another possible strategy to improve reconstruction that relates to our framework is using side information about the recovered signal, e.g., from estimates of similar signals. This approach was applied to improve the quality of MRI \cite{Fessler92Regularized, Hero99Minimax, Weizman15Compressed} and CT scans \cite{Chen08Prior}, and also in the general context of sparse recovery \cite{Friedlander12Recovering, Khajehnejad09Weighted, Vaswani10Modified, Weizman15Compressed}.
We demonstrate this approach, in combination with our proposed framework, for the recovery of a sparse vector under the discrete cosine transform (DCT), given information of its representation under the Haar transform. Our sampling matrix is $\matr{M} = \matr{A}\matr{D}^*$, where $\matr{A} \in \RR{700 \times 1024}$ is a random matrix with i.i.d. normally distributed entries, $\matr{D}^*$ is the (unitary) DCT transform (that is applied on the signal before multiplying the DCT coefficients with the random matrix $\matr{A}$) and $\matr{D}$ is the DCT dictionary. We use random patches of size $32\times 32$, normalized to have unit $\ell_2$ norm, from the standard {\em house} image. Note that such a patch is not exactly sparse either in the Haar or the DCT domains. See Fig.~\ref{fig:house} for an example of one patch of the house image.
Without considering the side information of the Haar transform, one may recover $\vect{x}$ (a representation of a patch in the DCT basis) by using PGD with the set $\mathcal K = \{\vect{z} : \norm{\vect{z}}_1 \le \norm{\vect{x}}_1 \}$. Given $\vect{x}$ we may recover the patch $\matr{D}\vect{x}$.
Assume that someone gives us oracle side information on the set of Haar columns corresponding to the largest coefficients that contain $95\%$ of the energy in a patch $\matr{D}\vect{x}$. While there are many ways to incorporate the side information in the recovery, we show here how IPGD can be used for this purpose.
Denoting by $\matr{P}_{\vect{x}, 95\%}^{oracle}$ the linear projection onto this set of columns, one may apply IPGD with $p = \matr{D}\matr{P}_{\vect{x}, 95\%}^{oracle}\matr{D}^*$ and $\mathcal K$. As $\norm{\matr{D}\vect{x}} = \norm{\vect{x}}$ (since $\matr{D}$ is unitary), we have that $\epsilon = 0.05$ in \eqref{eq:Kp_eps_ineq}. Figure~\ref{fig:side_inf} compares between PGD with $\mathcal K$ and IPGD with $p$ and $\mathcal K$. We average over $100$ different randomly selected sensing matrices and patches.
The number of columns in Haar that contain $95\%$ of the energy is roughly $d/2$. Thus, the Gaussian mean width $\omega_p(C)$ in this case is roughly the width of the tangent cone of the $\ell_1$ norm at a $k$-sparse vector in the space of dimension $d/2$, which is smaller than $\omega(C)_{\ell_1}(\vect{x})$. Thus, $\rho_p(C)$ is smaller than $\rho(C)$.
Clearly, for less energy preserved in $\vect{x}$ (i.e., bigger $\epsilon$) we need less columns from Haar, which implies a smaller Gaussian mean width and a faster convergence rate $\rho_p(\mathcal K)$. We have here a tradeoff between the approximation error we may allow $\epsilon$ and the convergence rate $\rho_p(\mathcal K)$ that improves as $\epsilon$ increases.
Since projections onto smaller sets lead to faster convergence we suggest as in the previous example to apply PGD with an oracle projection that uses less columns from the Haar basis at the first iteration (i.e., has larger $\epsilon$) and then adds columns gradually throughout the iterations. The third (red) line in Fig.~\ref{fig:side_inf} demonstrates this option, where the first iterations use a projection onto the columns that contain $50\%$ of the energy of $\vect{x}$ and then every $5$ iterations the next $50$ columns correspoding to the coefficients with the largest energy are added. We continue until the columns span $95\%$ of the energy of the signal. Thus, IPGD with changing projections converges faster than IPGD with a constant $p = \matr{D}\matr{P}_{\vect{x}, 95\%}^{oracle}\matr{D}^*$ but reaches the same plateau.
Typically, oracle information on the coefficients of $\vect{x}$ in the Haar basis is not accessible. Even though, it is still possible to use common statistics of the data to accelerate convergence. For example, in our case it is known that most of the energy of the signal is concentrated in the low-resolution Haar filters. Therefore, we propose to use IPGD with a projection $p$ that projects onto the first $512$ columns of the Haar basis. As before, it is possible to accelerate convergence by projecting first on a smaller number of columns and then increasing the number as the iterations proceed (in this case we add columns till IPGD coincides with PGD). These two options are presented in the fourth and fifth line of Fig.~\ref{fig:side_inf}, respectively. Both of these options provide faster convergence, where IPGD with a fixed projection $p$ incurs a higher error as it uses less accurate projections in the last iterations compared to PGD and IPGD with changing projections. The plateau of the latter is the same one of the regular PGD (which is not attained in the graph due to its early stop) but is achieved with a much smaller number of iterations.
\begin{table}
\footnotesize
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Image & Bicubic & OMP & IHT & LIPGD \\
\hline
baboon & 23.2 & 23.5 & 23.4 & \bf 23.6\\
bridge & 24.4 & 25.0 & 24.8 & \bf 25.1 \\
coastguard & 26.6 & 27.1 & 26.9 & \bf 27.2 \\
comic & 23.1 & 24.0 & 23.8 & \bf 24.2 \\
face & 32.8 & 33.5 & 33.2 & \bf 33.6 \\
flowers & 27.2 & 28.4 & 28.1 & \bf 28.7 \\
foreman & 31.2 & 33.2 & 32.3 & \bf 33.5 \\
lenna & 31.7 & 33.0 & 32.6 & \bf 33.2 \\
man & 27.0 & 27.9 & 27.7 & \bf 28.1 \\
monarch & 29.4 & 31.1 & 30.9 & \bf 31.6 \\
pepper & 32.4 & 34.0 & 33.6 & \bf 34.4 \\
ppt3 & 23.7 & 25.2 & 24.6 & \bf 25.5 \\
zebra & 26.6 & 28.5 & 28.0 & \bf 28.9 \\
\hline
\end{tabular}
\end{center}
\caption{PSNR of super-resolution by bicubic interpolation and a pair of dictionaries with various sparse coding methods.}
\label{tbl:sup_res}
\end{table}
\section{Learning the Projection -- Learned IPGD (LIPGD)}
\label{sec:LIPGD}
\begin{figure}[t]
\begin{center}
{{\includegraphics[width=0.48\textwidth]{LISTA_MM}}
}
\end{center}
\caption{The $\ell_1$ loss as a function of the iterations of ISTA, LISTA and LISTA-MM applied on patches from the {\em house} image. This demonstrates the faster convergence of the proposed LISTA-MM compared to LISTA and the fast convergence of LISTA compared to ISTA.}
\label{fig:LISTA}
\end{figure}
In many scenarios, we may not know what type of simple operator $p$ causes $\mathcal P_\mathcal K(p(\cdot))$ to approximate $\hat{\mathcal K}$ in the best possible way. Therefore, a useful strategy is to learn $p$ for a given dataset. Assuming a linear $p$, we may rewrite \eqref{eq:IPGD} as
\begin{eqnarray}
\label{eq:IPGD_net}
\vect{z}_{t+1} = \mathcal P_{\mathcal K}\left( p\left(\mu\matr{M}^*\vect{y}\right) +
p\left(\left(\matr{I} - \mu \matr{M}^*\matr{M}\right)\vect{z}_t \right)\right).
\end{eqnarray}
Instead of learning $p$ directly, we may learn two matrices $\matr{A}$ and $\matr{U}$, where the first replaces $p\mu\matr{M}^*$ and the second $p\left(\matr{I} - \mu \matr{M}^*\matr{M}\right)$.
This results in the iterations
\begin{eqnarray}
\label{eq:Learned_IPGD}
\vect{z}_{t+1} = \mathcal P_{\mathcal K}\left( \matr{A}\vect{y} +
\matr{U}\vect{z}_t \right),
\end{eqnarray}
which is very similar to those of LISTA in \eqref{eq:LISTA}. The only difference between \eqref{eq:Learned_IPGD} and LISTA is the non-linear part, which is an orthogonal projection in the first and a proximal mapping in the second.
We apply this method to replace the sparse coding step in the super-resolution algorithm proposed in \cite{Zeyde12single}, where a pair of low and high resolution dictionaries is used to reconstruct the patches of the high-resolution image from the low-resolution one. In the code provided by the authors of \cite{Zeyde12single}, orthogonal matching pursuit (OMP) \cite{MallatZhang93} with sparsity $3$ is used. The complexity of this strategy corresponds to IHT with $3$ iterations. The target sparsity we use with IHT is higher ($k=40$) as it was observed to provide better reconstruction results. Note that in IHT, unlike OMP, the number of iterations may be different than the sparsity level. For optimal hyperparameter selection (such as choosing the target sparsity level), we use the training set used for the training of the dictionary in \cite{Zeyde12single}, which contains $91$ images.
Since IHT does not converge with only $3$ iterations, we apply LIPGD to accelerate convergence. We use the same dictionary dimension as in \cite{Zeyde12single} ($30 \times 1000$ and $81\times1000$ for the low and high resolution dictionaries, respectively), and train an LIPGD network to infer the sparse code of the image patches in the low-resolution dictionary. Training of the weights is performed by stochastic gradient descent with batch-size $1000$ and Nesterov momentum \cite{Nesterov83Method} for adaptively setting the learning rate. We train the network using only the first $85$ images in the training set, keeping the last $6$ as a validation set. We reduce the training rate by a factor of $2$ if the validation error stops decreasing. The initial learning rate is set to $0.001$ and the Nesterov parameter to $0.9$.
We use the sparse representations of the training data calculated by IHT or LIPGD to generate the high-resolution dictionary as in \cite{Zeyde12single}.
Table~\ref{tbl:sup_res} summarizes the reconstruction results of regular bicubic interpolation, the OMP-based super-resolution technique of \cite{Zeyde12single} (with $3$ iterations) and its version with IHT and LIPGD (replacing OMP). It can be seen clearly that IHT leads to inferior results compared to OMP since it does not converge in $3$ iterations. LIPGD improves over both IHT and OMP as the training of the network allows it to provide good sparse approximation with only $3$ iterations. This demonstrates the efficiency of the proposed LIPGD technique, which has the same computational complexity of both OMP and IHT.
\section{Learning the Projection -- LISTA Mixture Model}
\label{sec:cs_DL}
Though the theory in this paper applies directly only to \eqref{eq:Learned_IPGD} (with some constraints on $\matr{A}$ and $\matr{U}$ that stem from the constraints on $p$), the fast convergence of LISTA may be explained by the resemblance of the two methods. The success of LISTA may be interpreted as learning to approximate the set $\hat{\mathcal K}$ in an indirect way by learning the linear operators $\matr{A}$ and $\matr{U}$.
In other words, it can be viewed as a method for learning a linear operator that together with the proximal mapping $\mathcal S_{f, \lambda}$ approximates a more accurate proximal mapping of a true unknown function $\hat{f}$ that leads to much faster convergence.
With this understanding, we argue that using multiple inexact projections may lead to faster convergence as each can approximate in a more accurate way different parts of the set $\hat{\mathcal K}$. In order to show this, we propose a LISTA mixture model (MM), similar to the Gaussian mixture model proposed in \cite{Yu12Solving}, in which we train several LISTA networks, one for each part of the dataset. Then, once we get a new vector, we apply all the networks on it in parallel (and therefore with negligible impact on the latency, which is very important in many applications) and chose the one that attains the smallest value in the objective of the minimization problem \eqref{eq:min_unconst}.
We test this strategy on the {\em house} image by extracting from it patches of size $5 \times 5$, adding random Gaussian noise to each of them with variance $\sigma^2= 25$ and then removing the DC and normalizing each. We take $7/9$ of the patches for training and $1/9$ for validation and testing. We train LISTA to minimize directly the objective \eqref{eq:min_unconst} as in \cite{Sprechmann15Learning} and stop the optimization after the error of the validation set increases.
For the LISTA-MM we use $6$ LISTA networks such that we train the first one on the whole data. We then remove $1/6$ of the data whose objective value in \eqref{eq:min_unconst} is the closest to the one ISTA attains after $1000$ iterations. We use this LISTA network as the initialization of the next one that is trained on the rest of the data. We repeat this process by removing in the same way the part of the data with the smallest relative error and then train the next network. After training $6$ networks we cluster the data points by selecting for each patch the network that leads to the smallest objective error for it in \eqref{eq:min_unconst} and fine tune each network for its corresponding group of patches. We repeat this process $5$ times. The objective error of \eqref{eq:min_unconst} as a function of the number of iterations/depth of the networks is presented in Fig.~\ref{fig:LISTA}. Indeed, it can be seen that partitioning the data, which leads to a better approximation, accelerates convergence.
Our proposed LISTA-MM strategy bears some resemblence to the recently proposed rapid and accurate image super resolution (RAISR) algorithm \cite{Romano17RAISR}. In this method, different filters are trained for different types of patches in natural images. This leads to improved quality in the attained up-scaled images with only minor overhead in the computational cost, leading to a very efficient super-resolution technique.
\section{Conclusion}
\label{sec:conc}
In this work we suggested an approach to trade-off between approximation error and convergence speed. This is accomplished by approximating complicated projections by inexact ones that are computationally efficient. We provided theory for the convergence of an iterative algorithm that uses such an approximate projection and showed that at the cost of an error in the projection one may achieve faster convergence in the first iterations. The larger the error the smaller the number of iterations that enjoy fast convergence. This suggests that if we have a budget for only a small number of iterations (with a given complexity), then it may be worthwhile to use inexact projections which can result in a worse solution in the long term but make better use of the given computational constraints. Moreover, we showed that even when we can afford a larger number of iterations, it may be worthwhile to use inexact projections in the first iterations and then change to more accurate ones at latter stages.
Our theory offers an explanation to the recent success of neural networks for
approximating the solution of certain minimization problems. These networks achieve similar accuracy to iterative techniques developed for such problems (e.g., ISTA for $\ell_1$ minimization) but with much smaller computational cost. We demonstrate the usage of this method for the problem of image super-resolution. In addition, our analysis provides a technique for estimating the solution of these minimization problems by using multiple networks but with fewer layers in each of them.
\appendices
\section{Proof of Theorem~\ref{thm:IPGD_error_t}}
\label{sec:IPGD_error_t_proof}
The proof of Theorem~\ref{thm:IPGD_error_t} relies on the following lemma.
\begin{lem}
\label{lem:Pc_IMM_diff_bound}
Under the same conditions of Theorem~\ref{thm:IPGD_error_t}
\begin{eqnarray}
&& \hspace{-0.35in}\norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)} \\ \nonumber && \hspace{0.8in} \le \rho_p(C)\norm{\vect{z}_t - \vect{x}} + \epsilon(1+\rho_p(C)) \norm{\vect{x}}.
\end{eqnarray}
\end{lem}
\noindent {\it Proof:}
Since $\vect{z}_t = \mathcal P_\mathcal K(p \vect{v})$ for a certain vector $\vect{v}$, we have
\begin{eqnarray}
\label{eq:IPGD_lem_conv_proof_1_pre}
&& \hspace{-0.28in} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)}
\\ && \nonumber \hspace{-0.24in}
\le \sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_{\mathcal K}(p\vect{v})- \vect{x}}
\le \norm{\vect{z}_t - \vect{x}}}} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\left(\mathcal P_{\mathcal K}(p\vect{v})- \vect{x} \right)\right)}
\\ && \nonumber \hspace{-0.24in}
= \sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(p\vect{v}- \vect{x})}
\le \norm{\vect{z}_t - \vect{x}}}} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p\vect{v} - \vect{x})\right)},
\end{eqnarray}
where the last equality follows from Lemma~\ref{lem:Kappa_D_rel}.
Using the triangle inequality with \eqref{eq:IPGD_lem_conv_proof_1_pre} leads to
\begin{eqnarray}
\label{eq:IPGD_lem_conv_proof_1}
&& \hspace{-0.32in} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)}
\\ && \nonumber \hspace{-0.28in} \le
\sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(p\vect{v}- \vect{x})}
\le \norm{\vect{z}_t - \vect{x}}}} \hspace{-0.3in} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p\vect{v} - p\vect{x})\right)}
\\ && \nonumber \hspace{-0.28in}
+ \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\frac{\mathcal P_D(\vect{x} - p\vect{x})}{\norm{\mathcal P_D(\vect{x} - p\vect{x})}}\right)}
\norm{\mathcal P_D(p\vect{x} - \vect{x})}.
\end{eqnarray}
We turn now to bound the first and second terms in the right-hand-side (rhs) of \eqref{eq:IPGD_lem_conv_proof_1}. For the second term, note that since $\mathcal P_D(\vect{x} - p\vect{x}) \in D$, we have
\begin{eqnarray}
\label{eq:IPGD_lem_conv_subproof_rho}
&& \hspace{-0.6in}\norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\frac{\mathcal P_D(\vect{x} - p\vect{x})}{\norm{\mathcal P_D(\vect{x} - p\vect{x})}}\right)} \\
&& \nonumber \le \sup_{\vect{v} \in D \cap \mathbb{B}^d} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\vect{v}\right)} \\
&& \nonumber \le \sup_{\vect{v}, \vect{u} \in C \cap \mathbb{B}^d} \norm{\vect{u}^*p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\vect{v}} \le \rho(\mathcal K),
\end{eqnarray}
where the second inequality follows from Lemma~\ref{lem:P_c_sup} and the fact that $D \subset C$. In the last inequality we replace $\vect{u}^*p = (p^*\vect{u})^*$ by $\tilde{\vect{u}}$ and take the supremum over it instead of over $\vect{u}$. In addition, $ \norm{\mathcal P_D(p\vect{x} - \vect{x})} \le \epsilon$ from \eqref{eq:Dp_eps_ineq}.
For the first term in the rhs of \eqref{eq:IPGD_lem_conv_proof_1} we use the inverse triangle inequality $\norm{\mathcal P_D(p\vect{v}- p\vect{x})} - \norm{\mathcal P_D(p\vect{x}- \vect{x}) } \le \norm{\mathcal P_D(p\vect{v}- \vect{x})}$ and \eqref{eq:Dp_eps_ineq}. Combining the results leads to
\begin{eqnarray}
\label{eq:IPGD_step_error_proof_2_convex}
&& \hspace{-0.3in} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)}
\\ && \nonumber \hspace{-0.3in}
\le \hspace{-0.14in} \sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(p(\vect{v}- \vect{x}))}
\le \norm{\vect{z}_t - \vect{x}}+ \epsilon\norm{\vect{x}}}} \hspace{-0.35in} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p(\vect{v} - \vect{x}))\right)}
\\ && \nonumber \hspace{2.4in}
+ \epsilon \rho(\mathcal K) \norm{\vect{x}}
\\ && \nonumber \hspace{-0.3in}
=\hspace{-0.14in} \sup_{\substack{\vect{u}\in C\cap \mathbb{B}^d,
\vect{v} \in \mathbb R^d \\ \text{ s.t. }\norm{\mathcal P_D(p\vect{v})}
\le 1}} \hspace{-0.15in}
\norm{\vect{u}^*p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p\vect{v})}\left(\norm{\vect{z}_t - \vect{x}}+ \epsilon\norm{\vect{x}}\right)
\\ && \nonumber \hspace{2.4in}
+ \epsilon \rho(\mathcal K) \norm{\vect{x}}
\\ && \nonumber \hspace{-0.3in}
=\hspace{-0.14in} \sup_{\vect{u} \in C\cap \mathbb{B}^d, p\vect{v} \in D\cap \mathbb{B}^d} \hspace{-0.15in}
\norm{(p\vect{u})^*\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)p\vect{v}}\left(\norm{\vect{z}_t - \vect{x}}+ \epsilon\norm{\vect{x}}\right)
\\ && \nonumber \hspace{2.4in}
+ \epsilon \rho(\mathcal K) \norm{\vect{x}}
\\ && \nonumber \hspace{-0.3in}
\le \hspace{-0.14in} \sup_{\vect{v},\vect{u} \in C\cap \mathbb{B}^d}
\norm{(p\vect{u})^*\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)p\vect{v}}\left(\norm{\vect{z}_t - \vect{x}}+ \epsilon\norm{\vect{x}}\right)
\\ && \nonumber \hspace{2.4in}
+ \epsilon \rho(\mathcal K) \norm{\vect{x}},
\end{eqnarray}
where the first equality follows from Lemma~\ref{lem:P_c_sup} and the second from the definition of a projection onto a set. The last inequality follows from the fact that $\{ p\vect{v} \in D \cap \mathbb{B}^d \} \subset D \cap \mathbb{B}^d \subset C \cap \mathbb{B}^d$. Reordering the terms and using the definition of $\rho_p(C)$ leads to the desired result.
\hfill $\Box$
We turn now to the proof of Theorem~\ref{thm:IPGD_error_t}
{\it Proof:} The IPGD error at iteration $t+1$ is,
\begin{eqnarray}
\label{eq:IPGD_error_def}
\norm{\vect{z}_{t+1} - \vect{x}} = \norm{\mathcal P_{\mathcal K}\left(p\left(\vect{z}_t + \mu \matr{M}^*(\vect{y} - \matr{M}\vect{z}_t)\right)\right) - \vect{x}}.
\end{eqnarray}
Using Lemma~\ref{lem:Kappa_D_rel} and the fact that $\vect{y} = \matr{M}\vect{x}$ we have
\begin{eqnarray}
\label{eq:IPGD_step_error_proof_ab}
&& \hspace{-0.33in} \norm{\vect{z}_{t+1} - \vect{x}} \hspace{-0.025in}
= \hspace{-0.025in} \norm{\mathcal P_{D}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x}) - \vect{x} + p\vect{x}\right) }
\\ \nonumber && \overset{(a)}{\le} \norm{\mathcal P_{D}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right) }
+ \norm{\mathcal P_{D}\left(\vect{x} - p\vect{x} \right) }
\\ \nonumber && \overset{(b)}{\le} \norm{\mathcal P_{C}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)}
+ \epsilon\norm{\vect{x}},
\end{eqnarray}
where $(a)$ follows from the convexity of $D$ and the triangle inequality; and $(b)$ from \eqref{eq:Dp_eps_ineq} and Lemma~\ref{lem:C_D_rel}. Using Lemma~\ref{lem:Pc_IMM_diff_bound} with \eqref{eq:IPGD_step_error_proof_ab} leads to
\begin{eqnarray}
\label{eq:IPGD_step_error_proof_iter}
\norm{\vect{z}_{t+1} - \vect{x}} \le \rho_p(\mathcal K)\norm{\vect{z}_t - \vect{x}} + \epsilon(2+\rho_p(\mathcal K)) \norm{\vect{x}}.
\end{eqnarray}
Applying the inequality in \eqref{eq:IPGD_step_error_proof_iter} recursively provides the desired result.
\hfill $\Box$
\section{Proof of Theorem~\ref{thm:IPGD_error_K}}
\label{sec:IPGD_error_K_proof}
The proof Theorem~\ref{thm:IPGD_error_K} relies on the following lemma.
\begin{lem}
\label{lem:PK_IMM_diff_bound}
Under the same conditions of Theorem~\ref{thm:IPGD_error_K},
\begin{eqnarray}
&& \hspace{-0.25in}\norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)} \\ \nonumber && \hspace{0.45in} \le \rho_p(\mathcal K)\norm{\vect{z}_t - \vect{x}} + \epsilon(2\rho(\mathcal K)+\rho_p(\mathcal K)) \norm{\vect{x}}.
\end{eqnarray}
\end{lem}
{\it Proof:}
Using Lemma~\ref{lem:Kappa_D_rel} and the fact that $\vect{z}_t = \mathcal P_\mathcal K (p\vect{v})$ for a certain vector $\vect{v} \in \RR{d}$ and then the triangle inequality, leads to
\begin{eqnarray}
\label{eq:IPGD_lemma_error_proof_1_nonconvex}
&& \hspace{-0.3in} \norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)}
\\ && \nonumber \hspace{-0.3in}
\le\sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(p\vect{v}- \vect{x})}
\le \norm{\vect{z}_t - \vect{x}}}} \hspace{-0.3in} \norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p\vect{v} - \vect{x})\right)}
\\ && \nonumber \hspace{-0.3in}
\le\sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(p\vect{v}- \vect{x})}
\le \norm{\vect{z}_t - \vect{x}}}} \hspace{-0.3in} \norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p\vect{v} - p\vect{x})\right)} +
\\ && \nonumber \hspace{-0.3in}
\norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\frac{\mathcal P_D(p\vect{v} - \vect{x}) - \mathcal P_D(p\vect{v} - p\vect{x})}{\norm{\mathcal P_D(p\vect{v} - \vect{x}) - \mathcal P_D(p\vect{v} - p\vect{x})}}\right)}
\\ && \nonumber \hspace{1.25in}
\cdot \norm{\mathcal P_D(p\vect{v} - \vect{x}) - \mathcal P_D(p\vect{v} - p\vect{x})}.
\end{eqnarray}
Using \eqref{eq:P_D_pv_x_pv_px_epsilon} and the same steps of the proof of Theorem~\ref{thm:PGD_error_K}, we may bound the second term in the rhs of \eqref{eq:IPGD_lemma_error_proof_1_nonconvex} by $2\epsilon \rho(\mathcal K) \norm{\vect{x}}$. This leads to
\begin{eqnarray}
\label{eq:IPGD_lemma_error_proof_2_nonconvex_pre}
&& \hspace{-0.3in} \norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)}
\\ && \nonumber \hspace{-0.3in}
\le\sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(p\vect{v}- \vect{x})}
\le \norm{\vect{z}_t - \vect{x}}}} \hspace{-0.3in} \norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p\vect{v} - p\vect{x})\right)}
\\ && \nonumber \hspace{1.465in}
+ 2\epsilon \rho(\mathcal K) \norm{\vect{x}}.
\end{eqnarray}
From the inverse triangle inequality together with \eqref{eq:P_D_pv_x_pv_px_epsilon}, we have that $\norm{\mathcal P_D(p\vect{v}- \vect{x})} \ge \norm{\mathcal P_D(p\vect{v}- p\vect{x})} - \epsilon \norm{\vect{x}}$. Thus,
\begin{eqnarray}
\label{eq:IPGD_lemma_error_proof_2_nonconvex}
&& \hspace{-0.3in} \norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})\right)}
\\ && \nonumber \hspace{-0.2in}
\le \hspace{-0.23in} \sup_{\substack{\vect{v} \in \mathbb R^d \text{ s.t. } \\ \norm{\mathcal P_D(p(\vect{v}- \vect{x}))}
\le \norm{\vect{z}_t - \vect{x}}+ \epsilon\norm{\vect{x}}}} \hspace{-0.4in} \norm{\mathcal P_{\mathcal K-\mathcal K}\left(p\left( \matr{I} - \mu \matr{M}^*\matr{M}\right)\mathcal P_D(p(\vect{v} - \vect{x}))\right)}
\\ && \nonumber \hspace{1.465in}
+ 2\epsilon \rho(\mathcal K) \norm{\vect{x}}
\\ && \nonumber \hspace{-0.2in}
\le
\rho_p(\mathcal K)\left(\norm{\vect{z}_t - \vect{x}}+ \epsilon\norm{\vect{x}}\right)
+ 2\epsilon \rho(\mathcal K) \norm{\vect{x}},
\end{eqnarray}
where the last inequality follows from the same line of argument used for deriving \eqref{eq:IPGD_step_error_proof_2_convex} in Lemma~\ref{lem:Pc_IMM_diff_bound} (with $\mathcal K-\mathcal K$ instead of $C$).
\hfill $\Box$
We now turn to the proof of Theorem~\ref{thm:IPGD_error_K}.
{\it Proof:} Denoting $\tilde{\vect{v}} = \left( \matr{I} - \mu \matr{M}^*\matr{M}\right)(\vect{z}_t - \vect{x})$, the IPGD error at iteration $t+1$ obeys
\begin{eqnarray}
\label{eq:IPGD_step_error_proof_dc}
&& \hspace{-0.5in} \norm{\vect{z}_{t+1} - \vect{x}}
= \norm{\mathcal P_{D}\left(p\tilde{\vect{v}} - \vect{x} +p \vect{x}\right) }
\\ \nonumber && \hspace{-0.2in}
\overset{(c)}{\le} \norm{\mathcal P_{D}\left(p\tilde{\vect{v}} \right) }
+ \norm{\mathcal P_{D}\left(p\tilde{\vect{v}}- \vect{x} +p \vect{x}\right) - \mathcal P_{D}\left(p\tilde{\vect{v}} \right) }
\\ \nonumber && \hspace{-0.2in}
\overset{(d)}{\le} \kappa_\mathcal K\norm{\mathcal P_{K-K}\left(p\tilde{\vect{v}} \right) }
+ \epsilon\norm{\vect{x} },
\end{eqnarray}
where $(c)$ follows from the triangle inequality; and $(d)$ from \eqref{eq:P_D_pv_x_px_epsilon} and Lemma~\ref{lem:C_D_rel}.
Using Lemma~\ref{lem:PK_IMM_diff_bound} with \eqref{eq:IPGD_step_error_proof_dc}, we get
\begin{eqnarray}
\label{eq:IPGD_step_error_proof_non_convex}
&& \hspace{-0.25in} \norm{\vect{z}_{t+1} - \vect{x}}
\\ \nonumber && \hspace{-0.1in} \le \rho_p(\mathcal K)\kappa_\mathcal K\norm{\vect{z}_t - \vect{x}} + \epsilon(2\rho(\mathcal K)\kappa_\mathcal K+\rho_p(\mathcal K)\kappa_\mathcal K+1) \norm{\vect{x}}.
\end{eqnarray}
Applying \eqref{eq:IPGD_step_error_proof_non_convex} recursively leads to the desired result.
\hfill $\Box$
\section*{Acknowledgments}
RG is partially supported by GIF grant no. I-2432-406.10/2016 and ERC-StG grant no. 757497 (SPADE). YE is partially supported by the ERC grant no. 646804-ERC-COG-BNYQ. AB is partially supported by ERC-StG RAPID. GS is partially supported by ONR, NSF, NGA, and ARO.
We thank Dr. Pablo Sprechmann for early work and insights into this line of research, and Prof. Ron Kimmel and Prof. Gilles Blanchard for insightful comments.
{\small
\bibliographystyle{plain}
|
\section{Introduction}
While the outer parts of protoplanetary discs are prone to the gravitational instability, the inner parts are stable to the gravitational perturbations \citep[e.g.,][]{rafikov2005}. It is known that the onset of the gravitational instability in an accretion disc occurs when the Toomre parameters becomes less than a threshold value around unity and the survival of the newly formed fragments is guaranteed when the cooling time-scale is less than a few dynamical time-scale \citep[e.g.,][]{gammie2001}. Although the {\it dissipationless} gravitational instability is able to explain some of the observational features of structure formation in the protoplanetary discs \citep[e.g.,][]{matzner2005}, presence of the dust particles can introduce new physical mechanisms in order dust particles clump together to form larger objects that may eventually growth into planet embryos \citep[e.g.,][]{chiang2010}. This is mainly because of the interactions between dust particles and the gas. Drag force is proportional to the relative velocity of dust and gas components. But the effect of this exchange of momentum is much stronger on the dust component simply because mass of gas is much larger than the total mass of dust particles. Dynamics of dust particles in a protoplanetray disc is not necessarily the same as gas component. They rotate slower than the local Keplerian velocity because of the pressure gradient which acts opposite to the direction of the central gravitational force. But an individual particle does not accelerate by the pressure gradient when its internal density is much larger than gas density and thereby dust particles rotate at full Keplerian velocity.
Although a few authors had already studied gravitational stability of accretion discs consisting of dust particles and gas \citep[e.g.,][]{coradini81,noh91}, during recent years specific types of instabilities have been identified for clumping of dust particles in the protoplanetary discs which are actually driven by the movement of dust particles through the gas \citep[e.g.,][]{youdin2005,youdin2007,Jac,armit,Laibe} or dust-gas interaction \citep[e.g.,][]{sekiya83,Cuzzi,youdin2011}. Streaming instability has been studied by many authors during recent years in the linear regime and its non-linear evolution investigated via direct numerical simulations.
Drag driven instability is known as secular gravitational instability which is actually the dissipative version of the classical gravitational instability for a dust layer in a fixed background gas component. Dynamics of the dust particles is mostly affected by gas-dust friction driven instabilities. Irrespective of the strength of the self-gravity, this instability is unconditional and can give raise to clumping of dust particles. There are simple theoretical explanations for this trend as have been clarified by \cite{goodman2001} and \cite{Cuzzi}. Radial perturbation leads to concentric rings of dust particles with slightly larger density comparing to their ambient dust density. Particles at the outer edge of a ring feel larger gravitational force due to the accumulated mass of the ring and thereby will rotate faster. Since the drag force is proportional to the velocity, inflow of dust particles increases at the outer edge of the ring. At the inner edge, the particles are orbiting at less than Keplerian velocity because of the extra outward gravitational force. Thus, the particles will therefore be energized by gas drag and will drift toward the ring. No matter how much self-gravity is weak, the mentioned process will eventually give raise to clumping of dust particles.
Most of the previous linear studies of secular gravitational instability assume that dust particles are moving in a fixed background gaseous component \citep[e.g.,][]{Cuzzi,youdin2011,Mich,shadmehri16}. In these models, dust grains are treated as {\it pressure-less fluid}. The nondimensional gas friction time or dimensionless stopping time which is defined as the product of the gas friction time and the Keplerian angular velocity determines gas-dust coupling via the drag force. When dimensionless stopping time is greater than unity, dust particles are decoupled from the gas component and it would not adequate to describe their dynamic using fluid approximation \citep[e.g.,][]{Jalali}.
Neglecting gas dynamics is justified by the fact that the total mass of dust particles is much smaller than the mass of the gaseous component of disc and so only dynamics of dust particles is modified because of drag force. Although this argument seems to be reasonable, just recently \cite{Taka} showed that long-wavelength perturbations are stable when the dynamical feedback from dust grains in the gas component is considered. Their analysis implies that we can not neglect small terms in the equation of motion for small growth rates. Thus, any physical agent that can modify gas dynamics may also affect dust dynamics indirectly via the drag force. Considering the important role of magnetic fields in the structure of protoplanetary discs, it is our motivation to study gravitational instability of a dust layer in a {\it magnetized} gaseous disc which has not been studied before to the best of our knowledge.
Structure of a protoplanetary disc strongly depends on the level of ionization and magnetic fields. External ionization sources such as X-ray radiation from the central star and cosmic rays can efficiently ionize surface layers of a disc. Most regions of a protoplanetary disc (PPD) are weakly ionized, however, which implies that the coupling between the disc material and the magnetic field to be incomplete. This will eventually lead to the non-ideal MHD effects which appear because of the drift velocity between neutral particles and ionized species. There are three non-ideal MHD effects, i.e. the Ohmic resistivity, Hall effect, and ambipolar diffusion. When the density is high and the ionization is very low, the Ohmic term is dominant, but the ambipolar diffusion term influences in the opposite limit. In between these extreme cases, the Hall term plays a significant role. All these non-ideal terms not only significantly modify growth rate of the magnetorotational instability and its non-linear evolution, but also dynamical structure of the disc and launching of winds and outflows are affected by these effects. In this study, we neglect possible role of the non-ideal effects for simplicity. An important mechanism for transporting angular momentum in an accretion disc which leads to accretion is known as magnetorotational instability and operates in weakly ionized discs \citep{balbus91}. Magnetic fields may also provide an efficient mechanism for launching jets or outflows from a disc. Moreover, dynamical structure of a disc is significantly modified in the presence of magnetic fields. Gravitational stability of an accretion disc in the presence of magnetic field has also been studied by many authors \citep[e.g.,][]{Elme,gamm,Fan97,Lizano2010, Lin14}. Many of the previous studies concentrated on analyzing gravitational stability of purely gaseous discs and do not consider dynamics of the dust particles explicitly. \cite{Lizano2010} extended the classical Toomre criterion to a magnetized disc by introducing a modified Toomre parameter which should be greater than one for a gravitationally stable disc. They showed that magnetic tension and pressure stabilize the disc against axisymmetric gravitational perturbations which means magnetic fields suppress gravitational instability in the protoplanetary discs.
In our study, we consider a disc consisting of the magnetized gas and dust where they are coupled via drag force and the particle-to-gas feedback is included. We then explore possible effects of the magnetic fields on gravitational stability of the dust layer using a linear perturbations analysis. In the next section, main assumptions and the basic equations of the model are presented. Linearized equations and the resulting dispersion relation are obtained in section 3. Numerical analysis of the unstable modes and their dependence on the input parameters including strength of the magnetic field are presented in section 4. We conclude with a summary of the results.
\section{General Formulation}
We consider a protoplanetray disc around a central star with mass $M$ as a system consisting of gas and dust components with the momentum exchange. It is assumed that the disc is so thin that the motion of both gas and dust fluids are in the plane of the disc. It means that we do not consider vertical motion of dust particles. Previous linear studies of drag-driven instability in a dust layer have been done in the shearing sheet approximation \citep{gold}. Here, we do not follow this approach. Our linear analysis is performed in cylindrical coordinates $(r,\phi , z)$ where the central star locates at its origin and time-evolution of the perturbations with wavelengths much smaller than the radial distance (i.e., WKB approximation) is studied. Our basic equations for the gas component is similar to \cite{Lizano2010} who studied gravitational stability of a thin and magnetized accretion disc. But we include the drag force due to the interaction with the dust fluid. Since we assume the dust particles are neutral, they do not feel magnetic force.
Thus, basic equations for the gas component are
\begin{equation}
\frac{\partial \Sigma}{\partial t}+{\nabla}.(\Sigma\mathbf{w})=0 ,
\end{equation}
\begin{displaymath}
\Sigma(\frac{\partial\mathbf{w}}{\partial t}+\mathbf{w}.\nabla\mathbf{w})=
\end{displaymath}
\begin{equation}
-\Sigma \nabla (\Phi-\frac{GM}{r})-c_s^2\nabla\Sigma+\frac{1}{4\pi} \int \mathbf{J}\times\mathbf{B} dz+\frac{\Sigma_d(\mathbf{w_d-w)}}{t_{\rm stop}},
\end{equation}
\begin{equation}
\frac{\partial\mathbf{B}}{\partial t}=\nabla\times(\mathbf{w}\times\mathbf{B}),
\end{equation}
\begin{equation}
\nabla.\mathbf{B}=0,
\end{equation}
where $\Sigma$, $\bf{w}$ and $c_s$ are surface density, velocity and the sound speed of gas, respectively. Also, $t_{\rm stop }$ is the stopping time (see below for its definition). It is assumed that the gas is isothermal. Magnetic field of gas is denoted by ${\bf B}$ and the current density is ${\bf J} = \nabla \times {\bf B}$. Note that $\Phi$ is the gravitational potential due to both gas and dust fluids. Note that the equations are integrated perpendicular to the disc so that vertically averaged physical quantities do not depend on the vertical coordinate $z$.
Also, the basic equations for the dust fluid are written as
\begin{equation}
\frac{\partial\Sigma_d}{\partial t}+\nabla.(\Sigma_d\mathbf{w_d})=D\nabla^2\Sigma_d ,
\end{equation}
\begin{displaymath}
\Sigma_d(\frac{\partial\mathbf{w_d}}{\partial t}+\mathbf{w_d}.\nabla\mathbf{w_d})=
\end{displaymath}
\begin{equation}
-\Sigma_d \nabla (\Phi-\frac{GM}{r})+\frac{\Sigma_d(\mathbf{w-w_d)}}{t_{\rm stop}},
\end{equation}
where ${\bf w_d}$ is dust velocity and $D$ is the radial diffusivity of the dust component because of the gas turbulence. The diffusion of dust particles due to stochastic forcing by gas turbulence has been studied by many authors \citep[e.g.,][]{youdin2007}. According to Equation (36) of Youdin \& Lithwick (2007), the radial diffusion coefficient $D$ is written as
\begin{equation}
D=\frac{1+\tau +4\tau^2}{(1+\tau^2)^2}D_g,
\end{equation}
where $D_g$ is is the strength of turbulent diffusion in the gas which can be defined as
\begin{equation}
D_g=\alpha c_s^2 \Omega^{-1},
\end{equation}
where $\alpha$ is the dimensionless measure of turbulent intensity. Strength of dust diffusion is measured by the dimensionless diffusivity coefficient $\xi$ as $\xi = D/( c_{\rm s}^2 \Omega^{-1})$. Moreover, $\tau$ is the dimensionless stopping time (see below). We note that equation of continuity with the diffusion term is not used commonly. In fact, one can start from the Boltzmann equation to obtain the above hydrodynamical equations which leads to viscosity in equation of motion instead of the diffusion term in equation of continuity. Following previous works \cite[e.g.,][]{Taka} we also used this problematic formulation, although these aspects of the work need further studies.
In the above equations, $t_{\rm stop }$ is the stopping time which is a time-scale for decay of relative velocity between the gas and the dust due to the drag force. We can then define nondimensional stopping time $\tau = t_{\rm stop} \Omega_{\rm K}$ \citep[e.g.,][]{miyake15}, where angular Keplerian velocity is $\Omega_{\rm K}=\sqrt{GM/r^3}$. If we assume that all dust particles are spherical with the same radius $a$ and homogeneous internal density $\rho_{\rm m}$, then the nondimensional stopping time becomes $\tau = [\rho_{\rm m} a /(\rho_{\rm g} c_{\rm s})] \Omega_{\rm K}$ where $\rho_{\rm g}$ is the gas density. Note that this relation is valid when the size of the particles is smaller than the mean free path of the gas. For instance, in the minimum mass solar nebula (MMSN) model of \cite{hayashi81} at the radial distances larger than 1 AU from the central star with one solar mass, the mean free path of gas is larger than 1 cm which implies that the above relation for the stopping time is applicable to the particles smaller than this length. Physical properties of the disc and dust distribution depend on the vertical location as well. But we do not consider vertical variation of the physical quantities and one can then evaluate the nondimensional stopping time at the midplane of MMSN \citep{miyake15}:
\begin{equation}
\tau = 1.8 \times 10^{-7} \left(\frac{a}{1 {\rm \mu m}} \right) \left(\frac{r}{1 {\rm AU}} \right)^{\frac{3}{2}}.
\end{equation}
The internal density of a dust particle is assumed to be $\rho_{\rm m}=2$ g ${\rm cm}^{-3}$ and the surface density and the sound speed obey power-law functions of the radial distance \citep{hayashi81}:
\begin{equation}
\Sigma (r) = 1.7 \times 10^{3} \left( \frac{r}{1 {\rm AU}}\right)^{-\frac{3}{2}} {\rm g} {\rm cm}^{-2},
\end{equation}
\begin{equation}
c_{\rm s} (r) = 1.0\times 10^{5} \left(\frac{r}{1 {\rm AU}} \right)^{-\frac{1}{4}} {\rm cm} {\rm s}^{-1}.
\end{equation}
Note that our study is a {\it local} linear perturbation analysis based on WKB approximation which means that we do not consider radial dependence of the initial equilibrium state. But the above physical profiles specify how properties of a disc can vary with the radial distance. It can then be used to calculate growth rate of the unstable modes at a certain radial distance. As for the initial magnetic field, we assume the disc is threaded by a net large-scale vertical field $B_{\rm z0}$ so that the ratio of the gas pressure to the magnetic pressure $\beta$ at the midplane of the disc is uniform through the disc. We then have
\begin{equation}
B_{\rm z0} (r) = 590 \left( \frac{\beta}{1000}\right)^{-\frac{1}{2}} \left(\frac{r}{1 {\rm AU}} \right)^{-\frac{13}{8}} {\rm m G}.
\end{equation}
Finally, our system of equations is closed with the Poisson equation for a thin disc which is written as
\begin{equation}\label{eq:pois}
\nabla^2\Phi=4 \pi G(\Sigma+\Sigma_d)\delta(z).
\end{equation}
Here, gravitational potential $\Phi$ due to both the gas and the dust components is considered.
\cite{Lizano2010} studied gravitational instability of a gaseous magnetized disc but without dust particles. They vertically averaged all basic equations including the Lorentz term in the equation of motion. We generalize their final main equations for a magnetized vertically averaged equations to include dust particles and their momentum exchange with the gas component.
Thus, the continuity equation for the gas is
\begin{equation}
\frac{\partial\Sigma}{\partial t}+\frac{1}{r}\frac{\partial}{\partial r}(r\Sigma u)+\frac{1}{r}\frac{\partial}{\partial\varphi}(\Sigma v)=0,
\end{equation}
where $u$ and $v$ are the radial and the azimuthal components of gas velocity ${\bf w}$. The components of radial and azimuthal Lorentz force are:
\begin{displaymath}
\int_{-H}^{H} \mathbf{J}\times\mathbf{B} dz=\int_{-H}^{H} [(B_z\frac{\partial B_r}{\partial z}-B_z\frac{\partial B_z}{\partial r}-\frac{B_\varphi}{r}\frac{\partial (rB_\varphi)}{\partial r}+\frac{B_\varphi}{r}\frac{\partial B_r}{\partial\varphi})\mathbf{e}_r
\end{displaymath}
\begin{equation}
-(\frac{B_z}{r}\frac{\partial B_z}{\partial\varphi}-B_z\frac{\partial B_\varphi}{\partial z}-\frac{B_r}{r}\frac{\partial(r B_\varphi)}{\partial r}+\frac{B_r}{r}\frac{\partial B_r}{\partial\varphi}) \mathbf{e}_{\varphi} ] dz,
\end{equation}
where $\mathbf{e}_r$ and $\mathbf{e}_{\varphi}$ are unit vectors in the radial and the azimuthal directions, respectively. We assume the toroidal component of the magnetic field is negligible, i.e. $B_\varphi =0$. This simplifying assumption not only simplifies the main equations, but it prevents emergence of the magnetorotational instability (MRI) modes in our analysis. In order to understand the dynamics in a magnetized disc, we note that MRI has a vital role. However, our purpose is to illustrate and understand the basic mechanism of the SGI with the magnetic field in the absence of MRI modes. Then the components of the Lorentz force become
\begin{equation}
\frac{1}{4\pi}\int_{-H}^{H} (\mathbf{J}\times\mathbf{B})_r dz=
\frac{1}{4\pi}\int_{-H}^{H} B_z\frac{\partial B^{+}_{r}}{\partial z}dz-\frac{1}{4\pi}\int_{-H}^{H} B_z\frac{\partial B_z}{\partial r}dz,
\end{equation}
and
\begin{equation}
\frac{1}{4\pi}\int_{-H}^{H} (\mathbf{J}\times\mathbf{B})_\varphi dz=-\frac{1}{4\pi}\int_{-H}^{H} (\frac{B_z}{r}\frac{\partial B_z}{\partial \varphi}+\frac{B^{+}_{r}}{r}\frac{\partial B^{+}_{r}}{\partial \varphi})dz.
\end{equation}
By integrating over $z$, we obtain
\begin{equation}
\frac{1}{4\pi}\int_{-H}^{H} (\mathbf{J}\times\mathbf{B})_r dz=\frac{B_z B_r^+}{2\pi}-\frac{H}{4\pi}\frac{\partial B_z^2}{\partial r},
\end{equation}
and
\begin{equation}
\frac{1}{4\pi}\int_{-H}^{H} (\mathbf{J}\times\mathbf{B})_\varphi dz=-\frac{H}{4\pi r}\frac{\partial}{\partial\varphi}(B_r^{+2}+B_z^2).
\end{equation}
Components of equation of motion for the gas are also written as
\begin{displaymath}
\Sigma[\frac{\partial u}{\partial t}+u\frac{\partial u}{\partial r}+\frac{v}{r}\frac{\partial u}{\partial \varphi}-\frac{v^2}{r}]=
\end{displaymath}
\begin{equation}
-c_s^2\frac{\partial\Sigma}{\partial r}-\Sigma\frac{\partial V}{\partial r}+\frac{B_zB_r^+}{2\pi}-\frac{H}{4\pi}\frac{\partial B_z^2}{\partial r}+\frac{\Sigma_d(u_d-u)}{t_{\rm stop}},
\end{equation}
and
\begin{displaymath}
\Sigma[\frac{\partial v}{\partial t}+u\frac{\partial v}{\partial r}+\frac{v}{r}\frac{\partial v}{\partial\varphi}+\frac{uv}{r}]=
\end{displaymath}
\begin{equation}
-\frac{c_s^2}{r}\frac{\partial\Sigma}{\partial\varphi}-\frac{1}{r}\Sigma(\frac{\partial V}{\partial\varphi})-\frac{H}{4\pi r}\frac{\partial}{\partial\varphi}(B_r^{+2}+B_z^2)+\frac{\Sigma_d(v_d-v)}{t_{\rm stop}}.
\end{equation}
The disc scale-height is $H=c_{\rm s}/\Omega_{\rm K}$. Also, $B_{r}^{+}$ is the radial component of the magnetic field at the surface of the disc. Note that $V$ is the gravitational potential due to the central star and the components of the disc itself, i.e. $V=-GM/r + \Phi $, where $\Phi$ satisfies Poisson's equation (\ref{eq:pois}).
The induction equation becomes
\begin{equation}
-\frac{\partial B_z}{\partial t}+\frac{1}{r}[\frac{\partial}{\partial r}(r B_z u)+\frac{\partial}{\partial\varphi}(B_z v)]=0,
\end{equation}
Continuity equation for the dust fluid is
\begin{displaymath}
\frac{\partial\Sigma_d}{\partial t}+\frac{1}{r}\frac{\partial}{\partial r}(r\Sigma_d u_d)+\frac{1}{r}\frac{\partial}{\partial\varphi}(\Sigma_d v_d)=
\end{displaymath}
\begin{equation}
D[\frac{1}{r}\frac{\partial}{\partial r}(r\frac{\partial \Sigma_d}{\partial r})+\frac{1}{r^2}\frac{\partial^2\Sigma_d}{\partial\varphi^2}],
\end{equation}
and the components of equation of motion for the dust fluid are
\begin{equation}
\Sigma_d[\frac{\partial u_d}{\partial t}+u_d\frac{\partial u_d}{\partial r}+\frac{v_d}{r}\frac{\partial u_d}{\partial \varphi}-\frac{v_d^2}{r}]=-\Sigma_d\frac{\partial V}{\partial r}+\frac{\Sigma_d(u-u_d)}{t_{\rm stop}},
\end{equation}
and
\begin{equation}
\Sigma_d[\frac{\partial v_d}{\partial t}+u_d\frac{\partial v_d}{\partial r}+\frac{v_d}{r}\frac{\partial v_d}{\partial\varphi}+\frac{u_d v_d}{r}]=-\frac{\Sigma_d}{r}(\frac{\partial V}{\partial\varphi})+\frac{\Sigma_d(v-v_d)}{t_{\rm stop}}.
\end{equation}
We note that dust particles are assumed to be neutral, and so, they do not experience magnetic force. Moreover, our dusty fluid is pressure-less and for this reason gradient of pressure does not appear in the above equation of motion.
\section{Linear Perturbations}
Having basic MHD equations including dust contributions, we can now perturb all physical quantities around a uniform equilibrium configuration and then investigate their fate provided that perturbations are much smaller than the initial state. This kind of linear analysis will lead to a dispersion relation to specify unstable modes and their growth rates. The subscripts 0 and 1 are used to denote the initial state and the perturbed quantities, respectively. Equilibrium states must satisfy continuity, motion and induction equations. We know that initial states are independent of $t$,$\varphi$ and $z$ and the velocities of dust and gas components are assumed to be same initially. We assume that $\Sigma_0$ is independent of $r$. Moreover, we assume that the radial component of the magnetic field at the surface of the disc is negligible for simplicity. Also, the initial vertical component of the magnetic field is considered to be independent of the radial distance. Then, equilibrium state satisfied all equations automatically except radial component of motion for gas and dust. The zeroth order of radial component of equation of motion for the gas is
\begin{equation}
\Omega^2r-\frac{c_s^2}{\Sigma_0}\frac{\partial \Sigma_0}{\partial r}-\frac{\partial V_0}{\partial r}+\frac{B_{0r}^+B_{0z}}{2\pi\Sigma_0}-\frac{H}{4\pi\Sigma_0}\frac{\partial B_{0z}^2}{\partial r}=0,
\end{equation}
which reduces to $\Omega^2 r-\frac{\partial V_0}{\partial r}=0$ subject to our mentioned simplifying assumptions.
Our linear perturbation of a physical quantity $X$ is $X=X_0 + X_1 e^{{\rm i} (\omega t +kr- m \varphi)}$, where $\omega$ is the frequency and $k$ is the radial wavenumber and $m$ is a positive integer for nonaxisymmetric perturbations and $m=0$ for axisymmetric modes. Here, we consider only axisymmetric perturbations. Following \cite{Lizano2010}, we also make further assumption that $|k|r \gg 1$, which means wavelength of the perturbations is much smaller than the radial distance.
Thus, the linearized dynamical equations for axisymmetric modes expand to
\begin{equation}
\omega\frac{\Sigma_1}{\Sigma_0}+k u_1=0,
\end{equation}
\begin{displaymath}
i\omega u_1-2\Omega v_1+ikc_s^2\frac{\Sigma_1}{\Sigma_0}+
\end{displaymath}
\begin{equation}
ikV_1+i(1+kH)\frac{B_{z0}B_{1z}}{2\pi \Sigma_0}-\frac{Z(u_{1d}-u_1)}{t_{\rm stop}}=0,
\end{equation}
\begin{equation}
i\omega v_1+u_1\frac{\kappa^2}{2\Omega}-\frac{Z(v_{1d}-v_1)}{t_{\rm stop}}=0,
\end{equation}
\begin{equation}
i\omega B_{1z}+ikB_{z0}u_1=0,
\end{equation}
\begin{equation}
(i \omega +Dk^2 )\frac{\Sigma_{1d}}{Z\Sigma_0}+iku_{1d}=0,
\end{equation}
\begin{equation}
i \omega u_{1d}-2\Omega_K v_{1d}+ikV_1-\frac{(u_1-u_{1d})}{t_{\rm stop}}=0,
\end{equation}
\begin{equation}
i \omega v_{1d}+u_{1d}\frac{\kappa^2}{2\Omega_K}-\frac{(v_{1}-v_{1d})}{t_{\rm stop}}=0,
\end{equation}
\begin{equation}
V_1+\frac{2\pi G}{k}(\frac{\Sigma_1}{1+kH}+\frac{\Sigma_{1d}}{1+kH_{d}})=0,
\end{equation}
where $H_{d}$ is the dust scale height $H_d=\sqrt{\frac{\alpha}{\tau}}H$ and $Z$ is the ratio of the dust density to the gas density or disc metallicity for the initial state, i.e. $Z=\Sigma_{0d}/ \Sigma_0$. Moreover, $\lambda$ is the dimensionless mass-to-flux ratio and is defined as
\begin{equation}
\lambda=\frac{2\pi G^\frac{1}{2}\Sigma_0}{B_{z0}}.
\end{equation}
The additional parameter resulting in our analysis is the magnetically modified Toomre parameter $Q_{\rm M}$, i.e.
\begin{equation}
Q_{\rm M}=\frac{\Theta^\frac{1}{2}c_s \kappa}{\pi G \epsilon \Sigma_0},
\end{equation}
where $\Theta=1+\frac{B_{z0}^2 H}{2\pi c_s^2 \Sigma_0}$ and $\epsilon = 1-\frac{1}{\lambda^2 }$. Also, $\kappa$ is the epicyclic frequency where in the absence of the magnetic effects it becomes the Keplerian angular velocity. But magnetic forces reduce the epicyclic frequency, though its exact value depends on the geometry of the magnetic configuration. \cite{Lizano2010} approximated the epicyclic frequency as $\kappa=f\Omega$ where $f$ is a number less than unity. Obviously, we can assume $f\simeq 1$ for the weak magnetic fields.
If we introduce the nondimensional growth rate and the nondimensional wavenumber as $x= i\omega /\Omega$ and $y = kH$, then we can re-write the above linearized equations:
\begin{equation}
x \frac{\Sigma_1}{\Sigma_0}+i\frac{y}{c_s}u_1=0 ,
\end{equation}
\begin{displaymath}
[x+\frac{Z}{f\tau}+(\frac{2y(1+y)}{\lambda^2 x})(\frac{\Theta^\frac{1}{2}}{Q_M\epsilon})]u_1-2v_1+
\end{displaymath}
\begin{equation}
iyc_s\frac{\Sigma_1}{\Sigma_0}+i\frac{y}{c_s}V_1-\frac{Z}{f\tau}u_{1d}=0 ,
\end{equation}
\begin{equation}
[x+\frac{Z}{f\tau}]v_1+\frac{1}{2}u_1-\frac{Z}{f\tau}v_{1d}=0,
\end{equation}
\begin{equation}
x B_{1z}+i\frac{y}{c_s}B_{z0}u_1=0,
\end{equation}
\begin{equation}
[ x+\xi y^2]\frac{\Sigma_{1d}}{Z\Sigma_0}+i\frac{y}{c_s}u_{1d}=0,
\end{equation}
\begin{equation}
[x+\frac{1}{f\tau}]u_{1d}-\frac{2}{f} v_{1d}+i\frac{y}{c_s}V_1-\frac{u_1}{f\tau}=0,
\end{equation}
\begin{equation}
[x+\frac{1}{f\tau}]v_{1d}+\frac{f}{2}u_{1d}-\frac{v_1}{f\tau}=0,
\end{equation}
\begin{equation}
V_1+\frac{2\Theta^\frac{1}{2}c_s^2 }{Q_M \epsilon \Sigma_0 y}(\frac{\Sigma_1}{1+y}+\frac{\Sigma_{1d}}{1+\sqrt{\frac{\alpha}{\tau}}y})=0.
\end{equation}
Thus, we have eight equations and eight unknowns, i.e. $\Sigma_{1}$, $\Sigma_{\rm 1d}$, $v_1$, $u_1$, $v_{\rm 1d}$, $u_{\rm 1d}$, $V_1$, $B_{\rm 1z}$. Since the above linearized equations are valid for the perturbations with a wavelength much smaller than radial distance, i.e. $kr\gg 1$, then we can consider perturbations which satisfy this inequality: $kH=y\gg H/r$. For instance, in a thin disc with $H/r = 0.1$, we consider only perturbations which are larger than this value, i.e. $y\gg 0.1$. In addition to this constraint of local approximation, the validity of the vertically integrated equations requires $k u<\Omega_{\rm K}$ \citep[e.g.,][]{wu,kato}. this requirement can be written as $kH<(\alpha H/r)^{-1}$ where $\alpha$ is the disc viscosity. Thus, our analysis is valid for $(H/r)\ll y < (\alpha H/r)^{-1}$. If we set $\alpha =0.01$ and $H/r =0.1$, then the valid range of nondimensional wavenumber becomes $0.1\ll y <10^3$. Also, validity of vertical integrated equations implies that the growth rates of the unstable modes are less than angular velocity \citep{kato}. This requirement is justified by the unstable modes as we will show.
Existence of a set of nontrivial solutions for the above linearized equations imply that the determinant of the coefficients becomes zero which gives us an algebraic equation involving the input parameters, growth rate and the wavenumber of the perturbations. Using MAPLE software, we found the dispersion relation. But the equation is very lengthy, and so, we do not bring it here. However, our analysis is based on the roots of this equation which can be calculated numerically. Obviously, unstable modes correspond to the roots with positive real part, i.e. ${\rm Re}(x) >0$. We generally found one or two unstable modes for a given set of the input parameters.
It is useful to re-write the input parameters as follows \citep{Lizano2010}:
\begin{equation}\label{eq:lambda}
\lambda=2.71\mu(\frac{N_{\rm H}}{10^{24} {\rm cm}^2})(\frac{B_{z0}}{\rm 1 mG})^{-1},
\end{equation}
\begin{equation}\label{eq:Theta}
\Theta=1+1.15\times 10^{-2}(\frac{B_{z0}}{\rm 1 mG})^2(\frac{H}{\rm 1 AU})(\frac{N_H}{10^{24} {\rm cm}^2})^{-1}(\frac{T}{\rm 1 K})^{-1},
\end{equation}
\begin{equation}\label{eq:epsilon}
\epsilon=1-1.36\times 10^{-1}(\frac{1}{\mu^2})(\frac{B_{z0}}{\rm mG})^2(\frac{N_{\rm H}}{10^{24} {\rm cm}^2})^{-2},
\end{equation}
\begin{equation}\label{eq:QM}
Q_M=2.12(\frac{\Theta^\frac{1}{2}}{\epsilon\mu^\frac{3}{2}})(\frac{\Omega}{10^{-2} {\rm km} {\rm s}^{-1} {\rm AU}^{-1}})(\frac{T}{\rm 1K})^\frac{1}{2}(\frac{N_{\rm H}}{10^{24} {\rm cm}^2})^{-1},
\end{equation}
where $\mu$ is the molecular weight, $N_H$ is the hydrogen column density, $T$ is the gas temperature and $\Omega$ is the angular velocity.
\begin{table}
\begin{center}
\begin{tabular}{c | c | c | c | c | c }
\hline
$B_{z0} {\rm (mG)}$& $\lambda$& $\Theta$& $\epsilon$& $f$& $Q_M$ \\ \hline
0& $\infty$& 1& 1& 1& 1.99\\
5& 4.35& 1.05& 0.94& 1& 2.16\\
10& 2.17& 1.21& 0.78& 1& 2.79\\
15& 1.45& 1.48& 0.52& 1& 4.9 \\
\end{tabular}
\end{center}
\caption{Input parameters for different values of the initial vertical magnetic field $B_{z0}$.}
\end{table}
\section{Analysis}
\begin{figure
\includegraphics[scale=0.7]{f1.eps}
\caption{ Dispersion relation of the unstable modes for different values of the magnetic strength and dimensionless stopping time. Each curve is labeled by its corresponding value of $B_{z0}$. Here, magnetic Toomre parameter is $Q_{M}=6$ which is larger than the threshold for the instability (see Table 1).}
\label{fig:1}
\end{figure}
\begin{figure
\includegraphics[scale=0.75]{f2.eps}
\caption{ Same as Figure \ref{fig:1}, but surface density of dust particles is larger. }
\label{fig:2}
\end{figure}
\begin{figure
\includegraphics[scale=0.4]{f3.eps}
\caption{ Growth rate of the unstable modes as a function of the normalized wavenumber in the minimum mass solar nebula at the radial distance 100 AU where the Toomre parameter is $Q_M =17$. Other input parameters are same as previous figures, and different values for the dimensionless stopping time $\tau$ are adopted. Each curve is labeled by the corresponding value of the vertical magnetic field and black curve is for the non-magnetized case. We found that for magnetic field strength larger than 2 mG, the parameter $\epsilon$ becomes negative which is not acceptable.}
\label{fig:3}
\end{figure}
We can now investigate axisymmetric unstable modes for different sets of the input parameters in order to explore possible effects of the magnetic field on the drag-driven instability. Nonzero values for $m$ do not affect essentially behavior of the unstable solutions. In Figures 1 and 2, we assume $H = 143 {\rm AU}$, $T=250$ K, $\mu = 2.33$ and $N_{\rm H}=3.46\times 10^{24} {\rm cm}^{-2}$ \citep{Lizano2010}. Corresponding to these input parameters, one can calculate the other parameters based on equations (\ref{eq:lambda})-(\ref{eq:QM}) for different values of the initial vertical magnetic field (Table 1). We first examine the dependence of the growth rate on the grain size, or dimensionless stopping time. Figure \ref{fig:1} shows unstable growth rate for particles with different sizes, ranging from strongly coupled particles with dimensionless stopping time $\tau =5\times 10^{-3}$ (top panel) and $\tau = 10^{-2}$ (middle panel) to a slightly less coupled case with $\tau = 10^{-1}$ (bottom panel). Each curve is labeled by its corresponding value of $B_{ z0}$. In this figure, the standard value of disc metallicity is adopted, i.e. $Z=0.01$, and Toomre parameter is $Q_M =6$ and $\alpha=10^{-6}$. Here, different values of $\tau$ are considered. Note that our adopted value of the Toomre parameter is greater than the threshold of the instability which means that the system is stable in the absence of dust particles. For some of the input parameters, we found two unstable roots where one root is much smaller than the other one. We actually displayed both roots, though the larger root which specifies the most unstable root is more interesting. We consider different values of the disc metalicity (i.e., $Z=0.01$ and $Z=0.1$) in our analysis to explore possible effects of its variations on the instability. Figure \ref{fig:1} shows that the instability occurs in the presence of the magnetic fields, and as the strength of the magnetic field increases the instability grows faster. This feature is understandable by the fact that the critical value of the Toomre parameter for the instability in the absence of the dust particles increase with the magnetic field (see Table 1). Since we assume a fixed value for the Toomre parameter, the system becomes closer to the threshold of the instability with increasing the magnetic field. Figure \ref{fig:2} is same as Figure \ref{fig:1}, but the disc metallicity is larger, i.e. $Z=0.1$. Again, the system is unstable in the presence of the dust particles. Figure \ref{fig:1} shows that when the disc contains aerodynamically well-coupled dust particles, even weak magnetic fields can destabilize the system considerably. For example, in the strongly coupled case (top panel), the system is unstable in the presence of the magnetic field. But as the level of dust-gas coupling reduces, the growth rate increases. Moreover, wavelength of the most unstable mode increases with the magnetic field strength when the particles are well-coupled to the gas. Note that in all cases, the corresponding modified Toomre parameter $Q_M$ is larger than one (see Table 1).
Figure 3 shows growth rate of the instability in the minimum mass solar nebula at the radial distance 100 AU where the Toomre parameter is $Q_M =17$. The rest of the input parameters are the same as previous figures. At the radial distance $100 AU$, when we have $\tau=0.04$, the size of the dust particles is $a=222$ $\mu {\rm m}$ and the most unstable wavelength is around $1 AU$ and the corresponding growth time is $0.13\times 10^6$ years. For $\tau=0.1$, the size of the dust particles is $a=555$ $\mu {\rm m}$ and the most unstable wavelength and the corresponding growth time become $2.45 AU$ and $ 10^6$ years, respectively.
Cosmic rays and radiation of the central star are the main sources of ionization in a protoplanetary disc. It is known that there is a region in a protoplanetary disc where neither cosmic rays can penetrate to ionize the gas nor the radiation of the central star is able to ionize the gas. This non-ionized region which is magnetically {\it inactive} is called dead zone \citep{Gam}. But interior to the dead zone or beyond that region, the gaseous component of the disc is magnetically active. Our analysis shows that drag-driven instability is more efficient in the magnetized regions comparing to the regions where magnetic fields does not play a significant role. Although the numerical values adopted in previous figures are certainly subject to uncertainties, our analysis serves as a proof of concept to illustrate the important role of the magnetic field in the drag-driven instability in protoplanetary discs.
\section{conclusion}
We surveyed linear instability of a dust layer in a magnetized gaseous disc for the perturbations with wavelengths much small than the radial distance. One of the interesting findings in this paper is that magnetic field can amplify the instability for even a weak gas-dust coupling. In particular, we showed that for well-coupled particles, even a weak magnetic field is able to amplify the instability and leads to a completely unstable system. Our study shows that the greatest response for axisymmetric perturbations occurs at large wavelengths. We also found that in the presence of magnetic fields enhancing the disc metallicity promotes the instability because this enhancement leads to stronger self-gravity of particles and slower radial drift.
Time-scale of drag driven instability should be shorter than radial drift time-scale if the instability is responsible for the planetesimal formation. Based on this physical constraint, the minimum dust abundance for planetesimal formation via secular gravitational instability has been estimated by \cite{Takeuchi}. Considering the destabilizing role of magnetic field, however, we think this minimum dust abundance is modified if magnetic fields are considered.
\section*{Acknowledgments}
We are very grateful to anonymous referee for his/her very useful comments and suggestions which greatly helped us to improve the paper. MS is grateful to Prof. Shu-ichiro Inutsuka for his useful comments on the early version of this manuscript.
\bibliographystyle{raa}
|
\section{Introduction}
For complex diseases, it is of significant interest to identify genetic risk factors. For etiology, biomarkers, and prognosis, the interactions between genetic and environmental risk factors, also referred to as G $\times$ E interactions, have important implications beyond the main effects. Extensive studies have been conducted to search for important G $\times$ E interactions \cite{cordell2009detecting,hunter2005gene,north2008importance,wu2014selective}. In this article, we focus on analyzing prognosis, which has an essential role in biomedical studies. It is conjectured that the proposed approach can be extended to some types of disease outcomes.
Denote $T$ as the survival time, $Z =(Z_{1},\cdots,Z_{p})^\top\in R^{p\times 1}$ as the $p$ genetic factors (G), and $X = (X_{1},\cdots,X_{q})^\top \in R^{q\times1}$ as the $q$ clinical/environmental risk factors (E). There are two families of methods for detecting the main G and E effects and G$\times$E interactions. The first conducts marginal analysis and analyzes one or a small number of G factors at a time \cite{hunter2005gene,shi2014penalized,thomas2010methods}. With a slight abuse of terminology, we use the generic phrase ``gene" for the G factor. In marginal analysis, for gene $k$, consider the model $T\sim \phi(\sum_{j=1}^qX_j\alpha_j+Z_k\beta_k+\sum_{j=1}^q X_jZ_k\xi_{jk})$, where model $\phi$ is known up to the regression coefficients, and $\alpha_j,\beta_k,\xi_{jk}$ are the unknown regression coefficients. Marginal analysis is easy to implement however cannot accommodate the joint effects of multiple main effects and interactions. The second family conducts joint analysis and describes the joint effects of all factors in a single model \cite{liu2013identification, wu2014integrative, zhu2014identifying}. More specifically, consider $T\sim \phi(\sum_{j=1}^qX_j\alpha_j+\sum_{k=1}^pZ_k\beta_k+\sum_{j=1}^q\sum_{k=1}^p X_jZ_k\xi_{jk})$.
For the simplicity of description, consider the simple linear regression setting. In the literature, most of the existing methods adopt loss functions built on {\it absolute error criteria}, with the most popular including the least squares (LS) and least absolute deviation (LAD). Under certain settings, it has been found that the {\it relative errors} are more sensible \cite{khoshgoftaar1992predicting, park1998relative, van2015relative}. The most distinguishable feature of the relative error-based approaches is that they are scale-free, which, as discussed in the published studies \cite{chen2010least}, can be advantageous in survival and other analysis.
There are at least two ways of defining relative error-based criteria. The first is defined based on the ratio of the error with respect to the target. The second is defined on the ratio of the error with respect to the predictor \cite{chen2010least}. Based on the two types of relative errors, researchers have proposed the least absolute relative errors (LARE) criterion and the least product relative errors (LPRE) criterion for linear multiplicative models. The LARE criterion is convex but not smooth. For its extensions and applications, we refer to \cite{li2014empirical, tsionas2014bayesian, zhang2013local} and followup studies. In comparison, the LPRE criterion is smooth and convex \cite{chen2013least}. Under low-dimensional settings, asymptotical properties of the LARE and LPRE estimates for linear multiplicative models have been established \cite{chen2010least, chen2013least}.
Different from the existing ones (which focus on the main effects), this study adopts the relative error-based criteria for analyzing interactions. Such new criteria may provide a useful alternative to the commonly-adopted absolute error-based criteria. In genetic data analysis, it is critical to identify the important main effects and interactions, which poses a variable (model) selection problem. In two recent studies \cite{xia2015regularized, zhang2013local}, variable selection based on the LARE has been studied. However, the existing studies are limited to the situation where the dimension of model is smaller than the sample size. To the best of our knowledge, there is a lack of study examining the relative error-based criteria under high-dimensional settings. Also different from the existing studies, we analyze prognosis data under right censoring, which introduces additional complexity.
\section{Methods}
\subsection{Model and relative error-based criteria}
For modeling a prognosis response, we consider the following linear multiplicative (accelerated failure time - AFT) model,
\begin{equation}\label{AFT}
T=\exp\Big(\sum_{j=1}^qX_j\alpha_j+\sum_{k=1}^pZ_k\beta_k+\sum_{j=1}^q\sum_{k=1}^p X_jZ_k\xi_{jk}\Big)\varepsilon,
\end{equation}
where $\varepsilon$ is the random error independent of $X$ and $Z$. This model provides a useful alternative to the Cox and other models. It can be especially preferred under high-dimensional settings. Let $U=(X^\top, Z^\top, (X\otimes Z)^\top)^\top$ and $\theta=(\alpha^\top, \beta^\top,\xi^\top)^\top$, then we can write model (\ref{AFT}) as
\begin{equation}\label{AFT2}
T=\exp(U^\top\theta)\varepsilon.
\end{equation}
First consider the case without censoring. Suppose that we have $n$ iid observations $\{t_{i},\textbf{x}_{i},\textbf{z}_{i}\}_{i=1}^n$, where $\textbf{x}_{i} = (x_{i1}, \cdots, x_{iq})^\top$ and $\textbf{z}_{i} = (z_{i1}, \cdots, z_{ip})^\top$. Denote $\textbf{u}_{i} = (\textbf{x}_{i}^\top, \textbf{z}_{i}^\top, (\textbf{x}_{i}\otimes \textbf{z}_{i})^\top)^\top$. With the logarithm transformation, model (\ref{AFT2}) can be rewritten as $\log(T)=U^\top\theta+\log(\varepsilon)$. The LS and LAD methods can be applied, which, respectively, minimize the objective functions $\sum_{i=1}^n (\log(t_i)-\textbf{u}_i^\top\theta)^2$ and $\sum_{i=1}^n |\log(t_i)-\textbf{u}_i^\top\theta|$. Both methods are built on the absolute errors.
As discussed in the literature, under certain scenarios, the relative error-based criteria can be more sensible. In this article, we consider the least absolute relative errors (LARE) \cite{chen2010least} and least product relative errors (LPRE) \cite{chen2013least} criteria. They have been relatively more popular in the relative error literature and deserve a higher priority. The LARE objective function is defined as
\begin{equation}\label{LARE}
LARE_n(\theta)=\sum_{n=1}^n \left\{\left|\frac{t_i-\exp(\textbf{u}_i^\top \theta)}{t_i}\right| + \left|\frac{t_i-\exp(\textbf{u}_{i}^\top \theta)}{\exp(\textbf{u}_{i}^\top \theta)}\right|\right\}.
\end{equation}
The LPRE objective function is defined as
\begin{equation}\label{LPRE}
LPRE_n(\theta)=\sum_{n=1}^n \left\{\left|\frac{t_i-\exp(\textbf{u}_i^\top \theta)}{t_i}\right| \times\left|\frac{t_i-\exp(\textbf{u}_{i}^\top \theta)}{\exp(\textbf{u}_{i}^\top \theta)}\right|\right\}.
\end{equation}
Now consider the realistic case with right censoring. For subject $i(=1,\ldots, n)$, let $c_i$ be the censoring variable which is independent of $\textbf{x}_i, \textbf{z}_i$, and $t_i$. We observe $y_i = \min(t_i,c_i)$ and $\delta_i = 1(t_i \le c_i)$. Without loss of generality, assume that the data $(y_i,\delta_i,\textbf{u}_i)$ have been sorted according to $y_i$ from the smallest to the largest.
\subsection{Penalized estimation and selection}
Consider the general relative error (GRE) criterion
\begin{equation}\label{GRE}
GRE_n(\theta)=\sum_{n=1}^n g\left\{\left|\frac{t_i-\exp(\textbf{u}_i^\top \theta)}{t_i}\right|, \left|\frac{t_i-\exp(\textbf{u}_{i}^\top \theta)}{\exp(\textbf{u}_{i}^\top \theta)}\right|\right\},
\end{equation}
where $g(a,b)$ is a bivariate function satisfying certain regularity conditions. When $g(a,b)=a+b$, the GRE criterion becomes the LARE \cite{chen2010least}; when $g(a,b)=ab$, it becomes the LPRE \cite{chen2013least}.
To accommodate right censoring in estimation, we adopt a weighted approach. Specifically, we first compute the Kaplan-Meier weights $\{w_i\}_{i=1}^n$ as
\begin{equation}
w_1=\frac{\delta_1}{n},w_i=\frac{\delta_i}{n-i+1}\prod_{j=1}^{i-1}
\Big(\frac{n-j}{n-j+1}\Big)^{\delta_j},i=2,\cdots,n.
\end{equation}
We propose the weighted objective function
\begin{equation}
Q_n(\theta)=\sum_{i=1}^n w_i g\left\{\left|\frac{y_i-\exp(\textbf{u}_i^\top \theta)}{y_i}\right|, \left|\frac{y_i-\exp(\textbf{u}_i^\top \theta)}{\exp(\textbf{u}_i^\top \theta)}\right| \right\}.
\end{equation}
In genetic interaction analysis, the dimension of unknown parameters can be much larger than the sample size. For regularized estimation and identification of important effects, we adopt penalization, where the objective function is
\begin{equation}
L_{n,\lambda}(\theta) = Q_n(\theta)+\varphi_\lambda(\theta).
\end{equation}
Here $\varphi_\lambda(\theta)$ is the penalty function. Adopting penalization for genetic interaction analysis has been pursued in recent literature. See for example \cite{bien2013lasso, liu2013identification, shi2014penalized}.
Multiple penalties are potentially applicable. Here we adopt the MCP \cite{zhang2010nearly}, which has been the choice of many high-dimensional studies including genetic interaction analysis. The penalty is defined as $\varphi_\lambda(t)=\lambda\int^{|t|}_0(1-x/(\gamma\lambda))_+dx$. $\gamma>0$ is the regularization parameter, and $\lambda$ is the tuning parameter.
It is noted that applying the MCP may lead to results not respecting the ``main effects, interactions" hierarchy, which has been stressed in some recent studies \cite{bien2013lasso}. The hierarchy postulates that the main effects corresponding to the identified interactions should be automatically identified. This can be achieved by replacing the MCP with for example sparse group penalties. However, we note that the computational cost of such penalties can be much higher. In addition, some published studies have demonstrated pure interactions without the presence of main effects \cite{caspi2006gene,zimmermann2011interaction}. In data analysis, when it is necessary to reinforce the hierarchy, we can refit and add back the main effects corresponding to the identified interactions (if these main effects are not identified in the first place).
\subsection{Computation}
For optimizing the penalized objective function, we propose combining the majorize-minimization (MM) algorithm \cite{hunter2005variable} with the coordinate descent (CD) algorithm \cite{wu2008coordinate}. The MM is used to approximate the objective function using its quadratic majorizer, while the CD is used for iteratively updating the estimate.
Specifically, when $g(a,b)=ab$, it is easy to compute the gradient and hessian matrix for $Q_n(\theta)$, and so approximation may not be needed. However when $g(a,b)=a+b$, computing the hessian matrix becomes difficult. With the estimate $\theta^{(s)}$ at the beginning of the $s+1$th iteration, we approximate $Q_n(\theta)$ by
\begin{eqnarray*}
Q_n(\theta; \theta^{(s)}) &=& \frac{1}{2}\sum_{i=1}^n w_i\left\{\frac{(1-y_i^{-1}\exp(\textbf{u}_i^\top \theta))^2}{|1-y_i^{-1}\exp(\textbf{u}_i^\top \theta^{(s)})|}+|1-y_i^{-1}\exp(\textbf{u}_i^\top \theta^{(s)})|\right.\\
&& \left.+ \frac{({1-y_i\exp(-\textbf{u}_i^\top \theta)})^2}{|{1-y_i\exp(-\textbf{u}_i^\top \theta^{(s)})}|} + |{1-y_i\exp(-\textbf{u}_i^\top \theta^{(s)})}| \right\}.
\end{eqnarray*}
It can be shown that $Q_n(\theta; \theta^{(s)})\geq Q_n(\theta)$, and the equality holds if and only if $\theta^{(s)}= \theta$. For the MCP, we use a quadratic approximation
\[
\varphi_{\lambda}(\theta; \theta^{(s)}) = \varphi_{\lambda}(\theta^{(s)})+ \frac{1}{2 |\theta^{(s)}|}\varphi'_{\lambda}(\theta^{(s)})(\theta^2-\theta^{(s)2}).
\]
By ignoring terms not related to $\theta$ in $Q_n(\theta; \theta^{(s)})+ \varphi_{\lambda}(\theta; \theta^{(s)})$, we have a smooth loss function $L_{n, \lambda}(\theta; \theta^{(s)})$, which is
\begin{equation}\label{app}
\sum_{i=1}^n w_i\left\{\frac{(1-y_i^{-1}\exp(\textbf{u}_i^\top \theta))^2}{|1-y_i^{-1}\exp(\textbf{u}_i^\top \theta^{(s)})|}+ \frac{({1-y_i\exp(-\textbf{u}_i^\top \theta)})^2}{|{1-y_i\exp(-\textbf{u}_i^\top \theta^{(s)})}|} \right\} + \frac{1}{ |\theta^{(s)}|}\varphi'_{\lambda}(\theta^{(s)})\theta^2.
\end{equation}
To solve the minimization problem $\theta^{(s+1)}= \arg \min_{\theta}L_{n, \lambda}(\theta; \theta^{(s)})$, we employ the coordinate descent algorithm. In summary, the algorithm proceeds as follows:
\medskip\noindent
Step 1. Initialize $s=0$. Compute $\theta^{(0)}$ as the Lasso estimate (which can be viewed as an extreme case of the MCP estimate).
\noindent
Step 2. Apply the CD algorithm to minimize the loss function $L_{n, \lambda}(\theta; \theta^{(s)})$ in (\ref{app}). Denote the estimate as $\theta^{(s+1)}$. Specially, the CD algorithm updates one coordinate at a time and treats the other coordinates as fixed. Define $u_{ij}$ as the $j$th component of $\textbf{u}_i$. For $j \in \{1, \cdots, p+q+pq\}$, defined $\vartheta_{i,-j}=\sum_{t<j}u_{it} \theta_{t}^{(s+1)} + \sum_{t>j}u_{it} \theta_{t}^{(s)}$, then
\begin{eqnarray*}
\theta_j^{(s+1)} &=& \arg\min_{\theta_j} \left\{\sum_{i=1}^n w_i \left[\frac{(1-y_i^{-1}\exp(\vartheta_{i,-j} +u_{ij}\theta_j ))^2}{|1-y_i^{-1}\exp(\textbf{u}_i^\top \theta^{(s)})|} \right.\right.\\
&& \left.\left.~~~~~~+ \frac{({1-y_i\exp(-\vartheta_{i,-j}-u_{ij}\theta_j )})^2}{|{1-y_i\exp(-\textbf{u}_i^\top \theta^{(s)})}|} \right] + \frac{1}{ |\theta_j^{(s)}|}\varphi'_{\lambda}(\theta_j^{(s)})\theta_j^2\right\} ~.
\end{eqnarray*}
\noindent
Step 3. Repeat Step 2 until convergence. We use the $L_2$-norm of the difference between two consecutive estimates less than $10^{-6}$ as the convergence criterion.
The proposed method involves tunings. For $\gamma$, published studies \cite{zhang2010nearly} suggest selecting from a small number of values or fixing it. In our simulation, we find that the estimation results are not sensitive to the value of $\gamma$. We follow published studies and set $\gamma=6$. The selection of $\lambda$ will be described in the following sections.
\section{Simulation}
Beyond evaluating performance of the proposed approach, we also use simulation to compare with the penalized weighted least squares (simply denoted as LS) and penalized weighted least absolute deviation (denoted as LAD) methods, which respectively have objective functions
\[\sum_{i=1}^n w_i(\log(y_i)-\textbf{u}_i^\top\theta)^2+\varphi_\lambda(\theta)
~~\mbox{and}
~~
\sum_{i=1}^n w_i|\log(y_i)-\textbf{u}_i^\top\theta|+\varphi_\lambda(\theta),\]
where $\{w_i\}_{i=1}^n$ and $\varphi_\lambda(\theta)$ are the same as defined before. \\
\noindent{\bf Simulation I.} In model
$t_i=\exp(\textbf{x}_i^\top\alpha+\textbf{z}_i^\top\beta+(\textbf{x}_i\otimes \textbf{z}_i)^\top\xi)\varepsilon_i,~ i=1,\cdots,n,$ $\textbf{z}_i$'s have a multivariate normal distribution with marginal means 0 and marginal variances 1. Denote the correlation coefficient between genes $j$ and $k$ as $\rho_{jk}$. Consider the following correlation structures: (i) independent, where $\rho_{jk}=0$ if $j\ne k$, (ii) AR(0.2), where $\rho_{jk}=0.2^{|j-k|}$; (iii) AR(0.8), where $\rho_{jk}=0.8^{|j-k|}$; (iv) Band1, where $\rho_{jk}=0.3$ if $|j-k|=1$ and $\rho_{jk}=0$ otherwise; and (v) Band2, where $\rho_{jk}=0.6$ if $|j-k|=1$, $\rho_{jk}=0.3$ if $|j-k|=2$, and $\rho_{jk}=0$ otherwise. We generate $\textbf{x}_i$'s from the standard multivariate normal distribution. We set $n=200$, $q=5$, and $p=500$. The dimension of genetic effects and interactions is much larger than the sample size. There are a total of 35 nonzero effects: 5 main effects of the E factors, 10 main effects of the G factors, and 20 interactions. The nonzero coefficients are randomly generated from $Uniform(0.4,1.2)$. We consider two error distributions: (i) $\log(\varepsilon)$ follows $N(0,1)$, and (ii) $\log(\varepsilon)$ follows $Unif(-2,2)$. The event times are computed from the AFT model. The censoring times are generated from a uniform distribution, with a censoring rate about 20\%.
\medskip\noindent{\bf Simulation II.} Data are first generated in the same manner as under Simulation I. To mimic discrete genetic data (for example SNPs), we dichotomize the simulated genetic data at -1 and 0.5 to create three levels.
\begin{table}[h]
\caption{Summary of Simulation II. In each cell, mean (sd) based on 200 replicates}
\vskip .1in
\label{table3}
\centering
\setlength{\tabcolsep}{3pt}
\begin{tabular}{c c c c c c}\hline
&&AUC&SE&TPR&FPR\\ \hline
\multicolumn{6}{c}{$\log(\varepsilon)\sim N(0,1)$} \\
\hline
independent&LARE&0.846(0.031)&19.53(3.321)&0.601(0.063)&0.098(0.013)\\
&LPRE&0.837(0.032)&19.47(3.130)&0.572(0.171)&0.095(0.134)\\
&LAD&0.833(0.029)&20.26(3.118)&0.564(0.117)&0.084(0.026)\\
&LS&0.854(0.020)&20.78(2.641)&0.562(0.109)&0.076(0.013)\\\hline
AR(0.2)&LARE&0.868(0.034)&17.55(3.252)&0.739(0.082)&0.103(0.018)\\
&LPRE&0.863(0.024)&16.68(3.671)&0.649(0.153)&0.062(0.027)\\
&LAD&0.847(0.027)&19.57(2.947)&0.564(0.100)&0.078(0.026)\\
&LS&0.860(0.024)&18.66(2.583)&0.628(0.086)&0.071(0.011)\\\hline
AR(0.8)&LARE&0.928(0.029)&7.655(2.611)&0.891(0.053)&0.062(0.027)\\
&LPRE&0.898(0.032)&7.755(2.990)&0.871(0.076)&0.066(0.021)\\
&LAD&0.911(0.022)&13.68(2.973)&0.758(0.098)&0.069(0.023)\\
&LS&0.901(0.026)&12.74(2.417)&0.779(0.104)&0.063(0.019)\\\hline
Band1&LARE&0.868(0.033)&18.51(3.316)&0.673(0.080)&0.078(0.022)\\
&LPRE&0.859(0.026)&17.78(3.560)&0.641(0.143)&0.059(0.023)\\
&LAD&0.850(0.031)&19.27(3.676)&0.629(0.119)&0.085(0.025)\\
&LS&0.864(0.022)&18.92(2.853)&0.616(0.074)&0.078(0.012)\\\hline
Band2&LARE&0.904(0.028)&10.82(2.571)&0.828(0.158)&0.060(0.017)\\
&LPRE&0.875(0.031)&11.39(2.922)&0.787(0.102)&0.055(0.021)\\
&LAD&0.872(0.033)&17.68(3.673)&0.685(0.108)&0.075(0.027)\\
&LS&0.880(0.025)&16.92(3.114)&0.725(0.081)&0.075(0.014)\\
\hline
\multicolumn{6}{c}{$\log(\varepsilon)\sim Unif(-2,2)$} \\
\hline
independent&LARE&0.840(0.032)&19.38(3.024)&0.634(0.073)&0.111(0.024)\\
&LPRE&0.845(0.022)&20.46(2.898)&0.582(0.169)&0.094(0.035)\\
&LAD&0.831(0.033)&21.29(3.453)&0.569(0.123)&0.081(0.027)\\
&LS&0.847(0.021)&21.03(3.258)&0.557(0.087)&0.080(0.018)\\\hline
AR(0.2)&LARE&0.832(0.029)&18.63(3.286)&0.696(0.076)&0.093(0.019)\\
&LPRE&0.850(0.022)&18.15(4.075)&0.616(0.082)&0.083(0.012)\\
&LAD&0.835(0.028)&19.49(2.958)&0.583(0.127)&0.082(0.028)\\
&LS&0.858(0.021)&20.52(3.063)&0.587(0.111)&0.076(0.018)\\\hline
AR(0.8)&LARE&0.913(0.031)&9.610(2.219)&0.833(0.128)&0.068(0.023)\\
&LPRE&0.889(0.025)&8.732(2.770)&0.857(0.105)&0.052(0.016)\\
&LAD&0.900(0.030)&15.85(2.980)&0.736(0.124)&0.072(0.026)\\
&LS&0.895(0.026)&14.60(2.970)&0.732(0.108)&0.007(0.029)\\\hline
Band1&LARE&0.850(0.028)&14.23(3.010)&0.714(0.082)&0.097(0.020)\\
&LPRE&0.856(0.023)&15.64(3.274)&0.624(0.120)&0.083(0.016)\\
&LAD&0.844(0.030)&20.94(3.371)&0.543(0.114)&0.077(0.024)\\
&LS&0.856(0.023)&19.70(2.899)&0.626(0.090)&0.076(0.011)\\\hline
Band2&LARE&0.868(0.032)&13.06(3.513)&0.782(0.163)&0.098(0.025)\\
&LPRE&0.864(0.030)&12.23(3.713)&0.763(0.148)&0.057(0.024)\\
&LAD&0.870(0.029)&17.23(3.555)&0.680(0.128)&0.073(0.027)\\
&LS&0.869(0.033)&16.46(2.470)&0.704(0.093)&0.071(0.010)\\
\hline
\end{tabular}
\end{table}
We evaluate the simulation results in two ways. First, we consider a sequence of $\lambda$ values, evaluate identification performance at each value, and then compute the overall AUC (area under the ROC -- receiver operating characteristic -- curve). In addition, we also select the optimal $\lambda$ using a cross validation approach and then compute the estimation squared error (SE), true positive rate (TPR), and false positive rate (TPR) at the optimal tuning. The summary based on 200 replicates is provided in Table 1 and 3 (Appendix), respectively. Simulation suggests that, when evaluated using AUC, the four methods have similar performance. Under Simulation I, the performance is also similar in terms of SE, TPR, and FPR. However, under Simulation II, the proposed LARE and LPRE can have better performance. In addition, it is also observed that LARE may outperform LPRE, at the cost of slightly higher computer time. Overall simulation suggests that the proposed approach, especially LARE, performs comparable to or better than the alternatives. Thus it provides a ``safe" choice for practical data analysis.
\section{Analysis of lung cancer prognosis data}
Lung cancer is the leading cause of cancer death worldwide. Genetic profiling studies have been extensively conducting, searching for genetic risk factors associated with lung cancer prognosis. Here we analyze the TCGA (The Cancer Genome Atlas) data on the prognosis of lung adenocarcinoma. The TCGA data were recently collected and published by NCI and have a high quality. The prognosis outcome of interest is overall survival. The dataset contains records on 468 patients, among whom 117 died during follow-up. The median follow-up time is 8 months.
\begin{table}[h]
\caption{Analysis of lung cancer data with LARE: main genetic effects and G$\times$E interactions. For the interactions, values in ``()'' are the stability results.}
\vskip .1in
\label{table5}
\centering
\setlength{\tabcolsep}{3pt}
\begin{tabular}{c c c c c c }\hline
&&\multicolumn{4}{c}{Interactions}\\
\cline{3-6}
Gene &Main effects&Age&Gender&Smoking pack year&Smoking history\\\hline
ADORA2B& -0.231 &&&& -0.260(0.76)\\
AKIRIN2& -0.281 \\
ASB12& -0.241 \\
C5ORF45& -0.042 \\
C14ORF93& -0.472\\
C16ORF93& -0.160& -0.293(0.91) \\
CAND1& 0.309 &-2.181(0.95) \\
CBWD2& 0.234 \\
CDR2& 0.210 \\
CIAPIN1& 0.187&& & -0.179(0.85) \\
DCP1B& 0.448 \\
DYRK2& -1.41& 0.758(0.66) \\
EIF4EBP1& 0.081& -0.001(0.81)\\
EMB& 0.224 \\
FDXR& 0.293 &&& -0.477(0.99) \\
GALK2& -0.158 &&& -0.240(0.75) \\
GOLGA7& -0.146 &-0.096(0.45) \\
HERPUD2& 0.121 \\
HOXC13& -0.248& -0.145(0.98) \\
ING1& -2.117 &&&& 2.154(0.97)\\
INO80B& -0.164 &&& -1.607(0.95) \\
KIF21B& -0.391 &-0.446(0.99) \\
KLHDC1& -0.011 &&& 0.382(0.98)\\
LIG4& -0.584 && 0.299(0.80) \\
LINC00471& 0.236 &&&& 0.114(0.94)\\
LINC00476& 0.258 &&&& 0.056(0.55)\\
LRRC45& -0.136 &-0.083(0.93) \\
MCAT& 0.103 &&& 0.180(0.96) \\
MVD& -0.348 \\
NCALD& 0.376 && -0.605(0.70) \\
OTUD1& 0.189 &&& 0.038(0.34) \\
PEX19& -0.444 &&&& 0.045(0.55)\\
PHLPP1& -0.439 \\
PNPLA2& -0.193 &0.014(0.55) \\
PPM1A& -0.124 &&& 0.166(0.89) \\
PPP2R2D& 0.157& -0.234(0.67) \\
RBM11& 0.032 && -0.291(0.71) \\
RNF6& -0.215 &0.199(0.90) \\
RNF126P1& 0.225 \\
RPS27& 0.134 &&&& -0.155(0.22)\\
SCAND2P& -0.002 &&& 0.329(0.35) \\
SERTAD4& -0.356 && 0.350(0.91) \\
SGSM3& 0.285 &&& -0.039(0.46)\\
SH3RF1& -0.096 \\
SLC25A2& -0.009 && -0.335(0.94) \\
SPCS3& -0.310 &&&& 0.340(0.66)\\
SPRED2& -0.260 \\
SRRM3& -0.317 &&&& -0.244(0.70)\\
TXN2& -0.339 &&& 0.012(0.46) \\
UBE4B& 0.418 &-0.497(0.53) \\
VPS13B& 0.065 &&& -0.108(0.99) \\
ZNF727& 0.401& -0.254(0.78) \\
\hline
\end{tabular}
\end{table}
Four E factors are included in analysis: age, gender, smoking pack years, and smoking history. All four have been suggested as associated with lung cancer prognosis in the literature. Among them, age and smoking pack years are continuous and normalized prior to analysis. Gender and smoking history are binary. A total of 436 subjects have complete E measurements. Among them, 110 died during follow-up, and the median follow-up time is 23 months. For the 326 censored subjects, the median follow-up time is 6 months.
Measurements on 18,897 gene expressions are available. To improve stability and reduce computational cost, we conduct marginal prescreening based on genes' univariate regression significance (p-value less than or equal to 0.1) and interquartile range (above the median of all interquartile ranges). Similar procedures have been adopted in the literature. A total of 819 gene expressions are included in downstream analysis. For each gene expression, we normalize to have mean 0 and variance 1.
We apply the proposed approach and select the optimal $\lambda$ using five-fold cross validation. The detailed identification and estimation results are presented in Tables 2 (LARE) and 5 (LPRE, Appendix). As previously described, it is possible that the main effects corresponding to the identified interactions are not identified. To respect the ``main effects, interactions" hierarchy, we add back such main effects and re-fit. Beyond the proposed, we also apply the LS and LAD methods. The summary of applying different methods is provided in Table \ref{table9} (Appendix). Detailed estimation and identification results using the alternatives are presented in Tables \ref{table7} and \ref{table8} (Appendix). Different methods identify different sets of main effects and interactions. It is interesting that all of the main effects and interactions identified by LPRE are identified by LARE. They may represent more reliable findings. The LAD method identifies much fewer effects.
To complement the identification and estimation analysis, we evaluate stability. Specifically, we randomly remove ten subjects and then analyze data. This procedure is repeated 200 times. We then compute the probability that an interaction term is identified. Such an evaluation has been conducted in the literature. The stability results are provided in Tables 2 and 5-7(Appendix). We can see that most of the identified interactions are relatively stable, with many having probabilities of being identified close to one.
\section{Discussion}
The identification of important G$\times$E interactions remains a challenging problem. In this article, we have introduced using the relative error criteria as loss functions. A penalized approach has been adopted for estimation and selection. Simulation shows that the proposed approach has performance comparable to or better than the alternatives. Thus it may be provide a useful alternative for data analysis. A limitation of this study is that the asymptotic properties have not been established. In the analysis of a lung cancer dataset, the LARE and LPRE results are relatively consistent but different from the alternatives. The identified interactions are reasonably stable. More examination of the findings is needed in the future.
\begin{acknowledgement}
We thank the participants of Joint 24th ICSA Applied Statistics Symposium and 13th Graybill Conference in Colorado and organizers of this proceedings. This study was supported in part by the National Science Foundation of China (Grant No. 11401561), National Social Science Foundation of China (13CTJ001, 13\&ZD148), NIH (CA016359, CA191383), and the U.S. VA Cooperative Studies Program of the Department of Veterans Affairs, Office of Research and Development.
\end{acknowledgement}
\bibliographystyle{spmpsci}
|
\section{Introduction}
\input{sec_introduction}
\subsection{Related Work}
\label{sec:related_work}
\input{sec_related_work}
\section{\textnormal{\textrm{SpanCCA}}\xspace:~An~Algorithm~for~Sparse Diagonal CCA}
\input{sec_algorithm}
\section{Beyond Sparsity: Structured CCA}
\input{sec_beyond_sparsity}
\section{Experiments}
\label{sec:experiments}
\input{sec_experiments}
\section{Discussion}
We presented a novel combinatorial algorithm for the sparse diagonal CCA problem and other constrained variants, with provable data-dependent global approximation guarantees and several attractive properties:
the algorithm is simple, embarrassingly parallelizable, with complexity that scales linearly in the dimension of the input data,
while it administers precise control on the sparsity of the extracted canonical vectors.
Further it can accommodate additional structural constraints by plugging in a suitable ``projection'' subroutine.
Several directions remain open.
We addressed the question of computing a single pair of sparse canonical vectors.
Numerically, multiple pairs can be computed successively employing an appropriate deflation step.
However, determining the kind of deflation most suitable for the application at hand,
as well as the sparsity level of each component can be a challenging task, leaving a lot of room for research.
\begin{small}
\subsection{Intuition}
The hardness of the sparse CCA problem \eqref{cca:def} lies in the detection of the optimal supports for the canonical vectors.
In the \emph{unconstrained} problem,
where only a unit $\ell_{2}$-norm constraint is imposed on $\mathbf{u}$ and $\mathbf{v}$, the optimal CCA pair coincides with the top singular vectors of the input argument~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
In the sparse variant,
if the optimal supports for $\mathbf{u}$ and $\mathbf{v}$ were known, computing the optimal solution would be straightforward:
the nonzero subvectors of $\mathbf{u}$ and $\mathbf{v}$ would coincide with the leading singular vectors of the $s_{x} \times s_{y}$ submatrix of $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, indexed by the two support sets.
Hence, the bottleneck lies in determining the optimal supports for $\mathbf{u}$ and $\mathbf{v}$.
\textbf{Exhaustive search}\,
A straightforward, brute-force approach is to exhaustively consider all possible supports for $\mathbf{u}$ and $\mathbf{v}$;
for each candidate pair solve the unconstrained CCA problem on the restricted input, and determine the supports for which the objective~\eqref{cca:def} is maximized.
Albeit optimal, this procedure is intractable as the number of candidate supports
$\binom{m}{s_{x}}\binom{n}{s_{y}}$ is overwhelming even for small values of $s_{x}$ and $s_{y}$.
\textbf{Thresholding}\,
On the other hand, a feasible pair of sparse canonical vectors $\mathbf{u}$, $\mathbf{v}$ can be extracted by {hard-thresholding} the solution to the unconstrained problem,
\textit{i.e.},
computing the leading singular vectors $\mathbf{u}$, $\mathbf{v}$ of~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, suppressing to zero all but the $s_{x}$ and $s_{y}$ largest in magnitude entries, respectively, and rescaling to obtain a unit $\ell_{2}$-norm solution.
Essentially, this heuristic resorts to unconstrained CCA for a \emph{guided} selection of the sparse support.
\textbf{Proposed method}\,
Our sparse CCA algorithm covers the ground between these two approaches.
Instead of relying on the solution to the unconstrained problem for the choice of the sparse supports,
it explores a principal subspace of the input matrix $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, spanned by its leading~${r \ge 1}$ singular vector pairs.
For~${r=1}$, its output coincides with that of the thresholding approach,
while for~${r = \min\lbrace{m, n\rbrace}}$ it approximates that of exhaustive search.
Effectively, we solve \eqref{cca:def} on a rank-$r$ approximation of the input~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
The key observation is that the low inner dimension of the argument matrix can be exploited to substantially reduce the search space:
our algorithm identifies an (approximately) optimal pair of supports for the low rank sparse CCA problem, without considering the entire collection of possible supports of cardinalities~$s_{x}$ and~$s_{y}$.
\subsection{Overview and Guarantees}
\label{sec:overview_guarantees}
\textnormal{\textrm{SpanCCA}}\xspace is outlined in Algorithm~\ref{algo:cca}.
The first step is to compute a rank-$r$ approximation $\mathbf{B}$ of the input $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$,
where $r$ is an accuracy input parameter,
via the truncated singular value decomposition (SVD) of $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.\footnote{
The low-rank approximation can be computed
using faster randomized approaches; see~\cite{halko2011finding}.
Here, for simplicity, we consider the exact case.}
From that point on, the algorithm operates exclusively on~$\mathbf{B}$ effectively solving a low-rank sparse diagonal CCA problem:
\begin{align}
\max_{
\mathbf{u} \in \mathcal{U},
\mathbf{v} \in \mathcal{V}} \quad
&
\mathbf{u}^\transpose \mathbf{B} \mathbf{v},
\label{cca:on-low-rank}
\end{align}
where~${\mathcal{U} \subseteq \mathbb{S}_{2}^{m-1}}$ and~${\mathcal{V} \subseteq \mathbb{S}_{2}^{n-1}}$ are defined in~\eqref{cca:sparse_constraints}.
As we discuss in the next section,
we can consider other constrained variants of CCA on potentially arbitrary, non-convex sets.
We do require, however, that there exist procedures to (at least approximately) solve the maximizations
\begin{align}
& \mathsf{P}_{\mathcal{U}}(\mathbf{a})
{\triangleq}
\argmax_{\mathbf{u} \in \mathcal{U}}\;
\mathbf{a}^{\transpose}\mathbf{u},
\label{lowrank:solve-for-x}
\\
& \mathsf{P}_{\mathcal{V}}(\mathbf{b})
{\triangleq}
\argmax_{\mathbf{v} \in \mathcal{V}}\;
\mathbf{b}^{\transpose}\mathbf{v},
\label{lowrank:solve-for-y}
\end{align}
for any given vectors ${\mathbf{a} \in \mathbb{R}^{m \times 1}}$ and ${\mathbf{b} \in \mathbb{R}^{n \times 1}}$.
Fortunately, this is the case for the sets of sparse unit $\ell_{2}$-norm vectors.
Algorithm~\ref{algo:sparse} outlines an efficient $O(m)$ procedure
that given $\mathbf{a} \in \mathbb{R}^{m\times 1}$ computes an exact solution to~\eqref{lowrank:solve-for-x} with at most $s \le m$ nonzero entries:
first it determines the~$s$ largest (in magnitude) entries of $\mathbf{a}$ (breaking ties arbitrarily), it zeroes out the remaining entries and re-scales the output to meet the $\ell_{2}$-norm requirement.
The main body of Algorithm~\ref{algo:cca} consists of a single iteration.
In the $i$th round, it independently samples a point, or equivalently direction, $\mathbf{c}_{i}$ from the $r$-dimensional unit $\ell_{2}$ sphere
and uses it to appropriately sample a point $\mathbf{a}_{i}$ in the range of~$\mathbf{B}$.
The latter is then used to compute a feasible solution pair $\mathbf{u}_{i}$, $\mathbf{v}_{i}$ via a two-step procedure:
first the algorithm computes $\mathbf{u}_{i}$ by ``projecting'' $\mathbf{a}_{i}$ onto $\mathcal{U}$ invoking Alg.~\ref{algo:sparse} as a subroutine to solve maximization~\eqref{lowrank:solve-for-x},
and then computes $\mathbf{v}_{i}$ by projecting $\mathbf{b}_{i}=\mathbf{B}^{\transpose}\mathbf{u}_{i}$ onto $\mathcal{V}$ in a similar fashion.
The algorithm repeats this procedure for $\mathrm{T}$ rounds and outputs the pair that achieves the maximum objective value in~\eqref{cca:on-low-rank}.
{We emphasize that consecutive rounds are completely independent and can be executed in parallel.}
\begin{algorithm}[t!]
\caption{\textnormal{\textrm{SpanCCA}}\xspace}
\label{algo:cca}
{
\begin{algorithmic}[1]
\INPUT:
${\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}}$, a real $m \times n$ matrix.\\
\quad\,\,\,\,\,$r \in \mathbb{N}_{+}$, the rank of the approximation to be used.\\
\quad\,\,\,\,\,$\mathrm{T} \in \mathbb{N}_{+}$, the number of samples/iterations.
\OUTPUT $\mathbf{u}_{\sharp} \in \mathcal{U}$, $\mathbf{v}_{\sharp} \in \mathcal{V}$
\STATE
${
\mathbf{U}, \mathbf{\Sigma}, \mathbf{V}
\leftarrow
\texttt{SVD}({\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}}, r)
}$
\hfill \COMMENT{
${
\mathbf{B} \gets \mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\transpose}
}$
}
\FOR{$i=1, \hdots, T$}
\STATE $\mathbf{c}_{i} \gets \texttt{randn(r)}$
\hfill \COMMENT{$\sim{}\mathcal{N}(\mathbf{0}, \mathbf{I}_{r \times r})$}
\STATE $\mathbf{c}_{i} \gets \mathbf{c}_{i} / \|\mathbf{c}_{i}\|_{2}$
\STATE $\mathbf{a}_{i} \gets {\mathbf{U}}{\mathbf{\Sigma}}\mathbf{c}_{i}$
\hfill \COMMENT{$\mathbf{a}_{i} \in \mathbb{R}^{m}$}
\STATE $\mathbf{u}_{i} \gets \argmax_{\mathbf{u} \in \mathcal{U}} \mathbf{a}_{i}^{\transpose}\mathbf{u}$
\hfill \COMMENT{$\mathsf{P}_{\mathcal{U}}(\cdot)$}
\STATE $\mathbf{b}_{i} \gets \mathbf{V}\mathbf{\Sigma}\mathbf{U}^{\transpose}\mathbf{u}_{i}$
\hfill \COMMENT{$\mathbf{b}_{i} \in \mathbb{R}^{n}$}
\STATE $\mathbf{v}_{i} \gets \argmax_{\mathbf{v} \in \mathcal{V}} \mathbf{b}_{i}^{\transpose}\mathbf{v}$
\hfill \COMMENT{$\mathsf{P}_{\mathcal{V}}(\cdot)$}
\STATE $\text{obj}_{i} \gets \mathbf{b}_{i}^{\transpose}\mathbf{v}_{i}$
\ENDFOR
\STATE $
i_{0}
\gets
\arg\max_{\substack{ i \in [T] } }
\text{obj}_{i}$
\STATE
$(\mathbf{u}_{\sharp}, \mathbf{v}_{\sharp})
\gets
(\mathbf{u}_{i_{0}}, \mathbf{v}_{i_{0}} )$
\end{algorithmic}
}
\end{algorithm}
\begin{algorithm}[t!]
\caption{
\small{
$\mathsf{P}_{\mathcal{U}}(\cdot)$ for
$\mathcal{U}
{\triangleq}
\bigl\lbrace \mathbf{u} \in \mathbb{S}_{2}^{m-1}: \|\mathbf{u}\|_{0} \le s\bigr\rbrace
$}
}
\label{algo:sparse}
\begin{algorithmic}[1]
\INPUT:
$\mathbf{a} \in \mathbb{R}^{d \times 1}$.
\OUTPUT
$
\mathbf{u}_{0} = \argmax_{\mathbf{u} \in \mathcal{U}} \mathbf{a}^{\transpose}\mathbf{u}$
\STATE $\mathbf{u}_{0} \gets \mathbf{0}_{d \times 1}$
\STATE $t \gets $ index of $s$th order element of $\texttt{abs}(\mathbf{a})$
\STATE $\mathcal{I} \gets \lbrace i: \lvert a_{i} \rvert \ge \lvert a_{t} \rvert\rbrace$
\STATE $\mathbf{u}_{0}[i] \gets \mathbf{a}[i], \forall i \in \mathcal{I}$
\STATE $\mathbf{u}_{0} \gets \mathbf{u}_{0} / \|\mathbf{u}_{0}\|_{2}$
\end{algorithmic}
\end{algorithm}
For a sufficiently large number~$\mathrm{T}$ of rounds (or samples)
the procedure guarantees that the output pair will be approximately optimal in terms of the objective for the low-rank problem~\eqref{cca:on-low-rank}.
That, it turn, translates to approximation guarantees for the full-rank sparse CCA problem~\eqref{cca:def}:
\begin{theorem}
\label{thm:main-algo-guarantees}
For any real ${m \times n}$ matrix~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$,
$\epsilon \in (0, 1)$, and $r \le \max\lbrace m, n\rbrace$,
Algorithm~\ref{algo:cca} with input $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, $r$,
and $\mathrm{T}=\widetilde{O}\bigl(2^{r \cdot \log_{2}(\sfrac{2}{\epsilon})} \bigr)$
outputs $\mathbf{u}_{\sharp} \in \mathcal{U}$
and $\mathbf{v}_{\sharp} \in \mathcal{V}$ such that
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
\ge
\mathbf{u}_{\star}^{\transpose} \mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}} \mathbf{v}_{\star}
-
\epsilon \cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
-
2 \sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}),
\nonumber
\end{align}
in time
$
\mathrm{T}_{\mathsmaller{\mathsf{SVD}}}(r) + O\mathopen{}\bigl(\mathrm{T} \cdot \bigl(\mathrm{T}_{\mathcal{U}}+\mathrm{T}_{\mathcal{V}}+r \cdot \max\lbrace m, n\rbrace \bigr)\bigr)
$.
\end{theorem}
Here,
$\mathbf{u}_{\star}$ and $\mathbf{v}_{\star}$ denote the unknown optimal pair of canonical vectors satisfying the desired constraints.
$\mathrm{T}_{\mathsmaller{\mathsf{SVD}}}(r)$ denotes the time to compute the rank-$r$ truncated SVD of the input~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, while $\mathrm{T}_{\mathcal{U}}$ and $\mathrm{T}_{\mathcal{V}}$ denote the time required to compute the maximizations~\eqref{lowrank:solve-for-x} and~\eqref{lowrank:solve-for-y}, respectively,
which in the case of Alg.~\ref{algo:sparse} are linear in the dimensions $m$ and $n$.
The first term in the additive error is due to the sampling approach of Alg.~\ref{algo:cca}.
The second term is due to the fact that the algorithm operates on the rank-$r$ surrogate matrix~$\mathbf{B}$.
Theorem~\ref{thm:main-algo-guarantees}
establishes a trade-off between the computational complexity of Alg.~\ref{algo:cca} and the quality of the approximation guarantees: the latter improves as $r$ increases, but the former depends exponentially in~$r$.
Finally,
in the special case where we impose sparsity constraints on only one of the two variables, say $\mathbf{u}$, while allowing the second variable, here $\mathbf{v}$, to be any vector with unit $\ell_{2}$ norm,
we obtain stronger guarantees.
\begin{theorem}
\label{thm:main-algo-guarantees-special-case}
If $\mathcal{V}=\lbrace \mathbf{v}: \|\mathbf{v}\|_{2}=1\rbrace$,
\textit{i.e.}, if no constraint is imposed on variable $\mathbf{v}$ besides unit length,
then Algorithm~\ref{algo:cca} under the same configuration as that in Theorem~\ref{thm:main-algo-guarantees}
outputs $\mathbf{u}_{\sharp} \in \mathcal{U}$
and $\mathbf{v}_{\sharp} \in \mathcal{V}$ such that
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
\ge
(1-\epsilon) \cdot
\mathbf{u}_{\star}^{\transpose} \mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}} \mathbf{v}_{\star}
-
2 \cdot \sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\nonumber
\end{align}
\end{theorem}
Theorem~\ref{thm:main-algo-guarantees-special-case} implies that due to the flexibility in the choice of the canonical vector $\mathbf{v}$,
Alg.~\ref{algo:cca} solves the low-rank problem~\eqref{cca:on-low-rank} within a multiplicative $(1-\epsilon)$-factor from the optimal;
the extra additive error term is once again due to the fact that the algorithm operates on the rank-$r$ approximation $\mathbf{B}$ instead of the original input~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
In this case, the optimal choice of $\mathbf{v}$ in~\eqref{lowrank:solve-for-y} is just a scaled version of the argument $\mathbf{b}$.
Finally, we note that if constraints need to be applied on $\mathbf{v}$ instead of $\mathbf{u}$, then the same guarantees can be obtained by applying Algorithm~\ref{algo:cca} on $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}$.
A formal proof for Theorem~\ref{thm:main-algo-guarantees-special-case} is provided in the Appendix, Sec.~\ref{sec:approx-proof}.
Overall, \textnormal{\textrm{SpanCCA}}\xspace is simple to implement and is trivially parallelizable: the main iteration can be split across and arbitrary number of processing units achieving a potentially linear speedup.
It is the first algorithm for sparse diagonal CCA with data-dependent global approximation guarantees.
As discussed in Theorem~\ref{thm:main-algo-guarantees}, the input accuracy parameter $r$ establishes a trade-off between the running time and the tightness of the theoretical guarantees.
Its complexity scales linearly in the dimensions of the input for any constant $r$,
but admittedly becomes prohibitive even for moderate values of $r$.
In practice, if the spectrum of the data exhibits sharp decay, we may be able to obtain useful approximation guarantees even for small values of $r$ such as $2$ or $3$.
Moreover, disregarding the theoretical guarantees, the algorithm can always be executed for any rank $r$ and an arbitrary number of iterations $\mathrm{T}$.
In Section \ref{sec:experiments}, we empirically show that moderate values for $r$ and $\mathrm{T}$ can potentially achieve better solutions compared to state of the art.
We note, however, that the tuning of those parameters needs to be investigated.
\subsection{Analysis}
Let $\mathbf{B} = \mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\transpose}$,
and $\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}, \mathbf{v}_{\mathsmaller{\mathtt{(B)}}}$ be a pair that maximizes --not necessarily uniquely-- the objective $\mathbf{u}^{\transpose}\mathbf{B}\mathbf{v}$ in~\eqref{cca:on-low-rank} over all feasible solutions.
We assume that $\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}\mathbf{B}\mathbf{v}_{\mathsmaller{\mathtt{(B)}}} > 0$.\footnote{Observe that this is always true for any nonzero argument $\mathbf{B}$ as long as at least one of the two variables $\mathbf{u}$ and $\mathbf{v}$ can take arbitrary signs.
It is hence true under vanilla sparsity constraints.}
Define the $r \times 1$ vector $\mathbf{c}_{\mathsmaller{\mathtt{(B)}}}{\triangleq}\mathbf{V}^{\transpose}\mathbf{v}_{\mathsmaller{\mathtt{(B)}}}$ and let $\rho$ denote its $\ell_{2}$ norm.
Then, $0<\rho\le 1$;
the upper bound follows from the fact that the $r$ columns of $\mathbf{V}$ are orthonormal and ${\|\mathbf{v}_{\mathsmaller{\mathtt{(B)}}}\|_{2} = 1}$, while the lower follows by the aforementioned assumption.
Finally, define $\overline{\mathbf{c}}_{\mathsmaller{\mathtt{(B)}}} = \mathbf{c}_{\mathsmaller{\mathtt{(B)}}} / \rho$, the projection of $\mathbf{c}_{\mathsmaller{\mathtt{(B)}}}$ on the unit $\ell_{2}$-sphere~$\mathbb{S}_{2}^{r-1}$.
\begin{definition}
For any $\epsilon \in (0,1)$, an $\epsilon$-net of $\mathbb{S}_{2}^{r-1}$ is a finite collection $N$ of points in $\mathbb{R}^{}$ such that for any $\mathbf{c} \in \mathbb{S}_{2}^{r-1}$, $N$ contains a point $\mathbf{c}\prime$ such that $\|\mathbf{c}\prime-\mathbf{c}\| \le \epsilon$.
\end{definition}
\begin{lemma}[\cite{vershynin2010introduction}, Lemma 5.2]
\label{lem:vershynin12}
For any $\epsilon \in (0, 1)$,
there exists an $\epsilon$-net of $\mathbb{S}_{2}^{r-1}$ equipped with the Euclidean metric, with at most
$\left(1+\sfrac{2}{\epsilon}\right)^r$ points.
\end{lemma}
Algorithm~\ref{algo:cca} runs in an iteration with $\mathrm{T}$ rounds.
In each round, it independently samples a point $\mathbf{c}_{i}$ from~$\mathbb{S}_{2}^{r-1}$,
by randomly generating a vector according to a spherical Gaussian distribution and appropriately scaling its length.
Based on Lemma~\ref{lem:vershynin12} and elementary counting arguments, for sufficiently large $\mathrm{T}$ the collection of sampled points forms a $\epsilon$-net of $\mathbb{S}_{2}^{r-1}$ with high probability:
\begin{lemma}
\label{lem:num-of-random-points}
For any $\epsilon, \delta \in (0, 1)$,
a set of $\mathrm{T}={O\mathopen{}\bigl( r (\sfrac{\epsilon}{4})^{-r} \cdot \ln{\sfrac{4}{{\epsilon}\cdot{\delta}}} \bigr)}$
randomly and independently drawn points uniformly distributed on~$\mathbb{S}_{2}^{r-1}$
suffices to construct an $\sfrac{\epsilon}{2}$-net of~$\mathbb{S}_{2}^{r-1}$ with probability at least $1 - \delta$.
\end{lemma}
It follows that there exists $i_{\star} \in [\mathrm{T}]$ such that
\begin{align}
\label{eq:net_property_main}
\|\mathbf{c}_{i_{\star}} - \overline{\mathbf{c}}_{\mathsmaller{\mathtt{(B)}}}\|_{2}
\le
\epsilon/2.
\end{align}
In the $i_{\star}$th round, the algorithm samples the point $\mathbf{c}_{i_{\star}}$ and computes a feasible pair $(\mathbf{u}_{i_{\star}}, \mathbf{v}_{i_{\star}})$ via the two step maximization procedure,
that is,
\begin{align}
\mathbf{u}_{i_{\star}}
{\triangleq}
\argmax_{\mathbf{u}\in \mathcal{U}}
\mathbf{u}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{i_{\star}}
\; \text{and} \;
\mathbf{v}_{i_{\star}}
{\triangleq}
\argmax_{\mathbf{v}\in \mathcal{V}}
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}.
\nonumber
\end{align}
We have:
\begin{align}
{\mathbf{u}}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}
\mathbf{B}
{\mathbf{v}}_{\mathsmaller{\mathtt{(B)}}}
&=
\rho \cdot
\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\overline{\mathbf{c}}_{\mathsmaller{\mathtt{(B)}}}
\nonumber\\&=
\rho \cdot
\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{i_{\star}}
+
\rho \cdot
\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\bigl(\overline{\mathbf{c}}_{\mathsmaller{\mathtt{(B)}}}-\mathbf{c}_{i_{\star}}\bigr)
\nonumber\\&\le
\rho \cdot
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{i_{\star}}
+
\rho \cdot
\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\bigl(\overline{\mathbf{c}}_{\mathsmaller{\mathtt{(B)}}}-\mathbf{c}_{i_{\star}}\bigr)
\nonumber\\&\le
\rho \cdot
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{i_{\star}}
+
\tfrac{\epsilon}{2} \cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
.
\label{base-inequality_in_main}
\end{align}
The first step follows by the definition of $\overline{\mathbf{c}}_{\mathsmaller{\mathtt{(B)}}}$
and the second by linearity.
The first inequality follows from the fact that $\mathbf{u}_{i_{\star}}$ maximizes the first term over all $\mathbf{u}\in \mathcal{U}$.
The last inequality follows straightforwardly from the fact
that $\|\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}\|_{2}=1$ and $\rho \le 1$ (see Lemma~\ref{lemma:abs-trace-XAY-ub}).
Using similar arguments,
\begin{align}
\rho \cdot
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{i_{\star}}
&=
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{V}^{\transpose}
\mathbf{v}_{\mathsmaller{\mathtt{(B)}}}
+
\rho \cdot
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\left(
\mathbf{c}_{i_{\star}} - \overline{\mathbf{c}}_{\mathsmaller{\mathtt{(B)}}}
\right)
\nonumber\\&\le
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{B}
\mathbf{v}_{i_{\star}}
+
\tfrac{\epsilon}{2} \cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\label{bound_on_sharp_x_in_main}
\end{align}
The inequality follows by the fact that $\mathbf{v}_{i_{\star}}$
maximizes the inner product with $\mathbf{B}^{\transpose}\mathbf{u}_{i_{\star}}$ over all ${\mathbf{v} \in \mathcal{V}}$,
as well as that $\|\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}\|_{2}=1$ and $\rho \le 1$.
Combining~\eqref{base-inequality_in_main} and~\eqref{bound_on_sharp_x_in_main}, we obtain
\begin{align}
\mathbf{u}_{i_{\star}}^{\transpose}
\mathbf{B}
\mathbf{v}_{i_{\star}}
\ge
\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}
\mathbf{B}
\mathbf{v}_{\mathsmaller{\mathtt{(B)}}}
-
\epsilon
\cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\label{eq:guarantees_for_low_rank}
\end{align}
Algorithm~\ref{algo:cca} computes multiple candidate solution pairs and outputs the pair $(\mathbf{u}_{\sharp}, \mathbf{v}_{\sharp})$ that achieves the maximum objective value.
The latter is at least as high as that achieved by
$(\mathbf{u}_{i_{\star}}, \mathbf{v}_{i_{\star}})$.
Inequality~\eqref{eq:guarantees_for_low_rank} establishes an approximation guarantee for the low-rank problem~\eqref{cca:on-low-rank}.
Those can be translated to guarantees on the original problem with input argument~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
Let $\mathbf{u}_{\star}$ and $\mathbf{v}_{\star}$ denote the (unknown) optimal solution of the sparse CCA problem \eqref{cca:def}.
By the definition of $\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}$ and $\mathbf{v}_{\mathsmaller{\mathtt{(B)}}}$, it follows that
\begin{align}
\mathbf{u}_{\mathsmaller{\mathtt{(B)}}}^{\transpose}\mathbf{B}\mathbf{v}_{\mathsmaller{\mathtt{(B)}}}
\ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{B}
\mathbf{v}_{\star}
&=
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
\mathbf{u}_{\star}^{\transpose}
(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}-\mathbf{B})
\mathbf{v}_{\star}
\nonumber\\& \ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
\sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\label{eq:bp03_main}
\end{align}
Similarly,
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
&=
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{B}
\mathbf{v}_{\sharp}
-
\mathbf{u}_{\sharp}^{\transpose}
(\mathbf{B}-\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
\mathbf{v}_{\sharp}
\nonumber\\
&\ge
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{B}
\mathbf{v}_{\sharp}
-
\sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\label{eq:bp02_main}
\end{align}
Combining~\eqref{eq:bp03_main} and~\eqref{eq:bp02_main}
with~\eqref{eq:guarantees_for_low_rank},
we obtain the approximation guarantees of Theorem~\ref{thm:main-algo-guarantees}.
The running time of Algorithm~\ref{algo:cca} follows straightforwardly by inspection.
The algorithm first computes the truncated singular value decomposition of inner dimension $r$ in time denoted by $\mathrm{T}_{\texttt{SVD}}(r)$.
Subsequently, it performs $\mathrm{T}$ iterations.
The cost of each iteration is determined by the cost of the matrix-vector multiplications and the running times $\mathrm{T}_{\mathcal{U}}$ and $\mathrm{T}_{\mathcal{V}}$ of the operators $\mathsf{P}_{\mathcal{U}}(\cdot)$ and $\mathsf{P}_{\mathcal{V}}(\cdot)$.
Note that matrix multiplications can exploit the available matrix decomposition and are performed in time ${r \cdot \max\lbrace m, n\rbrace}$.
Substituting the value of $\mathrm{T}$ with that specified in Lemma~\ref{lem:num-of-random-points} completes the proof of Thm.~\ref{thm:main-algo-guarantees}.
The proof of Theorem~\ref{thm:main-algo-guarantees-special-case} follows a similar path; see Appendix Sec.~\ref{sec:approx-proof}.
\section{Hardness}\label{sec:hardness}
We provide a proof for the NP-hardness of the constrained (and specifically sparse) CCA problem via a reduction from sparse PCA.
Recall that sparse PCA is the following optimization problem:
\begin{align}
\max_{
\substack{
\mathbf{x} :
\|\mathbf{x}\|_{0} = k\\
\|\mathbf{x}\|_{2} = 1
}
}
\mathbf{x}^{\transpose}
\mathbf{A}
\mathbf{x},
\label{pca:def}
\end{align}
where~$k$ is a given parameter and~$\mathbf{A}$ a given $n \times n$ positive semidefinite (PSD) matrix.
We show that the sparse PCA problem~\eqref{pca:def} reduces to the sparse CCA problem~\eqref{cca:def} and in particular the maximization
\begin{align}
\max_{
\substack{
\mathbf{x} :
\|\mathbf{x}\|_{0} = k,
\|\mathbf{x}\|_{2} = 1\\
\mathbf{y} :
\|\mathbf{y}\|_{0} = k,
\|\mathbf{y}\|_{2} = 1
}
}
\mathbf{x}^{\transpose}
\mathbf{A}
\mathbf{y}.
\label{cca:reduction:sparse}
\end{align}
The only difference between~\eqref{pca:def} and~\eqref{cca:reduction:sparse} is that in the latter the optimal values for the two variables $\mathbf{x}$ and $\mathbf{y}$ may be different.
If we add the constraint $\mathbf{x}=\mathbf{y}$ in
~\eqref{cca:reduction:sparse}, then then two maximizations are identical.
We show that this is not necessary: since $\mathbf{A}$ is PSD, the optimal solution of~\eqref{cca:reduction:sparse} will inherently satisfy $\mathbf{x}=\mathbf{y}$, and in turn the two maximizations are equivalent.
Let $\mathbf{U}, \mathbf{\Lambda}$ be the eigenvalue decomposition of~$\mathbf{A}$: the $n \times n$ matrix $\mathbf{U}$ contains the eigenvectors, while the $n \times n$ diagonal $\mathbf{\Lambda}$ contains the eigenvalues $\lambda_{1}, \hdots, \lambda_{n} \ge 0$ in decreasing order.
Let ($\mathbf{x}_{\star}$, $\mathbf{y}_{\star}$) be the optimal solution of~\eqref{cca:reduction:sparse}.
Further, let
$\overline{\mathbf{x}} = \mathbf{U}^{\transpose}\mathbf{x}_{\star}$,
and
$\overline{\mathbf{y}} = \mathbf{U}^{\transpose}\mathbf{y}_{\star}$.
Then,
\begin{align}
\mathbf{x}_{\star}^{\transpose}
\mathbf{A}
\mathbf{y}_{\star}
=
\mathbf{x}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Lambda}\mathbf{U}^{\transpose}
\mathbf{y}_{\star}
=
\overline{\mathbf{x}}^{\transpose}
\mathbf{\Lambda}
\overline{\mathbf{y}}
=
\sum_{i=1}^{n}
\overline{x}_{i} \overline{y}_{i} \lambda_{i}.
\end{align}
\begin{theorem}[Weighted Cauchy-Schwarz inequality; \citeappendix{cvetkovski:inequalities}, Theorem 10.1]
\label{thm:weighted_cs}
Let $a_{i}, b_{i} \in \mathbb{R}$ be real numbers and let ${m_{i} \in \mathbb{R}^{+}}$,
$i=1,2,\hdots, n$.
Then,
\begin{align}
\left(
\sum_{i=1}^{n}
a_{i} b_{i} m_{i}
\right)^{2}
\le
\left(
\sum_{i=1}^{n}
a_{i}^{2} m_{i}
\right)
\left(
\sum_{i=1}^{n}
b_{i}^{2} m_{i}
\right).
\nonumber
\end{align}
Equality occurs if and only if $\tfrac{a_{1}}{b_{2}} = \hdots = \tfrac{a_{n}}{b_{n}}$.
\end{theorem}
By Theorem~\ref{thm:weighted_cs},
\begin{align}
\left(
\mathbf{x}_{\star}^{\transpose}
\mathbf{A}
\mathbf{y}_{\star}
\right)^{2}
\le
\left(
\sum_{i=1}^{n}
\overline{x}_{i}^{2} \lambda_{i}
\right)
\left(
\sum_{i=1}^{n}
\overline{y}_{i}^{2} \lambda_{i}
\right),
\nonumber
\end{align}
with equality if and only if there exists a constant $c \in \mathbb{R}$ such that
$\overline{\mathbf{x}} = c \cdot \overline{\mathbf{y}}$,
and taking into account that both $\overline{\mathbf{x}}$ and $\overline{\mathbf{y}}$ are unit norm vectors, it follows that $\overline{\mathbf{x}} = \overline{\mathbf{y}}$.
Finally, since $\mathbf{U}$ is a full rank matrix (in fact orthonormal basis of $\mathbb{R}^{n}$), it follows that
$\mathbf{x}_{\star}=\mathbf{y}_{\star}$ must hold.
\section{Proofs}
\label{sec:approx-proof}
For the remainder of this section, we define
\begin{align}
\left(
\mathbf{u}_{\star}, \mathbf{v}_{\star}
\right)
{\triangleq}
\argmax_{\mathbf{u} \in \mathcal{U}, \mathbf{v} \in \mathcal{V}}
\mathbf{u}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v},
\nonumber
\end{align}
\textit{i.e.}, $\mathbf{u}_{\star}, \mathbf{v}_{\star}$ is a feasible pair that maximizes --not necessarily uniquely-- the objective.
Here, $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ is a given $m \times n$ real matrix, which plays the role of
Further, we assume that there exists procedures to compute the exact soluton to
$\mathsf{P}_{\mathcal{U}}(\cdot)$ and $\mathsf{P}_{\mathcal{V}}(\cdot)$ in~\eqref{lowrank:solve-for-x} and~\eqref{lowrank:solve-for-y}, running in time
$\mathrm{T}_{\mathcal{U}}$ and $\mathrm{T}_{\mathcal{V}}$, respectively.
The following results can be easily adapted for the case where these procedures yield approximate solutions.
\begin{lemma}
\label{core-rank-r-guarantees}
For any real ${m \times n}$ matrix~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ with $\text{rank}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})= r \le \max\lbrace m, n\rbrace$
and
$\epsilon \in (0, 1)$,
Algorithm~\ref{algo:cca} with input $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, $r$,
and $T=\widetilde{O}\bigl(2^{r \cdot \log_{2}(\sfrac{2}{\epsilon})} \bigr)$
outputs $\mathbf{u}_{\sharp} \in \mathcal{U}$
and $\mathbf{v}_{\sharp} \in \mathcal{V}$ such that
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
\ge
\mathbf{u}_{\star}^{\transpose} \mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}} \mathbf{v}_{\star}
-
\epsilon \cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}),
\nonumber
\end{align}
in time
$
\mathrm{T}_{\mathsmaller{\mathsf{SVD}}}(r)
+ O\mathopen{}\bigl(
T \cdot \bigl(\mathrm{T}_{\mathcal{U}}+\mathrm{T}_{\mathcal{V}}+r \cdot \max\lbrace m, n\rbrace \bigr)\bigr)
$.
\end{lemma}
\begin{proof}
In the sequel, $\mathbf{U}$, $\mathbf{\Sigma}$ and $\mathbf{V}$ are used to denote the $r$-truncated singular value decomposition of $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
Note that the lemma assumes that the accuracy parameter $r$ is equal to the rank of the input matrix $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ and hence ${\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}} = \mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\transpose}}$.
Recall that $\mathbf{u}_{\star}, \mathbf{v}_{\star}$ is a pair ---not necessarily unique--- that maximizes the objective $\mathbf{u}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}$ over all feasible solutions.
Define $\mathbf{c}_{\star} {\triangleq} \mathbf{V}^{\transpose}\mathbf{v}_{\star}$.
Note that
$\mathbf{c}_{\star}$ is a vector in $\mathbf{R}^{r \times 1}$ with $\|\mathbf{c}_{\star}\|_{2} \le 1$
since the $r$ columns of $\mathbf{V}$ are orthonormal and ${\|\mathbf{v}_{\star}\|_{2} = 1}$.
Finally, let $\overline{\mathbf{c}}_{\star} {\triangleq} \mathbf{c}_{\star} /\|\mathbf{c}_{\star}\|_{2}$.
Note that $\|\mathbf{c}_{\star}\|_{2} > 0$ since by assumption $\mathbf{u}_{\star}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}_{\star} > 0$.
Algorithm~\ref{algo:cca} operates in an iterative fashion. In each iteration, it independently considers a point~$\mathbf{c}$ selected randomly and uniformly from the $r$-dimensional $\ell_{2}$-unit sphere $\mathbb{S}_{2}^{r-1}$ and generates a candidate solution pair at each point.
For $\mathrm{T}=\widetilde{O}\bigl(2^{r \cdot \log_{2}(\sfrac{2}{\epsilon})} \bigr)$,
the collection of randomly sampled poins forms an $\sfrac{\epsilon}{2}$-net for $\mathbb{S}_{2}^{r-1}$.
By definition, the $\sfrac{\epsilon}{2}$-net contains a point $\widetilde{\mathbf{c}} \in \mathbb{R}^{r \times 1}$, such that
\begin{align}\label{eq:net_property}
\|\widetilde{\mathbf{c}} - \overline{\mathbf{c}}_{\star}\|_{2} \le \epsilon/2.
\end{align}
Let $(\widetilde{\mathbf{u}}, \widetilde{\mathbf{v}})$ be the candidate solution pair computed at $\widetilde{\mathbf{c}}$ by the two step maximization procedure, \textit{i.e.}, let
\begin{align}
\widetilde{\mathbf{u}}
\;{\triangleq}\;
\argmax_{\mathbf{u}\in \mathcal{U}}
\mathbf{u}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
\end{align}
and
\begin{align}
\widetilde{\mathbf{v}}
\;{\triangleq}\;
\argmax_{\mathbf{v}\in \mathcal{V}}
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}.
\label{def-of-nearXandY}
\end{align}
By the definition of $\mathbf{c}_{\star}$,
and letting $\rho{\triangleq} \|\mathbf{c}_{\star}\|_{2}$
\begin{align}
{\mathbf{u}}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
{\mathbf{v}}_{\star}
&=
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{\star}
\nonumber\\&=
\rho \cdot
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\overline{\mathbf{c}}_{\star}
\nonumber\\&=
\rho \cdot
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
+
\rho \cdot
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\bigl(\overline{\mathbf{c}}_{\star}-\widetilde{\mathbf{c}}\bigr)
\nonumber\\&\le
\rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
+
\rho \cdot
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\bigl(\overline{\mathbf{c}}_{\star}-\widetilde{\mathbf{c}}\bigr)
\nonumber\\&\le
\rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
+
\tfrac{\epsilon}{2} \cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
.
\label{base-inequality}
\end{align}
The first inequality follows from the fact that $\widetilde{\mathbf{u}}$ by definition maximizes the first term over all $\mathbf{u}\in \mathcal{U}$.
The last inequality is due to Lemma~\ref{lemma:abs-trace-XAY-ub} and the fact that $\|\mathbf{u}_{\star}\|_{2}=1$ and $\rho \le 1$.
We further upper bound the right hand side of~\eqref{base-inequality} as follows:
\begin{align}
& \rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
\nonumber\\&=
\rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\overline{\mathbf{c}}_{\star}
+
\rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\left(\widetilde{\mathbf{c}} - \overline{\mathbf{c}}_{\star}\right)
\nonumber\\&=
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{\star}
+
\rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\left(\widetilde{\mathbf{c}} - \overline{\mathbf{c}}_{\star}\right)
\nonumber\\&=
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{V}^{\transpose}
\mathbf{v}_{\star}
+
\rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\left(\widetilde{\mathbf{c}} - \overline{\mathbf{c}}_{\star}\right)
\nonumber\\&\le
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{V}^{\transpose}
\widetilde{\mathbf{v}}
+
\rho \cdot
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\left(\widetilde{\mathbf{c}} - \overline{\mathbf{c}}_{\star}\right)
\label{optimal_y_sharp}
\\&\le
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\widetilde{\mathbf{v}}
+
\tfrac{\epsilon}{2} \cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\label{bound_on_sharp_x}
\end{align}
Inequality~\eqref{optimal_y_sharp} follows by the fact that $\widetilde{\mathbf{v}}$ by definition~\eqref{def-of-nearXandY} maximizes the bilinear term $\mathbf{u}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}$ over all ${\mathbf{v} \in \mathcal{V}}$ when $\mathbf{u}=\widetilde{\mathbf{u}}$.
The last inequality is once again due to Lemma~\ref{lemma:abs-trace-XAY-ub} and the fact that $\|\mathbf{u}_{\star}\|_{2}=1$ and $\rho \le 1$.
Combining~\eqref{base-inequality} and~\eqref{bound_on_sharp_x}, we obtain
\begin{align}
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\widetilde{\mathbf{v}}
\ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
\epsilon
\cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\nonumber
\end{align}
Algorithm~\ref{algo:cca} computes multiple candidate solution pairs and outputs the one that maximizes the objective.
Therefore, the output pair $(\mathbf{u}_{\sharp}, \mathbf{v}_{\sharp})$ must achieve a value as least as high as that achieved by
$(\widetilde{\mathbf{u}}, \widetilde{\mathbf{v}})$,
which implies the desired guarantee.
The running time of Algorithm~\ref{algo:cca} follows straightforwardly by inspection.
The algorithm first computes the truncated singular value decomposition of inner dimension $r$ in time denoted by $\mathrm{T}_{\texttt{SVD}}(r)$.
Subsequently, it performs $T$ iterations.
The cost of each iteration is determined by the cost of the matrix-vector multiplications and the running times $\mathrm{T}_{\mathcal{U}}$ and $\mathrm{T}_{\mathcal{V}}$ of the operators $\mathsf{P}_{\mathcal{U}}(\cdot)$ and $\mathsf{P}_{\mathcal{V}}(\cdot)$.
Note that matrix multiplications can exploit the available singular value decomposition of $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ and are performed in time ${r \cdot \max\lbrace m, n\rbrace}$.
Substituting the value of $\mathrm{T}$, completes the proof.
\end{proof}
\begin{reptheorem}{thm:main-algo-guarantees}
For any real ${m \times n}$ matrix~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$,
$\epsilon \in (0, 1)$, and $r \le \max\lbrace m, n\rbrace$,
Algorithm~\ref{algo:cca} with input $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, $r$,
and $T=\widetilde{O}\bigl(2^{r \cdot \log_{2}(\sfrac{2}{\epsilon})} \bigr)$
outputs $\mathbf{u}_{\sharp} \in \mathcal{U}$
and $\mathbf{v}_{\sharp} \in \mathcal{V}$ such that
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
\ge
\mathbf{u}_{\star}^{\transpose} \mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}} \mathbf{v}_{\star}
-
\epsilon \cdot \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
-
2 \cdot \sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}),
\nonumber
\end{align}
in time
$
\mathrm{T}_{\mathsmaller{\mathsf{SVD}}}(r) + O\mathopen{}\bigl(T \cdot \bigl(\mathrm{T}_{\mathcal{U}}+\mathrm{T}_{\mathcal{V}}+r \cdot \max\lbrace m, n\rbrace \bigr)\bigr)
$.
\end{reptheorem}
\begin{proof}
Recall that Algorithm~\ref{algo:cca} with input an $m \times n$ matrix $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ and accuracy parameter $r$,
first computes a rank-$r$ truncated singular value decomposition $\mathbf{U}$, $\mathbf{\Sigma}$, $\mathbf{V}$ and operates on that principal subspace of $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
Let $\mathbf{B}$ be the $m \times n$ best rank-$r$ approximation of $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ under the spectral norm.
Then
$\mathbf{B}=\mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\transpose}$.
One can easily verify that running Algorithm~\ref{algo:cca} with input $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ and accuracy parameter $r$, is equivalent to applying the algorithm on $\mathbf{B}$ with the same parameters.
By Lemma~\ref{core-rank-r-guarantees},
Algorithm~\ref{algo:cca} outputs
$\mathbf{u}_{\sharp}, \mathbf{v}_{\sharp}$
such that
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{B}
\mathbf{v}_{\sharp}
\ge
\widehat{\mathbf{u}}_{\star}^{\transpose}
\mathbf{B}
\widehat{\mathbf{v}}_{\star}
-
\epsilon \cdot \sigma_{1}(\mathbf{B}),
\label{eq:guar-lowrank}
\end{align}
where
\begin{align}
\left(
\widehat{\mathbf{u}}_{\star},
\widehat{\mathbf{v}}_{\star}
\right)
{\triangleq}
\argmax_{\mathbf{u} \in \mathcal{U}, \mathbf{v} \in \mathcal{V}}
\mathbf{u}^{\transpose}\mathbf{B}\mathbf{v}
\nonumber
\end{align}
is a pair that optimally solves the maximization on the rank-$r$ matrix $\mathbf{B}$.
By the optimality of the pair
$\widehat{\mathbf{u}}_{\star},
\widehat{\mathbf{v}}_{\star}$ for the rank-$r$ problem,
it follows that
\begin{align}
\widehat{\mathbf{u}}_{\star}^{\transpose}
\mathbf{B}
\widehat{\mathbf{v}}_{\star}
\ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{B}
\mathbf{v}_{\star}.
\label{eq:optimal_low_better}
\end{align}
Recall that $\mathbf{u}_{\star}$, $\mathbf{v}_{\star}$ is the pair that optimally solves the maximization on the original input matrix~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
Further,
\begin{align}
\mathbf{u}_{\star}^{\transpose}
\mathbf{B}
\mathbf{v}_{\star}
&=
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
\mathbf{u}_{\star}^{\transpose}
(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}-\mathbf{B})
\mathbf{v}_{\star}
\nonumber\\& \ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
\bigl\lvert
\mathbf{u}_{\star}^{\transpose}
(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}-\mathbf{B})
\mathbf{v}_{\star}
\bigr\rvert
\nonumber\\& \ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
\sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\label{eq:bp03}
\end{align}
Combining~\eqref{eq:bp03} with
\eqref{eq:guar-lowrank} and
\eqref{eq:optimal_low_better},
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{B}
\mathbf{v}_{\sharp}
\ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
\sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
-
\epsilon \cdot \sigma_{1}(\mathbf{B}).
\nonumber
\end{align}
Finally,
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{B}
\mathbf{v}_{\sharp}
&=
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
-
\mathbf{u}_{\sharp}^{\transpose}
(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}-\mathbf{B})
\mathbf{v}_{\sharp}
\nonumber\\
&\le
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
+
\bigl\lvert
\mathbf{u}_{\sharp}^{\transpose}
(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}-\mathbf{B})
\mathbf{v}_{\sharp}
\bigr\rvert.
\nonumber\\
&\le
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
+
\sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}).
\label{eq:bp02}
\end{align}
Combining with the previous inequality,
we obtain
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
\ge
\mathbf{u}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\star}
-
2 \cdot \sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
-
\epsilon \cdot \sigma_{1}(\mathbf{B}).
\nonumber
\end{align}
Noting that ${\sigma_{1}(\mathbf{B}) = \sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})}$ completes the proof of the approximation guarantee.
The running time of the algorithm is established in Lemma~\ref{core-rank-r-guarantees}.
\end{proof}
\begin{lemma}
\label{same-lemma-for-the-special-case}
For any real ${m \times n}$ matrix~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$ with $\text{rank}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})= r \le \max\lbrace m, n\rbrace$
and
$\epsilon \in (0, 1)$,
if $\mathcal{V} = \left\{\mathbf{v}~:~\|\mathbf{v}\|_2 = 1\right\}$,
then
Algorithm~\ref{algo:cca} with input $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, $r$,
and $\mathrm{T}=\widetilde{O}\bigl(2^{r \cdot \log_{2}(\sfrac{2}{\epsilon})} \bigr)$
outputs $\mathbf{u}_{\sharp} \in \mathcal{U}$
and $\mathbf{v}_{\sharp} \in \mathcal{V}$ such that
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
\ge
(1-\epsilon) \cdot \mathbf{u}_{\star}^{\transpose} \mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}} \mathbf{v}_{\star}
\nonumber
\end{align}
in time
$
\mathrm{T}_{\mathsmaller{\mathsf{SVD}}}(r) + O\mathopen{}\bigl(\mathrm{T} \cdot \bigl(\mathrm{T}_{\mathcal{U}}+\mathrm{T}_{\mathcal{V}}+r \cdot \max\lbrace m, n\rbrace \bigr)\bigr)
$.
\end{lemma}
\begin{proof}
The lemma focuses on the special case where the feasible region for the variable $\mathbf{v}$ coincides with the set of vectors with unit $\ell_{2}$ norm,
\textit{i.e.},
\begin{align}
\mathcal{V}
=
\lbrace
\mathbf{v}: \|\mathbf{v}\|_{2}=1
\rbrace.
\label{Y_unit_norm}
\end{align}
The feasible region $\mathcal{U}$ for $\mathbf{u}$ is arbitrary, assuming once again that there exists an efficient operator $\mathsf{P}_{\mathcal{U}}(\cdot)$.
By the Cauchy-Schwarz inequality,
for any $\mathbf{u}_{0} \in \mathbb{R}^{m \times 1}$,
\begin{align}
\mathbf{u}_{0}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}
&\le
\|\mathbf{u}_{0}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\|_{2},
\quad \forall \mathbf{v} \in \mathcal{V}.
\end{align}
In fact, equality is achieved when $\mathbf{v}$ is aligned with
$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{0}$,
\textit{i.e.}, for
$\mathbf{v} = \mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{0}/\|\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{0}\|_{2} \in \mathcal{V}$.
In turn, for any $\mathbf{u}_{0} \in \mathbb{R}^{m \times 1}$,
\begin{align}
\max_{\mathbf{v} \in \mathcal{V}}
\mathbf{u}_{0}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}
&=
\mathbf{u}_{0}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{0}/\|\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{0}\|_{2}
\nonumber\\
&=
\|\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{0}\|_{2}
\nonumber\\
&=
\|\mathbf{V}\mathbf{\Sigma}\mathbf{U}^{\transpose}\mathbf{u}_{0}\|_{2}
\nonumber\\
&=
\|\mathbf{\Sigma}\mathbf{U}^{\transpose}\mathbf{u}_{0}\|_{2},
\label{eq:y_objective_as_x_norm}
\end{align}
where the last equality follows from the fact that the $r$ columns of $\mathbf{V}$ are orthonormal.
We now proceed in a fashion very similar to that in the proof of Lemma~\ref{core-rank-r-guarantees}.
Recall that $\mathbf{u}_{\star}, \mathbf{v}_{\star}$ is a pair that maximizes --not necessarily uniquely-- the objective $\mathbf{u}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}$ over all feasible solutions,
and define $\mathbf{c}_{\star} {\triangleq} \mathbf{V}^{\transpose}\mathbf{v}_{\star}$.
Note that here,
\begin{align}
\mathbf{c}_{\star}
{\triangleq}
\mathbf{V}^{\transpose}\mathbf{v}_{\star}
&=
\mathbf{V}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{\star}/\|\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}^{\transpose}\mathbf{u}_{\star}\|_{2}
\nonumber\\&=
\mathbf{\Sigma}\mathbf{U}^{\transpose}\mathbf{u}_{\star}/\|\mathbf{V}\mathbf{\Sigma}\mathbf{U}^{\transpose}\mathbf{u}_{\star}\|_{2}
\nonumber\\&=
\mathbf{\Sigma}\mathbf{U}^{\transpose}\mathbf{u}_{\star}/\|\mathbf{\Sigma}\mathbf{U}^{\transpose}\mathbf{u}_{\star}\|_{2}
\end{align}
and hence, $\|\mathbf{c}_{\star}\|_{2}=1$.
Following similar reasoning as in the proof of Lemma~\ref{core-rank-r-guarantees},
Algorithm~\ref{algo:cca} considers a point $\widetilde{\mathbf{c}} \in \mathbb{R}^{r \times 1}$, such that
\begin{align}
\|\widetilde{\mathbf{c}} - \overline{\mathbf{c}}_{\star}\|_{2} \le \epsilon.
\nonumber
\end{align}
Let $(\widetilde{\mathbf{u}}, \widetilde{\mathbf{v}})$ be the candidate solution pair computed at $\widetilde{\mathbf{c}}$ by the two step maximization procedure.
We have,
\begin{align}
{\mathbf{u}}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
{\mathbf{v}}_{\star}
&=
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\mathbf{c}_{\star}
\nonumber\\&=
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
+
\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\bigl(\mathbf{c}_{\star}-\widetilde{\mathbf{c}}\bigr)
\nonumber\\&\le
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
+
\|\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}\|_{2}
\|\mathbf{c}_{\star}-\widetilde{\mathbf{c}}\|_{2}
\nonumber\\&\le
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
+
\epsilon \cdot \|\mathbf{u}_{\star}^{\transpose}
\mathbf{U}\mathbf{\Sigma}\|_{2}
\label{eq:special-case-upper-bound}
.
\end{align}
where the first inequality follows from the fact that $\widetilde{\mathbf{u}}$ by definition maximizes the first term at $\widetilde{\mathbf{c}}$ over all $\mathbf{u} \in \mathcal{U}$ and the Cauchy-Schwarz inequality.
The key difference from the proof of Lemma~\ref{core-rank-r-guarantees},
is that the term $\|\mathbf{u}_{\star}^{\transpose}\mathbf{U}\mathbf{\Sigma}\|_{2}$ in the right-hand side coincides with the optimal objective value $ {\mathbf{u}}_{\star}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}{\mathbf{v}}_{\star}$ as follows from~\eqref{eq:y_objective_as_x_norm}.
For comparison, note that in the proof of Lemma~\ref{core-rank-r-guarantees} it was loosely upper bounded by~${\sigma_{1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})}$.
Continuing from~\eqref{eq:special-case-upper-bound},
\begin{align}
\left(1-\epsilon\right)
\cdot
{\mathbf{u}}_{\star}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
{\mathbf{v}}_{\star}
&\le
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
.
\label{eq:almost_there}
\end{align}
But, once again by the Cauchy-Schwarz inequality,
\begin{align}
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
&\le
\|\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}\|_{2}
\|\widetilde{\mathbf{c}}\|_{2}
\nonumber\\&=
\|\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}\|_{2}
\end{align}
and by~\eqref{eq:y_objective_as_x_norm},
\begin{align}
\widetilde{\mathbf{u}}^{\transpose}
\mathbf{U}\mathbf{\Sigma}
\widetilde{\mathbf{c}}
&=
\max_{\mathbf{v} \in \mathcal{V}}
\widetilde{\mathbf{u}}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}
=
\widetilde{\mathbf{u}}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\widetilde{\mathbf{v}}
.
\label{eq:last-cookie}
\end{align}
Combining~\eqref{eq:last-cookie} with~\eqref{eq:almost_there},
\begin{align}
\widetilde{\mathbf{u}}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\widetilde{\mathbf{v}}
\ge
(1-\epsilon) \cdot
\mathbf{u}_{\star}^{\transpose}\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}\mathbf{v}_{\star}
\end{align}
Recalling that Algorithm~\ref{algo:cca} outputs the candidate pair that maximizes the objective among all computed feasible points implies the desired result.
\end{proof}
\begin{reptheorem}{thm:main-algo-guarantees-special-case}
For any real ${m \times n}$ matrix~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$
and
$\epsilon \in (0, 1)$,
if $\mathcal{V} = \left\{\mathbf{v}~:~\|\mathbf{v}\|_2 = 1\right\}$,
then
Algorithm~\ref{algo:cca} with input $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$, $r$,
and $T=\widetilde{O}\bigl(2^{r \cdot \log_{2}(\sfrac{2}{\epsilon})} \bigr)$
outputs $\mathbf{u}_{\sharp} \in \mathcal{U}$
and $\mathbf{v}_{\sharp} \in \mathcal{V}$ such that
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}
\mathbf{v}_{\sharp}
\ge
(1-\epsilon) \cdot \mathbf{u}_{\star}^{\transpose} \mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}} \mathbf{v}_{\star}
-
2 \cdot \sigma_{r+1}(\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}})
\nonumber
\end{align}
in time
$
\mathrm{T}_{\mathsmaller{\mathsf{SVD}}}(r) + O\mathopen{}\bigl(T \cdot \bigl(\mathrm{T}_{\mathcal{U}}+\mathrm{T}_{\mathcal{V}}+r \cdot \max\lbrace m, n\rbrace \bigr)\bigr)
$.
\end{reptheorem}
\begin{proof}
The Theorem follows from Lemma~\ref{same-lemma-for-the-special-case}.
The proof is similar to that of Theorem~\ref{thm:main-algo-guarantees}.
The main difference lies in substituting \eqref{eq:guar-lowrank} with
\begin{align}
\mathbf{u}_{\sharp}^{\transpose}
\mathbf{B}
\mathbf{v}_{\sharp}
\ge
(1-\epsilon) \cdot
\mathbf{u}_{\mathsmaller{(B)}\star}^{\transpose}
\mathbf{B}
\mathbf{v}_{\mathsmaller{(B)}\star}.
\label{Y_unit:low-rank-guarantee}
\end{align}
The remainder of the proof easily follows.
\end{proof}
\section{Auxiliary Lemmata}
\begin{lemma}
\label{holder-consequence}
Let $a_1,\hdots , a_n$
and $b_1, \hdots, b_n$ be $2n$ real numbers and let $p$ and
$q$ be two numbers such that ${1/p} + {1/q} = 1$ and $p>1$. We have
\begin{align}
\bigl\lvert
\sum_{i=1}^{n} a_{i}b_{i}
\bigr\rvert
\le
\bigl( \sum_{i=1}^{n} \lvert a_{i}\rvert^{p} \bigr)^{1/p}
\cdot
\bigl( \sum_{i=1}^{n} \lvert b_{i}\rvert^{q} \bigr)^{1/q}.
\nonumber
\end{align}
\end{lemma}
\begin{lemma}
\label{lemma:inner-prod-ub}
For any $\mathbf{A}, \mathbf{B} \in \mathbb{R}^{n \times k}$,
\begin{align}
\bigl\lvert
\langle \mathbf{A}, \mathbf{B} \rangle
\bigr\rvert
{\triangleq}
\bigl\lvert
{\textsc{Tr}}\bigl( \mathbf{A}^{\transpose} \mathbf{B}\bigr)
\bigr\rvert
\le
\|\mathbf{A}\|_{{\textnormal{\tiny{F}}}}
\|\mathbf{B}\|_{{\textnormal{\tiny{F}}}}.
\nonumber
\end{align}
\end{lemma}
\begin{proof}
Treating $\mathbf{A}$ and $\mathbf{B}$ as vectors,
the lemma follows immediately from Lemma~\ref{holder-consequence} for $p=q=2$.
\end{proof}
\begin{lemma}
\label{lemma:frob-of-matrix-prod}
For any two real matrices $\mathbf{A}$ and $\mathbf{B}$ of appropriate dimensions,
\begin{align}
\|\mathbf{A}\mathbf{B}\|_{{\textnormal{\tiny{F}}}}
\le
\min\mathopen{}
\bigl\lbrace
\|\mathbf{A}\|_{2} \|\mathbf{B}\|_{{\textnormal{\tiny{F}}}}, \;
\|\mathbf{A}\|_{{\textnormal{\tiny{F}}}} \|\mathbf{B}\|_{2}
\bigr\rbrace.
\nonumber
\end{align}
\end{lemma}
\begin{proof}
Let $\mathbf{b}_{i}$ denote the $i$th column of $\mathbf{B}$.
Then,
\begin{align}
\|\mathbf{A}\mathbf{B}\|_{{\textnormal{\tiny{F}}}}^{2}
& =
\sum_{i} \| \mathbf{A} \mathbf{b}_{i} \|_{2}^{2}
\le
\sum_{i} \| \mathbf{A}\|_{2}^{2} \|\mathbf{b}_{i} \|_{2}^{2}
\nonumber\\&=
\| \mathbf{A}\|_{2}^{2} \sum_{i} \|\mathbf{b}_{i} \|_{2}^{2}
=
\| \mathbf{A}\|_{2}^{2} \|\mathbf{B} \|_{{\textnormal{\tiny{F}}}}^{2}.
\nonumber
\end{align}
Similarly, using the previous inequality,
\begin{align}
\|\mathbf{A}\mathbf{B}\|_{{\textnormal{\tiny{F}}}}^{2}
=
\|\mathbf{B}^{\transpose}\mathbf{A}^{\transpose}\|_{{\textnormal{\tiny{F}}}}^{2}
\le
\|\mathbf{B}^{\transpose}\|_{2}^{2}\|\mathbf{A}^{\transpose}\|_{{\textnormal{\tiny{F}}}}^{2}
=
\|\mathbf{B}\|_{2}^{2}\|\mathbf{A}\|_{{\textnormal{\tiny{F}}}}^{2}.
\nonumber
\end{align}
The desired result follows combining the two upper bounds.
\end{proof}
\begin{lemma}
\label{lemma:abs-trace-XAY-ub}
For any real
$m \times k$ matrix $\mathbf{X}$, $m \times n$ matrix $\mathbf{A}$, and $n \times k$ matrix $\mathbf{Y}$,
\begin{align}
\bigl\lvert
{\textsc{Tr}}\bigl( \mathbf{X}^{\transpose} \mathbf{A} \mathbf{Y}\bigr)
\bigr\rvert
\le
\|\mathbf{X}\|_{{\textnormal{\tiny{F}}}} \cdot \|\mathbf{A} \|_{2} \cdot \|\mathbf{Y}\|_{{\textnormal{\tiny{F}}}}.
\nonumber
\end{align}
\end{lemma}
\begin{proof}
We have
\begin{align}
\bigl\lvert
{\textsc{Tr}}\bigl( \mathbf{X}^{\transpose} \mathbf{A} \mathbf{Y}\bigr)
\bigr\rvert
\le
\|\mathbf{X}\|_{{\textnormal{\tiny{F}}}} \cdot \|\mathbf{A}\mathbf{Y}\|_{{\textnormal{\tiny{F}}}}
\le
\|\mathbf{X}\|_{{\textnormal{\tiny{F}}}} \cdot \|\mathbf{A} \|_{2} \cdot \|\mathbf{Y}\|_{{\textnormal{\tiny{F}}}},
\nonumber
\end{align}
with the first inequality following from Lemma~\ref{lemma:inner-prod-ub}
on $\lvert \langle \mathbf{X},\, \mathbf{A}\mathbf{Y}\rangle\rvert$ and the second from Lemma~\ref{lemma:frob-of-matrix-prod}.
\end{proof}
\begin{lemma}
\label{lemma:abs-trace-XAY-ub-orthogonal}
For any real
$m \times n$ matrix $\mathbf{A}$,
and pair of
$m \times k$ matrix $\mathbf{X}$ and $n \times k$ matrix $\mathbf{Y}$
such that $\mathbf{X}^{\transpose}\mathbf{X}=\mathbf{I}_{k}$ and $\mathbf{Y}^{\transpose}\mathbf{Y}=\mathbf{I}_{k}$
with $k \le \min\lbrace {m},\; {n}\rbrace$, the following holds:
\begin{align}
\bigl\lvert
{\textsc{Tr}}\bigl( \mathbf{X}^{\transpose} \mathbf{A} \mathbf{Y}\bigr)
\bigr\rvert
\le
\sqrt{k}\cdot
\bigl( \sum_{i=1}^{k} \sigma_{i}^{2}\bigl(\mathbf{A}\bigr) \bigr)^{1/2}.
\nonumber
\end{align}
\end{lemma}
\begin{proof}
By Lemma~\ref{lemma:inner-prod-ub},
\begin{align}
\lvert \langle \mathbf{X},\, \mathbf{A}\mathbf{Y}\rangle\rvert
&=
\bigl\lvert
{\textsc{Tr}}\bigl( \mathbf{X}^{\transpose} \mathbf{A} \mathbf{Y}\bigr)
\bigr\rvert
\nonumber\\&
\le
\|\mathbf{X}\|_{{\textnormal{\tiny{F}}}} \cdot \|\mathbf{A}\mathbf{Y}\|_{{\textnormal{\tiny{F}}}}
=
\sqrt{k} \cdot \|\mathbf{A}\mathbf{Y}\|_{{\textnormal{\tiny{F}}}}.
\nonumber
\end{align}
where the last inequality follows from the fact that $\|\mathbf{X}\|_{{\textnormal{\tiny{F}}}}^{2} = {\textsc{Tr}}\bigl(\mathbf{X}^{\transpose}\mathbf{X}\bigr) = {\textsc{Tr}}\bigl(\mathbf{I}_{k}\bigr) = k$.
Further, for any $\mathbf{Y}$ such that $\mathbf{Y}^{T}\mathbf{Y}= \mathbf{I}_{k}$,
\begin{align}
\|\mathbf{A} \mathbf{Y}\|_{{\textnormal{\tiny{F}}}}^{2}
\le
\max_{
\substack{
\widehat{\mathbf{Y}} \in \mathbb{R}^{n \times k}\\
\widehat{\mathbf{Y}}^{\transpose}\widehat{\mathbf{Y}} = \mathbf{I}_{k}
}
}
\|\mathbf{A} \widehat{\mathbf{Y}} \|_{{\textnormal{\tiny{F}}}}^{2}
=
\sum_{i=1}^{k} \sigma_{i}^{2}(\mathbf{A}).
\end{align}
Combining the two inequalities, the result follows.
\end{proof}
\newpage
\begin{lemma}
For any real
$m \times n$ matrix $\mathbf{A}$,
and any $k \le \min\lbrace {m},\; {n}\rbrace$,
\begin{align}
\max_{
\substack{
{\mathbf{Y}} \in \mathbb{R}^{n \times k}\\
{\mathbf{Y}}^{\transpose}{\mathbf{Y}} = \mathbf{I}_{k}
}
}
\|\mathbf{A} {\mathbf{Y}} \|_{{\textnormal{\tiny{F}}}}
=
\left(\sum_{i=1}^{k} \sigma_{i}^{2}(\mathbf{A}) \right)^{1/2}.
\nonumber
\end{align}
The above equality is realized when the $k$ columns of~$\mathbf{Y}$ coincide with the $k$ leading right singular vectors of~$\mathbf{A}$.
\end{lemma}
\begin{proof}
Let $\mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\transpose}$ be the singular value decomposition of $\mathbf{A}$;
$\mathbf{U}$ and $\mathbf{V}$ are $m \times m$ and $n \times n$ unitary matrices respectively, while $\Sigma$ is a diagonal matrix with $\Sigma_{jj} = \sigma_{j}$, the $j$th largest singular value of $\mathbf{A}$, $j=1, \hdots, d$, where $d {\triangleq} \min\lbrace {m}, {n} \rbrace$.
Due to the invariance of the Frobenius norm under unitary multiplication,
\begin{align}
\|\mathbf{A} \mathbf{Y}\|_{{\textnormal{\tiny{F}}}}^{2}
=
\|\mathbf{U}\mathbf{\Sigma}\mathbf{V}^{\transpose} \mathbf{Y}\|_{{\textnormal{\tiny{F}}}}^{2}
=
\|\mathbf{\Sigma}\mathbf{V}^{\transpose} \mathbf{Y}\|_{{\textnormal{\tiny{F}}}}^{2}.
\label{frob-norm-unitary}
\end{align}
Continuing from~\eqref{frob-norm-unitary},
\begin{align}
\|\mathbf{\Sigma}\mathbf{V}^{\transpose} \mathbf{Y}\|_{{\textnormal{\tiny{F}}}}^{2}
&=
{\textsc{Tr}}\bigl(\mathbf{Y}^{\transpose}\mathbf{V}\mathbf{\Sigma}^{2}\mathbf{V}^{\transpose} \mathbf{Y}\bigr)
\nonumber\\
&=
\sum_{i=1}^{k} \mathbf{y}_{i}^{\transpose}
\left(
\sum_{j=1}^{d} \sigma_{j}^{2} \cdot \mathbf{v}_{j} \mathbf{v}_{j}^{\transpose}
\right)
\mathbf{y}_{i}
\nonumber\\
&=
\sum_{j=1}^{d}
\sigma_{j}^{2} \cdot \sum_{i=1}^{k}
\left( \mathbf{v}_{j}^{\transpose} \mathbf{y}_{i}\right)^{2}.
\nonumber
\end{align}
Let $z_{j} {\triangleq} \sum_{i=1}^{k} \bigl( \mathbf{v}_{j}^{\transpose} \mathbf{y}_{i}\bigr)^{2}$, $j=1, \hdots, d$.
Note that each individual $z_{j}$ satisfies
\begin{align}
0 \le z_{j} {\triangleq} \sum_{i=1}^{k} \bigl( \mathbf{v}_{j}^{\transpose} \mathbf{y}_{i}\bigr)^{2}
\le
\|\mathbf{v}_{j}\|^{2} = 1,
\nonumber
\end{align}
where the last inequality follows from the fact that the columns of $\mathbf{Y}$ are orthonormal.
Further,
\begin{align}
\sum_{j=1}^{d} z_{j}
&=
\sum_{j=1}^{d}\sum_{i=1}^{k} \bigl( \mathbf{v}_{j}^{\transpose} \mathbf{y}_{i}\bigr)^{2}
=
\sum_{i=1}^{k}\sum_{j=1}^{d} \bigl( \mathbf{v}_{j}^{\transpose} \mathbf{y}_{i}\bigr)^{2}
\nonumber\\&=
\sum_{i=1}^{k}\|\mathbf{y}_{i}\|^{2} = k.
\nonumber
\end{align}
Combining the above, we conclude that
\begin{align}
\|\mathbf{A}\mathbf{Y}\|_{{\textnormal{\tiny{F}}}}^{2}
=
\sum_{j=1}^{d}
\sigma_{j}^{2} \cdot z_{j}
\le \sigma_{1}^{2} + \hdots + \sigma_{k}^{2}.
\label{frob-prod-ub}
\end{align}
Finally, it is straightforward to verify that if $\mathbf{y}_{i} = \mathbf{v}_{i}$, $i=1, \hdots, k$, then~\eqref{frob-prod-ub} holds with equality.
\end{proof}
\subsection{Breast Cancer Dataset}
The breast cancer dataset~\cite{chin2006genomic} consists of gene expression and DNA copy number measurements on a set of $89$ tissue samples.
Among others, it contains
a $89 \times 2149$ matrix (DNA) with CGH spots for each sample
and a $89 \times 19672$ matrix (RNA) of genes,
along with information for the chromosomal locations of each CGH spot and each gene.
As described in~\cite{witten2009penalized},
this dataset can be used to perform integrative analysis of gene expression and DNA copy number data,
and in particular to identify sets of genes that have expression that is correlated with a set of chromosomal gains or losses.
\begin{figure}[t!]
\centering
\includegraphics[width=0.99\linewidth]{Figures/breast_data_compariston.pdf}
\vspace{1em}
{\small
\rowcolors{2}{white}{black!05!white}
\begin{tabular}{lccc}
\toprule
& Avg Exec. Time & Configuration \\
\cmidrule{1-1} \cmidrule{2-2} \cmidrule{3-3} \cmidrule{4-4}
PMD & $\sim 44$ seconds & $10$ rand. restarts\\
\textnormal{\textrm{SpanCCA}}\xspace & $\sim 24$ seconds & $T=10^4$, $r=3$.\\
\bottomrule
\end{tabular}
}
\caption{%
Comparison of \textnormal{\textrm{SpanCCA}}\xspace and the PMD algorithm~\cite{witten2009penalized}.
We configure PMD with $\ell_{1}$-norm thresholds $c_{1}=c \cdot \sqrt{m}$ and $c_{2}=c \cdot \sqrt{n}$,
and consider various values of the constant $c \in (0,1)$.
For each~$c$, we run PMD $10$ times, select the canonical vectors $\mathbf{x}$, $\mathbf{y}$ that achieve the highest objective value and count their nonzero entries (depicted as percentage of the corresponding dimension).
Finally, we run our \textnormal{\textrm{SpanCCA}}\xspace algorithm with ${T=10^4}$ and ${r=3}$, using the latter as target sparsities,
and compare the objective values achieved by the two methods.
Execution times remain approximately the same for all target sparsity values (equiv. all~$c$).
}
\label{fig:breast_comparison}
\end{figure}
We run our algorithm on the breast cancer dataset and compare the output with the PMD algorithm of~\cite{witten2009penalized};
PMD is regarded as state of the art by practitioners and has been used --in its original form or slightly modified-- in several neuroscience and biomedical applications; see also Section \ref{sec:related_work}.
The input to both algorithms is the $m \times n$ matrix $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}=\mathbf{X}^{\transpose}\mathbf{Y}$ (${m=2149}$, ${n=19672}$),
where $\mathbf{X}$ and $\mathbf{Y}$ are obtained from the aforementioned DNA and RNA matrices upon feature standardization.
Recall that PMD is an iterative, alternating optimization scheme,
where the sparsity of the extracted components $\mathbf{x}$ and $\mathbf{y}$ is implicitly controlled by enforcing upper bounds $c_{1}$ and $c_{2}$ on their $\ell_{1}$ norm, respectively, with $1\le c_{1} \le \sqrt{m}$ and $ 1\le c_{2} \le \sqrt{n}$.
Here, for simplicity, we set ${c_{1} = c \sqrt{m}}$ and ${c_{2} = c \sqrt{n}}$ and consider multiple values of the constant $c$ in $(0, 1)$.
Note that under this configuration, for any given value of~$c$, we expect that the extracted components will be approximately equally sparse, relatively to their dimension.
For each~$c$, we first run the PMD algorithm $10$ times with random initializations,
determine the pair of components $\mathbf{x}$ and $\mathbf{y}$ that achieves the highest objective value, and count the number of nonzero entries of both components as a percentage of their corresponding dimension.
Subsequently, we run \textnormal{\textrm{SpanCCA}}\xspace (Alg.~\ref{algo:cca}) with parameters ${T=10^4}$, ${r=3}$, and target sparsity equal to that of the former PMD output.
Recall that our algorithm administers precise control on the number of nonzero entries of the extracted components.
Figure~\ref{fig:breast_comparison} depicts the objective value achieved by the two algorithms,
as well as the corresponding sparsity level of the extracted components.
SpanCCA achieves a higher objective value in all cases.
Finally, note that under the above configuration,
both algorithms run for a few seconds per target sparsity,
with \textnormal{\textrm{SpanCCA}}\xspace running approximately half the time of PMD.
\subsection{Brain Imaging Dataset}
\label{sec:brain_imaging_dataset}
We analyzed functional statistical maps and behavioral variables from $497$ subjects available from the Human Connectome Project (HCP)~\cite{van2013wu}.
The HCP consists of high-quality imaging and behavioral data, collected from a large sample of healthy adult subjects, motivated by the goal of advancing knowledge between human brain function and its association to behavior.
We apply our algorithm to investigate the shared co-variation between patterns of brain activity as measured by the experimental tasks, and behavioral variables.
We selected the same subset of behavioral variables examined by~\cite{smith2015positive}, which include scores from psychological tests, physiological measurements, and self reported behavior questionnaires
($\mathbf{Y}$ dataset with dimensions $497 \times 38$).
For each subject, we collected statistical maps corresponding to ``$n$-back'' task.
These statistical maps summarize the activation of each voxel in response to the experimental manipulation.
In the ``$n$-back'' task, designed to measure working memory, items are presented one at a time and subjects identify each item that repeats relative to the item that occurred $n$ items before.
Further details on all tasks and variables are available in the HCP documentation~\cite{van2013wu}.
We used the pre-computed 2back - 0back statistical contrast maps provided by the HCP.
Standard preprocessing included motion correction, image registration to the MNI template (for comparison across subjects), and general linear model analysis, resulting in $91 \times 109 \times 91$ voxels.
Voxels are then resampled to $61 \times 73 \times 61$ using the \texttt{nilearn} python package\footnote{\url{http://nilearn.github.io/}} and applying standard brain masks, resulting in $65598$ voxels after masking non-grey matter regions.
($\mathbf{X}$ dataset with dimensions $497 \times 65598$).
We apply our \textnormal{\textrm{SpanCCA}}\xspace algorithm on the HCP data with arbitrarily selected parameters ${\mathrm{T}}=10^6$ and $r=5$.
We set the target sparsity at $15\%$ for each canonical vector.
Figure~\ref{fig:brain_behavior} depicts the brain regions and the behavioral factors corresponding to the nonzero weights of the extracted canonical pair.
The map identifies a set of fronto-parietal regions known to be involved in executive function and working memory, which are the major functions isolated by the 2 back - 0-back contrast. In addition, it identifies deactivation in the default mode areas (medial prefrontal and parietal), which is also associated with engagement of difficult cognitive functions. The behavioral variables associated with activation of this network are all related to various aspects of intelligence; the Penn Matrix Reasoning Test (\texttt{PMAT24}, a measure of fluid intelligence), picture vocabulary (\texttt{PicVocab}, a measure of language comprehension), and reading ability (\texttt{ReadEng}).
\begin{figure*}[t!]
\centerline{%
\setlength{\fboxsep}{0pt}%
\setlength{\fboxrule}{0pt}%
\begin{minipage}[td]{0.72\linewidth}
\raisebox{-0.5\height}{%
\fbox{%
\includegraphics[width=1\textwidth]{Figures/2BK-0BK_stat_factor0_copy.pdf}%
}%
}%
\end{minipage}%
\definecolor{dark-gray}{gray}{0.30}
\begin{minipage}[t]{0.28\linewidth}%
\centering
\raisebox{+0.1\height}{%
\color{dark-gray}
\fbox{\begin{tabular}{ll}%
\toprule
\multicolumn{2}{l}{Behavioral Factor \& Weight}\\
\midrule
\texttt{PMAT24\_A\_CR} &$0.487$\\
\texttt{PicVocab\_AgeAdj} &$0.448$\\
\texttt{ReadEng\_AgeAdj} &$0.440$\\
\texttt{PicVocab\_Unadj} &$0.433$\\
\texttt{ReadEng\_Unadj} &$0.426$\\
\bottomrule
\end{tabular}}}
\end{minipage}%
\caption{%
Brain regions and behavioral factors selected by the sparse left and right canonical vectors extracted by our \textnormal{\textrm{SpanCCA}}\xspace algorithm.
Target sparsity is set at $15\%$ for each canonical vector
and \textnormal{\textrm{SpanCCA}}\xspace is configured to run for ${\mathrm{T}}=10^6$ samples operating on a rank $r=5$ approximation of the input data.
The map identifies a set of fronto-parietal regions known to be involved in executive function and working memory
and deactivation in the default mode areas (medial prefrontal and parietal), which is also associated with engagement of difficult cognitive functions. The behavioral variables identified to be positively correlated with the activation of this network are all related to various aspects of intelligence.
}
\label{fig:brain_behavior}
\end{figure*}
\paragraph{Parallelization}
To speed up execution,
our prototypical \texttt{Python} implementation of \textnormal{\textrm{SpanCCA}}\xspace exploits the \texttt{multiprocessing} module:
$\mathrm{N}$ independent worker processes are spawned, and each one independently performs $\mathrm{T}/\mathrm{N}$ rounds of the main iteration of Alg.~\ref{algo:cca} returning a single canonical vector pair.
The main process collects and compares the candidate pairs to determine the final output.
To demonstrate the parallelizability of our algorith, we run \textnormal{\textrm{SpanCCA}}\xspace for the aforementioned task on the brain imaging data for various values of the number $\mathrm{N}$ of workers on a single server with $36$ physical processing cores\footnote{Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz} and approximately $250\mathrm{Gb}$ of main memory.
In Figure~\ref{fig:multicore_scaling} (top panel), we plot the run time with respect to the number of workers used.
The bottom panel depicts the achieved speedup factor:
using the execution time on~$5$ worker processes as a reference value,
the speedup factor is the ratio of the execution time on $5$ processes over that on $\mathrm{N}$.
As expected, the algorithm achieved a speedup factor that grows almost linearly in the number of available processors.
\begin{figure}[t!]
\centering
\includegraphics[width=0.97\linewidth]
{Figures/spancca_multicore_scaling_T100000_m65598_n38_MC20_exectime.pdf}
\includegraphics[width=0.97\linewidth]{Figures/spancca_multicore_scaling_T100000_m65598_n38_MC20_speedup.pdf}
\caption{%
Speedup factors and corresponding total execution time, achieved by the prototypical parallel implementation of SpannCCA (Alg.~\ref{algo:cca}) as a function of the number of worker processes or equivalently the number of processors used.
Depicted values are medians over $20$ executions, each with ${T=10^5}$ and $r=5$, on the $65598 \times 38$ example discussed in section~\ref{sec:brain_imaging_dataset}.
A speedup factory approximately linear in the number of workers is achieved.
}
\label{fig:multicore_scaling}
\end{figure}
\subsection{Our contributions}
We present a novel and efficient combinatorial algorithm for sparse diagonal CCA in \eqref{cca:def}.
The main idea is to reduce the exponentially large search space of candidate supports of the canonical vectors, by exploring a low-dimensional principal subspace of the input data.
Our algorithm runs in polynomial time --in fact linear-- in the dimension of the input.
It administers precise control over the sparsity of the extracted canonical vectors and can extract components for multiple sparsity values on a single run.
It is simple and trivially parallelizable;
we empirically demonstrate that it achieves an almost linear speedup factor in the number of available processing units.
The algorithm is accompanied with theoretical data-dependent global approximation guarantees with respect to the CCA objective~\eqref{cca:def};
this is the first approach with this kind of global guarantees.
The latter depend on the rank $r$ of the low-dimensional space and the spectral decay of the input matrix $\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
The main weakness is an exponential dependence of the computational complexity on the accuracy parameter $r$.
In practice, however, disregarding the theoretical approximation guarantees, our algorithm can be executed for any allowable time window.
Finally, we note that our approach is similar to that of~\cite{asteris2014nonnegative} for sparse PCA.
The latter has a similar formulation with~\eqref{cca:def} but is restricted to a positive semidefinite argument~$\mathbf{\Sigma_{\mathsmaller{\mathrm{X}\mathrm{Y}}}}$.
Our main technical contribution is extending those algorithmic ideas and developing theoretical approximation guarantees for the bilinear maximization~\eqref{cca:def}, where the input matrix can be arbitrary.
|
\section{Introduction}
Magnetic-dipole transitions between the ground state fine structure components in hollow shell lanthanides are strongly shielded from
external electric fields by the closed outer $5s^2$ and $6s^2$ shells. In the solid state these well resolved transitions protected from
intra-crystal electric fields are widely used in various active media doped by Er$^{3+}$, Tm$^{3+}$, and other ions lasing in the
near-infrared and infrared spectral ranges~\cite{Zharkov1975, Barnes1993}. Such shielding can also facilitate the use of inner-shell
transitions in optical frequency metrology due to low sensitivity to external electric fields and collisions~\cite{Boettger2001}.
In 1983 Alexandrov et. al.~\cite{ref:AlexandrovEnglish1983} showed that the collisional broadening of the inner-shell magnetic dipole
transition in the Tm atom $[\text{Xe}]4f^{13}6s^2 (J=7/2) \rightarrow [\text{Xe}]4f^{13}6s^2 (J=5/2)$, where $J$ is the total electronic angular momentum, at the wavelength of 1.14\,$\mu$m in He buffer
gas is suppressed by at least 500 times compared to the outer shell transitions. Note, that in the early era of optical atomic
clocks the dominating systematic uncertainty was the collisional shift in a cloud of laser cooled atoms~\cite{Wilpers2002, Ido2005}.
One could expect better performance using inner-shell transitions in lanthanides, but this study was hampered by difficulties with
their laser cooling. It was shown later that for Tm-He collisions shielding strongly reduces the spin relaxation~\cite{Hancox2004}
but it does not reduce the spin relaxation rate in Tm-Tm collisions due to the anisotropic nature of the magnetic dipole-dipole
interaction~\cite{Connolly2010}.
The problem with atom-atom collisions in optical clocks was solved after the invention of an optical lattice clock~\cite{Katori2003,
Takamoto2003} which resulted in rapid progress of accuracy and stability over the last decade~\cite{Ludlow2015}. Today, lattice clocks
based on Sr~\cite{Bloom2014} and Yb~\cite{Hinkley2013} demonstrate unprecedented fractional frequency instabilities in the low
$10^{-18}$ range. One of the important limiting factors is the shift caused by black-body radiation (BBR)~\cite{Beloy2014,
Safronova2013, Ushijima2015}.
Optical clock community continues an intensive search for alternative candidates aiming for lower sensitivity to BBR and other shifts,
simplicity of manipulation and better accuracy~\cite{Kulosa2015, McFerran2014}. Since hollow-shell lanthanides are expected to show
small differential static polarizabilities of the states with different configurations of the $4f$ electrons, one expects small BBR
shift of the inner-shell magnetic dipole transitions. Taking into account large natural lifetimes of the clock levels, these
transitions can be successfully used in optical lattice clocks. Recent progress in laser cooling of Er~\cite{Aikawa2012},
Dy~\cite{ref:DyBEC:Lev:PRL}, and Tm~\cite{Sukachev2010, Sukachev2014} and frequency stabilized laser systems~\cite{Alnis2008,
Kessler2012} open the way for experimental implementation of these ideas.
Similar to other lanthanides, laser cooling of Tm is achieved in two stages.
The first cooling stage is done at the strong 410.6\,nm
transition which routinely allows reaching subdoppler temperature of 80\,$\mu$K in a cloud of $2\times10^6$ atoms~\cite{Sukachev2010}.
The second cooling stage at the weak 530.7\,nm transition results in the Doppler-limited temperature of 9\,$\mu$K~\cite{Vishnyakova2016}.
This temperature is low enough to load atoms in a shallow optical trap or a lattice as was demonstrated in \cite{Sukachev2014} using
532\,nm laser radiation.
Relevant Tm levels are shown in Fig.~\ref{fig_levels}.
Further cooling of atoms is possible by the optimization of the cooling sequence~\cite{Frish-PhD-2014} or by evaporative cooling~\cite{Ketterle1996}. These experiments stimulated further study of the inner shell transition $[\text{Xe}]4f^{13}6s^2 (J=7/2, \,F=4) \rightarrow [\text{Xe}]4f^{13}6s^2 (J=5/2, \,F=3)$, where $F$ is the total atomic angular momentum, for its application in optical lattice clocks.
In this article, the level $[\text{Xe}]4f^{13}6s^2 (J=7/2)$ will be referred to as the ``lower clock level'' while the level $[\text{Xe}]4f^{13}6s^2 (J=5/2)$ as the ``upper clock level''.
In the next sections, we analyze effects which may impact the performance of such clocks.
First, the only stable isotope $^{169}$Tm is a boson and the clock transition is subject to collisional shifts. The related scattering length depends on the poorly known Tm-Tm potential at small distances and is very sensitive to the calculation uncertainty of the long-range potentials~\cite{Dalibard1998,Gribakin1993}. This difficulty can be overcome if Tm atoms are loaded in a 2D-optical lattice with a small filling factor canceling Tm-Tm collisions.
Second, to avoid intensity-dependent effects we calculated the dynamic polarizabilities of the upper and lower clock levels and defined a candidate for the ``magic'' wavelength (sec.~\ref{sec:polarizabilities}).
Third, the large ground-state dipole moment of Tm atoms induces a frequency shift due to
magnetic dipole-dipole interaction. Preparing Tm atoms in the $|m = 0\rangle$ (here $m$ is a magnetic quantum number) state cancels this shift but magnetic relaxation still limits the interrogation time of the clock transition and should be
taken into account (sec.~\ref{sec:magnetic}).
In sec.~\ref{sec:budget} we present the error budget of the proposed Tm optical clock.
\begin{figure}
\includegraphics[width=1\linewidth]{1.pdf}
\caption{\label{fig_levels} Relevant energy levels of $^{169}$Tm. The strong transition at 410.6\,nm is used for the first-stage laser
cooling and detecting the ground state populations and the weak transition at 530.7\,nm is used for the second-stage cooling. The
proposed clock transition $4f^{13}6s^2 (J=7/2, F=4) \rightarrow 4f^{13}6s^2 (J=5/2, F=3)$ is at the wavelength of 1.14\,um.}
\end{figure}
In the experimental part (sec.~\ref{sec:experiment}), we demonstrate direct excitation of the clock transition at 1.14\,$\mu$m and measure the lifetime of the upper clock level in a 1D optical lattice formed by 532\,nm laser
radiation~\cite{Golovizin2015}. Also, we experimentally evaluate the dynamic polarizability of the Tm ground state at 532\,nm by
excitation of parametric resonances in the optical dipole trap.
\section{Polarizabilities}
\label{sec:polarizabilities}
\label{sec:polarizability:theory}
To find the magic wavelength and to estimate the BBR and the van der Waals shifts, one should know the energy shifts $\varDelta E$ of the
clock states in an external monochromatic electric field $\vec{E} = \sfrac{1}{2} \vec{\mathcal{E}} e^{-i\omega
t}+\text{c.c.} $ at the angular frequency $\omega$
\begin{equation}
\varDelta E(\omega) = -\frac{\alpha(\omega)}{4} |\mathcal{E}|^2-\frac{\gamma(\omega)}{64} |\mathcal{E}|^4 +...,
\label{eq1}
\end{equation}
where $\alpha(\omega)$ is the dynamic polarizability and $\gamma(\omega)$ is the hyperpolarizability, both depending on $m$
and the polarization of the field. To our knowledge, there are only a few publications where the polarizability of Tm levels was
analyzed. In \cite{RETensorPol, TmTensorPol} the authors measured the static tensor polarizability, while a theoretical calculation of
static polarizabilities without accounting for a fine-structure interaction is presented in~\cite{Chu2007}. In this section we will
calculate polarizabilities of the clock states.
To suppress the site-dependent frequency shift from varying light polarization in the lattice, we suggest loading Tm atoms into a 2D
optical lattice formed by 4 laser beams with the same linear polarization as shown in Fig.\,\ref{latticeGeometry}. This guarantees that the trapping light polarization is the same for all lattice sites. Further in this paper, we consider only the transition
$|J=7/2,F=4, m=0\rangle \rightarrow |J=5/2,F=3,m=0\rangle$ which is free from the frequency shifts induced by the magnetic dipole-dipole interaction (see sec.\,\ref{sec:magnetic}). Since both levels have $m=0$, the contribution from the vector polarizability for this transition also vanishes and the total
polarizability $\alpha$ can be separated into the scalar $\alpha^s$ and the tensor $\alpha^t$ parts~\cite{Lepers2014} as follows:
\begin{equation}
\begin{split}
\alpha_{JFm}(\omega) &= \alpha^s_{JF}(\omega) + \alpha^t_{JF}(\omega) \frac{3m^2-F(F+1)}{F(2F-1)}\,,\\
\alpha^s_{JF}(\omega) &=\frac{1}{2F+1}\sum\limits_{m=-F}^{m=F} \alpha_{JFm}(\omega)\,,\\
\alpha^t_{JF}(\omega) &= \alpha_{JF,m=F}(\omega)-\alpha^s_{JF}(\omega).
\end{split}
\label{eq2}
\end{equation}
For consistency with other papers, we will calculate the polarizabilities in atomic units (a.u.); 1\,a.u. =
$4\pi \epsilon_0 a_0^3$ = $1.65\times 10^{-41}$\,J/(V/m)$^2$,
where $a_0$ is the Bohr radius and $\epsilon_0$ is the vacuum permittivity (for conversion to another units, see~\cite{Mitroy2010}).
\begin{figure}
\includegraphics[width=1\linewidth]{2.pdf}
\caption{\label{latticeGeometry}
Geometry of a 2D optical lattice. The lattice is formed by 4 horizontal laser beams (lattice, red) with linear vertical
polarization. A uniform external magnetic field $B_0$ is applied along the vertical axis. An interrogating laser beam (clock, blue)
lies in the optical lattice plane (horizontal) to eliminate frequency shifts by the Doppler-effect and photon recoil. The $B$-component of the
interrogating light should be vertical to excite the $|m=0 \rangle \rightarrow |m=0\rangle$ clock transition (the clock transition is of a magnetic dipole type).
}
\end{figure}
\subsection{Discrete spectrum}
The contribution of a discrete spectrum is given by~\cite{Lepers2014, Angel1968}
\begin{multline}\label{eq:alphaFM}
\alpha_{Fm}(\omega)=\frac{3}{2} \frac{c^3\hbar^4}{a_0^3} \sideset{}{'} \sum_{F'}\frac{2F_u+1}{(E_{F'}-E_F)^2}
\left(
\begin{matrix}
F_u & 1 & F_d\\
-m & 0 & m
\end{matrix}
\right)^2\\
\times \frac{A_{F_u \to F_d}}{(E_{F'}-E_F)^2-(\hbar\omega)^2},
\end{multline}
where $c$ is the speed of light, $\hbar$ is the reduced Planck’s constant, and $E_F$ and $E_{F'}$ are energies of levels $|F\rangle$
and $|F'\rangle$, respectively. The summation is over all levels $F'$. For each term, $F_u = F'$ and $F_d = F$ if $E_{F'} > E_{F}$
and vice versa. $A_{F_u \to F_d}$ is a transition probability (spontaneous decay rate) from $|F_u\rangle$ to $|F_d\rangle$.
Assuming $JI$-coupling between the total electron momentum $J$ and the nuclear spin $I$, the scalar polarizability is independent of
$F$~\cite{Lepers2014, Angel1968}:
\begin{multline}
\alpha^s_{JF}(\omega)=\alpha^s_{J}(\omega)=\frac{1}{2J+1}\sum\limits_{m_J=-J}^{m_J=J} \alpha_{Jm_J}(\omega)=\\
\frac{1}{2} \frac{c^3}{a_0^3} \sideset{}{'}\sum_{J'}\frac{2J_u+1}{2J_d+1}\frac{1}{(\omega_{J'J})^2} \frac{A_{J_u \to
J_d}}{(\omega_{J'J})^2-\omega^2},
\end{multline}
where $\omega_{J'J}=(E_{J'}-E_J)/\hbar$.
The tensor polarizability equals
\begin{multline}\label{eq:alphaJFt}
\alpha^t_{JF}(\omega) = \alpha^t_{J}(\omega) \times (-1)^{I+J+F} \left\{
\begin{matrix}
F & J & I\\
J & F & 2
\end{matrix}
\right\} \\
\times \sqrt{\frac{F(2F-1)(2F+1)(2J+3)(2J+1)(J+1)}{(2F+3)(F+1)(2J-1)J}},
\end{multline}
where
\begin{multline}
\alpha^t_{J}(\omega) = \frac{3c^3}{a_0^3} \sideset{}{'} \sum\limits_{J'} \frac{2J_u+1}{\omega_{J'J}^2} \frac{A_{J_u\rightarrow
J_d}}{\omega_{J'J}^2-\omega^2} (-1)^{J+J'} \\
\times \left\{
\begin{matrix}
1 & 1 & 2\\
J_d & J_d & J_u
\end{matrix}
\right\} \sqrt{\frac{5J(2J-1)}{6(J+1)(2J+1)(2J+3)}}.
\end{multline}
Note that
$\alpha^t_{\sfrac{7}{2},\,4} = \alpha^t_{\sfrac{7}{2}}$ and $\alpha^t_{\sfrac{5}{2},\,3} = \alpha^t_{\sfrac{5}{2}}$.
As an input for calculation of $\alpha$ one should have transition probabilities $A_{J'\rightarrow J}$ from the level of interest
to all others.
Though many transition wavelengths in the spectral range from 250\,nm to 807\,nm and their probabilities were measured
in~\cite{TmTransitions, Anderson1996} by Fourier-transform spectroscopy and time-resolved laser-induced fluorescence, there is still
a number of transitions in the UV, visible, and IR spectral ranges which are essential for calculation and have unknown probabilities. We used the
numerical package COWAN~\cite{Cowan1981} to calculate transition wavelengths and probabilities in the spectral range from 250\,nm to
1200\,nm (see Appendix A).
The most self-consistent approach for calculation of the differential static
polarizability of the clock levels is to use only the numerically calculated wavelengths and probabilities.
A slight modification of this approach is used for calculation of the magic wavelengths
(sec.~\ref{sec:mw}).
As expected from general considerations concerning the inner shell transitions, the static scalar polarizabilities for the clock levels are nearly equal. Our calculation shows that they differ by less than $0.1$\,a.u. and are equal to 138\,a.u.
Note, that the calculated static tensor polarizability of $-2.7$\,a.u. for the lower $|J=7/2\rangle$ clock level is in good agreement with
the known experimental value of $-2.7(2)$\,a.u.~\cite{RETensorPol}. For the upper $|J=5/2\rangle$ clock level our calculations give $-2.3$ atomic units for the tensor static polarizability.
\subsection{Continuous spectrum}
To determine a contribution of the continuous spectrum (ignoring hyperfine interaction) to the polarizability we used the formula~\cite{Veseth1992}:
\begin{equation}
\alpha^s_{\text{cont}}(\omega) = \frac{c}{2\pi^2}\int^{\infty}_{\omega_I}
\frac{\sigma(\omega')d\omega'}{(\omega'-\omega_n)^2-\omega^2},
\end{equation}
where $\omega_I$ is the photoionization limit and $\sigma(\omega)$ is the photoionization cross section of the energy level.
The ionization cross-section was numerically calculated using the package FAC ~\cite{FAC2011} and the results are shown in the upper panel of
Fig.\,\ref{fig_ion}. Using these results, we evaluated the polarizabilities $\alpha^s_\text{cont}(\omega)$ for the clock levels resulting from transitions to continuous spectrum (Fig.\,\ref{fig_ion}, lower panel).
The contributions $\alpha^s_\text{cont}(\omega)$ are small compared to the contribution from the discrete spectrum and differ only
by 2\,a.u. for the two fine structure components $|J=7/2\rangle$ and $|J=5/2\rangle$. This means that the transitions to continuum basically do not influence positions of magic wavelengths. We also assume that the corresponding contribution to the tensor polarizability is even
smaller and will neglect it in further analysis. Since the discrete spectrum gives the equal static scalar polarizabilities for the clock levels, we expect the continuous spectrum to contribute to polarizabilities of the both levels equally as well.
Thus, a rough estimation of the error of calculated contribution of the continuous spectrum to the differential polarizability is about the difference between $\alpha^s_\text{cont}(0)$ for the clock levels, i.e., 2\,a.u.
Unfortunately, we do not know of any experimental data on the photoionization
cross sections for Tm atoms and therefore can't rigorously estimate the error.
\begin{figure}
\includegraphics[width=1\linewidth]{3.pdf}
\caption{\label{fig_ion} Upper panel: The photoionization cross sections for $|J=7/2\rangle$ (solid blue) and $|J=5/2\rangle$ (dashed red) clock levels. Peaks
around 28 and 35\,eV correspond to strong resonance enhancement~\cite{ref:Whitfield2008}.
Lower panel: Contribution to the scalar polarizability $\alpha^s_\text{cont}(\omega)$ from the continuous spectra. The hyperfine interaction is not taken into account.}
\end{figure}
\subsection{Magic wavelength}
\label{sec:mw}
The magic wavelengths for optical traps providing the vanishing total light shift of the clock transition (\ref{eq1}) are widely used in
optical clocks~\cite{Katori2003}. To determine the magic wavelengths one should search for the crossing points of the dynamic
polarizabilities for the upper and lower clock levels (neglecting the contribution from the hyperpolarizability in the first
approximation). Position of the magic wavelengths strongly depend on energies and probabilities of the resonances in the atom. In
general, we can not use only results of our calculations because of insufficient accuracy provided by the COWAN package (see Appendix A).
To solve this problem first we tried a ``combined'' approach. Calculated transitions were assigned with experimental ones which can
be done unambiguously for wavelengths $\lambda>500$\,nm. It turns to be impossible for the shorter wavelengths ($\lambda<500$\,nm) due to
higher density of transitions. Then we combined the calculated spectrum for $\lambda<500$\,nm, the available experimental data for
$\lambda>500$\,nm, and calculated probabilities for known transitions but with not measured probabilities. After detailed study we
concluded that it is a questionable approach because the calculated and experimentally measured transition probabilities sometimes
differ by an order of magnitude (see Fig.\,\ref{fig:COWAN}, lower panel). This difference
impacts the calculated polarizabilities in a wide spectral range impeding reliable prediction of the magic wavelengths.
To our opinion, a more reliable approach bases on maximal use of calculation results: We took the calculated spectrum and substituted
the predicted wavelengths with correct ones known from the experiment for all transitions with $\lambda>500$\,nm. As for the
probabilities, we used the calculated ones except the case when the probability is smaller than $10^5$\,s$^{-1}$. This method give
reliable results for the magic wavelengths in the near IR region with low density of strong transitions.
Selected approach predicts the reliable
candidate for the magic wavelength at 807\,nm with an attractive lattice potential (Fig.~\ref{fig:someMW}).
Its presence is caused by
the weak transition from the $|J=5/2\rangle$ clock level $ 4f^{13}(^2F^o)6s^2 (J=5/2)$ with the energy $8771.24\,\text{cm}^{-1}$ to another
level $4f^{12}(^3F_4)5d_{3/2}6s^2 (J=5/2)$, with the energy $21161.4\,\text{cm}^{-1}$ at 807.1\,nm ~\cite{TmTransitions}. At the same
time, there are no allowed transition from the $|J=7/2\rangle$ clock level in the vicinity of 807\,nm.
Taking into account the uncertainty in the contribution of the continuous spectra to the evaluated differential polarizability of $\pm
1$\,a.u., the proposed magic wavelength should be blue-detuned from the transition 807.1\,nm by 0.1\,nm --- 1\,nm.
The figure-of-merit for an optical lattice comes from its depth, the off-resonant scattering rate and the magnetic dipole-dipole
relaxation rate.
The optical lattice depth in kelvin is given by:
\begin{equation}
U\,[\text{K}] = \alpha[\text{a.u.}]\frac{2\pi a_0^3}{c\,k_B}I\,[\text{W/m$^2$}],
\end{equation}
where $I$ is the field intensity in lattice anti-nodes given in W/m$^2$ and $k_B$ is the Boltzmann constant.
The spontaneous decay following the off-resonant excitation by the lattice field perturbs the coherence of the clock levels and should be taken into account. The off-resonance scattering rate for the transition $|m=0\rangle \rightarrow |m=0\rangle$ can be estimated as~\cite{Lepers2014}:
\begin{multline}
\varGamma(\omega)_{0\rightarrow 0} = I \sideset{}{'} \sum \limits_{F'} \frac{\omega_{F'F}^2+\omega^2}{\left[\omega_{F'F}^2-\omega^2 \right]^2}
\frac{3\pi c^2\,A_{F' \rightarrow F}}{\hbar \omega_{F'F}^3} \\
\times
\left(
\begin{matrix}
F_u & 1 & F_d\\
0 & 0 & 0
\end{matrix}
\right)^2 (2F_u+1) \varGamma_{F_u},
\label{eq:alpha-scal-im}
\end{multline}
where $\varGamma_{F_u}$ is the inverse lifetime of $|F_u \rangle$ level.
\begin{figure}
\includegraphics[width=1\linewidth]{4.pdf}
\caption{The magic wavelengths for the 1.14\,$\mu$m clock transition in Tm atom ($m=0\rightarrow m=0$) around 807\,nm
calculated for the linear vertical polarization of the trapping light (see Fig.\,\ref{latticeGeometry}). Blue solid curve is the polarizability of the lower
$|J=7/2,F=4,m=0\rangle$ clock state, red dashed curve is the polarizability of the upper $|J=5/2,F=3,m=0\rangle$ clock state. The dashed
blue lines show the anticipated uncertainty in the calculation of the differential polarizability which may impact the position of
the magic wavelength.
}
\label{fig:someMW}
\end{figure}
The optical lattice at 807\,nm can be formed by a Ti:sapphire laser beam.
With 0.5\,W output power focused in the beam
waist of 50\,$\mu$m (radius at $1/e^2$ intensity level) corresponding to $I=50$\,kW/cm$^2$ in the retro-reflected configuration, one expects the trap depth of 20\,$\mu$K.
This is enough to capture Tm atoms from a narrow-line MOT. Even for the smallest expected detuning from the 807.1\,nm resonance of
0.1\,nm, the off-resonant scattering rate is less than $0.1$\,s$^{-1}$.
\subsection{Hyperpolarizability}
\label{sec:hyper}
The magic wavelength depends not only on the differential polarizability of the clock states, but also on the differential
hyperpolarizability (\ref{eq1}) and light intensity $I$. The scalar hyperpolarizability $\gamma(\omega)$ is given by~\cite{Bishop1994} :
\begin{equation}
\label{eq:hyper1}
\gamma(\omega) = \frac{1}{4} \left( \gamma^+(\omega) + \gamma^-(\omega) \right),
\end{equation}
where
\begin{multline}
\gamma^{+}=\frac{4}{\hbar^3} \sideset{}{'} \sum_{m,k,n}(D_z)_{gm}(D_z)_{mk}(D_z)_{kn}(D_z)_{ng}\\
\times\Bigg(\frac{4\omega_{mg}\omega_{ng}}{\omega_{kg}\left(\omega^2-\omega_{mg}^2\right)\left(\omega^2-\omega_{ng}^2\right)}\\
+\frac{1}{(\omega_{mg}-\omega)(\omega_{kg}-2\omega)(\omega_{ng}-\omega)}\\
+\frac{1}{(\omega_{mg}+\omega)(\omega_{kg}+2\omega)(\omega_{ng}+\omega)}\Bigg)
\label{eqXp2}
\end{multline}
and
\begin{multline}
\gamma^{-}=\frac{8}{\hbar^3} \sideset{}{'} \sum_{m,n}\left|(D_z)_{mg}\right|^2\left|(D_z)_{ng}\right|^2 \\
\times \frac{\omega_{mg}\left(\omega^2+3\omega_{ng}^2\right)}{\left(\omega^2-\omega_{mg}^2\right)\left(\omega^2-\omega_{ng}^2\right)^2}\,,
\end{multline}
where $(D_z)_{i,j}$ is a matrix element of the $z$-projection of the dipole moment between levels $i$ and $j$.
For calculation of hyperpolarizability we used the transition matrix elements, their signs, and the transition wavelengths obtained by the COWAN package for all transitions except the 807.1\,nm one. For this transition, we used the experimentally measured wavelength and probability; the sign of the transition matrix element was taken from the numerical calculations.
This exception is done to improve accuracy of the magic wavelength
prediction.
The light shifts for the clock levels $|J=7/2\rangle$ and $|J=5/2 \rangle$ coming from hyperpolarizability in the optical lattice at $\lambda= 807$\,nm and $I = 50$\,kW/cm$^2$ is shown in Fig.\,\ref{fig:hyper}.
As follows from the previous section, the magic wavelength is blue detuned from the 807.1\,nm resonance by more than
0.1\,nm, which makes the hyperpolarizability shift to be less than $0.5$\,Hz. Corresponding correction to the magic wavelength is
negligible. Still, hyperpolarizability contributes to the clock frequency uncertainty which is discussed later in
sec.\,\ref{sec:budget}.
\begin{figure}
\includegraphics[width=1\linewidth]{5.pdf}
\caption{Light shifts of the lower $|J=7/2\rangle$ (blue, solid) and upper $|J=5/2\rangle$ (red, dashed) clock levels caused by the hyperpolarizability. Vertical red line denotes the position of the resonance.
Calculations are done for an intensity of $I=50$\,kW/cm$^2$; the hyperfine interaction is ignored.}
\label{fig:hyper}
\end{figure}
\section{Magnetic interactions}
\label{sec:magnetic}
\subsection{Magnetic dipole-dipole interaction}
\label{sec:mdd}
The magnetic moment of the thulium ground state equals $4 \mu_B$ ($\mu_B$ is the Bohr magneton) which causes a magnetic dipole-dipole
interaction between atoms. The interaction potential between two atoms is
\begin{equation}
U_{dd}(r)=\mu_0 (g_F \mu_B)^2 \frac{\hat{\vec{F_1}} \cdot \hat{\vec{F_2}}-3(\hat{\vec{F_1}} \cdot \hat{\vec{r}}) (\hat{\vec{F_2}}
\cdot \hat{\vec{r}})}{4\pi r^3},
\label{mdd}
\end{equation}
where $\vec{F}_{1,2}$ are the total atomic angular momenta, $\mu_0$ is the magnetic permeability of vacuum, $\vec{r}$ is the vector pointing
from one atom to another and $g_F$ is the Land\'{e} g-factor of the ground state. For the Tm ground state $g_F \approx 1$.
For spatially non-uniform
atom distributions over optical lattice sites, the magnetic
interaction may lead to inhomogeneous broadening and frequency shifts of the clock transition, both of them being of the same order of
magnitude. These shifts correspond to the interaction energy between neighboring atoms.
For two Tm atoms loaded in the adjacent sites of optical lattice at 800\,nm and prepared in the $|m=4 \rangle$ magnetic state, the
interaction energy (\ref{mdd}) corresponds to the frequency shift of
\begin{equation}
\label{eq:mdd_shift}
\varDelta f_{dd} \approx \frac{\mu_0 (m\, \mu_B)^2 }{4\pi r^3 h} \,
\end{equation}
which is of the order of 10\,Hz. The shift is large and difficult to predict due to randomness of
the lattice site occupation. Further, we will analyze only the transition $|m=0\rangle \rightarrow |m=0\rangle$ which is insensitive
to this shift.
Magnetic dipole-dipole interactions also limit the interrogation time because of spin relaxation: the atomic ensemble prepared
into a pure polarized state will gradually lose its polarization.
To evaluate the corresponding relaxation time, we solved the Shr\"{o}dinger equation with the interaction (\ref{mdd}) for 2, 3, 4, and 5
spatially-fixed Tm atoms in ground state ($F_1 = F_2 = 4$) prepared in the initial $|m=0\rangle \otimes ... \otimes |m=0\rangle $ state at the vanishing external
magnetic field. The spatial separation of $a=400$\,nm corresponds to an 800-nm optical lattice. The relative positions of the atoms
are shown in Fig.\,\ref{fig_dipolar}, upper panel. The lower panel of Fig.\,\ref{fig_dipolar} shows dynamics of the spin state for the central atom, marked blue
in the upper panel of Figure.
For 2, 3, and 4 atoms the Shr\"{o}dinger equation was solved exactly. For 5 atoms the Hilbert space is too large and we restricted our
calculation to the subspace $m_i = \left\{-2,-1,0,1,2\right\}$ . To estimate validity of this approach we also solved the
Shr\"{o}dinger equation for 2, 3, and 4 atoms in the restricted subspace. The inset in the lower panel of Fig.\,\ref{fig_dipolar} shows good agreement between
approximate and exact solutions for the first 50\,ms of the evolution. In the steady-state, it is reasonable to assume that all spin projections are equiprobable. Consequently, average probability to find the central spin in the $|m=0 \rangle$ state equals 1/9 for the full Hilbert space and 1/5 for the truncated space. This explains the discrepancy between the exact and approximate solutions at longer times (> 50\,ms).
The characteristic relaxation time was derived by setting the probability to find the central spin in the initial $|m=0\rangle$ state
to 0.7. It equals 20\,ms, 13\,ms, 11\,ms, and 10\,ms for 2,3,4, and 5 spins, respectively (Fig.\,\ref{fig_dipolar}).
External magnetic field reduces spin relaxation because some spinflip processes require additional energy. A significant reduction of the spin
relaxation is expected if the Zeeman splitting becomes larger than the kinetic energy $E_K$ of atoms. At the experimentally achieved
temperature of $T\sim 10$\,$\mu$K, the kinetic energy equals $E_K=k_B T \sim 100$\,kHz$\times h$. This energy corresponds to a magnetic
field of $B\approx E_K /( h \mu_B)$ or approximately 100\,mG. We show in the next section that such a bias magnetic field will cause
a significant Zeeman shift and cannot not be applied during clock operation. As mentioned in the Introduction, the temperature can be lowered to a few microkelvin which will reduce the threshold magnetic field to a few tens of milligauss which is sufficient for a target clock
accuracy.
Assuming that the lattice filling factor is less than unity and taking into account the influence of weak magnetic bias field
(10\,mG), we conclude that the spin relaxation time should be larger than 10\,ms. This sets the bound for the interrogation time of
the clock transition and, correspondingly, the Fourier limit of its spectral linewidth of $<10$\,Hz. As a result, the spin
relaxation should not considerably impact the performance of the proposed optical clock.
\begin{figure}
\includegraphics[width=1\linewidth]{6.pdf}
\caption{Spin relaxation dynamics.
Upper panel: Configurations of atoms used in the simulation, $a=400$\,nm is the inter-atomic separation, quantization axes is
perpendicular to the plane of the sketch;
Main panel: Probability to find the central atom (blue) in the state $|m=0 \rangle$ for 2, 3, 4, and 5 atoms in the 2D optical lattice.
Note, that the probabilities do not go to zero at longer times because at steady state there is almost uniform distribution among
magnetic sublevels.
Inset: Difference of the probabilities to find the first atom in the state $|m=0\rangle$ for exact and approximate solutions for 2,3, and 4 atoms. The horizontal dashed line represents the difference of the steady-state probabilities (see the main text).
}
\label{fig_dipolar}
\end{figure}
\subsection{Interaction with an external magnetic field}
\label{sec:zeeman}
To selectively address the $|m=0\rangle \rightarrow |m=0\rangle$ transition, an external static magnetic field $B$ has to be applied. The
Hamiltonian describing hyperfine interaction for the $^{169}$Tm atom ($I=1/2$) in the presence of the external magnetic field $B$
is~\cite{Giglberger1967}:
\begin{equation}\label{eq:HFS}
H=h A \hat{\vec{I}}\cdot\hat{\vec{J}} - g_I \mu_N \hat{\vec{I}}\cdot \hat{\vec{B}}-g_J \mu_B \hat{\vec{J}}\cdot \hat{\vec{B}},
\end{equation}
where $A$ is the hyperfine constant, $g_I$ is the nuclear Land\'{e} g-factor, $\mu_N$ is the nuclear magneton, and $g_J$ is the electronic
Land\'{e} g-factor. The well-known Breit-Rabi formula gives eigenvalues for the special case of $J=1/2$.
Making the formal substitution $I\leftrightarrow J$, $g_I \leftrightarrow g_J$, $\mu_N \leftrightarrow \mu_B$ in (\ref{eq:HFS}) one can
use the Breit-Rabi expression for $I=1/2$~\cite{Giglberger1967}.
The frequency shift of the clock transition $|m=0\rangle \rightarrow |m=0\rangle$ is given by
\begin{multline}
\label{eq:quad_zeeman}
\varDelta f_{0 \rightarrow 0} = \frac{(g_{5/2}\mu_B - g_I\mu_N)^2 B^2}{4 h^2 \varDelta W_{5/2} } \\
- \frac{(g_{7/2}\mu_B - g_I\mu_N)^2 B^2}{4 h^2\varDelta W_{7/2}} =\beta \, B^2
\end{multline}
and
\begin{equation}
\beta = -257(1)\text{\,Hz}/\text{G}^2,
\end{equation}
where $\varDelta W_{7/2} = 1496.550(1)$\,MHz, $\varDelta W_{5/2} = 2114.946(1)$\,MHz are the hyperfine frequency splittings of the $|J=7/2\rangle$
and $|J=5/2\rangle$ clock levels~\cite{Leeuwen1980}, and $g_{7/2} = 1.141189(3)$~\cite{Giglberger1967},
$g_{5/2}=0.855(1)$~\cite{Blaise1965} are their Land\'{e} g-factors, $g_I
= 0.462(3)$~\cite{Giglberger1967}.
\section{ \texorpdfstring{T\MakeLowercase{m}}{Tm} clock uncertainty}
\label{sec:budget}
Here we will discuss the most significant sources of uncertainty for the proposed Tm clock.
\subsection{Black body radiation}
\label{subsec:BBR}
The frequency shift of the clock transition due to the AC-Stark shift induced by BBR is given by
\begin{multline}
\varDelta f_{BBR} =
\int \limits_{\omega=0}^{\infty} \frac{ a_0^3\omega^3}{\pi^2 c^2} \frac{\left( \alpha_{gr}^s (\omega) - \alpha_{cl}^s(\omega)
\right)}{e^{\frac{\hbar \omega}{k_B T}-1}} d\omega \\
\approx \varDelta \alpha^s_0 \frac{a_0^3\pi^2 k_B^4}{15 c^3 \hbar^4}\, T^4 =
1.17 \times 10^{-12} \, \varDelta \alpha^s_0 \text{[a.u.]}\,T^4\text{[K]}
\label{eq:mdd1}
\end{multline}
where $\varDelta \alpha^s_0$ is the differential scalar static polarizability of the clock levels in atomic units, $T$ is the temperature
in kelvin.
Our calculations (see sec.~\ref{sec:polarizability:theory}) gives $\varDelta \alpha^s_0=2$\,a.u. which results in $\varDelta f_{BBR}
=20$\,mHz at $T=300$\,K. It corresponds to a fractional frequency shift of the clock transition of $8\times 10^{-17}$ which is
much less than for the Sr atom and is comparable to Al$^+$ clock transition~\cite{Mitroy2010}. Uncertainty of the ambient temperature of
3\,K will introduce a frequency uncertainty of $3\times10^{-18}$ (0.8\,mHz).
Since there are no strong transitions from the clock levels in the infrared region the dynamic BBR
shift is negligibly small~\cite{middelmann2011}.
\subsection{Second order Zeeman shift}
\label{sec:quad_zeeman}
According to (\ref{eq:quad_zeeman}), the frequency shift of the clock transition in the external magnetic field of $B=10$\,mG
corresponds to $-25.7(1)$\,mHz or $10^{-16}$ in fractional units. One can accurately measure the bias field $B$ by monitoring the
Zeeman shift of the $|m=-4\rangle \rightarrow |m=-3\rangle$ and $|m=4\rangle \rightarrow |m=3\rangle$ transitions~\cite{Rosenband2007}. The frequency splitting of
these magnetic sensitive transitions in a magnetic field is equal to
$\xi = 2(4 g_{J=7/2,F=4} - 3 g_{J=5/2,F=3})\mu_B/h = 6.00(2)$ \,MHz/G, where $g_{J=7/2,F=4}$ and $g_{J=5/2,3=4}$ are the Land\'{e} g-factors of the states $|J=7/2, F = 4 \rangle$ and $|J=5/2, F=3\rangle$, respectively.
Given that the linewidths of both transitions are smaller than $\delta f_{4 \rightarrow 3} = 100$\,Hz (the broadening due to the magnetic interaction (\ref{eq:mdd1}) is included), the bias magnetic field $B_0$ can
be measured {\it{in situ}} with the uncertainty of $\varDelta B = \delta f_{4 \rightarrow 3} / \xi \le 0.1$\,mG. Since the magnetic field
can be stabilized at the same level over the interrogation sequence~\cite{ref:IonOpticalClocks:NIST:Science}, we take 0.1\,mG
as an upper limit for the bias magnetic field instability and estimate the quadratic Zeeman shift's contribution as $2\beta B_0 \varDelta B_0
= 0.5$\,mHz ($2\times 10^{-18}$ in fractional units) after correction.
\subsection{Dynamic light shifts}
\label{sec:lightshifts}
Fluctuations $\delta I$ of the laser intensity cause shifts and broadening of the clock transition originated from a non-zero differential
hyperpolarizability $\varDelta\gamma$:
\begin{equation}
\label{eq:hypererror}
\varDelta f_I = -\frac{\delta I}{I}\left(\frac{\varDelta \alpha}{4} I + 2 \frac{\varDelta\gamma}{64} I^2 \right)= -\frac{\delta I}{I}
\frac{\varDelta\gamma}{64} I^2\,,
\end{equation}
where we take into account that
\begin{equation}
\frac{\varDelta \alpha}{4} I + \frac{\varDelta\gamma}{64} I^2 =0
\end{equation}
at the magic wavelength.
Previously, we estimated that $\varDelta\gamma/64 \times I^2 $ is less than $0.5$\,Hz for the given lattice parameters (see sec.\,\ref{sec:hyper}).
Stabilizing the laser intensity at the level of $10^{-3}$ the uncertainty in the frequency of the clock transition can be reduced to
0.5\,mHz or $2\times 10^{-18}$ in fractional units.
\subsection{Van der Waals and quadrupole interactions}
The electrostatic van der Waals interaction between two neutral atoms shifts the clock frequency by
$
-(C_{6} a_B^6 E_H)/(h \,r^6),
$
where $C_6$ is the van der Waals coefficient in atomic units, $a_B$ is the Bohr radius, and $E_H$ is the Hartree energy.
Following~\cite{Kotochigova2011}, we estimated $C_6 \sim 6000$\,a.u. for $|J=7/2\rangle$ level.
For an atomic separation of $r=400$\,nm (atoms are placed in the 800-nm optical lattice, less than one atom per site) the
van der Waals frequency shift is less than 0.1\,mHz which corresponds to $4\times 10^{-19}$ in fractional units.
To estimate the contribution from the quadrupole-quadrupole interaction, we calculated the quadruple moment for the ground state of Tm atom
using the COWAN package. The result is $D\sim 0.5\,e a_B^2$ ($e$ is the elementary charge). The corresponding frequency shift is $D^2
/ (4\pi \epsilon_0 r^5 h) < 0.1$\,mHz.
\subsection{Line pulling and geometrical effects}
In an external bias magnetic field of $B=10$\,mG, the $|J=7/2,F=4\rangle \rightarrow |J=5/2,F=3\rangle$ transition will be split into magnetic
components. The line pulling effect~\cite{Marchi1984} can perturb the magnetic-insensitive clock $|m=0\rangle \rightarrow |m=0\rangle$ transition.
Imperfect co-alignment of the magnetic field $B$ and the polarization of the interrogating laser beam (Fig.\,\ref{latticeGeometry}) leads to
excitation of $|m=0\rangle \rightarrow |m=\pm1\rangle$ transitions and also can cause the line pulling effect.
In both cases, the separation from the clock transition to the nearest transition is not less than $10^3\gamma \approx $ 20\,kHz,
where $\gamma = 20$\,Hz is the upper bound for the expected transition linewidth. The corresponding incoherent line pulling is
negligible ($<10^{-8}$\,Hz) and does not impact the clock performance.
For reading the clock transition, absorption spectroscopy is typically used and we do not expect a contribution from the coherent line
pulling ~\cite{Beyer2015}.
Another systematic effect related to the geometry can come from misalignment of the lattice light polarization and the bias magnetic
field (see Fig.\,\ref{latticeGeometry}). The shift results from the differential tensor polarizability of the clock levels $\varDelta
\alpha^t$ and scales as the square of the misalignment angle~\cite{Nicholson2015}. It was shown that the corresponding relative
frequency shift can be reduced to less than $2\times 10^{-18}$ by proper alignment~\cite{Nicholson2015}.
\subsection{Uncertainty budget}
\label{sec:Uncertainty}
\renewcommand*{\arraystretch}{1.4}
\begin{table}[t]
\caption{ Uncertainty budget for a Tm optical clock operating at the 1.14 $\mu$m magnetic dipole transition. Atoms are trapped in the
optical lattice at the magic wavelength close to 807\,nm with the light intensity of 50\,kW/cm$^2$.}
\label{budget}
\begin{tabular}{>{\centering\arraybackslash}b{3.1cm} >{\raggedleft\arraybackslash}b{1.5cm} >{\raggedleft\arraybackslash}b{1.8cm} >{\raggedleft\arraybackslash}b{1.8cm} }
\hline\hline
Contribution & Frequency shift, mHz & Uncertainty after correc- tion, mHz & Uncertainty in fractional units, $10^{-18}$\\
\hline
BBR ($T=300\pm3$\,K) & 20 & 0.8 & 3 \\
Zeeman shift ($B=10.0\pm0.1$\,mG) & $-26$ & 0.5 & 2 \\
Light shift due to hyperpolarizability ($\delta I/I=10^{-3}$) & 0 & 0.5 & 2 \\
Light shift due to tensor polarizability & 0.5 & 0.5 & 2 \\
van der Waals and quadrupole interaction & 0.1 & 0.1 & 0.4 \\
\hline
Total & $-6 $ & 1.2 & $<5$\\
\hline\hline
\end{tabular}
\end{table}
The list of dominant frequency shifts and corresponding uncertainties is presented in Table\,\ref{budget}. The major line
shifts are the BBR shift and the second-order Zeeman shift. All of these can be well characterized and corrected to a high degree
using moderate assumptions and established experimental techniques. Light shift can also be controlled at low $10^{-18}$ level by
intensity stabilization of the light field. As a result, the systematic frequency uncertainty of the proposed Tm optical clock at
$1.14$\,$\mu$m can be reduced to $5 \times 10^{-18}$ in fractional units.
\section{Experiment}
\label{sec:experiment}
The experimental section describes our measurement of the $|J=5/2\rangle$ clock level lifetime in a dilute cloud of cold Tm atoms.
Formerly, the decay from this level was studied in Tm atoms implanted in solid and liquid $^4$He \cite{ishikawa1997}. Strong shielding of
inner shells and the high symmetry of the perturbing field of the He matrix give the impressive result of 75(3)\,ms for the
lifetime of the $|J=5/2\rangle$ clock level. Note that the level was populated by a cascade decay from highly excited levels.
In contrast to \cite{ishikawa1997}, we directly excite the $|J=7/2,F=4\rangle \rightarrow |J=5/2,F=3\rangle$ transition by spectrally narrow laser radiation at 1.14\,$\mu$m in an ensemble of Tm atoms trapped in a 1D optical lattice and measure the lifetime of the $|J=5/2\rangle$ clock level monitoring its decay to the ground $|J=7/2,F=4\rangle$ state.
Besides that, we evaluated dynamic polarizability of Tm atoms at 532\,nm by exciting parametric resonances in an optical lattice.
\subsection{Lifetime of the $|J=5/2\rangle$ clock level}
\label{sec:lifetime}
The lifetime of $|J=5/2\rangle$ clock level is measured by excitation of the magnetic dipole transition at 1.14\,$\mu$m in a 1D optical
lattice. About $10^6$ thulium atoms are laser cooled down to 20\,$\mu$K in a narrow-line MOT operating at
530.7\,nm~\cite{Sukachev2014} and then are recaptured by the 1D optical lattice. The lattice is formed by a retro-reflected focused
532\,nm cw laser beam (waist radius is 50\,$\mu$m, laser power is 3\,W) and superimposed with the atomic cloud. The trap
depth is calculated to be 400\,$\mu$K for the ground $|J=7/2,F=4\rangle$ state which provides a recapture efficiency of
40\%~\cite{Sukachev2014}.
After recapture (pulse 1 in Fig.\,\ref{fig_lifetime}), we switch the MOT off and wait for 20\,ms to let uncaptured atoms escape.
Then, a resonant 1.14\,$\mu$m laser pulse of 30\,ms (pulse 2) is applied to excite atoms to the $|J=5/2,F=3\rangle$
level~\cite{Golovizin2015}. The laser is actively stabilized to a high-finesse ULE cavity~\cite{Alnis2008} which narrows the laser
spectral linewidth down to $\sim 10$\,Hz. After the interrogation pulse, a resonant 410.6\,nm laser pulse of 1\,ms (pulse 3) is
applied to remove atoms from the $|J=7/2\rangle$ ground state (Fig.~\ref{fig_levels}).
Atoms exited to the $|J=5/2,F=3\rangle$ decay back to the $|J=7/2,F=4\rangle$ ground state, which population is monitored by a fluorescence signal induced by a delayed 410.6\,nm probe pulse (pulse 4).
The increase of the population of the $|J=7/2,F=4\rangle$ ground state is described by the exponential function
\begin{equation}
N(t) = N_{0}(1-\exp(-{t}/{\tau})),
\label{eq:exp}
\end{equation}
where $\tau$ is the lifetime of the the excited $|J=5/2,F=3\rangle$ state and $N_{0}$ is the initial number of atoms in this state. By fitting the experimental data presented in
Fig.~\ref{fig_lifetime}, we measure $\tau=112(4)$~ms. It is the lower bound for the $|J=5/2\rangle$ level natural lifetime since the measured lifetime can be reduced by additional weak losses from the $|J=5/2\rangle$ level in the optical lattice. These losses may be related to optical or magnetic Feshbach resonances~\cite{Chin2010}.
Thus, the natural linewidth of the clock transition $|J=7/2\rangle \rightarrow |J=5/2\rangle$ is expected to be not broader than 1.4\,Hz which is consistent with
the previous measurement in $^4$He matrix \cite{ishikawa1997} and the theoretical prediction of 1.14\,Hz~\cite{Kolachevsky2007}. The
natural linewidth of the transition does not limit the performance of the proposed optical clock (see sec.~\ref{sec:budget}), because for most routinely operating optical clocks the Fourier-limited spectral linewidth of the clock transition is on the order of 10\,Hz.
In the current experimental arrangement we observed the spectral linewidth of the $|J=7/2,F=4\rangle \rightarrow |J=5/2,F=3\rangle$ transition of 1\,MHz at the low power limit~\cite{Vishnyakova2016}. It is due to Zeeman splitting in the un-compensated laboratory field ($\sim 0.5$\,MHz) and inhomogeneous power broadening caused by different dynamic polarizabilities of the $|J=7/2,F=4\rangle$ and $|J=5/2,F=3\rangle$ levels at 532\,nm ($\sim 0.4$\,MHz).
\begin{figure}[t]
\includegraphics[width=1\linewidth]{./7.pdf}
\caption{\label{fig_lifetime} Measurement of the lifetime of the $|J=5/2\rangle$ level. Red dots are the normalized number of atoms decayed to the ground $|J=7/2,F=4\rangle$ level.
The solid curve is a fit by (\ref{eq:exp}).
The inset shows the pulse sequence described in the text.
}
\end{figure}
\subsection{Parametric resonances}
\label{sec:polarizability532nm}
The dynamic polarizability of the $|J=7/2\rangle$ ground state at the lattice wavelength can be evaluated by the excitation of parametric
resonances in the lattice and by monitoring the corresponding losses~\cite{Friebel1998}. This method is very sensitive to the laser
beam parameters (waist size, astigmatism) and does not allow accurate comparison to our calculations. Still, it gives the proper
order of magnitude for the polarizability and provides unambiguous proof that Tm atoms are localized in the 1D optical lattice.
At low temperatures, atomic motion in the optical lattice becomes quantized and the corresponding axial $f_a$ and radial $f_r$ oscillation
frequencies at the center of the lattice are given by
\begin{equation}
\begin{split}
\label{eq:parametric0}
f_a =& \frac{4}{w_0 \lambda}\sqrt{\frac{2 a_0^3 \alpha^s P}{c m_0 }}\,,\\
f_r =&\frac{4}{\pi w_0^2} \sqrt{\frac{a_0^3 \alpha^s P}{c m_0}}\,,
\end{split}
\end{equation}
where $P=4$\,W is the optical power of the laser beam forming the 1D lattice, $m_0$ is the Tm atomic mass, $w_0$ is the beam waist radius (at
$1/e^2$ intensity level), $\lambda=532$\,nm is the lattice wavelength, and $\alpha^s$ is the scalar polarizability at $\lambda=532$\,nm of the $|J=7/2,F=4\rangle$ level in atomic units.
According to~\cite{Landau:Mechanics}, harmonic modulation of the trap depth at frequencies $2f / n$ (here $f$ is one of the
eigenfrequencies (\ref{eq:parametric0}) and $n$ is an integer) will cause parametric excitation of the resonances and corresponding
trap losses.
To excite parametric resonances in the 532\,nm optical lattice, we harmonically modulated the laser power and, correspondingly, the
trap depth by an acousto-optical modulator (AOM) at the level of 10\%. The number of atoms remaining in the optical lattice after
100\,ms of parametric excitation was monitored by resonance fluorescence at 410.6\,nm. The corresponding spectrum is shown in
Fig.~\ref{fig:parametric}. The low frequency parametric resonances at $400(40)$\,Hz and $900(150)$\,Hz correspond to the radial
oscillations at $f_r$ and $2f_r$ frequencies. The high frequency resonances at $230(40)$\,kHz and $420(50)$\,kHz are related to axial
oscillations at $f_a$ and $2f_a$ in the tight potential wells of the lattice. Higher order parametric resonances are much weaker and
broader~\cite{Landau:Mechanics} and were not observed.
The scalar polarizability can be deduced from (\ref{eq:parametric0}) by excluding $w_0$:
\begin{equation}
\alpha^s = \frac{f_a^4 \lambda^4 c m_0}{64 f_r^2 a_0^3 \pi^2 P}\,.
\end{equation}
From the measured frequencies we estimate this value to be $360^{+300}_{-200}$ a.u. which agrees with the calculated
polarizability of 600\,a.u. within error bars. The main sources of uncertainty are astigmatism in the lattice beams, axial and radial
misalignment of the waist positions of the lattice beams, and error in our determination of the parametric resonance frequency.
\begin{figure}
\includegraphics[width=1\linewidth]{8.pdf}
\caption{Excitation of the parametric resonances in the 532\,nm 1D optical lattice. Green triangles are axial resonances, red circles are radial ones; dashed curves are guides to the eyes.}
\label{fig:parametric}
\end{figure}
In conclusion, the experimental results for the scalar dynamic polarizability $\alpha^s$ at 532\, nm in the optical lattice and in
the dipole trap are consistent with the calculated value of 600\,a.u. Although the experiment does not allow to test the accuracy of
our calculations, it unambiguously proves trapping Tm atoms in the optical lattice at 532\,nm.
\section{Summary}
We considered the possibility to use the inner-shell transition $[\text{Xe}]4f^{13}6s^2 (J=7/2,\,F=4,\,m=0) \rightarrow [\text{Xe}]4f^{13}6s^2
(J=5/2,\,F=3,\,m=0)$ in the Tm atom at $\lambda=1.14\,\mu$m as a candidate for an optical lattice clock. The transition wavelengths and
probabilities for two clock levels $|J=7/2\rangle$ and $|J=5/2\rangle$ in the spectral range 250\,nm --- 1200\,nm are calculated using the COWAN
package which allows deducing of the differential dynamic polarizability and suggests that the magic wavelength is at around 807\,nm with
an attractive optical potential. Our calculations show a reasonable correspondence with existing experimental data and significantly extend
it to the UV and IR spectral ranges.
The suggested clock transition demonstrates a low sensitivity to the BBR shift which provides a clock frequency instability at the low $10^{-18}$
level competing with the best known optical clocks. We also evaluated other feasible contributions to clock performance (magnetic
interactions, light shifts, van der Waals, and quadrupole shifts) which, after reasonable assumptions, can be lowered to the $10^{-18}$
level. Together with the relative simplicity of laser cooling and trapping Tm atoms, our results demonstrate that Tm is a promising candidate for optical clock
applications. One of the disadvantages is the relatively low carrier frequency of only $2.6\times10^{14}$\,Hz which requires longer
integration time to reach the same instability as Sr and Yb lattice clocks.
Our experiments with direct excitation of the clock transition by spectrally narrow laser radiation at $\lambda=1.14\,\mu$m set a
lower limit for the upper clock level lifetime of 112\,ms which corresponds to the natural linewidth of $<1.4$\,Hz. Experiments were
done in a 1D optical lattice at 532\,nm. Modulating the trap depth and analyzing the
corresponding parametric resonances frequencies, we deduced the scalar polarizability of the Tm ground state at 532\,nm which shows reasonable agreement with our calculations.
To experimentally study the magic wavelength and analyze systematic shifts, we plan to change the trapping wavelength to 806\,nm ---
807\,nm using a tunable Ti:sapphire laser. This will also simplify our study of Feshbach resonances and may open a way to study and
control dipole-dipole interactions using narrow band excitation of the clock transition at 1.14\,nm.
\begin{acknowledgments}
The work is supported by RFBR grants \#15-02-05324 and \#16-29-11723. We are grateful to S.\,Kanorski and V.\,Belyaev for invaluable
technical support.
\end{acknowledgments}
|
\section{Introduction}
The scattering theory in periodic diffractive structures, which are known as
diffraction gratings, has many significant applications in optical industry
\cite{bcm-01, bdc-josa95}. The time-harmonic problems have been
studied extensively in diffraction gratings by many researchers for acoustic,
electromagnetic, and elastic waves \cite{a-jiea99, a-mmas99, b-sjna95, b-sjam97,
cf-tams91, df-jmaa92, eh-mmas10, eh-mmmas12, lwz-ip15, wbllw-sjna15}. The
underlying equations of these waves are the Helmholtz equation, the Maxwell
equations, and the Navier equation, respectively. This paper is concerned with
the numerical solution of the elastic wave scattering problem in such a periodic
structure. The problem has two fundamental challenges. The first one is to
truncate the unbounded physical domain into a bounded computational domain. The
second one is the singularity of the solution due to nonsmooth grating surfaces.
Hence, the goal of this work is two fold to overcome these two issues. First, we
adopt the perfectly matched layer (PML) technique to handle the domain
truncation. Second, we use an a posteriori error analysis and design a finite
element method with adaptive mesh refinements to deal with the singularity of
the solution.
The research on the PML technique has undergone a tremendous development since
B\'{e}renger proposed a PML for solving the time-dependent Maxwell equations
\cite{b-jcp94}. The basis idea of the PML technique is to surround the domain of
interest by a layer of finite thickness fictitious material which absorbs all
the waves coming from inside the computational domain. When the waves
reach the outer boundary of the PML region, their energies are so small that
the simple homogeneous Dirichlet boundary conditions can be imposed. Various
constructions of PML absorbing layers have been proposed and investigated
for the acoustic and electromagnetic wave scattering problems \cite{bw-sjna05,
bp-mc07, cm-sjsc98, ct-g01, cw-motl94, hsz-sjma03, ls-c98, ty-anm98}. The PML
technique is much less studied for the elastic wave scattering problems
\cite{hsb-jasa96}, especically for the rigorous convergence analysis. We refer
to \cite{bpt-mc10, cxz-mc} for recent study on convergence analysis of
the elastic obstacle scattering problem. Combined with the PML technique, an
effective adaptive finite element method was proposed in \cite{bcw-josa05,
cw-sjna03} to solve the two-dimensional diffraction grating problem where the
one-dimensional grating structure was considered. Due to the competitive
numerical performance, the method was quickly adopted to solve many other
scattering problems including the obstacle scattering problems \cite{cl-sjna05,
cc-mc08} and the three-dimensional diffraction grating problem \cite{blw-mc10}.
Based on the a posteriori error analysis, the adaptive finite element PML method
provides an effective numerical strategy which can be used to solve a variety of
wave propagation problems which are imposed in unbounded domains.
In this paper, we explore the possibility of applying such an adaptive finite
element PML method to solve the diffraction grating problem of elastic
waves. Specifically, we consider the incidence of a time-harmonic elastic plane
wave on a one-dimensional grating surface, which is assumed to be elastically
rigid. The open space, which is above the surface, is assumed to be filled with
a homogeneous and isotropic elastic medium. Using the quasi-periodicity of the
solution and the transparent boundary condition, we formulate the scattering
problem equivalently into a boundary value problem in a bounded domain. The
conservation of energy is proved for the model problem and is used to verify
our numerical results when the exact solutions are not available. Following the
complex coordinate stretching, we study the truncated PML problem which is an
approximation to the original scattering problem. We develop the transparent
boundary condition for the truncated PML problem and show that it has a unique
weak solution which converges exponentially to the solution of the original
scattering problem. Moreover, an a posteriori error estimate is deduced for the
discrete PML problem. It consists of the finite element error and the PML
modeling error. The estimate is used to design the adaptive finite element
algorithm to choose elements for refinements and to determine the PML
parameters. Numerical experiments show that the proposed method can effectively
overcome the aforementioned two challenges.
This paper presents a nontrivial application of the adaptive finite element
PML method for the grating problem from the Helmholtz (acoustic) and Maxwell
(electromagnetic) equations to the Navier (elastic) equation. The elastic wave
equation is complicated due to the coexistence of compressional and shear waves
that have different wavenumbers and propagate at different speeds. In view of
this physical feature, we introduce two scalar potential functions to split the
wave field into its compressional and shear parts via the Helmholtz
decomposition. As a consequence, the analysis is much more sophisticated than
that for the Helmholtz equation or the Maxwell equations. We believe that
this work not only enriches the range of applications for the PML technique but
also is a valued contribution to the family of numerical methods for solving
elastic wave scattering problems.
The paper is organized as follows. In section 2, we introduce the model problem
of the elastic wave scattering by a periodic surface and formulate it into a
boundary value problem by using a transparent boundary condition. The
conservation of the total energy is proved for the propagating wave modes. In
section 3, we introduce the PML formulation and prove the well-posedness and
convergence of the truncated PML problem. Section 4 is devoted to the finite
element approximation and the a posteriori error estimate. In section 5, we
discuss the numerical implementation of our adaptive algorithm and present some
numerical experiments to illustrate the performance of the proposed method. The
paper is concluded with some general remarks and directions for future research
in section 6.
\section{Problem formulation}
In this section, we introduce the model problem and present an exact transparent
boundary condition to reduce the problem into a boundary value problem in a
bounded domain. The energy distribution will be studied for the reflected
propagating waves of the scattering problem.
\subsection{Navier equation}
Consider the elastic scattering of a time-harmonic plane wave by a periodic
surface $S$ which is assumed to be Lipschitz continuous and elastically rigid.
In this work, we consider the two-dimensional problem by assuming that the
surface is invariant in the $z$ direction. The three-dimensional problem will
be studied as a separate work. Figure \ref{pg} shows the problem geometry in one
period. Let $\boldsymbol{x}=[x, y]^\top\in\mathbb{R}^2$. Denote
by $\Gamma=\{\boldsymbol{x}\in\mathbb{R}^2: 0<x<\Lambda,\, y=b\}$ the
artificial boundary above the scattering surface, where $\Lambda$ is the period
and $b$ is a constant. Let $\Omega$ be the bounded domain which is enclosed from
below and above by $S$ and $\Gamma$, respectively. Finally, denote by
$\Omega^{\rm e}=\{\boldsymbol{x}\in\mathbb{R}^2: 0<x<\Lambda,\, y>b\}$ the
exterior domain to $\Omega$.
The open space, which is above the grating surface, is assumed to be filled with
a homogeneous and isotropic elastic medium with a unit mass density.
The propagation of a time-harmonic elastic wave is governed by the Navier
equation
\begin{equation}\label{une}
\mu\Delta\boldsymbol{u}+(\lambda+\mu)\nabla\nabla\cdot\boldsymbol{u}
+\omega^2\boldsymbol{u}=0\quad\text{in} ~ \Omega\cup\Omega^{\rm e},
\end{equation}
where $\omega>0$ is the angular frequency, $\mu$ and $\lambda$ are the Lam\'{e}
constants satisfying $\mu>0$ and $\lambda+\mu>0$, and $\boldsymbol{u}=[u_1,
u_2]^\top$ is the displacement vector of the total field which satisfies
\begin{equation}\label{bc}
\boldsymbol{u}=0\quad\text{on} ~ S.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.32\textwidth]{pg}
\caption{Geometry of the scattering problem.}
\label{pg}
\end{figure}
Let the surface be illuminated from above by either a time-harmonic
compressional plane wave
\[
\boldsymbol{u}_{\rm inc}(\boldsymbol x)=[\sin\theta,\,-\cos\theta]^\top
e^{{\rm i}\kappa_1(x\sin\theta - y\cos\theta)},
\]
or a time-harmonic shear plane wave
\[
\boldsymbol{u}_{\rm inc}(\boldsymbol x)=[\cos\theta,\,\sin\theta]^\top e^{{\rm
i}\kappa_2 (x\sin\theta-y\cos\theta)},
\]
where $\theta\in(-\pi/2, \pi/2)$ is the incident angle and
\begin{equation}\label{wn}
\kappa_1=\frac{\omega}{\sqrt{\lambda+2\mu}},\quad
\kappa_2=\frac{\omega}{\sqrt{\mu}}
\end{equation}
are the compressional and shear wavenumbers, respectively. It can be verified
that the incident wave also satisfies the Navier equation:
\begin{equation}\label{uine}
\mu\Delta\boldsymbol{u}_{\rm inc}+(\lambda
+\mu)\nabla\nabla\cdot\boldsymbol{u}_{\rm inc}
+\omega^2\boldsymbol{u}_{\rm inc}=0\quad\text{in} ~ \Omega\cup\Omega^{\rm e}.
\end{equation}
\begin{rema}
Our method works for either the compressional plane incident wave, or
the shear plane incident wave, or any linear combination of these two plane
incident waves. For clarity, we will take the compressional plane incident wave
as an example to present the results in our subsequent analysis.
\end{rema}
Motivated by uniqueness, we are interested in a quasi-periodic solution of
$\boldsymbol{u}$, i.e., $\boldsymbol{u}(x, y)e^{-{\rm i}\alpha x}$ is periodic
in $x$ with period $\Lambda$ where $\alpha=\kappa_1\sin\theta$. In addition, the
following radiation condition is imposed: the total displacement
$\boldsymbol{u}$ consists of bounded outgoing waves plus the incident wave
$\boldsymbol{u}_{\rm inc}$ in $\Omega^{\rm e}$.
We introduce some notation and Sobolev spaces. Let $\boldsymbol{u}=[u_1,
u_2]^\top$ and $u$ be a vector and scalar function, respectively. Define the
Jacobian matrix of $\boldsymbol{u}$ as
\[
\nabla\boldsymbol{u}=\begin{bmatrix}
\partial_x u_1 & \partial_y u_1\\
\partial_x u_2 & \partial_y u_2
\end{bmatrix}
\]
and two curl operators
\[
{\rm curl}\boldsymbol{u}=\partial_x u_2 - \partial_y u_1,\quad {\bf
curl} u=[\partial_y u, -\partial_x u]^\top.
\]
Define a quasi-periodic functional space
\[
H^1_{S, \rm qp}(\Omega)=\{u\in H^1(\Omega): u(\Lambda, y)=u(0, y)e^{{\rm
i}\alpha\Lambda},\, u=0 ~\text{on}~S\},
\]
which is a subspace of $H^1(\Omega)$ with the norm $\|\cdot\|_{H^1(\Omega)}$.
For any quasi-periodic function $u$ defined on $\Gamma$, it admits the Fourier
series expansion
\[
u(x)=\sum_{n\in\mathbb{Z}}u^{(n)}e^{{\rm i}\alpha_n x}, \quad
u^{(n)}=\frac{1}{\Lambda}\int_0^\Lambda u(x)e^{-{\rm i}\alpha_n x}{\rm
d}x,\quad\alpha_n=\alpha+n\left(\frac{2\pi}{\Lambda}\right).
\]
We define a trace functional space $H^s(\Gamma)$ with the norm
given by
\[
\|u\|_{H^s(\Gamma)}=\Bigl(\Lambda \sum_{n\in\mathbb{Z}}(1+\alpha_n^2)^s
|u^{(n)}|^2\Bigr)^{1/2}.
\]
Let $H^1_{S, \rm qp}(\Omega)^2$ and $H^s(\Gamma)^2$ be the Cartesian product
spaces equipped with the corresponding 2-norms of $H^1_{S, \rm qp}(\Omega)$ and
$H^s(\Gamma)$, respectively. It is known that $H^{-s}(\Gamma)^2$ is the
dual space of $H^s(\Gamma)^2$ with respect to the $L^2(\Gamma)^2$ inner
product
\[
\langle \boldsymbol{u},
\boldsymbol{v}\rangle_{\Gamma}=\int_{\Gamma}\boldsymbol{u}\cdot
\bar{\boldsymbol{v}}\,{\rm d}x,
\]
where the bar denotes the complex conjuate.
\subsection{Boundary value problem}
We wish to reduce the problem equivalently into a boundary value problem in
$\Omega$ by introducing an exact transparent boundary condition on $\Gamma$.
The total field $\boldsymbol{u}$ consists of the incident field
$\boldsymbol{u}_{\rm inc}$ and the diffracted field $\boldsymbol{v}$, i.e.,
\begin{equation}\label{tf}
\boldsymbol{u}=\boldsymbol{u}_{\rm inc}+\boldsymbol{v}.
\end{equation}
Noting \eqref{tf} and subtracting \eqref{uine} from \eqref{une}, we obtain the
Navier equation for the diffracted field
$\boldsymbol{v}$:
\begin{equation}\label{vne}
\mu\Delta\boldsymbol{v}+(\lambda +\mu)\nabla\nabla\cdot\boldsymbol{v}
+\omega^2\boldsymbol{v}=0\quad\text{in} ~ \Omega^{\rm e}.
\end{equation}
For any solution $\boldsymbol{v}$ of \eqref{vne}, we introduce the Helmholtz
decomposition to split it into the compressional and shear parts:
\begin{equation}\label{hdv}
\boldsymbol{v}=\nabla\phi_1 +{\bf curl}\phi_2,
\end{equation}
where $\phi_1$ and $\phi_2$ are scalar potential functions. Substituting
\eqref{hdv} into \eqref{vne} gives
\[
\nabla\left((\lambda +2\mu)\Delta\phi_1 +\omega^2\phi_1 \right)+{\bf
curl}(\mu\Delta\phi_2 +\omega^2\phi_2)=0,
\]
which is fulfilled if $\phi_j$ satisfy the Helmholtz equation
\begin{equation}\label{he}
\Delta\phi_j +\kappa_j^2\phi_j=0,
\end{equation}
where $\kappa_j$ is the wavenumber defined in \eqref{wn}.
Since $\boldsymbol{v}$ is a quasi-periodic function, we have from \eqref{hdv}
that $\phi_j$ is also a quasi-periodic function in the $x$ direction with
period $\Lambda$ and it has the Fourier series expansion
\begin{equation}\label{fse}
\phi_j(x, y)=\sum_{n\in\mathbb{Z}}\phi_j^{(n)}(y)e^{{\rm i}\alpha_n x}.
\end{equation}
Plugging \eqref{fse} into \eqref{he} yields
\begin{equation}\label{ode}
\frac{{\rm d}^2\phi_j^{(n)}(y)}{{\rm d}y^2}+\bigl(\beta_j^{(n)}\bigr)^2
\phi_j^{(n)}(y)=0, \quad y>b,
\end{equation}
where
\begin{equation}\label{beta}
\beta^{(n)}_j=
\begin{cases}
(\kappa_j^2 - \alpha_n^2)^{1/2},\quad &|\alpha_n|<\kappa_j,\\[2pt]
{\rm i}(\alpha_n^2 - \kappa_j^2)^{1/2},\quad &|\alpha_n|>\kappa_j.
\end{cases}
\end{equation}
Note that $\beta_1^{(0)}=\beta=\kappa_1\cos\theta$. We assume that $\kappa_j\neq
|\alpha_n|$ for all $n\in\mathbb{Z}$ to exclude possible resonance. Noting
\eqref{beta} and using the bounded outgoing radiation condition, we obtain the
solution of \eqref{ode}:
\[
\phi_j^{(n)}(y)=\phi_j^{(n)}(b) e^{{\rm i}\beta_j^{(n)}(y-b)},
\]
which gives Rayleigh's expansion for $\phi_j$:
\begin{equation}\label{pfre}
\phi_j(x, y)=\sum_{n\in\mathbb{Z}}\phi_j^{(n)}(b) e^{{\rm i}\bigl(\alpha_n
x+\beta_j^{(n)}(y-b)\bigr)},\quad y>b.
\end{equation}
Combining \eqref{pfre} and the Helmholtz decomposition \eqref{hdv} yields
\begin{equation}\label{vre}
\boldsymbol{v}(x, y)={\rm i}\sum_{n\in\mathbb{Z}}
\begin{bmatrix}
\alpha_n\\[5pt]
\beta_1^{(n)}
\end{bmatrix}
\phi_1^{(n)}(b) e^{{\rm i}\bigl(\alpha_n x
+\beta_1^{(n)}(y-b)\bigr)} +\begin{bmatrix}
\beta_2^{(n)}\\[5pt]
-\alpha_n
\end{bmatrix}
\phi_2^{(n)}(b) e^{{\rm i}\bigl(\alpha_n x +\beta_2^{(n)}(y-b)\bigr)}.
\end{equation}
On the other hand, as a quasi-periodic function, the diffracted field
$\boldsymbol{v}$ also has the Fourier series expansion
\begin{equation}\label{vfe}
\boldsymbol{v}(x, b)=\sum_{n\in\mathbb{Z}}\boldsymbol{v}^{(n)}(b) e^{{\rm
i}\alpha_n x}.
\end{equation}
From \eqref{vfe} and \eqref{vre}, we obtain a linear
system of algebraic equations for $\phi_j^{(n)}(b)$:
\[
\begin{bmatrix}
{\rm i}\alpha_n & {\rm i}\beta_2^{(n)}\\[5pt]
{\rm i}\beta_1^{(n)} & -{\rm i}\alpha_n
\end{bmatrix}
\begin{bmatrix}
\phi_1^{(n)}(b)\\[5pt]
\phi_2^{(n)}(b)
\end{bmatrix}=
\begin{bmatrix}
v_1^{(n)}(b)\\[5pt]
v_2^{(n)}(b)
\end{bmatrix}.
\]
Solving the above equations via Cramer's rule gives
\begin{subequations}\label{pffc}
\begin{align}
\phi_1^{(n)}(b)&=-\frac{\rm i}{\chi^{(n)}}\bigl(\alpha_n
v_1^{(n)}(b)+\beta_2^{(n)} v_2^{(n)}(b)\bigr),\\
\phi_2^{(n)}(b)&=-\frac{\rm i}{\chi^{(n)}}\bigl(\beta_1^{(n)}
v_1^{(n)}(b)-\alpha_n v_2^{(n)}(b)\bigr),
\end{align}
\end{subequations}
where
\begin{equation}\label{chi}
\chi^{(n)}=\alpha_n^2 + \beta_1^{(n)}\beta_2^{(n)}.
\end{equation}
Plugging \eqref{pffc} into \eqref{vre}, we obtain Rayleigh's expansion for the
diffracted field $\boldsymbol{v}$ in $\Omega^{\rm e}$:
\begin{align}\label{vr}
\boldsymbol{v}(x, y)=\sum_{n\in\mathbb{Z}} \frac{1}{\chi^{(n)}}
&\begin{bmatrix}
\alpha_n^2 & \alpha_n\beta_2^{(n)}\\[5pt]
\alpha_n \beta_1^{(n)} & \beta_1^{(n)}\beta_2^{(n)}
\end{bmatrix}
\boldsymbol{v}^{(n)}(b)e^{{\rm i}\bigl(\alpha_n
x+\beta_1^{(n)}(y-b)\bigr)}\notag\\
&+\frac{1}{\chi^{(n)}}
\begin{bmatrix}
\beta_1^{(n)}\beta_2^{(n)} & - \alpha_n\beta_2^{(n)}\\[5pt]
-\alpha_n \beta_1^{(n)} & \alpha_n^2
\end{bmatrix}
\boldsymbol{v}^{(n)}(b)e^{{\rm i}\bigl(\alpha_n
x+\beta_2^{(n)}(y-b)\bigr)}.
\end{align}
Given a vector field $\boldsymbol{v}=[v_1, v_2]^\top$, we define a differential
operator on $\Gamma$:
\begin{equation}\label{do}
\mathscr{D}\boldsymbol{v}=\mu\partial_y\boldsymbol{v}+(\lambda +\mu)[0,
1]^\top\nabla\cdot\boldsymbol{v}=[\mu\partial_y v_1, (\lambda
+\mu)\partial_x v_1+(\lambda +2\mu)\partial_y v_2]^\top.
\end{equation}
By \eqref{do}, and \eqref{vr}, we deduce the transparent
boundary condition
\[
\mathscr{D}\boldsymbol{v}=\mathscr{T}\boldsymbol{v}:=\sum_{n\in\mathbb{Z}}M^{(n)
} \boldsymbol{v}^{(n)}(b)e^{{\rm i}\alpha_n x}\quad\text{on} ~ \Gamma,
\]
where the matrix
\[
M^{(n)}=\frac{\rm i}{\chi^{(n)}}
\begin{bmatrix}
\omega^2 \beta_1^{(n)} & \mu\alpha_n \chi^{(n)} -\omega^2\alpha_n\\[5pt]
\omega^2\alpha_n-\mu\alpha_n \chi^{(n)}& \omega^2\beta_2^{(n)}
\end{bmatrix}.
\]
Equivalently, we have the transparent boundary condition for the total
field $\boldsymbol{u}$:
\[
\mathscr{D}\boldsymbol{u}=\mathscr{T}\boldsymbol{u}+\boldsymbol{f}\quad\text{on
} ~ \Gamma,
\]
where $\boldsymbol{f}=\mathscr{D}\boldsymbol{u}_{\rm
inc}-\mathscr{T}\boldsymbol{u}_{\rm inc}$.
The scattering problem can be reduced to the following boundary value problem:
\begin{equation}\label{bvp}
\begin{cases}
\mu\Delta\boldsymbol{u}+(\lambda+\mu)\nabla
\nabla\cdot\boldsymbol{u}+\omega^2\boldsymbol{u}=0\quad&\text{in}~ \Omega,\\
\boldsymbol{u}=0\quad&\text{on} ~ S,\\
\mathscr{D}\boldsymbol{u}=\mathscr{T}\boldsymbol{u}+\boldsymbol{f}\quad&\text{on
} ~ \Gamma.
\end{cases}
\end{equation}
The weak formulation of \eqref{bvp} reads as follows: Find $\boldsymbol{u}\in
H^1_{S, \rm qp}(\Omega)^2$ such that
\begin{equation}\label{wp}
a(\boldsymbol{u}, \boldsymbol{v})=\langle\boldsymbol{f},
\boldsymbol{v}\rangle_{\Gamma},\quad\forall\,\boldsymbol{v}\in H^1_{S, \rm
qp}(\Omega)^2,
\end{equation}
where the sesquilinear form $a: H^1_{S, \rm qp}(\Omega)^2\times H^1_{S, \rm
qp}(\Omega)^2\to\mathbb{C}$ is defined by
\begin{align}\label{sf}
a(\boldsymbol{u}, \boldsymbol{v})=\mu\int_\Omega
\nabla\boldsymbol{u}:\nabla\bar{\boldsymbol
v}\,{\rm d}\boldsymbol{x}+(\lambda+\mu)\int_\Omega (\nabla\cdot\boldsymbol{u}
)(\nabla\cdot\bar{\boldsymbol v})\,{\rm d}\boldsymbol{x}\notag\\
-\omega^2\int_\Omega \boldsymbol{u}\cdot\bar{\boldsymbol v}\, {\rm
d}\boldsymbol{x}- \langle T\boldsymbol{u}, \boldsymbol{v}\rangle_{\Gamma}.
\end{align}
Here $A:B={\rm tr}(A B^\top)$ is the Frobenius inner product of square matrices
$A$ and $B$.
The existence of a unique weak solution $\boldsymbol{u}$ of \eqref{wp} is
discussed in \cite{eh-mmas10}. In this paper, we assume that the variational
problem \eqref{wp} admits a unique solution. It follows from the general theory
in \cite{ba-73} that there exists a constant $\gamma_1>0$ such that the
following inf-sup condition holds
\begin{equation}\label{ifc}
\sup_{0\neq \boldsymbol{v}\in H^1_{S, \rm qp}(\Omega)^2}\frac{|a(\boldsymbol{u},
\boldsymbol{v})|}{\|\boldsymbol{v}\|_{H^1(\Omega)^2}}\geq \gamma_1
\|\boldsymbol{u}\|_{H^1(\Omega)^2},\quad\forall\,\boldsymbol{u}\in H^1_{S, \rm
qp}(\Omega)^2.
\end{equation}
\subsection{Energy distribution}
We study the energy distribution for the propagating reflected wave modes of
the displacement. The result will be used to verify the accuracy of our
numerical method when the analytic solution is not available.
Denote by $\boldsymbol{\nu}=(\nu_1, \nu_2)^\top$ and $\boldsymbol{\tau}=(\tau_1,
\tau_2)^\top$ the unit normal and tangential vectors on $S$, where
$\tau_1=\nu_2$ and $\tau_2=-\nu_1$. Let
$\Delta_j^{(n)}=|\kappa_j^2-\alpha_n^2|^{1/2}$ and $U_j=\{n:
|\alpha_n|<\kappa_j\}$. We point out that $U_1$ and $U_2$ are the
collections of all the propagating modes for the compressional and shear waves,
respectively. It is clear to note that $\beta_j^{(n)}=\Delta_j^{(n)}$ for $n\in
U_j$ and $\beta_j^{(n)}={\rm i}\Delta_j^{(n)}$ for $n\notin U_j$.
Consider the Helmholtz decomposition for the total field:
\begin{equation}\label{hdt}
\boldsymbol{u}=\nabla\varphi_1+{\bf curl}\varphi_2.
\end{equation}
Substituting \eqref{hdt} into \eqref{une}, we may verify that $\varphi_j$ also
satisfies the Helmholtz equation
\[
\Delta \varphi_j+\kappa_j^2\varphi_j=0\quad\text{in} ~ \Omega\cup\Omega^{\rm
e}.
\]
Using the boundary condition \eqref{bc}, we have
\[
\partial_{\boldsymbol\nu}\varphi_1-\partial_{\boldsymbol\tau}
\varphi_2=0\quad\text{and}\quad \partial_{\boldsymbol\nu}\varphi_2+\partial_{
\boldsymbol\tau}\varphi_1=0 \quad\text{on} ~ S.
\]
Correspondingly, we introduce the Helmholtz decomposition for the incident
field:
\[
\boldsymbol{u}_{\rm inc}=\nabla\psi_1 + {\bf curl}\psi_2,
\]
which gives explicitly that
\[
\psi_1=-\frac{1}{\kappa_1^2}\nabla\cdot\boldsymbol{u}_{\rm inc}=-\frac{\rm
i}{\kappa_1}e^{{\rm i}(\alpha x-\beta y)},\quad\psi_2=\frac{1}{\kappa_2^2}{\rm
curl}\boldsymbol{u}_{\rm inc}=0.
\]
Hence we have
\[
\varphi_1=\phi_1+\psi_1, \quad \varphi_2=\phi_2.
\]
Using the Rayleigh expansions \eqref{pfre}, we get
\begin{align}
\label{vphi1} \varphi_1(x, y)&=r_0 e^{{\rm i}(\alpha x-\beta
y)} +\sum_{n\in\mathbb{Z}}r_1^{(n)} e^{{\rm i}\bigl(\alpha_n
x+\beta_1^{(n)}y\bigr)},\\
\label{vphi2}\varphi_2(x, y)&=\sum_{n\in\mathbb{Z}}r_2^{(n)} e^{{\rm
i}\bigl(\alpha_n
x+\beta_2^{(n)}y\bigr)},
\end{align}
where
\[
r_0=-\frac{\rm i}{\kappa_1}, \quad r_1^{(n)}=\phi_1^{(n)}(b)e^{-{\rm
i}\beta_1^{(n)} b},\quad r_2^{(n)}=\phi_2^{(n)}(b)e^{-{\rm i}\beta_2^{(n)}b}.
\]
The grating efficiency is defined by
\begin{equation}\label{ge}
e_1^{(n)}=\frac{\beta_1^{(n)}|r_1^{(n)}|^2}{\beta|r_0|^2},\quad
e_2^{(n)}=\frac{\beta_2^{(n)}|r_2^{(n)}|^2}{\beta|r_0|^2},
\end{equation}
where $e_1^{(n)}$ and $e_2^{(n)}$ are the efficiency of the $n$-th
order reflected modes for the compressional wave and the shear wave,
respectively. We have the following conservation of energy.
\begin{theo}\label{ce}
The total energy is conserved, i.e.,
\[
\sum_{n\in U_1}e_1^{(n)}+\sum_{n\in U_2}e_2^{(n)}=1.
\]
\end{theo}
\begin{proof}
Consider the following coupled problem:
\begin{equation}\label{ce-s1}
\begin{cases}
\Delta\varphi_j + \kappa_j^2\varphi_j=0 &\quad\text{in} ~ \Omega,\\
\partial_{\boldsymbol\nu}\varphi_1-\partial_{\boldsymbol\tau}
\varphi_2=0&\quad\text{on} ~ S,\\
\partial_{\boldsymbol\nu}\varphi_2+\partial_{\boldsymbol\tau}\varphi_1=0
&\quad\text{on} ~ S.
\end{cases}
\end{equation}
It is clear to note that $\bar{\varphi}_j$ also satisfies the problem
\eqref{ce-s1} since the wavenumber $\kappa_j$ is real. Using Green's
theorem and quasi-periodicity of the solution, we have
\begin{align}\label{ce-s2}
0=&\int_\Omega(\bar{\varphi}_1\Delta\varphi_1-\varphi_1\Delta\bar{\varphi}_1)\,{
\rm
d}\boldsymbol{x}+(\bar{\varphi}_2\Delta\varphi_2-\varphi_2\Delta\bar{\varphi}
_2)\,{\rm d}\boldsymbol{x}\notag\\
=&\int_S(\bar{\varphi}_1\partial_{\boldsymbol\nu}\varphi_1-\varphi_1\partial_{
\boldsymbol\nu}\bar{\varphi}_1)\,{\rm d}s+\int_S
(\bar{\varphi}_2\partial_{\boldsymbol\nu}\varphi_2-\varphi_2\partial_{
\boldsymbol\nu}\bar{\varphi}_2)\,{\rm d}s\notag\\
& +\int_\Gamma (\bar{\varphi}_1\partial_y
\varphi_1-\varphi_1\partial_y\bar{\varphi}_1)\,{\rm d}x+\int_\Gamma
(\bar{\varphi}_2\partial_y\varphi_2-\varphi_2\partial_y\bar{\varphi}_2)\,{\rm
d}x.
\end{align}
It follows from integration by parts and the boundary conditions on $S$ in
\eqref{ce-s1} that
\begin{align*}
& \int_S \bar{\varphi}_1\partial_{\boldsymbol\nu}\varphi_1\,{\rm d}s=\int_S
\bar{\varphi}_1\partial_{\boldsymbol\tau}\varphi_2\,{\rm d}s=-\int_S
\varphi_2\partial_{\boldsymbol\tau}\bar{\varphi}_1\,{\rm d}s=\int_S
\varphi_2\partial_{\boldsymbol\nu}\bar{\varphi}_2\,{\rm d}s,\\
& \int_S \bar{\varphi}_2\partial_{\boldsymbol\nu}\varphi_2\,{\rm d}s=-\int_S
\bar{\varphi}_2\partial_{\boldsymbol\tau}\varphi_1\,{\rm d}s=\int_S
\varphi_1\partial_{\boldsymbol\tau}\bar{\varphi}_2\,{\rm d}s=\int_S
\varphi_1\partial_{\boldsymbol\nu}\bar{\varphi}_1\,{\rm d}s,
\end{align*}
which yields after taking the imaginary part of \eqref{ce-s2} that
\begin{equation}\label{ce-s3}
{\rm Im}\int_\Gamma (\bar{\varphi}_1\partial_y\varphi_1 +
\bar{\varphi}_2\partial_y\varphi_2)\,{\rm d}x=0.
\end{equation}
It follows from \eqref{vphi1} and \eqref{vphi2} that we have
\begin{align*}
\varphi_1(x, b)&=r_0 e^{{\rm i}(\alpha x-\beta
b)} +\sum_{n\in U_1}r_1^{(n)} e^{\bigl({\rm i}\alpha_n
x+{\rm i}\Delta_1^{(n)}b\bigr)} +\sum_{n\notin U_1}r_1^{(n)} e^{\bigl({\rm
i}\alpha_n x-\Delta_1^{(n)}b\bigr)},\\
\varphi_2(x, b)&=\sum_{n\in U_2}r_2^{(n)} e^{\bigl({\rm i}\alpha_n
x+{\rm i}\Delta_2^{(n)}b\bigr)}+\sum_{n\notin U_2}r_2^{(n)} e^{\bigl({\rm
i}\alpha_n x-\Delta_2^{(n)}b\bigr)},
\end{align*}
and
\begin{align*}
\partial_y\varphi_1(x, b)&=-{\rm i}\beta r_0 e^{{\rm i}(\alpha x-\beta
b)} +\sum_{n\in U_1} {\rm i}\Delta_1^{(n)} r_1^{(n)} e^{\bigl({\rm i}\alpha_n
x+{\rm i}\Delta_1^{(n)}b\bigr)} -\sum_{n\notin U_1} \Delta_1^{(n)} r_1^{(n)}
e^{\bigl({\rm i}\alpha_n x-\Delta_1^{(n)}b\bigr)},\\
\partial_y\varphi_2(x, b)&=\sum_{n\in U_2}{\rm i}\Delta_2^{(n)} r_2^{(n)}
e^{\bigl({\rm i}\alpha_n x+{\rm i}\Delta_2^{(n)}b\bigr)}-\sum_{n\notin
U_2}\Delta_2^{(n)} r_2^{(n)} e^{\bigl({\rm i}\alpha_n x-\Delta_2^{(n)}b\bigr)}.
\end{align*}
Substituting the above four functions into \eqref{ce-s3} and using the
orthogonality of Fourier series, we get
\[
\sum_{n\in U_1}\Delta_1^{(n)}|r_1^{(n)}|^2 +\sum_{n\in
U_2}\Delta_2^{(n)}|r_2^{(n)}|^2=\beta |r_0|^2,
\]
which completes the proof.
\end{proof}
In practice, the grating efficiencies \eqref{ge} can be computed from
\eqref{pffc} once the scattering problem is solved and the diffracted field
$\boldsymbol{v}$ is available on $\Gamma$.
\section{The PML problem}
In this section, we shall introduce the PML formulation for the scattering
problem and establish the well-posedness of the PML problem. An error estimate
will be shown for the solutions between the original scattering problem and the
PML problem.
\subsection{PML formulation}
Now we turn to the introduction of an absorbing PML layer. As is shown in
Figure \ref{pg_pml}, the domain $\Omega$ is covered by a chunck of PML layer
of thickness $\delta$ in $\Omega^{\rm e}$. Let $\rho(\tau)=\rho_1(\tau)+{\rm
i}\rho_2(\tau)$ be the PML function which is continuous and satisfies
\[
\rho_1=1, \quad \rho_2=0 \quad\text{for}~ \tau<b\quad\text{and}\quad
\rho_1\geq 1,\quad \rho_2>0\quad\text{otherwise}.
\]
We introduce the PML by complex coordinate stretching:
\begin{equation}\label{cs}
\hat{y}=\int_0^y \rho(\tau) {\rm d}\tau.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.36\textwidth]{pg_pml}
\caption{Geometry of the PML problem.}
\label{pg_pml}
\end{figure}
Let $\hat{\boldsymbol x}=(x, \hat{y})$. Introduce the new field
\begin{equation}\label{nf}
\hat{\boldsymbol u}(\boldsymbol{x})=\begin{cases}
\boldsymbol{u}_{\rm inc}(\boldsymbol{x})+(\boldsymbol{u}(\hat{\boldsymbol
x})-\boldsymbol{u}_{\rm inc}(\hat{\boldsymbol x})),\quad&
\boldsymbol{x}\in\Omega^{\rm e},\\
\boldsymbol{u}(\hat{\boldsymbol x}) ,\quad
&\boldsymbol{x}\in\Omega.
\end{cases}
\end{equation}
It is clear to note that $\hat{\boldsymbol u}({\boldsymbol
x})=\boldsymbol{u}(\boldsymbol{x})$ in $\Omega$ since $\hat{\boldsymbol
x}=\boldsymbol{x}$ in $\Omega$. It can be verified from \eqref{une} and
\eqref{cs} that $\hat{\boldsymbol u}$ satisfies
\[
\mathscr{L}(\hat{\boldsymbol u}-\boldsymbol{u}_{\rm inc})=0\quad\text{in}
~\Omega\cup \Omega^{\rm e}.
\]
Here the PML differential operator
\[
\mathscr{L}\boldsymbol{u}:=\begin{bmatrix}
(\lambda+2\mu)\partial_x (\rho(y)\partial_x u_1)+\mu\partial_y
(\rho^{-1}(y)\partial_y u_1)+(\lambda+\mu)\partial^2_{xy} u_2+\omega^2
\rho(y)u_1\\[5pt]
\mu\partial_x(\rho(y)\partial_x u_2)+(\lambda+2\mu)\partial_y(\rho^{-1}(y)
\partial_y u_2)+(\lambda+\mu)\partial^2_{xy} u_1+\omega^2 \rho(y)u_2
\end{bmatrix}.
\]
Define the PML region
\[
\Omega^{\rm PML}=\{\boldsymbol{x}\in\mathbb{R}^2: 0<x<\Lambda,\,
b<y<b+\delta\}.
\]
Clearly, we have from \eqref{nf} and \eqref{vr} that the outgoing wave
$\hat{\boldsymbol u}(\boldsymbol{x})-\boldsymbol{u}_{\rm
inc}(\boldsymbol{x})$ in $\Omega^{\rm e}$ decays exponentially as
$y\to \infty$. Therefore, the homogeneous Dirichlet boundary condition can be
imposed on
\[
\Gamma^{\rm PML}=\{\boldsymbol{x}\in\mathbb{R}^2:
0<x<\Lambda,\,y=b+\delta\}
\]
to truncate the PML problem. Define the computational domain for the PML
problem $D=\Omega\cup\Omega^{\rm PML}$. We arrive at the following truncated PML
problem: Find a quasi-periodic solution $\hat{\boldsymbol u}$ such that
\begin{equation}\label{pmlp}
\begin{cases}
\mathscr{L}\hat{\boldsymbol u}=\boldsymbol{g}&\quad\text{in} ~ D,\\
\hat{\boldsymbol u}=\boldsymbol{u}_{\rm inc}&\quad\text{on} ~
\Gamma^{\rm PML},\\
\hat{\boldsymbol u}=0&\quad\text{on} ~ S,
\end{cases}
\end{equation}
where
\[
\boldsymbol{g}=\begin{cases}
\mathscr{L}\boldsymbol{u}_{\rm inc}&\quad\text{in} ~
\Omega^{\rm PML},\\
0&\quad\text{in} ~ \Omega.
\end{cases}
\]
Define $H^1_{0, \rm qp}(D)=\{u\in H^1_{\rm qp}(D):
u=0~\text{on}~S\cup\Gamma^{\rm
PML}\}$. The weak formulation of the PML problem \eqref{pmlp} reads as
follows: Find $\hat{\boldsymbol u}\in H^1_{S, \rm qp}(D)^2$ such that
$\hat{\boldsymbol u}=\boldsymbol{u}_{\rm inc}$ on $\Gamma^{\rm PML}$ and
\begin{equation}\label{twp}
b_D(\hat{\boldsymbol u}, \boldsymbol{v})=-\int_D
\boldsymbol{g}\cdot\bar{\boldsymbol v}{\rm d}\boldsymbol{x},\quad\forall\,
\boldsymbol{v}\in H^1_{0, \rm qp}(D)^2.
\end{equation}
Here for any domain $G\subset\mathbb{R}^2$, the sesquilinear form $b_G: H^1_{\rm qp}(G)^2\times
H^1_{\rm qp}(G)^2\to\mathbb{C}$ is defined by
\begin{align*}
b_G(\boldsymbol{u}, \boldsymbol{v})=\int_G (\lambda+2\mu)
(\rho\partial_x u_1\partial_x\bar{v}_1+\rho^{-1}\partial_y
u_2\partial_y\bar{v}_2)+\mu(\rho^{-1}\partial_y
u_1\partial_y\bar{v}_1+\rho\partial_x u_2\partial_x\bar{v}_2)\\
+(\lambda+\mu)(\partial_x u_2\partial_y\bar{v}_1+\partial_x
u_1\partial_y\bar{v}_2)-\omega^2\rho(u_1\bar{v}_1+u_2\bar{v}_2)\,{\rm
d}\boldsymbol{x}.
\end{align*}
We will reformulate the variational problem \eqref{twp} in the domain
$D$ into an equivalent variational formulation in the domain $\Omega$, and
discuss the existence and uniqueness of the weak solution to the equivalent weak
formulation. To do so, we need to introduce the transparent boundary condition
for the truncated PML problem.
\subsection{Transparent boundary condition of the PML problem}
Let $\hat{\boldsymbol v}(\boldsymbol{x})=\boldsymbol{v}(\hat{\boldsymbol
x})=\boldsymbol{u}(\hat{\boldsymbol x})-\boldsymbol{u}_{\rm
inc}(\hat{\boldsymbol x})$. It is clear to note that $\hat{\boldsymbol v}$
satisfies the Navier equation in the complex coordinate
\begin{equation}\label{cvne}
\mu\Delta_{\hat{\boldsymbol x}}\hat{\boldsymbol v}+(\lambda
+\mu)\nabla_{\hat{\boldsymbol x}}\nabla_{\hat{\boldsymbol
x}}\cdot\hat{\boldsymbol v} +\omega^2\hat{\boldsymbol v}=0\quad\text{in} ~
\Omega^{\rm PML},
\end{equation}
where $\nabla_{\hat{\boldsymbol x}}=[\partial_x, \partial_{\hat y}]^\top$ with
$\partial_{\hat y}=\rho^{-1}(y)\partial_y$.
We introduce the Helmholtz decomposition to the solution of \eqref{cvne}:
\begin{equation}\label{chdv}
\hat{\boldsymbol v}=\nabla_{\hat{\boldsymbol x}}\hat{\phi}_1+{\bf
curl}_{\hat{\boldsymbol x}}\hat{\phi}_2,
\end{equation}
where ${\bf curl}_{\hat{\boldsymbol x}}=[\partial_{\hat y}, -\partial_x]^\top$
and $\hat{\phi}_j(\boldsymbol{x})=\phi_j(\hat{\boldsymbol{x}})$ satisfies the
Helmholtz equation
\begin{equation}\label{che}
\Delta_{\hat{\boldsymbol{x}}}\hat{\phi}_j +\kappa_j^2\hat{\phi}_j=0.
\end{equation}
Due to the quasi-periodicity of the solution, we have the Fourier series
expansion
\begin{equation}\label{cfse}
\hat{\phi}_j(x, y)=\sum_{n\in\mathbb{Z}}\hat{\phi}_j^{(n)}(y)e^{{\rm i}\alpha_n
x}.
\end{equation}
Substituting \eqref{cfse} into \eqref{che} yields
\begin{equation}\label{code}
\rho^{-1}\frac{\rm d}{{\rm d}y}\Bigl(\rho^{-1}\frac{{\rm
d}}{{\rm d}y}\hat{\phi}_j^{(n)}(y)\Bigr)+(\beta_j^{(n)})^2
\hat{\phi}_j^{(n)}(y)=0.
\end{equation}
The general solutions of \eqref{code} is
\begin{align*}
\hat{\phi}_j^{(n)}(y)=A_j^{(n)} e^{{\rm
i}\beta_j^{(n)}\int_{b}^y\rho(\tau){\rm d}\tau} + B_j^{(n)}e^{-{\rm
i}\beta_j^{(n)}\int_{b}^y \rho(\tau){\rm d}\tau}.
\end{align*}
Denote by
\begin{equation}\label{zeta}
\zeta=\int_b^{b+\delta}\rho(\tau){\rm d}\tau.
\end{equation}
The coefficients $A_j^{(n)}$ and $B_j^{(n)}$ can be uniquely determined by
solving the following linear equations
\begin{equation}\label{lev}
\begin{bmatrix}
\alpha_n & \alpha_n & \beta_2^{(n)} & -\beta_2^{(n)}\\[5pt]
\beta_1^{(n)} &-\beta_1^{(n)} & -\alpha_n & -\alpha_n\\[5pt]
\alpha_ne^{{\rm i}\beta_1^{(n)}\zeta} & \alpha_ne^{-{\rm
i}\beta_1^{(n)}\zeta}&
\beta_2^{(n)}e^{{\rm i}\beta_2^{(n)}\zeta} & -\beta_2^{(n)}e^{-{\rm
i}\beta_2^{(n)}\zeta}\\[5pt]
\beta_1^{(n)}e^{{\rm i}\beta_1^{(n)}\zeta} & -\beta_1^{(n)}e^{-{\rm
i}\beta_1^{(n)}\zeta}&-\alpha_ne^{{\rm i}\beta_2^{(n)}\zeta} &
-\alpha_ne^{-{\rm i}\beta_2^{(n)}\zeta}
\end{bmatrix}
\begin{bmatrix}
A_1^{(n)}\\[5pt]
B_1^{(n)}\\[5pt]
A_2^{(n)}\\[5pt]
B_2^{(n)}
\end{bmatrix}
=\begin{bmatrix}
-{\rm i}\hat{v}_1^{(n)}(b)\\[5pt]
-{\rm i}\hat{v}_2^{(n)}(b)\\[5pt]
0\\[5pt]
0
\end{bmatrix},
\end{equation}
where we have used the Helmholtz decomposition \eqref{chdv} and the homogeneous
Dirichlet boundary condition
\[
\hat{\boldsymbol v}(x,b+\delta)=0\quad\text{on} ~ \Gamma^{\rm PML}
\]
due to the PML absorbing layer. Solving the linear equations \eqref{lev}, we
obtain
\begin{align*}
A_1^{(n)}=&\frac{{\rm i}}{2\chi^{(n)}\hat{\chi}^{(n)}}
\Big\{-\chi^{(n)}(\varepsilon_1^{(n)}+2)
(\alpha_n\hat{v}_1^{(n)}(b)+\beta_2^{(n)}\hat{v}_2^{(n)}(b))\\
&+2\beta_2^{(n)}(\varepsilon_1^{(n)}
+2\delta_1^{(n)})(1+\delta_2^{(n)}
-\eta^{(n)})(\alpha_n\beta_1^{(n)}\hat{v}_1^{(n)}(b)
+\alpha_n^2\hat{v}_2^{(n)}(b))\Big\},\\
B_1^{(n)}=&\frac{{\rm i}}{2\chi^{(n)}\hat{\chi}^{(n)}}
\Big\{\chi^{(n)}\varepsilon_1^{(n)}
(\alpha_n\hat{v}_1^{(n)}(b)-\beta_2^{(n)}\hat{v}_2^{(n)}(b))\\
&+2(\varepsilon_1^{(n)}\delta_2^{(n)}
+2(\delta_1^{(n)}+\delta_1^{(n)}\delta_2^{(n)})
(\alpha_n\beta_1^{(n)}\beta_2^{(n)}\hat{v}_1^{(n)}(b)
-\alpha_n^2\beta_2^{(n)}\hat{v}_2^{(n)}(b))\Big\},\\
A_2^{(n)}=&\frac{{\rm i}}{2\chi^{(n)}\hat{\chi}^{(n)}}
\Big\{\chi^{(n)}[\varepsilon_1^{(n)}\eta^{(n)}
-2(\varepsilon_1^{(n)}+1)(1+\delta_2^{(n)})]
(\beta_1^{(n)}\hat{v}_1^{(n)}(b)-\alpha_n\hat{v}_2^{(n)}(b))\\
&+2\varepsilon_1^{(n)}(1+\delta_2^{(n)}-\eta^{(n)})((\beta_1^{(n)})^2\beta_2^{
(n)}\hat{v}_1^{(n)}(b)-\alpha_n^3\hat{v}_2^{(n)}(b))\Big\},\\
B_2^{(n)}=&\frac{{\rm i}}{2\chi^{(n)}\hat{\chi}^{(n)}}
\Big\{\chi^{(n)}[2\delta_2^{(n)}(\varepsilon_1^{(n)}+1)
-\varepsilon_1^{(n)}\eta^{(n)}]
(\beta_1^{(n)}\hat{v}_1^{(n)}(b)+\alpha_n\hat{v}_2^{(n)}(b))\\
&-2\delta_2^{(n)}(\varepsilon_1^{(n)}+2)
((\beta_1^{(n)})^2\beta_2^{(n)}\hat{v}_1^{(n)}(b)
+\alpha_n^3\hat{v}_2^{(n)}(b))\Big\},
\end{align*}
where
\begin{equation}\label{vepsn}
\begin{cases}
\varepsilon_j^{(n)}&=\coth(-{\rm i}\beta_j^{(n)}\zeta)-1,\\
\delta_j^{(n)}&=(e^{{\rm i}\beta_2^{(n)}\zeta}-e^{{\rm
i}\beta_1^{(n)}\zeta})/
(e^{-{\rm i}\beta_j^{(n)}\zeta}-e^{{\rm i}\beta_j^{(n)}\zeta}),\\
\eta^{(n)}&=\delta_2^{(n)}/\delta_1^{(n)}=(e^{-{\rm
i}\beta_1^{(n)}\zeta}-e^{{\rm i}\beta_1^{(n)}\zeta})/(e^{-{\rm
i}\beta_2^{(n)}\zeta}-e^{{\rm i}\beta_2^{(n)}\zeta})
\end{cases}
\end{equation}
and
\begin{equation}\label{hchi}
\hat{\chi}^{(n)}=\chi^{(n)}+4(\delta_2^{(n)}-\delta_1^{(n)}
-\delta_1^{(n)}\delta_2^{(n)})\alpha_n^2\beta_1^{(n)}\beta_2^{(n)}/\chi^{(n)}.
\end{equation}
Here, the hyperbolic cotangent function is defined as
\[
\coth(t)=(e^t+e^{-t})/(e^t-e^{-t}).
\]
Following the Helmholtz decomposition \eqref{chdv} again, we have
\begin{align}\label{cvre}
\hat{\boldsymbol v}(x, y)={\rm i} \sum_{n\in\mathbb{Z}}
&\begin{bmatrix}
\alpha_n\\[5pt]
\beta_1^{(n)}
\end{bmatrix}
A_1^{(n)}e^{{\rm i}\bigl(\alpha_n x
+\beta_1^{(n)}\int_{b}^y\rho(\tau){\rm
d}\tau\bigr)}+\begin{bmatrix}
\alpha_n\\[5pt]
-\beta_1^{(n)}
\end{bmatrix}
B_1^{(n)}e^{{\rm i}\bigl(\alpha_n x
-\beta_1^{(n)}\int_{b}^y\rho(\tau){\rm d}\tau\bigr)}\notag\\
&\begin{bmatrix}
\beta_2^{(n)}\\[5pt]
-\alpha_n
\end{bmatrix}
A_2^{(n)}e^{{\rm i}\bigl(\alpha_n x
+\beta_2^{(n)}\int_{b}^y\rho(\tau){\rm d}\tau\bigr)}-\begin{bmatrix}
\beta_2^{(n)}\\[5pt]
\alpha_n
\end{bmatrix}
B_2^{(n)}e^{{\rm i}\bigl(\alpha_n x
-\beta_2^{(n)}\int_{b}^y\rho(\tau){\rm d}\tau\bigr)}.
\end{align}
Combining \eqref{cvre} and \eqref{do}, we derive the transparent boundary
condition for the PML problem on $\Gamma$:
\[
\mathscr{D}\hat{\boldsymbol v}=\mathscr{T}^{\rm
PML}\hat{\boldsymbol v}:=\sum_{n\in\mathbb{Z}}\hat{M}^{(n)}
\hat{\boldsymbol v}^{(n)}(b)e^{{\rm i}\alpha_n x},
\]
where the matrix
\[
\hat{M}^{(n)}=
\begin{bmatrix}
\hat{m}_{11}^{(n)}&\hat{m}_{12}^{(n)}\\[5pt]
\hat{m}_{21}^{(n)}&\hat{m}_{22}^{(n)}
\end{bmatrix}.
\]
Here the entries are
\begin{align*}
\hat{m}_{11}^{(n)}&=\frac{{\rm i}\omega^2\beta_1^{(n)}}{\hat{\chi}^{(n)}}
+\frac{{\rm i}\omega^2\beta_1^{(n)}}{\chi^{(n)}\hat{\chi}^{(n)}}
\Big[\varepsilon_1^{(n)}\alpha_n^2
+(\varepsilon_1^{(n)}\eta^{(n)}+2\delta_2^{(n)})
\beta_1^{(n)}\beta_2^{(n)}\Big],\\
\hat{m}_{12}^{(n)}&={\rm i}\mu\alpha_n
-\frac{{\rm i}\omega^2\alpha_{n}}{\hat{\chi}^{(n)}}
-\frac{{\rm
i}\omega^2\alpha_{n}\beta_1^{(n)}\beta_2^{(n)}}{\chi^{(n)}\hat{\chi}^{(n)}}
\Big[\varepsilon_1^{(n)}(1+2\delta_2^{(n)}-\eta^{(n)})
+2\delta_2^{(n)}\Big],\\
\hat{m}_{21}^{(n)}&=-{\rm i}\mu\alpha_n+\frac{{\rm
i}\omega^2\alpha_n}{\hat{\chi}^{(n)}}
-\frac{{\rm
i}\omega^2\alpha_n\beta_1^{(n)}\beta_2^{(n)}}{\chi^{(n)}\hat{\chi}^{(n)}}
\Big[\varepsilon_1^{(n)}(1+2\delta_2^{(n)}-\eta^{(n)}
)+2(2\delta_1^{(n)}+2\delta_1^{(n)}\delta_2^{(n)}-\delta_2^{
(n)})\Big],\\
\hat{m}_{22}^{(n)}&=\frac{{\rm i}\omega^2\beta_2^{(n)}}{\hat{\chi}^{(n)}}
+\frac{{\rm i}\omega^2\beta_2^{(n)}}{\chi^{(n)}\hat{\chi}^{(n)}}
\Big[\varepsilon_1^{(n)}\beta_1^{(n)}\beta_2^{(n)}
+(\varepsilon_1^{(n)}\eta^{(n)}+2\delta_2^{(n)})\alpha_{n}
^2\Big].
\end{align*}
Equivalently, we have the transparent boundary condition for the total field
$\hat{\boldsymbol u}$ on $\Gamma$:
\[
\mathscr{D} \hat{\boldsymbol u}=\mathscr{T}^{\rm PML}\hat{\boldsymbol
u}+\boldsymbol{f}^{\rm PML},
\]
where $\boldsymbol{f}^{\rm PML}=\mathscr{D}\hat{\boldsymbol u}_{\rm
inc}-\mathscr{T}^{\rm PML}\hat{\boldsymbol u}_{\rm inc}$.
The PML problem can be reduced to the following boundary value problem:
\begin{equation}\label{cbvp}
\begin{cases}
\mu\Delta\boldsymbol{u}^{\rm PML}+(\lambda+\mu)\nabla
\nabla\cdot\boldsymbol{u}^{\rm
PML}+\omega^2\boldsymbol{u}^{\rm PML}=0\quad&\text{in}~ \Omega,\\
\boldsymbol{u}^{\rm PML}=0\quad&\text{on} ~ S,\\
\mathscr{D}\boldsymbol{u}^{\rm PML}=\mathscr{T}^{\rm PML}\boldsymbol{u}^{\rm
PML}+\boldsymbol{f}^{\rm PML}\quad&\text{on}
~ \Gamma.
\end{cases}
\end{equation}
The weak formulation of \eqref{cbvp} is to find $\boldsymbol{u}^{\rm PML}\in
H^1_{S, \rm qp}(\Omega)^2$ such that
\begin{equation}\label{cwp}
a^{\rm PML}(\boldsymbol{u}^{\rm PML},
\boldsymbol{v})=\langle\boldsymbol{f}^{\rm PML},
\boldsymbol{v}\rangle_{\Gamma},\quad\forall\,\boldsymbol{v}\in H^1_{S, \rm
qp}(\Omega)^2,
\end{equation}
where the sesquilinear form $a^{\rm PML}: H^1_{S, \rm qp}(\Omega)^2\times
H^1_{S, \rm qp}(\Omega)^2\to\mathbb{C}$ is defined by
\begin{align}\label{csf}
a^{\rm PML}(\boldsymbol{u}, \boldsymbol{v})=\mu\int_\Omega
\nabla\boldsymbol{u}:\nabla\bar{\boldsymbol
v}{\rm d}\boldsymbol{x}+(\lambda+\mu)\int_\Omega (\nabla\cdot\boldsymbol{u}
)(\nabla\cdot\bar{\boldsymbol v})\,{\rm d}\boldsymbol{x}\notag\\
-\omega^2\int_\Omega \boldsymbol{u}\cdot\bar{\boldsymbol v}\, {\rm
d}\boldsymbol{x}- \langle\mathscr{T}^{\rm PML}\boldsymbol{u},
\boldsymbol{v}\rangle_{\Gamma}.
\end{align}
The following lemma establishes the relationship between the variational
problem \eqref{cwp} and the weak formulation \eqref{twp}. The proof is
straightforward based on our constructions of the transparent boundary
conditions for the PML problem. The details of the proof is omitted for
briefty.
\begin{lemm}
Any solution $\hat{\boldsymbol u}$ of the variational problem \eqref{twp}
restricted to $\Omega$ is a solution of the variational \eqref{cwp}; conversely,
any solution $\boldsymbol{u}^{\rm PML}$ of the variational problem \eqref{cwp}
can be uniquely extended to the whole domain to be a solution $\hat{\boldsymbol
u}$ of the variational problem \eqref{twp} in $D$.
\end{lemm}
\subsection{Convergence of the PML solution}
Now we turn to estimating the error between $\boldsymbol{u}^{\rm PML}$ and
$\boldsymbol{u}$. The key is to estimate the error of the boundary operators
$\mathscr{T}^{\rm PML}$ and $\mathscr{T}$.
Let
\[
\Delta^{-}_j=\min\{\Delta_j^{(n)}: n\in
U_j\},\quad \Delta^{+}_j=\min\{\Delta_j^{(n)}: n\notin U_j\}.
\]
Denote
\begin{align*}
F=&\max_{j=1, 2} \left(\frac{\Delta_j^{-}}{e^{\frac{1}{2}\Delta_j^{-}{\rm
Im}\zeta} - 1},\, \frac{\Delta_j^{+}}{e^{\frac{1}{2}\Delta_j^+{\rm
Re}\zeta}-1}\right)\\
&\quad\times \max \left\{12\kappa_2,16\kappa_2^4, 8+2\kappa_2^2,
\frac{16\kappa_2^3}{\kappa_1^2},\frac{24(16+\kappa_2^2)^2}{\kappa_1^2}\right\}.
\end{align*}
The constant $F$ will be used to control the modeling error between the
PML problem and the original scattering problem. Once the incoming plane wave
$\boldsymbol{u}_{\rm inc}$ is fixed, the quantities $\Delta_j^{-},
\Delta_j^{+}$ are fixed. Thus the constant $F$ approaches to zero
exponentially as the PML parameters ${\rm Re}\zeta$ and ${\rm Im}\zeta$ tend to
infinity. Recalling the definition of $\zeta$ in \eqref{zeta}, we know that
${\rm Re}\zeta$ and ${\rm Im}\zeta$ can be calculated by the medium property
$\rho(y)$, which is usually taken as a power function:
\[
\rho(y)= 1+\sigma\left(\dfrac{y-b}{\delta}\right)^m \quad\text{if}~ y\geq b,
\quad m\geq 1.
\]
Thus we have
\[
{\rm Re}\zeta=\left(1+\frac{{\rm Re}\sigma}{m+1}\right)\delta, \quad {\rm
Im}\zeta=\left(\frac{{\rm Im}\sigma}{m+1}\right)\delta.
\]
In practice, we may pick some appropriate PML parameters $\sigma$ and $\delta$
such that ${\rm Re}\zeta\geq 1$.
\begin{lemm}\label{boe}
For any $\boldsymbol{u}, \boldsymbol{v}\in H^1_{S, \rm qp}(\Omega)^2$, we have
\[
|\langle (\mathscr{T}^{\rm PML}-\mathscr{T})\boldsymbol{u},
\boldsymbol{v}\rangle_{\Gamma}|\leq \hat{F}\|\boldsymbol{u}\|_{L^2(\Gamma)^2}
\|\boldsymbol{v}\|_{L^2(\Gamma)^2},
\]
where $\hat{F}=17\omega^2 F/\kappa_1^4$.
\end{lemm}
\begin{proof}
For any $\boldsymbol{u}, \boldsymbol{v}\in H^1_{S, \rm qp}(\Omega)^2$, we have
the following Fourier series expansions:
\[
\boldsymbol{u}(x, b)=\sum_{n\in\mathbb{Z}}\boldsymbol{u}^{(n)}(b)e^{{\rm
i}\alpha_n x},\quad
\boldsymbol{v}(x, b)=\sum_{n\in\mathbb{Z}}\boldsymbol{v}^{(n)}(b)e^{{\rm
i}\alpha_n x},
\]
which gives
\[
\|\boldsymbol{u}\|^2_{L^2(\Gamma)^2}=\Lambda\sum_{n\in\mathbb{Z}}|\boldsymbol
{ u }^{(n)} (b)|^2 , \quad
\|\boldsymbol{v}\|^2_{L^2(\Gamma)^2}=\Lambda\sum_{n\in\mathbb{Z}}
|\boldsymbol{v} ^{(n)} (b)|^2.
\]
It follows from the orthogonality of Fourier series, the Cauchy--Schwarz
inequality, and Proposition \ref{me} that we have
\begin{align*}
&|\langle (\mathscr{T}^{\rm PML}-\mathscr{T})\boldsymbol{u},
\boldsymbol{v}\rangle_{\Gamma}|=\bigg{|}\Lambda\sum_{n\in\mathbb{Z}}\bigl(
(M^{(n)}-\hat{M}^{(n)}) \boldsymbol{u}^{(n)}(b)\bigr)\cdot\bar{\boldsymbol
v}^{(n)}(b)\bigg{|}\\
&\leq\Bigl(\Lambda\sum_{n\in\mathbb{Z}}\|M^{(n)}-\hat{M}^{(n)}\|_2^2\,
|{\boldsymbol u}^{(n)}(b)|^2\Bigr)^{1/2}\Bigl(\Lambda\sum_{n\in\mathbb{Z}}|
\boldsymbol{v}^{(n)}(b)|^2
\Bigr)^{1/2}\leq\hat{F}\|\boldsymbol{u}\|_{L^2(\Gamma)^2}\|\boldsymbol{v}\|_
{ L^2(\Gamma)^2},
\end{align*}
which completes the proof.
\end{proof}
Let $a=\min_{y}\{\boldsymbol{x}\in S\}$. Denote
$\tilde{\Omega}=\{\boldsymbol{x}\in\mathbb{R}^2: 0<x<\Lambda,\,a<y<b\}$.
\begin{lemm}\label{tr}
For any $\boldsymbol{u}\in H^1_{S, \rm qp}(\Omega)^2$, we have
\[
\|\boldsymbol{u}\|_{L^2(\Gamma)^2}\leq \|\boldsymbol{u}\|_{H^{1/2}(\Gamma)^2}
\leq\gamma_2\|\boldsymbol{u}\|_{H^1(\Omega)^2},
\]
where $\gamma_2=(1+(b-a)^{-1})^{1/2}$.
\end{lemm}
\begin{proof}
First we have
\begin{align*}
(b-a)|u(b)|^2&=\int_a^b|u(y)|^2{\rm
d}y+\int_a^b\int_y^b\frac{\rm d}{{\rm d}t}|u(t)|^2{\rm d}t{\rm d}y\\
&\leq \int_a^b|u(y)|^2{\rm d}y +(b-a)\int_a^b 2|u(y)|
|u'(y)|{\rm d}y,
\end{align*}
which gives by applying the Cauchy--Schwarz inequality that
\[
(1+\alpha_n^2)^{1/2}|u(b)|^2\leq \gamma_2^2 (1+\alpha_n^2)\int_a^b |u(y)|^2{\rm
d}y+\int_a^b |u'(y)|^2{\rm d}y.
\]
Given $\boldsymbol{u}\in H^1_{S, \rm qp}(\Omega)^2$, we consider the zero
extension
\[
\tilde{\boldsymbol u}=\begin{cases}
\boldsymbol{u}&\quad\text{in} ~ \Omega,\\
0 &\quad\text{in} ~ \tilde{\Omega}\setminus\bar{\Omega},
\end{cases}
\]
which has the Fourier series expansion
\[
\tilde{\boldsymbol u}(x, y)=\sum_{n\in\mathbb{Z}}\tilde{\boldsymbol
u}^{(n)}(y)e^{{\rm i}\alpha_n x}\quad\text{in} ~ \tilde{\Omega}.
\]
By definitions, we have
\[
\|\tilde{\boldsymbol u}\|^2_{H^{1/2}(\Gamma)^2}=\Lambda\sum_{n\in\mathbb{Z}}
(1+\alpha_n^2)^{1/2}|\tilde{\boldsymbol u}^{(n)} (b)|^2
\]
and
\[
\|\tilde{\boldsymbol
u}\|^2_{H^1(\tilde{\Omega})^2}=\Lambda\sum_{n\in\mathbb{Z}} \int_a^b
(1+\alpha_n^2)|\tilde{\boldsymbol u}^{(n)}(y)|^2+|\boldsymbol{u}^{(n)'} (y)|^2
{\rm d}y.
\]
Noting $\|\boldsymbol{u}\|_{H^{1/2}(\Gamma)^2}=\|\tilde{\boldsymbol
u}\|_{H^{1/2}(\Gamma)^2}$ and
$\|\boldsymbol{u}\|_{H^1(\Omega)^2}=\|\tilde{\boldsymbol
u}\|_{H^1(\tilde{\Omega})^2}$, we complete the proof by combining the above
estimates.
\end{proof}
\begin{theo}\label{se}
Let $\gamma_1$ and $\gamma_2$ be the constants in the inf-sup condition
\eqref{ifc} and in Lemma \ref{tr}, respectively. If
$\hat{F}\gamma^2_2<\gamma_1$, then the PML variational problem
\eqref{cwp} has a unique weak solution $\boldsymbol{u}^{\rm PML}$, which
satisfies the error estimate
\begin{align}\label{ee}
\|\boldsymbol{u}-\boldsymbol{u}^{\rm PML}\|_\Omega:=\sup_{0\neq
\boldsymbol{v}\in H^1_{S,
\rm qp}(\Omega)^2}\frac{|a(\boldsymbol{u}-\boldsymbol{u}^{\rm PML},
\boldsymbol{v})|}{\|\boldsymbol{v}\|_{H^1(\Omega)^2}}
\leq \hat{F}\gamma_2 \|\boldsymbol{u}^{\rm PML}-\boldsymbol{u}_{\rm
inc}\|_{L^2(\Gamma)^2},
\end{align}
where $\boldsymbol{u}$ is the unique weak solution of the variational problem
\eqref{wp}.
\end{theo}
\begin{proof}
It suffices to show the coercivity of the sesquilinear form $a^{\rm PML}$
defined in \eqref{csf} in order to prove the unique solvability of the weak
problem \eqref{cwp}. Using Lemmas \ref{boe}, \ref{tr} and the assumption
$\hat{F}\gamma^2_2<\gamma_1$, we get for any $\boldsymbol{u},
\boldsymbol{v}$ in $H^1_{S, \rm qp}(\Omega)^2$ that
\begin{align*}
|a^{\rm PML}(\boldsymbol{u}, \boldsymbol{v})|&\geq |a(\boldsymbol{u},
\boldsymbol{v})|-|\langle (\mathscr{T}^{\rm PML}-\mathscr{T})\boldsymbol{u},
\boldsymbol{v}\rangle_{\Gamma}|\\
&\geq|a(\boldsymbol{u}, \boldsymbol{v})|-\hat{F}\gamma_2^2\|\boldsymbol{u}\|_{
H^1(\Omega)^
2} \|\boldsymbol{v}\|_{H^1(\Omega)^2}\\
&\geq\bigl(\gamma_1-\hat{F}\gamma_2^2\bigr)\|\boldsymbol{u}\|_{
H^1(\Omega)^2}\|\boldsymbol{v}\|_{H^1(\Omega)^2}.
\end{align*}
It remains to show the error estimate \eqref{ee}. It follows from
\eqref{wp}--\eqref{sf} and \eqref{cwp}--\eqref{csf} that
\begin{align*}
a(\boldsymbol{u}-\boldsymbol{u}^{\rm PML},
\boldsymbol{v})&=a(\boldsymbol{u}, \boldsymbol{v})-a(\boldsymbol{u}^{\rm PML},
\boldsymbol{v})\\
&=\langle\boldsymbol{f},
\boldsymbol{v}\rangle_{\Gamma}-\langle\boldsymbol{f}^{\rm PML},
\boldsymbol{v}\rangle_{\Gamma}+a^{\rm PML}(\boldsymbol{u}^{\rm PML},
\boldsymbol{v})-a(\boldsymbol{u}^{\rm PML}, \boldsymbol{v})\\
&=\langle (\mathscr{T}^{\rm PML}-\mathscr{T})\boldsymbol{u}_{\rm inc},
\boldsymbol{v}\rangle_{\Gamma}-\langle (\mathscr{T}^{\rm
PML}-\mathscr{T})\boldsymbol{u}^{\rm PML},
\boldsymbol{v}\rangle_{\Gamma}\\
&=\langle (\mathscr{T}-\mathscr{T}^{\rm PML})(\boldsymbol{u}^{\rm
PML}-\boldsymbol{u}_{\rm inc}), \boldsymbol{v}\rangle_{\Gamma},
\end{align*}
which completes the proof upon using Lemmas \ref{boe} and \ref{tr}.
\end{proof}
We remark that the error estimate \eqref{ee} is a posteriori in nature as it
depends only on the PML solution $\boldsymbol{u}^{\rm PML}$, which makes a
posteriori error control possible. Moreover, the PML approximation error can be
reduced exponentially by either enlarging the thickness $\delta$ of the PML
layers or enlarging the medium parameters ${\rm Re}\sigma$ and ${\rm
Im}\sigma$.
\section{Finite element approximation}
In this section, we consider the finite element approximation of the PML
problem \eqref{twp} and deduce the a posterior error estimate.
\subsection{The discrete problem}
Let $\mathcal{M}_h$ be a regular triangulation of the domain $D$. Every triangle
$T\in\mathcal{M}_h$ is considered as closed. We assume that any element $T$ must
be completely included in $\overline{\Omega^{\rm PML}}$ or $\overline{\Omega}$.
In order to introduce a finite element space whose functions are quasi-periodic
in the $x$ direction, we require that if $(0, y)$ is a node on the left
boundary, then $(\Lambda, y)$ is also a node on the right boundary, and vice
versa. Let $V_h(D)\subset H^1_{\rm qp}(D)$ be a conforming finite element space,
and $\mathring{V}_h(D)=V_h(D) \cap H^1_{0,\rm qp}(D)$.
Denote by $\Pi_h:C(\bar{D})^2\rightarrow V_h(D)^2$ the Scott--Zhang
interpolation operator, which has the following properties:
\[
\|\boldsymbol{v}-\Pi_h\boldsymbol{v}\|_{L^2(T)^2}
\leq Ch_{T}\|\nabla\boldsymbol{v}\|_{F(\tilde{T})},\quad
\|\boldsymbol{v}-\Pi_h\boldsymbol{v}\|_{L^2(e)^2}
\leq Ch_{e}^{1/2}\|\nabla\boldsymbol{v}\|_{F(\tilde{e})},
\]
where $h_T$ is the diameter of the triangle $T$, $h_e$ is the length of the
edge $e$, $\tilde T$ and $\tilde e$ are the unions of all elements which have
nonempty intersection with the element $T$ and the edge $e$, respectively, and
the Frobenius norm of the Jacobian matrix $\nabla \boldsymbol{v}$ is defined by
\[
\|\nabla\boldsymbol{v}\|_{F(G)}
=\left(\sum\limits_{j=1}^{2}\int_G|\nabla v_j|^2{\rm
d}\boldsymbol{x}\right)^{1/2}.
\]
The finite element approximation to the problem (\ref{twp}) reads as follows:
Find $\hat{\boldsymbol{u}}_h\in V_h(D)^2$
such that $\hat{\boldsymbol{u}}_h=\Pi_h\boldsymbol{u}_{\rm inc}$ on $\Gamma^{\rm
PML}$, $\hat{\boldsymbol{u}}_h=0$ on $S$, and
\begin{equation}\label{femvp}
b_D(\hat{\boldsymbol{u}}_h,\boldsymbol{v}_h)
=-\int_{D}\boldsymbol{g}\cdot\bar{\boldsymbol{v}}_h{\rm d}\boldsymbol{x},
\quad\forall\ \boldsymbol{v}_h\in\mathring{V}_h(D)^2.
\end{equation}
Following the general theory in \cite{ba-73}, the existence of a
unique solution of the discrete problem (\ref{femvp}) and the finite element
convergence analysis depend on the following discrete inf-sup condition:
\begin{equation}\label{isc}
\sup\limits_{0\neq\mathring{V}_h(D)^2}
\frac{|b(\hat{\boldsymbol{u}}_h,\boldsymbol{v}_h)|}
{\|\boldsymbol{v}_h\|_{H^1(D)^2}}
\geq \gamma_0\|\hat{\boldsymbol{u}}_h\|_{H^1(D)^2},
\quad\forall\ \hat{\boldsymbol{u}}_h\in\mathring{V}_h(D)^2,
\end{equation}
where the constant $\gamma_0>0$ is independent of the finite element mesh size.
Since the continuous problem \eqref{twp} has a unique solution by Theorem
\ref{se}, the sesquilinear form $b:H^1_{\rm qp}(D)^2\times H^1_{\rm
qp}(D)^2\rightarrow \mathbb{C}$ satisfies the continuous inf-sup condition. Then
a general argument of Schatz \cite{sch-74} implies that \eqref{isc} is valid
for sufficiently small mesh size $h<h^{*}$. Thanks to \eqref{isc}, an
appropriate a priori error estimate can be derived and the estimate depends on
the regularity of the PML solution $\boldsymbol{u}^{\rm PML}$. We assume that
the discrete problem \eqref{femvp} admits a unique
solution $\hat{\boldsymbol{u}}_h\in V_h(D)^2$, since we are interested in the a
posteriori error estimate and the associated adaptive algorithm.
Denote by $\mathcal{B}_h$ the set of all sides that do not lie on $\Gamma^{\rm
PML}$ and $S$. For any $T\in\mathcal{M}_h$, we introduce the residual
\[
R_T:=(\mathscr{L}\hat{\boldsymbol{u}}_h+\boldsymbol{g})|_{T}=
\begin{cases}
\mathscr{L}(\hat{\boldsymbol{u}}_h-\boldsymbol{u}_{\rm inc})|_T&
\text{if}\ T\in\Omega^{\rm PML},\\
\mathscr{L}\hat{\boldsymbol{u}}_h|_T&\text{otherwise}.
\end{cases}
\]
For any interior side $e\in\mathcal{B}_h$ which is the common side of $T_1$ and
$T_2$, we define the jump residual across $e$ as
\[
J_e=\mathscr{D}_{\boldsymbol\nu}\hat{\boldsymbol{u}}_h|_{T_1}-\mathscr{D}_{
\boldsymbol\nu} \hat{\boldsymbol{u}}_h|_{T_2},
\]
where the unit normal vector $\boldsymbol\nu$ on $e$ points from
$T_2$ to $T_1$ and the differential operator
\[
\mathscr{D}_{\boldsymbol\nu}\boldsymbol{v}=\mu\partial_{\boldsymbol\nu}
\boldsymbol{v}+(\lambda+\mu)(\nabla\cdot\boldsymbol{v})\boldsymbol\nu.
\]
Define
\[
\Gamma_{\rm left}=\{\boldsymbol x\in\partial D : x=0\},\quad
\Gamma_{\rm right}=\{\boldsymbol x\in\partial D : x=\Lambda\}.
\]
If $e=\Gamma_{\rm left}\cap\partial T$ for some element $T\in\mathcal{M}_h$
and $e'$ be the corresponding side on $\Gamma_{\rm right}$, which is also a side
for some element $T'$, then we define the jump residual as
\begin{align*}
J_e=&\Big[\mu\partial_x(\hat{\boldsymbol{u}}_h|_T)
+(\lambda+\mu)[1, 0]^\top\nabla_{\hat{\boldsymbol{x}}}
\cdot(\hat{\boldsymbol{u}}_h|_T)\Big]-e^{-{\rm
i}\alpha\Lambda}\Big[\mu\partial_x(\hat{\boldsymbol{u}}_h|_{T'})
+(\lambda+\mu)[1, 0]^\top\nabla_{\hat{\boldsymbol{x}}}
\cdot(\hat{\boldsymbol{u}}_h|_{T'})\Big],\\
J_{e'}=&e^{{\rm
i}\alpha\Lambda}\Big[\mu\partial_x(\hat{\boldsymbol{u}}_h|_T)
+(\lambda+\mu)[1, 0]^\top\nabla_{\hat{\boldsymbol{x}}}
\cdot(\hat{\boldsymbol{u}}_h|_T)\Big]-\Big[\mu\partial_x(\hat{
\boldsymbol{u}}_h|_{T'})+(\lambda+\mu)[1, 0]^\top\nabla_{\hat{\boldsymbol{x}}}
\cdot(\hat{\boldsymbol{u}}_h|_{T'})\Big].
\end{align*}
For any $T\in\mathcal{M}_h$, denote by $\eta_T$ the local error estimator:
\[
\eta_T=h_T\|R_T\|_{L^2(T)^2}+\bigg(\frac{1}{2}\sum\limits_{e\subset \partial
T}h_e\|J_e\|_{L^2(e)^2}^2\bigg)^{1/2}.
\]
The following theorem is the main result of this paper.
\begin{theo}\label{thmerr}
There exists a positive constant $C$ such that the following a posteriori
error estimate holds
\begin{align*}
\|\boldsymbol{u}-\hat{\boldsymbol{u}}_h\|_{H^1(\Omega)^2}
\leq&\gamma_2\hat{F}\|\hat{\boldsymbol{u}}_h-\boldsymbol{u}_{\rm
inc}\|_{L^2(\Gamma)^2}+\gamma_2 C_2\|\Pi_h{\boldsymbol{u}_{\rm
inc}}-\boldsymbol{u}_{\rm inc}\|_{L^2(\Gamma^{\rm PML})^2}\\
&+C(1+\gamma_2C_1)\bigg{(}\sum\limits_{T\in\mathcal{M}_h}\eta_T^2\bigg{)}^{1/2},
\end{align*}
where the constants $\hat{F}$, $\gamma_2$, and $C_j$ are defined in Lemmas
\ref{boe}, \ref{tr}, \ref{extend est}, \ref{bext}, respectively.
\end{theo}
\subsection{A posteriori error analysis}
For any $\boldsymbol{v}\in H^1_{\rm qp}(\Omega)^2$, we denote by
$\tilde{\boldsymbol v}$ the extension of $\boldsymbol{v}$ such that
$\tilde{\boldsymbol v}=\boldsymbol{v}$ in $\Omega$ and $\tilde{\boldsymbol v}$
satisfies the following boundary value problem
\begin{equation}\label{extend}
\begin{cases}
\mu\Delta_{\hat{\boldsymbol{x}}}\bar{\tilde{\boldsymbol v}}
+(\lambda+\mu)\nabla_{\hat{\boldsymbol{x}}}\nabla_{\hat{\boldsymbol{x}}}\cdot
\bar{\tilde{\boldsymbol v}}+\omega^2\bar{\tilde{\boldsymbol v}}=0
&\quad\text{in}\, \Omega^{\rm PML},\\
\tilde{\boldsymbol v}(x,b)=\boldsymbol{v}(x,b) &\quad\text{on} ~
\Gamma,\\
\tilde{\boldsymbol v}(x,b+\delta)=0 &\quad\text{on} ~ \Gamma^{\rm PML}.
\end{cases}
\end{equation}
\begin{lemm}\label{dp}
For any $\boldsymbol{u}$, $\boldsymbol{v}\in H^1_{\rm qp}(\Omega)^2$ we have
\[
\int_{\Gamma}\mathscr{T}^{\rm PML}\boldsymbol{u}\cdot\bar{\boldsymbol{v}}{\rm
d}x=\int_{\Gamma}\boldsymbol{u}\cdot \mathscr{D}\bar{\tilde{\boldsymbol v}}{\rm
d}x.
\]
\end{lemm}
\begin{proof}
Introduce a function $\hat{\boldsymbol{w}}\in H_{\rm qp}^1(\Omega^{\rm
PML})^2$ which satisfies
\[
\begin{cases}
\mu\Delta_{\hat{\boldsymbol{x}}}\hat{\boldsymbol{w}}+(\lambda+\mu)\nabla_{\hat{
\boldsymbol{x}}}\nabla_{\hat{\boldsymbol{x}}}\cdot\hat{\boldsymbol{w}}
+\omega^2\hat{\boldsymbol{w}}=0 &\quad\text{in} ~ \Omega^{\rm PML},\\
\hat{\boldsymbol{w}}(x,b)=\boldsymbol{u}(x,b)&\quad\text{on} ~ \Gamma,\\
\hat{\boldsymbol{w}}(x,b+\delta)=0&\quad\text{on} ~ \Gamma^{\rm PML}.
\end{cases}
\]
Using the definitions of the operators $\mathscr{T}^{\rm PML}$ and
$\mathscr{D}$, we have
\[
\mathscr{T}^{\rm PML}\boldsymbol{u}=\mathscr{D}\hat{\boldsymbol{w}}\quad
\text{on} ~ \Gamma.
\]
On the other hand, it follows from Green's formula and the extension that
\begin{align*}
\int_{\Gamma}\boldsymbol{u}\cdot\mathscr{D}
\bar{\tilde{\boldsymbol v}}{\rm d}x&
=\int_{\Gamma}\hat{\boldsymbol{w}}\cdot
\mathscr{D}\bar{\tilde{\boldsymbol v}}{\rm d}x=-\int_{\Omega^{\rm PML}}
\Big[\mu\nabla_{\hat{\boldsymbol{x}}}\bar{\tilde{\boldsymbol v}}:
\nabla_{\hat{\boldsymbol{x}}}\hat{\boldsymbol{w}}
+(\lambda+\mu)(\nabla_{\hat{\boldsymbol{x}}}\cdot\bar{\tilde{\boldsymbol
v }})(\nabla_{\hat{\boldsymbol{x}}}\cdot\hat{\boldsymbol{w}})
-\omega^2\bar{\tilde{\boldsymbol v}}\cdot\hat{\boldsymbol{w}}\Big]{\rm
d}\boldsymbol x \\
&=\int_{\Omega^{\rm
PML}}\Big[\mu\Delta_{\hat{\boldsymbol{x}}}\hat{\boldsymbol{w} }
+(\lambda+\mu)\nabla_{\hat{\boldsymbol{x}}}\nabla_{\hat{\boldsymbol{x}}}\cdot
\hat{\boldsymbol{w}}+\omega^2\hat{\boldsymbol{w}}\Big]\cdot\bar{\tilde{
\boldsymbol v}}{\rm d}\boldsymbol x
+\int_{\Gamma}\mathscr{D}\hat{\boldsymbol{w}}\cdot\bar{\tilde{\boldsymbol
v}}{\rm d}x\\
&=\int_{\Gamma}\mathscr{D}\hat{\boldsymbol{w}}\cdot\bar{\tilde{\boldsymbol
v}}{\rm d}x=\int_{\Gamma}\mathscr{T}^{\rm
PML}\boldsymbol{u}\cdot\bar{\tilde{\boldsymbol v}}{\rm d}x,
\end{align*}
which completes the proof.
\end{proof}
Define $\mathring{H}^1_{\rm qp}(D)=\{v\in H^1_{\rm qp}(D): v=0 ~ \text{on} ~
\Gamma^{\rm PML}\}$. The following two lemmas are concerned with the stability
of the extension. The proofs are given in Appendix.
\begin{lemm}\label{extend est}
Let $\boldsymbol{v}\in H^1_{\rm qp}(\Omega)^2$ and $\tilde{\boldsymbol v}\in
\mathring{H}^1_{\rm qp}(D)^2$ be its extension satisfying \eqref{extend}. Then
there exists a positive constant $C_1$ such that
\[
\|\nabla\tilde{\boldsymbol v}\|_{F(\Omega^{\rm PML})}\leq
\gamma_2 C_1\|\boldsymbol{v}\|_{H^1(\Omega)^2},
\]
\end{lemm}
\begin{lemm}\label{bext}
Let $\boldsymbol{v}\in H^1_{\rm qp}(\Omega)^2$ and $\tilde{\boldsymbol v}\in
\mathring{H}^1_{\rm qp}(D)^2$ be its extension satisfying \eqref{extend}. Then
there exists a positive constant $C_2$ such that
\[
\|\mathscr{D}\tilde{\boldsymbol v}\|_{L^2(\Gamma^{\rm
PML})^2}\leq \gamma_2 C_2\|\boldsymbol{v}\|_{H^1(\Omega)^2}.
\]
\end{lemm}
For simplicity, we shall write $\tilde{\boldsymbol v}$ as
$\boldsymbol{v}$ in the rest of the paper since no confusion of the
notation is incurred.
\begin{lemm}[Error representation formula]\label{erf}
For any $\boldsymbol{v}\in H^1_{S, \rm qp}(\Omega)^2$, which is extended to be a
function in $H^1_{0, \rm qp}(D)^2$ according to
\eqref{extend}, and $\boldsymbol{v}_h\in\mathring{V}_h(D)^2$, we have
\begin{align*}
a(\boldsymbol{u}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})
=&-\int_D \boldsymbol{g}\cdot(\bar{\boldsymbol v}-\bar{\boldsymbol v}_h){\rm
d}\boldsymbol{x}-b_D(\hat{\boldsymbol{u}}_h,\boldsymbol{v}-\boldsymbol{v}_h)\\
&+\int_{\Gamma}(\mathscr{T}-\mathscr{T}^{\rm PML})
(\hat{\boldsymbol{u}}_h-\boldsymbol{u}_{\rm inc})
\cdot\bar{\boldsymbol{v}}{\rm d}x\label{errform}+\int_{\Gamma^{\rm PML}}
(\Pi_h\boldsymbol{u}_{\rm inc}-\boldsymbol{u}_{\rm inc})
\cdot \mathscr{D}_{\hat{\boldsymbol{x}}}\bar{\boldsymbol v}{\rm d}x.
\end{align*}
\end{lemm}
\begin{proof}
It follows from \eqref{sf} and \eqref{twp} that
\begin{align*}
a(\boldsymbol{u}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})
&=a(\boldsymbol{u}-\hat{\boldsymbol{u}},\boldsymbol{v})
+a(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})\\
&=\int_{\Gamma}(\mathscr{T}-\mathscr{T}^{\rm
PML})(\hat{\boldsymbol{u}}-\boldsymbol{u}_{\rm
inc})\cdot\bar{\boldsymbol{v}}{\rm d}x+a^{\rm
PML}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})
-\int_{\Gamma}(\mathscr{T}-\mathscr{T}^{\rm
PML})(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h)
\cdot\bar{\boldsymbol{v}}{\rm d}x\\
&=\int_{\Gamma}(\mathscr{T}-\mathscr{T}^{\rm
PML})(\hat{\boldsymbol{u}}_h-\boldsymbol{u}_{\rm
inc})\cdot\bar{\boldsymbol{v}}{\rm d}x
+a^{\rm PML}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v}).
\end{align*}
Using \eqref{sf} and Lemma \ref{dp} give
\[
a^{\rm
PML}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})=b_{\Omega
}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})-\int_{\Gamma}
\mathscr{T}^{\rm PML}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h)
\cdot\boldsymbol{v}{\rm
d}x=b_{\Omega}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,
\boldsymbol{v})-\int_{\Gamma}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h)
\cdot \mathscr{D}\bar{\boldsymbol{v}}{\rm d}x.
\]
Since $\mathscr{L}\bar{\boldsymbol{v}}=0$ in $\Omega^{\rm PML}$, we deduce by
Green's formula that
\[
b_{\Omega^{\rm
PML}}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})
=-\int_{\Gamma}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h)
\cdot \mathscr{D}\bar{\boldsymbol{v}}{\rm d}x
+\int_{\Gamma^{\rm PML}}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h)
\cdot \mathscr{D}_{\hat{\boldsymbol x}}\bar{\boldsymbol v}{\rm d}x.
\]
Applying \eqref{twp} and \eqref{femvp} yields
\begin{align*}
a^{\rm PML}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})
=&b_{D}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})
-\int_{\Gamma^{\rm PML}}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h)
\cdot \mathscr{D}_{\hat{\boldsymbol x}}\bar{\boldsymbol{v}}{\rm d}x\\
=&-\int_D \boldsymbol{g}(\bar{\boldsymbol v}-\bar{\boldsymbol v}_h){\rm
d}\boldsymbol{x}-b_D (\hat{\boldsymbol{u}}_h,\boldsymbol{v}-\boldsymbol{v}_h)
-\int_{\Gamma^{\rm PML}}(\hat{\boldsymbol{u}}-\hat{\boldsymbol{u}}_h)
\cdot \mathscr{D}_{\hat{\boldsymbol x}}\bar{\boldsymbol{v}}{\rm d}x,
\end{align*}
which completes the proof.
\end{proof}
Clearly, it suffices to evaluate all the terms in the error representation
formula in order to show the posteriori error estimate in the Theorem
\ref{thmerr}. Now we present the proof as follows.
\begin{proof}
Taking $\boldsymbol{v}_h=\Pi_h\boldsymbol{v}_h\in H_{0, \rm qp}^1(D)^2$ in
Lemma \ref{erf} for the error representation formula, we have
\begin{align*}
a(\boldsymbol{u}-\hat{\boldsymbol{u}}_h,\boldsymbol{v})
=&-\int_D \boldsymbol{g}(\bar{\boldsymbol v}-\bar{\boldsymbol
v}_h){\rm d}\boldsymbol{x}
-b_{D}(\hat{\boldsymbol{u}}_h,\boldsymbol{v}-\boldsymbol{v}_h)\\
&+\int_{\Gamma}(\mathscr{T}-\mathscr{T}^{\rm
PML})(\hat{\boldsymbol{u}}_h-\boldsymbol{u}_{\rm inc})\cdot\boldsymbol{v}{\rm
d}x+ \int_{\Gamma^{\rm PML}}(\Pi_h{\boldsymbol{u}}_{\rm
inc}-\boldsymbol{u}_{\rm inc})
\cdot \mathscr{D}_{\hat{\boldsymbol x}}\bar{\boldsymbol{v}}{\rm d}x\\
=& J_1+J_2+J_3+J_4.
\end{align*}
It follows from integration by parts that
\[
J_1+J_2=\sum\limits_{T\in\mathcal{M}_h}\Big(
\int_TR_T\cdot(\bar{\boldsymbol v}-\Pi_h\bar{\boldsymbol v})
{\rm d}\boldsymbol{x}+\sum\limits_{e\subset\partial T}\frac{1}{2}\int_e J_e
\cdot(\bar{\boldsymbol v}-\Pi_h\bar{\boldsymbol v}){\rm d}x\Big),
\]
which gives after using the interpolation estimates and Lemma \ref{extend est}
that
\[
|J_1+J_2|\leq
C\sum\limits_{T\in\mathcal{M}_h}\eta_T \|\nabla\boldsymbol{v}\|_{
F(\tilde{T})}\leq C(1+\gamma_2C_1)\left(\sum\limits_{T\in\mathcal{M}_h}
\eta_T^2\right)^{1/2}
\|\boldsymbol{v}\|_{H^1(\Omega)^2}.
\]
By Lemmas \ref{boe} and \ref{tr}, we obtain
\[
|J_3|\leq \hat{F}\|\hat{\boldsymbol{u}}_h-\boldsymbol{u}_{\rm
inc}\|_{L^2(\Gamma)^2} \|\boldsymbol{v}\|_{L^2(\Gamma)^2}
\leq \gamma_2\hat{F}\|\hat{\boldsymbol{u}}_h-\boldsymbol{u}_{\rm
inc}\|_{L^2(\Gamma)^2}\|\boldsymbol{v}\|_{H^1(\Omega)^2}.
\]
Finally, it follows from Lemmas \ref{tr} and \ref{bext} that
\begin{align*}
|J_4|\leq & C_2 \|\Pi_h\boldsymbol{u}_{\rm inc}
-\boldsymbol{u}_{\rm inc}\|_{L^2(\Gamma^{\rm PML})^2}
\|\boldsymbol{v}\|_{L^2(\Gamma)^2}\\
\leq & \gamma_2 C_2 \|\Pi_h\boldsymbol{u}_{\rm inc}
-\boldsymbol{u}_{\rm inc}\|_{L^2(\Gamma^{\rm PML})^2}
\|\boldsymbol{v}\|_{H^1(\Omega)^2}.
\end{align*}
The proof is completed by combining the above estimates
\end{proof}
\section{Numerical experiments}
According to the discussion in section 3, we choose the PML medium property as
the power function and need to specify the thickness $\delta$ of the layers
and the medium parameter $\sigma$. Recall from Theorem \ref{thmerr} that the a
posteriori error estimate consists of two parts: the PML error $\epsilon_{\rm
PML}$ and the finite element discretization error $\epsilon_{\rm FEM}$, where
\begin{align}
\label{epml} \epsilon_{\rm PML}&=\hat{F}\|\boldsymbol{u}^{\rm
PML}_h-\boldsymbol{u}_{\rm inc}\|_{L^2(\Gamma)^2},\\
\label{efem}\epsilon_{\rm FEM}&=\|\boldsymbol{u}^{\rm
PML}_h-\boldsymbol{u}_{\rm inc}\|_{L^2(\Gamma^{\rm PML})^2}+
\left(\sum_{T\in\mathcal{M}_h}\eta^2_T\right)^{1/2}.
\end{align}
In our implementation, we first choose $\delta$ and $\sigma$ such that
$\hat{F} \Lambda^{1/2}\leq 10^{-8}$, which makes the PML error negligible
compared with the finite element discretization error. Once the PML region and
the medium property are fixed, we use the standard finite element adaptive
strategy to modify the mesh according to the a posteriori error estimate
\eqref{efem}. For any $T\in\mathcal{M}_h$, we define the local a posteriori
error estimator
\[
\hat{\eta}_T=\eta_T +\|\Pi_h\boldsymbol{u}_{\rm
inc}-\boldsymbol{u}_{\rm inc}\|_{L^2(\Gamma^{\rm PML}\cap\partial T)^2}.
\]
The adaptive FEM algorithm is summarized in Table \ref{alg}.
\begin{table}
\hrulefill
\begin{tabular}{ll}
1 & Given the tolerance $\epsilon > 0, \tau\in (0,1)$;\\
2 & Choose $\delta$ and $\sigma$ such that $\hat{F} \Lambda^{1/2}\leq
10^{-8}$;\\
3 & Construct an initial triangulation $\mathcal{M}_h$ over $\Omega$ and
compute error estimators;\\
4 & While $\epsilon_h>\epsilon$ do\\
5 & \qquad choose $\hat{\mathcal{M}}_h\subset\mathcal{M}_h$ according to the
strategy $\eta_{\hat{\mathcal{M}}_h}>\tau\eta_{\mathcal{M}_h}$;\\
6 & \qquad refine all the elements in $\hat{\mathcal{M}}_h$ and obtain a new
mesh denoted still by $\mathcal{M}_h$;\\
7 & \qquad solve the discrete problem \eqref{femvp} on the new mesh
$\mathcal{M}_h$;\\
8 & \qquad compute the corresponding error estimators;\\
9 & End while.
\end{tabular}
\hrulefill
\caption{The adaptive FEM algorithm.}
\label{alg}
\end{table}
In the following, we present two examples to demonstrate the
competitive numerical performance of the proposed algorithm. We choose
$\lambda=1$ and $\mu=2$. The implementation of the adaptive algorithm is based
on FreeFem++-cs \cite{h-jnm12}.
{\em Example} 1. We consider the simplest periodic structure, a straight line.
In this situation, the exact solution is available, which allows us to test the
accuracy of the numerical algorithm. Assume that a plane compressional plane
wave $\boldsymbol{u}_{\rm inc}=[\sin\theta,\,-\cos\theta]^\top e^{{\rm i}(\alpha
x-\beta y)}$ is incident on the straight line $y=0$, where
$\alpha=\kappa_1\sin\theta, \beta=\kappa_1\cos\theta, \theta\in(-\pi/2,\,\pi/2)$
is the incident angle. It follows from the Navier equation, Helmholtz
decomposition, and outgoing radiation condition that we obtain the exact
solution
\[
\boldsymbol{u}(x, y)=\boldsymbol{u}_{\rm inc}(x, y)-[\alpha,\,
\beta]^\top R_1 e^{{\rm i}(\alpha x+\beta y)}-[\beta_2^{(0)}, \, -\alpha]^\top
R_2 e^{{\rm i}(\alpha x+\beta_2^{(0)}y)},
\]
where $\beta_2^{(0)}=(\kappa_2^2-\alpha^2)^{1/2}$ and
\[
R_1=\left(\frac{\alpha\sin\theta-\beta_2^{(0)}\cos\theta}{\alpha^2+\beta\beta_2^
{(0)}} \right),
\quad
R_2=\left(\frac{\alpha\cos\theta+\beta\sin\theta}{\alpha^2+\beta\beta_2^{(0)}}
\right).
\]
In our experiment, the parameters are chosen as $\theta=\pi/6$,
$\omega=2\pi$, and the domain $\Omega=(0,1)\times(0,1)$.
Figure \ref{ex1:err} shows the curves of
$\log \|\nabla(\boldsymbol{u}-\hat{\boldsymbol u}_k)\|_{F(\Omega)}$ versus
$\log N_k$ for both the a priori and the a posteriori error estimates, where
$N_k$ is the number of nodes of the mesh $\mathcal{M}_k$. The result shows that
the meshes and the associated numerical complexity are quasi-optimal for the
proposed method, i.e., $\log \|\nabla(\boldsymbol{u}-\hat{\boldsymbol
u}_k)\|_{F(\Omega)}=CN_k^{-1/2}$ is valid asymptotically.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{ex1err}
\includegraphics[width=0.4\textwidth]{ex1est}
\caption{Example 1: Quasi-optimality of the a priori (left) and a posteriori
(right) error estimates.}
\label{ex1:err}
\end{figure}
{\em Example 2}. This example is concerned with the scattering of the
compressional plane wave
$\boldsymbol{u}_{\rm inc}=[\sin\theta,\,-\cos\theta]^\top e^{{\rm i}(\alpha
x-\beta y)}$ on a grating surface with a sharp angle. The problem geometry
is shown in Figure \ref{ex2:geo}. The parameters are chosen the same as those
for Example 1. Since there is no exact solution for this example, we plot in
Figure \ref{ex2:err} the curves of $\log\|\nabla(\boldsymbol{u}-\hat{\boldsymbol
u}_k)\|_{F(\Omega)}$ versus $\log N_k$ for the a posteriori error estimate,
where $N_k$ is the number of nodes of the mesh $\mathcal{M}_k$. Again, the
result shows that the meshes and the associated numerical complexity are
quasi-optimal for the proposed method. To verify Theorem \ref{ce}, we plot
in Figure \ref{ex2:eff} the grating efficiencies and the errors of the total
efficiency for different PML thickness. Figure \ref{ex3:mesh} shows the mesh and
the amplitude of the associated solution after 6 adaptive iterations when the
grating efficiency is stabilized. The mesh has 8491 nodes. This example shows
clearly the ability of the proposed method to capture the singularity of the
solution.
\begin{figure}
\center
\includegraphics[width=0.3\textwidth]{geo}
\caption{Example 2: Geometry of the domain}
\label{ex2:geo}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.4\textwidth]{ex2est}
\caption{Example 2: Quasi-optimality of the a posteriori error estimates for
different PML thickness.}
\label{ex2:err}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.4\textwidth]{eff1}
\includegraphics[width=0.4\textwidth]{eeff3}
\caption{Example 2: (left) Grating efficiency with $\delta=1.0$; (right)
Robustness of grating efficiency with respect to the thickness of PML
layers}
\label{ex2:eff}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.25\textwidth]{ex2mesh}
\hspace{2.5cm}
\includegraphics[width=0.35\textwidth]{ex2solu}
\caption{Example 2: The mesh (left) and the surface plot of the amplitude
of the associated solution (right) after 6 adaptive iterations. The mesh has
8491 nodes.}\label{ex3:mesh}
\end{figure}
\section{Concluding remarks}
We presented an adaptive finite element method with the PML absorbing layer
technique for the elastic wave scattering problem in a periodic structure. We
showed that the truncated PML problem has a unique weak solution which
converges exponentially to the solution of the original problem by increasing
the PML parameters. We deduced the a posteriori error estimate for the PML
solution which serves as a basis for the adaptive finite element approximation.
Numerical results show that the proposed method is effective to solve the
diffractive grating problem of elastic waves. The method can be directly
applied to solve the diffraction grating problems with other interface and/or
boundary conditions. We are also currently extending the method to the
three-dimensional problem where biperiodic structures need to be considered.
|
\section{Introduction}
Symplectic dynamics has been an intense research area for the last fifty or sixty years.
It all started with Poincar\'{e}--Birkhoff Theorem in the first decades of last century.
Arnold conjecture and related questions were pivotal to the revolution in this field, though.
The existence of extra fixed points for Hamiltonian symplectomorphisms, statement of Arnold conjecture, has
been proved in most of the instances.
The Hamiltonian flows also tend to have special dynamical properties. For instance, the existence of periodic orbits
on a dense set of energy levels (see Hofer and Zehnder \cite{hoferzehnder}).
There is another instance of Hamiltonian rigidity, that is the Conley conjecture (proven in several cases,
see \cite{ginzburggurel}). The conjecture states that a Hamiltonian symplectomorphism has an infinite number of
geometrically different periodic orbits. The mantra behind this note is to point out that the removal of the
Hamiltonian condition destroys most of the rigidity in dynamics. The prototypical example is $\mathds{T}^2$ with the symplectic
(non Hamiltonian) irrational translation: it is minimal and, therefore, it does not have a single periodic orbit. We will generalize this example.
Rigidity in contact dynamics is more subtle.
Hamiltonian contact flows do correspond to a special kind of Hamiltonian symplectic flows.
Therefore, the existence result of periodic orbits in a generic energy level was conjecturally strengthen to
all the levels in the convex case. The generalization of this fact is the content of the famous Weinstein conjecture. It has been a powerful engine developing
contact topology in the last 30 years. It has been proven in several cases.
However, the behaviour of a discrete contact dynamical system remained vague. Only, recently, an analogue of the Arnold conjecture
was stated: the translated points conjecture (see \cite{sandon}).
In this case, our aim is to show that the interplay between the Reeb flow and the Hamiltonian contactomorphisms,
which is the content of the conjecture, is probably the only natural rigid phenomenum in discrete Hamiltonian contact dynamics.
In particular, we will show that (positive) Hamiltonian contactomorphisms can be pretty wild from a dynamical point of view.
The approximation by conjugation method, introduced by Anosov and Katok, is a remarkable technique to
produce diffeomorphisms with prescribed properties on the asymptotical distribution of their orbits.
In their seminal paper \cite{anosovkatok}, Anosov and Katok proved the existence of ergodic transitive diffeomorphisms
on any closed manifold. The name of the method comes from the fact that the maps are constructed as limits
of conjugates $h R_{\alpha} h^{-1}$ of a non--trivial $\mathbb{S}^1$--action $\{R_{\alpha}: \alpha \in \mathbb{S}^1\}$ on the manifold.
This direct approach was replaced by an abstract one by Fathi and Herman in \cite{fathiherman}. They applied
the Baire category theorem to the closure of the previous set of conjugates to prove the existence of minimal and uniquely ergodic
diffeomorphisms in closed manifolds that admit a locally free $\mathbb{S}^1$--action. A short overview of the conjugation method
together with further applications is found in \cite{fayadkatok}.
In this article, the arguments in \cite{fathiherman} are adapted to the symplectic and contact setting.
Denote by $\mathrm{Symp}(M, \omega)$ the group of symplectomorphisms of a symplectic manifold $(M, \omega)$
and $\mathrm{Symp}_0(M, \omega)$ the connected component of the identity map. Similarly, denote by $\mathrm{Cont}(M, \xi)$
the group of contactomorphisms of a contact manifold $(M, \xi)$ and $\mathrm{Cont}_0(M, \xi)$ the connected component
of the identity. All these groups are known to be $C^1$--closed, hence $C^{\infty}$--closed, in $\mathrm{Diff}(M)$.
As it will be clear from the contents of the article a more ambitious goal would be to adapt the statements
to the Hamiltonian symplectomorphism group\footnote{As is well--known, the Hamiltonian contactomorphism group
corresponds to the identity component of the contactomorphisms group, i.e. any contact vector field is Hamiltonian.}. This encounters two obstacles:
\begin{itemize}
\item $\mathrm{Ham}(M, \omega)$ is not known to be $C^{\infty}$--closed in $\mathrm{Diff}(M)$ as this is the content of the Flux
conjecture. However, this can be dealt with in particular cases (for instance, assuming that $M$ is simply connected).
\item A Hamiltonian $\mathbb{S}^1$--action is never locally free. Therefore, the conjugation method, particularly simple in \cite{fathiherman}, needs to be sophisticated \cite{fayadkatok}.
This requires more subtle symplectic topology results to be developed in a forthcoming work.
\end{itemize}
A map is called minimal if every orbit is dense and is called uniquely ergodic provided it has
just one invariant probability measure.
The main theorems of this article are:
\begin{theorem}\label{thm:symplectic}
If a symplectic manifold $(M, \omega)$ admits a locally free $\mathbb{S}^1$--action by symplectomorphisms, then
there exists $\psi \in \mathrm{Symp}_0(M)$ such that $\psi$ is minimal.
\end{theorem}
\begin{theorem}\label{thm:contact}
If a contact manifold $(M, \xi)$ admits a locally free $\mathbb{S}^1$--action by contactomorphisms, then
there exists $\psi \in \mathrm{Cont}_0(M, \xi)$ such that $\psi$ is strictly ergodic.
\end{theorem}
Given a contact manifold $(M, \xi)$, $\xi = \ker(\alpha)$ cooriented, a path/loop of contactomorphisms
$\{\psi_t\}$ is \emph{positive} if $\alpha (\partial \psi/\partial t) > 0$ everywhere.
By definition, this is equivalent to the Hamiltonian $H$ which generates this path/loop being everywhere positive.
The $\mathbb{S}^1$--action on $M$ is said positive if the loop of contactomorphisms it defines is positive.
\begin{corollary}
If the $\mathbb{S}^1$--action is positive then $\psi$ from Theorem \ref{thm:contact} is generated by a positive path.
\end{corollary}
\begin{proof}
This follows from a general fact for contact manifolds that admit a positive loop.
The group of contactomorphisms is locally path connected (see \cite[Chapter 6]{banyaga}) hence $\mathrm{Cont}_0(M, \xi)$ is path connected.
Let $\{\varphi_t\}_{t = 0}^1$ be a path of contactomorphisms from $\varphi_0 = \mathrm{id}$ to $\varphi_1 = \psi$.
Denote by $\{R_{t}\}_{t=0}^1$ the loop of contactomorphisms given by the action.
A computation (see \cite{casalspresas})
shows that, if $H, G: M \times [0,1] \to \mathds{R}$ are the Hamiltonians which generate $\{\varphi_t\}$ and $\{R_t\}$,
the composition path $\{\varphi_t \circ R_{mt}\}_{t = 0}^1$ is generated by the Hamiltonian
$$F(p, t) = H(p, t) + m e^{-f_t} G(\varphi_t^{-1}(p), t),$$
where $\varphi_t^* \alpha = e^{f_t} \alpha$ and $m \ge 1$ is arbitrary. Incidentally, notice that $G$ is independent of $t$.
Then, since $G$ is strictly positive, if $m \ge 1$ is sufficiently large, $F > 0$.
\end{proof}
It will be obvious from the version of the conjugation method used in this note that:
\begin{corollary}\label{cor:id}
The diffeomorphisms constructed in Theorems \ref{thm:symplectic} and \ref{thm:contact} can be chosen $C^{\infty}$--close
to the identity.
\end{corollary}
\textbf{Examples}
{\bf 1. Prequantum contact manifolds.}
By a result of Thomas \cite{thomas}, a contact manifold admits a locally free $\mathbb{S}^1$--action if and only if
it is the prequantization $\mathbb{S}^1$--bundle over a symplectic orbifold (also called circle orbibundle).
{\bf 2. Manifolds not admitting a positive $\mathbb{S}^1$--action.}
The spherical cotangent bundle, $\mathbb{S} T^* M$, of a manifold $M$ (also referred to as oriented projectivization of $T^* M$)
carries a canonical cooriented contact structure. There do not exist positive paths of Legendrian isotopies connecting
different fibers of $\mathbb{S} T^* M$ provided the universal cover of $M$ is $\mathds{R}^n$ (see \cite{colin}) or, more generally,
an open manifold (see \cite{chernov}).
As a consequence, $\mathbb{S} T^* M$ does not admit positive loops.
However, it may admit $\mathbb{S}^1$--actions: any $\mathbb{S}^1$--action on $M$ by diffeomorphisms lifts to a $\mathbb{S}^1$--action on $\mathbb{S} T^* M$ by contactomorphisms so Theorem \ref{thm:contact} can be applied.
This is the case, for instance, of standard $\mathds{T}^3$ as it is the spherical cotangent bundle over $\mathds{T}^2$.
{\bf 3. Symplectic examples.}
If $M$ admits a symplectic free $\mathbb{S}^1$--action,
then it admits a fibration $H\colon M \to \mathbb{S}^1$ such that the action preserves its fibers. Moreover,
there is a $C^{\infty}$-small perturbation $\widetilde{\omega}$ of the original symplectic form for which the action is still symplectic
and such that $\ker (\iota_{X} \widetilde{\omega}) = \ker (dH)$,
where $X$ denotes the infinitesimal generator of the action.
Consequently, there exists a symplectic fibration $\widehat{H}\colon M/\mathbb{S}^1 \to \mathbb{S}^1$.
Partial converse results are available \cite{marisafdez}.
\section*{Acknowledgements}
The authors express their gratitude to Viktor Ginzburg and Daniel Peralta for their useful comments during
the preparation of the manuscript.
The authors have been supported by the Spanish Research Projects SEV-2015-0554 and MTM2013-42135 and by the
ERC Starting Grant 335079 from Daniel Peralta.
\section{Conjugation method}\label{sec:conjugationmethod}
The conjugation method was introduced by Anosov and Katok \cite{anosovkatok} in 1970.
They constructed ergodic diffeomorphisms on every closed manifold as limits of volume preserving periodic transformations.
Later, Fathi and Herman \cite{fathiherman} developed the method to prove the existence of
minimal and uniquely ergodic volume preserving diffeomorphisms in manifolds which admit a locally free $\mathbb{S}^1$--action.
Katok had previously announced \cite{katokrussian} the theorem provided the action is free.
This section, following the approach of \cite{fathiherman}, presents how the conjugation method is used
to find minimal and uniquely ergodic diffeomorphisms.
Given a closed manifold $N$, $\mathrm{Diff}(N)$ denotes the group of diffeomorphisms of $N$ equipped with the $C^{\infty}$--topology.
As our aim is to adapt the method to work for subgroups of $\mathrm{Diff}(N)$, an abstract subgroup $\mathcal{G}(N) < \mathrm{Diff}(N)$ is considered.
These subgroups must be defined, at least, for the manifold in consideration, later denoted $M$, and some quotients
of $M$ over the free action of a finite group.
In \cite{fathiherman}, $\mathcal{G}(N)$ was set to be equal to $\mathrm{Diff}(N)$ or $\mathrm{Diff}_{\mu}(N)$
(diffeomorphisms preserving a prescribed measure $\mu$ of positive $C^{\infty}$ density).
In this work, $\mathcal{G}(N)$ will later be set to be $\mathrm{Symp}_0(M, \omega)$ or $\mathrm{Cont}_0(M, \xi)$ and the underlying geometric structure
will be preserved by the finite quotients under consideration.
Although the previously cited subgroups are closed, this discussion does not assume $\mathcal{G}(N)$ is necessarily closed,
$\overline{\mathcal{G}(N)}$ denotes the closure of $\mathcal{G}(N)$.
The topology considered in the group of diffeomorphisms and their subsets will always be the $C^{\infty}$--topology.
Let $M$ be a closed manifold and, for simplicity, put $\mathcal{G} = \mathcal{G}(M)$.
Assume $M$ admits a locally free $\mathbb{S}^1$--action on $M$ by diffeomorphisms in $\mathcal{G}$, i.e.
a group homomorphism $\mathbb{S}^1 \rightarrow \mathcal{G}: \alpha \mapsto R_{\alpha}$ such that for every $x \in M$ the stabilizer subgroup
$S_x = \{\alpha: R_{\alpha}(x) = x\}$ is a discrete subgroup of $\mathbb{S}^1$ (in particular, it is finite).
Throughout this article, $\mathbb{S}^1$ is identified to $\mathds{R}/\mathds{Z}$.
The union of all stabilizers $S = \cup_{x \in M} S_x$ is still a finite subgroup of $\mathbb{S}^1$.
Indeed, for any $x \in M$ consider a small neighborhood $V_x$ of $S_x$ in $\mathbb{S}^1$ such that any subgroup of $\mathbb{S}^1$ contained
in $V_x$ is a subgroup of $S_x$.
By continuity one gets that $S_y \subset V_x$ in some neighborhood of $x$, so $S_y < S_x$.
The conclusion follows from the compactness of $M$.
The goal is to prove the existence of elements of $\overline{\mathcal{G}}$ satisfying certain dynamical properties. Let $\mathcal{A}$ be the subset of
$\overline{\mathcal{G}}$ composed of such elements. The conjugation method aims to prove that $\mathcal{A}$ is not empty by showing that
the elements of $\mathcal{A}$ are generic in the subgroup
\begin{equation}\label{eq:Cdefinition}
\mathcal{C} = \overline{\{g R_{\alpha} g^{-1}: \alpha \in \mathbb{S}^1, g \in \mathcal{G}\}}
\end{equation}
of $\overline{\mathcal{G}}$, that is, $\mathcal{A} \cap \mathcal{C}$ is a dense $G_{\delta}$ subset of $\mathcal{C}$
(intersection of countably many open and dense subsets).
Corollary \ref{cor:id} is then a trivial consequence of this approach.
Recall that since $\mathrm{Diff}(M)$ is a Baire space and $\mathcal{C}$ is closed in $\overline{\mathcal{G}}$ and so in $\mathrm{Diff}(M)$, $\mathcal{C}$ is also a Baire space.
Thus, it is enough to find a countable family $\{A_j\}_{j \in J}$ of open and dense subsets of $\mathcal{C}$
such that
$$\bigcap_{j \in J} A_j = \mathcal{A} \cap \mathcal{C}.$$
In the proofs below, a family of open and dense subsets $\{A_i\}_{i \in I}$ of $\mathcal{C}$ is defined.
It has a countable coinitial subfamily $\{A_j\}_{j \in J}$, that is, for every $i \in I$ there is $j \in J$
such that $A_j \subset A_i$.
Consequently, $\cap_{i \in I} A_i = \cap_{j \in J} A_j$.
Additionally, for any $g \in \mathcal{G}$ and $i \in I$, $g A_i g^{-1} \subset A_{k}$ for some $k \in I$.
In order to prove the density of $A_i$ in $\mathcal{C}$, one needs to show that every element $g R_{\alpha} g^{-1}$
belongs to the closure of $A_i$. In view of the previous properties on the family $\{A_i\}_{i \in I}$,
it is enough to show that $R_{\alpha} \in \overline{A_i}$ for any $\alpha \in \mathbb{S}^1$ and $i \in I$.
A number $\alpha \in \mathds{R}/\mathds{Z}$ is said to be \emph{coprime} with $S$ if $S \cap \langle \alpha \rangle = \{0\}$,
where $\langle \alpha \rangle$ denotes the subgroup of $\mathbb{S}^1 \cong \mathds{R}/\mathds{Z}$ generated by $\alpha$.
Since the set $S$ of stabilizers is finite and $\mathds{Q}/\mathds{Z}$ is dense in $\mathbb{S}^1$, it suffices to check that
$R_{\alpha} \in \overline{A_i}$ for $\alpha \in \mathds{Q}/\mathds{Z}$ coprime with $S$ in order to prove it for all $\alpha \in \mathds{R}/\mathds{Z}$.
\subsection{Minimal diffeomorphisms}
This subsection shows how the conjugation method was applied in \cite{fathiherman} to obtain diffeomorphisms
with all their orbits dense.
\begin{definition}
A homeomorphism $f \colon X \to X$ is said to be minimal if every orbit is dense, i.e.
$M = \overline{\{f^n(x): n \in \mathds{Z}\}}$ for every $x \in X$.
\end{definition}
Minimality is a topological property which is easily seen to be equivalent to the non-existence of proper
closed subsets invariant under the dynamics.
\begin{lemma}\label{lem:minimalequivalence}
Let $X$ be a compact topological space.
A homeomorphism $f\colon X \to X$ is minimal if and only if for every open set $U \subset X$ there exists $n \in \mathds{N}$
such that
$$U \cup f(U) \cup \ldots \cup f^n(U) = X.$$
\end{lemma}
\begin{proof}[Sketch of the proof.]
Notice that the complement of $\cup_{m \in \mathds{Z}} f^m(U)$ in $X$ is closed and $f$--invariant, it must be empty if $f$ is minimal.
The open cover $\{f^m(U)\}_{m \in \mathds{Z}}$ of $X$ has, by compactness, a finite subcover whose elements, after applying $f$ conveniently,
are forward iterates of $U$.
\end{proof}
Let us apply the conjugation method to obtain minimal diffeomorphisms in $\overline{\mathcal{G}}$.
Denote by $\mathcal{M}$ the set of minimal diffeomorphisms. Consider the sets
$$W_{U, n} = \{g \in \mathcal{G}: U \cup g(U) \cup \ldots \cup g^n(U) = M\},$$
where $U$ ranges over the open subsets of $M$ and $n$ over the positive integers.
By definition, $W_{U,n}$ is an open subset of $\mathrm{Diff}(M)$.
Consequently, $\mathcal{M}_U = \overline{\mathcal{G}} \, \cap \, \left(\cup_{n \ge 1} W_{U, n}\right)$ is an open subset of $\overline{\mathcal{G}}$.
By Lemma \ref{lem:minimalequivalence}, $\mathcal{M}$ is the intersection of all open sets $\mathcal{M}_U$.
Since $M$ has a countable basis of open sets $\{U_i\}_{i\in \mathds{N}}$, the subfamily $\{\mathcal{M}_{U_i}\}_{i \in \mathds{N}}$ is coinitial
and $\mathcal{M} = \cap_i \mathcal{M}_{U_i}$. Moreover, $g \mathcal{M}_U g^{-1} = \mathcal{M}_{g^{-1}(U)}$.
Thus, in order to check that $\mathcal{M}_U$ is dense in $\mathcal{C}$ it is enough to prove that $R_{\alpha}$ is accumulated by elements
of $\mathcal{M}_U$ for every $\alpha \in (\mathds{Q}/\mathds{Z}) \setminus S$.
Given any $\alpha = p/q \notin S$, $\mathrm{gcd}(p,q) = 1$, $F_q$ denotes the subgroup of $\mathbb{S}^1$ generated by $p/q$.
The quotient of $M$ under the action of $F_q$ is a manifold $\widehat{M} = M/F_q$. The $\mathbb{S}^1$--action on $M$ induces,
by the identification $\mathbb{S}^1/F_q \cong \mathbb{S}^1$, another locally free $\mathbb{S}^1$--action on $\widehat{M}$.
This action is given by $\{\widehat{R}_{\beta}: \beta \in \mathbb{S}^1\}$, where $\widehat{R}_{q\alpha}$
is the projection of $R_{\alpha}$ onto $\widehat{M}$.
The subgroup $\widehat{\mathcal{G}} := \mathcal{G}(\widehat{M}) < \mathrm{Diff}(\widehat{M})$ must satisfy the following hypothesis:
\medskip
\textbf{(H1)}
Any element of $\widehat{\mathcal{G}}$ has a lift in $\mathcal{G} = \mathcal{G}(M)$.
\medskip
Next condition must be also fulfilled. These hypothesis are discussed in Section \ref{sec:proofs}.
Note that it is an exercise to check them in the case $\mathcal{G}(N) = \mathrm{Diff}(N)$ or $\mathcal{G}(N) = \mathrm{Diff}_{\mu}(N)$,
where $\mu$ is a probability measure on $N$ of positive $C^{\infty}$ density.
\medskip
\textbf{(H2)} For any locally free $\mathbb{S}^1$--action on a manifold $N$, and any open $V \subset N$, there exists an
element $f \in \mathcal{G}(N)$ such that $f^{-1}(V)$ meets all the orbits of the action.
\medskip
Fix an open set $U \subset M$. For the previous $\alpha = p/q$, denote by $\widehat{U}$ the projection of $U$ onto $\widehat{M}$
and let $\widehat{h}$ be the map supplied by (H2) for $N = \widehat{M}$,
$V = \widehat{U}$ and the induced $\mathbb{S}^1$--action on $\widehat{M}$.
There is a lift $h \in \mathcal{G}$ of $\widehat{h}$ whose existence is guaranteed by (H1). It satisfies
$h R_{\alpha} h^{-1} = R_{\alpha}$ and $h^{-1}(U)$ meets all the orbits of the $\mathbb{S}^1$--action on $M$ because $\alpha$
is coprime with $S$.
The following lemma generalizes to $\mathbb{S}^1$--actions a simple fact:
the iterates of an open interval under an irrational rotation eventually cover $\mathbb{S}^1$.
\begin{lemma}\label{lem:irrationalminimal}
For any $\beta \notin \mathds{Q}/\mathds{Z}$, $h R_{\beta} h^{-1} \in \mathcal{M}_U$.
\end{lemma}
\begin{proof}
Let $\gamma$ denote a orbit of the action and $W$ an open subset of $M$, $\gamma \cap W \neq \emptyset$.
There exists $n \ge 1$ such that $W \cup R_{\beta}(W) \cup \ldots \cup R_{n\beta}(W)$ cover $\gamma$.
The same is true for orbits close to $\gamma$.
Take $W = h^{-1}(U)$, then $W$ meets every orbit.
Since the orbit space of the action is compact, there exists $m \ge 1$ such that
$W \cup R_{\beta}(W) \cup \ldots \cup R_{\beta}^m(W)$ contains every orbit $\gamma$.
Thus $h R_{\beta} h^{-1} \in W_{U, m} \cap \mathcal{C} \subset \mathcal{M}_U$.
\end{proof}
Take a sequence of irrational numbers $\beta_n \to \alpha$.
Clearly, $h R_{\beta_n} h^{-1} \xrightarrow{\mathcal{C}^{\infty}} h R_{\alpha} h^{-1} = R_{\alpha}$.
Since from Lemma \ref{lem:irrationalminimal}, $h R_{\beta_n} h^{-1}$ belongs to $\mathcal{M}_U$ and also, by definition, to $\mathcal{C}$,
the map $R_{\alpha} \in \overline{\mathcal{M}_U}$ and the conclusion follows.
In conclusion, the existence of a minimal diffeomorphism in $\overline{\mathcal{G}}$ is guaranteed provided (H1) and (H2) are satisfied.
\subsection{Strictly ergodic diffeomorphisms}
In this subsection, the conjugation method is applied to obtain a diffeomorphism
whose orbits are uniformly distributed along $M$ in a measure theory sense.
\begin{definition}
Let $X$ be a compact metric space. A homeomorphism $f\colon X \to X$ is called uniquely ergodic if it has only one
invariant probability measure. If the invariant measure has full support $f$ is called strictly ergodic.
\end{definition}
Next lemma follows from two dynamical facts: the support of an $f$--invariant measure is itself invariant under $f$
and any compact invariant subset of $X$ admits an invariant measure supported on it.
\begin{lemma}
A map is strictly ergodic if and only if it is uniquely ergodic and minimal.
\end{lemma}
Once strict ergodicity has been split into two properties, let us take care of unique ergodicity.
\begin{proposition}\label{prop:uecharacterization}
Let $X$ be a compact metric space and $f\colon X \to X$ be a homeomorphism. The following statements are equivalent.
\begin{itemize}
\item $f$ is uniquely ergodic.
\item For every map $u \in C^0(X, \mathds{R})$,
$\frac{1}{n} \left(\sum_{k = 0}^{n-1} u \circ f^k\right)$ converges uniformly
as $n \to \infty$ to a constant (equal to $\int_X u \, d\mu$,
where $\mu$ is the only invariant probability measure).
\end{itemize}
\end{proposition}
\begin{proof}
See Walters \cite[Theorem 6.19]{walters}
\end{proof}
Since the space of real--valued continuous functions over $M$ is separable, Proposition \ref{prop:uecharacterization}
can be used to show that the set of uniquely ergodic diffeomorphisms $\mathcal{E}$ is a $G_{\delta}$ subset of $\mathrm{Diff}(M)$.
Indeed, let $u \in C^0(X, \mathds{R})$, $\varepsilon > 0$. Define
$$\mathcal{E}(u, \varepsilon) = \biggl\{f \in \mathcal{C}: \exists \, n \ge 1, c \in \mathds{R} \text{ such that }
\biggl|\biggl| \frac{1}{n} \sum_{k = 0}^{n-1} u \circ f^k - c\biggr|\biggr| < \varepsilon \biggr\},$$
where $\mathcal{C}$, the closure of the set of conjugates of the action, was defined in (\ref{eq:Cdefinition}).
Trivially, the sets $\mathcal{E}(u, \varepsilon)$ are open.
Fix a dense sequence $\{u_i\}$ in $C^0(X, \mathds{R})$.
\begin{lemma}
$$\mathcal{E} \cap \mathcal{C} = \bigcap_i \bigcap_{k \ge 1} \mathcal{E}(u_i, 1/k).$$
\end{lemma}
\begin{proof}
Simply notice that if $\varepsilon_1 > \varepsilon_0 > 0$ and $\Vert u_1 - u_0 \Vert < \varepsilon_1 - \varepsilon_0$
then $\mathcal{E}(u_0, \varepsilon_0) \subset \mathcal{E}(u_1, \varepsilon_1)$.
\end{proof}
Furthermore, note that $g \, \mathcal{E}(u, \varepsilon) \, g^{-1} = \mathcal{E}(u \circ g, \varepsilon)$.
Thus, in order to show that $\mathcal{E}$ is dense in $\mathcal{C}$ it suffices to check that $R_{\alpha}$ belongs to
$\overline{\mathcal{E}(u, \varepsilon)}$ for every rational $\alpha$ not contained in $S$.
Henceforth, assume $u \in C^0(M, \mathds{R})$ and $\varepsilon > 0$ are fixed and the following hypothesis is satisfied.
\medskip
\textbf{(H3)}
Given a locally free $\mathbb{S}^1$--action $\{R_{\alpha}\}_{\alpha \in \mathbb{S}^1}$ on a manifold $N$, an open set $V \subset N$
and $\varepsilon > 0$, there exists an element $f \in \mathcal{G}(N)$ such that
$$\lambda(\{\alpha \in \mathbb{S}^1: R_{\alpha}(y) \notin f^{-1}(V)\}) < \varepsilon$$
for every $y \in N$ ($\lambda$ denotes the Lebesgue measure in $\mathbb{S}^1$ of total mass equal to 1).
\medskip
There are some unavoidable bothering technical issues to check this condition
(see \cite[Section 6]{fathiherman}).
As in the minimal case, fix $\alpha = p/q \notin S$, $\gcd(p, q) = 1$, and write $F_q = \langle \alpha \rangle$. Define
$$\widehat{u} = \frac{1}{q} \sum_{k = 0}^{q-1} u \circ R_{k/q}.$$
Clearly, $\widehat{u}$ can be seen as a function on the quotient $\widehat{M} = M/F_q$.
Consider an arbitrary $\eta > 0$ and fix $\delta > 0$ such that $\delta (1 + ||u||) < \eta$.
Take $y_0 \in \widehat{M}$ and note that
$\widehat{U} = \{y \in \widehat{M}: |\widehat{u}(y) - \widehat{u}(y_0)| < \delta\}$
is a non--empty open subset of $\widehat{M}$.
Apply (H3) to find $\widehat{h} \in \widehat{\mathcal{G}}$ such that for every $y \in \widehat{M}$,
$\lambda(\{\alpha \in \mathbb{S}^1: \widehat{R}_{\alpha}(y) \notin \widehat{h}^{-1}(U)\}) < \delta$,
where $\widehat{R}$ denotes the $\mathbb{S}^1$--action on $\widehat{M}$ induced by $R$. Then,
\[
\left|\int_{\mathbb{S}^1} \widehat{u} \circ \widehat{h} \circ \widehat{R}_{\beta}(y) d\beta - \widehat{u}(y_0)\right| \le
\delta + \delta ||\widehat{u}|| \le \delta (1 + ||u||) < \eta.
\]
for every $y \in \widehat{M}$.
Choose a lift $h \in \mathcal{G} = \mathcal{G}(M)$ of $\widehat{h}$, whose existence is guaranteed by (H1). Then,
$h \circ R_{\alpha} = R_{\alpha} \circ h$, and, for every $x \in M$,
$$\int_{\mathbb{S}^1} u \circ h \circ R_{\theta}(x) d\theta =
\int_{\mathbb{S}^1} \widehat{u} \circ \widehat{h} \circ \widehat{R}_{\beta}(\widehat{x}) d\beta,$$
where $\widehat{x}$ denotes the projection of $x$ onto $\widehat{M}$.
In sum, we have proved the following lemma.
\begin{lemma}\label{lem:averagingmap}
For any $\eta > 0$, there exists $h \in \mathcal{G}$ and a constant $c \in \mathds{R}$ such that
$$\left| \int_{\mathbb{S}^1} u \circ h \circ R_{\theta}(x) d\theta - c\right| < \eta$$
holds for every $x \in M$.
\end{lemma}
\begin{proposition}
$R_{\alpha} \in \overline{\mathcal{E}(u, \varepsilon)}$.
\end{proposition}
\begin{proof}
Since an irrational rotation is uniquely ergodic, for any $\beta \notin \mathds{Q}/\mathds{Z}$ and $x \in M$
$$\lim_{n \to \infty} \frac{1}{n} \sum_{k = 0}^{n-1} v \circ R_{\beta}^k(x) = \int_{\mathbb{S}^1} v \circ R_{\theta} d\theta$$
holds for every $v \in C^0(M, \mathds{R})$.
Furthermore, the convergence is uniform in $x \in M$.
In particular, if $h$ comes from Lemma \ref{lem:averagingmap}
\[
\lim_{n \to \infty} \frac{1}{n} \sum_{k = 0}^{n-1} u \circ h \circ R_{\beta}^k \left( h^{-1} (x) \right) =
\int_{\mathbb{S}^1} (u \circ h) \circ R_{\theta}(h^{-1}(x)) d\theta.
\]
and the convergence is uniform.
If $\eta$ from Lemma \ref{lem:averagingmap} is chosen sufficiently small and $n$ is large
\begin{align*}
\left\Vert \frac{1}{n} \sum_{k = 0}^{n-1} u \circ \left( h R_{\beta}^k h^{-1} \right) - c \right\Vert &
\le \left\Vert \frac{1}{n} \sum_{k = 0}^{n-1} u \circ h \circ R_{\beta}^k h^{-1} -
\int_{\mathbb{S}^1} u \circ h \circ R_{\theta} d\theta \right\Vert + \\
& + \left\Vert \int_{\mathbb{S}^1} u \circ h \circ R_{\theta} d\theta - c \right\Vert < \varepsilon,
\end{align*}
so $h R_{\beta} h^{-1} \in \mathcal{E}(u, \varepsilon)$.
The conclusion is obtained taking a sequence of irrational $\beta_n \to \alpha$ because
$h R_{\beta_n} h^{-1} \xrightarrow{\mathcal{C}^{\infty}} h R_{\alpha} h^{-1} = R_{\alpha}$.
\end{proof}
In conclusion, $\overline{\mathcal{G}}$ contains strictly ergodic diffeomorphisms as long as it satisfies (H1) and (H3).
\section{Proofs of main theorems}\label{sec:proofs}
The strategy of the proofs of Theorems \ref{thm:symplectic} and \ref{thm:contact} is the
conjugation method which was explained in Section \ref{sec:conjugationmethod}.
There have been some marked points in the arguments, where hypothesis on the diffeomorphisms subgroups $\mathcal{G}$
were assumed and must be checked separately.
Theorem \ref{thm:symplectic} follows once it is shown that (H1) and (H2) are valid for the connected component of the
group of symplectic diffeomorphisms which contains the identity.
Analogously, to prove Theorem \ref{thm:contact} (H1) and (H3) must be shown to hold true for
the connected component of the identity in the group of contactomorphisms.
Notice, incidentally, that (H3) implies (H2).
\begin{definition}
The action of a group $G < \mathrm{Diff}(M)$ is said to be $n$--transitive if for every $x_1, \ldots, x_n$ and
$y_1, \ldots, y_n$ there is $g \in G$ such that $g(x_i) = y_i$, $i = 1\ldots n$.
\end{definition}
The following result is well-known, see Boothby \cite{boothby} for a detailed proof.
\begin{theorem}\label{thm:transitivity}
The group of symplectic/contact diffeomorphisms acts $n$--transitively for any $n \ge 1$.
Furthermore, the same is true for the group of Hamiltonian symplectomorphisms and Hamiltonian contactomorphisms.
\end{theorem}
\medskip
\textbf{Hypothesis (H1):}
\emph{Any element of $\widehat{\mathcal{G}} = \mathcal{G}(\widehat{M})$ has a lift in $\mathcal{G} = \mathcal{G}(M)$.}
\medskip
This is obviously true for contactomorphisms and symplectomorphisms.
Recall that $\widehat{M}$ was defined as the quotient of $M$ under the free action of the group generated
by $R_{p/q}$.
\bigskip
\textbf{Hypothesis (H2):}
\emph{Given any locally free $\mathbb{S}^1$--action on a manifold $N$, and any open $V \subset N$, there exists an
element $f \in \mathcal{G}(N)$ such that $f^{-1}(V)$ meets all the orbits of the action.}
\medskip
It will be now checked that this hypothesis is verified in the case
$(N^{2n}, \omega)$ is a symplectic manifold, the action is symplectic and $\mathcal{G}(N) = \mathrm{Symp}_0(N, \omega)$.
A Darboux flow box for a symplectic vector field $X$ is a Darboux chart $(U_i, \theta_i)$,
$$\theta_i \colon U_i \to Q(\delta, \rho) := [-\delta, \delta] \times [-\rho, \rho] \times B^2(\rho) \times \ldots B^2(\rho) \subset \mathds{R}^{2n}$$ such that $\theta^*_i \bigl(\frac{\partial}{\partial x_1} \bigr) = X$.
Henceforth, $X$ will be the infinitesimal generator of the $\mathbb{S}^1$--action.
Darboux flow boxes do exist at any point of $M$.
The next statement follows from compactness of $M$.
\begin{lemma}\label{lem:flowboxes}
There exist $\varepsilon, r > 0$ and $\{(U_i, \vartheta_i)\}_{i = 1}^m$
a finite set of pairwise disjoint Darboux flow boxes such that
$\vartheta_i: U_i \to Q(\varepsilon, r + \varepsilon)$ and the union of the codimension--1 disks
$$\bigsqcup_{i=1}^m \vartheta_i^{-1}(\{0\}\times [-r,r] \times B^2(r) \times \ldots \times B^2(r))$$
touches all the orbits of $X$.
\end{lemma}
This lemma allows to work in a local fashion in $(\mathds{R}^{2n}, \omega_0)$ so as to apply the following
squeezing result, which will be proved in Section \ref{sec:packing}.
\begin{proposition}\label{prop:folding}
Let $r > 0$, $D = \{0\} \times [-r, r] \times B^2(r) \times \ldots \times B^2(r) \subset \mathds{R}^{2n}$
and $\varepsilon > 0$. For any $\delta > 0$, there exists a Hamiltonian symplectomorphism $\psi$
with support in
$Q(\varepsilon, r) = [-\varepsilon, \varepsilon] \times [-r-\varepsilon, r + \varepsilon] \times B^2(r + \varepsilon) \times
\ldots B^2(r + \varepsilon)$ such that $\psi(D) \subset P^{2n}(\delta, \ldots, \delta)$.
\end{proposition}
Apply Lemma \ref{lem:flowboxes} to obtain $r, \varepsilon > 0$ and a family of Darboux flow boxes $\{(U_i, \vartheta_i)\}_{i =1}^m$.
By Theorem \ref{thm:transitivity}, there is a Hamiltonian symplectomorphism $\phi$ such that $\phi^{-1}(V)$ contains the centers
$\vartheta_i^{-1}(\textbf{0})$ of the flow boxes.
Let $\psi$ be the squeezing map given by Proposition \ref{prop:folding}. Define a Hamiltonian
symplectomorphism $\varphi$ in $N$ which is equal to $\vartheta_i \circ \psi \circ \vartheta_i^{-1}$ in
$U_i$, for any $1 \le i \le m$, and to the identity elsewhere.
Then $(\phi \circ \varphi)^{-1}(V)$ meets all the orbits of the $\mathbb{S}^1$--action.
\begin{remark}
Note that the map $\phi \circ \varphi$ in $\mathrm{Symp}_0(N)$ realizing (H2) is actually a Hamiltonian symplectomorphism.
Ultimately, this boils down to the fact that both $n$--transitivity and Proposition \ref{prop:folding} are realized
by Hamiltonian symplectomorphisms.
\end{remark}
\bigskip
\textbf{Hypothesis (H3):}
\emph{Given a locally free $\mathbb{S}^1$--action $\{R_{\alpha}\}_{\alpha \in \mathbb{S}^1}$ on a manifold $N$, an open set $V \subset N$
and $\varepsilon > 0$, there exists an element $f \in \mathcal{G}(N)$ such that
$$\lambda(\{\alpha \in \mathbb{S}^1: R_{\alpha}(y) \notin f^{-1}(V)\}) < \varepsilon$$
for every $y \in N$ ($\lambda$ denotes the Lebesgue measure in $\mathbb{S}^1$ of total mass 1).}
\medskip
As was said before, there are some technicalities difficulties to overcome in the proof even in the case
$\mathcal{G}(N) = \mathrm{Diff}(N)$ or $\mathrm{Diff}_{\mu}(N)$ (see \cite{fathiherman}). Notice also that (H3) is out of reach
in the symplectic case. Indeed, the volume is always preserved under symplectic transformations so
it is not possible to modify $V$ to swallow ``most'' of all the orbits.
Henceforth, assume that $(N^{2n+1}, \xi)$ is a contact manifold, $R_{\alpha}$ is a contactomorphism
for every $\alpha \in \mathbb{S}^1$ and $\mathcal{G}(N) = \mathrm{Cont}_0(N, \xi)$.
Let $\{(U_j, \theta_j\colon U_j \to \theta_j(U_j) \subset \mathds{R}^{2n+1})\}$ be a finite atlas of $N$ by Darboux charts,
that is, $\theta_j\colon (U_j, \xi_{|U_j}) \to (\mathds{R}^{2n+1}, \xi_{0})$ is a contactomorphism.
The standard contact distribution in $\mathds{R}^{2n+1}$ is denoted $\xi_0 = \ker(dz - \sum_{k=1}^n y_k dx_k)$.
The following packing result is well--known in the field of contact geometry,
a proof accessible to non--experts is presented in Section \ref{sec:packing}.
The term \emph{cuboid} is here used to name the closure of a domain in $\mathds{R}^{2n+1}$ enclosed by
$2n+1$ couples of parallel hyperplanes in general position.
\begin{lemma}\label{lem:contactinflating}
Let $p \in \mathds{R}^{2n+1}$, $V_p$ a neighborhood of $p$ and $Y$ be a vector field in $V_p$ such that $Y(p) \neq 0$.
There exists a cuboid $C$ centered at $p$ such that
\begin{itemize}
\item $C$ is contained in $V_p$,
\item $Y$ is transverse to the faces of $C$ and
\item for any neighborhood $W$ of $\partial C$ and any ball $B \subset C$ centered at $p$
it is possible to find a Hamiltonian contactomorphism $\varphi \colon (\mathds{R}^{2n+1}, \xi_0) \to (\mathds{R}^{2n+1}, \xi_0)$ such that:
\begin{enumerate}
\item $\mathrm{supp}(\varphi) \subset C$.
\item $\varphi(B)$ contains $C \setminus W$.
\end{enumerate}
\end{itemize}
\end{lemma}
For every $j$ and every $x \in U_j$
apply Lemma \ref{lem:contactinflating} to $\theta_j(x)$ and the vector field $(\theta_j)_*(X)$
to obtain a cuboid $Q^j_x$, centered at $x$, which is further assumed to lie within $\theta_j(U_j)$.
Define $C^j_x = \theta_j^{-1}(Q^j_x)$ and note that $X$ is transversal to $\partial C^j_x$.
Since $N$ is compact and $\{\mathrm{int}(C^j_x): x \in U_j\}$ is an open cover of $N$,
there is a finite set of different points $\{p_i\}_{i = 1}^m$ whose associated $C_{p_i}$ cover $N$.
By transversality, any orbit of the action meets $\Delta = \bigcup_{i = 1}^m \partial C_{p_i}$ at a finite number of points.
\begin{lemma}\label{lem:skeleton}
If $W$ is a sufficiently small neighborhood of $\Delta$ then
$$\lambda(\{\alpha \in \mathbb{S}^1: R_{\alpha}(y) \in W\}) < \varepsilon$$
for every $y \in N$.
\end{lemma}
\begin{proof}
For any $y_0 \in N$, the orbit $\mathcal{O}(y_0) = \{R_{\alpha}(y_0): \alpha \in \mathbb{S}^1\}$ meets $\Delta$ in finitely many points,
$R_{\alpha_1}(y_0), \ldots, R_{\alpha_k}(y_0)$. Take a neighborhood $U$ of the union of these points
small enough so that $\lambda(\{\alpha \in \mathbb{S}^1: R_{\alpha}(y) \in U\}) < \varepsilon$ for every $y \in N$.
Clearly, if $W^{y_0}$ is a sufficiently small neighborhood of $\Delta$, $W^{y_0} \cap \mathcal{O}(y_0) \subset U$
and the same is true for orbits of points in a neighborhood of $y_0$. A compactness argument yields the result.
\end{proof}
Fix $W$ from the previous lemma and note that to conclude (H3) it is enough to inflate $V$ to cover $N \setminus W$.
This strategy splits into two steps. Firstly, choose $s$ different points $\{q_i\}$ in $V$. By Theorem \ref{thm:transitivity}
there is a Hamiltonian contactomorphism $\phi\colon (N, \xi) \to (N, \xi)$ such that $\phi(p_i) = q_i$.
Then, $\{p_i\} \subset \phi^{-1}(V)$.
Let $B_i$ be a small ball centered at $p_i$ and contained in $\phi^{-1}(V)$.
Recall from Lemma \ref{lem:contactinflating} the properties of $Q^i_x$, which are inherited by
$C_{p_i}$. In particular, there are Hamiltonian contactomorphisms $\varphi_i\colon (N, \xi) \to (N, \xi)$
supported in $C_{p_i}$ such that $\varphi_i(B_i) \supset C_{p_i} \setminus W$.
Consider
$\varphi = \varphi_1 \circ \ldots \circ \varphi_m.$
\begin{lemma}
$$\varphi \left( \cup_{i = 1}^s B_{i} \right) \cup W = N.$$
\end{lemma}
\begin{proof}
For a point $p$ in the support of $\varphi_i$ there are two non--exclusive possibilities:
$\varphi_i(p) \in W$ or $p \in B_{i}$. The conclusion follows from the fact that any point of $N$
belongs to some $C_{p_i}$ hence to the support of one or more $\varphi_i$.
\end{proof}
As a consequence, $\varphi \circ \phi^{-1} (V)$ contains $N \setminus W$. Thus, Lemma \ref{lem:skeleton} concludes (H3).
\begin{remark}
Notice that the map $\varphi \circ \psi$ is the composition of two Hamiltonian contactomorphisms.
\end{remark}
\section{Packing lemmas}\label{sec:packing}
This section discusses the two packing results (Proposition \ref{prop:folding} for (H2) in the symplectic case and
Lemma \ref{lem:contactinflating} for (H3) in the contact case) which eventually led to the proofs of the main theorems.
\subsection{Contact}
In the contact case, the goal is to construct a ``box'' within which any small ball may be inflated
(by a contact transformation) to take up all the space inside but a small margin. This is an easy consequence
of the basic fact that special dilations preserve the standard contact structure.
\begin{lemmacontactinflating}
Let $p \in \mathds{R}^{2n+1}$, $V_p$ a neighborhood of $p$ and $Y$ be a vector field in $V_p$ such that $Y(p) \neq 0$.
There exists a cuboid $C$ centered at $p$ such that
\begin{itemize}
\item $C$ is contained in $V_p$,
\item $Y$ is transverse to the faces of $C$ and
\item for any neighborhood $W$ of $\partial C$ and any ball $B \subset C$ centered at $p$
it is possible to find a Hamiltonian contactomorphism $\varphi \colon (\mathds{R}^{2n+1}, \xi_0) \to (\mathds{R}^{2n+1}, \xi_0)$ such that:
\begin{enumerate}
\item $\mathrm{supp}(\varphi) \subset C$.
\item $\varphi(B)$ contains $C \setminus W$.
\end{enumerate}
\end{itemize}
\end{lemmacontactinflating}
\begin{proof}
If $p = (\textbf{x}_0, \textbf{y}_0, z_0)$, the affine map
\[
\tau(\textbf{x}, \textbf{y}, z) = (\textbf{x} - \textbf{x}_0, \textbf{y} - \textbf{y}_0, z - z_0 + \textbf{x}_0 \cdot (\textbf{y} - \textbf{y}_0))
\]
is a contactomorphism in $(\mathds{R}^{2n+1}, \xi_{0})$ which maps $p$ to the origin.
Thus, assume without lose of generality $p = \textbf{0}$.
The vector field $V(\textbf{x}, \textbf{y}, z) = (\textbf{x}, \textbf{y}, 2z)$ is a contact vector field because its flow
$\psi_t(\textbf{x}, \textbf{y}, z) = (e^t\textbf{x}, e^t\textbf{y}, e^{2t}z)$ preserves the standard contact structure $\xi_{0}$.
Denote $H_V$ the contact Hamiltonian associated to $V$.
Let $C_r$ be the cuboid of size $r > 0$ centered at $\textbf{0}$ and
generated by the set of linear 1--forms $\{\lambda_1 = dx_1, \ldots, \lambda_{2n+1} = dz\}$, that is,
\[
C_r = \{v \in \mathds{R}^{2n+1} : |\lambda_i(v)| < r \enskip \forall i\}.
\]
A computation shows that $V$ points outwards $C_{r}$.
As a consequence, the image under the flow $\psi_t$ of any neighborhood $B$ of the origin eventually covers $C_{r}$.
Note that the statement remains valid as long as the vector field $V$ points outwards every such cuboid.
In case $\lambda_i(Y(\textbf{0})) = 0$, replace $\lambda_i$ by a sufficiently close linear 1--form $\tilde{\lambda}_i$.
For $r_0 > 0$ sufficiently small, $V$ still point outwards the modified cuboids
$\widetilde{C}_r = \{v \in \mathds{R}^{2n+1} : |\widetilde{\lambda}_i(v)| < r \enskip \forall i\}$
and the faces of $\widetilde{C}_r$,
$\partial \widetilde{C}_r$, are transversal to $Y$ if $r \le r_0$. Take $C = \widetilde{C}_{r_0}$.
Fix now a ball $B$ centered at \textbf{0} and a neighborhood $W$ of $\partial C$. As was noticed before,
$\psi_t(B) \subset C \setminus W$ for large $t$.
Consider now a smooth function $H\colon \mathds{R}^{2n+1} \to \mathds{R}$ equal to $H_V$ inside $C \setminus W$ which vanishes outside $C$.
The contact vector field $V'$ associated to $H$ is then equal to $V$ inside $C \setminus W$ and vanishes outside $C$.
Consequently, the flow $\varphi_t$ generated by $V'$ satisfies the properties in the statement.
\end{proof}
\subsection{Symplectic}
This subsection contains the proof of Proposition \ref{prop:folding}, that is,
it is devoted to show how to squeeze a large codimension--1 disk into a small ball in a symplectic fashion.
Denote by $B^2(r)$ the closed 2--ball of radius $r$ and $P^{2n}(r_1, \ldots, r_n) = B^2(r_1) \times \ldots \times B^2(r_n)$.
The proof presented here adapts the following non--trivial result to answer the question.
\begin{lemma}\label{lem:folding}
Given $s, \rho > 0$ there exists $\eta = \eta(s, \rho) > 0$ and a Hamiltonian symplectomorphism $\phi$
such that $\phi$ embeds $B^2(\eta) \times B^2(s)$ into $B^2(\rho) \times B^2(1)$.
Furthermore, $\phi$ can be assumed to be supported in $B^2(c \rho) \times B^2(s + c)$ for a constant $c > 1$
independent of $s, \rho$.
\end{lemma}
There exist several approaches to this result in the literature. One could use the $h$--principle for isosymplectic embeddings to
obtain an embedding of the disk $B^2(s)$ and then extend it to a neighborhood thanks to the Symplectic Neighborhood Theorem.
The parametric version of this symplectic embedding theorem (see \cite[Section 12.1]{eliashberg}) would provide the result.
However, we prefer a more hands--on approach using symplectic folding.
Following Schlenk (\cite[Remark 3.3.1]{schlenk}), the nature of this deformation is local
and each step of the construction is induced by a Hamiltonian flow.
Indeed, a careful look through the folding shows that all deformations take place inside an arbitrary neighborhood
of the figures apart from the stretching in the base, where an extra space proportional to the size of the fibers is required.
The constant $c$ in the lemma is introduced to make up for it.
For our purpose, the extra space around the codimension--1 disk in which the transformation is supported must be
arbitrarily small. The following technical lemma shows how to
scale properly Lemma \ref{lem:folding} and to fold the disk in $n-1$ directions.
\begin{lemma}\label{lem:sabana}
Given $r, \varepsilon > 0$, for every $0 < \delta < \varepsilon$, there exists $\sigma > 0$ and a Hamiltonian
symplectomorphism $\varphi$ such that $\varphi$ embeds $P^{2n}(\sigma, r, \ldots, r)$
into $P^{2n}(\delta, \ldots, \delta)$ and the support of $\varphi$ is contained in
$P^{2n}(\varepsilon, r+\varepsilon, \ldots r+\varepsilon)$.
\end{lemma}
\begin{proof}
The goal is to find $\sigma$ so that the result of folding in each of the $n-1$ ``thick'' directions the thin polydisk
is contained into the target cube. The squeezing map of Lemma \ref{lem:folding} has to be scaled properly.
Set $\rho_1 = 1$ and consider $\lambda_1 > 0$ small enough so that the following inequalities are satisfied for
$\rho = \rho_1$ and $\lambda = \lambda_1$:
\begin{equation}\label{eq:1}
\lambda, \lambda \rho < \delta, \enskip \enskip \enskip \lambda c, \lambda c \rho < \varepsilon,
\end{equation}
where $c$ comes from Lemma \ref{lem:folding}.
Define $s_1 = r / \lambda_1$. Lemma \ref{lem:folding} yields $\rho_2 := \eta(s_1,\rho_1)$ and $\phi_{1}$.
The map $\widehat{\phi}_{1}(x) = \lambda_1 \phi_{1}(x/\lambda_1)$ defines a Hamiltonian symplectomorphism
such that:
\begin{itemize}
\item $\widehat{\phi}_{1}$ embeds $B^2(\lambda_1 \rho_2) \times B^2(\lambda_1 s_1)$ into
$B^2(\lambda_1 \rho_1) \times B^2(\lambda_1)$ so,
the set of inequalities (\ref{eq:1}) implies $B^2(\lambda_1 \rho_2) \times B^2(r)$ is sent inside $P^4(\delta, \delta)$.
\item The support of $\widehat{\phi}_{1}$ is contained in $B^2(\lambda_1 c \rho_1) \times B^2(\lambda_1(s_1 + c))$
which, by (\ref{eq:1}), is in turn contained in $B^2(\varepsilon) \times B^2(r + \varepsilon)$.
\end{itemize}
Setting $\sigma = \lambda_1 \rho_2$ would already prove the lemma for $n = 2$. Let us continue
the argument pursuing the general case.
For $k \ge 2$, define inductively $\rho_k = \eta(s_{k-1}, \rho_{k-1})$,
and $\lambda_k > 0$ smaller than $\lambda_{k-1}$ and such that inequalities (\ref{eq:1})
are satisfied for $\rho = \rho_k$ and $\lambda = \lambda_k$.
Define $\widehat{\phi}_{k} = \lambda_k \phi_{k}(x/\lambda_k)$ and $s_k = r/\lambda_k$.
For every $k \ge 1$, the following properties are satisfied:
\begin{itemize}
\item $\widehat{\phi}_{k}$ embeds $B^2(\lambda_k \rho_{k+1}) \times B^2(r) =
B^2(\lambda_k \eta(s_k, \rho_k)) \times B^2(\lambda_k s_k)$ inside $B^2(\lambda_k \rho_k) \times B^2(\lambda_k)$.
\item The support of $\widehat{\phi}_{k}$ is contained in $B^2(\lambda_k c) \times B^2(\lambda_k(s_k + c))$
so, by (\ref{eq:1})
\begin{equation}\label{eq:2}
\mathrm{supp}(\widehat{\phi}_{k}) \subset B^2(\varepsilon) \times B^2(r + \varepsilon).
\end{equation}
\end{itemize}
For any $1 \le k \le n-1$, $E_k$ denotes the linear subspace of $\mathds{R}^{2n} = \mathds{R}^2 \times \ldots \times \mathds{R}^2$
spanned by the $1^{st}$ and $(n-k+1)^{th}$ factors.
Define $\widehat{\varphi}_{k}$ as the map which acts as $\widehat{\phi}_{k}$ in $E_k$ and as the identity in the other directions.
Evidently, $\widehat{\varphi}_{k}$ is again a Hamiltonian symplectomorphism.
In view of (\ref{eq:2}),
after a suitable cut-off we can obtain another Hamiltonian symplectomorphism $\varphi_k$ which coincides with $\widehat{\varphi}_k$
in $P^{2n}(\delta, r, \ldots, r)$ and is supported on $P^{2n}(\varepsilon, r + \varepsilon, \ldots, r + \varepsilon)$.
Define $\sigma = \lambda_{n-1} \rho_{n}$. Then,
\begin{equation*}
\left.\begin{aligned}
& P^{2n}(\sigma, r, r, \ldots, r) = P^{2n}(\lambda_{n-1}\rho_{n}, \lambda_{n-1}s_{n-1}, r, \ldots, r)
\xhookrightarrow{\varphi_{n-1}}\\
& \xhookrightarrow{\varphi_{n-1}} P^{2n}(\lambda_{n-1} \rho_{n-1}, \lambda_{n-1}, r, \ldots, r)\\
& P^{2n}(\lambda_{n-1} \rho_{n-1}, \lambda_{n-1}, r, \ldots, r)
\subset P^{2n}(\lambda_{n-2} \rho_{n-1}, \delta, r, \ldots, r) \xhookrightarrow{\varphi_{n-2}} \ldots \\
& \qquad \qquad \qquad \qquad \qquad \ldots \xhookrightarrow{\varphi_{2}} P^{2n}(\lambda_2 \rho_2, \delta, \ldots, \delta, r) \subset
P^{2n}(\lambda_1 \rho_2, \delta, \ldots, \delta, r) \\
& P^{2n}(\lambda_1 \rho_2, \delta, \ldots, r) \xhookrightarrow{\varphi_{1}} P^{2n}(\lambda_1 \rho_1, \delta, \ldots, \delta)
\subset P^{2n}(\delta, \ldots, \delta).
\end{aligned}\right.
\end{equation*}
Thus, in order to conclude the lemma it suffices to define
$$\varphi = \varphi_1 \circ \ldots \circ \varphi_{n-1}.$$
\end{proof}
\begin{propositionfolding}
Let $r > 0$, $D = \{0\} \times [-r, r] \times B^2(r) \times \ldots \times B^2(r) \subset \mathds{R}^{2n}$
and $\varepsilon > 0$. For any $\delta > 0$, there exists a Hamiltonian symplectomorphism $\psi$
with support in
$[-\varepsilon, \varepsilon] \times [-r-\varepsilon, r + \varepsilon] \times B^2(r + \varepsilon) \times
\ldots B^2(r + \varepsilon)$ such that $\psi(D) \subset P^{2n}(\delta, \ldots, \delta)$.
\end{propositionfolding}
\begin{proof}
Fix $\sigma$ from the previous lemma.
The Hamiltonian $H(\textbf{x}, \textbf{y}) = -x_1 y_1$ induces a flow in $\mathds{R}^{2n}$ which carries $\{0\} \times [-r, r] \times A^{2n-2}$ onto
$\{0\} \times [-\sigma, \sigma] \times A^{2n-2}$, for arbitrary $A^{2n-2}$.
Applying an appropriate cut-off to $H$ we can assume the flow is supported in
$[-\varepsilon, \varepsilon] \times [-r-\varepsilon, r + \varepsilon] \times B^2(r+ \varepsilon) \times \ldots \times B^2(r+ \varepsilon)$.
It is enough to compose the time--$t$ map of the flow, for large $t > 0$,
with $\varphi$ from Lemma \ref{lem:sabana} to obtain the desired map.
\end{proof}
|
\section{Introduction}
Efforts to find the chemical imprints in the oldest stars of the Milky Way left by the first
stellar generations (hereafter, first stars) have focused on very metal-poor halo stars
with [Fe/H]~$\sim-3$. Some cosmological simulations have suggested that at least half
of the first stars should have formed in the Galactic bulge (e.g. Tumlinson 2010).
Consequently, this Galactic component is a potential source of interesting targets to be explored.
These simulations suggest that the oldest stars, which have formed at the highest density peaks
(bulge), have enriched the surrounding interstellar medium (ISM) on very short timescales.
Chemical evolution models also suggest that the old bulge formed on short timescales
(e.g. Grieco et al. 2012).
Barbuy et al. (2009, 2014) searched for evidence of the signatures of formation of the first
stellar generations in the old bulge globular cluster NGC 6522, potentially the oldest Milky Way
globular cluster. The results were discussed in the framework of the early fast-rotating massive
stars, coined Spinstars, or mass transfer from AGB stars (Chiappini et al. 2011; Ness et al. 2014).
The central parts of the Galaxy, where the oldest stars most probably preferentially reside,
have not been targeted extensively, partly due to the very small fraction of metal-poor stars
in the predominantly metal-rich bulge region. The situation started to change with the Bulge Radial
Velocity Assay (BRAVA; Kunder et al. 2012) and the Abundances and Radial velocity Galactic Origins
Survey (ARGOS; Freeman et al. 2013), and the new data being obtained by the Apache Point Observatory
Galactic Evolution Experiment (APOGEE; Majewski et al. 2015) in the near infrared. In particular,
the Galactic bulge ARGOS Survey, an AAOmega/AAT spectroscopic survey that measured radial
velocities, metallicities and [$\alpha$/Fe] ratio of about 28, 000 stars (Freeman et al. 2013),
opened the opportunity to explore the bulge field metal-poor stars. The rather large number of
targets observed by ARGOS provided the opportunity of identifying bulge stars with metallicities
[Fe/H]~$\approx-1$, and estimating oversolar [$\alpha$/Fe] ratios.
Using ARGOS targets, we began a pilot project aimed at obtaining detailed chemical abundances of
metal-poor bulge field stars to look for possible chemical imprints of the first stars.
One of these imprints could be an overabundance of dominantly s-process elements (e.g. Meynet et al. 2006;
Pignatari et al. 2008; Chiappini et al. 2011; Frischknecht et al. 2012, 2015;
Barbuy et al. 2014; Cecutti et al. 2013; Cescutti \& Chiappini 2014).
Indeed, enhancements of Sr, Y, and Zr relative to Ba and La, and an excess of Ba or La relative to Eu,
in very old stars of the Milky Way can be attributed to the s-process activation in early generations of
fast rotating massive stars, which pollute the primordial material prior to the formation of the oldest
bulge (halo) field stars. An alternative possibility is an s-process contribution from massive AGB stars
bound in a binary system (e.g. Bisterzo et al. 2010). Otherwise the idea that has been more widely accepted
is that these elements were produced by the r-process in early times (Truran 1981). An extra process is also
claimed to produce the enhancement of the lightest heavy elements relative to the heaviest elements, in
literature called the lighter element primary process (LEPP; Travaglio et al. 2004; Bisterzo et al. 2014)
or weak r-process (Wanajo \& Ishimaru 2006). The astrophysical scenarios with neutrino-driven winds are
considered the most promising sites (Wanajo 2013; Arcones \& Thielemann 2013; Fujibayashi et al. 2015;
Niu et al. 2015).
The aim of this work is to obtain detailed chemical constraints from field bulge stars. Here, we analyse five
of these stars at high spectral resolution, using UVES spectra. We derive element abundances of C, N,
the $\alpha$-elements O, Mg, Si, Ca, and Ti, the odd-Z elements Na and Al, the dominantly s-elements Y, Zr, Ba,
and La, and the r-element Eu. The observations are described in Sect. 2. Photometric effective temperatures are
derived in Sect. 3. Spectroscopic parameters are derived in Section 4 and abundance ratios are computed in Sect. 5.
A discussion is presented in Sect. 6 and conclusions are drawn in Sect. 7.
\begin{table}
\caption{Log of the spectroscopic observations: date, time, exposure time,
seeing, and air mass
at the beginning and at the end of the observation.}
\label{log}
\scalefont{0.87}
\centering
\begin{tabular}{cccccc}
\hline\hline
\noalign{\smallskip}
\hbox{Run} & \hbox{Date} & \hbox{Time} &\hbox{Exp.} &\hbox{Seeing} &\hbox{Airmass} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{} & \hbox{} & \hbox{} & \hbox{(s)} & \hbox{(``)} & \hbox{} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{1} & 2012-07-12 & 04:07:43.1 & 2775 & 0.8$-$0.8 & 1.0$-$1.1\\
\hbox{2} & 2012-08-02 & 02:59:47.1 & 2775 & 0.3$-$0.7 & 1.0$-$1.1\\
\hbox{3} & 2012-07-21 & 03:59:12.6 & 2775 & 1.0$-$1.3 & 1.0$-$1.1\\
\hbox{4} & 2012-07-23 & 02:46:14.3 & 2775 & 0.7$-$1.0 & 1.0$-$1.0\\
\hbox{5} & 2012-08-02 & 03:57:27.9 & 2775 & 1.0$-$1.5 & 1.1$-$1.2\\
\hbox{6} & 2012-08-03 & 23:31:53.7 & 2775 & 0.8$-$0.8 & 1.2$-$1.1\\
\hbox{7} & 2012-08-22 & 01:22:09.9 & 2775 & 0.7$-$0.8 & 1.0$-$1.1\\
\hbox{8} & 2012-08-21 & 23:56:35.2 & 2308 & 0.9$-$0.9 & 1.0$-$1.0\\
\hbox{9} & 2012-08-22 & 02:24:20.8 & 2775 & 0.7$-$0.7 & 1.1$-$1.2\\
\hbox{10} & 2012-08-23 & 01:48:05.3 & 2775 & 1.0$-$1.1 & 1.0$-$1.1\\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table}
\begin{table}[ht!]
\caption{Geocentric radial velocity in each of the ten exposure runs with corresponding
heliocentric radial velocities and mean heliocentric radial velocity, in \kms.}
\label{vradial}
\centering
\begin{tabular}{crrrr}
\hline\hline
\noalign{\smallskip}
\hbox{run} & \hbox{RV$\rm_{G}$} & \hbox{RV$\rm_{B}$} & \hbox{RV$\rm_{G}$} & \hbox{RV$\rm_{B}$} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
&\multicolumn{2}{c}{221}&\multicolumn{2}{c}{224}\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{run1} & $-$96.9$\pm$2.1 & $-$106.7$\pm$2.1 & $-$105.9$\pm$2.1 & $-$115.6$\pm$2.1 \\
\hbox{run2} & $-$88.3$\pm$1.6 & $-$107.1$\pm$1.6 & $-$96.7$\pm$2.0 & $-$115.4$\pm$2.0 \\
\hbox{run3} & $-$92.4$\pm$1.8 & $-$106.3$\pm$1.8 & $-$101.1$\pm$2.0 & $-$115.0$\pm$2.0 \\
\hbox{run4} & $-$66.9$\pm$2.1 & $-$81.5$\pm$2.1 & $-$101.6$\pm$3.8 & $-$116.2$\pm$3.8 \\
\hbox{run5} & $-$87.5$\pm$2.4 & $-$106.3$\pm$2.4 & $-$96.3$\pm$2.3 & $-$115.1$\pm$2.3 \\
\hbox{run6} & $-$87.8$\pm$1.4 & $-$106.9$\pm$1.4 & $-$96.2$\pm$1.8 & $-$115.3$\pm$1.8 \\
\hbox{run7} & $-$81.9$\pm$2.0 & $-$107.1$\pm$2.0 & $-$90.4$\pm$2.1 & $-$115.5$\pm$2.1 \\
\hbox{run8} & $-$81.7$\pm$3.4 & $-$106.7$\pm$3.4 & $-$89.3$\pm$3.0 & $-$114.3$\pm$3.0 \\
\hbox{run9} & $-$80.8$\pm$1.6 & $-$106.1$\pm$1.6 & $-$89.3$\pm$1.6 & $-$114.5$\pm$1.6 \\
\hbox{run10} & $-$81.9$\pm$1.4 & $-$107.4$\pm$1.4 & $-$90.2$\pm$1.5 & $-$115.7$\pm$1.5 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{Mean} & & $-$106.7$\pm$2.1 & & $-$115.3$\pm$2.4 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
&\multicolumn{2}{c}{230}&\multicolumn{2}{c}{235}\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{run1} & $-$71.1$\pm$2.1 & $-$80.9$\pm$2.1 & 145.4$\pm$1.4 & 135.6$\pm$1.4 \\
\hbox{run2} & $-$62.2$\pm$2.1 & $-$80.9$\pm$2.1 & 154.3$\pm$1.6 & 135.6$\pm$1.6 \\
\hbox{run3} & $-$66.2$\pm$1.9 & $-$80.1$\pm$1.9 & 150.6$\pm$1.6 & 136.7$\pm$1.6 \\
\hbox{run4} & ------------- & ------------- & 149.7$\pm$1.5 & 135.1$\pm$1.5 \\
\hbox{run5} & ------------- & ------------- & 155.5$\pm$1.6 & 136.6$\pm$1.6 \\
\hbox{run6} & $-$61.5$\pm$2.5 & $-$80.6$\pm$2.5 & 155.0$\pm$1.5 & 135.9$\pm$1.5 \\
\hbox{run7} & $-$55.9$\pm$1.9 & $-$81.0$\pm$1.9 & 160.9$\pm$1.6 & 135.7$\pm$1.6 \\
\hbox{run8} & $-$54.7$\pm$4.4 & $-$79.7$\pm$4.4 & 162.0$\pm$1.6 & 137.0$\pm$1.6 \\
\hbox{run9} & $-$54.7$\pm$1.9 & $-$80.0$\pm$1.9 & 162.2$\pm$1.6 & 136.9$\pm$1.6 \\
\hbox{run10} & $-$55.5$\pm$1.8 & $-$80.9$\pm$1.8 & 161.0$\pm$1.4 & 135.6$\pm$1.4 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{Mean} & & $-$80.5$\pm$2.3 & & 136.1$\pm$1.6 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
& \multicolumn{2}{c}{238} & & \\
\hbox{run1} & $-$137.1$\pm$1.5 & $-$146.9$\pm$1.5 & & \\
\hbox{run2} & $-$128.0$\pm$1.6 & $-$146.7$\pm$1.6 & & \\
\hbox{run3} & $-$132.3$\pm$1.5 & $-$146.3$\pm$1.5 & & \\
\hbox{run4} & $-$132.6$\pm$1.8 & $-$147.3$\pm$1.8 & & \\
\hbox{run5} & $-$127.4$\pm$1.6 & $-$146.3$\pm$1.6 & & \\
\hbox{run6} & $-$127.5$\pm$1.7 & $-$146.6$\pm$1.7 & & \\
\hbox{run7} & $-$121.5$\pm$1.5 & $-$146.7$\pm$1.5 & & \\
\hbox{run8} & $-$121.0$\pm$1.9 & $-$146.0$\pm$1.9 & & \\
\hbox{run9} & $-$120.8$\pm$1.4 & $-$146.1$\pm$1.4 & & \\
\hbox{run10} & $-$121.5$\pm$1.4 & $-$147.0$\pm$1.4 & & \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{Mean} & & $-$146.6$\pm$1.7 & & \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table}
\begin{table*}[ht!]
\caption[1]{Identifications, coordinates, magnitudes, and reddening.
$JHK_{s}$ from both 2MASS and VVV surveys are given. }
\small
\scalefont{0.85}
\begin{flushleft}
\tabcolsep 0.15cm
\begin{tabular}{ccccccrrrrrrrrrr}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hline
\noalign{\smallskip}
{\rm star} & 2MASS ID & \hbox{$\alpha$(J2000)} & \hbox{$\delta$(J2000)} & \hbox{l($^{\circ}$)} & \hbox{b($^{\circ}$)} & $V$ & $J$ & $H$ & $K_{\rm s}$ & $J_{\rm VVV}$
& {\rm $H_{\rm VVV}$} & {\rm $K_{\rm s VVV}$} & E($B-V$)$^{a}$ & E($B-V$)$^{b}$ & \hbox{S/N}\\
\noalign{\vskip 0.2cm}
\noalign{\hrule\vskip 0.2cm}
\noalign{\vskip 0.2cm}
221 & 18033285-3117421 & 18:03:32.80 & $-$31:17:42.04 & 359.92 & $-$4.54 & 17.9 & 13.95 & 13.08 & 12.84 & 13.85 & 13.08 & 12.84 & 0.73 & 0.85 & 101 \\
224 & 18034522-3117379 & 18:03:45.18 & $-$31:17:37.72 & 359.94 & $-$4.57 & 17.9 & 13.96 & 13.10 & 12.84 & 13.77 & 13.02 & 12.79 & 0.77 & 0.90 & 79 \\
230 & 18033933-3114044 & 18:03:39.28 & $-$31:14:04.24 & 359.98 & $-$4.53 & 18.6 & 14.70 & 13.94 & 13.56 & 14.62 & 13.85 & 13.61 & 0.76 & 0.89 & 65 \\
235 & 18032741-3109441 & 18:03:27.40 & $-$31:09:43.96 & 0.02 & $-$4.45 & 16.2 & 12.63 & 11.81 & 11.60 & 12.58 & 11.71 & 11.57 & 0.72 & 0.83 & 175 \\
238 & 18031238-3106210 & 18:03:12.35 & $-$31:06:20.92 & 0.05 & $-$4.38 & 17.0 & 13.14 & 12.24 & 12.02 & 13.03 & 12.25 & 11.95 & 0.71 & 0.82 & 152 \\
\noalign{\smallskip} \hline
\end{tabular}
\tablebib{$^{a}$Schlafly \& Finkbeiner (2011), $^{b}$Schlegel et al. (1998)}
\end{flushleft}
\label{starmag}
\end{table*}
\begin{table*}
\caption{Photometric temperatures derived using the calibrations by Alonso et al. (1999)
for several colours and the final temperature adopted.
Colours from 2MASS and VVV catalogues were used, with reddening E($B-V$)
based on Schlegel et al. (1998) (first line) and Schlafly \& Finkbeiner (2011)
(second line) for each star.}
\label{temp}
\centering
\begin{tabular}{crcccccc}
\hline\hline
\noalign{\smallskip}
\hbox{Star} & \hbox{\Teff($V-K$)} & \hbox{\Teff($V-K$)} & \hbox{\Teff($J-H$)} & \hbox{\Teff($J-H$)} & \hbox{\Teff($J-K$)} & \hbox{\Teff($J-K$)} & \hbox{Average \Teff}\\
\hbox{} & \hbox{2MASS} & \hbox{VVV} & \hbox{2MASS} & \hbox{VVV} & \hbox{2MASS} & \hbox{VVV} & \hbox{(K)} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{221} & 4179.3 & 4172.5 & 4281.9 & 4304.9 & 4404.8 & 4478.9 & 4303.7\\
\hbox{} & 4401.3 & 4392.8 & 4400.9 & 4425.4 & 4576.7 & 4658.7 & 4476.0\\
\hbox{224} & 4249.3 & 4211.3 & 4348.7 & 4415.8 & 4436.3 & 4644.4 & 4384.3\\
\hbox{} & 4502.9 & 4455.1 & 4479.0 & 4550.9 & 4621.7 & 4853.9 & 4577.3\\
\hbox{230} & 4244.7 & 4267.8 & 4666.3 & 4348.2 & 4379.3 & 4526.9 & 4405.5\\
\hbox{} & 4493.2 & 4522.1 & 4817.5 & 4476.7 & 4555.6 & 4719.7 & 4597.5\\
\hbox{235} & 4476.1 & 4443.6 & 4437.9 & 3997.6 & 4640.7 & 4445.2 & 4406.9\\
\hbox{} & 4764.5 & 4723.9 & 4564.5 & 4096.8 & 4834.1 & 4617.7 & 4600.3\\
\hbox{238} & 4182.7 & 4135.7 & 4180.5 & 4258.3 & 4336.5 & 4261.4 & 4225.9\\
\hbox{} & 4396.8 & 4338.9 & 4289.0 & 4371.5 & 4495.3 & 4412.7 & 4384.0\\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table*}
\section {Observations and reductions}
We used the UVES spectrograph (Dekker et al. 2000), in FLAMES-UVES mode, for the observation
of five metal-poor ([Fe/H]~$\sim-1$) bulge stars, at a high resolution of R~=~45, 000 with a
slit width of 0.8''. Centring the wavelength at 5800~{\rm \AA}, the spectral wavelength range
4800~$-$~6800~{\rm \AA} with a gap at 5708~$-$~5825~{\rm \AA} was obtained. The red chip
(5800~$-$~6800~{\rm \AA}) has ESO CCD\#20, an MIT backside illuminated, with 4096~x~2048 pixels,
and pixel size 15~x~15~$\mu$m. The blue chip (4800~$-$~5800~{\rm \AA}) uses ESO Marlene EEV CCD\#44,
backside illuminated, with 4102~x~2048 pixels, and pixel size 15~x~15~$\mu$m. The pixel scale is
0.0147~{\rm \AA}/pix, with $\sim7.5$ pixels per resolution element at 6000~{\rm \AA}.
The log of observations is given in Table \ref{log}. The data were reduced using the UVES
pipeline, within ESO/Reflex software (Ballester et al. 2000; Modigliani et al. 2004).
The spectra were flatfielded, optimally-extracted and wavelength calibrated with the FLAMES-UVES
pipeline. The spectra were normalized, corrected for radial-velocity shift, and combined to
produce the final average data. Figure \ref{espectros} ilustrates the quality of the spectra
for the five sample stars.
\begin{figure}
\centering
\includegraphics[width=\hsize]{espectros.eps}
\caption{Portion of the final data. The dashed lines show \ion{Ni}{I} (red),
\ion{Fe}{I} (blue), and \ion{Mg}{I} (green) lines located in this wavelength range.}
\label{espectros}
\end{figure}
\subsection{Radial velocities}
In Table \ref{vradial}, we report the geocentric and heliocentric radial velocities measured
with IRAF/FXCOR for each of the 10 runs, together with their mean values. A solar spectrum
was adopted as the template. We used a solar synthetic spectrum to confirm the correction for
the radial-velocity shift. We note that the heliocentric radial velocity for the star 221 obtained
in run 4 is excessively different in comparison to others, so it was excluded from the final
average spectrum. In addition, the IRAF routine was not able to measure the radial velocities
for the star 230 using two runs (4 and 5), and they were also discarded.
\section{Photometric stellar parameters}
\subsection{Temperatures}
The selected stars, their OGLE and 2MASS designations, coordinates,
magnitudes, and S/N values corresponding to an average of clean windows
in the range 6400-6500~{\AA}, are given in Table~\ref{starmag}.
$V$ magnitudes computed using individual reddening from Schlegel et al. (1998)
in the direction of each star are adopted from the ARGOS survey
(Freeman et al. 2013), $JHK_s$ magnitudes from 2MASS (Skrutskie et al. 2006),
and VVV surveys (Saito et al. 2012). In this section we derive the photometric
temperatures, to compare with results from the ARGOS survey, as explained below.
We calculated photometric temperatures based on three colours: ($V-K$), ($J-H$), and ($J-K$).
Calibrations by Alonso et al. (1999) were applied, with reddening E($B-V$) computed with the
Galactic reddening and extinction calculator from the Infrared Processing and Analysis Center
(IRSA)\footnote{http://irsa.ipac.caltech.edu/applications/DUST/}. Results based on
Schlegel et al. (1998) and on Schlafly \& Finkbeiner (2011) were used. The extinction laws
given by Rieke \& Lebofsky (1985) were adopted. Colours from 2MASS were transformed into
the ESO photometric system and from this into the TCS (Telescopio Carlos S\'anchez) system,
following the relations established by Carpenter (2001) and Alonso et al. (1998).
The VVV $JHK_{s}$ colours were transformed to the 2MASS $JHK_{s}$ system, using relations by
Soto et al. (2013).
The derived photometric effective temperatures are listed in Table~\ref{temp}. In our sample,
the values obtained using the reddening from Schlafly \& Finkbeiner (2011) are
$\Delta$\Teff~$=182\pm13$~K higher than the results with reddening from Schlegel et al. (1998).
The more affected temperature is the one obtained with the colour ($V-K$), for which the average
difference is $\Delta$\Teff~$=243\pm22$~K. The differences between temperatures derived with
the VVV and 2MASS $JHK_{s}$ colours are $\Delta$\Teff~$=12\pm45$~K, which indicates that there
is no significant trend in the temperature owing to the survey chosen for our sample.
The outlier is the star 235: the temperatures obtained with 2MASS $JHK_{s}$ colours are
$\Delta$\Teff~$=232\pm9$~K higher than the results from the VVV survey.
Table \ref{paramargo} shows the parameters obtained from the ARGOS survey to our set of stars.
As described in Freeman et al. (2013), the effective temperatures were derived from the ($J-K$)
colours using the calibration from Bessell et al. (1998) with interstellar reddening from
Schlegel et al. (1998). The values for \logg, [Fe/H], and [$\alpha$/Fe] were determined by
comparing the observed spectra with a grid of synthetic spectra computed in LTE with the
code MOOG 2010 (Sneden 1973) and using 1D model atmospheres, described in
Castelli \& Kurucz (2004). A constant microturbulence velocity $\xi=2.0$~\kms~was adopted in
their method.
The photometric temperatures derived in this work are systematically lower than the
results from ARGOS: $\Delta$\Teff$=388\pm54$~K if the reddening from Schlegel et al. (1998)
is adopted and $\Delta$\Teff$=206\pm59$~K for the reddening maps of Schlafly \& Finkbeiner (2011).
These differences are smaller when our temperatures derived with the ($J-K$) colours are compared:
$\Delta$\Teff$=278\pm58$~K and $\Delta$\Teff$=98\pm66$~K, respectively. Indeed,
Freeman et al. (2013) report a mean temperature lower by 100~K when the empirical calibration
from Alonso et al. (1999) is applied, and this difference is higher (up to 200~K) for the
most metal-poor stars in the sample.
\begin{table}
\caption{Galactocentric velocities, atmospheric parameters, and enhancement
in $\alpha-$elements for the present sample from the Galactic bulge ARGOS Survey.}
\label{paramargo}
\centering
\begin{tabular}{crcccc}
\hline\hline
\noalign{\smallskip}
\hbox{Star} & \hbox{Vgal} & \hbox{\Teff} & \hbox{\logg} & \hbox{[Fe/H]} & \hbox{[$\alpha$/Fe]} \\
\hbox{} & \hbox{(\kms)} & \hbox{(K)} & \hbox{[cgs]} & \hbox{} & \hbox{} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{221} & $-$96.06 & 4669.54 & 1.6 & $-$0.80 & 0.40 \\
\hbox{224} &$-$105.17 & 4757.18 & 1.8 & $-$0.82 & 0.35 \\
\hbox{230} & $-$71.73 & 4697.27 & 1.4 & $-$0.84 & 0.03 \\
\hbox{235} & 146.01 & 4929.72 & 2.4 & $-$0.80 & 0.52 \\
\hbox{238} &$-$136.31 & 4613.03 & 2.4 & $-$0.80 & 0.33 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table}
\section{Spectroscopic stellar parameters}
\subsection{Equivalent widths}
To derive the atmospheric parameters, we measured the equivalent
widths (EW) of selected iron lines, using the IRAF software.
We decided to retain the lines with $10<$~EW~$<100$~m{\AA} located in
the range 6100~$-$~6800~{\AA} to derive the atmospheric parameters,
which is the spectral region with the highest S/N ratio available.
The EW values measured manually were adopted since this method allows
a better continuum placement and an individual evaluation of each line.
Table. \ref{EW_measurements} presents the complete list of lines,
describing the atomic data, the EW measured with IRAF,
and the individual iron abundance derived using the atmospheric
parameters adopted.
\begin{table*}
\caption{Spectroscopic parameters adopted for each star. We also present the number of iron lines
used for each star to derive the atmospheric parameters.
The difference $\Delta_{II-I}=[\ion{Fe}{II}/\hbox{H}]~-~[\ion{Fe}{I}/\hbox{H}]$ shows the quality in the ionization
equilibrium, and the parameters from the linear fitting $[\ion{Fe}{I}/\hbox{H}]=a_{EW}*log(EW/\lambda)+b_{EW}$
for microturbulence and $[\ion{Fe}{I}/\hbox{H}]=a_{exc}*exc.pot+b_{exc}$ for effective temperature.}
\label{param}
\scalefont{0.75}
\centering
\begin{tabular}{crcccccrcrrrrr}
\hline\hline
\noalign{\smallskip}
\hbox{Star} & \hbox{\Teff} & \hbox{\logg} & \hbox{[FeI/H]} & \hbox{[FeII/H]} & \hbox{[Fe/H]$_{model}$} & \hbox{$\xi$} &
\hbox{\#\ion{Fe}{I}} & \hbox{\#\ion{Fe}{II}} & \hbox{$\Delta_{II-I}$}
& \hbox{a$_{EW}$} & \hbox{b$_{EW}$} & \hbox{a$_{exc}$} & \hbox{b$_{exc}$}\\
\hbox{} & \hbox{(K)} & \hbox{[cgs]} & \hbox{} & \hbox{} & \hbox{} & \hbox{(\kms)} & \hbox{} & \hbox{} & \hbox{}
& \hbox{} & \hbox{} & \hbox{} & \hbox{} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{221} & 4620 & 2.0 & $-$0.88$\pm$0.24 & $-$0.91$\pm$0.17 & $-$0.90 & 1.0 & 31 & 2 & $-$0.035 & $+$0.00$\pm$0.12 & $+$6.63$\pm$0.55 & $+$0.004$\pm$0.045 & $+$6.61$\pm$0.15 \\
\hbox{224} & 5000 & 3.5 & $-$0.62$\pm$0.36 & $-$0.60$\pm$0.30 & $-$0.65 & 0.8 & 61 & 4 & $+$0.015 & $+$0.02$\pm$0.10 & $+$6.96$\pm$0.48 & $+$0.013$\pm$0.044 & $+$6.84$\pm$0.15 \\
\hbox{230} & 4960 & 3.0 & $-$1.04$\pm$0.40 & $-$1.03$\pm$0.18 & $-$1.10 & 0.8 & 58 & 5 & $+$0.0038 & $+$0.02$\pm$0.11 & $+$6.57$\pm$0.52 & $+$0.087$\pm$0.056 & $+$6.17$\pm$0.19 \\
\hbox{235} & 4680 & 2.2 & $-$0.94$\pm$0.15 & $-$0.95$\pm$0.09 & $-$0.95 & 1.1 & 45 & 7 & $-$0.012 & $+$0.00$\pm$0.03 & $+$6.59$\pm$0.17 & $-$0.002$\pm$0.024 & $+$6.57$\pm$0.08 \\
\hbox{238} & 4720 & 2.9 & $-$0.43$\pm$0.18 & $-$0.42$\pm$0.07 & $-$0.50 & 1.0 & 33 & 5 & $-$0.0077 & $+$0.00$\pm$0.07 & $+$7.07$\pm$0.35 & $+$0.027$\pm$0.031 & $+$6.98$\pm$0.11 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table*}
\subsection{Atmospheric parameters}
The photometric temperatures, together with the gravity and
metallicity values from the ARGOS survey as given in Table \ref{paramargo},
are adopted as a first guess to calculate the excitation
and ionization equilibria of \ion{Fe}{I} and \ion{Fe}{II} lines.
The MARCS spherical model atmosphere grids (Gustafsson et al. 2008) with 1~M$_{\odot}$
and the code Turbospectrum (Alvarez \& Plez 1998) in the equivalent width mode
were used, with solar abundances adopted from Asplund et al. (2009).
Applying an automatic routine on a grid of models with $\Delta$\Teff~$=20$~K,
$\Delta$\logg~$=0.1$~[cgs], and $\Delta\xi=0.1$~\kms, the final surface
gravity \logg~was chosen to minimize $[\ion{Fe}{II}/\hbox{H}]~-~[\ion{Fe}{I}/\hbox{H}]$,
the final microturbulence velocity $\xi$ was chosen to minimize the dependence of
[\ion{Fe}{I}/H] on log~(EW/$\lambda$), and the final temperature was obtained by
the excitation equilibrium. The grid was recomputed successively with a new metallicity
in each step, and the range used for each parameter was selected to avoid local solutions.
Figure \ref{221_235} shows the excitation and ionization equilibria for two different
typical cases: the star 221, which presents a spectrum with a low S/N; and the star 235,
which presents a high-quality spectrum. Lines with abundances out of the region limited by
$\pm3\sigma$, where $\sigma$ is the standard error of the mean, were removed from the computations
of final metallicities.
The black dots are the
abundances obtained from \ion{Fe}{I} lines, and the red squares are the results
from the \ion{Fe}{II} lines. The blue dashed lines represent the linear fit to data,
and the blue dotted lines are the same function moved vertically by .
The derived stellar parameters are reported in Table \ref{param}, as well as the number
of \ion{Fe}{I} and \ion{Fe}{II} lines retained, the difference
$\Delta_{II-I}=[\ion{Fe}{II}/\hbox{H}]~-~[\ion{Fe}{I}/\hbox{H}]$ obtained with the final models,
and the parameters from the linear fit to data in each case. The angular coefficients
and the values of $\Delta_{II-I}$ are null within the error bar, which indicate that there
are no relevant trends in the excitation and ionization equilibria.
The parameters can be compared with the results derived from the mid-resolution survey ARGOS,
reported in Table \ref{paramargo}.
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{221_235.eps}}
\caption{Ionization and excitation equilibria of Fe lines for the stars 221 and 235,
using the newly derived atmospheric parameters. The black dots are the abundances
obtained from the \ion{Fe}{I} lines, the red squares are those from the \ion{Fe}{II} lines,
the blue dashed lines represent the linear fit to data, and the dotted blue lines
are the same linear fit moved vertically by $\pm3\sigma$, where $\sigma$ is the
standard error of the mean.}
\label{221_235}
\end{figure}
\section{Abundance ratios}
\begin{table*}
\caption{Mean LTE abundances of the elements derived in the present work.}
\label{finalAbund}
\scalefont{1.0}
\centering
\begin{tabular}{crrrrrrrrrrr}
\hline\hline
\noalign{\smallskip}
\hbox{Element} & \hbox{A(X)$_{\odot}$} & \hbox{A(X)} & \hbox{[X/Fe]} & \hbox{A(X)} & \hbox{[X/Fe]} & \hbox{A(X)} & \hbox{[X/Fe]} & \hbox{A(X)} & \hbox{[X/Fe]} & \hbox{A(X)} & \hbox{[X/Fe]} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
& & \multicolumn{2}{c}{221} & \multicolumn{2}{c}{224} & \multicolumn{2}{c}{230} & \multicolumn{2}{c}{235} & \multicolumn{2}{c}{238} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{\ion{Fe}{I}*} & $+$7.50 & $+$6.62 & $-$0.88 & $+$6.88 & $-$0.62 & $+$6.46 & $-$1.04 & $+$6.56 & $-$0.94 & $+$7.07 & $-$0.43 \\
\hbox{\ion{Fe}{II}*} & $+$7.50 & $+$6.59 & $-$0.91 & $+$6.90 & $-$0.60 & $+$6.47 & $-$1.03 & $+$6.55 & $-$0.95 & $+$7.08 & $-$0.43 \\
\hbox{\ion{O}{I}} & $+$8.69 & $+$8.40 & $+$0.61 & ------- & ------- & $+$8.25 & $+$0.60 & $+$8.40 & $+$0.66 & $+$8.85 & $+$0.59 \\
\hbox{\ion{Na}{I}} & $+$6.24 & $+$5.68 & $+$0.33 & $+$5.58 & $-$0.06 & $+$4.90 & $-$0.31 & $+$5.21 & $-$0.09 & $+$5.93 & $+$0.12 \\
\hbox{\ion{Mg}{I}} & $+$7.60 & $+$7.22 & $+$0.52 & $+$7.50 & $+$0.51 & $+$7.19 & $+$0.63 & $+$7.13 & $+$0.47 & $+$7.45 & $+$0.28 \\
\hbox{\ion{Al}{I}} & $+$6.45 & $+$6.00 & $+$0.45 & $+$6.29 & $+$0.45 & $+$5.77 & $+$0.36 & $+$5.79 & $+$0.28 & $+$6.23 & $+$0.21 \\
\hbox{\ion{Si}{I}} & $+$7.51 & $+$6.92 & $+$0.31 & $+$7.25 & $+$0.35 & $+$6.98 & $+$0.50 & $+$6.78 & $+$0.22 & $+$7.23 & $+$0.14 \\
\hbox{\ion{Ca}{I}} & $+$6.34 & $+$5.67 & $+$0.23 & $+$6.01 & $+$0.28 & $+$5.46 & $+$0.16 & $+$5.75 & $+$0.36 & $+$6.10 & $+$0.18 \\
\hbox{\ion{Ti}{I}} & $+$4.95 & $+$4.38 & $+$0.33 & $+$4.71 & $+$0.37 & $+$4.28 & $+$0.37 & $+$4.39 & $+$0.39 & $+$4.87 & $+$0.34 \\
\hbox{\ion{Ti}{II}} & $+$4.95 & $+$4.51 & $+$0.46 & $+$4.80 & $+$0.46 & $+$4.34 & $+$0.42 & $+$4.38 & $+$0.38 & $+$4.82 & $+$0.30 \\
\hbox{\ion{Y}{I}} & $+$2.21 & $+$1.60 & $+$0.29 & $+$1.95 & $+$0.12 & $+$1.90 & $+$0.73 & $+$1.60 & $+$0.34 & $+$1.92 & $+$0.14 \\
\hbox{\ion{Zr}{I}} & $+$2.56 & $+$2.40 & $+$0.72 & $+$2.60 & $+$0.63 & $+$2.40 & $+$0.86 & $+$2.33 & $+$0.69 & $+$2.60 & $+$0.45 \\
\hbox{\ion{Ba}{II}} & $+$2.18 & $+$1.60 & $+$0.32 & $+$1.80 & $+$0.23 & $+$1.63 & $+$0.48 & $+$1.90 & $+$0.67 & $+$2.10 & $+$0.35 \\
\hbox{\ion{La}{II}} & $+$1.10 & $+$0.30 & $+$0.10 & $+$1.03 & $+$0.54 & ------- & ------- & $+$0.68 & $+$0.53 & $+$0.90 & $+$0.23 \\
\hbox{\ion{Eu}{II}} & $+$0.52 & $+$0.00 & $+$0.38 & $+$0.35 & $+$0.44 & $-$0.10 & $+$0.42 & $+$0.05 & $+$0.48 & $+$0.56 & $+$0.47 \\
\noalign{\smallskip}
\hline
\end{tabular}
\tablefoot{ *: [X/H] is used in place of [X/Fe].}
\end{table*}
A line-by-line fitting was carried out to derive the abundances,
using the spectrum synthesis code Turbospectrum (Alvarez \& Plez 1998),
which includes scattering
in the blue and UV domain, molecular dissociative equilibrium, and collisional
broadening by H, He, and H$_{2}$, following Anstee \& O'Mara (1995),
Barklem \& O'Mara (1997), and Barklem et al. (1998). The atomic line lists
were adopted from the Vienna Atomic Line Database compilation (VALD3; Piskunov et al. 1995),
together with the Turbospectrum molecular line lists (B. Plez, private communication).
For lines used to derive abundances, as reported in Table \ref{linelist},
the oscillator strengths were adopted from Barbuy et al. (2014), except
when described. Hyperfine structures were adopted for the lines relevantly
affected by this effect. Tables \ref{finalAbund} and \ref{CN} show the adopted final
abundances.
\subsection{Carbon and nitrogen}
To evaluate the adopted line list in the regions selected for
carbon and nitrogen abundances, we used the Arcturus spectrum (Hinkle et al. 2000)
as a reference star. Our benchmark analysis is based on the stellar parameters
described in Mel\'endez et al. (2003): \Teff~$=4275$~K, \logg~$=1.55$~[cgs],
[Fe/H]~$=-0.54$, and $\xi=1.65$~\kms. We adopted chemical abundances from
Ram\'irez \& Allende Prieto (2011) and Mel\'endez et al. (2003),
as presented in Table \ref{arcturus_abund}.
\begin{table}
\caption{Adopted Arcturus abundaces.}
\label{arcturus_abund}
\centering
\begin{tabular}{c c c c}
\hline\hline
El. & A(X)$_{Arcturus}$ & El. & A(X)$_{Arcturus}$ \\
\hline
\noalign{\smallskip}
C & 8.32$^{[1]}$ & Ca & 5.94$^{[1]}$ \\
N & 7.68$^{[2]}$ & Sc & 2.81$^{[1]}$ \\
O & 8.66$^{[2]}$ & Ti & 4.66$^{[1]}$ \\
Na & 5.82$^{[1]}$ & V & 3.58$^{[1]}$ \\
Mg & 7.47$^{[1]}$ & Cr & 4.99$^{[1]}$ \\
Al & 6.26$^{[1]}$ & Mn & 4.74$^{[1]}$ \\
Si & 7.30$^{[1]}$ & Co & 4.71$^{[1]}$ \\
K & 4.99$^{[1]}$ & Ni & 5.73$^{[2]}$ \\
\hline
\end{tabular}
\tablebib{[1]: Ram\'irez \& Allende Prieto (2011);
[2]: Mel\'endez et al. (2003).}
\end{table}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{arcturus.eps}}
\caption{\textbf{Upper panel:} Fit to the C$_{2}$(1,0) molecular
bandhead at 5635.3~{\AA} in Arcturus. Observations (black crosses)
are compared with synthetic spectra computed with the adopted abundances
(blue solid lines) from literature. Red marks show the positions of molecular lines.
\textbf{Lower panel:} Fit to the CN(6,2) molecular bandhead at
6478.5~{\AA} in Arcturus. Symbols are the same as in the upper panel.}
\label{arcturus}
\end{figure}
To measure the carbon abundances we used the C$_{2}(0,1)$ molecular bandhead.
The region is extended and a mean abundance was derived from the
overall fit, however the bandhead at 5635.3~{\rm \AA} received
more weight in the fitting procedure. The line list
for $^{12}$C$_{2}$, $^{13}$C$_{2}$, and $^{12}$C$^{13}$C was
adopted from Wahlin \& Plez (2005), which contains transitions from the
Swan (d$^{3}\Pi$~-~a$^{3}\Pi$) electronic band. The solar isotopic
fraction for $^{12}$C (98.9\%) and $^{13}$C (1.1\%) was adopted (Asplund et al. 2009).
Figure \ref{arcturus} shows in the upper panel the synthetic spectrum
computed for Arcturus (blue solid line), which is in very good agreement with
observations. For the sample stars, the C$_{2}(0,1)$ molecular bandhead is located
in the region observed with the blue chip, showing a lower S/N.
An example can be seen in the fit to star 235 shown in Fig. \ref{carbon}
(upper panel). The C abundances were adopted as upper limits.
The derived abundances are presented in Table \ref{CN}. Beers \& Christlieb (2005)
defined carbon-enhanced metal-poor stars (CEMP) as having [C/Fe]~$>+1.0$,
but Aoki et al. (2007) presents a new definition, which takes into account the
mixing events in evolved stars and the consequently lower carbon abundance on their
surface. Following Aoki et al. (2005, 2007), we assumed the mass of the stars
to be $0.8$~M$_\odot$ to calculate the luminosities
$L/L_{\odot}\propto(M/M_{\odot})(g/g_{\odot})^{-1}(T_{eff}/T_{eff\odot})^{4}$,
and in Fig. \ref{carbon} (lower panel) we show the [C/Fe] abundance ratios
as a function of the luminosity log($L/L_{\odot}$) for our sample. The limits for
CEMP stars are also presented, showing that our sample consists of carbon-normal
metal-poor stars (non-CEMP).
For comparison, we included carbon abundances of bulge stars from the literature.
The open black squares represent seven stars in the globular cluster M~62 (NGC~6266)
studied in high-resolution by Yong et al. (2014). This object is located at J(2000)
$\alpha=17^{h}01^{m}12.60^{s}$ and $\delta=-30^{\circ}06'44.5''$ (Di Criscienzo et al. 2006),
or $l=353.5746^{\circ}$ and $b=+7.3196^{\circ}$, therefore projected in the bulge.
The open black stars represent the results of the high-resolution abundance analysis from
Barbuy et al. (2014) for the globular cluster NGC~6522, which is located at J(2000)
$\alpha=18^{h}03^{m}34.08^{s}$ and $\delta=-30^{\circ}02'02.3''$, or $l=1.0246^{\circ}$
and $b=-3.9256^{\circ}$ (Barbuy et al. 2009), and therefore also projected in the bulge.
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{carbon.eps}}
\caption{\textbf{Upper panel:} Fit to the C$_{2}$(1,0) bandhead
at 5635.3~{\AA} in star 235. Observations (black crosses) are
compared with synthetic spectra computed with the different abundances
indicated in the figure (blue dashed lines), as well as with the adopted
abundance (red solid lines), also indicated.
\textbf{Lower panel:} comparison of the [C/Fe] abundance ratios derived
in the present sample (filled red circles) with stars from the
bulge blobular clusters M~62 (Yong et al. 2014, open black squares)
and NGC~6522 (Barbuy et al. 2014, open black stars). The dashed black
line corresponds to the limit for carbon-enhanced stars, as defined
by Aoki et al. (2007).}
\label{carbon}
\end{figure}
The nitrogen abundance was derived using the CN A$^{2}\Pi$~-~X$^{2}\Sigma$
red system, based on the CN$(6,2)$~6478.48~{\rm \AA} bandhead. The CN line list is a
compilation by B. Plez (private communication), using data from Cerny et al. (1978),
Kotlar et al. (1980), Larsson et al. (1983), Bauschlicher et al. (1988),
Ito et al. (1988a, 1988b), Prasad \& Bernath (1992), Prasad et al. (1992), and
Rehfuss et al. (1992). All four isotope combinations $^{12}$C$^{14}$N, $^{12}$C$^{15}$N,
$^{13}$C$^{14}$N, and $^{13}$C$^{15}$N were treated with nitrogen solar isotopic fraction
$^{14}$N (99.8\%) and $^{15}$N (0.2\%) from Asplund et al. (2009).
The synthetic spectrum computed for Arcturus (blue solid line) in this
region is shown in Fig. \ref{arcturus} (lower panel), showing good agreement with
observations.
For the sample stars, the selected molecular transitions are weak and the noise becomes
more evident, as shown in Fig. \ref{N_O} (upper left panel) for star 238. Table \ref{CN} shows
the derived N abundances which, owing to the previous discussion, must be used with caution.
The difficulty in defining the local continuum does not permit us to determine the N abundance
in the star 224.
\begin{table}
\caption{Carbon and nitrogen abundances [X/Fe] from C$_{2}$ and
CN bandheads.}
\label{CN}
\scalefont{0.9}
\centering
\begin{tabular}{ccccccc}
\hline\hline
\noalign{\smallskip}
\hbox{Species} & \hbox{$\lambda$({\rm\AA})} & \hbox{221} & \hbox{224} & \hbox{230} & \hbox{235} & \hbox{238}\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{[C/Fe] C$_{2}(0,1)$} & 5635.3 & $<+$0.2 & $<+$0.2 & $<+$0.1 & $<$0.0 & $<+$0.1 \\
\hbox{[N/Fe] CN$(6,2)$} & 6478.5 & $+$0.82 & ------- & $+$0.71 & $+$0.97 & $+$0.35 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{N_O.eps}}
\caption{\textbf{Upper left panel:} Fit to the
CN$(6,2)$~6478.48~{\rm \AA} bandhead in star 238.
\textbf{Upper right panel:} Fit to the [\ion{O}{I}]~6300.3~{\AA}
line in star 235. Symbols are the same as in Fig. \ref{carbon}
(upper panel). \textbf{Lower panel:} [O/Fe] abundance
ratio as a function of the metallicity for the five sample stars
(filled red circles), compared with literature abundances from
Bensby et al. (2013; filled black triangles), Barbuy et al. (2014;
open black stars), Johnson et al. (2014; open grey circles),
and Barbuy et al. (2015; filled grey circles).}
\label{N_O}
\end{figure}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{Mg_Si_Ca_Ti.eps}}
\caption{\textbf{Upper left panel:} Fit to the
\ion{Mg}{I}~5528.4~{\AA} line in star 235.
\textbf{Upper right panel:} Fit to the \ion{Si}{I}~6243.8~{\AA}
line in star 224.
\textbf{Lower left panel:} Fit to the \ion{Ca}{I}~6439.1~{\AA}
line in star 235. \textbf{Lower right panel:} Fit to the
\ion{Ti}{II}~6559.6~{\AA} line in star 224.
Symbols are the same as in Fig. \ref{carbon} (upper panel).}
\label{Mg_Si_Ca_Ti}
\end{figure}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{Y_Ba_La_Eu.eps}}
\caption{\textbf{Upper left panel:} Fit to the \ion{Y}{I}~6435.0~{\AA}
line in star 235. \textbf{Upper right panel:} Fit to the
\ion{Ba}{II}~6141.7~{\AA} line in star 230.
\textbf{Lower left panel:} Fit to the \ion{La}{II}~6390.5~{\AA}
line in star 224. \textbf{Lower right panel:} Fit to the
\ion{Eu}{II}~6437.6~{\AA} line in star 238.
Symbols are the same as in Fig. \ref{carbon} (upper panel).}
\label{Y_Ba_La_Eu}
\end{figure}
\subsection{Alpha elements}
The oxygen abundance was derived using the forbidden line [\ion{O}{I}]~6300.3~{\AA},
as shown in Fig. \ref{N_O} (upper right panel) for the star 235.
We inspected the individual spectra, before combining, to check for possible blends
with telluric lines, and we remove them from the final average when necessary.
For star 224, the [\ion{O}{I}] line was strongly contaminated
in all individual spectra, and consequently the oxygen abundance was not derived.
Some individual spectra were also discarded owing to the higher noise level surrounding
the [\ion{O}{I}] line, compared with the average, which allowed a better placement of the
continuum.
In Fig. \ref{N_O} (lower panel) we compare the [O/Fe] abundance ratios in the
sample stars with the result in the bulge globular cluster NGC~6522 (Barbuy et al. 2014),
with microlensed bulge dwarfs and subgiants stars from Bensby et al. (2013) selected to have
ages older than 11 Gyr, with selected red giant branch stars in the Galactic bulge from Johnson
et al. (2014), and with the giant stars from Barbuy et al. (2015). The solar oxygen abundance
A(O)$_{\odot}=8.69$ (Asplund et al. 2009) adopted in our results is $0.08$~dex lower than
A(O)$_{\odot}=8.77$ adopted in Barbuy et al. (2014, 2015) and, to ensure consistency among the
abundance results, we shifted their values. This figure shows that our abundances are in agreement
with previous results for bulge stars.
We checked four \ion{Mg}{I} lines located at 5528.4~{\AA}, 6318.7~{\AA}, 6319.24~{\AA}, and 6765.4~{\AA}
to derive the magnesium abundance. The line at 5528.4~{\AA} is in the portion of the spectra with more noise,
obtained with the blue chip, and was only useful in star 235, as shown in Fig. \ref{Mg_Si_Ca_Ti} (upper left panel).
The silicon abundance was measured using ten \ion{Si}{I} lines, two of them (5665.5~{\AA}
and 5690.4~{\AA}) located in the wavelengths measured with the blue chip, but with individual abundances that are
consistent with results from the lines in the red portion of the spectra. In the upper right panel of
Fig. \ref{Mg_Si_Ca_Ti}, we show the fit to the line at 6243.8~{\AA} in star 224.
The calcium abundance was derived after checking 19 \ion{Ca}{I} lines. The transition located at
5601.3~{\AA} is the only line in the blue spectra and was used only for stars 235 and 238, giving individual
abundances that are in agreement with the other lines. The result obtained from the \ion{Ca}{I}~6439.1~{\AA}
line in star 235 is shown in Fig. \ref{Mg_Si_Ca_Ti} (lower left panel).
It was possible to inspect 14 \ion{Ti}{I} lines, with only the \ion{Ti}{I}~5689.5~{\AA} line located in the
blue portion of the spectra. For the ionized species, six \ion{Ti}{II} lines were checked to obtain the
titanium abundance, but three of them (5336.8~{\AA}, 5381.0~{\AA}, and 5418.7~{\AA}) are located in
wavelengths of the blue portion of the spectra and they were only used for stars 235 and 238. In the lower right
panel of Fig. \ref{Mg_Si_Ca_Ti}, we show the \ion{Ti}{II}~6559.6~{\AA} line measured in star 224.
This line is located in the blue wing of the H$\alpha$ line, and it was necessary to take the hydrogen
line in the spectrum synthesis into account.
\subsection{Odd-Z elements Na, Al}
The sodium abundances are based on four \ion{Na}{I} lines,
located at 4982.8~{\AA}, 5688.2~{\AA}, 6154.2~{\AA},
and 6160.7~{\AA}. We did not use the resonance lines
\ion{Na}{I}~5889.95~{\rm \AA} (D2) and \ion{Na}{I}~5895.92~{\rm \AA}
(D1) because they are very sensitive to non-LTE effects.
The only stable isotope
$^{23}$Na has nuclear spin I~$=3/2$\footnote{Adopted from the
Particle Data Group (PDG) collaboration: http://pdg.lbl.gov/}
and therefore exhibits hyperfine structure (HFS).
The hyperfine coupling constants are adopted from Das \& Natarajan (2008)
and Marcassa et al. (1998). When not
available in the literature, the hyperfine constants for a given level
were assumed to be null. The line splitting was computed by employing
a code made available by McWilliam et al. (2013).
For aluminum, only \ion{Al}{I}~6696.0~{\AA} and \ion{Al}{I}~6698.7~{\AA}
lines were available. The stable isotope $^{27}$Al has nuclear spin
I~$=5/2$ and we adopted the hyperfine coupling constants from Nakai et al. (2007)
and Belfrage et al. (1984) to compute the HFS.
\subsection{Heavy elements}
We derive the abundances of the neutron-capture elements Y, Zr, Ba, La,
and the reference r-element Eu.
We preferentially used lines of ionized species, since
these elements are mostly in this form.
For strontium, we evaluated six \ion{Sr}{I}
lines: 6408.5~{\AA}, 6504.0~{\AA}, 6546.8~{\AA}, 6550.2~{\AA},
6617.3~{\AA}, and 6791.0~{\AA}. All transitions are too weak for
abundance purposes, consequently no [Sr/Fe] result is presented.
The best yttrium line \ion{Y}{II}~6795.4~{\AA} is located in the border of
the \'echelle spectrum and shows clearly fringes that prevent its use.
The most reliable line from our spectra is \ion{Y}{II}~5544.6~{\AA},
which is located in the blue portion of the spectra and was not useful
for the abundance determination. Consequently, the Y abundance in the sample
is based on the \ion{Y}{I}~6435.0~{\AA} line, as shown in the upper left
panel of Fig. \ref{Y_Ba_La_Eu} for star 235.
For zirconium, we checked three \ion{Zr}{II} lines located at
5112.3~{\AA}, 5350.1~{\AA}, and 5350.3~{\AA}, but none is reliable
for measuring abundances. Due to a lack of useful ionized lines, we measured
abundances from three lines of \ion{Zr}{I}: 6127.47~{\rm \AA}, 6134.58~{\rm \AA},
and 6143.25~{\rm \AA}. Oscillator strengths were adopted from van der Swaelmen et al. (2013).
The barium abundance was measured using the \ion{Ba}{II}~6141.7~{\AA}
and \ion{Ba}{II}~6496.9~{\AA} lines. As well known, the HFS
(nuclear spin I~$=3/2$) and the isotopic splitting are important effects to
be taken into account in Ba transitions. Following Barbuy et al. (2014),
the hyperfine coupling constants were adopted
experimentally from Rutten (1978) and Biehl (1976).
According to Asplund et al. (2009), the major contribution comes from the
isotope $^{138}$Ba (71.698\%), followed
by $^{137}$Ba (11.232\%), $^{136}$Ba (7.854\%), $^{135}$Ba (6.592\%), and
$^{134}$Ba (2.417\%). The isotopes $^{130}$Ba and $^{132}$Ba together represent
less than 0.11\% and they were ignored in the computations.
In addition, to compute the profile for \ion{Ba}{II}~6141.7~{\AA}, it is
important to include a blend with the \ion{Fe}{I}~6141.7~{\AA},
for which we adopted \loggf~$=-1.60$ (Barbuy et al. 2014).
The fit to this line for star 230 is shown in
Fig. \ref{Y_Ba_La_Eu} (upper right panel).
The lanthanum abundance is a contribution of two stable isotopes.
The most relevant is $^{139}$La, with $99.909\%$ in the solar material,
and the only isotope included in the computations since $^{138}$La
contributes less than $0.1\%$ (Asplund et al. 2009).
The HFS values were computed with coupling constants A and B adopted from
Lawler et al. (2001a) and Biehl (1976), with nuclear spin I~$=7/2$.
The final abundances are based on three \ion{La}{II} lines, located
at 6320.4~{\AA}, 6390.5~{\AA}, and 6774.3~{\AA}.
In Fig. \ref{Y_Ba_La_Eu} we show the fit to the \ion{La}{II}~6320.5~{\AA}
line for star 224 (lower left panel).
Europium is the heaviest element measured in the sample stars.
The solar isotopic fraction $^{151}$Eu~$=47.81\%$ and $^{153}$Eu~$=52.19\%$
(Asplund et al. 2009) was adopted, with nuclear spin I~$=5/2$.
We computed the HFS using coupling constants A and B from Lawler et al. (2001b).
The final europium abundances were derived from the \ion{Eu}{II}~6437.6~{\AA}
and \ion{Eu}{II}~6645.1~{\AA} lines, and in Fig. \ref{Y_Ba_La_Eu} we show
(lower right panel) the result for the \ion{Eu}{II}~6437.6~{\AA} line
in star 238.
\subsection{Uncertainties on the derived abundances}
The typical errors in the spectroscopic atmospheric parameters are
$\Delta$\Teff~$=100$~K, $\Delta$\logg~$=0.1$~dex, and $\Delta\xi=0.2$~\kms.
Since the stellar parameters are not independent, the quadratic sum of
the abundance uncertainties that arise from each of these three sources
independently will add significant covariance terms to the final error budget.
We solved this problem by creating a new atmospheric model with a $100$~K lower
temperature, determining the corresponding surface gravity \logg~and microturbulent
velocity $\xi$ by the spectroscopic method. The difference between the
abundances derived with this new model and the nominal model in each
star are expected to represent the total error budget arising from the
stellar parameters.
The observational uncertainties are assumed as the standard error of the mean
obtained with the abundances from individual lines. For elements with three or
less lines used to determine the average, we adopted the Fe observational
error as a representative value. The final error is the quadratic sum of the
uncertainty from the atmospheric parameters and the observational error.
Table \ref{error} shows the results in star 235 as an example.
It is important to note, as described already in Table \ref{param}, that
the observation errors in stars 221, 224, and 230 are significantly larger
in comparison with stars 235 and 238, as a consequence of differences
in the S/N.
\begin{table}
\caption{Observational and atmospheric errors in star 235, as well as the final uncertainties.}
\label{error}
\centering
\begin{tabular}{ccccc}
\hline\hline
\noalign{\smallskip}
\hbox{Element} & \hbox{$\Delta_{obs}$} & \hbox{$\Delta_{atm}$} & \hbox{$\Delta_{final[X/H]}$} & \hbox{$\Delta_{final[X/Fe]}$} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
& \hbox{(dex)} & \hbox{(dex)} & \hbox{(dex)} & \hbox{(dex)} \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{\ion{Fe}{I}} & 0.15 & 0.08 & 0.17 & ---- \\
\hbox{\ion{Fe}{II}} & 0.09 & 0.06 & 0.11 & ---- \\
\hbox{C(C$_{2}$)} & 0.09 & 0.04 & 0.10 & 0.14 \\
\hbox{N(CN)} & 0.09 & 0.14 & 0.17 & 0.19 \\
\hbox{[\ion{O}{I}]} & 0.09 & 0.12 & 0.15 & 0.18 \\
\hbox{\ion{Na}{I}} & 0.09 & 0.02 & 0.09 & 0.14 \\
\hbox{\ion{Mg}{I}} & 0.09 & 0.03 & 0.09 & 0.14 \\
\hbox{\ion{Al}{I}} & 0.09 & 0.05 & 0.10 & 0.14 \\
\hbox{\ion{Si}{I}} & 0.03 & 0.02 & 0.04 & 0.11 \\
\hbox{\ion{Ca}{I}} & 0.10 & 0.02 & 0.10 & 0.14 \\
\hbox{\ion{Ti}{I}} & 0.06 & 0.08 & 0.10 & 0.14 \\
\hbox{\ion{Ti}{II}} & 0.04 & 0.07 & 0.08 & 0.13 \\
\hbox{\ion{Y}{I}} & 0.09 & 0.07 & 0.11 & 0.15 \\
\hbox{\ion{Zr}{I}} & 0.09 & 0.08 & 0.12 & 0.16 \\
\hbox{\ion{Ba}{II}} & 0.09 & 0.02 & 0.09 & 0.14 \\
\hbox{\ion{La}{II}} & 0.09 & 0.07 & 0.11 & 0.15 \\
\hbox{\ion{Eu}{II}} & 0.09 & 0.05 & 0.10 & 0.14 \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{table}
\section{Discussion}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{Na_Al_compara.eps}}
\caption{[Na/Fe] (\textbf{upper panel}) and [Al/Fe] (\textbf{lower panel})
abundance ratios as a function of the metallicity for the five sample
stars (filled red circles), compared with literature abundances
from Yong et al. (2014; open black squares), Barbuy et al. (2014; open
black stars), Bensby et al. (2013; filled black triangles),
Johnson et al. (2012; grey crosses), and Johnson et al. (2014; open grey circles).}
\label{Na_Al_compara}
\end{figure}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{Mg_Si_compara.eps}}
\caption{[Mg/Fe] (\textbf{upper panel}) and [Si/Fe] (\textbf{lower panel})
abundance ratios as a function of the metallicity for the five sample
stars, compared with literature abundances. Symbols are the same as
in Fig. \ref{Na_Al_compara}.}
\label{Mg_Si_compara}
\end{figure}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{Ca_Ti_compara.eps}}
\caption{[Ca/Fe] (\textbf{upper panel}) and [Ti/Fe] (\textbf{lower panel})
abundance ratios as a function of the metallicity for the five sample
stars, compared with literature abundances. Symbols are the same as
in Fig. \ref{Na_Al_compara}.}
\label{Ca_Ti_compara}
\end{figure}
The sample stars analysed at high spectral resolution are confirmed to be
moderately metal-poor with $-1.04\leq$~[Fe/H]~$\leq-0.43$
with no enhancement in [C/Fe], and some stars show high
nitrogen abundances [N/Fe]: $+0.82\pm0.26$, $+0.71\pm0.34$, $+0.97\pm0.19$,
and $+0.35\pm0.20$, for 221, 230, 235, and 238, respectively.
In the context of APOGEE (Majewski et al. 2015), Schiavon et al. (2015) recently reported the
discovery of a population of Galactic bulge field stars
with high values of [N/Fe], which is correlated with [Al/Fe]
and anticorrelated with [C/Fe], typical of globular cluster stars
(Hesser et al. 1982; Gratton, Carretta \& Bragaglia 2012).
The N-rich stars in our sample could be related to the population
newly discovered by Schiavon et al. (2015). According to the authors,
abundance ratios [N/Fe]~$>+0.6$ cannot be
explained by the CN-cycle mixing scenario, and the contamination by
mass transfer binaries mechanism can only account for, at most,
$25$\% of their sample. Possible scenarios for the origin of the
N-rich stars are: i) dissolution of an early population of globular
clusters (Belokurov et al. 2006; Shapiro et al. 2010; Kruijssen 2015; Bournaud 2016);
ii) a shared (or similar) molecular cloud responsible for forming
these stars and the globular cluster (Longmore et al. 2014; Schiavon et al. 2015);
iii) these stars are among the oldest in the Galaxy and their abundances are imprints
of the first stellar generations (Tumlinson 2010; Chiappini et al. 2011).
To better place our results in the context of the Galactic
bulge, we present comparisons with literature abundances in
bulge stars. As already described in Sect. 5.1, we adopted results from
Yong et al. (2014) of seven stars in the globular cluster M~62,
the ninth most luminous Galactic globular cluster, which also
presents an extended horizontal branch. The stars
were observed with the High Dispersion Spectrograph
(HDS; Noguchi et al. 2002) on the Subaru Telescope
and with the Magellan Inamori Kyocera Echelle spectrograph (MIKE; Bernstein et al. 2003) at
the Magellan-II Telescope. The authors found a
good agreement between the scaled-solar r-process distribution
and the derived abundances for the elements heavier than La,
as well as an enhancement in Y, Zr, and Ba in comparison
with the solar r-process pattern. According to Yong et al. (2014),
these results are incompatible with the s-process in AGB stars
and suggest the fast-rotating massive stars as a possible
solution.
Also discussed in Sect. 5.1, Barbuy et al. (2014) analysed
four stars in the globular cluster NGC~6522, which
appears to be the oldest known Milky Way globular cluster.
The targets were observed at the VLT using the UVES spectrograph (Dekker et al. 2000)
in FLAMES-UVES mode. They found an enhancement in
s-process-dominant elements, suggesting spinstars as a possibility to form these elements,
besides the usual explanations of mass transfer from s-process-rich AGB stars and extra
mechanisms as the weak r-process as possible scenarios to explain the abundance signatures.
Ness et al. (2014) found similar results to Barbuy et al. (2014), but these authors insist
that the abundances of this cluster were measured to be similar to bulge field stars, halo stars,
and other Galactic globular clusters of the same metallicity. We note that NGC~6522 appears to be
among the oldest globular clusters, and as such it should show signatures as one of the main pieces
of the sub-systems that first formed in the central parts of the Galaxy.
In addition, we selected 62 red giant stars analysed in
Johnson et al. (2012), observed in Plaut's low-extinction
window. Using the Hydra multi-fiber spectrograph on the CTIO Blanco
$4$~m telescope, the stars were observed at
$l=-1^{\circ}$ and $b=-8.5^{\circ}$ (field 1) and at
$l=0^{\circ}$ and $b=-8^{\circ}$ (field 2).
Another 156 red giant branch stars in two Galactic bulge fields
centred near $l=+5.25^{\circ}$ and $b=-3.02^{\circ}$ and
$l=0^{\circ}$ and $b=-12^{\circ}$ analysed in
Johnson et al. (2014), using FLAMES-GIRAFFE spectra, were
selected in the comparison.
In Bensby et al. (2013), 58 microlensed bulge dwarfs and
subgiants stars were analysed. The authors
estimated the stellar ages based on isochrones
(Demarque et al. 2004) and probability distribution
functions (Bensby et al. 2011), so we selected
22 stars with ages older than 11 Gyr, avoiding the younger
stellar populations present in the bulge.
Finally, 56 other bulge giant stars were selected from
Van der Swaelmen et al. (2016), who analysed the
heavy elements in this sample. Already studied
in Zoccali et al. (2006), Lecureur et al. (2007),
and Barbuy et al. (2013, 2015), the observations were
performed with the multi-fibre spectrograph FLAMES-UVES,
at the UT2 Kuyen VLT/ESO telescope. The stars are located
at the Baade's Window ($l=1.14^{\circ}$, $b=-4.2^{\circ}$),
at the Blanco field ($l=0^{\circ}$, $b=-12^{\circ}$),
at the field of NGC~6553 ($l=5.2^{\circ}$, $b=-3^{\circ}$),
and in an additional field at $l=0.2^{\circ}$ and $b=-6^{\circ}$.
In Fig. \ref{Na_Al_compara} we show the comparisons for
the odd-Z elements sodium (upper panel) and aluminum
(lower panel), while the $\alpha$-elements Mg, Si, Ca, and Ti
are presented in Figs. \ref{Mg_Si_compara} and \ref{Ca_Ti_compara}.
As a general behaviour, the derived abundances in the sample
stars are in agreement with the literature, for objects with
the same metallicities, within the error bars.
Chemical similarities between globular cluster primary
stellar populations and field stars for a given metallicity
were studied in other works (see e.g. Renzini 2008;
Gratton et al. 2012; Schiavon et al. 2015).
Chemical inhomogeneity among stars within individual
globular clusters are well known for elements like C, N,
O, Na, Mg, and Al (see Kraft 1994; Gratton et al. 2004, 2012;
M\'esz\'aros et al. 2015), and several models are claimed
to explain this observational signature (Fenner et al. 2004;
Decressin et al. 2007; Renzini 2008; Marcolini et al. 2009;
Hopkins 2014; Renzini et al. 2015). To test these so-called abundance
anomalies, we checked if the Na~$-$~O, Al~$-$~O, and Al~$-$~Mg
anti-correlations were present in our sample, as well as
the Na~$-$~Al correlation, but no significant trend was found.
The $\alpha$-elements abundaces are enhanced, as typical of
chemical enrichment from core-collapse supernovae
(Woosley \& Weaver 1995; Nomoto et al. 2013, and references therein).
Regarding the heavy elements, we derived abundances of Y, Zr, Ba,
La, and Eu in our sample, and, in Figs. \ref{Y_Ba_compara} and
\ref{La_Eu_compara}, we show the results in comparison to values
from literature. A good agreement is observed
among the selected stars and, analogously to $\alpha$-elements,
enhancement in the heavy elements was obtained in the sample.
The only exception to this average behaviour in literature is
observed in Bensby et al. (2013), for which the major fraction
of stars shows solar values of [Y/Fe] and [Ba/Fe] abundance ratio.
The behaviour of the [Eu/Fe] abundance ratio is similar with that
observed for the $\alpha$-elements, which would be expected
from the main r-process. In the solar material, the r-process is responsible
for $94\pm0.4$\% of the total Eu abundance (Bisterzo et al. 2014).
However, most of the Y, Zr, Ba, and La available today in the solar system and in the
Galaxy has been produced by the s-process in low-mass AGB stars (Sneden et al. 2008, and
references therein) which, owing to the typical long lifetimes of low-mass stars, would
not have had time to evolve and pollute the gas before forming the sample stars, which are
probably very old. Possible scenarios for the enrichment of Y, Zr, Ba, and La derived
in our sample are:
i) Early enrichment from r-process only, where an extra mechanism
is required to produce excesses of the lightest
trans-Fe elements with respect to second peak elements
such as Ba and La (e.g. Travaglio et al. 2004; Wanajo et al. 2011; Arcones \& Montes 2011;
Arcones \& Thielemann 2013; Fujibayashi et al. 2015; Niu et al. 2015);
ii) s-process elements from AGB stars bounded in a binary system,
polluting the observed stars via AGB-mass transfer
(Bisterzo et al. 2014);
iii) s-process activation in early generations of Spinstars (Meynet et al. 2006; Pignatari et al. 2008;
Frischknecht et al. 2012; Frischknecht et al. 2015), which pollutes the primordial material before
forming the oldest bulge field stars.
Figure \ref{YxBa_compara} shows the [Y/Ba] vs. [Fe/H] diagram for bulge
stars from our sample compared with selected results from the literature.
From Bensby et al. (2013), we only retained the star enhanced in [Y/Fe].
We included the average [Y/Ba]$_{r}=-0.42\pm0.12$
abundance ratio value obtained (yellow region) from six halo metal-poor r-element-rich
stars (HD~221170, HD~115444, CB~22892-052, HE~1523-0901,
BD~17~3248, and CS~31082-001), compiled in Sneden et al. (2008),
as a representative value of the main r-process.
We also show the mean value
of the [Y/Ba] ratio obtained from six halo metal-poor r-process stars showing enhancement
in the lightest heavy elements: HD~88609 (Honda et al. 2007), BD~4~2621 (Johnson 2002),
HD~4306 (Honda et al. 2004), HD~237846 (Roederer et al. 2010),
HD~122563 (Honda et al. 2007), and HD~140283 (Siqueira-Mello et al. 2015).
The average value is [Y/Ba]$_{E}=+0.56\pm0.18$ (illustrated in the figure by the cyan
region - where E stands for ``enhanced''). The source of this enhancement can be manifold
(see discussion below). We note that the cyan region is meant to only show a reference value since,
even for the same nucleosynthetic sources, the expected level of enhancement in the bulge and
halo will be different (see Barbuy et al. 2014).
The figure shows that the value derived in star 235 agrees with those observed in the
r-element-rich stars, and it can be explained by a pattern arising in a
so-called main r-process. Considering the error bars, the same conclusion
could be claimed to explain the [Y/Ba] abundance in star 238.
On the other hand, the derived abundances in 221, 224, and 230 are barely explained using
only the main r-process. If the bulge stars with [Fe/H]~$\sim-1$ trace the same early phases of chemical
enrichment as halo stars with [Fe/H]~$\sim-3$, the sample stars 235 and
238 may be classified as r-process enhanced stars, analogous to the r-I and r-II metal-poor
halo stars (Beers \& Christlieb 2005). In fact, using the latter authors' definitions that are
based on the [Ba/Eu] abundance ratio, objects 235 and 238 must be classified as r/s and r-I stars,
respectively. This is the first time that these kind of stars are identified in the Galactic bulge.
Figure \ref{ZrxBa_compara} presents the same comparison for zirconium.
In the upper panel the [Zr/Fe] abundance ratios derived in the sample
are compared with results from the literature. Johnson et al. (2012)
identified evidence of two separate sequences: a group of stars enhanced
in [Zr/Fe], and another group moderately poor. Clearly our sample stars
are members of the enhanced group. In the lower panel we show the
[Zr/Ba] vs. [Fe/H] diagram for the sample stars and the selected results from
literature. The selected metal-poor r-element-rich and the enhanced stars
in lightest heavy elements were also used to define [Zr/Ba]$_{r}=-0.18\pm0.12$
(yellow region) and [Zr/Ba]$_{E}=+0.95\pm0.15$ (cyan region), respectively.
The figure shows that the Zr abundances derived in stars 235 and 238 also agree
with those observed in the r-element-rich stars, while stars 221, 224, and 230
require extra mechanism(s) to explain the abundance ratios.
For the Galactic halo, Roederer et al. (2010) show several metal-poor stars located
in the region between these two extremes abundance regimes, suggesting also a continuous
range of r-process nucleosynthesis patterns. On the other hand,
Niu et al. (2015) more recently suggest that the weak r-process
and the main r-process are two distinct astrophysical processes.
The fundamental challenge that we are facing is that in the early galaxy a
multitude of different nucleosynthesis processes may have contributed to the production
of the elements at the first neutron-magic peak beyond iron, including Sr, Y, and Zr.
Together with the s-process in fast rotating massive stars (Frischknecht et al. 2016) and the weak r-process
(e.g., Farouqi et al. 2009), other sources could be at play such as: the electron capture supernovae
(e.g., Wanajo et al. 2011), or the $\alpha$-rich freezout in most energetic core-collapse
supernovae (e.g., Woosley \& Hoffman 1992), and the intermediary neutron capture i-process
(Dardelet et al. 2015, and references therein). In addition, neutrino-winds in core-collapse supernovae can
host a large variety of processes that can produce elements in the same mass region
(e.g., Fr\"ohlich et al. 2006, Farouqi et al. 2010, Roberts et al. 2010, Arcones \& Montes 2011).
It is thus crucial to measure as many heavy elements as possible to isolate the different nucleosynthesis sources.
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{Y_Ba_compara.eps}}
\caption{[Y/Fe] (\textbf{upper panel}) and [Ba/Fe] (\textbf{lower panel})
abundance ratios as a function of the metallicity for the five sample
stars, compared with literature abundances. Symbols are the same as
in Fig. \ref{Na_Al_compara}, in addition to the abundances from
Van der Swaelmen et al. (2016; filled grey circles).}
\label{Y_Ba_compara}
\end{figure}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{La_Eu_compara.eps}}
\caption{[La/Fe] (\textbf{upper panel}) and [Eu/Fe] (\textbf{lower panel})
abundance ratios as a function of the metallicity for the five sample
stars, compared with literature abundances. Symbols are the same as
in Figs. \ref{Na_Al_compara} and \ref{Y_Ba_compara}.}
\label{La_Eu_compara}
\end{figure}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{YxBa_compara.eps}}
\caption{[Y/Ba] vs. [Fe/H] diagram for the sample stars
and results from literature. Symbols are the same as
in Figs. \ref{Na_Al_compara}. The yellow and cyan
regions correspond to the main r-process signature
and the abundance ratio from metal-poor stars enhanced in
the lightest heavy elements (see text for details).}
\label{YxBa_compara}
\end{figure}
\begin{figure}
\centering
\resizebox{90mm}{!}{\includegraphics[angle=0]{ZrxBa.eps}}
\caption{\textbf{Upper panel}: [Zr/Fe] abundance ratios as a
function of the metallicity for the five sample stars, compared with
literature abundances. Symbols are the same as in Figs. \ref{Na_Al_compara}.
\textbf{Lower panel:} [Zr/Ba] vs. [Fe/H] diagram for the
sample stars and results from literature. Symbols are the same as
in Figs. \ref{Na_Al_compara}. The yellow and cyan
regions correspond to the main r-process signature
and the abundance ratio from metal-poor stars enhanced in
the lightest heavy elements (see text for details).}
\label{ZrxBa_compara}
\end{figure}
\section{Conclusions}
We have carried out a pilot project with the goal of providing detailed abundances
for moderately metal-poor Galactic bulge stars that are believed to host imprints left by the first
stellar generations. In this work, we were able to obtain detailed abundances for five moderately
metal-poor and [$\alpha$/Fe]~$>0$ stars from one field of the ARGOS survey. Our high-resolution
FLAMES-UVES spectra have confirmed three out of five stars to have metallicities [Fe/H]~$<-0.8$.
All stars are confirmed to be $\alpha$-enhanced: overabundances of the typical
$\alpha$'s Mg, Si, Ca, Ti, and of the odd-Z element Al are clearly detected.
Three sample stars exhibit high [N/Fe] abundance ratios. Similar high-[N/Fe] bulge stars
were recently found in APOGEE (see Schiavon et al. 2015). According to the latter authors, abundance
ratios [N/Fe]~$>+0.6$ cannot be explained by the CN-cycle mixing scenario.
The sample stars show enhancements in [Y/Fe], [Zr/Fe], [Ba/Fe], [La/Fe], and [Eu/Fe]. We found that
three of our stars also show [Y/Ba] and [Zr/Ba] ratios slightly higher than expected from a pure main
r-process nucleosynthesis. These results are very similar to the recent reported chemical pattern found in some stars of the
oldest Milky Way globular cluster, NGC~6522 (Barbuy et al. 2014).
Considering the sample stars as an old population, whereas an enhancement in Eu would be expected
from the rapid neutron capture process (or r-process), consistent with the observed
[$\alpha$/Fe] enhancements, the observed anomalous enrichment of the
dominantly s-process elements [Y/Ba] and [Zr/Ba] in these stars are more difficult to understand using
standard nucleosynthesis processes.
Finally, previous to the present work, some excesses of s-process-typical elements in the Galactic bulge
had been found only in globular clusters (Barbuy et al. 2009; Chiappini et al. 2011; Barbuy et al. 2014; Yong et al. 2014).
The goal of our pilot project was to also look for the existence of these stars in the field.
Although our sample is very small, three of our stars seem to show not only excesses of the lightest heavy elements with respect to iron,
but also enhancement in the [Y/Ba] and [Zr/Ba] abundance ratios. The s-process activation in fast-rotating massive stars
and/or other extra mechanisms are possible solutions to these anomalies. There is a debate in the literature
about the origin of heavy elements in the oldest stars, such that
future large samples are urgently needed to futher explore the impact of these findings in our understanding of
the nature of the first stellar generations.
\begin{acknowledgements}
CS, BB, and EC acknowledge grants from CAPES, CNPq and FAPESP.
MP acknowledges support from NSF grants PHY 02-16783 and PHY 09-22648
(Joint Institute for Nuclear Astrophysics, JINA), NSF grant PHY-1430152
(JINA Center for the Evolution of the Elements) and EU MIRG-CT-2006-046520.
The continued work on codes and in disseminating data is made possible through
funding from STFC and EU-FP7-ERC-2012-St Grant 306901 (RH, UK).
MP acknowledges support from the Lendulet-2014 Programme of the Hungarian Academy of
Sciences and from SNF (Switzerland).
The research leading to these results has received funding from
the European Research Council under the European Union's Seventh
Framework Programme (FP/2007-2013) / ERC Grant Agreement n. 306901.
R. Hirschi acknowledges support from the World Premier International
Research Center Initiative (WPI Initiative), MEXT, Japan.
This work has made use of the VALD database, operated at
Uppsala University, the Institute of Astronomy RAS in Moscow,
and the University of Vienna.
\end{acknowledgements}
|
\section{Introduction}
Given a graph $H$, let $f(n,H)$ be the maximum number of edges not contained in any monochromatic
copy of $H$ in a 2-edge-coloring of the complete graph $K_n$, and let $ex(n,H)$ be the {\it Tur\'an number} of $H$,
i.e., the maximum number of edges in an $n$-vertex $H$-free graph.
The problem of determining $f(n,H)$ was motivated by counting the number of monochromatic cliques,
and we refer interested readers to \cite{KS2} for a thoughtful discussion on the background and related topics.
(For results on monochromatic cliques, see \cite{G59,T97,E97,EFGJL,KS1,Y08,CKPSTY}.)
If one considers the 2-edge-coloring of $K_n$ in which one of the colors induces the largest $H$-free graph,
then it is easy to see that for any $H$ and $n$, we have
\begin{align}\label{equ:f>=ex}
f(n,H)\ge ex(n,H).
\end{align}
Erd\H{o}s, Rousseau and Schelp (see \cite{E97}) showed that $f(n,K_3)=ex(n,K_3)$ for sufficiently large $n$,
and this also can be derived from a result of Pyber in \cite{P86} for $n\ge 2^{1500}$.
The generalization of this result was suggested by Erd\H{o}s in \cite{E97}.
Keevash and Sudakov \cite{KS2} studied general graphs and asked that if,
for large $n$, the above lower bound \eqref{equ:f>=ex} is tight.
\begin{prob} (\cite{KS2})\label{prob:main}
Let $H$ be a fixed graph. Is it true that for $n$ sufficiently large, $f(n,H)=ex(n,H)$?
\end{prob}
The authors of \cite{KS2} confirmed it for $H$ being any edge-color-critical graph or a $C_4$,
and in fact, quite amazingly, they were able to determine the value of $f(n,H)$ for every $n$ when $H$ is a $K_3$ or $C_4$.
We quote from their remark \cite{KS2} that ``for bipartite graphs the situation is less clear,
as even the asymptotics of the Tur\'an numbers are known only in a few cases".
In this paper, we provide an affirmative answer to Problem \ref{prob:main} for an abundant infinite family of bipartite graphs.
A vertex $w$ in a bipartite graph $H$ is called {\it weak}, if
$$ex(n,H-w)=o(ex(n,H)).$$
The notation of weak vertices is explicitly defined in the literature and has been well studied (see \cite{S84}).
We call a bipartite graph $H$ {\it reducible}, if it contains a weak vertex $w$
such that $H-w$ is connected. For instance all even cycles are reducible.
Our main theorem is as follows.
\begin{theorem}\label{thm:main}
Let $H$ be a reducible bipartite graph. Then for sufficiently large $n$, $f(n,H)=ex(n,H)$.
Moreover, a 2-edge-colorings of $K_n$ achieves the maximum number $f(n,H)$
if and only if one of the color classes induces an extremal graph for $ex(n,H)$.
\end{theorem}
We point out that the ``moreover" part is new for $C_4$,
while its analog is not true for edge-color-critical graphs as noticed in \cite{KS2}.
Let $\mathcal{C}^*$ be the family of bipartite graphs,
each of which contains a cycle and a vertex $w$ whose deletion will result in a tree.
It is easy to see that all graphs in $\mathcal{C}^*$, including even cycles and Theta graphs,\footnote{The Theta
graphs $\theta_{k,l}$ denotes the graph consisting of $k$ internally disjoint paths of length $l$ between
two fixed endpoints, for $k,l\ge 2$.} are reducible.
Based on the current knowledge on degenerated Tur\'an numbers,
we collect some reducible graphs in the coming result.
\begin{coro}\label{Coro}
For $n$ sufficiently large, $f(n,H)=ex(n,H)$ holds for every $H$ as following: even cycles $C_{2l}$,
Theta graphs $\theta_{k,l}$,
and complete bipartite graphs $K_{s,t}$ for $t>s^2-3s+3$ or $(s,t)\in\{(3,3),(4,7)\}$.
\end{coro}
The rest of the paper is organized as follows. In the next section, we prove Theorem \ref{thm:main}
in full and then derive Corollary \ref{Coro}. In Section 3, we generalize Theorem \ref{thm:main} to multi-colorings.
In the final section, we close this paper by mentioning some related problems.
\section{Reducible bipartite graphs}
Let $H$ be a fixed graph and $c$ be a $k$-edge-coloring of $K_n$.
An edge of $K_n$ is called {\it NIM-H},
if it is not contained in any monochromatic copy of $H$ in $c$.
Let $E_c$ denote the set of all NIM-H edges of $K_n$.
For $A,B\subseteq V(K_n)$, by $(A,B)$ we denote the complete bipartite graph with two parts $A$ and $B$.
In this section, we establish Theorem \ref{thm:main} and Corollary \ref{Coro}.
To do so, we prove the following stronger result.
\begin{theorem}\label{thm:str}
Let $H$ be a reducible bipartite graph.
If $c$ is a 2-edge-coloring of $K_n$ such that $E_c$ contains a red edge and a blue edge,
then $|E_c|=o(ex(n,H))$.
\end{theorem}
\begin{pf}
Let $h=|V(H)|$, $(X,Y)$ be the bipartition of $H$, and $w\in X$ be a weak vertex of $H$
such that $ex(n,H-w)=o(ex(n,H))$ and $H-w$ is connected. Note that $(X-w,Y)$ is the unique bipartition of $H-w$, as $H-w$ is connected.
We first define a red star $S_0$ in $K_n$ (i.e., all edges in the star are red), which contains at least one red NIM-H edge, as follows.
If there exist vertices incident with a red NIM-H edge and at least $h$ red edges,
then pick one such vertex $x$ and form a star $S_0$ consisting of the center $x$ and $h$ red neighbors of $x$
such that $xv$ is a red NIM-H edge for some $v\in V(S_0)$.
Otherwise every vertex incident with a red NIM-H edge has less than $h$ red neighbors,
then pick one such vertex $x$ with maximum number of red neighbors and let $S_0$ consist of $x$ and all its red neighbors.
Similarly as above, we define a blue star $S_1$ in $K_n$, which contains at least one blue NIM-H edge.
Let $x, y$ be the centers of the stars $S_0, S_1$, respectively.
Note that $S_0$ and $S_1$ may share some common vertices. We let $$S=V(S_0)\cup V(S_1)=\{s_1,s_2,...,s_t\}.$$
So $t=|S|\le 2h+2$. For $z\in V(K_n)\backslash S$, let $\vec{\epsilon}(z)=(\epsilon_1,\epsilon_2,...,\epsilon_t)$ be the vector
such that
\begin{equation}
\epsilon_i=\left\{\begin{array}{ll}
0 & \text{if } zs_i \text{ is red}, \\
1 & \text{if } zs_i \text{ is blue}.
\end{array}\right.
\end{equation}
For $\vec{v}\in \{0,1\}^t$, let $A_{\vec{v}}$ denote the set of all vertices $z\in V(K_n)\backslash S$
such that $\vec{\epsilon}(z)=\vec{v}$.
Observe that all edges between $s_i\in S$ and $A_{\vec{v}}$ must be monochromatic.
We now consider the numbers of NIM-H edges adjacent to sets $A_{\vec{v}}$.
The first claim implies that the number of NIM-H edges adjacent to $A_{\vec{0}}\cup A_{\vec{1}}$ is $O(n)$.
\medskip
{\bf Claim 1}: $|A_{\vec{0}}|<h$ and $|A_{\vec{1}}|<h$.
By symmetry, it suffices to consider $A_{\vec{0}}$.
We notice that all edges in $(A_{\vec{0}},S)$ are red.
Suppose for a contradiction that $|A_{\vec{0}}|\ge h$.
If the red star $S_0$ has less than $h+1$ vertices, then it is clear
that no vertex in $V(K_n)\backslash S$ can be adjacent to $x$, implying that $A_{\vec{0}}=\emptyset$.
So the red star $S_0$ has exactly $h+1$ vertices.
We see that all edges in $(A_{\vec{0}}\cup \{x\},S_0-\{x\})$ are red.
From this, one can easily find a red copy of $H$ which uses one NIM-H edge of $x$,
contradicting the definition of NIM-H edges. This proves claim 1.
\medskip
{\bf Claim 2}: For $\vec{v}\in \{0,1\}^t-\{\vec{0},\vec{1}\}$, the number of NIM-H edges contained in $A_{\vec{v}}$ is at most $2\cdot ex(n,H-w)$.
As $\vec{v}\notin \{\vec{0},\vec{1}\}$, there exist $a,b\in S$ such that all edges
in $(a,A_{\vec{v}})$ are red and all edges in $(b,A_{\vec{v}})$ are blue.
If the red NIM-H edges in $A_{\vec{v}}$ form a copy $K=(X-w,Y)$ of $H-w$,
then $K\cup \{a\}$ would contain a red copy of $H$ with some NIM-H edges, a contradiction.
Therefore, neither the red NIM-H edges nor the blue NIM-H edges can form a copy of $H-w$.
This proves claim 2.
\medskip
{\bf Claim 3}: For $\vec{v},\vec{u}\in \{0,1\}^t-\{\vec{0},\vec{1}\}$, the number of NIM-H edges in
$(A_{\vec{v}},A_{\vec{u}})$ is at most $2\cdot ex(n,H-w)$.
Suppose that the red NIM-H edges in $(A_{\vec{v}},A_{\vec{u}})$ form a copy $K=(X-w,Y)$ of $H-w$.
By symmetry, we assume that $X-w\subseteq A_{\vec{v}}$ and $Y\subseteq A_{\vec{u}}$.
Since $\vec{u}\neq \vec{1}$, there exists $a\in S$ such that all edges in $(a,A_{\vec{u}})$ are red.
Adding $a$ and all red edges in $(a,Y)$ to $K$ would result in a red copy of $H$, a contradiction.
Therefore, neither the red NIM-H edges nor the blue NIM-H edges in $(A_{\vec{v}},A_{\vec{u}})$ can form a copy of $H-w$.
Claim 3 is finished.
\medskip
Each edge in $E_c$ is either adjacent to $S\cup A_{\vec{0}}\cup A_{\vec{1}}$
or contained in $A_{\vec{v}}$ or $(A_{\vec{v}},A_{\vec{u}})$ for some $\vec{v},\vec{u}\in \{0,1\}^t-\{\vec{0},\vec{1}\}$.
Since $|S|=t\le 2h+2$, there are at most $2^{2h+2}$ sets $A_{\vec{v}}$. Combining the above claims, we have
\begin{align*}
|E_c|&\le (|S|+2h)\cdot n +2^{2h+2}\cdot 2\cdot ex(n,H-w)+(2^{2h+2})^2\cdot 2\cdot ex(n,H-w)\\
&\le 2^{4h+6}\cdot ex(n,H-w)=o(ex(n,H)).
\end{align*}
This finishes the proof of Theorem \ref{thm:str}.
\qed
\end{pf}
\bigskip
We are ready to prove Theorem \ref{thm:main}.
\medskip
\noindent{\it Proof of Theorem \ref{thm:main}.}
We have seen $f(n,H)\ge ex(n,H)$ from \eqref{equ:f>=ex}.
Let $c$ be a 2-edge-coloring of $K_n$ such that $|E_c|=f(n,H)$.
If $E_c$ contains a red edge and a blue edge, then by Theorem \ref{thm:str},
we have $ex(n,H)\le f(n,H)=|E_c|=o(ex(n,H))$, a contradiction.
So we may assume that all NIM-H edges are red.
It then becomes clear that $E_c$ does not contain any copy of $H$,
implying that $f(n,H)=|E_c|\le ex(n,H)$. This proves that $f(n,H)=ex(n,H)$ for large $n$.
It also follows that $|E_c|=ex(n,H)$. So $E_c$ must induce an extremal graph for $ex(n,H)$.
We claim that except these edges in $E_c$, no other edge can be red.
Suppose not, say $e\in E(K_n)-E_c$ is red.
Then $E_c\cup \{e\}$ induces an $n$-vertex graph with more than $ex(n,H)$,
which must contain a copy of $H$. But this $H$ contains all red edges
and in particular some NIM-H edges from $E_c$, a contradiction. This proves the claim.
Now we see that all red edges of $c$ induces an extremal graph for $ex(n,H)$.
To prove the ``moreover" part, it remains to show that
if all red edges of $c$ induces an extremal graph for $ex(n,H)$, then $|E_c|=f(n,H)=ex(n,H)$.
Since all red edges surely are NIM-H, we have $|E_c|\ge ex(n,H)$.
So we need to show that no blue edge can be NIM-H. This, again, can be derived from Theorem \ref{thm:str}.
We have finished the proof. \qed
\bigskip
We conclude this section by showing Corollary \ref{Coro}.
Recall the seminal theorem of K\H{o}v\'ari-S\'os-Tur\'an \cite{KST} and
the best known general lower bound on Tur\'an number of $K_{s,t}$ that
$$\Omega(n^{2-\frac{s+t-2}{st-1}})\le ex(n,K_{s,t})\le \frac12(t-1)^{1/s}n^{2-1/s}+\frac12(s-1)n.$$
We also need the result (see \cite{KRS,ARS}) that $ex(n,K_{s,t})\ge \Omega(n^{2-1/s})$ for $t>(s-1)!$.
\medskip
\noindent{\it Proof of Corollary \ref{Coro}.}
In view of Theorem \ref{thm:main}, it is enough to show that every graph $H$ in the list is reducible.
As it is clear that even cycles and Theta graphs are reducible, we only need to consider $K_{s,t}$.
When $t>(s-1)!$, it holds that $$\frac{ex(n,K_{s-1,t})}{ex(n,K_{s,t})}=O\left(\frac{n^{2-1/(s-1)}}{n^{2-1/s}}\right)=o(1),$$
and when $t>s^2-3s+3$, we have $\frac{s+t-2}{st-1}<\frac{1}{s-1}$, implying that
$$\frac{ex(n,K_{s-1,t})}{ex(n,K_{s,t})}= O\left(\frac{n^{2-1/(s-1)}}{n^{2-(s+t-2)/(st-1)}}\right)=o(1).$$
Therefore, $K_{s,t}$ is reducible whenever $t>\min\{s^2-3s+3,(s-1)!\}$, finishing the proof.\qed
\section{Generalization to multi-colorings}
In this section, we consider multi-color versions of Theorem \ref{thm:main}.
For $k\ge 3$, let $f_k(n,H)$ denote the maximum number of edges not contained in any monochromatic
copy of $H$ in a $k$-edge-coloring of the complete graph $K_n$.
Given a bipartition $(X,Y)$ of bipartite $H$, let $ex^*(m,n,H)$ denote the maximum number of edges of
graphs $G$, where $G$ is a spanning subgraph of $K_{m,n}$ and has no copies of $H=(X,Y)$ with $X$ contained in the $m$-part.\footnote{The Zarankiewicz function $z(m,n,s,t)$ is just the same as $ex^*(m,n,K_{s,t})$.}
We first prove a general lower bound for every bipartite graph $H$ that
\begin{align}\label{equ:fk>=ex}
f_k(n,H)\ge (k-1)\cdot ex(n,H)- O\left(ex(n,H)/n\right)^2=(k-1-o(1))\cdot ex(n,H).
\end{align}
\noindent{\it Proof.}
Let $G$ be an $n$-vertex $H$-free extremal graph for $ex(n,H)$.
For a permutation $\pi$ on $V(G)$, let $G(\pi)$ be obtained from $G$
by permuting all edges according to $\pi$, i.e., $E(G(\pi))=\pi(E(G))$.
Take $k-1$ random permutations $\pi_1,\pi_2,...,\pi_{k-1}$ and consider the overlap $E_{ij}=E(G(\pi_i))\cap E(G(\pi_j))$.
Since the probability that each $e\in \binom{V}{2}$ belongs to $G(\pi_i)$ equals $ex(n,H)/\binom{n}{2}$,
the expectation of $\sum_{i,j} |E_{ij}|$ is at most $\binom{k-1}{2} ex(n,H)^2/\binom{n}{2}$.
Therefore, there exist permutations $\pi_1,\pi_2,...,\pi_{k-1}$ such that
the total overlap $\sum_{i,j} |E_{ij}|$ is at most $\binom{k-1}{2} ex(n,H)^2/\binom{n}{2}$.
We then define a $k$-edge-coloring $c$ of $K_n$ as following.
Color the edges of $G(\pi_1)$ by color 1; and for $2\le i\le k-1$,
color the edges in $E(G(\pi_i))-\cup_{1\le j\le i-1} E_{ij}$ by color $i$;
and lastly, color all edges of $K_n$ not in $\cup_{1\le i\le k-1} E(G(\pi_i))$ by color $k$.
This implies that
$f_k(n,H)\ge |\cup_{1\le i\le k-1} E(G(\pi_i))|\ge (k-1)\cdot ex(n,H)-\sum_{i,j}|E_{ij}|,$
which is at least $(k-1)\cdot ex(n,H)- O\left(ex(n,H)/n\right)^2$. \qed
\begin{theorem}\label{thm:k-C4}
For $n$ sufficiently large, we have $$(k-1-o(1))\cdot ex(n,C_4)\le f_k(n,C_4)\le (k-1)\cdot ex(n,C_4).$$
\end{theorem}
\begin{theorem}\label{thm:multi}
Let $H$ be a bipartite graph with a vertex $w$ such that
$ex^*(n,n,H-w)=o(ex(n,H)).$
Then for sufficiently large $n$, $$(2-o(1))\cdot ex(n,H)\le f_3(n,H)\le 2\cdot ex(n,H).$$
Such graphs $H$ include even cycles $C_{2l}$ and
complete bipartite graphs $K_{s,t}$ for $t>s^2-3s+3$ or $(s,t)\in\{(3,3),(4,7)\}$.
\end{theorem}
\begin{pf} (For both Theorems \ref{thm:k-C4} and \ref{thm:multi}.)
The lower bound follows from $\eqref{equ:fk>=ex}$.
First we prove an analog of Theorem \ref{thm:str}.
Let $(X,Y)$ be the partition of $H$ which $ex^*(n,n,H-w)$ refers to. Let $w\in X$ and $h=|V(H)|$.
Call an edge with color $i$ as an $i$-edge for convenience.
{\bf Claim:} Let $c$ be a $k$-edge-coloring of $K_n$. If $E_c$ contains a NIM-$H$ $i$-edge
for each $i\in [k]$, then $|E_c|\le (k-2)\cdot ex(n,H)+o(ex(n,H))$.
The proof of this claim will follow the same lines of Theorem \ref{thm:str}.
For each color $i\in [k]$, we define a star $S_i$ in $K_n$ consisting of $i$-edges,
among which there is at least one NIM-H $i$-edge.
If there exist vertices incident with a NIM-H $i$-edge and at least $h$ $i$-edges,
then pick one such vertex $x_i$ and form a star $S_i$ with the center $x_i$ and consisting of $h$ $i$-edges
such that there exists at least one NIM-H $i$-edge $x_iv$ for some $v\in V(S_i)$.
Otherwise every vertex incident with a NIM-H $i$-edge has less than $h$ $i$-neighbors,
then pick one such vertex $x_i$ with maximum number of $i$-neighbors and let $S_i$ consist of $x_i$ and all its $i$-neighbors.
Let $S=\cup_{i\in [k]}V(S_i)=\{s_1,s_2,...,s_t\}.$
So $t=|S|\le k(h+1)$. For $z\in V(K_n)\backslash S$,
let $\vec{\epsilon}(z)=(\epsilon_1,\epsilon_2,...,\epsilon_t)$ be the vector
such that $\epsilon_i=j$ iff $zs_i$ is colored by $j$.
For $\vec{v}\in [k]^t$, let $A_{\vec{v}}$ denote the set of all vertices $z\in V(K_n)\backslash S$
such that $\vec{\epsilon}(z)=\vec{v}$.
For some $I\subseteq [k]$, we say $A_{\vec{v}}$ is {\it $I$-feasible}, if for each $i\in I$ there exists some coordinate in $\vec{v}$ being $i$,
and subject to this, $I$ is maximal. We then establish the following three assertions.
\medskip
{\bf (1).} For each $i\in [k]$, we have $|A_{\vec{i}}|<h$.
Note that all edges in $(A_{\vec{i}}, S)$ are of color $i$. If $|A_{\vec{i}}|\ge h$,
then the complete bipartite graph $(A_{\vec{i}}\cup \{x_i\}, S_i-\{x_i\})$ contains
a copy of $H$ of all $i$-edges with at least one NIM-H $i$-edge (incident to $x_i$),
a contradiction. This proves (1).
\medskip
{\bf (2).} For $i\in I$, the $I$-feasible set $A_{\vec{v}}$ has no more than $ex(n,H-w)$ NIM-H $i$-edges.
Suppose that the NIM-H $i$-edges in $A_{\vec{v}}$ form a copy $K$ of $H-w$.
Since $i\in I$, there exists some $a\in S$ such that the edges
in $(a, A_{\vec{v}})$ are all of color $i$.
Then adding $a$ into $K$ would give a copy of $H$ of color $i$
which also contains NIM-H edges, a contradiction.
This shows that there are no more than $ex(n,H-w)$ NIM-H $i$-edges in $A_{\vec{v}}$, establishing (2).
\medskip
{\bf (3).} Let $A_{\vec{u}}$ be $I$-feasible and $A_{\vec{v}}$ be $J$-feasible.
For $i\in I\cup J$, there are no more than $ex^*(n,n,H-w)$ NIM-H $i$-edges in $(A_{\vec{u}},A_{\vec{v}})$.
For $i\in I\cup J$, there exists some coordinate in $\vec{u}$ or $\vec{v}$ being $i$.
By symmetry, say this coordinate is from $\vec{v}$. Then there exists some $a\in S$
such that all edges from $a$ to $A_{\vec{v}}$ are of color $i$.
Suppose that the NIM-H $i$-edges in $(A_{\vec{u}},A_{\vec{v}})$
contains a copy $K$ of $H-w=(X-w,Y)$ with $X-w\subseteq A_{\vec{u}}$ and $Y\subseteq A_{\vec{v}}$.
Then $\{a\}\cup K$ would contain a copy of $H$ of color $i$ with some NIM-H edges,
a contradiction. Thus $(A_{\vec{u}},A_{\vec{v}})$ has no more than
$ex^*(|A_{\vec{u}}|,|A_{\vec{v}}|,H-w)\le ex^*(n,n,H-w)$ NIM-H $i$-edges. This proves (3).
\medskip
Observe that the NIM-H edges not in (2) and (3)
are of the following three types:
\begin{itemize}
\item[(i).] NIM-H edges which are adjacent to $S$ or $A_{\vec{i}}$ for some $i\in [k]$,
\item[(ii).] NIM-H $i$-edges in $I$-feasible set $A_{\vec{v}}$, where $i\in I^c$ and $|I|\ge 2$, and
\item[(iii).] NIM-H $i$-edges between $I$-feasible set $A_{\vec{u}}$ and $J$-feasible set $A_{\vec{v}}$,
where $i\notin I\cup J$ (or equivalently $i\in I^c\cap J^c$) and $|I|,|J|\ge 2$.
\end{itemize}
Note that $|S|=t\le k(h+1)$. So there are at most $k^{k(h+1)}$ sets $A_{\vec{v}}$, which is constantly many.
Also note that $ex(n,H-w)=O(ex^*(n,n,H-w))=o(ex(n,H))$,
implying that $H$ must not be a forest and thus $ex(n,H)=\Omega(n^{1+c})$ for some $c>0$.
These, combining with the above assertions, imply that
there are just $o(ex(n,H))$ NIM-H edges contained in (2), (3) and (i).
To complete the proof of the claim, it then suffices to show that the number $N^*$ of NIM-H edges
in (ii) and (iii) is at most $(k-2)\cdot ex(n,H)+o(ex(n,H))$.
For $i\in [k]$, denote $B_i$ to be the union of all $I$-feasible sets $A_{\vec{v}}$ satisfying $i\in I^c$ and $|I|\ge 2$.
It is straightforward to verify that the NIM-H $i$-edges in (ii) and (iii) must be contained in $B_i$.
Thus the number of NIM-H edges in (ii) and (iii) is $N^*\le \sum_{i=1}^k ex(b_i,H),$ where $b_i=|B_i|$.
Note each $b_i\le n$ and $\sum_{i=1}^k b_i\le (k-2)\cdot\sum |A_{\vec{v}}|\le (k-2)n$,
as every vertex in $I$-feasible sets with $|I|\ge 2$ only can appear in at most $k-2$ $B_i$'s.
When $k=3$, we have $b_1+b_2+b_3\le n$, so it is clear that $N^*\le \sum_{i=1}^3 ex(b_i,H)\le ex(n,H).$
When $H=C_4$, using the well-known result (see \cite{ERS,B66,KST}) that $ex(n,C_4)=(1/2+o(1))\cdot n^{3/2}$,
it holds that $N^*\le \sum_{i=1}^k ex(b_i,C_4)\le \frac{1}{2}\sum_{i=1}^k b_i^{3/2}+o(n^{3/2})$.
Subject to $b_i\le n$ and $\sum_{i=1}^k b_i\le (k-2)n$, by convexity, we have
$N^*\le \frac{k-2}{2}\cdot n^{3/2}+o(n^{3/2})=(k-2)\cdot ex(n,C_4)+o(ex(n,C_4))$, which is desired.
This completes the proof of the claim.
\medskip
Next we prove the upper bound of $f_k(n,H)$.
Let $c$ be a $k$-edge-coloring of $K_n$ such that $|E_c|=f_k(n,H)$.
If $E_c$ contains a NIM-H $i$-edge for every $i\in [k]$,
then by the claim, we have
$(k-1-o(1))\cdot ex(n,H)\le f_k(n,H)=|E_c|\le (k-2)\cdot ex(n,H)+o(ex(n,H))$, a contradiction.
Therefore $E_c$ has at most $k-1$ colors.
The set of NIM-H edges of the same color contains none copy of $H$ and thus is of size at most $ex(n,H)$.
Thus, $f_k(n,H)=|E_c|\le (k-1)\cdot ex(n,H)$.
It remains to verify that $ex^*(n,n,H-w)=o(ex(n,H))$ holds for $H$ being an even cycle or $K_{s,t}$
for $t>\min\{s^2-3s+3,(s-1)!\}$. This follows by the same proof of Corollary \ref{Coro},
using $ex^*(m,n,K_{s,t})\le (t-1)^{1/s}mn^{1-1/s}+(s-1)n$ (see \cite{KST}).
This proves Theorems \ref{thm:k-C4} and \ref{thm:multi}. \qed
\end{pf}
\section{Concluding remarks}
In Theorem \ref{thm:main} we prove that $f(n,H)$ equals $ex(n,H)$ for bipartite graphs $H$
having weak vertices (for sufficiently large $n$).
Simonovits asked in \cite{S84} to ``characterize those bipartite graphs which have weak vertices" and this remains unclaimed.
It seems that for $k\ge 3$, the function $f_k(n,H)$ has a different behavior between bipartite and non-bipartite graphs.
For bipartite $H$, it may be reasonable to ask if $f_k(n,H)=(k-1)\cdot ex(n,H)$ holds for sufficiently large $n$.
For non-bipartite graphs, the situation is more complicate.
We speculate that the following 3-edge-coloring $c$ of $K_n$ (which also is the extremal configuration as in \cite{CKPSTY})
achieves the maximum of $f_3(n,K_3)$:
let $V(K_n)=V_1\cup V_2\cup...\cup V_5$, where $||V_i|-|V_j||\le 1$, and
color all edges in $(V_i,V_{i+1})$ by red, all edges in $(V_i,V_{i+2})$ by blue and all edges in each $V_i$ by green.
For more discussion and other related problems, we direct readers to \cite{KS2}.
|
\section{Introduction}
It is shown in \cite{Me} that the additive group of a model of true arithmetic cannot have $\mathbb{Z}$ as a direct summand. On the other hand,
various models of arithmetic with quantifier-free induction (open induction, IOpen) and of IOpen with the condition of normality are known whose
additive group does have $\mathbb{Z}$ as a direct summand. We ask how strong an arithmetic theory needs to be to rule out $\mathbb{Z}$
as a direct summand of the additive group of a model. In this note, we show that $IE_{2}$, i.e. arithmetic with induction restricted
to formulas with one bounded existential quantifier followed by a bounded universal quantifier and an open formula, suffices.
We start by noting that IOpen does not suffice to rule out $\mathbb{Z}$ as a direct summand of the additive group of a nonstandard model:
\begin{thm}
There are nonstandard $M$ with $M\models IOpen$ and $H\subset M$ such that $(M,+)=H\oplus\mathbb{Z}$
\end{thm}
\begin{proof}
The integer parts of real closed fields constructed by Morgues-Ressayre (\cite{MR}) obviously have $\mathbb{Z}$ as a direct summand.
\end{proof}
We can even demand that these models are normal:\\
\begin{thm}
There are nonstandard $M$ with $M\models NOI$ and $H\subset M$ such that $(M,+)=H\oplus\mathbb{Z}$.
\end{thm}
\begin{proof}
Applying Proposition $1$ of \cite{GA} to $\mathbb{R}((\mathbb{Q}))$ gives an example.
\end{proof}
\section{Main Result}
We now show that $IE_{2}$ suffices to rule out $\mathbb{Z}$ as a direct summand of the additive group of a nonstandard model.
\begin{defini}
$E_2$ is the class of formulas in the language $\mathcal{L}$ of arithmetic of the form $\exists{x<t_1}\forall{y<t_2}\phi(x,y,\vec{z})$, where $t_1$ is a term not containing $x$, $t_2$ is a term not containing $y$ and $\phi$ is an open formula.\\
$IE_2$ is the axiomatic system consisting of the basic axioms of arithmetic together with induction for $E_2$-formulas.
\end{defini}
\begin{thm}
Let $M\models IE_2$ be nonstandard. Then there is no $H\subset M$ such that $(M,+)=H\oplus\mathbb{Z}$.
\end{thm}
We will prove this by three intermediate results. Assume for the rest of this section that $H$ is such a group complement of $\mathbb{Z}$, we work for a contradiction.\\
\begin{lemma}
Every element $n$ of $H$ is divisible by every standard prime.
\end{lemma}
\begin{proof}
Assume wlog that $n>0$. Let $\mathbb{P}$ denote the standard primes. Suppose for a contradiction that $n\in H$ and $p\in\mathbb{P}$ are such that $p$ does not divide $n$.
Then $IE_{2}$ proves that there is $m$ such that $pm<n<p(m+1)$, so such an $m$ exists in $M$. Let $m^{\prime}\in H$ such that $d:=m^{\prime}-m\in\mathbb{Z}$.
Then, since $p$ is standard, we must have $pm^{\prime}\in H$. Hence $pm^{\prime}-n\in H$ as well.
Furthermore, we have $\mathbb{Z}\ni p(m^{\prime}-m-1)=pm^{\prime}-p(m+1)<pm^{\prime}-n<pm^{\prime}-pm=p(m^{\prime}-m)=pd\in\mathbb{Z}$, so $pm^{\prime}-n\in\mathbb{Z}$. Therefore, we get $pm^{\prime}-n\in H\cap\mathbb{Z}$.
But $H\cap\mathbb{Z}=\{0\}$, since $0$ must be an element of every subgroup and hence a group complement of $\mathbb{Z}$ cannot contain any other element of $\mathbb{Z}$. So we conclude that $pm^{\prime}-n=0$, i.e. $n=pm^{\prime}$, which
implies that $n$ is indeed divisible by $p$, a contradiction.\\
\end{proof}
\begin{corollary}
For any $m\in M$, there is $z\in\mathbb{Z}$ such that $m\equiv_{p}z$ for all standard primes $p$.
\end{corollary}
\begin{proof}
Let $m\in M$, so $m$ can be written in the form $h+z$ for some $h\in H$ and some $z\in\mathbb{Z}$. By the last lemma, $h\equiv_{p}0$ for all standard primes $p$, hence $m\equiv_{p}z$ for all standard primes $p$.
\end{proof}
\begin{lemma}
In $M$, there is an infinite irreducible $q$ such that $q\equiv3$ mod $5$.
\end{lemma}
\begin{proof}
Consider the formula $A(n):=\exists{m,k<2n}\forall{a,b<2n}(n<C\vee(m>n\wedge m=5k+3\wedge(ab=m\implies(a=1\vee b=1))))$. It is obviously $E_2$. $A(n)$ says that, unless $n<C$, there is a prime between $n$ and $2n$ which is congruent to $3$ modulo $5$.
By the well-known asymptotic variant of Dirichlet's theorem (such as the Siegel-Walfisz-Theorem, see e.g. Satz 3.3.3 on p. 114 of \cite{Br}), the number $\pi(x;3,5)$ of primes below $x$ which are congruent to $3$
modulo $5$ is $\frac{x}{4log(x)}(1+\mathcal{O}(\frac{1}{log(x)}))$. It follows that, for sufficiently large $x$,
we have $\pi(2x;3,5)-\pi(x;3,5)>0$, so there is such a prime between $x$ and $2x$. Let $C$ be large enough that this holds for $x\geq C$. Then $A(n)$ holds for all standard natural numbers $n$.\\
As $M\models IE_2$, $M$ satisfies $E_2$-overspill. Hence there is a nonstandard element $n^{\prime}$ of $M$ such that $M\models A(n^{\prime})$. As $n^{\prime}$ is infinite, $n^{\prime}>C$, so there is an irreducible $q$ between $n^{\prime}$ and $2n^{\prime}$
leaving residue $3$ modulo $5$, as desired.
\end{proof}
\textbf{Remark}: This Lemma fails in models of mere IOpen: The methods in \cite{MM} can be used to construct nonstandard models of IOpen in which there are unboundedly many primes, but all nonstandard primes leave residue $1$ modulo $5$.
Now we can prove the theorem: By the corollary, there must be some standard integer $z$ such that $q\equiv_{p}z$ for all standard primes $p$. As $q$ is irreducible and infinite, $q$ is not divisible by any standard prime. Hence $z$ is not divisible by
any standard prime. So $z\in\{-1,1\}$. But $z\equiv_{5}q\equiv_{5}3$, hence this is impossible. Contradiction.\\
An immediate consequence is that the integer parts constructed in \cite{MR} or \cite{GA} can never be models of $IE_{2}$:\\
\begin{corollary}
Let $K$ be a non-archimedean real closed field, and let $Z$ be an integer part of $K$ generated by one of the constructions described in \cite{MR} or \cite{GA}. Then $(Z^{\geq0},+,\cdot)\not\models IE_2$.
\end{corollary}
\begin{proof}
All of these $IP$'s have $\mathbb{Z}$ as a direct summand.
\end{proof}
By a well-known result of V. Pratt, primality testing is in $NP$. Therefore, there is a $\Sigma_{1}^{b}$-definition of primality, where a $\Sigma_{1}^{b}$-formula is a formula starting with one bounded existential quantifier followed
by logarithmically bounded quantifiers (see e.g. \cite{HP}). Hence, we can reformulate our $A(n)$ as a $\Sigma_{1}^{b}$-formula, which gives us the following result:\\
\begin{corollary}
If $M\models I\Sigma_{1}^{b}$ is nonstandard, then $\mathbb{Z}$ is not a direct summand of $(M,+)$.
\end{corollary}
\textbf{Question}: The obvious next question is now whether $IE_1$ is already sufficient to exclude $\mathbb{Z}$ as a direct summand of a nonstandard model.
(This would, in particular, follow if $IE_{1}=IE_{2}$, which is still wide open.) It would also follow if there
was an $E_1$-definition of primality. Thus, in particular, it is a consequence of bounded Hilbert's $10$th problem stating that every $NP$ predicate is expressible by a bounded diophantine equation.\\
\section{Generalization}
Instead of $\mathbb{Z}$, we can consider other initial segments. It turns out that our arguments above allow two immediate generalizations.\\
\begin{thm}
(i) Let $M\models I\Delta_{0}+EXP$ and let $N$ be a cut of $M$ (i.e. a proper initial segment closed under the successor function). Then there is no $H\subseteq M$ such that $(M,+)=H\oplus N$.\\
(ii) Let $M\models IE_{2}$ and let $N\models PA$ be an initial segment of $M$. Then there is no $H\subseteq M$ such that $(M,+)=H\oplus N$.
\end{thm}
\begin{proof}
(i) Assume otherwise, and define functions $\rho_{1}:M\rightarrow H$ and $\rho_{2}:M\rightarrow N$ by $x=\rho_{1}(x)+\rho_{2}(x)$ for all $x\in M$. Then $\rho_{2}$ is a ring homomorphism from $M$ to $N$.
(In particular, on sees that $N$ must be closed under addition and multiplication and hence in fact be a model of $I\Delta_{0}$.) Now we can use the strategy of \cite{Me}: $I\Delta_{0}+EXP$ proves that every positive number is the sum
of four squares. Therefore, if $m_{1}<m_{2}$ are elements of $M$, then there are $x_1,x_2,x_3,x_4$ in $M$ such that $m_2-m_1=x_{1}^{2}+x_{2}^{2}+x_{3}^{2}+x_{4}^{2}$ and thus
$\rho_{2}(m_{2})-\rho_{2}(m_{1})=\rho_{2}(m_{2}-m_{1})=\rho_{2}(x_{1})^{2}+\rho_{2}(x_{2})^{2}+\rho_{2}(x_{3})^{2}+\rho_{2}(x_{4})^{2}>0$, so $\rho_{2}(m_{2})>\rho_{2}(m_{1})$. Hence $\rho_{2}$ preserves the ordering $M$.
But, unless $H=\{0\}$ (and hence $N$ is not a proper initial segment), there are $m_{1},m_{2}\in M$ with $\rho_{2}(m_{1})=\rho_{2}(m_{2})$, a contradiction.\\
(ii) Here, we re-use our argument from above: If such $H$ existed, then every element of $H$ would be divisible by every element of $M$. Therefore, for any $m\in M$, there would be $n\in N$ such that
$m\equiv_{k}n$ for all $k\in M$. But, as $PA$ proves the Dirichlet theorem used above, it follows by $IE_{2}$-overspill that $M$ contains an irreducible element $a$ such that $a\equiv_{5}2$ and hence
$a$ is not congruent to any element of $N$ modulo all elements of $M$, a contradiction.
\end{proof}
\textbf{Remark}: In (i), $I\Delta_{0}+EXP$ can be replaced by the weaker system $I\Delta_{0}+\Omega_{1}$, which is sufficient to prove Lagrange's theorem. In (ii), $PA$ can be replaced with any fragment of arithmetic
strong enough to prove the asymptotic version of Dirichlet's theorem.\\
|
\section{Introduction}
\textbf{If a photon interacts with a member of an entangled photon pair via a so-called Bell-state measurement (BSM), its state is teleported over principally arbitrary distances onto the second member of the pair \cite{Bennett93}. Starting in 1997 \cite{Boschi98, Zeilinger_telep, furusawa}, this puzzling prediction of quantum mechanics has been demonstrated many times \cite{Pirandola15}; however, with one very recent exception \cite{Hensen15}, only the photon that received the teleported state, if any, travelled far while the photons partaking in the BSM were always measured closely to where they were created. Here, using the Calgary fibre network, we report quantum teleportation from a telecommunication-wavelength photon, interacting with another telecommunication photon after both have travelled over several kilometres in bee-line, onto a photon at 795~nm wavelength. This improves the distance over which teleportation takes place from 818~m to 6.2~km. Our demonstration establishes an important requirement for quantum repeater-based communications \cite{Sangouard_review} and constitutes a milestone on the path to a global quantum Internet \cite{kimble}.}
While the possibility to teleport quantum states, including the teleportation of entangled states, has been verified many times using different physical systems (see Ref. [\citenum{Pirandola15}] for a recent review), the maximum distance over which teleportation is possible --- which we define to be the spatial separation between the BSM and the photon, at the time of this measurement, that receives the teleported state --- has so far received virtually no experimental attention. To date, only two experiments have been conducted in a setting that resulted in a teleportation distance that exceeds the laboratory scale \cite{Hensen15,landry}, even if in a few demonstrations the bee-line distance travelled by the photon that receives the teleported state has been much longer \cite{Ma12,Yin12}.
The reason to stress the importance of distances is linked to the ability of exploiting teleportation in various quantum information applications. One important example is the task of extending quantum communication distances using quantum repeaters \cite{Sangouard_review}, most of which rely on the creation of light-matter entanglement, e.g. by creating an entangled two-photon state out of which one photon is absorbed by a quantum memory for light \cite{Lvovsky}, and entanglement swapping \cite{Zuckowski}. The latter shares the Bell state measurement (BSM) with standard teleportation; however, the photon carrying the state to be teleported is itself a member of an entangled pair. Entanglement swapping is therefore sometimes referred-to as teleportation of entanglement. To be useful in such a repeater, two entangled photon pairs must be created far apart, and the BSM, which heralds the existence of the two partaking photons and hence of the remaining members of the two pairs, should, for optimal performance, take place approximately halfway in-between these two locations.
\begin{figure*}
\centering
\includegraphics[width=2 \columnwidth]{map_v1.eps}
\caption{\textbf{Aerial view of Calgary.} Alice is located in Manchester, Bob at the University of Calgary (UofC), and Charlie in a building next to City Hall in Calgary downtown. The teleportation distance --- in our case the distance between Charlie and Bob --- is 6.2 km. All fibres belong to the Calgary telecommunication network but, during the experiment, they only carry signals created by Alice, Bob or Charlie and were otherwise ``dark".
}
\label{fig:map}
\end{figure*}
Yet, due to the difficulty to guarantee indistinguishability of the two interacting photons after their transmission through long and noisy quantum channels \cite{Rubenok13}, entanglement swapping or standard teleportation in the important midpoint configuration has only been reported very recently outside the laboratory \cite{Hensen15}. This work exploited the heralding nature of the BSM for the first loophole-free violation of a Bell inequality --- a landmark result that exemplifies the importance of this configuration. However, the two photons featured a wavelength of about 637~nm, which, due to high loss during transmission through optical fibre, makes it impossible to extend the transmission distance to tens, let alone hundreds, of kilometers. In all other demonstrations, either the travel distances of the two photons were small, or they were artificially increased using fibre on spool \cite{landry, Riedmatten04,felix14}, effectively increasing travel time and transmission loss --- and hence decreasing communication rates --- rather than real separation. Here we report the first demonstration of quantum teleportation over several kilometers in the mid-point configuration and with photons at telecommunication wavelength.
An areal map of Calgary, identifying the locations of Alice, Bob and Charlie, is shown in Fig.~\ref{fig:map}, and Fig.~\ref{fig:setup} depicts the schematics of our experimental setup. Alice, located in Manchester (a Calgary neighbourhood), prepares phase-randomized attenuated laser pulses at 1532 nm wavelength with different mean photon numbers $\mu_\mathrm{A}\ll 1$ in various time-bin qubit states \mbox{$\ket{\psi}_A=\alpha\ket{e}+\beta e^{i\phi}\ket{\ell}$}, where $\ket{e}$ and $\ket{\ell}$ denote early and late temporal modes, respectively, $\phi$ is a phase-factor, and $\alpha$ and $\beta$ are real numbers that satisfy $\alpha^2+\beta^2=1$. Using 6.2 km of deployed fibre, she sends her qubits to Charlie, who is located 2.0 km away in a building next to Calgary City Hall.
Bob, located at the University of Calgary (UofC) 6.2 km from Charlie, creates pairs of photons --- one at 1532 nm and one at 795 nm wavelength --- in the maximally time-bin entangled state \mbox{$\ket{\phi^+}=2^{-1/2}(\ket{e,e}+\ket{\ell,\ell})$}. He sends the telecommunication-wavelength photons through 11.1 km of deployed fibre to Charlie, where they are projected jointly with the photons from Alice onto the maximally entangled state \mbox{$\ket{\psi^-}=2^{-1/2}(\ket{e,\ell}-\ket{\ell,e})$}. This leads to the 795 nm wavelength photon at Bob's acquiring the state $\ket{\psi}_B=\sigma_{y}\ket{\psi}_A$, where $\sigma_{y}$ is the Pauli operator describing a bit-flip combined with a phase-flip. In other words, Charlie's measurement results in the teleportation of Alice's photon's state, modulo a unitary transformation, over 6.2 km distance onto Bob's 795 nm wavelength photon.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{exp-setup_v4.eps}
\caption{\textbf{Schematics of the experimental setup.}
\textbf{a,} Alice's setup: An intensity modulator (IM) tailors 20 ps-long pulses of light at an 80~MHz rate out of 10~ns-long, phase randomized laser pulses at 1532~nm wavelength. Subsequently, a widely unbalanced fibre interferometer with Faraday mirrors (FM), active phase control (see the Methods sections) and path-length difference equivalent to 1.4 ns~travel time difference creates pulses in two temporal modes or bins. Following their spectral narrowing by means of a 6~GHz wide fibre Bragg grating (FBG) and attenuation to the single-photon level the time-bin qubits are sent to Charlie via a deployed fibre --- referred to as a quantum channel (QC) --- featuring 6~dB loss.
\textbf{b}, Bob's setup: Laser pulses at 1047~nm wavelength and 6~ps duration from a mode-locked laser are frequency doubled (SHG) in a periodically~poled lithium-niobate (PPLN) crystal and passed through an actively phase-controlled Mach-Zehnder interferometer (MZI) that introduces the same 1.4~ns delay as between Alice's time-bin qubits. Spontaneous parametric down-conversion (SPDC) in another PPLN crystal and pump rejection using an interference filter (not shown) results in the creation of time-bin entangled photon-pairs \cite{Brendel} at 795~and 1532~nm wavelength with mean probability $\mu_\mathrm{SPDC}$ up to 0.06. The 795~nm and 1532~nm (telecommunication-wavelength) photons are separated using a dichroic mirror (DM), and subsequently filtered to 6~GHz by a Fabry-Perot (FP) cavity and an FBG, respectively. The telecom photons are sent through deployed fibre featuring 5.7~dB loss to Charlie, and the state of the 795~nm wavelength photons is analyzed using another interferometer --- again introducing a phase-controlled travel-time difference of 1.4~ns --- and two single photon detectors based on Silicon avalanche photodiodes (Si-APD) with 65\% detection efficiency.
\textbf{c,} Charlie's setup: A beamsplitter (BS) and two WSi superconducting nanowire single photon detectors \cite{SNSPD} (SNSPD), cooled to 750~mK in a closed-cycle cryostat and with 70\% system detection efficiency, allow the projection of bi-photon states --- one from Alice and one from Bob --- onto the $\ket{\psi^-}$ Bell state. To ensure indistinguishability of the two photons at the BSM, we actively stabilize the photon arrival times and polarization, the latter involving a polarization tracker and polarizing beamsplitters (PBS), as explained in the Methods. Various synchronization tasks are performed through deployed fibres, referred to as classical channels CC, and aided by dense-wavelength division multiplexers (DWDM), photo-diodes (PD), arbitrary waveform generators (AWG), and field-programmable gate-arrays (FPGA), with details in the Methods.
}
\label{fig:setup}
\end{figure*}
To confirm successful quantum teleportation, Bob then performs a variety of projective measurements on this photon, whose outcomes, conditioned on a successful BSM at Charlie's, are analyzed using different approaches (see the Methods section for more information on how data is taken). We point out that the 795 nm photons are measured prior to the BSM, thus realizing a scenario where teleportation is achieved \textit{a posteriori} \cite{ma2012a,megedish2013a}.
The main difficulty in long-distance quantum teleportation is to ensure the required indistinguishability between the two photons subjected to the BSM at Charlie's despite them being created by independent sources and having travelled over several kilometres of deployed fibre. As we show in Fig.~\ref{fig:indistinguishability}, varying environmental conditions during the measurements significantly impact the polarization and arrival times of the photons --- of particular concern being variations of path-lengths differences. Without the active feedback (see the Methods) that our setup performs in an automized way, quantum teleportation would be impossible.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{pol+hom_v4.eps}
\caption{\textbf{Indistinguishability of photons at Charlie's.} \textbf{a}, Fluctuations of the count rate of a single SNSPD at the output of Charlie's BS with and without polarization feedback \textbf{b}, Inset: rate of coincidences between counts from SNSPDs as a function or arrival time difference, displaying a Hong-Ou-Mandel (HOM) dip \cite{HOM} when photon-arrival times at the BS are equal. Orange filled circles: The change in the generation time of Alice's qubits that is applied to ensure they arrive at Charlie's BSM at the same time as Bob's. Green empty squares: Coincidence counts per 10 s with timing feedback engaged, showing locking to the minimum of the HOM dip (see Methods). All error bars (one standard deviation) are calculated assuming Poissonian detection statistics.}
\label{fig:indistinguishability}
\end{figure}
To verify successful teleportation, first, Alice creates photons in an equal superposition of $\ket{e}$ and $\ket{\ell}$ with a fixed phase, and Bob makes projection measurements onto states described by such superpositions with various phases. Conditioned on a successful BSM at Charlie's, we find sinusoidally varying triple-coincidence count rates with a visibility of ($38 \pm 4$)\% and an average of 17.0 counts per minute. This result alone already represents a strong indication of quantum teleportation: assuming that the teleported state is a statistical mixture of a pure state and white noise, the visibility consistent with the best classical strategy and assuming Alice creates single photons is 33\% \cite{Massar95}. However, here we use this result merely to establish absolute phase references for Alice's and Bob's interferometers.
Being able to control the absolute phase values, we can now create photons in, and project them onto, well defined states, e.g. $\ket{e}$, $\ket{\ell}$, $\ket{\pm}\equiv 2^{-1/2}(\ket{e}\pm\ket{\ell})$, and $\ket{\pm i}\equiv 2^{-1/2}(\ket{e}\pm i\ket{\ell})$. This allows us to reconstruct the density matrices $\rho_\mathrm{out}$ of various quantum states after teleportation, and, in turn, calculate the fidelities $F={}_\mathrm{B}\bra{\psi}\rho_\mathrm{out}\ket{\psi}_\mathrm{B}$ with the expected states $\ket{\psi}_\mathrm{B}$. The results, depicted in Figs.~\ref{fig:QST} and ~\ref{fig:fidelity}, show that the fidelity for all four prepared states exceeds the maximum classical value of 66\% \cite{Massar95}. In particular, the average fidelity $\langle F\rangle= \left[F_e+F_l+2(F_++F_{+i})\right]/6$ = (78$\pm$1)\% violates this threshold by 12 standard deviations.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{QST-matrix_v1.eps}
\caption{\textbf{Density matrices of four states after teleportation.} Shown are the real and imaginary parts of the reconstructed density matrices for four different input states created at Alice's. The mean photon number per qubit is $\mu_\mathrm{A}=0.014$, and the mean photon pair number is $\mu_\mathrm{SPDC}=0.045$. The state labels denote the states expected after teleportation.}
\label{fig:QST}
\end{figure}
One may conclude that this result shows the quantum nature of the disembodied state transfer between Charlie and Bob. However, strictly speaking, the 66\% bound only applies to Alice's state being encoded into a single photon, while our demonstration, as others before, relied on attenuated laser pulses. To extract the appropriate experimental value, we therefore take advantage of the so-called decoy-state method, which was developed for quantum key distribution (QKD) to assess an upper bound on the error rate introduced by an eavesdropper on single photons emitted by Alice \cite{decoy05, Wang}. Here, we rather use it to characterize how a quantum channel --- in our case the concatenation of the direct transmission from Alice to Charlie and the teleportation from Charlie to Bob --- impacts on the fidelity of quantum states encoded into individual photons \cite{Sinclair15}. Towards this end, we vary the mean number of photons per qubit emitted at Alice's between three optimized values, $\mu_\mathrm{A}\in \{0, 0.014, 0.028\}$, and calculate error rates and transmission probabilities for each value independently. The results, also depicted in Fig.~\ref{fig:fidelity}, show again that the fidelities for all tested states exceed the maximum value of 2/3 achievable in classical teleportation. We note the good agreement between the measured values and those predicted by our model developed (inspired by \cite{felix14}) that takes into account various, independently characterized system imperfections (no fit). This allows us to identify that deviations of the measured fidelities from unity --- i.e. from ideal teleportation --- are mostly due to remaining distinguishability of the two photons subjected to the BSM at Charlie's, followed by multi-pair emissions by the pair-source. Finally, by averaging the single-photon fidelities over all input states, weighted as above, we find $\langle F^{(1)}\rangle \geq (80\pm 2)$\% --- as before significantly violating the threshold between classical and quantum teleportation.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fidelity_v3.eps}
\caption{Individual and average fidelities of four teleported states with expected (ideal) states, measured using quantum state tomography (QST) and the decoy-state method (DSM). For the DSM we set $\mu_\mathrm{SPDC}=0.06$. Error bars (one standard deviation) are calculated assuming Poissonian detection statistics and using Monte-Carlo simulation.}
\label{fig:fidelity}
\end{figure}
Our measurements establish the possibility for quantum teleportation over many kilometres in the important mid-point configuration --- as is required for extending the distance of quantum communications using quantum repeaters. We emphasize that both photons travelling to Charlie are at telecommunication wavelength, making it possible to extend the Alice-Bob distance from its current value of 8 kilometres by at least one order of magnitude. This corresponds to the distance of an elementary link, which includes teleportation of entanglement, at which communication links based on spectrally multiplexed quantum repeaters start to outperform direct qubit transmission \cite{Sinclair15,krovi15}.
We also note that the 795 nm photon, both in terms of central wavelength as well as spectral width, is compatible with quantum memory for light --- a key element of a quantum repeater --- in cryogenically-cooled thulium-doped crystals, in particular Tm:YGG, whose spectral acceptance exceeds the bandwidth needed for most practical applications \cite{Thiel14}, and using which we have recently demonstrated storage times of up to 100 $\mu$sec \cite{Sinclair16}. Finally, we note that quantum teleportation involves the interesting aspect of Alice transferring her quantum state in a disembodied fashion to Bob without him ever receiving any physical particle. In other words, Bob is only sending photons (all of them members of an entangled pair) and thus better able to protect his system from any outside interference, e.g. from an adversary. This points to similar considerations of security as measurement-device-independent QKD \cite{Lo}, albeit in a more flexible quantum network setting that could allow, e.g., distributed quantum computing \cite{kimble}. These key features make our demonstration an important step towards long-distance quantum communication, and ultimately a global quantum Internet.
We note that, a related experimental demonstration has been reported in a concurrent manuscript \cite{Pan16}.
\hspace{1cm}
\noindent
\textbf{Acknowledgements}\\
The authors thank Tyler Andruschak and Harpreet Dhillon from the City of Calgary for providing access to the fibre network and for help during the experiment, Vladimir Kiselyov for technical support, and Pascal Lefebvre for help with aligning the entangled photon pair source. This work was funded through Alberta Innovates Technology Futures (AITF), the National Science and Engineering Research Council of Canada (NSERC), and the Defense Advanced Research Projects Agency (DARPA) Quiness program (contract no. W31P4Q-13-l-0004). WT furthermore acknowledges funding as a Senior Fellow of the Canadian Institute for Advanced Research (CIFAR), and VBV and SWN acknowledge partial funding for detector development from the Defense Advanced Research Projects Agency (DARPA) Information in a Photon (InPho) program. Part of the detector research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.
\\
\noindent
\textbf{These authors contributed equally to this work}\\
Raju Valivarthi, Marcel.li {Grimau Puigibert}, Qiang Zhou, and Gabriel H. Aguilar.\\
\noindent
\textbf{Author contributions}\\
The SNSPDs were fabricated and tested by VBV, FM, MDS, and SWN. The experiment was conceived and guided by WT. The setup was developed, measurements were performed and the data analyzed by RV, MGP, QZ, GHA, and DO. The manuscript was written by WT, RV, MGP, QZ, GHA, and DO.\\
\noindent
\textbf{Competing financial interests}\\
The authors declare no competing financial interests.
\newpage
\section*{Methods}
\subsection{Synchronization}
For the following discussion, please refer to experimental setup outlined in Fig.~\ref{fig:setup}. Charlie is connected via pairs of optical fibres both to Alice and to Bob. In each pair, one fibre --- referred to as the quantum channel (QC) --- carries single photons, while the other --- referred to as the classical channel (CC) --- distributes various strong optical signals whose purpose will be described in the following. In addition, Alice and Bob are directly connected via an optical fibre that transmits narrow-line-width laser light at 1550~nm in order to lock all interferometers. This is crucial for maintaining a common phase reference for the qubit states generated at Alice's and Bob's, and analyzed at Bob's. In each interferometer, the power of the locking laser in one output arm (measured on a classical detector) is used to derive a feedback signal to a piezo-element that regulates the path-length difference of the interferometer to maintain a fixed phase. Additionally, all interferometers are kept in temperature controlled boxes.
The master clock for all devices is derived from detection of the mode-locked laser pulses (80~MHz) and converted back into an optical signal for distribution through the CC via Charlie to Alice.
\subsection{Stabilization to ensure photon indistinguishability}
For a successful BSM, the photons arriving at Charlie's from Alice and Bob need to be indistinguishable despite being generated by independent and different sources, and having travelled through several kilometres of deployed fibre. To ensure that the photons have the same spectral profile, they are sent through separate, temperature-stabilized fibre Bragg gratings (FBG) that narrow their spectra to 6~GHz. The spectral overlap of the FBGs at Alice and Bob needs to be set only once. However, due to temperature-dependent properties of fibre such as birefringence and length, the polarization and arrival time of the photons change with external environmental conditions, making it difficult to implement the BSM in a real-world environment. Towards this end, we apply feedback mechanisms to compensate for drifts in polarization and arrival time.
\subsubsection{Timing}
The short duration of our photons ($\sim70$~ps) prevents us from using the SNSPDs (featuring a time jitter of $\sim150$ ps) to directly determine their arrival times with the required precision to adjust the difference to zero. Instead, we compensate for arrival-time drifts with the novel approach of observing the degree of quantum-interference (Hong-Ou-Mandel or HOM effect \cite{HOM}) of the photons. The signals from the two SNSPDs (which are used to perform the BSM) are also sent to a HOM analyzing unit (see Fig.~\ref{fig:setup}) that monitors the number of coincidences between detections corresponding to either both photons arriving in an early time bin mode, or both in a late bin. As shown in the inset of Fig.~\ref{fig:indistinguishability}b, the HOM interference causes photon bunching and thus the coincidences to be reduced when the photons arrive at the beam splitter at the same time. Hence, to counteract the drift in travel time of the photons, we vary the qubit generation time at Alice with a precision of about $\sim$4~ps to keep the coincidence count rate at the minimum value of around 750 per 10 sec., as shown in Fig.~\ref{fig:indistinguishability}b. In practice Alice's time-shift is triggered at Charlie's by shifting the phase of the master clock signal that he forwards to Alice. Fig.~\ref{fig:indistinguishability}b, shows that, during a typical 1.5~hour measurement, we apply a time shift of $\sim$200 ps to compensate drifts in timing. Since the shift is larger than the duration of the photons, the teleportation protocol would fail without the active stabilization.
\subsubsection{Polarization}
Because photons from Alice and Bob pass through polarizing beam-splitters (PBS) at Charlie's, their polarization indistinguishability is naturally satisfied. However, correct photon polarizations must be set and maintained to maximize the transmission through the PBSs, or else the channel loss will vary over time. In our system, the QC between Bob and Charlie experiences only a slow drift, which allows for manual compensation using a polarization controller --- located at Bob's --- once a day. However, an automated polarization feedback system is required for the channel between Alice and Charlie, which drifts significantly on the time-scale of the experiment. To that end, we monitor the count rate of an additional SNSPD, located in the reflection port of the PBS at Charlie's, with a field-programmable gate-arrays (FPGA) so as to generate a feedback signal that minimizes the rate by adjusting the polarization by means of a polarization tracker (also located at Charlie's). As seen in Fig.~\ref{fig:indistinguishability}, the intensity fluctuations in 1.5 hours (a typical time scale to acquire results for one qubit setting) are limited to 5$\%$ with feedback, and to about 15$\%$ without feedback.
\subsection{Data collection}
Using Alice's qubits and the 1532 nm-members of the entangled pairs, Charlie performs $\ket{\psi^{-}}$ Bell-state projections. Such a projection occurs when one SNSPD detects a photon arriving in the early time-bin and the other SNSPD records a photon in the late time-bin. Successful Bell-state projection measurements are communicated via the CC and using classical laser pulses to Bob's, who converts them back to electrical signals. Each signal is then used to form a triple coincidence with the detection signal of the 795~nm wavelength photon exiting Bob's qubit analyzer. Towards this end, the latter is delayed using a variable electronic delay-line (VEDL) implemented on an FPGA by the time it takes the 1532~nm entangled photon to travel from Bob to Charlie plus the travel time of the BSM signal back to Bob.
|
\section{Introduction}
Let $X(t),t\ge 0$ be a centered stationary Gaussian processes with continuous trajectories,
unit variance and correlation function \KD{$r(\cdot)$} satisfying Pickands's condition
\BQN \label{PickandsC}
1-r(t) \sim a\abs{t}^{\alpha} , \quad t \downarrow 0, \quad a>0, \quad \text{ and } r(t)< 1, \forall t\not=0,
\EQN
with $\alpha \in (0,2]$; in our notation $\sim$ means asymptotic equivalence when the argument tends to 0 or $\IF$. \\
In the seminal contribution {\cite{PicandsA}} it was shown that
under \eqref{PickandsC}, for any $T$ positive
\begin{eqnarray}\label{eq1.2}
\pk{\sup_{t\in [0,T]} X(t)>u}\sim T\HH_{\alpha}a^{1/\alpha}u^{2/\alpha}\pk{X(0)> u}, \quad \ u\rightarrow\infty,
\end{eqnarray}
with the classical Pickands constant $\HH_{\alpha}$ defined by
$$\HH_{\alpha}=\lim_{T\rightarrow\infty} T^{-1} \EE{ \sup_{t\in[0,T]} e^{\sqrt{2}B_{\alpha}(t)-t^{\alpha}}},$$
where ${B_{\alpha}(t),t\ge 0}$ is a standard fractional Brownian motion with Hurst index $\alpha/2\in (0,1]$, see
\cite{PicandsA,Pit72,Pit96,DE2002,DI2005,DE2014,DiekerY,DEJ14,Pit20, Tabis, DM, SBK} for various properties of $\HH_\alpha$.\\
\KD{The above finding was extended in various directions, including
$\alpha(t)$-locally-stationary Gaussian processes \eMM{(see \cite{MR2462286})},
and general non-stationary Gaussian processes and random fields, \eMM{see e.g., \cite{Pit20}.}
A particularly important place in this theory is taken by the result of
Piterbarg \eMM{and Prisja{\v{z}}njuk} \cite{MR0494458}, where
the exact tail asymptotics of $\sup_{t\in [0,T]} X(t)$
is derived in the case that
the variance function $\sigma^2$ of a centered Gaussian process $X$ has a unique point of maximum in $[0,T]$, \eMM{say
$t_0$}.
More precisely, \eMM{for the correlation function it is assumed therein that} for some $\alpha \in (0,2]$
\BQN \label{countP}
1- \eM{r(s,t)} \sim a\abs{t-s}^{\alpha}, \quad s, t \downarrow 0, \quad a>0,
\EQN
whereas the behaviour of the variance function around the unique maximizer $t_0$ of
$\sigma^2(t)$ over $[0,T]$ \eMM{such that} $\sigma(t_0)=1$, is supposed to satisfy
\BQN\label{eqA1}
1- \sigma(t)&\sim & b\abs{t}^{\beta}, \quad t\downarrow 0, \quad b>0, \quad \beta \in (0,\IF).
\EQN
}
{\eMM{Assume} further that the following H\"older continuity condition
\BQN\label{HOLDER}
\EE{(X(t)-X(t))^2} \le C \abs{t-s}^\nu
\EQN
is \eMM{valid} for all $s,t \in [0, \theta]$ with some $\theta\in (0,T]$ and $\nu\in (0,2]$},
\Kd{by \cite{MR0494458}, for $\alpha < \beta$}
\BQN\label{PLE:a}
\pk{ \sup_{t\in [0,T]} X(t)>u} \sim \HH_{{{\alpha}}}\frac{ a^{1/\alpha}}{b^{1/\beta}}\Gamma(1/{{\beta}}+1) u^{2/\alpha - 2/\beta}
\pk{X(0)> u},
\EQN
and \Kd{for} $\alpha = \beta$
\BQN\label{PLE:b}
\pk{ \sup_{t\in [0,T]} X(t)>u} \sim \PP_{\alpha}^{b/a}\pk{X(0)> u},
\EQN
where $\PP_{\alpha}^d,d>0$ is the Piterbarg constant defined by
$$ \PP_{\alpha }^d = \limit{S} \EE{ \sup_{t \in [0,S]} e^{\sqrt{2} B_\alpha(t)- (1+ d) t^\alpha}} \in (0,\IF).$$
\eMM{When $\alpha > \beta$, then} \eqref{PLE:b} holds with 1 instead of $\PP_{\alpha}^{b/a}$;
see \Kd{also} Theorem 2.1 in \cite{DEJ14} \eMM{for the case $T=\IF$}.\\
We note in passing that in fact {the H\"older continuity} (\ref{HOLDER})
is not needed to derive the \Kd{asymptotics} of (\ref{eq1.2}), which will be shown later in our main theorems;
{necessary and sufficient conditions that guarantee the
global H\"older continuity of $X$ are presented in the deep contribution \cite{vitisari}.}\\
The original Pickands assumption \eqref{PickandsC}, and its counterpart \eqref{countP}
can be relaxed to $1- r$ being regularly varying at 0 with index $\alpha \in \eMM{(0},2]$, see \cite{QuallsW, Berman92}. Specifically,
in the case of \HEH{a} non-stationary $X$ we shall assume
for some \eMM{non-negative} $\LL\in \mathcal{R}_{\alpha/2}, \alpha \in (0,2]$
\BQN\label{eR}
1- r(s,t) &\sim& \LL^2(\abs{t-s}), \quad s,t\downarrow 0.
\EQN
Here $f\in \mathcal{R}_\gamma$ means that the function $f$ is regularly varying at 0 with index $\gamma$, see \cite{Res, EKM97, Soulier} for details.\\
\Kd{The first goal of} this contribution is \Kd{an extension of} Piterbarg's \eMM{results} to a \KD{more general setup, that is to suppose that}
\BQN \label{eqsvv}
1- \sigma(t)\sim \vv^2(t), \quad t\downarrow 0,
\EQN
where $\vv\geq 0$ and $\vv \in \mathcal{R}_{\beta/2}, \beta>0$.
\Kd{In \netheo{mainT} we show}
that the \eMM{asymptotic} tail behaviour of
$\sup_{t\in [0,T]} X(t)$ can be determined under the assumption that $1- \sigma$ can be compared with $1- \rho$, \eMM{namely if
further}
\BQN\label{gamma}
\lim_{t\downarrow 0} \frac{\vv^2(t)}{\LL^2(t)}= \gamma \in [0,\IF].
\EQN
\KD{Note that}, in Piterbarg's result mentioned above the limit $\gamma$ \eMM{is assumed to exist}.
\Kd{
Then we analyze tail distribution asymptotics of supremum of centered Gaussian random field $X(s,t),s\in[-T_1,T_1], t\in[\-T_2,T_2]$
with unique point that maximizes its variance function, say $(0,0)$.
Although extremes of Gaussian random fields with regularly varying correlation function are discussed in \cite{QuallsW},
see also \cite{ChengA,ChengC, YiminA,YiminB,StamatPopLet, Soja,MR3413855,Polnik} for new developments on extremes of Gaussian random fields,
most of the results in the existing literature are focused on the analysis of fields with {\it locally additive} dependence structure,
that is if
$$\Var(X(0,0))-\Var(X(s,t))\sim A_1|s|^{\beta_1} +A_1|t|^{\beta_2}$$
and
$$
1-{\rm Corr}(X(s,t),X(s_1,t_1))\sim B_1|s-s_1|^{\alpha_1}+B_2|t-t_1|^{\alpha_2}$$
as $s,s_1\to 0,$ $t,t_1\to 0.$
It appears that the investigation of fields that do not satisfy this properties is considerably more delicate
and leads to qualitatively new results.
In Section 3 we derive several novel results \eM{concerned with the exact tail asymptotics of the
maximum of centered} Gaussian random fields when both variance and correlation functions
are regularly varying and \HEH{do not} possess a locally additive strucuture.}\\
Brief outline of the rest of the paper: Our main result for extremes of Gaussian processes is displayed in the Section 2, whereas Section 3 covers Gaussian random fields. The proofs of the theorems are presented in Section 4 and some technical results and their proofs are relegated to Appendix A and B.
\section{Gaussian Processes}
Before continuing with our investigation, we mention first that there are indeed important cases of
Gaussian processes that satisfy our general setup in Section 1. \eMM{Indeed}, as remarked in \cite{HP99} and \cite{HP2004}, for any function $\rho^2 \in \eMM{\mathcal{R}_\alpha}, \alpha\in (0,2]$ there exists a centered stationary Gaussian process $Y$ with continuous trajectories, unit variance and correlation function $r$ satisfying \eqref{eR}.
Clearly, for any continuous function $\sigma(t),t\ge 0$ the process
$X(t)=\sigma(t) Y(t),t\ge 0$ has continuous trajectories and variance function $\sigma^2$.
One \eMM{instance for the properties} of $\sigma$ is to assume that \eqref{eqsvv}
holds with $$\vv^2(t)= |\ln t|^c t^\beta, \quad A>0, c\in \mathbb{R}, \beta>0.$$ For such $\sigma$,
only the case $c=0$ can be dealt with using Piterbarg's result mentioned in the Introduction. It is tempting to write
$$\vv^2(t)= (|\ln t|^{c/\beta} t) ^\beta.$$
\KD{Since in Piterbarg's result condition \eqref{eqA1}
explains the asymptotic expansion in \eqref{PLE:a}}
the $u^{-2/\beta}$ term when $\alpha < \beta$, the above could imply that
\eqref{PLE:a} still holds if we replace $u^{-2/\beta}$ by $|\ln u|^{-2c/\beta^2} u^{-2/\beta} $.\\
Detailed calculations show that \eMM{this intuition does not lead to the correct result, and in fact}
the problem is much more complicated.
Indeed,
\KD{the tail asymptotics of the supremum is determined
in this case}
in terms of the (unique) asymptotic inverse of $v$, which is given by (see Example 1.24 in \cite{Soulier} or Lemma 2 in \cite{Gira})
\BQNY
\overleftarrow{v}(t)\sim \left(\frac{\beta}{2}\right)^{c/\beta}|\ln t|^{-c/\beta}t^{2/\beta}, \quad t\downarrow 0,
\EQNY
\COM{Consequently,
\BQNY
\int_0^{\overleftarrow{\vv}(u^{-1}g(u))}e^{-u^2v^2(t)}dt\sim \left(\frac{\beta}{2}\right)^{\frac{c}{\beta}}\Gamma(1/\beta+1)u^{-2/\beta}(\ln u)^{-c/\beta}.
\EQNY
}
\KD{where} $\overleftarrow{f}$ denotes the asymptotic (unique) inverse of $f\in \mathcal{R}_\gamma$.\\
Hereafter all regularly varying functions at 0 are assumed to be ultimately non-negative as $t\rw 0$.
\KD{Further $\Psi(u) \sim e^{-u^2/2} /(\sqrt{2 \pi }u)$, as $u \to \IF$, denotes the tail distribution of an $N(0,1)$ random variable, and
\eM{we set
$$\PP_{\alpha}^\infty=:1, \quad \PP_{\alpha}^\infty[0,S]=:1, \quad \alpha \in (0, 2], S>0.$$
}
}
We state next \eMM{the main result of this section}.
\BT \label{mainT}
Let $X(t),t\ge 0$ be a centered Gaussian process with continuous trajectories and
variance {function} $\sigma^2$ having unique maximum at 0 with $\sigma(0)=1$.
Suppose that $\sigma$ satisfies \eqref{eqsvv} and the correlation function $r$ of $X$ satisfies \eqref{eR}.
Assume further that
\KD{condition \eqref{gamma} is valid} for some $\gamma \in [0,\IF]$. \\
i) If $\gamma=0$, then
\BQNY
\pk{ \sup_{t\in [0,T]} X(t)>u}\sim \ehc{\Gamma(1/\beta+1)\mathcal{H}_{\alpha}
\frac{\overleftarrow{v}(1/u)}{\overleftarrow{\LL}(1/u)}}
\Psi(u).
\EQNY
ii) If $\gamma\in (0,\IF]$, then
\BQNY
\pk{ \sup_{t\in [0,T]} X(t)>u}\sim \mathcal{P}_{\alpha}^\gamma\Psi(u) .
\EQNY
\ET
\BRM
i) If the maximum point of the variance is not 0, but an inner point, say $t_0\in (0,T)$
such that $\sigma(t_0)=1$, then
the results of \netheo{mainT} remain valid with $\mathcal{H}_\alpha$
replaced by $\widehat{\mathcal{H}}_\alpha$
and $\mathcal{P}_\alpha^\gamma$ replaced by $\widehat{\mathcal{P}}_\alpha^\gamma$,
where
$$ \widehat{\mathcal{H}}_\alpha= 2 \mathcal{H}_\alpha, \quad
\widehat{ \PP}_{\alpha }^\gamma = \limit{S} \EE{ \sup_{t \in [-S,S]} e^{\sqrt{2} B_\alpha(t)- (1+ \gamma) t^\alpha}},
\quad \widehat \PP_{\alpha }^{\IF}=1.$$
ii) Since \netheo{mainT} remain valid if we substitute $v$ by an asymptotically equivalent $v^*$, we can assume that
$v^2(t)= \ell_\sigma(t) t^\beta$ with $\ell_\sigma$ a normalized slowly varying function (see e.g., \cite{BI1989, Soulier}).
Similarly, let $\rho^2(t)=\ell_\rho(t) t^\alpha$ with $\ell_\rho$ another normalized slowly varying function.
Set next
$$
\ell_{\rho,\alpha}(x)= \sqrt{\ell_\rho(x^{1/\alpha})}, \quad \ell_{\sigma,\beta}(x)= \sqrt{\ell_\sigma(x^{1/\beta})}.$$
If further $\ell_{\sigma,\beta}^{\sharp}$ and
$ \ell_{\rho,\alpha}^{\sharp}$ denote the asymptotic inverses of $\ell_{\sigma, \beta}$ and $\ell_{\rho, \alpha}$, respectively then we have
$$v(x) = \ell_{\sigma,\beta}(x^\beta)x^{\beta/2}, \quad \rho(x) = \ell_{\rho,\alpha}(x^\alpha)x^{\alpha/2}$$
and thus by Example 1.24 in \cite{Soulier} as $t\to 0$
$$ \overleftarrow{v}(t) \sim [\ell_{\sigma, \beta}^{\sharp}(t)]^{2/\beta}t^{2/\beta} , \quad \overleftarrow{\rho}(t) \sim [\ell_{\rho,\beta}^{\sharp}(t)]^{2/\alpha}t^{2/\alpha}.$$
Consequently,
$$ \frac{\overleftarrow{v}(1/u)}{\overleftarrow{\rho}(1/u)} \sim u^{2/\alpha- 2/\beta}
\frac{[\ell_{\sigma, \beta}^{\sharp}(1/u)]^{2/\beta}}
{[\ell_{\rho,\beta}^{\sharp}(1/u)]^{2/\alpha}}, \quad u\to \IF.$$
\ERM
\netheo{mainT} is useful also when dealing with additive Gaussian random field. Specifically, assume that for \Kd{$T_1,T_2>0$}
$$ X(s,t)= \eta_1(s) + \eta_2(t), \quad s\in [-T_1,T_1], t\in [-T_2,T_2],$$
with $\eta_1,\eta_2$ two independent centered Gaussian random processes with continuous trajectories.
If both $\eta_1$ and $\eta_2$ are stationary satisfying \eqref{PickandsC}, or $\eta_1$ and $\eta_2$ satisfy the conditions of \netheo{mainT}, then
$$ \pk{\sup_{t\in \Kd{[-T_i,T_i]}} \eta_i(t) > u} \sim \mathcal{L}_i(u) u^{\tau_i} e^{- u^2/2} $$
for some $\tau_i \ge -1$ with $\mathcal{L}_i(x)= C, x\ge 0$ if $\tau_i=-1$ and $\mathcal{L}_i$ is slowly varying at infinity if $\tau_i>-1$. Hence, since
$$ \sup_{s\in [-T_1,T_1], t\in [-T_2,T_2]} X(s,t)= \sup_{s\in [-T_1,T_1]} \eta_1(s) + \sup_{t\in [-T_2,T_2]} \eta_2(t),$$
then Lemma 2.3 in \cite{Farkas15} implies
\BQN\label{nuse}
\pk{ \sup_{s\in [-T_1,T_1], t\in [-T_2,T_2]} X(s,t)> u} \sim \sqrt{2 \pi} \mathcal{L}_1(u) \mathcal{L}_2(u) u^{\tau_1+ \tau_2-1} e^{-u^2/4}, \quad u\to \IF.
\EQN
In the particular case that $\eta_i$'s satisfy the conditions of \netheo{mainT} with $\rho_i, v_i, i=1,2$ instead of $\rho$ and $v$, where
\BQN \label{gammai}
\lim_{t\downarrow 0} \frac{\vv^2_i(t)}{\LL^2_i(t)}= \gamma_i \in [0,\IF], i=1,2,
\EQN
\eMM{then } \KD{ \eqref{nuse}} can be given more explicitly, see \netheo{th.T1} below. \\
\COM{
\BK \label{korrA} For the additive Gaussian random field $X$ above we have:\\
i) If $\gamma_1=\gamma_2=0$, then
\BQNY
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}\sim 4\prod_{i=1}^2 \Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\Psi(u).
\EQNY
ii) If $\gamma_1=0, \gamma_2 \in (0,\IF]$, then
\BQNY
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}\sim2\Gamma(1/\beta_1+1)\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\Psi(u).
\EQNY
iii) If $\gamma_1, \gamma_2 \in (0,\IF]$, then
\BQNY
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}\sim\widehat{\mathcal{P}}^{\gamma_1}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}\Psi(u).
\EQNY
\EK
}
As we show in the next section, general Gaussian random fields are much more complex to deal with, and the results cannot be
derived from \netheo{mainT}.
\section{Gaussian Random Fields}
Extremes of \Kd{locally additive} Gaussian random fields with regularly varying correlation function are discussed in \cite{QuallsW}.
\KD{However there} are no results in the literature if the variance function is determined in terms of regularly varying functions
\Kd{and the dependence structure is non additive}.
In order to motivate our study, we consider first the additive Gaussian random field $X(s,t)= \eta_1(s)+ \KD{\eta_2}(t),
\Kd{ s\in [-T_1,T_1], t\in [-T_2,T_2]}$ \KD{introduced in Section 2}.
\KD{Thus, using that the variance function} $\sigma^2(s,t)$ of $X(s,t)$ is simply given by
$$\sigma^2(s,t)= \sigma_1^2(s)+ \sigma_2^2(t),$$
if $\eta_1,\eta_2$ satisfy the assumptions of \netheo{mainT}, then
$\sigma(s,t)$ achieves its unique maximum at $(0,0)$.\\
In this section we shall discuss an extension of \netheo{mainT} to
\begin{eqnarray}
\pk{\sup_{(s,t) \in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}, \label{task.2}
\end{eqnarray}
as $u\to \IF$,
where $X(s,t),(s,t) \in[-T_1,T_1]\times[-T_2,T_2]$
is a centered Gaussian random field, with variance function that is maximal on a unique point \Kd{but possess dependence structure that
is more complex than the additive one discussed above.}
\Kd{In particular, we suppose that}
\BQN\label{Cor2}
1- r(s,t,s_1,t_1) \sim \LL_{1}^2(\abs{a_{11}(s-s_1)+a_{12}(t-t_1)})+ \LL_{2}^2(\abs{a_{21}(s-s_1)+a_{22}(t-t_1)})
\EQN
as $s,s_1,t,t_1 \to 0$ with $\LL_{i}\geq 0$ and $\LL_{i}\in \mathcal{R}_{\alpha_i/2}, \alpha_i \in (0,2],i=1,2$.
For the variance function $\sigma^2(s,t)=Var(X(s,t))$ we shall assume that it attains \Kd{its} maximum at the unique point $(0,0)$ with $\sigma(0,0)=1$ and further
\BQN\label{Var2}
1- \sigma(s,t) \sim \vv_{1}^2 (\abs{b_{11}(s-s_1)+b_{12}(t-t_1)})+ \vv_{2}^2(\abs{b_{21}(s-s_1)+b_{22}(t-t_1)}) , \quad s,t\downarrow0,
\EQN
where $\vv_i\geq 0$ and $\vv_i\in \mathcal{R}_{\beta_i/2},\beta_i>0, i=1,2$.\\
\Kd{We note that recent results for the case that the variance function $\sigma^2$ is maximal on a line, which is the case for instance
if $\eta_1$ is stationary with unit variance 1 and $\eta_2$ satisfies the assumptions of \netheo{mainT},
are obtained in \cite{nonhomoANN}.
}
\KD{For further analysis it is useful to introduce the following matrices
\BQNY
A=\left(\begin{array}{cc}
a_{11}& a_{12}\\
a_{21}& a_{22}\\
\end{array}
\right), \ \
B=\left(\begin{array}{cc}
b_{11}& b_{12}\\
b_{21}& b_{22}\\
\end{array}
\right).
\EQNY
}
Let us observe that the assumption of uniqueness of
the maximizer of $\sigma(\cdot, \cdot)$ implies that ${\rm rank}(B)=2$.
We shall assume that \eqref{gammai} holds
and furthermore the following limits
\BQN\label{ette}
\lim_{t\downarrow 0}\frac{\rho_2^2(t)}{\rho_1^2(t)}= \eta \in [0,\IF], \ \ \lim_{t\downarrow 0}\frac{v_2^2(t)}{v_1^2(t)}=\theta \in [0,\IF]
\EQN
exist.
\KD{
It appears that the rank of matrix $A$ plays the key role
for the asymptotics of (\ref{task.2}), as $u\to\infty$.
Thus, in what follows, we shall distinguish
between two scenarios, when
${\rm rank}(A)=2$ and ${\rm rank}(A)=1$.
We exclude from further analysis the degenerated case of ${\rm rank}(A)=0$.
}
\subsection{\underline{Scenario I: ${\rm rank}(A)=2$}}
Suppose that $A$ is invertible and observe that
$Y(s,t):=X((A^{-1}(s,t)^\top )^\top )$
has, under (\ref{Cor2}), (\ref{Var2}), correlation function such that
\BQN\label{Cor0}
1-r_{Y}(s,s_1,t,t_1)\sim \rho_1^2(|s-s_1|)+\rho^2_2(|t-t_1|), \ \ s,s_1,t,t_1\rw 0,
\EQN
and variance function $\sigma_Y^2$ satisfying
\BQN\label{Var0}
1-\sigma_Y(s,t)\sim v_1^2(|c_{11}s+c_{12}t|)+v_2^2(|c_{21}s+c_{22}t|), \ \ s,t\rw 0,
\EQN
with
$$ C=\left(\begin{array}{cc}
c_{11}& c_{12}\\
c_{21}& c_{22}\\
\end{array}
\right)
=BA^{-1}.$$
Therefore, with no loss of generality, in this section we tacitly assume that
$X$ satisfies (\ref{Cor2}) with
$$ A=\left(\begin{array}{cc}
1& 0\\
0& 1\\
\end{array}
\right)=:I.$$
Next, define an additive fractional Brownian field $W$ by
$$W(s,t)=\sqrt{2}B_{\alpha}(s)+\sqrt{2}\widetilde{B}_{\alpha}(t)- \abs{s}^\alpha- \abs{t}^\alpha, $$
where
$B_{\alpha}(t)$ and $\widetilde{B}_{\alpha}(t)$ are
independent standard fBm's with index $\alpha\in (0,2]$.
For a given matrix $D=(d_{ij})_{i,j=1,2}$, we define the generalized Piterbarg constant
$$\widehat{\mathcal{P}}_{\alpha}^{\gamma_1,\gamma_2, D}:=
\lim_{S\rw \IF}\EE{\sup_{(s,t)\in [-S,S]^2} e^{
W(s,t)-\gamma_1|d_{11}s+d_{12}t|^{\alpha}
-\gamma_2|d_{21}s+d_{22}t|^{\alpha}}},$$
where $\gamma_1, \gamma_2>0$. Note that if $det(\mathrm{D})\neq 0$, then there exists $\gamma_3>0$ such that
$$\gamma_1|d_{11}s+d_{12}t|^{\alpha}
+\gamma_2|d_{21}s+d_{22}t|^{\alpha}\geq \gamma_3(|s|^{\alpha}+|t|^{\alpha}), \ \ s,t\in \mathbb{R},$$
which implies that
$\widehat{\mathcal{P}}_{\alpha}^{\gamma_1,\gamma_2, D}\leq \left(\widehat{\mathcal{P}}_{\alpha}^{\gamma_3}\right)^2<\IF.$
Moreover,
\KD{for $D=I$} we have
$$
\widehat{\mathcal{P}}_{\alpha}^{\gamma_1,\gamma_2, I}=\widehat{ \mathcal{P}}_{\alpha}^{\gamma_1} \widehat{\mathcal{P}}_{\alpha}^{\gamma_2}.
$$
Let \eMM{for $S_1,S_2$ non-negative}
$$\mathcal{H}_\alpha^{\gamma_1,\gamma_2,b}(S_1,S_2):=
\EE{\sup_{(s+bt,t)\in [-S_1,S_1]\times[0,S_2]} e^{ W(s,t) -\gamma_1|s+bt|^{\alpha}-\gamma_2|t|^{\alpha}
}} ,$$
and
$$\widehat{\mathcal{H}}_\alpha^{\gamma_1,\gamma_2,b}(S_1,S_2):=\EE{\sup_{(s+bt,t)\in [-S_1,S_1]\times[-S_2,S_2]} e^{ W(s,t) -\gamma_1|s+bt|^{\alpha}-\gamma_2|t|^{\alpha}
}} .$$
In order to simplify the notation we set
$$\mathcal{H}_\alpha^{\gamma_1,\gamma_2,b}(S_1):=\mathcal{H}_\alpha^{\gamma_1,\gamma_2,b}(S_1,S_1), \ \ \widehat{\mathcal{H}}_\alpha^{\gamma_1,\gamma_2,b}(S_1)=\widetilde{\mathcal{H}}_\alpha^{\gamma_1,\gamma_2,b}(S_1,S_1), \ \ \mathcal{H}_\alpha^{\gamma_1,b}(S_1)=\mathcal{H}_\alpha^{\gamma_1,0,b}(S_1,S_1),$$
and
$$ \mathcal{H}_\alpha^{\gamma_1,b}:=\lim_{S\rw\IF}S^{-1}\mathcal{H}_\alpha^{\gamma_1,b}(S).$$
Now let us proceed to the analysis of (\ref{task.2}) for four special cases
whose proofs \eMM{are all different}, and
to which
one can reduce all other scenarios
(as will be advocated at the end of this section).
\\
Since below $A$ is taken to be the identity matrix, the cases discussed below are defined by the different choices of the matrix $B$.\\
$\diamond$ {\underline{Case 1.}}
We say that $X$ is {\it locally additive}, if both \eqref{Cor2} and \eqref{Var2} hold with $A=B=I$.
\COM{
(\ref{Cor2}) and
(\ref{Var2}) are satisfied with
$$ A=\left(\begin{array}{cc}
1& 0\\
0& 1\\
\end{array}
\right), \quad
B=\left(\begin{array}{cc}
1 & 0\\
0& 1\\
\end{array}
\right)
.$$
\\
\\
}
\KD{The result below holds} for any $\theta, \eta \in [0,\IF]$ defined in \eqref{ette}.
\BT\label{th.T1}
Suppose that $X$ is \KD{a} {\it locally additive} Gaussian random field. \\
i) If $\gamma_1=\gamma_2=0$, then
\BQNY \label{T1a}
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}\sim 4\prod_{i=1}^2 \left(\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\right)\Psi(u).
\EQNY
ii) If $\gamma_1=0, \gamma_2 \in (0,\IF]$, then
\BQNY
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}\sim2\Gamma(1/\beta_1+1)\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\Psi(u).
\EQNY
iii) If $\gamma_1, \gamma_2 \in (0,\IF]$, then
\BQNY
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}\sim\widehat{\mathcal{P}}^{\gamma_1}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}\Psi(u).
\EQNY
\ET
\begin{remark}\label{remN}
We note that by the use of change of coordinates
Theorem \ref{th.T1} covers all the combinations of values of $\gamma_1, \gamma_2$.
\end{remark}
$\diamond$ {\underline{Case 2.}} Here we shall assume that
(\ref{Cor2}) and
(\ref{Var2}) are satisfied with
\begin{eqnarray}
A=I, \quad B=\left(\begin{array}{cc}
1& b_{12}\\
0& 1\\
\end{array}
\right),\ \rm{with}\ b_{12}\neq 0. \label{case2}
\end{eqnarray}
\BT\label{th.T2}
\KD{Suppose that \eqref{ette} is satisfied with $\eta\in(0,\infty)$, $\theta=0$ and (\ref{case2}) holds}.
i) If $\gamma_1=\gamma_2=0$, then
\[
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
4 \prod_{i=1}^2 \left(\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\right) \Psi(u).
\]
ii) If $\gamma_2=0, \gamma_1\in (0,\IF]$, then
\BQNY
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
2\Gamma(1/\beta_1+1) \mathcal{H}_{\alpha_1}^{\gamma_1, b_{12} \eta^{-1/\alpha_1}}
\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\Psi(u) .
\EQNY
iii) If $\gamma_2 \in (0,\IF], \gamma_1=\IF$, then
\[
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
\widehat{\mathcal{P}}_{\eMM{\alpha_1}}^{\gamma_2(|b_{12}|^{\alpha_1}\eta^{-1}+1)^{-1}}
\Psi(u) .
\]
\ET
\begin{remark}
The above theorem \KD{covers} all the possible combinations of values
of $\gamma_1,\gamma_2$, since
the assumption that $\eta\in(0,\infty)$, $\theta=0$
excludes cases
$\gamma_1=0,\gamma_2 \in (0,\IF]$ and
$\gamma_1\in(0,\infty),\gamma_2 \in (0,\IF]$.
\KD{Although the same asymptotics \eMM{are imposed} in $i)$ of Theorem \ref{th.T1} and $i)$ of Theorem \ref{th.T2},
their proofs require \HEH{a} substantially different approach. Thus we did not combine
\HEH{those} cases in one result.}
\end{remark}
$\diamond$ {\underline{Case 3.}}
The assumptions on $A$ and $B$ are the same as in Case 2 above, however
\COM{ case We say that
$X\in T_3$ if
(\ref{Cor2}) and
(\ref{Var2}) are satisfied with
$ A=\left(\begin{array}{cc}
1& 0\\
0& 1\\
\end{array}
\right)$,
$
B=\left(\begin{array}{cc}
1& b_{12}\\
0& 1\\
\end{array}
\right)
$, $b_{12}\neq 0$, and
}
we shall suppose that $\eta=0$, $\theta\in(0,\infty)$. \\
Since $\theta\in (0,\IF)$, we set $\beta=\beta_1=\beta_2$. Let $\mu \in (-\IF,\IF)$ be the point at which
$|1+b_{12}t|^{\beta}+\theta |t|^{\beta}$
attains its minimum over $(-\IF,\IF)$. We have $\mu\in [-1/|b_{12}|, 1/|b_{12}|]$.
Further, set
\BQN\label{M} \Mba =\inf_{t\in (-\IF,\IF)}\left(|1+b_{12}t|^{\beta}+\theta |t|^{\beta}\right)
\EQN
and define the two-sided Piterbarg-type constant
$$\widehat{\PP}_{\beta }^{g_s} = \limit{S}\widehat{\PP}_{\beta }^{g_s}[-S,S], \ \ \text{with} \ \ \widehat{\PP}_{\beta }^{g_s}[-S,S]
=\EE{ \sup_{t \in \Kd{[-S,S]}} e^{\sqrt{2} B_\beta(t)- t^\beta-g_s(t)}}, \quad S>0, s\geq 0,$$
where
$$g_s(t)=\theta^{-1} \gamma_2\left(|s+b_{12}t|^{\beta}+\theta|t|^{\beta}-|(1+b_{12}\mu) s|^{\beta}-\theta|\mu s|^{\beta}\right), s\geq 0, t\in \mathbb{R}.$$
Further, set
$$\mathcal{I}_\beta := \int_{-\IF}^{\IF}\widehat{\mathcal{P}}_{\beta}^{g_{|s|}}
e^{-\frac{\gamma_2 \Mba }{\theta}|s|^{\beta}}ds
\in (0,\IF).$$
The finiteness of $\eMM{\mathcal{I}_\beta}$ follows from the fact that for any $\epsilon>0$, there exists a positive constant $c_\epsilon>0$ such that
$$g_s(t)+\epsilon |s|^{\beta}\geq c_{\epsilon}|t|^{\beta}, \ \ s\geq 0, t\in \mathbb{R}
$$
implying that $\widehat{\mathcal{P}}_{\beta}^{g_{s}}\leq \widehat{\mathcal{P}}_{\beta}^{c_\epsilon}e^{\epsilon s^{\beta}}<\IF$, \eMM{and thus} for $\epsilon \in (0, \theta^{-1} \gamma_2 \Mba )$
$$\mathcal{I}_\beta\leq 2\int_0^\IF
\widehat{\mathcal{P}}_{\beta}^{g_{s}}e^{-\frac{\gamma_2 \Mba }{\theta}s^{\beta}}ds\leq 2\widehat{\mathcal{P}}_{\beta}^{c_\epsilon}\int_0^\IF
e^{-(\frac{\gamma_2 \Mba }{\theta}-\epsilon)s^{\beta}}ds<\IF.$$
\BT\label{th.T3}
\KD{Suppose that \eqref{case2} holds and \eqref{ette} is satisfied}
with $\eta=0$, $\theta\in(0,\infty)$.\\
i) If $\gamma_1=\gamma_2=0$, then
\[
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
4\prod_{i=1}^2 \left(\Gamma(1/\KD{\beta}+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\right) \Psi(u) .
\]
ii) If $\gamma_1=0, \gamma_2 \in (0,\IF]$, then
\[
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u} \sim
\mathcal{H}_{\alpha_1} \left(\frac{\gamma_2}{\theta}\right)^{1/\beta}
\mathcal{I}_\beta \frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\Psi(u),
\]
iii)) If $\gamma_1=0, \gamma_2 =\IF$, then
\[
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u} \sim
2\Gamma(1/\beta+1)\left( \Mba \right)^{-1/\beta}\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\vv_1}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u),
\]
iv) If $\gamma_1\in (0,\IF], \gamma_2=\IF$, then
\[
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
\widehat{\mathcal{P}}^{\gamma_1 \Mba }_{\alpha_1}
\Psi(u).
\]
\ET
\begin{remark}
Analogously to the Case 2,
the assumption that $\eta=0$, $\theta\in(0,\infty)$
excludes case
$\gamma_1\in (0,\IF],\gamma_2\in[0,\infty)$.
\end{remark}
$\diamond$ {\underline{Case 4.}}
Here we still assume that $A=I$ but there are no restrictions on the invertible $B$.
\COM{We say that
$X\in T_4$ if
(\ref{Cor2}) and
(\ref{Var2}) are satisfied with
$ A=\left(\begin{array}{cc}
1& 0\\
0& 1\\
\end{array}
\right)$
and
$
B=\left(\begin{array}{cc}
b_{11}& b_{12}\\
b_{21}& b_{22}\\
\end{array}
\right)
$,
and $\eta,\theta\in(0,\infty)$.
\\
\\
This case leads to the following asymptotics.
}
\BT\label{th.T4}
Suppose that \eqref{Cor2} and \eqref{Var2} hold with $A=I$ and $B$ an invertible matrix,
and
\eqref{ette} is satisfied with $\eta,\theta\in(0,\infty)$. \\
i) If $\gamma_1=\gamma_2=0$, then
\BQN\label{gg0}
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
\frac{4}{\abs{ \mathrm{det}(B)}}
\prod_{i=1}^2 \left(\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\right) \Psi(u) .
\EQN
ii) If $\gamma_1,\gamma_2 \in (0,\IF)$ or $\gamma_1=\gamma_2=\IF$, then
\BQN \label{f1}
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
\widehat{\mathcal{P}}_{\KD{\alpha_1}}^{\gamma_1,\gamma_1\theta, B_{\eta,\alpha_1}}
\Psi(u) ,
\EQN
where $\widehat{\mathcal{P}}_{\alpha_1}^{\gamma_1,\theta\gamma_1, B_{\eta,\alpha_1}}=1$ if $\gamma_1=\gamma_2= \IF$ and
$B_{\eta,\KD{\alpha_1}}=\left(\begin{array}{cc}
b_{11}& b_{12}\eta^{-1/\alpha_1}\\
b_{21}& b_{22}\eta^{-1/\alpha_1}\\
\end{array}\right).$
\ET
\subsubsection{Discussion} As mentioned above, all other cases for
${\rm rank}(A)=2$ can be reduced to
the analysis of the field of one of types covered by Case 1-4. For the sake of transparency, let us first consider
$ A=I$
and $B$ such that exactly one element $b_{ij}$
equals to 0. With no loss of generality, by a change of variables, we can assume that
$$
B=\left(\begin{array}{cc}
1& b_{12}\\
0& 1\\
\end{array}
\right)
, \quad b_{12}\neq 0.$$
Then the following holds:\\
\begin{itemize}
\item[$\diamond$] \underline{$\theta=\IF$:}
The asymptotics of (\ref{task.2})
in this case is covered by Case 1 above, since by Lemma \ref{simple} we obtain
$$v_1^2(|s+b_{12}t|)+v_2^2(|t|)\sim v_1^2(|s|)+v_2^2(|t|), \quad s, t\rw 0.$$
\item[$\diamond$] \underline{$\eta=\IF$:} Let $Z(s,t)=X(s-b_{12}t,t)$, which is a {\it locally additive} Gaussian random field.
Indeed, it follows from Lemma \ref{simple} that
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1-b_{12}(t-t_1)|)+\rho_2^2(|t-t_1|)\sim \rho_1^2(|s-s_1|)+\rho_2^2(|t-t_1|),\quad s,t,s_1,t_1\rw 0, $$
and $1-\sigma_Z(s,t)\sim v_1^2(|s|)+v_2^2(|t|), s, t\rw 0.$
\item[$\diamond$] \underline{$\theta=0, \eta=0$:} Let $Z(s,t)=X(s,\frac{t-s}{b_{12}})$.
Then, again by Lemma \ref{simple}, $Z$ is a {\it locally additive} Gaussian random field
with
$$1-r_Z(s,t,s_1,t_1)\sim\rho_1^2(|s-s_1|)+|b_{12}|^{-\alpha_2}\rho_2^2(|t-t_1|),\quad s,t,s_1,t_1\rw 0, $$
and
$1-\sigma_Z(s,t)\sim |b_{12}|^{-\beta_2}v_2^2(|s|)+v_1^2(|t|), s, t\rw 0.$
\item[$\diamond$] \underline{$\theta=0$, $\eta\in (0,\infty)$:}
This is covered by Case 2 above.
\item[$\diamond$] \underline{$\theta\in(0,\infty)$, $\eta=0$:}
This is covered by Case 3 above.
\item[$\diamond$] \underline{$\theta\in(0,\infty)$, $\eta\in(0,\infty)$:}
This is covered by Case 4 above.
\\
\end{itemize}
Next, let $A=I$ and
\COM{Suppose now that
$ A=\left(\begin{array}{cc}
1& 0\\
0& 1\\
\end{array}
\right)$
and}
$b_{ij}\neq 0$ for $i,j=1,2$.
With no loss of generality we can assume that
$$B=\left(\begin{array}{cc}
1& b_{12}\\
b_{21}& 1\\
\end{array}
\right)
, \quad b_{12}b_{21}\neq 0.$$
Let us observe that $det(B)=1-b_{12}b_{21}\neq 0$, which will be used in several places below.
Then the following holds:
\begin{itemize}
\item[$\diamond$] \underline{$\theta=0$, $\eta=0$:} Let $Z(s,t)=X(s, \frac{t-s}{b_{12}})$.
Again by Lemma \ref{simple} $Z$ is a {\it locally additive} Gaussian random field
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|)+|b_{12}|^{-\alpha_2}\rho_2^2(|t-t_1|), \quad s,t,s_1,t_1\rw 0, $$
and
$1-\sigma_Z(s,t)\sim \left|\frac{detB}{b_{12}}\right|^{\beta_2}v_2^2(|s|)+v_1^2(|t|), \quad s, t\rw 0.$
\item[$\diamond$] \underline{$\theta=0$, $\eta\in (0,\IF)$:}
This is Case 2 with $v_2^2$ replaced by $|det (B)|^{\beta_2}v_2^2$.
Indeed, by Lemma \ref{simple}, we have
\BQNY v_1^2(|s+b_{12}t|)+v_2^2(|b_{21}s+t|)&=&v_1^2(|s+b_{12}t|)+v_2^2(|b_{21}(s+b_{12}t)+(1-b_{12}b_{21})t|)\\
&\sim& v_1^2(|s+b_{12}t|)+|det (B)|^{\beta_2}v_2^2(|t|), \quad s, t\rw 0.
\EQNY
\item[$\diamond$] \underline{$\theta=0$, $\eta=\IF$:} Let $Z(s,t)=X(s-b_{12}t,t)$.
Again, by Lemma \ref{simple}, $Z$ is a {\it locally additive} Gaussian random field with
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|)+\rho_2^2(|t-t_1|),\quad s,t,s_1,t_1\rw 0, $$
and
$1-\sigma_Z(s,t)\sim v_1^2(|s|)+v_2^2(|b_{21}s+(1-b_{12}b_{21})t|)\sim v_1^2(|s|)+|det (B)|^{\beta_2}v_2^2(|t|), \quad s, t\rw 0.$
\item[$\diamond$] \underline{$\theta\in (0,\IF), \eta=0$:} Let $Z(s,t)=X(s,t-b_{21}s)$.
This is Case 3 with
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|)+\rho_2^2(|t-t_1|), \quad s,t,s_1,t_1\rw 0, $$
and
$1-\sigma_Z(s,t)\sim |det (B)|^{\beta_1}v_1^2(|s+b_{12}(det (B))^{-1}t|)
+v_2^2(|t|), \quad s, t\rw 0.$
\item[$\diamond$] \underline{$\theta\in(0,\infty)$, $\eta\in(0,\infty)$:}
This is covered by Case 4.
\item[$\diamond$] \underline{$\theta\in(0,\infty)$, $\eta=\IF$:} Let $Z(s,t)=X(\frac{t-s}{b_{21}},s)$.
This is Case 3 with
$$1-r_Z(s,t,s_1,t_1)\sim \rho_2^2(|s-s_1|)+|b_{21}|^{-\alpha_1}\rho_1^2(|t-t_1|), \quad s,t,s_1,t_1\rw 0, $$
and
$1-\sigma_Z(s,t)\sim \left|\frac{det (B)}{b_{21}}\right|^{\beta_1}v_1^2(|s+(-det (B))^{-1}t|)+v_2^2(|t|), s, t\rw 0.$
\item[$\diamond$] \underline{$\theta=\IF$, $\eta=0$:} Let $Z(s,t)=X(s,t-b_{21}s)$.
This is a {\it locally additive} Gaussian random field with $v_1^2$ substituted by $|det (B)|^{\beta_1}v_1^2$.
\item[$\diamond$] \underline{$\theta=\IF$, $\eta\in (0,\IF)$:} By Lemma \ref{simple}
we have that this is Case 2 with
$$1-r_X(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|)+\rho_2^2(|t-t_1|), \quad s,t,s_1,t_1\rw 0, $$
and
\BQNY
1-\sigma_X(s,t)&=&v_2^2(|b_{21}s+t|)+v_1^2(|b_{21}^{-1}(b_{21}s+t)+(b_{12}-b_{21}^{-1})t|)\\
&\sim&v_2^2(|b_{21}s+t|)+v_1^2((b_{12}-b_{21}^{-1})t|)\\
&\sim& |b_{21}|^{\beta_2}v_2^2(|s+(b_{21})^{-1}t|)+\left|\frac{det (B)}{b_{21}}\right|^{\beta_1}v_1^2(|t|), \quad s, t\rw 0.
\EQNY
\item[$\diamond$] \underline{$\theta=\IF$, $\eta=\IF$:} Let $Z(s,t)=X(\frac{s-t}{b_{21}},t)$.
We have that $Z$ is a {\it locally additive} Gaussian random field with
$$1-r_Z(s,t,s_1,t_1)\sim |b_{21}|^{-\alpha_1}\rho_1^2(|s-s_1|)+\rho_2^2(|t-t_1|),\quad s,t,s_1,t_1\rw 0, $$
and
$1-\sigma_Z(s,t)\sim v_2^2(|s|)+ \left|\frac{det (B)}{b_{21}}\right|^{\beta_1}v_1^2(|t|), s, t\rw 0.$
\end{itemize}
\subsection{\underline{Scenario II: ${\rm rank}(A)=1$}}
Suppose that ${\rm rank}(A)=1$. \Kd{Clearly} it suffices to consider Gaussian random fields with
covariance {function} that satisfies (\ref{Cor2}) with
$ A=\left(\begin{array}{cc}
1& 0\\
0& 0\\
\end{array}
\right)$
and variance function satisfying (\ref{Var2}).
We begin with the analysis of
two special cases, to
which all other structures of
field $X$ can be reduced.
\\
\\
$\diamond$ {\underline{Case 5.}}
Here we shall assume that (\ref{Cor2}) and
(\ref{Var2}) are satisfied with
\begin{eqnarray}
A=\left(\begin{array}{cc}
1& 0\\
0& 0\\
\end{array}
\right)
\ {\rm{and}} \
B=I.
\label{case5}
\end{eqnarray}
\COM{
B=\left(\begin{array}{cc}
1& 0\\
0& 1\\
\end{array}
\right).
$
\\}
\BT\label{th.T5}
\KD{Suppose that \eqref{case5} holds.}
\\
i) If $\gamma_1=0$, then
\BQNY
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim 2\Gamma(1/\beta_1+1)\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\Psi(u) .
\EQNY
ii) If $\gamma_1 \in (0,\IF]$, then
\BQN \label{f1}
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
\widehat{\mathcal{P}}^{\gamma_1}_{\alpha_1} \Psi(u).
\EQN
\ET
$\diamond$ {\underline{Case 6.}} Here we shall assume that
(\ref{Cor2}) and
(\ref{Var2}) are satisfied with
\begin{eqnarray}
A=\left(\begin{array}{cc}
1& 0\\
0& 0\\
\end{array}
\right), \quad
B=\left(\begin{array}{cc}
1& b_{12}\\
0& 1\\
\end{array}
\right),
\ \rm{and}\ b_{12}\neq 0. \label{case6}
\end{eqnarray}
\BT\label{th.T6}
\KD{Suppose that \eqref{case6} holds and \eqref{ette} is satisfied with $\theta \in (0,\IF)$.} \\
i) If $\gamma_1=0$, then
\BQN
\pk{ \sup_{(s,t)\in[-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
2 (\Mbb )^{-1/\beta_1} \Gamma(1/\beta_1+1)\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\Psi(u) .
\EQN
ii) If $\gamma_1 \in (0,\IF]$, then
\BQN \label{f1}
\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}
\sim
\widehat{\mathcal{P}}^{\eMM{\gamma_1} \Mbb }_{\alpha_1}
\Psi(u)
\EQN
with $M_\beta$ defined in (\ref{M}).
\ET
\subsubsection{Discussion}
Having analyzed the above special cases, we are now in position to
give the asymptotics of (\ref{task.2}) for general structure of $X$. Suppose first, analogously to Scenario I, that $X$ satisfies
(\ref{Cor2}) and
(\ref{Var2}) with
$ A=\left(\begin{array}{cc}
1& 0\\
0& 0\\
\end{array}
\right)$
and
exactly one element of matrix $B$ equals 0.
With no loss of generality we can assume that
$
B=\left(\begin{array}{cc}
1& b_{12}\\
0 & 1\\
\end{array}
\right)
$, $b_{12}\neq 0$.
Then the following holds.
\begin{itemize}
\item[$\diamond$] \underline{$\theta=0$:}
Let $Z(s,t)=X(s,\frac{t-s}{b_{12}})$. Then, by Lemma \ref{simple}, this is Case 5
with
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|), s,t,s_1,t_1\rw 0, \ \ 1-\sigma_Z(s,t)\sim |b_{12}|^{-\beta_2}v_2^2(|s|)+v_1^2(|t|),
\quad s,t\rw 0.$$
\item[$\diamond$] \underline{$\theta\in(0,\infty)$:}
This is case 6.
\item[$\diamond$] \underline{$\theta=\infty$:}
The asymptotics of (\ref{task.2})
in this case is the same as the asymptotis derived in Case 5.
Indeed, by Lemma \ref{simple}, we have
$$v_1^2(|s+b_{12}t|)+v_2^2(|t|)\sim v_1^2(|s|)+v_2^2(|t|), \quad s,t\rw 0.$$
\end{itemize}
Finally we discuss the other case
where the matrix $B$ is such that $b_{ij}\neq 0$ for $i,j=1,2$.
Again with no loss of generality we can assume that
$
B=\left(\begin{array}{cc}
1& b_{12}\\
b_{21} & 1\\
\end{array}
\right)
$, $b_{12}, b_{21}\neq 0$.
Then the following holds with $det(B)=1-b_{12}b_{21}\neq 0$:
\begin{itemize}
\item[$\diamond$] \underline{$\theta=0$:}
Let $Z(s,t)=X(s,\frac{t-s}{b_{12}})$.
This is covered by Case 5.
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|), s,t,s_1,t_1\rw 0, \ \ 1-\sigma_Z(s,t)\sim \left|\frac{det (B)}{b_{12}}\right|^{\beta_2}v_2^2(|s|)+v_1^2(|t|), \quad s,t\rw 0.$$
\item[$\diamond$] \underline{$\theta\in(0,\infty)$:}
Let $Z(s,t)=X(s,t-b_{21}s)$. Then, by
Lemma \ref{simple}, $Z$ is as in Case 6 with
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|), \quad s,t,s_1,t_1\rw 0, \ \
1-\sigma_Z(s,t)\sim |det (B)|^{\beta_1}v_1^2(|s+b_{12}(det (B))^{-1}t|)
+v_2^2(|t|), \quad s,t\rw 0.$$
\item[$\diamond$] \underline{$\theta=\infty$:}
Let $Z(s,t)=X(s, t-b_{21}s)$ . This is Case 5 with
$$1-r_Z(s,t,s_1,t_1)\sim \rho_1^2(|s-s_1|), s,t,s_1,t_1\rw 0, \ \ 1-\sigma_Z(s,t)\sim |det( B)|^{\beta_1}v_1^2(|s|)+v_2^2(|t|),
\quad s,t\rw 0.$$
\end{itemize}
\section{Proofs}
\KD{
In the rest of this section by
$\mathbb{Q}, \mathbb{Q}_i>0, i=1,2,...$
we denote constants that may differ from line to line.
}
\def\gu{\ehc{\xi(u)}}
\prooftheo{mainT}
We set, for $u>1$ and $\gu:=u^{-1} \ln u$
\KD{
$$E_u=[0,\overleftarrow{\vv}(\gu )],
\quad I_k(u)=[kS\overleftarrow{\rho}(u^{-1}), (k+1)S\overleftarrow{\rho}(u^{-1})],
k\in \mathbb{N}\cup \{0\}$$
}
and, for given $\ve\in (0,1/2)$, define
$$u_{k,\epsilon}^{-}=u(1+(1-\epsilon)\inf_{t\in I_{k}(u)}\vv^2(t)), \quad u_{k,\epsilon}^{+}=u(1+(1+\epsilon)\sup_{t\in I_{k}(u)}\vv^2(t)), \quad N(u)=\left[\frac{\overleftarrow{\vv}(\gu
)}{\overleftarrow{\rho}(u^{-1})S}\right]+1.$$
For $L>0$ sufficiently small
\BQN\label{new1}
\EE{(\overline{X}(t)-\overline{X}(t))^2}\leq 2(1-r(s,t))\leq 4\rho^2(|t-s|)\leq \mathbb{Q}|t-s|^{\alpha/2}, \ \ s,t\in [0,L],
\EQN
which ensures the H\"{o}lder condition in a neighborhood of $0$.
By the fact that
$\sup_{t\in [\overleftarrow{\vv}(\gu ),T]}\sigma(t)\leq 1-\mathbb{Q}(\gu)^2$
for $u$ sufficiently large, (\ref{new1}), Theorem 8.1 in \cite{Pit96}, \Kd{we have}
\BQNY\label{proof2}
\pk{ \sup_{t\in [\overleftarrow{\vv}(\gu ),L]} X(t)>u}
\le
\mathbb{Q} T u^{4/\alpha}\Psi\left(\frac{u}{1-\mathbb{Q} (\gu)^2}\right).
\EQNY
Moreover, in light of Borell inequality (see e.g., \cite{GennaBorell}) and the fact that $\sup_{t\in [L,T]}\sigma(t)\leq 1-\delta$ with $\delta>0$,
$$\pk{ \sup_{t\in [\overleftarrow{\vv}(\gu ),L]} X(t)>u}
\le
e^{-\frac{(u-a)^2}{2(1-\delta)}}$$
with $a=\mathbb{E}\left(\sup_{t\in [0,T]}X(t)\right)$.
Consequently, for all large $u$ we have
\BQN\label{proof1}
\pi(u)\leq \pk{ \sup_{t\in [0,T]} X(t)>u}\leq \pi(u)+ \mathbb{Q} T u^{4/\alpha}\Psi\left(\frac{u}{1-\mathbb{Q} (\gu)^2}\right),
\EQN
where $\pi(u)=\pk{ \sup_{t\in [0, \overleftarrow{\vv}( \gu )]} X(t)>u}$.
\COM{
For $L>0$ sufficiently small
\BQNY
\EE{(\overline{X}(t)-\overline{X}(t))^2}\leq 2(1-r(s,t))\leq 4\rho^2(|t-s|)\leq \mathbb{Q}|t-s|^{\alpha/2}, \ \ s,t\in [0,L].
\EQNY
By the fact that
$\sup_{t\in [\overleftarrow{\vv}(\gu ),T]}\sigma(t)\leq 1-\mathbb{Q}(\gu)^2$
for $u$ sufficiently large and
in view of (\ref{HOLDER}) and Theorem 8.1 in \cite{Pit96}, we have
\BQN\label{proof2}
\pk{ \sup_{t\in [\overleftarrow{\vv}(\gu ),L]} X(t)>u}\leq \pk{ \sup_{t\in [\overleftarrow{\vv}(\gu ),L]} \overline{X}(t)>\frac{u}{1-\mathbb{Q} (\gu)^2}}\leq \mathbb{Q} T u^{4/\alpha}\Psi\left(\frac{u}{1-\mathbb{Q} (\gu)^2}\right)
\EQN
for some $\mathbb{Q}>0$.
Since \KD{$0$} is the unique point on $[0,T]$ such that $\sigma(\KD{0})=1$ is maximal,
there exists $\delta\in(0,1)$ such that
$\sup_{t\in [L,T]}\sigma(t)\leq 1-\delta$,
which together with Borel theorem (see, e.g., Theorem D.1 in \cite{Pit96})
implies that there exists a constant $a>0$ such that for $u$ sufficiently large,
\BQN\label{proof0}
\pk{ \sup_{t\in [\theta,T]} X(t)>u}\leq 2\Psi\left(\frac{u-a}{1-\delta}\right).
\EQN
\BQN\label{proof2}
\pk{ \sup_{t\in [\overleftarrow{\vv}(\gu ),L]} X(t)>u}\leq \pk{ \sup_{t\in [\overleftarrow{\vv}(\gu ),L]} \overline{X}(t)>\frac{u}{1-\mathbb{Q} (\gu)^2}}\leq \mathbb{Q} T u^{4/\alpha}\Psi\left(\frac{u}{1-\mathbb{Q} (\gu)^2}\right)
\EQN
}
Next we give the exact asymptotics of $\pi(u)$ subject to three different scenarios.
\iffalse As mentioned in \cite{HP2004},
for any $0<\epsilon<1/2$, we can find
stationary Gaussian processes $\{Y_{\pm\epsilon}(t), t\in \mathbb{R}\}$ with continuous trajectories, unit variance and correlation functions satisfying
$r_{\pm\epsilon}(t)=1-(1\pm \epsilon)\LL^2(t), \ \ t\downarrow 0.$\\ \fi
{\underline{Case i) $\gamma=0$}}. For any $u$ positive we have
\BQN\label{proof3}
\sum_{k=0}^{N(u)-1}\pk{ \sup_{t\in I_{k}(u)} X(t)>u}-\sum_{i=1}^2\Lambda_i(u)\leq \pi(u)\leq \sum_{k=0}^{N(u)}\pk{ \sup_{t\in I_{k}(u)} X(t)>u},
\EQN
where
$$\Lambda_1(u)=\sum_{k=0}^{N(u)}\pk{ \sup_{t\in I_{k}(u)} X(t)>u, \sup_{t\in I_{k+1}(u)} X(t)>u},$$ and $$\Lambda_2(u)=\sum_{0\leq k,l\leq N(u), l\geq
k+2}\pk{ \sup_{t\in I_{k}(u)} X(t)>u, \sup_{t\in I_{l}(u)} X(t)>u}.$$
\KD{
The main difference in comparison with the proofs of the classical
cases considered in the literature, as e.g., in \cite{Pit96},
is contained in the approximation given below.
}
By uniform convergence theorem (UCT) for regularly varying functions, see e.g., \cite{BI1989}, we have
$$\sup_{s,t\in I_{k}(u), 1\leq k\leq N(u)}\left|\frac{\vv^2(s)}{\vv^2(t)}-\left(\frac{s}{t}\right)^{\alpha}\right|\rw 0, \ \ u\rw\IF,$$
which implies that for any $\epsilon>0$ and for $u$ sufficiently large,
$$\frac{\vv^2(s)}{\vv^2(t)}\geq\left(\frac{k}{k+1}\right)^{\alpha}-\epsilon/2, \ \ s,t \in I_k(u), 1\leq k\leq N(u).$$
Thus for any $\epsilon>0$, there exists $k_\epsilon\in \mathbb{N}$ such that
$$\inf_{t\in I_{k}(u)}\vv^2(t)\geq (1-\epsilon)\sup_{t\in I_{k}(u)}\vv^2(t), \ \ k_\epsilon\leq k\leq N(u).$$
Let $\eMM{X_{u,k}}(t)=\overline{X}(kS\overleftarrow{\rho}(u^{-1})+t), t\in I_{0}(u)$ with $k\in \mathcal{K}_u=\{k, 0\leq k\leq N(u)\}$ and $h_k(u)=u_{k,\epsilon}^{-}$. In light of Lemma \ref{Pickands1}, we have
\BQN\label{callup}
\lim_{u\rw\IF}\sup_{0\leq k\leq N(u)}\left|(\Psi(u_{k,\epsilon}^{-}))^{-1}\pk{ \sup_{t\in I_{0}(u)}\eMM{X_{u,k}(t)}>u_{k,\epsilon}^{-}}
- \mathcal{H}_{\alpha}[0,S]\right|=0.
\EQN
\eMM{Consequently, as $u\to \IF$}
\BQNY\label{proof4}
\sum_{k=0}^{N(u)}\pk{ \sup_{t\in I_{k}(u)} X(t)>u}&\leq& \sum_{k=0}^{N(u)}\pk{ \sup_{t\in I_{k}(u)} \overline{X}(t)>u_{k,\epsilon}^{-}}\nonumber\\
&\leq& \sum_{k=0}^{N(u)}\pk{ \sup_{t\in I_{0}(u)}\eMM{X_{u,k}}(t)
>u_{k,\epsilon}^{-}}\nonumber\\
&\sim& \sum_{k=0}^{N(u)}\mathcal{H}_{\alpha}[0,S]\Psi(u_{k,\epsilon}^{-})\nonumber\\
&\sim& \mathcal{H}_{\alpha}[0,S]\Psi(u)\sum_{k=0}^{N(u)}e^{-u^2(1-\epsilon)\inf_{t\in I_{k}(u)}\vv^2(t)}.
\label{q1}
\EQNY
\eMM{Further by Lemma \ref{integ}}
\BQNY
\sum_{k=0}^{N(u)}\pk{ \sup_{t\in I_{k}(u)} X(t)>u}&\leq&
\mathcal{H}_{\alpha}[0,S]\Psi(u)\left(k_\epsilon+\frac{1}{\overleftarrow{\LL}(u^{-1})S}\sum_{k=k_\epsilon}^{N(u)}\int_{t\in I_k(u)}e^{-(1-\epsilon)^2u^2\vv^2(t)}dt\right),\nonumber\\
&\sim&
\mathcal{H}_{\alpha}[0,S]
\left(k_\epsilon+\frac{1}{\overleftarrow{\LL}(u^{-1})S}\int_0^{\overleftarrow{\vv}(\gu)}e^{-(1-\epsilon)^2u^2v^2(t)}dt\right)\Psi(u)\nonumber\\
&\sim& \Gamma(1/\beta+1)\mathcal{H}_{\alpha}\frac{\overleftarrow{\vv}(u^{-1})}{\overleftarrow{\LL}(u^{-1})}\Psi(u), \ \ u\rw\IF, S\rw\IF, \epsilon\rw 0.
\label{q1}
\EQNY
Similarly, we obtain
\BQN\label{proof5}
\sum_{k=0}^{N(u)-1}\pk{ \sup_{t\in I_{k}(u)} X(t)>u}\geq
\Gamma(1/\beta+1)\mathcal{H}_{\alpha}\frac{\overleftarrow{\vv}(u^{-1})}{\overleftarrow{\LL}(u^{-1})}\Psi(u)
\KD{(1+o(1))}, \ \ u\rw\IF, S\rw\IF.
\label{q2}
\EQN
{Next we focus on $\Lambda_i(u), i=1,2$. Let $\hat{u}_{k,-\epsilon}=\min (u_{k,-\epsilon}^-, u_{k+1,-\epsilon}^-)$. Then by (\ref{callup}) the fact that
\BQNY
&&\pk{ \sup_{t\in I_{k}(u)} X(t)>u, \sup_{t\in I_{k+1}(u)} X(t)>u}\\
&& \ \ \leq \pk{ \sup_{t\in I_{k}(u)}\overline{X}(t)>\hat{u}_{k,-\epsilon}}+\pk{\sup_{t\in
I_{k+1}(u)}\overline{X}(t)>\hat{u}_{k,-\epsilon}}-\pk{\sup_{t\in I_{k}(u)\cup I_{k+1}(u)} \overline{X}(t)>\hat{u}_{k,-\epsilon}},
\EQNY
we have
$$\lim_{S\rw\IF}\lim_{u\rw\IF}\sup_{0\leq k\leq N(u)}\frac{\pk{ \sup_{t\in I_{k}(u)} X(t)>u, \sup_{t\in I_{k+1}(u)} X(t)>u}}{\mathcal{H}_{\alpha}[0,S]\Psi(\hat{u}_{k,-\epsilon})}\leq \lim_{S\rw\IF}\left(2-\frac{\mathcal{H}_{\alpha}[0,2S]}{\mathcal{H}_{\alpha}[0,S]}\right)=0.$$}
{Therefore,
\BQNY\label{proof6}
\Lambda_1(u) &=&
o(1)\sum_{k=0}^{N(u)}\mathcal{H}_{\alpha}[0,S]\Psi(\hat{u}_{k,-\epsilon})\leq o(1)\sum_{k=0}^{N(u)}2\mathcal{H}_{\alpha}[0,S]\Psi(u_{k,-\epsilon})\\
&=& o\left(\frac{\overleftarrow{\vv}(u^{-1})}{\overleftarrow{\LL}(u^{-1})}\Psi(u)\right), \ \ u\rw\IF, S\rw\IF.
\EQNY
}
By (\ref{eR}) and applying Lemma \ref{uni} in Appendix, we have \HEH{(note that below $k, l$ take values up to $N_u$, therefore an uniform
upper bound for approximating the summuands derived in \nelem{uni} is essential)}
{\BQNY\label{proof7}
\Lambda_2(u)&\leq& \sum_{0\leq k,l\leq N(u), l\geq k+2} \pk{ \sup_{t\in I_{k}(u)} \overline{X}(t)>u_{k,-\epsilon}^-, \sup_{t\in I_{l}(u)} \overline{X}(t)>u_{l,-\epsilon}^-}\\
&\leq& \sum_{0\leq k,l\leq N(u), l\geq k+2} \mathbb{Q}S^2\Psi\left(\hat{u}_{k,l,-\epsilon}\right)
e^{-\mathbb{Q}_1|(l-k)S|^{\alpha/2}}\nonumber\\
&\leq& \mathbb{Q}S^2\sum_{0\leq k\leq N(u)}\Psi\left(u_{k,-\epsilon}^{-}\right)\sum_{l=1}^\IF
e^{-\mathbb{Q}_1(lS)^{\alpha/2}}\nonumber\\
&\leq& \mathbb{Q}S^2 e^{-\mathbb{Q}_2S^{\alpha/2}}\sum_{k=0}^{N(u)}\Psi\left(u_{k,-\epsilon}^-\right)\nonumber\\
&=&o\left(\frac{\overleftarrow{\vv}(u^{-1})}{\overleftarrow{\LL}(u^{-1})}\Psi(u)\right), \ \ u\rw\IF, S\rw\IF, \nonumber
\EQNY
with $\hat{u}_{k,l,-\epsilon}=\min(u_{k,-\epsilon}^-, u_{l,-\epsilon}^-)$.}
By the above \HEH{calculations} both $\Lambda_1(u)$ and $\Lambda_2(u)$ are negligible.
\eMM{Hence the results displayed in (\ref{proof2})-(\ref{proof7}) establish} the claim.\\
{\underline{Case ii) $\gamma \in (0,\IF]$ }.
\KD{
The proof of this case
is the same as the proof of the corresponding counterpart of Theorem D2 in \cite{Pit96},
with the exception that
\[
\pi(u)
\sim
\pk{ \sup_{t\in I_{0}(u)} X(t)>u}
\sim
\mathcal{P}_{\alpha}^{\gamma}[0, S]\Psi(u), \ \ u\rw\IF,
\]
where the last asymptotics follows by
Lemma \ref{Pickands1}.
}
This completes the proof. \QED
\subsection{Proofs of Theorems \ref{th.T1}, \ref{th.T2}, \ref{th.T3}, \ref{th.T4}, \ref{th.T5} \text{and} \ref{th.T6}}
Define next for $S,u$ positive
\begin{eqnarray*}
I_{k,l}(u)&=&[\overleftarrow{\LL_1}(u^{-1})kS, \overleftarrow{\LL_1}(u^{-1})(k+1)S]\times[\overleftarrow{\LL_2}(u^{-1})lS,
\overleftarrow{\LL_2}(u^{-1})(l+1)S], k, l \in \mathbb{N}\cup \{0\},\\
I_k(u)&=&[\overleftarrow{\LL_1}(u^{-1})kS, \overleftarrow{\LL_1}(u^{-1})(k+1)S], \quad
J_k(u)=[\overleftarrow{\LL_2}(u^{-1})kS,\overleftarrow{\LL_2}(u^{-1})(k+1)S],
\end{eqnarray*}
\KD{for $k\in\mathbb{N}\cup\{0\}$,}
and
$$ N_1(u)=\left[\frac{\overleftarrow{v_1}(u^{-1}\ln u)}{ \overleftarrow{\LL_1}(u^{-1})S}\right], \quad
N_2(u)=\left[\frac{ \overleftarrow{v_2}(u^{-1}\ln u)}{ \overleftarrow{\LL_2}(u^{-1})S}\right].$$
Additionally, let
$$\mathbb{V}_1(u)=\{(k,l,k_1,l_1): -N_1(u)-2\leq k\leq k_1\leq N_1(u)+1, -N_2(u)-2\leq l,l_1\leq N_2(u)+1, I_{k,l}\cap I_{k_1,l_1}=\emptyset\},$$
$$\mathbb{V}_2(u)=\{(k,l,k_1,l_1): -N_1(u)-2\leq k\leq k_1\leq N_1(u)+1, -N_2(u)-2\leq l,l_1\leq N_2(u)+1, (k,l)\neq (k_1,l_1), I_{k,l}\cap I_{k_1,l_1}\neq\emptyset\},$$
$$u^-_{k,l,\epsilon}=u(1+(1-\epsilon)\inf_{(s,t)\in I_{k,l}(u)}\left(v_1^2(|b_{11}s+b_{12}t|)+v_2^2(|b_{21}s+b_{22}t|)\right)),$$
$$u^{+}_{k,l,\epsilon}=u(1+(1+\epsilon)\sup_{(s,t)\in I_{k,l}(u)}\left(v_1^2(|b_{11}s+b_{12}t|)+v_2^2(|b_{21}s+b_{22}t|)\right)),$$
$$u^{1,-}_{k,\epsilon}=u(1+(1-\epsilon)\inf_{s\in I_{k}(u)}v_1^2(|s|)), \ \ u^{1,+}_{k,\epsilon}=u(1+(1+\epsilon)\sup_{s\in I_{k}(u)}v_1^2(|s|)),$$
$$u^{2,-}_{l,\epsilon}=u(1+(1-\epsilon)\inf_{s\in J_{l}(u)}v_2^2(|s|)), \ \ u^{2,+}_{l,\epsilon}=u(1+(1+\epsilon)\sup_{s\in J_{l}(u)}v_2^2(|s|)), k,l\in \mathbb{Z},$$ where $u^{\pm}_{k,l,\epsilon}$ varies according to $B$.
\KD{
In what follows \eMM{for a given Gaussian random random field $Z$ we write $\overline{Z}$ for the standardised random field.}
}
\KD{\\
The general strategy of proofs of Theorems \ref{th.T1}, \ref{th.T2}, \ref{th.T3}, \ref{th.T4}, \ref{th.T5} \text{and} \ref{th.T6}
agrees from the double-sum technique developed for Gaussian random fields
in e.g., \cite{Pit96}.
However the variance-covariance structure of some cases substantially differs
from the families of Gaussian random fields analyzed in \cite{Pit96} and requires a case-specific approach,
on which we focus below.
\\
Observe that for all Cases 1-6
\BQN\label{F1}
\pi_1(u)\leq\pk{ \sup_{(s,t)\in [-T_1,T_1]\times[-T_2,T_2]} X(s,t)>u}\leq \pi_1(u) +\pk{ \sup_{(s,t)\in \left([-T_1,T_1]\times[-T_2,T_2]\right)
\setminus D_u} X(s,t)>u},
\EQN
where
$$\pi_1(u)=\pk{ \sup_{(s,t)\in D_u} X(s,t)>u},\text{with} \ \ D_u=[-\overleftarrow{v_1}\left(u^{-1}\ln u\right), \overleftarrow{v_1}\left(u^{-1}\ln
u\right)]\times[-\overleftarrow{v_2}\left(u^{-1}\ln u\right), \overleftarrow{v_2}\left(u^{-1}\ln u\right)].$$
For Case 1-Case 3 and Case 5-Case 6, by (\ref{Var2}) for $u$ sufficiently large we have
$$\sup_{(s,t)\in \left([-T_1,T_1]\times[-T_2,T_2]\right) \setminus D_u}\sigma(s,t)\leq 1-\mathbb{Q}u^{-2}\ln^2 u.$$
For Case 4, in light of (\ref{Var2}) and Lemma \ref{simple1}, we have
$$\sup_{(s,t)\in \left([-T_1,T_1]\times[-T_2,T_2]\right) \setminus D_u}\sigma(s,t)\leq 1-\mathbb{Q}\inf_{(s,t)\in \left([-T_1,T_1]\times[-T_2,T_2]\right) \setminus D_u}\left(v_1^2(|s|)+v_2^2(|t|)\right)\leq 1-\mathbb{Q}u^{-2}\ln^2 u.$$
\eMM{It follows by the fact that $(0,0)$ is the unique maximizer of $\sigma$, Theorem 8.1 in \cite{Pit96} and Borell theorem
that}
\BQN\label{neww}
\pk{ \sup_{(s,t)\in \left([-T_1,T_1]\times[-T_2,T_2]\right)
\setminus D_u} X(s,t)>u} \le \mathbb{Q}T_1 T_2u^{4/\alpha_1+4/\alpha_2}\Psi\left(\frac{u}{1-2u^{-2}\ln^2 u}\right).
\EQN
\COM{
point thatBy the assumption of the Similarly as in (\ref{eq3}), there exists $\delta>0$ sufficiently small such that
$$\EE{(\overline{X}(s,t)-\overline{X}(s_1,t_1))^2}\leq \mathbb{Q}_1(|s-s_1|^{\alpha_1/2}+|t-t_1|^{\alpha_2/2}), \ \ (s,t), (s_1,t_1)\in [-\delta,\delta]^2.$$
Thus in light of Theorem 8.1 in \cite{Pit96}, we have, for $u$ sufficiently large and $\delta$ sufficiently small
\BQN\label{F2}
\pk{ \sup_{(s,t)\in [-\delta,\delta]^2 \setminus D_u} X(s,t)>u}&\leq& \pk{ \sup_{(s,t)\in [-\delta,\delta]^2 \setminus D_u}
\overline{X}(s,t)>\frac{u}{1-2u^{-2}\ln^2 u}} \nonumber \\
&\leq& \mathbb{Q}T T_1u^{4/\alpha_1+4/\alpha_2}\Psi\left(\frac{u}{1-2u^{-2}\ln^2 u}\right).
\EQN
There exists a positive constant $0<\delta_1<1$ such that for all Case 1-Case 6
$$\sup_{(s,t)\in \left([-T_1,T_1]\times[-T_2,T_2]\right) \setminus [-\delta,\delta]^2}\sigma(s,t)\leq 1-\delta_1,$$
which together with Borell theorem gives that
\BQN\label{F3}
\pk{ \sup_{(s,t)\in \left([-T_1,T_1]\times[-T_2,T_2]\right) \setminus[-\delta,\delta]^2} X(s,t)>u}\leq 2\Psi\left(\frac{u-a}{1-\delta_1}\right),
\EQN
with $a>0$.
\\
}
Therefore, for all Cases 1-6 we focus on the asymptotics of $\pi_1(u)$ as $u\rw\IF$,
proving that it delivers the asymptotics of (\ref{task.2}) as $u\to\infty$.
}
\\
\prooftheo{th.T1}\\
\underline{ Case i).}
\KD{Suppose that $\gamma_1=\gamma_2=0.$}
For any $0<\epsilon<1/2$ and $u$ large enough we have
\BQN\label{main}
\pi_{1,\epsilon}(u)-\sum_{i=1}^2\Lambda_i'(u)\leq \pi_1(u)\leq \pi_{1,-\epsilon}(u),
\EQN
with
\BQNY
\pi_{1,\pm \epsilon}(u):&=&\sum_{k=-N_1(u)\pm 2}^{N_1(u)\mp 1}\sum_{l=-N_2(u)\pm 2}^{N_2(u)\mp 1}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^{\pm}\right),\\
\Lambda_1'(u):&=&\sum_{(k,l,k_1,l_1)\in \mathbb{V}_1(u) }\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-, \sup_{(s,t)\in
I_{k_1,l_1}(u)}\overline{X}(s,t)>u_{k_1,l_1,\epsilon}^-\right),\\
\Lambda_2'(u):&=&\sum_{(k,l,k_1,l_1)\in \mathbb{V}_2(u) }\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-, \sup_{(s,t)\in
I_{k_1,l_1}(u)}\overline{X}(s,t)>u_{k_1,l_1,\epsilon}^-\right).
\EQNY
By UCT, for any $\epsilon>0$, there exist two constants $k_\epsilon, l_\epsilon\in \mathbb{N}$ such that
\BQN\label{eq4}
&&\inf_{t\in I_{k}(u)}\vv_1^2(t)\geq (1-\epsilon)\sup_{t\in I_{k}(u)}\vv_1^2(t), \ \ \inf_{t\in J_{l}(u)}\vv_2^2(t)\geq (1-\epsilon)\sup_{t\in J_{l}(u)}\vv_2^2(t),
\EQN
hold for $k_\epsilon\leq |k|\leq N_1(u)+2, l_\epsilon\leq |l|\leq N_2(u)+2.$
Let
$$X_{u,k,l}(s,t)=\overline{X}(kS\overleftarrow{\rho}_1(u^{-1})+s,lS\overleftarrow{\rho}_2(u^{-1})+t), \mathcal{K}_u=\{(k,l), |k|\leq N_1(u)+2, |l|\leq N_2(u)+2\},$$
$$h_{k,l}(u)=u_{k,l,\epsilon}^-, \mathcal{E}_u=I_{0,0}(u), d_u=0.$$
One can easily check that conditions of Lemma \ref{PIPI} \eMM{are} satisfied \eMM{implying that}
\BQN\label{uniform}
\lim_{u\rw\IF}\sup_{(k,l)\in \mathcal{K}_u}\left|(\Psi(u_{k,l,\epsilon}^-))^{-1}\mathbb{P}\left(\sup_{(s,t)\in
I_{0,0}(u)}\overline{X}(kS\overleftarrow{\rho}_1(u^{-1})+s,lS\overleftarrow{\rho}_2(u^{-1})+t)>u_{k,l,\epsilon}^-\right)-
\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\right|=0.\EQN
Further, using Lemma \ref{integ} we have
\BQN\label{new}
\pi_{1,-\epsilon}(u
&& \ = \sum_{k=-N_1(u)-2}^{N_1(u)+ 1}\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}\mathbb{P}\left(\sup_{(s,t)\in
I_{0,0}(u)}\overline{X}(kS\overleftarrow{\rho}_1(u^{-1})+s,lS\overleftarrow{\rho}_2(u^{-1})+t)>u_{k,l,\epsilon}^-\right)\nonumber\\
&& \ \ \sim \sum_{k=-N_1(u)- 2}^{N_1(u)+ 1}\sum_{l=-N_2(u)- 2}^{N_2(u)+
1}\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u_{k,l,\epsilon}^-)\nonumber\\
&& \ \ \sim\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u)\left(R_1(u)+R_2(u)+\sum_{k_\epsilon\leq|k|\leq N_1(u)+2}
\frac{1}{\overleftarrow{\LL_1}(u^{-1})S}\int_{s\in I_k(u)}e^{-(1-\epsilon)^2u^2v_1^2(|t|)}dt\right.\nonumber\\
&& \ \ \ \left.\times\sum_{l_\epsilon\leq |l|\leq N_2(u)+2}\frac{1}{\overleftarrow{\LL_2}(u^{-1})S}\int_{t\in J_l(u)}e^{-(1-\epsilon)^2u^2v_2^2(t)}dt\right)\nonumber\\
&& \ \ \sim
4\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u)\frac{1}{\overleftarrow{\LL_1}(u^{-1})S}\int_0^{
\overleftarrow{v_1}(u^{-1}\ln u)}e^{-(1-\epsilon)^2u^2v^2_1(t)}dt\nonumber\\
&& \ \ \ \ \ \times \frac{1}{\overleftarrow{\LL_2}(u^{-1})S}\int_0^{
\overleftarrow{v_2}(u^{-1}\ln u)}e^{-(1-\epsilon)^2u^2v^2_2(t)}dt\nonumber\\
&& \ \ \sim
4(1-\epsilon)^{-1/\beta_1-1/\beta_2}\Gamma(1/\beta_1+1)\Gamma(1/\beta_2+1)S^{-2}\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]
\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\frac{\overleftarrow{\vv}_2(1/u)}{\overleftarrow{\LL}_2(1/u)}\Psi(u)\nonumber\\
&& \ \ \sim 4\Gamma(1/\beta_1+1)\Gamma(1/\beta_2+1)\prod_{i=1}^2\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\frac{\overleftarrow{\vv}_2(1/u)}{\overleftarrow{\LL}_2(1/u)}\Psi(u), \ \ u\rw \IF,
S\rw\IF, \epsilon\rw 0,\label{mainpart1}
\EQN
where
$$R_1(u)=\sum_{|k|\leq k_\epsilon}\sum_{|l|\leq N_2(u)+2}e^{-(1-\epsilon)u^2\inf_{s\in I_{k}(u)}v_1^2(|s|)-(1-\epsilon)u^2\inf_{t\in J_l(u)}v_2^2(|t|)},$$
$$R_2(u)=\sum_{|k|\leq N_1(u)+2}\sum_{|l|\leq l_\epsilon}e^{-(1-\epsilon)u^2\inf_{s\in I_{k}(u)}v_1^2(|s|)-(1-\epsilon)u^2\inf_{t\in J_l(u)}v_2^2(|t|)}.$$
Note that (\ref{new}) holds since in light of Lemma \ref{integ} we have
\BQNY
R_1(u)&\leq& (2k_\epsilon+1) \left(2l_\epsilon+1+\frac{1}{\overleftarrow{\LL_2}(u^{-1})S}\int_0^{
\overleftarrow{v_2}(u^{-1}\ln u)}e^{-(1-\epsilon)^2u^2v^2_2(t)}dt\right)\\
&\sim&(2k_\epsilon+1)(1-\epsilon)^{-1/\beta_2} \frac{\overleftarrow{\vv}_2(1/u)}{\overleftarrow{\LL}_2(1/u)S}=
o\left(\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\frac{\overleftarrow{\vv}_2(1/u)}{\overleftarrow{\LL}_2(1/u)}\right), \ \ u\rw\IF,
\EQNY
and
\BQNY
R_2(u)&\leq& (2l_\epsilon+1) \left(2k_\epsilon+1+\frac{1}{\overleftarrow{\LL_1}(u^{-1})S}\int_0^{
\overleftarrow{v_1}(u^{-1}\ln u)}e^{-(1-\epsilon)^2u^2v^2_1(t)}dt\right)\\
&\sim&(2l_\epsilon+1)(1-\epsilon)^{-1/\beta_1} \frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)S}\\
&=&
o\left(\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\frac{\overleftarrow{\vv}_2(1/u)}{\overleftarrow{\LL}_2(1/u)}\right), \ \ u\rw\IF.
\EQNY
Similarly,
\BQN\label{Addition1}
\pi_{1,\epsilon}(u)&\geq& \sum_{k=-N_1(u)+1}^{N_1(u)-1}\sum_{l=-N_2(u)+1}^{N_2(u)- 1}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^+\right)\nonumber\\ &\sim& 4\Gamma(1/\beta_1+1)\Gamma(1/\beta_2+1)\prod_{i=1}^2\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\frac{\overleftarrow{\vv}_2(1/u)}{\overleftarrow{\LL}_2(1/u)}\Psi(u),
\EQN
as $u\rw\IF, S\rw\IF, \epsilon\rw 0$.
\KD{Next we prove that both $\Lambda'_1(u),\Lambda'_2(u)$ are asymptotically negligible.}
\iffalse we derive that, for $u$ large enough
\BQN\label{Var3}
2\leq Var\left(\overline{X}(s,t)+\overline{X}(s_1,t_1)\right)&=&4-2(1-r_X(s,t,s_1,t_1))\leq 4-(\LL_1^2(|s-s_1|)+\LL_2^2(|t-t_1|))\nonumber\\
&\leq& 4-\mathbb{Q}_1u^{-2}(|k-k_1|^{\alpha_1/2}S^{\alpha_1/2}+|l-l_1|^{\alpha_2/2}S^{\alpha_2/2}),
\EQN
holds for $(s,t)\in I_{k,l}(u), (s_1,t_1)\in I_{k_1,l_1}(u), (k,l,k_1,l_1)\in \mathbb{V}_1(u).$
In light of (\ref{Cor2}) and the fact that $\LL^2_i, i=1,2$ are regularly varying at zero, we have
that
\BQN\label{Cor3}
&&1-Cor\left(\overline{X}(s,t)+\overline{X}(s_1,t_1), \overline{X}(s',t')+\overline{X}(s_1',t_1')\right)\nonumber\\
&\quad&\quad\leq \LL_1^2(|s-s'|)+\LL_1^2(|s_1-s_1'|)+\LL_2^2(|t-t'|)+\LL_2^2(|t_1-t_1'|)\nonumber\\
&\quad&\quad\leq C_1u^{-2}\left(\left|\frac{s-s'}{\overleftarrow{\LL_1}(u^{-1})}\right|^{\alpha_1/2}+\left|\frac{s_1-s_1'}
{\overleftarrow{\LL_1}(u^{-1})}\right|^{\alpha_1/2}+\left|\frac{t-t'}{\overleftarrow{\LL_2}(u^{-1})}\right|^{\alpha_2/2}+\left|\frac{t-t'}
{\overleftarrow{\LL_2}(u^{-1})}\right|^{\alpha_2/2}\right)
\EQN
holds with $C_1=2(S^{\alpha_1/2}+S^{\alpha_2/2})$ for $ (s,t,S,t_1), (s',t',s_1',t_1')\in I_{k,l}(u)\times I_{k_1,l_1}(u)$ with $|k|, |k_1|\leq
N_1(u)+2, |l|, |l_1|\leq N_2(u)+2$ and $u$ large enough.
Let $X_u^*(s,t,s_1,t_1), (s,t,s_1,t_1)\in D_u\times D_u$ be a stationary Gaussian random field such
that its correlation function satisfies
\BQNY
&&1-Cor\left(X_u^*(s,t,s_1,t_1), X_u^*(s',t',s_1',t_1')\right)\\
&=&2C_1u^{-2}\left(\left|\frac{s-s'}{\overleftarrow{\LL_1}(u^{-1})}\right|^{\alpha_1/2}+\left|\frac{s_1-s_1'}
{\overleftarrow{\LL_1}(u^{-1})}\right|^{\alpha_1/2}+\left|\frac{t-t'}{\overleftarrow{\LL_2}(u^{-1})}\right|^{\alpha_2/2}+\left|\frac{t-t'}
{\overleftarrow{\LL_2}(u^{-1})}\right|^{\alpha_2/2}\right)(1+o(1))
\EQNY
for $(s,t), (s_1,t_1), (s',t'), (s_1',t_1')\in D_u$ and $u\rw\IF$.\fi
\eMM{From (\ref{Cor2})}, applying Lemma \ref{uni} in the Appendix,
with
$$\hat{u}_{k,l,k_1,l_1,\epsilon}=\min(u_{k,l,\epsilon}^-, u_{k_1,l_1,\epsilon}^-), \quad
\beta^*=\min(\alpha_1,\alpha_2),$$
we obtain
\BQN\label{lambda1}
\Lambda_1'(u)
&\leq& \mathbb{Q}S^{4} \sum_{(k,l,k_1,l_1)\in \mathbb{V}_1(u) }
\Psi\left(\hat{u}_{k,l,k_1,l_1,\epsilon}\right)e^{-\mathbb{Q}_1(|k-k_1|^2+|l-l_1|^2)^{\beta^*}S^{\beta^*/2}}\nonumber\\
&\leq& \mathbb{Q}S^{4}
\sum_{k=-N_1(u)-2}^{N_1(u)+1}\sum_{l=-N_2(u)-2}^{N_2(u)+1}\Psi(u_{k,l,\epsilon}^-)\sum_{m+n\geq 1, m,n \geq 0}
e^{-\mathbb{Q}_1(|m|^2+|n|^2)^{\beta^*}S^{\beta^*/2}}\nonumber\\
&\leq &\mathbb{Q}S^{4}
\sum_{k=-N_1(u)-2}^{N_1(u)+1}\sum_{l=-N_2(u)-2}^{N_2(u)+1}\Psi(u_{k,l,\epsilon}^-)e^{-\mathbb{Q}_2S^{\beta^*/2}}\nonumber\\
&=&o\left(\pi_{1,\epsilon}(u)\right), \ \ u\rw\IF, S\rw\IF.
\EQN
Now we focus on $\Lambda_2(u)$. Without loss of generality, we assume that $k_1=k+1$. Then let, for $k_1, l_1 \in \mathbb{N}\cup \{0\}$,
$$I^{(1)}_{k_1,l_1}=[\overleftarrow{\LL_1}(u^{-1})k_1S, \overleftarrow{\LL_1}(u^{-1})\left(k_1S+\sqrt{S}\right)]
\times[\overleftarrow{\LL_2}(u^{-1})l_1S, \overleftarrow{\LL_2}(u^{-1})(l_1+1)S],$$
$$I^{(2)}_{k_1,l_1}=[\overleftarrow{\LL_1}(u^{-1})\left(k_1S+\sqrt{S}\right),
\overleftarrow{\LL_1}(u^{-1})(k_1+1)S]\times[\overleftarrow{\LL_2}(u^{-1})l_1S, \overleftarrow{\LL_2}(u^{-1})(l_1+1)S].$$ For $(k,l,k_1,l_1)\in
\mathbb{V}_2(u), k_1=k+1$, we have
\BQNY
&&\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-, \sup_{(s,t)\in
I_{k_1,l_1}(u)}\overline{X}(s,t)>u_{k_1,l_1,\epsilon}^-\right)\\
&& \ \ \leq
\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-, \sup_{(s,t)\in
I^{(1)}_{k_1,l_1}(u)}\overline{X}(s,t)>u_{k_1,l_1,\epsilon}^-\right)\\
&& \ \ \ \ +
\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-, \sup_{(s,t)\in
I^{(2)}_{k_1,l_1}(u)}\overline{X}(s,t)>u_{k_1,l_1,\epsilon}^-\right)\\
&& \ \ :=p_{k,l,k_1,l_1}^{(1)}(u)+p_{k,l,k_1,l_1}^{(2)}(u).
\EQNY
It follows from Lemma \ref{PIPI} that
\BQNY
p_{k,l,k_1,l_1}^{(1)}(u)\leq \mathbb{P}\left( \sup_{(s,t)\in I^{(1)}_{k_1,l_1}(u)}\overline{X}(s,t)>u_{k_1,l_1,\epsilon}^-\right)\sim
\mathcal{H}_{\alpha_1}[0,\sqrt{S}]\mathcal{H}_{\alpha_2}[0,S]\Psi(u_{k,l,\epsilon}^-).
\EQNY
Further, since each $I_{k,l}(u)\times I_{k_1,l_1}(u)$ has at most $8$ neighbors, we have that
\BQNY
\lefteqn{\sum_{(k,l,k_1,l_1)\in \mathbb{V}_2(u)}p_{k,l,k_1,l_1}^{(1)}(u
\ \ \leq 8\sum_{k=-N_1(u)-2}^{N_1(u)+1}\sum_{l=-N_2(u)-2}^{N_2(u)+1}\Psi(u_{k,l,\epsilon}^-)}\\
&& \ \ \ \ \ \times \left(\mathcal{H}_{\alpha_1}[0,\sqrt{S}]
\mathcal{H}_{\alpha_2}[0,S]+
\mathcal{H}_{\alpha_1}
[0, S]\mathcal{H}_{\alpha_2}[0,\sqrt{S}]\right)\\
&& \ \ \leq 8\left(\frac{\mathcal{H}_{\alpha_1}[0,\sqrt{S}]}{\mathcal{H}_{\alpha_1}[0,S]}
+\frac{\mathcal{H}_{B_{\alpha_2}}
[0,\sqrt{S}]}{\mathcal{H}_{\alpha_2}[0,S]}\right)\nonumber\\
&& \ \ \ \ \ \times
\sum_{k=-N_1(u)-2}^{N_1(u)+1}\sum_{l=-N_2(u)-2}^{N_2(u)+1}\mathcal{H}_{\alpha_1}[0,S]
\mathcal{H}_{\alpha_2}[0,S]\Psi(u_{k,l,\epsilon}^-)\\
&& \ \ =o\left(\pi_{1,\epsilon}(u)\right), \ \ u\rw\IF, S\rw\IF.
\EQNY
In light of Lemma \ref{uni}, we have
\BQNY
\sum_{(k,l,k_1,l_1)\in \mathbb{V}_2(u)}p_{k,l,k_1,l_1}^{(2)}(u)
& \leq& \mathbb{Q}S^{4}e^{-\mathbb{Q}_1S^{\beta^*/4}}
\sum_{(k,l,k_1,l_1)\in \mathbb{V}_2(u)}\Psi(\hat{u}_{k,l,k_1,l_1,\epsilon})\\
&\leq& \mathbb{Q}S^{4}e^{-\mathbb{Q}_1S^{\beta^*/4}}
\sum_{k=-N_1(u)-2}^{N_1(u)+1}\sum_{l=-N_2(u)-2}^{N_2(u)+1}\Psi(u_{k,l,\epsilon}^-)\\
& =& o\left(\pi_{1,\epsilon}(u)\right), \ \ u\rw\IF, S\rw\IF.
\EQNY
Consequently,
\BQN\label{lambda2}
\Lambda_2'(u)=o\left(\pi_{1,\epsilon}(u)\right), \ \ u\rw\IF, S\rw\IF.
\EQN
Combing (\ref{main}), (\ref{mainpart1}), (\ref{Addition1}) (\ref{lambda1}) with (\ref{lambda2}), we derive that
$$\pi_1(u)\sim 4\Gamma(1/\beta_1+1)\Gamma(1/\beta_2+1)\prod_{i=1}^2\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_1(1/u)}{\overleftarrow{\LL}_1(1/u)}\frac{\overleftarrow{\vv}_2(1/u)}{\overleftarrow{\LL}_2(1/u)}\Psi(u), \ \ u\rw \IF,$$
\eMM{hence the claim follows}. \\
\underline{ Case ii) $\gamma_1=0, \gamma_2\in (0,\IF)$.} Let in the sequel
$$\widetilde{I}_{k,0}(u)=I_{k,0}(u)\cup I_{k,-1}(u), \quad \mathbb{V}_1^{(1)}(u)=\{(k,k_1):
-N_1(u)-2\leq k<k_1\leq N_1(u)+1, k_1-k\geq 2\},$$
and $$\mathbb{V}_2^{(1)}(u)=\{(k,k_1): -N_1(u)-2\leq k<k_1\leq N_1(u)+1, k_1=k+1\}.$$
\CE{For} any $0<\epsilon<1$ and all $u$ large enough
\BQN\label{main1}
\pi_{1,\epsilon}^{(1)}(u)-\sum_{i=1}^2\Lambda_i^{(1)}(u)\leq \pi_1(u)\leq \pi_{1,-\epsilon}^{(1)}(u)+\pi_{1,-\epsilon}^{(2)}(u),
\EQN
with
\BQNY
\pi_{1,\pm \epsilon}^{(1)}(u):&=&\sum_{k=-N_1(u)\pm 2}^{N_1(u)\mp1}\mathbb{P}\left(\sup_{(s,t)\in
\widetilde{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1\pm\epsilon)v^2_2(t)}>u_{k,\epsilon}^{1,\pm}\right)\nonumber\\
\pi_{1,-\epsilon}^{(2)}(u):&=&\sum_{k=-N_1(u)-2}^{N_1(u)+1}\sum_{|l|=1, l\neq -1}^{N_2(u)+1}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-\right)\nonumber\\
\HEH{\Lambda_i}^{(1)}(u):&=&\sum_{(k,k_1)\in \mathbb{V}_i^{(1)}(u) }\mathbb{P}\left(\sup_{(s,t)\in \widetilde{I}_{k,0}(u)}\overline{X}(s,t)>u_{k,\epsilon}^{1,-},
\sup_{(s,t)\in \widetilde{I}_{k_1,0}(u)}\overline{X}(s,t)>u_{k_1,\epsilon}^{1,-}\right), \quad \HEH{i=1,2,}\nonumber\\
\EQNY
Set further $X_{u,k}(s,t)=\overline{X}(k\overleftarrow{\rho}_1(u^{-1})S+s,t)$ and define
$$ \mathcal{K}_u=\{k, |k|\leq N_1(u)+2\}, \quad \mathcal{E}_u=\tilde{I}_{0,0}(u), \quad h_k(u)=u_{k,\epsilon}^{1,-}, \quad d_u(s,t)=(1-\epsilon)v^2_2(t).$$
Using Lemma {\ref{PIPI}}, we have
$$\lim_{u\rw\IF}\sup_{k\in \mathcal{K}_u}\left|(\Psi(u_{k,\epsilon}^{1,-}))^{-1}\mathbb{P}\left(\sup_{(s,t)\in
\widetilde{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)v^2_2(t)}>u_{k,\epsilon}^{1,-}\right)-\mathcal{H}_{\alpha_1}[0,
S]\widehat{\mathcal{P}}^{\gamma_2(1-\epsilon)}_{\alpha_2}(S)\right|=0.$$
Further, by Lemma \ref{integ}, we have
\BQN\label{PG1}
\pi_{1,- \epsilon}^{(1)}(u)
&\sim&\sum_{k=-N_1(u)-2}^{N_1(u)+1}\mathcal{H}_{\alpha_1}[0,
S]\widehat{\mathcal{P}}^{\gamma_2(1-\epsilon)}_{\alpha_2}(S)\Psi(u_{k,\epsilon}^{1,-})\nonumber\\
&\leq& 2\mathcal{H}_{\alpha_1}[0,
S]\widehat{\mathcal{P}}^{\gamma_2(1-\epsilon)}_{\alpha_2}(S)
\left( 2k_{\epsilon}+1+\frac{\Psi(u)}{\overleftarrow{\LL}_1(u^{-1})S}\int_0^{\overleftarrow{v_1}(u^{-1}\ln u)}e^{-(1-\epsilon)^2u^2v^2_1(t)}dt\right)(1+o(1))\nonumber\\
&\sim&
2\Gamma(1/\beta_1+1)\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}\frac{\overleftarrow{\vv}_1(u^{-1})}{\overleftarrow{\LL}_1(u^{-1})
}\Psi(u), \ \ u\rw \IF, \epsilon\rw 0, S\rw\IF.
\EQN
Similarly, as $u\rw \IF, \eMM{\ve \to 0}, S\rw\IF,$
\BQN\label{Addition2}
\pi_{1, \epsilon}^{(1)}(u)
\sim
2\Gamma(1/\beta_1+1)\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}\frac{\overleftarrow{\vv}_1(u^{-1})}{\overleftarrow{\LL}_1(u^{-1})
}\Psi(u).
\EQN
Moreover, by Lemma \ref{PIPI}
\BQN
\pi_{1,-\epsilon}^{(2)}(u)
&\sim& \sum_{k=-N_1(u)-2}^{N_1(u)+2}\sum_{|l|=1, l\neq -1}^{N_2(u)+1}\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,
S]\Psi(u_{k,l,\epsilon}^-)\nonumber\\
&\leq & 2\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0, S]\Psi(u)
\sum_{k=-N_1(u)-2}^{N_1(u)+2}e^{-(1-\epsilon)u^2\inf_{s\in I_k(u)}v_1^2
(|s|)} \sum_{|l|=1, l\neq -1}^{N_2(u)+1}e^{-(1-\epsilon)u^2v_2^2(\overleftarrow{\LL_2}(u^{-1}))|lS|^{\beta_2'}}\nonumber\\
&\leq&2\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0, S]\Psi(u)\sum_{k=-N_1(u)-2}^{N_1(u)+ 1}
e^{-(1-\epsilon)u^2\inf_{s\in I_k(u)}v_1^2
(|s|)}\sum_{|l|=1, l\neq -1}^{N_2(u)+1}e^{-(1-2\epsilon)\gamma_2|lS|^{\beta_2'}}\label{Addition3}\\
&\leq& 4\prod_{i=1}^2\mathcal{H}_{\alpha_i}\frac{\overleftarrow{\vv}_1(u^{-1})\Psi(u)}{(1-\epsilon)^{2/\beta_1}\overleftarrow{\LL}_1(u^{-1})
}S^2 e^{-\mathbb{Q}_4S^{\beta_2'}}\nonumber\\
&=&o\left(\pi_{1,\epsilon}^{(1)}(u)\right)
\label{PG2}
\EQN
as $u\rw\IF, S\rw\IF$ and $\eMM{\ve \to 0}$, with $0<\beta_2'<\beta_2$.
Note that in (\ref{Addition3}) we use Potter's \HEH{bounds} (see e.g., \cite{Res})
for regularly varying function $v_2(t)$ at zero to derive that, for $u$ and $S$ large enough,
\BQN\label{A0}
\frac{v_2^2(\overleftarrow{\LL_2}(u^{-1})lS)}{v_2^2(\overleftarrow{\LL_2}(u^{-1}))}\geq \left(lS\right)^{\beta_2'}
\EQN
holds for $1\leq l\leq N_2(u) $.
\iffalse Referring to (\ref{Var3}), we have for $(s,t)\in \widetilde{I}_{k,0}(u), (s_1,t_1)\in I_{k_1,0}(u), (k,k_1)\in \mathbb{V}_1^{(1)}(u)$
\BQN\label{Var4}
2\leq Var\left(\overline{X}(s,t)+\overline{X}(s_1,t_1)\right)
\leq 4-\mathbb{Q}u^{-2}|k-k_1|^{\alpha_1/2}S^{\alpha_1/2}.
\EQN
\fi
Using Lemma \ref{uni}, we have
\BQN\label{lambda11}
\Lambda_1^{(1)}(u)
&\leq& \mathbb{Q}S^{4}
\sum_{(k,k_1)\in \mathbb{V}_1^{(1)}(u) }\Psi
\left(u_{k,k_1,\epsilon}^1\right)e^{-\mathbb{Q}_1(|k-k_1|S)^{\beta^*/2}}\nonumber\\
&\leq& \mathbb{Q}S^{4}
\sum_{k=-N_1(u)-2}^{N_1(u)+1}\Psi(u_{k,\epsilon}^{1,-})\sum_{m\geq 1} e^{-\mathbb{Q}_1m^{\beta^*/2}S^{\beta^*/2}}\nonumber\\
&\leq & \mathbb{Q}S^{4}
\sum_{k=-N_1(u)-2}^{N_1(u)+1}\Psi(u_{k,\epsilon}^{1,-})e^{-\mathbb{Q}_2S^{\beta^*/2}}\nonumber\\
&=&o\left(\pi_{1,\epsilon}^{(1)}(u)\right),\ \ u\rw\IF, S\rw\IF,
\EQN
with $\hat{u}_{k,k_1,\epsilon}=\min(u_{k,\epsilon}^{1,-}, u_{k_1,\epsilon}^{1,-})$ and $\beta^*=\min (\alpha_1,\alpha_2)$.
Using Lemma \ref{PIPI} yields that
\BQN\label{lambda21}
\Lambda_2^{(1)}(u)&\leq&\sum_{k=-N_1(u)-2}^{N_1(u)+1}\Bigg[
\pk{\sup_{(s,t)\in
\widetilde{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)v^2_2(t)}>u_{k,\epsilon}^{1,-}} \\
&& +\pk{ \sup_{(s,t)\in
\widetilde{I}_{k+1,0}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)v^2_2(t)}>u_{k+1,\epsilon}^{1,-}} \nonumber\\
&& -\pk{\sup_{(s,t)\in \widetilde{I}_{k,0}(u)\cup
\widetilde{I}_{k+1,0}(u)}\frac{\overline{X}(s,t)}{1+(1+\epsilon)v^2_2(t)}>\check{u}_{k,k+1,\epsilon}} \Bigg]\nonumber\\
&\leq&
\left((1+\epsilon)2\mathcal{H}_{B_{\alpha_1}}[0,S]-(1-\epsilon)\mathcal{H}_{B_{\alpha_1}}[0,2S]\right)\nonumber\\
&\ \ &\times\widehat{\mathcal{P}}^{\gamma_2(1-\epsilon)}_{\alpha_2}(S)
\sum_{k=-N_1(u)-2}^{N_1(u)+1}\Psi(u_{k,\epsilon}^{1,-})\nonumber\\
& =&o\left(\pi_{1,\epsilon}^{(1)}(u)\right), \ \ u\rw\IF, S\rw\IF,
\EQN
with $\check{u}_{k,k_1,\epsilon}=\max(u_{k,\epsilon}^{1,-}, u_{k_1,\epsilon}^{1,-})$.
Combining (\ref{main1}), (\ref{PG1}), (\ref{Addition2}), (\ref{PG2}), (\ref{lambda11}) with (\ref{lambda21}) leads to
$$\pi_1(u) \sim
2\Gamma(1/\beta_1+1)\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}\frac{\overleftarrow{\vv}_1(u^{-1})}{\overleftarrow{\LL}_1(u^{-1})
}\Psi(u), \ \ u\rw \IF.$$
The proof is \HEH{completed} by inserting the above asymptotic into (\ref{F1}).\\
\underline{Case ii) $\gamma_1=0, \gamma_2=\IF$.} In this case (\ref{main1}), (\ref{PG1}), (\ref{Addition2}), (\ref{PG2}), (\ref{lambda11}) and
(\ref{lambda21}) still hold except for the fact that in light of Lemma {\ref{PIPI}}, we have to replace both
$\widehat{\mathcal{P}}^{\gamma_2(1-\epsilon)}_{\alpha_2}(S)$ and $\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}$ by
1 in (\ref{PG1}) and (\ref{Addition2}). \\%Thus the claim of case iii) is established.\\
\underline{Case iii) $\gamma_1, \gamma_2\in (0,\IF)$.} Let $\widehat{I}_{0,0}(u)=I_{0,0}(u)\cup I_{-1,0}(u)\cup I_{0,-1}(u)\cup I_{-1,-1}(u)$. It follows
straightforwardly that for any $0<\epsilon<1/2$ and $u$ large enough
\BQN\label{main2}
\pi_{1,\epsilon}^{(3)}(u)\leq \pi_1(u)\leq \pi_{1,-\epsilon}^{(3)}(u)+\pi_{1,-\epsilon}^{(4)}(u)
\EQN
with
$$
\pi_{1,\pm\epsilon}^{(3)}(u)=\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{0,0}(u)}\frac{\overline{X}(s,t)}{(1+(1\pm\epsilon)v^2_1(s))(1+(1\pm\epsilon)v^2_2(t))}>u\right),$$
and
$$\pi_{1,-\epsilon}^{(4)}(u)=\sum_{|k|=1, k\neq -1}^{N_1(u)+2}\sum_{|l|=1, l\neq -1}^{N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-\right).
$$
By Lemma {\ref{PIPI}}, it follows that
\BQN\label{PG3}
\pi_{1,\pm\epsilon}^{(3)}(u)\sim {\prod_{i=1}^2} \widehat{\mathcal{P}}^{(1\pm\epsilon)\gamma_i}_{\alpha_i}(S)\Psi(u), \ \ u\rw\IF, S\rw\IF.
\EQN
In addition,
using Lemma \ref{PIPI} and (\ref{A0}),
the same argument as given in the derivation of
the upper bound for $\pi_{1,-\epsilon}^{(2)}(u)$ yields
\BQN\label{PG4}
\pi_{1,-\epsilon}^{(4)}(u)
=o(\pi_{1, \pm\epsilon}^{(3)}(u))
\EQN
as $u\rw\IF$ and $S\rw\IF$.
Combination of (\ref{main2}) and (\ref{PG3}) with (\ref{PG4})
leads to $$\pi_1(u)\sim \prod_{i=1}^2 \widehat{\mathcal{P}}^{\gamma_i}_{\alpha_i}\Psi(u), \ \ u\rw\IF, S\rw\IF,$$
\eMM{hence the proof of this case is complete}.\\
\underline{Case iii) $\gamma_1\in (0,\IF), \gamma_2=\IF$.}
\KD{The proof follows the same lines} as given in previous case,
with \KD{$\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}$} replaced by 1.\\
\underline{Case iii) $\gamma_1, \gamma_2=\IF$.} Similarly, (\ref{main2}), (\ref{PG3}) and (\ref{PG4}) hold with
\KD{$\widehat{\mathcal{P}}^{\gamma_1}_{\alpha_1},\widehat{\mathcal{P}}^{\gamma_2}_{\alpha_2}$ replaced by 1.}
\QED
\prooftheo{th.T4}
Similarly as in (\ref{main})
\BQN\label{E0}
\pi_{1}^+(u)-\Lambda^{(1)}(u)\leq \pi_1(u)\leq \pi_{1}^-(u),
\EQN
\KD{with}
\BQNY
\pi_{1}^{\pm}(u):&=&\sum_{k=-N_1(u)\pm 2}^{N_1(u)\mp 1}\sum_{l=-N_2(u)\pm 2}^{N_2(u)\mp 1}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^{\pm}\right),\\
\Lambda^{(1)}(u):&=&\sum_{(k,l,k_1,l_1)\in \mathbb{V}_1(u)\cup \mathbb{V}_2(u) }\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}X(s,t)>u, \sup_{(s,t)\in
I_{k_1,l_1}(u)}X(s,t)>u\right).
\EQNY
Since \KD{$B$} is \HEH{non-singular matrix},
then there exists a positive constant $\mu>0$ such that for any $s, t$,
$$|b_{11}s+b_{12}t|+|b_{21}s+b_{22}t|\geq \mu \left(|s|+|t|\right).$$
Thus, for $(s,t)\in I_{k,l}(u)$ with $|k|\geq k_0\geq 2$ and $|l|\geq l_0\geq 2$
$$|b_{11}s+b_{12}t|+|b_{21}s+b_{22}t|\geq \mu S\left((k_0-1)\overleftarrow{\LL_1}(u^{-1})+(l_0-1)\overleftarrow{\LL_2}(u^{-1})\right).$$
\def\AS{\HEH{a(s,t)}}
By UCT,
for any $(s,t), (s',t')\in I_{k,l}(u)$ with $k_0\leq |k|\leq N_1(u)+2, l_0\leq |l|\leq N_2(u)+2$ and $u$ large enough \HEH{set $\AS:= v_1^2(|b_{11}s+b_{12}t|)+v_2^2(|b_{21}s+b_{22}t|)$)}
$$\frac{\AS }{v_1^2(|b_{11}s+b_{12}t|+|b_{21}s+b_{22}t|)}\geq
(1-\epsilon/3)\left(\nu^{\beta_1}+\theta(1-\nu)^{\beta_1}\right) $$
and
\BQNY\frac{ \NE{a(s',t') }
}{v_1^2(|b_{11}s+b_{12}t|+|b_{21}s+b_{22}t|)}\leq
\frac{(1-\epsilon/3)^2}{1-\epsilon}\left((\nu+\delta)^{\beta_1}+\theta(1-\nu+\delta)^{\beta_1}\right) ,
\EQNY
where
\BQNY \NE{\nu}=\frac{|b_{11}s+b_{12}t|}{|b_{11}s+b_{12}t|+|b_{21}s+b_{22}t|}\in [0,1],
\EQNY
and
$$0\leq \delta\leq
\frac{\left(|b_{11}|+|b_{12}|+|b_{21}|+|b_{22}|\right)\left(\overleftarrow{\LL_1}(u^{-1})+\overleftarrow{\LL_2}(u^{-1})\right)S}
{|b_{11}s+b_{12}t|+|b_{21}s+b_{22}t|}\leq
\frac{2\left(\sum_{i,j=1}^2|b_{ij}|\right)(1+\eta^{-1/\alpha_1})}{\mu\left((k_0-1)+(l_0-1)\eta^{-1/\alpha_1}\right)}\rw 0, $$
as $k_0, l_0\rw\IF.$
\KD{Using that for any $0<\epsilon<1/2$, when $k_0$ and $l_0$ are large enough
\BQNY
\left(\nu^{\beta_1}+\theta(1-\nu)^{\beta_1}\right)\geq (1-\epsilon/3)\left((\nu+\delta)^{\beta_1}+\theta(1-\nu+\delta)^{\beta_1}\right), \ \ \nu\in [0,1]
\EQNY
}
for any
$0<\epsilon<1/2$, there exists $k_\epsilon, l_\epsilon$ such that for any $k_\epsilon\leq |k|\leq N_1(u)+2, l_\epsilon\leq |l|\leq N_2(u)+2$ and $u$
large enough
\NE{$$a(s,t)\geq (1-\epsilon)a(s',t'), \quad (s,t), (s',t')\in I_{k,l}(u),$$}
which is equivalent to
\BQN\label{Basic1}
\inf_{(s,t)\in I_{k,l}(u)
\AS
\geq (1-\epsilon)\sup_{(s,t)\in
I_{k,l}(u)}\AS
.
\EQN
\underline{ Case i).}
Using (\ref{uniform}) and by (\ref{Basic1}), we have
\BQNY
\pi_1^{-}(u)
&\sim& \sum_{k=-N_1(u)- 2}^{N_1(u)+ 1}\sum_{l=-N_2(u)- 2}^{N_2(u)+
1}\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u_{k,l,\epsilon}^-)\nonumber\\
&\sim&\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u)\sum_{k=-N_1(u)-2}^{N_1(u)+ 1}\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}
e^{-(1-\epsilon)u^2\inf_{(s,t)\in I_{k,l}(u)} \AS}\\% \left(v_1^2(|b_{11}s+b_{12}t|)+v_2^2(|b_{21}s+b_{22}t|)\right)}\\
&\leq& \prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u)
\left(R_3(u)+R_4(u)+\right.\nonumber\\
&& \ +\left.\frac{1}{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})S^2}\sum_{|k|=k_\epsilon}^{N_1(u)+2}\sum_{|l|=l_\epsilon}^{N_2(u)+ 2}
\int_{(s,t)\in I_{k,l}(u)}e^{-(1-\epsilon)^2u^2 \AS
}dsdt\right),
\EQNY
where
$$R_3(u)=\sum_{|k|\leq k_\epsilon}\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}
e^{-(1-\epsilon)u^2\inf_{(s,t)\in I_{k,l}(u)} \AS
},$$
and
$$R_4(u)=\sum_{|l|\leq l_\epsilon}\sum_{k=-N_1(u)-2}^{N_1(u)+ 1}
e^{-(1-\epsilon)u^2\inf_{(s,t)\in I_{k,l}(u)}\AS
}.$$
\KD{By linear transformation
$(s',t')^{\top}=B(s,t)^\top$
and Lemma \ref{integ}, we have}
\BQNY
&&\frac{1}{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})}\sum_{|k|=k_\epsilon}^{N_1(u)+2}\sum_{|l|=l_\epsilon}^{N_2(u)+ 2}
\int_{(s,t)\in I_{k,l}(u)}e^{-(1-\epsilon)^2u^2 \A
}dsdt\\
&& \ \ \ \leq \frac{1}{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})}\int_{-2\overleftarrow{v_1}(u^{-1}\ln
u)}^{2\overleftarrow{v_1}(u^{-1}\ln u)}\int_{-2\overleftarrow{v_2}(u^{-1}\ln u)}^{2\overleftarrow{v_2}(u^{-1}\ln
u)}e^{-(1-\epsilon)^2u^2 \AS
}dsdt\\
&& \ \ \ \leq\frac{1}{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})}\frac{1}{\abs{ det(B)}}\int_{-\mathbb{Q}\overleftarrow{v_1}(u^{-1}\ln
u)}^{\mathbb{Q}\overleftarrow{v_1}(u^{-1}\ln u)}\int_{-\mathbb{Q}\overleftarrow{v_2}(u^{-1}\ln u)}^{\mathbb{Q}\overleftarrow{v_2}(u^{-1}\ln
u)}e^{-(1-\epsilon)^2u^2v_1^2(|s'|)}e^{-(1-\epsilon)^2u^2v_2^2(|t'|)}ds'dt'\\
&& \ \ \ =\frac{1}{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})}\frac{4}{\abs{ det(B)}}\int_{0}^{\mathbb{Q}\overleftarrow{v_1}(u^{-1}\ln
u)}\int_{0}^{\mathbb{Q}\overleftarrow{v_2}(u^{-1}\ln u)}e^{-(1-\epsilon)^2u^2v_1^2(|s'|)}e^{-(1-\epsilon)^2u^2v_2^2(|t'|)}ds'dt'\\
&& \ \ \ \sim (1-\epsilon)^{-1/\beta_1-1/\beta_2}\frac{4}{\abs{
det(B)}}\Gamma(1/\beta_1+1)\Gamma(1/\beta_2+1)\frac{\overleftarrow{\vv_1}(u^{-1})\overleftarrow{\vv_2}(u^{-1})}
{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})}\rw \IF, \ \ u\rw\IF.
\EQNY
Moreover, in light of Lemma \ref{simple1}, there exists a constant $\kappa_1>0$ such that
\BQN\label{eq5}
\kappa_1v^2_1(|s|)+\kappa_1v^2_2(|t|)\leq \AS
, \ \ s,t\in \mathbb{R}.
\EQN
Thus we have
\BQNY
R_3(u)&\leq& \sum_{|k|\leq k_\epsilon}\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}
e^{-(1-\epsilon)u^2\inf_{(s,t)\in I_{k,l}(u)}\kappa_1\left(v_1^2(|s|)+v_2^2(|t|)\right)}\\
&\leq& (2k_\epsilon+1)\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}
e^{-(1-\epsilon)u^2\inf_{t\in J_l(u)}\kappa_1v_2^2(|t|)}\\
&\leq& \mathbb{Q}\frac{\overleftarrow{\vv_2}(u^{-1})}
{\overleftarrow{\LL_2}(u^{-1})}=o\left(\frac{\overleftarrow{\vv_1}(u^{-1})\overleftarrow{\vv_2}(u^{-1})}
{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})}\right), \ \ u\rw\IF.
\EQNY
Similarly,
$$R_4(u)\leq \mathbb{Q}_1\frac{\overleftarrow{\vv_1}(u^{-1})}
{\overleftarrow{\LL_1}(u^{-1})}=o\left(\frac{\overleftarrow{\vv_1}(u^{-1})\overleftarrow{\vv_2}(u^{-1})}
{\overleftarrow{\LL_1}(u^{-1})\overleftarrow{\LL_2}(u^{-1})}\right), \ \ u\rw\IF.$$
Therefore,
\BQN\label{E1}
\pi_1^-(u)\leq \frac{4}{\abs{ det(B)}}\prod_{i=1}^2 \Bigl[\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\Bigr]\Psi(u)(1+o(1)), \ \ u\rw\IF, S\rw\IF,\epsilon\rw 0.
\EQN
\KD{In the same way we obtain that}
\BQN\label{E2}
\pi_1^{+}(u)\geq \frac{4}{\abs{ det(B)}}\prod_{i=1}^2 \Bigl[\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\Bigr]\Psi(u)(1+o(1)), \ \ u\rw\IF, S\rw\IF.
\EQN
\KD{Due to} (\ref{eq5}), letting
\BQN\label{Y}
Y(s,t)=\frac{\overline{X}(s,t)}{1+\frac{\kappa_1}{2}v_1^2(|s|)+\frac{\kappa_1}{2}v_2^2(|t|)}, \ \ (s,t)\in \mathbb{R}^2,
\EQN we have
\BQNY
\Lambda^{(1)}(u)&\leq &\sum_{(k,l,k_1,l_1)\in \mathbb{V}_1(u)\cup \mathbb{V}_2(u) }\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}Y(s,t)>u, \sup_{(s,t)\in
I_{k_1,l_1}(u)}Y(s,t)>u\right).
\EQNY
\KD{The same argument as given in the proof of Theorem \ref{th.T1} leads to}
\BQN\label{E3}
\Lambda^{(1)}(u)=o(\pi_1^-(u)), \ \ u\rw\IF, S\rw\IF.
\EQN
Inserting (\ref{E1})-(\ref{E3}) into (\ref{E0}) yields
$$\pi_1(u)\sim \frac{4}{\abs{ det(B)}}\prod_{i=1}^2 \Bigl[\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}
\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\Bigr]\Psi(u),$$
which together with (\ref{F1}) completes the proof.\\
\underline{Case ii) $\gamma_1,\gamma_2\in (0,\IF)$.} Using the same notation for $\widehat{I}_{0,0}(u)$ as that in the proof of \KD{Theorem \ref{th.T1}} for case iii)
$\gamma_1, \gamma_2\in (0,\IF)$, (\ref{main2}) holds with
$$
\pi_{1,\pm\epsilon}^{(3)}(u)=\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{0,0}(u)}\frac{\overline{X}(s,t)}{1+(1\pm\epsilon) \AS
}>u\right),$$
and
$$\pi_{1,-\epsilon}^{(4)}(u)=\sum_{|k|=1, k\neq -1}^{N_1(u)+2}\sum_{|l|=1, l\neq -1}^{N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}X(s,t)>u\right).
$$
Noting that
\BQNY
&&u^2\left(v^2_1(|b_{11}\overleftarrow{\LL}_1(1/u)s+b_{12}\overleftarrow{\LL}_2(1/u)t|)
+v^2_2(|b_{21}\overleftarrow{\LL}_1(1/u)s+b_{22}\overleftarrow{\LL}_2(1/u)t|)\right)\\ \ \ &&\rw
\gamma_1|b_{11}s+b_{12}\eta^{-1/\alpha_1}t|^{\alpha_1}+\theta\gamma_1|b_{21}s+b_{22}\eta^{-1/\alpha_1}t|^{\alpha_1}, \ \ u\rw\IF
\EQNY
uniformly with respect to $s,t\in [-S,S]^2$,
it follows from Lemma \ref{PIPI} that
\BQN
\pi_{1,\pm\epsilon}^{(3)}(u)\sim\widehat{\mathcal{P}}_\alpha^{(1\pm \epsilon)\gamma_1, (1\pm\epsilon)\theta\gamma_1, B_{\eta,\alpha}}(S)\Psi(u)\sim
\widehat{\mathcal{P}}_\alpha^{\gamma_1,\theta\gamma_1, B_{\eta,\alpha}}\Psi(u), \ \ u\rw\IF, S\rw\IF, \epsilon\rw 0.
\EQN
Moreover, by Lemma \ref{simple1} and (\ref{PG4}), with $Y$ defined by (\ref{Y}),
\BQNY
\pi_{1,-\epsilon}^{(4)}(u)\leq \sum_{|k|=1, k\neq -1}^{N_1(u)+2}\sum_{|l|=1, l\neq -1}^{N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}Y(s,t)>u\right)\KD{=o\left(\Psi(u)\right)},
\EQNY
\KD{as $u\to\infty, S\rw\IF$}.
Thus
$\pi_1(u)\sim\widehat{\mathcal{P}}_\alpha^{\gamma_1,\theta\gamma_1, B_{\eta,\alpha}}\Psi(u),$
which completes the proof.\\
\underline{Case ii) $\gamma_1=\gamma_2=\IF$.}
\KD{The proof follows by the same argument as the proof of Case ii) $\gamma_1,\gamma_2\in (0,\IF)$,
with $\widehat{\mathcal{P}}_\alpha^{\gamma_1,\theta\gamma_1, B_{\eta,\alpha}}$
replaced by 1}.
\QED
\prooftheo{th.T2}
\KD{This scenario requires a modification of set $D_u$.}
Let
$D_u^{(1)}=\{(s,t), |s+b_{12}t|\leq \overleftarrow{\vv}_1(u^{-1}\ln u), |t|\leq \overleftarrow{\vv}_2(u^{-1}\ln u) \}$.
\eMM{It follows that (\ref{F1}})-(\ref{neww}) also hold with $D_u$ replaced by
$D_u^{(1)}$. In this scenario, denote $\alpha=\alpha_1=\alpha_2$.\\%Next we focus on $\pi_1(u)$. \\
\underline{Case i) $\gamma_1=\gamma_2=0$.}
\KD{Let }
$$E_{l}^+(u)=\{k: I_{k,l}(u)\subset D_u^{(1)}\}, E_{l}^-(u)=\{k:
I_{k,l}(u)\cap D_u^{(1)}\neq \emptyset\},$$ $$E^{(1)}(u)=\{(k,l, k_1,l_1), k\leq k_1, I_{k,l}(u)\cap D_u^{(1)}\neq \emptyset, I_{k_1,l_1}(u)\cap D_u^{(1)}\neq
\emptyset \ \ \text{and} \ \ I_{k,l}(u)\cap I_{k',l'}(u)= \emptyset\},$$
$$E^{(2)}(u)=\{(k,l, k_1,l_1), k\leq k_1, I_{k,l}(u)\cap D_u^{(1)}\neq \emptyset, I_{k_1,l_1}(u)\cap D_u^{(1)}\neq \emptyset, (k,l)\neq (k_1,l_1)\ \ \text{and} \ \ I_{k,l}(u)\cap
I_{k',l'}(u) \neq\emptyset\}.$$
It follows that
\BQN\label{E4}
\pi_{2}^+(u)-\sum_{i=1}^2\Lambda_i^{(2)}(u)\leq \pi_1(u)\leq \pi_{2}^-(u),
\EQN
where
\BQNY
\pi_{2}^{\pm}(u):&=&\sum_{l=-N_2(u)\pm 2}^{N_2(u)\mp 1}\sum_{k\in E^{\pm}_{l}(u)}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^{\pm}\right),\\
\Lambda_i^{(2)}(u):&=&\sum_{(k,l,k_1,l_1)\in E^{(i)}(u)}\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^- , \sup_{(s,t)\in
I_{k_1,l_1}(u)}\overline{X}(s,t)>u_{k_1,l_1,\epsilon}^-\right).
\EQNY
Using (\ref{uniform}), we have
\BQNY
\pi_2^-(u)
&=&\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}\sum_{k\in E^{-}_{0}(u)}\mathbb{P}\left(\sup_{(s,t)\in I_{0,0}(u)}\overline{X}(\overleftarrow{\LL_1}(u^{-1})kS+s,\overleftarrow{\LL_2}(u^{-1})lS+t)>u_{k,l,\epsilon}^-\right)\nonumber\\
&\sim& \sum_{l=-N_2(u)-2}^{N_2(u)+ 1}\sum_{k\in E^{-}_{l}(u)}\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u_{k,l,\epsilon}^-)\nonumber\\
&\sim&\prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u)\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}\sum_{k\in E^{-}_{l}(u)}
e^{-(1-\epsilon)u^2\inf_{(s,t)\in I_{k,l}(u)}\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}
\EQNY
We observe that, for $u$ sufficiently large and all $|l|\leq N_2(u)+2$,
$$E_{l}^{-}\subset \Biggl\{k\in \mathbb{N}, |k-\left[b_{12}l\frac{\overleftarrow{\LL}_2(1/u)}{\overleftarrow{\LL}_1(1/u)}\right]|\leq
N_1(u)+2(2+\left[|b_{12}|\eta^{-1/\alpha}\right]) \Biggr\},$$
and
$$E_l^+\supset\Biggl \{k\in \mathbb{N}, |k-\left[b_{12}l\frac{\overleftarrow{\LL}_2(1/u)}{\overleftarrow{\LL}_1(1/u)}\right]|\leq
N_1(u)-2(2+\left[|b_{12}|\eta^{-1/\alpha}\right])\Biggr\}.$$
By UCT, we have \KD{that} for any $\epsilon>0$ there exists $l_\epsilon>0$ such that
\BQNY
\inf _{t\in J_l(u)}v_2^2(|t|)\geq (1-\epsilon) \sup_{t\in J_l(u)}v_2^2(|t|)
\EQNY
holds for $l_\epsilon\leq |l|\leq N_2(u)+2$.
Moreover, for any $\epsilon>0$ there exists $k_\epsilon>0$ such that
\BQNY
\inf _{t\in I_{k,l}(u)}v_1^2(|s+b_{12}t|)\geq (1-\epsilon) \sup_{s\in I_k(u)}v_1^2(|s+b_{12}l\overleftarrow{\LL}_2(u^{-1})S|)
\EQNY
hold for $|l|\leq N_2(u)+2$ and
$$k\in E_{l,\epsilon}^-(u):= \Bigg\{k, k_\epsilon\leq
|k-\left[b_{12}l\frac{\overleftarrow{\LL}_2(1/u)}{\overleftarrow{\LL}_1(1/u)}\right]|\leq N_1(u)+2(1+\left[|b_{12}|\eta^{-1/\alpha}\right]) \Bigg\}.$$ Therefore, in light of Lemma \ref{integ}, we have
\BQNY
&&\sum_{l=-N_2(u)-2}^{N_2(u)+ 1}\sum_{k\in E^{-}_{l}(u)}
e^{-(1-\epsilon)u^2\inf_{(s,t)\in I_{k,l}(u)}\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}\\
&& \ \ \ \leq \sum_{l=-N_2(u)-2}^{N_2(u)+ 1}e^{-(1-\epsilon)u^2\inf_{t\in J_l(u)}v_2^2(|t|)}\sum_{k\in E^{-}_{l}(u)}
e^{-(1-\epsilon)u^2\inf_{(s,t)\in I_{k,l}(u)}v_1^2(|s+b_{12}t|)}\\
&& \ \ \ \leq \frac{1}{\overleftarrow{\LL}_2(u^{-1})S}\sum_{|l|\geq l_\epsilon}^{N_2(u)+ 2}\int_{t\in
J_l(u)}e^{-(1-\epsilon)^2u^2v_2^2(|t|)}dt \\
&& \ \ \times \left(2k_\epsilon+1+\frac{1}{\overleftarrow{\LL}_1(1/u)S} \sum_{k\in E_{l,\epsilon}(u)}
\int_{s\in I_k(u)}e^{-(1-\epsilon)^2u^2v_1^2(|s+b_{12}l\overleftarrow{\LL}_2(1/u)S|)}ds\right)\\
&&\ \ + \sum_{|l|=0}^{l_\epsilon}\Bigg(2k_\epsilon+1+\frac{1}{\overleftarrow{\LL}_1(1/u)S}\sum_{k\in E_{l,\epsilon}(u)}
\int_{s\in I_k(u)}e^{-(1-\epsilon)^2u^2v_1^2(|s+b_{12}l\overleftarrow{\LL}_2(1/u)S|)}ds\Bigg)\\
&& \ \ \ \leq \frac{2(1+o(1))}{\overleftarrow{\LL}_2(u^{-1})S} \int_{0}^{\mathbb{Q}\overleftarrow{v_2}(u^{-1}\ln
u)}e^{-(1-\epsilon)^2u^2v_2^2(|t|)}dt\Bigg (2k_\epsilon+1+\frac{2}{\overleftarrow{\LL}_1(1/u)S}\int_{0}^{\mathbb{Q}\overleftarrow{v_1}(u^{-1}\ln u)}
e^{-(1-\epsilon)^2u^2v_1^2(|s|)}ds\Bigg)\\
&& \ \ \ \sim (1-\epsilon)^{-1/\beta_1-1/\beta_2}\frac{4}{S^2}\prod_{i=1}^2
\Gamma(1/\beta_i+1)\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}, \ \ u\rw\IF.
\EQNY
Consequently,
\KD{
\BQN\label{E5}
\pi_2^-(u)&\leq&
4\prod_{i=1}^2\left[\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\right]
\Psi(u)(1+o(1)), \ \ u\rw\IF, S\rw\IF, \epsilon\rw 0.
\EQN
Let $E_{l,\epsilon}^+(u):=\{k, k_\epsilon\leq |k-\left[bl\frac{\overleftarrow{\LL}_2(1/u)}{\overleftarrow{\LL}_1(1/u)}\right]|\leq
N_1(u)-2(1+\left[|b|\eta^{-1/\alpha}\right])\}$.
}
Similarly,
\BQN\label{E6}
\pi_2^+(u)&\geq& \prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u)\sum_{l=-N_2(u)+2}^{N_2(u)-1}\sum_{k\in E^{+}_{l}(u)}
e^{-(1-\epsilon)u^2\sup_{(s,t)\in I_{k,l}(u)}\left(v_1^2(|s+bt|)+v_2^2(|t|)\right)}\nonumber\\
&\geq& \prod_{i=1}^2\mathcal{H}_{\alpha_i}[0,S]\Psi(u)\sum_{|l|=k_\epsilon}^{N_2(u)-2}e^{-(1-\epsilon)u^2\sup_{t\in
J_l(u)}v_2^2(|t|)}\sum_{k\in E^{+}_{l,\epsilon}(u)}
e^{-(1-\epsilon)u^2\sup_{(s,t)\in I_{k,l}(u)}v_1^2(|s+bt|)}\nonumber\\
&\sim& 4\prod_{i=1}^2\left[\Gamma(1/\beta_i+1)\mathcal{H}_{\alpha_i}\frac{\overleftarrow{\vv}_i(1/u)}{\overleftarrow{\LL}_i(1/u)}\right]
\Psi(u), \ \ u\rw\IF, S\rw\IF, \epsilon\rw 0.
\EQN
\KD{
Following the same argumentation as given
in
(\ref{lambda1}) and (\ref{lambda2}),
we get that
$\Lambda^{(2)}_i(u)
=o\left(\pi_2^-(u)\right), \ \ i=1,2, u\rw\IF,$
which together with (\ref{E5}) and (\ref{E6}) completes the proof.}\\
\underline{Case ii) $\gamma_2=0, \gamma_1\in (0,\IF)$.}
We first
\KD{introduce}
\begin{eqnarray*}
L_{0,l}^*(u)&=&\{(s,t), |s+b_{12}t|\leq \overleftarrow{\LL}_1(1/u)S, t\in [l\overleftarrow{\LL}_2(1/u)S, (l+1)\overleftarrow{\LL}_2(1/u)S]\},\\
L_{k,l}(u)&=&\{(s,t), k\overleftarrow{\LL}_1(1/u)S\leq s+b_{12}t\leq (k+1)\overleftarrow{\LL}_1(1/u)S, t\in [l\overleftarrow{\LL}_2(1/u)S,
(l+1)\overleftarrow{\LL}_2(1/u)S]\},
\end{eqnarray*}
$$u_{k,l,\epsilon,*}^-=u\left(1+(1-\epsilon)\inf_{(s,t)\in L_{k,l}(u)}\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)\right)$$
with $k,l \in \mathbb{Z}$.
Then we have
\BQN
\pi_3^+(u)+\sum_{i=1}^2\Lambda_i^{(3)}(u)\leq \pi_1(u)\leq \pi_3^-(u)+\pi_4(u),
\EQN
where
\BQNY
\pi_3^{\pm}(u)&:=&\sum_{l=-N_2(u)\pm 2}^{N_2(u)\mp 1}\mathbb{P}\left(\sup_{(s,t)\in L_{0,l}^*(u)}\frac{\overline{X}(s,t)}{1+(1\pm
\epsilon)v_1^2(|s+b_{12}t|)}>u_{l,\epsilon}^{2,\pm}\right),\\
\pi_4(u)&:=& \sum_{|k|\leq N_1(u)+2, k\neq 0, -1} \sum_{l=-N_2(u)\pm 2}^{N_2(u)\mp 1}\mathbb{P}\left(\sup_{(s,t)\in
L_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon,*}^-\right),\\
\Lambda_1^{(3)}(u)&:=&\sum_{-N_2(u)-2\leq l+1<l_1\leq N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in L_{0,l}(u)}\overline{X}(s,t)>u_{l,\epsilon}^{2,-}, \sup_{(s,t)\in
L_{0,l_1}(u)}\overline{X}(s,t)>u_{l_1,\epsilon}^{2,-}\right)\\
\Lambda_2^{(3)}(u)&:=&\sum_{l=-N_2(u)-2}^{N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in L_{0,l}(u)}\frac{\overline{X}(s,t)}{1+(1-
\epsilon)v_1^2(|s+b_{12}t|)}>u_{l,\epsilon}^{2,-}, \sup_{(s,t)\in L_{0,l+1}(u)}\frac{\overline{X}(s,t)}{1+(1- \epsilon)v_1^2(|s+b_{12}t|)}>u_{l+1,\epsilon}^{2,-}\right).
\EQNY
Let
$$X_l(s,t)=\overline{X}(-b_{12}l\overleftarrow{\rho}_2(u^{-1})S+s,l\overleftarrow{\rho}_2(u^{-1})S+t),\quad \mathcal{K}_u=\{l, |l|\leq N_2(u)+2\}, \quad \mathcal{E}_u=L_{0,0}^*(u), $$ $$ h_l(u)=u_{l,\epsilon}^{2,-}, \quad d_u(s,t)=(1-\epsilon)v_1^2(|s+b_{12}t|)$$
Since
$$\lim_{u\rw\IF}\sup_{l\in \mathcal{K}_u}\left|(u_{l,\epsilon}^{2,-})^2v_1^2(|\overleftarrow{\rho}_1(1/u)s+b_{12}\overleftarrow{\rho}_2(1/u)t|)-
\gamma_1|s+b_{12}\eta^{-1/\alpha}t|\right|=0, $$
uniformly over any compact set,
by Lemma \ref{PIPI}, we have
$$\lim_{u\rw\IF}\sup_{l\in\mathcal{K}_u}\left|(\Psi(u_{l,\epsilon}^{2,-}))^{-1}\mathbb{P}\left(\sup_{(s,t)\in L_{0,0}^*(u)}\frac{\overline{X}(-b_{12}l\overleftarrow{\LL}_2(1/u)S+s,l\overleftarrow{\LL}_2(1/u)S+t)}{1+(1-
\epsilon)v_1^2(|s+b_{12}t|)}>u_{l,\epsilon}^{2,-}\right)-\mathcal{H}_\alpha^{(1-\epsilon)\gamma_1, b_{12}\eta^{-1/\alpha}}(S)\right|=0.$$
Thus we have,
\BQN\label{eq2}
\pi_3^{-}(u)
&=& \sum_{l=-N_2(u)- 2}^{N_2(u)+ 1}\mathbb{P}\left(\sup_{(s,t)\in L_{0,0}^*(u)}\frac{\overline{X}(-b_{12}l\overleftarrow{\LL}_2(1/u)S+s,l\overleftarrow{\LL}_2(1/u)S+t)}{1+(1-
\epsilon)v_1^2(|s+b_{12}t|)}>u_{l,\epsilon}^{2,-}\right)\nonumber\\
&\sim& \sum_{l=-N_2(u)- 2}^{N_2(u)+ 1}\Psi(u_{l,\epsilon}^{2,-})\mathcal{H}_\alpha^{(1-\epsilon)\gamma_1, b_{12}\eta^{-1/\alpha}}(S)\nonumber\\
&\sim&\mathcal{H}_\alpha^{(1-\epsilon)\gamma_1, b_{12}\eta^{-1/\alpha}}(S) \Psi(u)\sum_{l=-N_2(u)- 2}^{N_2(u)+
1}e^{-(1-\epsilon)^2u^2\inf_{t\in J_l(u)}v_2^2(|t|)}\nonumber\\
&\sim& \frac{\mathcal{H}_\alpha^{\gamma_1, b_{12}\eta^{-1/\alpha}}(S)}{S}2\Gamma(1/\beta_2+1)\Psi(u)\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}(1+o(1)), \
\ u\rw\IF, \upsilon, \epsilon\rw 0.
\EQN
Moreover, in light of \cite{nonhomoANN},
$$\lim_{S\rw\IF} \frac{\mathcal{H}_\alpha^{\gamma_1, b_{12}\eta^{-1/\alpha}}(S)}{S}=\mathcal{H}_\alpha^{\gamma_1, b_{12}\eta^{-1/\alpha}}\in (0,\IF).$$
Thus we have
\BQN\label{E7}
\pi_3^{-}(u)\leq \mathcal{H}_\alpha^{\gamma_1, b_{12}\eta^{-1/\alpha}}2\Gamma(1/\beta_2+1)\Psi(u)\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}(1+o(1)), \ \ u\rw\IF, S\rw\IF, \eMM{\epsilon\rw 0} .
\EQN
Similarly,
\BQN\label{E12}
\pi_3^{+}(u)\geq \mathcal{H}_\alpha^{\gamma_1, b_{12}\eta^{-1/\alpha}}2\Gamma(1/\beta_2+1)\Psi(u)\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}(1+o(1)), \ \ u\rw\IF, S\rw\IF, \eMM{\epsilon\rw 0}.
\EQN
Note that for $u$ sufficiently large
\BQN\label{eq6}
L_{0,0}(u)\subset [-(1+2|b_{12}|\eta^{-1/\alpha})\overleftarrow{\rho}_1(1/u)S, (1+2|b_{12}|\eta^{-1/\alpha})\overleftarrow{\rho}_1(1/u)S]\times [0,
\overleftarrow{\rho}_2(1/u)S]=:J_{0,0}(u).
\EQN
Thus,
with $S_2=(1+2|b|\eta^{-1/\alpha})S$, by (\ref{uniform}) with $u_{k,l,\epsilon,*}^-$ instead of $u_{k,l,\epsilon}^-$ , we obtain
\BQN\label{E8}
\pi_4(u)&=&\sum_{|k|\leq N_1(u)+2, k\neq 0, -1} \sum_{l=-N_2(u)- 2}^{N_2(u)+ 1}\mathbb{P}\left(\sup_{(s,t)\in
L_{0,0}(u)}\overline{X}(k\overleftarrow{\LL}_1(1/u)S-b_{12}l\overleftarrow{\LL}_2(1/u)S+s,l\overleftarrow{\LL}_2(1/u)S+t)>u_{k,l,\epsilon,*}^-\right)\nonumber\\
&\leq& \sum_{|k|\leq N_1(u)+2, k\neq 0, -1} \sum_{l=-N_2(u)- 2}^{N_2(u)+ 1}\mathbb{P}\left(\sup_{(s,t)\in
J_{0,0}(u)}\overline{X}(k\overleftarrow{\LL}_1(1/u)S-b_{12}l\overleftarrow{\LL}_2(1/u)S+s,l\overleftarrow{\LL}_2(1/u)S+t)>u_{k,l,\epsilon,*}^-\right)\nonumber\\
&\sim& \sum_{|k|\leq N_1(u)+2, k\neq 0, -1} \sum_{l=-N_2(u)- 2}^{N_2(u)+
1}\widehat{\mathcal{H}}_\alpha(S_2)\mathcal{H}_{\alpha}(S)\Psi(u)e^{-(1-\epsilon)u^2\inf_{(s,t)\in L_{k,l}(u)}\left(v_1^2(|s+bt|)+v_2^2(|t|)\right)}\nonumber\\
&\leq& \widehat{\mathcal{H}}_\alpha(S_2)\mathcal{H}_{\alpha}(S)\Psi(u)\sum_{1\leq|k|\leq
N_1(u)+2}e^{-\mathbb{Q}u^2v_1^2(\overleftarrow{\rho}_1(1/u)|k|S)} \sum_{l=-N_2(u)- 2}^{N_2(u)+ 1}e^{-(1-\epsilon)u^2\inf_{t\in J_l(u)}v_2^2(|t|)}\nonumber\\
&\leq& 2\Gamma(1/\beta_2+1)(1-\epsilon)^{-2/\beta_2}\frac{\widehat{\mathcal{H}}_\alpha(S_2)}{S}\frac{\mathcal{H}_{\alpha}(S)}{S}
\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}\Psi(u)S\sum_{1\leq|k|\leq N_1(u)+2}e^{-\mathbb{Q}_1|kS|^{\beta_1/2}}\nonumber\\
&\leq&\mathbb{Q}_2
\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}\Psi(u)e^{-\mathbb{Q}_3S^{\beta_1/2}}=o\left(\pi_3^-(u)\right), \ \ u\rw\IF, S\rw\IF.
\EQN
\KD{
Following the same idea as given in
the proof of Theorem \ref{th.T1}, we get that
}
$
\Lambda_1^{(3)}(u)+\Lambda_2^{(3)}(u)
=
o\left(\pi_3^-(u)\right)
$, as $\to\infty$,
which completes the proof of this case.
\\
\underline{Case ii) $\gamma_2=0, \gamma_1=\IF$.} It follows straightforwardly that, for any $x>0$ and $u$ sufficiently large,
\BQN
\mathbb{P}\left(\sup_{|t|\leq \overleftarrow{\vv}_2(u^{-1}\ln u)}X(-b_{12}t,t)>u\right)\leq \pi_1(u)\leq \mathbb{P}\left(\sup_{(s,t)\in
D_u^{(1)}}\frac{\overline{X}(s,t)}{1+x\rho_1^2(|s+b_{12}t|)+v_2^2(|t|)}>u\right).
\EQN
Using that the Gaussian random field on the right hand side of the above
satisfies case $\gamma_2=0, \gamma_1=x\in (0,\IF)$, by (\ref{eq2}) and (\ref{E8}) we get for $S$ sufficiently large
\BQNY
\mathbb{P}\left(\sup_{(s,t)\in D_u^{(1)}}\frac{\overline{X}(s,t)}
{1+x\rho_1^2(|s+b_{12}t|)+v_2^2(|t|)}>u\right)
&\leq&
\frac{\mathcal{H}_\alpha^{x,b_{12}\eta^{-1/\alpha}}(S)2\Gamma(1/\beta_2+1)}{S}\Psi(u)\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}(1+o(1)).
\EQNY
It follows that for any $S$ positive
\BQNY\lim_{x\rw\IF}\mathcal{H}_\alpha^{x, b_{12}\eta^{-1/\alpha}}(S)&=&\lim_{x\rw\IF}\EE{\sup_{(s+ b_{12}\eta^{-1/\alpha}t,t)\in [-S,S]\times[0,S]} e^{ W(s,t) -x|s+ b_{12}\eta^{-1/\alpha}t|^{\alpha}
}}\\
&=&\EE{\sup_{(s+ b_{12}\eta^{-1/\alpha}t,t)\in \{0\}\times[0,S]} e^{ W(s,t)
}}\\&=&
\mathcal{H}_\alpha(\left(|b_{12}|^{\alpha}\eta^{-1}+1\right)^{1/\alpha}S).
\EQNY
Hence, as $u\rw\IF, x\rw\IF, S\rw\IF$
$$\mathbb{P}\left(\sup_{(s,t)\in D_u^{(1)}}
\frac{\overline{X}(s,t)}{1+x\rho_1^2(|s+b_{12}t|)+v_2^2(|t|)}>u\right)\leq
2\left(|b_{12}|^{\alpha}\eta^{-1}+1\right)^{1/\alpha}\mathcal{H}_\alpha\Gamma(1/\beta_2+1)\Psi(u)\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}(1+o(1)).$$
Further, for the process $X(-b_{12}t,t)$, we have
\BQN\label{E9}
1-\sqrt{Var\left(X(-b_{12}t,t)\right)}&\sim& v_2^2(|t|), \ \ t\rw 0,\nonumber\\
1-Corr\left(X(-b_{12}t,t), X(-b_{12}s,s)\right)&\sim& \left(|b_{12}|^{\alpha}\eta^{-1}+1\right)\rho_2^2(|t-s|), \ \ s,t\rw 0.
\EQN
Thus in light of Theorem \ref{mainT}, we have
$$
\KD{\mathbb{P}\left(\sup_{|t|\leq \overleftarrow{\vv}_2(u^{-1}\ln u)}X(-b_{12}t,t)>u\right)}
\sim
2\left(|b_{12}|^{\alpha}\eta^{-1}+1\right)^{1/\alpha}\mathcal{H}_\alpha\Gamma(1/\beta_2+1)\Psi(u)\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)}.$$
Consequently,
$$\pi_1(u)\sim
2\left(|b_{12}|^{\alpha}\eta^{-1}+1\right)^{1/\alpha}\mathcal{H}_\alpha\Gamma(1/\beta_2+1)\Psi(u)\frac{\overleftarrow{v}_2(1/u)}{\overleftarrow{\rho}_2(1/u)},$$
which completes the proof.\\
\underline{Case iii) $\gamma_2\in (0,\IF), \gamma_1=\IF$.}
Let
$\widehat{I}_{0,0}^*(u)=\{(s,t), |s+b_{12}t|\leq \overleftarrow{\LL}_1(1/u)S, |t|\leq
\overleftarrow{\LL}_2(1/u)S\}.$
Then for $u$ sufficiently large, we have
\BQN\label{E10}
\mathbb{P}\left(\sup_{(s,t)\in D_u^{(1)}}X(-b_{12}t,t)>u\right)\leq \pi_1(u)\leq \pi_{1,-\epsilon}^{(5)}(u)+\pi_{1,-\epsilon}^{(6)}(u)
\EQN
with
$$
\pi_{1,-\epsilon}^{(5)}(u)=\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{0,0}^*(u)}\frac{\overline{X}(s,t)}{1+x\rho^2_1(|s+b_{12}t|)+(1-\epsilon)v^2_2(|t|)}>u\right),$$
and
$$\pi_{1,-\epsilon}^{(6)}(u)=\sum_{|k|=1, k\neq -1}^{N_1(u)+2}\sum_{|l|=1, l\neq -1}^{N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in
L_{k,l}(u)}\overline{X}(s,t)>u_{k,l,\epsilon}^-\right).
$$
Since
$$u^2(x\rho^2_1(|\overleftarrow{\rho}_1(1/u)s+b_{12}\overleftarrow{\rho}_2(1/u)t|)+(1-\epsilon)v^2_2(|\overleftarrow{\rho}_2(1/u)t|))\rw x|s+b_{12}\eta^{-1/\alpha}t|^{\alpha}+(1-\epsilon)\gamma_2|t|^{\alpha}, \ \ u\rw\IF$$
uniformly on any compact set,
then, by Lemma \ref{PIPI},
$$\pi_{1,-\epsilon}^{(5)}(u)\sim \widehat{\mathcal{H}}_{\alpha}^{x,\gamma_2,b_{12}\eta^{-1/\alpha}}(S)\Psi(u), \ \ u\rw\IF, \epsilon\rw 0.
$$
\KD{Moreover, by the same argument as given in case ii), we have
$$\lim_{x\rw\IF}\widehat{\mathcal{H}}_{\alpha}^{x,\gamma_2,b_{12}\eta^{-1/\alpha}}(S)=\widehat{\mathcal{P}}_\alpha^
{\gamma_2\left(|b_{12}|^\alpha\eta^{-1}+1\right)^{-1}}\left(\left(|b|^\alpha\eta^{-1}+1\right)^{1/\alpha}S\right).$$
Then}
\BQNY
\pi_{1,-\epsilon}^{(5)}(u)\sim \widehat{\mathcal{P}}_\alpha^
{\gamma_2\left(|b_{12}|^\alpha\eta^{-1}+1\right)^{-1}}\Psi(u), \ \ u\rw\IF, x\rw\IF, \epsilon\rw 0, S\rw\IF.
\EQNY
\KD{Using that
$L_{0,0}(u)\subset J_{0,0}(u),$ with $J_{0,0}(u)$ defined by (\ref{eq6}),
and following the same steps as in (\ref{E8}), we get}
\BQNY
\pi_{1,-\epsilon}^{(6)}(u)
=o(\Psi(u)), \ \
u\rw\IF, S\eMM{\rw \IF}.
\EQNY
Hence, from Theorem \ref{mainT} and (\ref{E9})
$$\mathbb{P}\left(\sup_{(s,t)\in D_u^{(1)}}X(-b_{12}t,t)>u\right)\sim \widehat{\mathcal{P}}_\alpha^
{\gamma_2\left(|b_{12}|^\alpha\eta^{-1}+1\right)^{-1}}\Psi(u), \ \ u\rw\IF,$$
which establishes the claim.\\
\underline{Case iv) $\gamma_2=\IF, \gamma_1=\IF$.}
Clearly, (\ref{E10}) holds with
$$
\pi_{1,-\epsilon}^{(5)}(u):=\mathbb{P}\left(\sup_{(s,t)\in \widehat{I}_{0,0}(u)}
\frac{\overline{X}(s,t)}{1+x\rho^2_1(|s+b_{12}t|)+y\rho^2_2(|t|)}>u\right), x,
y>0.$$
Moreover,
$$\pi_{1,-\epsilon}^{(5)}(u)\sim \widehat{\mathcal{H}}_{\alpha}^{x,y,b_{12}\eta^{-1/\alpha}}(S)\Psi(u)$$
and
$$\lim_{y\rw\IF}\lim_{x\rw\IF}\widehat{\mathcal{H}}_{\alpha}^{x,y,b_{12}\eta^{-1/\alpha}}(S)=\lim_{y\rw\IF}\widehat{\mathcal{P}}_\alpha^
{y\left(|b_{12}|^\alpha\eta^{-1}+1\right)^{-1}}\left(\left(|b_{12}|^\alpha\eta^{-1}+1\right)^{1/\alpha}S\right)=1.$$
Hence
$\pi_{1,-\epsilon}^{(5)}(u)\sim \Psi(u), u\rw\IF, x\rw\IF, y\rw\IF.$
\KD{The rest of the proof is the same as the case
$\gamma_2\in (0,\IF), \gamma_1=\IF$.}
\QED
\prooftheo{th.T3}.
We focus on $\pi_1(u)$ as $u\rw\IF$. \\
\underline{ Case i)}
\KD{
The proof of this case follows line by line the same arguments as given in the proof
of Case i) of Theorem \ref{th.T6}.
}
\\
\underline{Case ii) $\gamma_1=0, \gamma_2\in (0,\IF)$.}
\KD{
First we introduce some new notation.}
Let
$$u_{k,\epsilon}^{{* -}}=1+(1 {-} 3\epsilon)\inf_{t\in I_k(u)}\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right),$$
and
$$\widehat{I}_{k,0}(u)=I_k(u)\times \left(J_{-1}(u)\cup J_0(u)\right), \ \
v(s,t)=v_1^2(|s+b_{12}t|)+v_2^2(|t|)-v_1^2(|(1+b_{12}\mu)s|)-v_2^2(|\mu s|), \ \
(s,t)\in D_u,
$$
where $\mu$ is defined right before Theorem \ref{th.T3}.
For any $0<x<y<\frac{S}{2|b_{12}|}$ and $0<\epsilon<1/4$, we have
\BQN\label{EQ3}
\pi_5^+(u)+\Lambda(u)\leq\pi_1(u)\leq \pi_5^-(u)+\pi_6(u)+\pi_7(u)+\pi_8(u),
\EQN
where
\BQNY
\pi_5^\pm(u):&=&\sum_{k\in E_{x,y}^\pm(u)}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1\pm\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right),\\
\pi_6(u):&=&\sum_{k\in E_{0,x}(u)}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right),\\
\pi_7(u):&=&\sum_{k\in E_{y,\IF}(u)}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right),\\
\pi_8(u):&=&\sum_{|k|=0}^{N_1(u)+2}\sum_{|l|=1, l\neq -1}^{N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in
I_{k,l}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right),\\
\Lambda(u):&=&\sum_{k<k_1\in E_{x,y}^-(u)}\mathbb{P}\left(\sup_{(s,t)\in \widehat{I}_{k,0}(u)}X(s,t)>u, \sup_{(s,t)\in
\widehat{I}_{k_1,0}(u)}X(s,t)>u\right),
\EQNY
with
\begin{eqnarray*}
E_{0,x}
&=&
\{k, |k|\leq N_1(u)+2, I_{k}(u)\cap [-\overleftarrow{\LL_2}(u^{-1})x, \overleftarrow{\LL_2}(u^{-1})x]\neq \KD{\emptyset}\},\\
E_{x,y}^-
&=&
\{k, |k|\leq N_1(u)+2, I_{k}(u)\cap \left([-\overleftarrow{\LL_2}(u^{-1})y, -\overleftarrow{\LL_2}(u^{-1})x
]\cup[\overleftarrow{\LL_1}(u^{-1})x, \overleftarrow{\LL_1}(u^{-1})y ]\right)\neq \KD{\emptyset}\}, \\
E_{x,y}^+
&=&
\{k, |k|\leq N_1(u)+2, I_{k}(u)\subset \left([-\overleftarrow{\LL_2}(u^{-1})y, -\overleftarrow{\LL_2}(u^{-1})x
]\cup[\overleftarrow{\LL_1}(u^{-1})x, \overleftarrow{\LL_1}(u^{-1})y ]\right)\},\\
E_{y,\IF}^-
&=&
\{k, |k|\leq N_1(u)+2, I_{k}(u)\cap \left([-\IF, -\overleftarrow{\LL_2}(u^{-1})y]
\cup[\overleftarrow{\LL_1}(u^{-1})y, \IF]\right)\neq \KD{\emptyset}\}.
\end{eqnarray*}
We observe that for $|s|\in [\frac{i-1}{n}\overleftarrow{\LL_2}(u^{-1}), \frac{i+2}{n}\overleftarrow{\LL_2}(u^{-1})]$ with $x/2\leq \frac{i}{n}\leq 2y$ and and
$|t|\in [0, \overleftarrow{\vv}_2(\ln u/u)]$
\BQN\label{EQ2}
&&1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)\nonumber\\
&& \ \ \geq \left[1+(1-3\epsilon)\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)\right]\left[1+(1-3\epsilon)v(i\overleftarrow{\LL_2}(u^{-1})/n,t)\right],
\EQN
whose proofs is postponed in the Appendix.
Let
$$X_{u,k}(s,t)=\overline{X}(k\overleftarrow{\LL}_1(1/u)S+s,t), \quad \mathcal{K}_u=E_{i/n,(i+1)/n }^-, \quad \mathcal{E}_u=\widehat{I}_{0,0}(u),$$
$$d_u(s,t)= (1-3\epsilon)u^2v(i\overleftarrow{\LL_2}(u^{-1})/n, t), \quad h_k(u)=u_{k,\epsilon}^{*-}.$$
We note that
$$\lim_{u\rw\IF}\sup_{k\in \mathcal{K}_u, t\in[-S,S]}\left|(u_{k,\epsilon}^{*-})^2v(i\overleftarrow{\LL_2}(u^{-1})/n, \overleftarrow{\LL_2}(u^{-1})t)-g_{i/n}(t)\right|=0.$$
Thus in light of Lemma \ref{PIPI}, we have
$$\lim_{u\rw\IF}\sup_{x/2\leq \frac{i}{n}\leq 2y}\sup_{k\in E_{i/n,(i+1)/n }^-}\left|(\Psi(u_{k,\epsilon}^{*-}))^{-1}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{0,0}(u)}\frac{\overline{X}(k\overleftarrow{\LL}_1(1/u)S+s,t)}{1+(1-3\epsilon)
v(i\overleftarrow{\LL_2}(u^{-1})/n,t)}>u_{k,\epsilon}^{*-}\right)-\mathcal{H}_{\alpha_1}(S)\widehat{\mathcal{P}}_{\beta}^{(1-3\epsilon)g_{i/n}}
(S)\right|=0.$$
Thus for $[nx]-1\leq i\leq [ny]$,
it follows that
\BQNY
&&\sum_{k\in E_{i/n, (i+1)/n}^-}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right)\\
&& \ \ \ \leq\sum_{k\in E_{i/n, (i+1)/n}^-}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{\left[1+(1-3\epsilon)\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu
s|)\right)\right]\left[1+(1-3\epsilon)v(i\overleftarrow{\LL_2}(u^{-1})/n,t)\right]}>u\right)\nonumber\\
&& \ \ \ \leq\sum_{k\in E_{i/n, (i+1)/n}^-}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1-3\epsilon)v(i\overleftarrow{\LL_2}(u^{-1})/n,t)}>u_{k,\epsilon}^{*-}\right)\nonumber\\
&&\ \ \ =\sum_{k\in E_{i/n, (i+1)/n}^-}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{0,0}(u)}\frac{\overline{X}(k\overleftarrow{\LL}_1(1/u)S+s,t)}{1+(1-3\epsilon)v(i\overleftarrow{\LL_2}(u^{-1})/n,t)}>u_{k,\epsilon}^{*-}\right)\nonumber\\
&& \ \ \ \sim \sum_{k\in E_{i/n, (i+1)/n}^-}\mathcal{H}_{\alpha_1}(S)\widehat{\mathcal{P}}_{\beta}^{(1-3\epsilon)g_{i/n}}
(S)\Psi(u_{k,\epsilon}^{*-})\nonumber\\
&& \ \ \ \sim\mathcal{H}_{\alpha_1}(S)\widehat{\mathcal{P}}_{\beta}^{(1-3\epsilon)g_{i/n}}
(S)\Psi(u)\sum_{k\in E_{i/n, (i+1)/n}^-}e^{-u^2(1-3\epsilon)\inf_{t\in I_k(u)}\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu
s|)\right)}\nonumber\\
&& \ \ \ \leq\frac{\mathcal{H}_{\alpha_1}(S)}{S}\widehat{\mathcal{P}}_{\beta}^{(1-3\epsilon)g_{i/n}}
(S)\frac{\Psi(u)}{\overleftarrow{\LL_1}(u^{-1})}2\int_{i\overleftarrow{\LL_2}(u^{-1})/n}^{(i+1)
\overleftarrow{\LL_2}(u^{-1})/n}e^{-(1-4\epsilon)\frac{\gamma_2 \Mba }{\theta} u^2 \rho_2^2(|s|)}ds(1+o(1)),
\EQNY
as $u\rw\IF$, with $ \Mba $ defined right before Theorem {\ref{th.T3}}.
\KD{Using the same arguments as in} the proof of Lemma \ref{integ}, we have
\BQNY
\int_{i\overleftarrow{\LL_2}(u^{-1})/n}^{(i+1)
\overleftarrow{\LL_2}(u^{-1})/n}e^{-(1-4\epsilon)\frac{\gamma_2 \Mba }{\theta} u^2 \rho_2^2(|s|)}ds &\sim&
\frac{2}{\beta}\overleftarrow{\LL_2}(u^{-1})\int_{(i/n)^{\beta/2}}^{\left((i+1)/n\right)^{\beta/2}}t^{2/\beta-1}
e^{-(1-4\epsilon)\frac{\gamma_2 \Mba }{\theta}t^2}dt\\
&\sim& \overleftarrow{\LL_2}(u^{-1})\int_{i/n}^{(i+1)/n}
e^{-(1-4\epsilon)\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt.
\EQNY
Hence
\BQNY
&&\sum_{k\in E_{i/n, (i+1)/n}^-}\mathbb{P}\left(\sup_{(s,t)\in
\widehat{I}_{k,0}(u)}\frac{\overline{X}(s,t)}{1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right)\nonumber\\
&& \ \ \ \leq 2\frac{\mathcal{H}_{\alpha_1}(S)}{S}\widehat{\mathcal{P}}_{\beta}^{(1-3\epsilon)g_{i/n}}
(S)\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\int_{i/n}^{(i+1)/n}
e^{-(1-4\epsilon)\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt(1+o(1))\nonumber\\
&& \ \ \ \leq 2\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}_{\beta}^{g_{i/n}}
\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\int_{i/n}^{(i+1)/n}
e^{-\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt(1+o(1)), \ \ u\rw\IF, S\rw\IF, \eMM{\epsilon\rw 0}.
\EQNY
Further, by the continuity of $\widehat{\mathcal{P}}_{\beta}^{g_{s}} $ over $s\in [x/2, 2y]$, we have
\BQN\label{EQ4}
\pi_5^-(u)
&\leq& 2\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\sum_{i=[nx]-1}^{[ny]+1}\int_{i/n}^{(i+1)/n}
\widehat{\mathcal{P}}_{\beta}^{g_{i/n}}e^{-\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt(1+o(1))\nonumber\\
&\leq&2\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\int_x^y
\widehat{\mathcal{P}}_{\beta}^{g_{t}}e^{-\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt(1+o(1)), \ \ u\rw\IF,
S\rw\IF, \eMM{\epsilon\rw 0}, n\rw\IF.
\EQN
Similarly,
\BQN\label{EQ5}
\pi_5^+(u)
\geq 2\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\int_x^y
\widehat{\mathcal{P}}_{\beta}^{g_{t}}e^{-\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt(1+o(1)), \ \ u\rw\IF,
S\rw\IF, \eMM{\epsilon\rw 0}, n\rw\IF.
\EQN
Next we focus on $\pi_6(u)$.
In light of (\ref{eq5}) and (\ref{Y}), we have
\BQNY
\pi_6(u)\leq\sum_{k\in E_{0,x}(u)}\mathbb{P}\left(\sup_{(s,t)\in \widehat{I}_{k,0}(u)}Y(s,t)>u\right).
\EQNY
Hence, following case ii) $\gamma_1=0, \gamma_2\in (0,\IF)$ in Theorem \ref{th.T1}, we hav
\BQN\label{EQ6}
\pi_6(u)&\leq& 2\mathcal{H}_{\alpha_1}(S)\widehat{\mathcal{P}}^{\gamma_2\kappa_1/2}_{\beta}(S)
\frac{\Psi(u)}{\overleftarrow{\LL}_1(u^{-1})S}\int_0^{x\overleftarrow{\LL_2}(u^{-1})}e^{-\frac{\kappa_1}{2}u^2v_1^2(t)}dt(1+o(1))\nonumber\\
&\leq& 2\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2\kappa_1/2}_{\beta}
\frac{\overleftarrow{\LL}_2(u^{-1})}{\overleftarrow{\LL}_1(u^{-1})}\Psi(u)\int_0^{x}e^{-\frac{\kappa_1}{2}\frac{\gamma_2}{\theta}t^{\beta}}dt(1+o(1)), \ \
u\rw\IF, S\rw\IF.
\EQN
Similarly,
\BQN\label{EQ7}
\pi_7(u)
\leq 2\mathcal{H}_{\alpha_1}\widehat{\mathcal{P}}^{\gamma_2\kappa_1/2}_{\beta}
\frac{\overleftarrow{\LL}_2(u^{-1})}{\overleftarrow{\LL}_1(u^{-1})}\Psi(u)\int_y^{\IF}e^{-\frac{\kappa_1}{2}\frac{\gamma_2}{\theta}t^{\beta}}dt(1+o(1)), \
\ u\rw\IF, S\rw\IF.
\EQN
Moreover, by Lemma {\ref{simple1}}, (\ref{lambda11}) and (\ref{lambda21}),
\BQN\label{EQ8}
\Lambda(u)\leq \sum_{k<k_1\in E_{x,y}^-(u)}\mathbb{P}\left(\sup_{(s,t)\in \widehat{I}_{k,0}(u)}Y(s,t)>u, \sup_{(s,t)\in
\widehat{I}_{k_1,0}(u)}Y(s,t)>u\right)=o(\pi_5^+(u)), \ \ u\rw\IF, S\rw\IF.
\EQN
Moreover, it follows from Lemma \ref{simple1} and (\ref{PG2}) that
\BQN\label{EQ9}
\pi_8(u)\leq \sum_{|k|=0}^{N_1(u)+2}\sum_{|l|=1, l\neq -1}^{N_2(u)+2}\mathbb{P}\left(\sup_{(s,t)\in I_{k,l}(u)}Y(s,t)>u\right)=o(\pi_5^+(u)), \ \ u\rw\IF,
S\rw\IF.
\EQN
Inserting (\ref{EQ4})--(\ref{EQ9}) into (\ref{EQ3}), we have
$$\pi_1(u)\geq 2\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\int_x^y
\widehat{\mathcal{P}}_{\beta}^{g_{t}}e^{-\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt(1+o(1), \ \ u\rw\IF, S\rw\IF,$$
and
\BQN\label{EQ10}
\pi_1(u)&\leq& 2\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\left(\int_x^y
\widehat{\mathcal{P}}_{\alpha_2}^{g_{t}}e^{-\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt
+\widehat{\mathcal{P}}^{\gamma_2\kappa_1/2}_{\beta}\int_0^{x}e^{-\frac{\kappa_1}{2}\frac{\gamma_2}{\theta}t^{\beta}}dt\right.\nonumber\\
&& \ \ \ \ \left.+
\widehat{\mathcal{P}}^{\gamma_2\kappa_1/2}_{\beta}\int_y^{\IF}e^{-\frac{\kappa_1}{2}\frac{\gamma_2}{\theta}t^{\beta}}dt\right)(1+o(1)),
\EQN
as $u\rw\IF, S\rw\IF$.
Letting $x\rw 0$ and $y\rw\IF$ leads to
$$\pi_1(u)\sim 2\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\LL_2}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\int_0^\IF
\widehat{\mathcal{P}}_{\beta}^{g_{t}}e^{-\frac{\gamma_2 \Mba }{\theta}t^{\beta}}dt(1+o(1), \ \ u\rw\IF, S\rw\IF.$$
which, together with the fact that
$\overleftarrow{\LL_2}(u^{-1})\sim \left(\frac{\gamma_2}{\theta}\right)^{1/\beta}\overleftarrow{\vv_1}(u^{-1}),$
derives the claim.
This completes the proof.\\
\underline{Case iii) $\gamma_1=0, \gamma_2=\IF$}
Let $X_z^\epsilon(s,t), (s,t)\in \mathbb{R}^2, z>0, \epsilon>0$ be homogeneous Gaussian random fields with
correlation function
$$1-Corr\left(X_z^\epsilon(s,t), X_z^\epsilon(S,t_1) \right)\sim (1+\epsilon)\rho_1^2(|s-s_1|)(1+o(1))+\frac{1}{z}v_2^2(|t-t_1|)(1+o(1)), \ \ |s-s_1|,
|t-t_1|\rw\IF.$$
Thus, by Slepian inequality
\BQN\label{EQ12}
\pi_9^+(u)\leq\pi_1(u)\leq \pi_9^-(u),
\EQN
where
\BQNY
\pi_9^+(u):&=&\mathbb{P}\left(\sup_{|s|\leq \overleftarrow{\vv_1}(u^{-1})}X(s,\mu s)>u\right),\\
\pi_9^-(u):&=&\mathbb{P}\left(\sup_{(s,t)\in D_u}\frac{X_z^{\epsilon}(s,t)}{1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right).
\EQNY
\KD{It is straightforward to check that
$\frac{X_z^{\epsilon}(s,t)}{1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}$
satisfies assumptions of}
Case ii) $\gamma_1=0, \gamma_2=(1-\epsilon)z\in (0,\IF)$. Thus
\BQNY
\pi_9^-(u)&\leq& 2\left(\frac{z}{\theta}\right)^{1/\beta}\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\vv_1}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)\left(\int_0^{\IF}
\widehat{\mathcal{P}}_{\beta}^{g_{t,z}}e^{-\frac{z \Mba }{\theta}t^{\beta}}dt
\right)(1+o(1)), \ \ u\rw\IF, \eMM{ S\rw\IF, \ve \to 0},
\EQNY
with $$g_{s,z}(t)=\frac{z}{\theta}\left(|s+bt|^{\beta}+\theta|t|^{\beta}-|(1+b\mu) s|^{\beta}-\theta|\mu s|^{\beta}\right), s\geq 0, t\in \mathbb{R}.$$
Replacing $t$ by $z^{-1/\beta} s$ in the above integral yields
$$z^{1/\beta} \int_0^{\IF} \widehat{\mathcal{P}}_{\beta}^{g_{t,z}}e^{-\frac{z \Mba }{\theta}t^{\beta}}dt=\int_0^{\IF}
\widehat{\mathcal{P}}_{\beta}^{g_{z^{-1/\beta} s,z}}e^{-\frac{ \Mba }{\theta}s^{\beta}}ds.$$
Note that for any $\epsilon>0$, there exists a positive constant
$M_\epsilon>0$ such that \KD{for} $z$ sufficiently large
$$g_{z^{-1/\beta}s,z}(t)+\epsilon|s|^\beta=\frac{1}{\theta}\left(|s+b_{12}tz^{1/\beta}|^{\beta}+\theta|tz^{1/\beta}|^{\beta}-|(1+b_{12}\mu) s|^{\beta}-\theta|\mu
s|^{\beta}\right)+\epsilon|s|^\beta\geq M_\epsilon z|t|^{\beta}, \ \ t\in \mathbb{R},$$
which implies that
$$\widehat{\mathcal{P}}_{\beta}^{g_{z^{-1/\beta} s,z}}\leq e^{\epsilon |s|^\beta}\widehat{\mathcal{P}}_{\beta}^{M_{\epsilon}z}.$$
Since
$$\lim_{z\rw\IF}\widehat{\mathcal{P}}_{\beta}^{M_{\epsilon}z}=1,$$ then using dominated convergence theorem
\BQNY
\limsup_{z\rw\IF}\int_0^{\IF} \widehat{\mathcal{P}}_{\beta}^{g_{z^{-1/\beta} s,z}}e^{-\frac{ \Mba }{\theta}s^{\beta}}ds&\leq&
\limsup_{z\rw\IF}\int_0^{\IF} \widehat{\mathcal{P}}_{\beta}^{M_{\epsilon}z}e^{-\left(\frac{ \Mba }{\theta}-\epsilon\right)s^{\beta}}ds\\
&\rw&
\left(\frac{ \Mba }{\theta}\right)^{-1/\beta}\Gamma(1/\beta+1), \ \ \epsilon\rw 0.
\EQNY
Thus we conclude that
\BQN\label{EQ13}
\pi_9^-(u)&\leq& 2\Gamma(1/\beta+1)\left( \Mba \right)^{-1/\beta}\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\vv_1}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u)(1+o(1)), \ \ u\rw\IF, \eMM{ S\rw\IF, \ve \to 0}.
\EQN
Next we focus on $\pi_9^+(u)$. One can easily check that
the variance and correlation \NE{functions of $X(s,\mu s)$ satisfy}
$$1-Var\left(X(s, \mu s)\right)\sim v_1^2(|(1+b_{12}\mu) s|)+v_2^2(|\mu s|)\sim \Mba v_1^2(|s|), \ \ s\rw 0,$$ and
$$1-Corr\left(X(s, \mu s), X(s_1, \mu s_1)\right)\sim \rho_1^2(|s-s_1|)+\rho_2^2(|\mu(s-s_1)|)\sim \rho_1^2(|s-s_1|), \ \ s, s_1\rw 0.$$
In light of Theorem \ref{mainT}, we have
$$\pi_9^+(u)\sim 2\Gamma(1/\beta+1)\left( \Mba \right)^{-1/\beta}\mathcal{H}_{\alpha_1}
\frac{\overleftarrow{\vv_1}(u^{-1})}{\overleftarrow{\LL_1}(u^{-1})}\Psi(u), \ \ u\rw\IF,$$
which combined with
(\ref{EQ12}) and (\ref{EQ13}) establishes the proof.\\
\underline{Case iv) $\gamma_1\in (0,\IF), \gamma_2=\IF$.} \ \ Let $Z(s,t)$ be a homogeneous Gaussian random field with variance $1$ and correlation
function satisfying
$$1-Corr\left(Z(s,t), Z(s_1,t_1)\right)\sim 2\rho_1^2(|s-s_1|)+\rho_1^2(|t-t_1|), \ \ |s-s_1|\rw 0, |t-t_1|\rw 0,$$
and
$\widehat{I}_{0,0}(u)=[-\overleftarrow{\LL_1}(u^{-1})S, \overleftarrow{\LL_1}(u^{-1})S]\times [-\overleftarrow{\LL_1}(u^{-1})S_1,
\overleftarrow{\LL_1}(u^{-1})S_1]. $
Then, by Slepian's inequality and Lemma \ref{simple1},
\BQN
\pi_{10}^+(u)\leq \pi_1(u)\leq \pi_{10}^-(u)+\pi_{11}(u),
\EQN
where
\BQNY
\pi_{10}^{\pm}(u)&=&\mathbb{P}\left(\sup_{(s,t)\in \widehat{I}_{0,0}(u)}\frac{\overline{X}(s,t)}{1+(1\pm
\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right),\\
\pi_{11}(u)&=&\mathbb{P}\left(\sup_{(s,t)\in
\left(D_u-\widehat{I}_{0,0}(u)\right)}\frac{Z(s,t)}{1+\frac{\kappa_1}{2}\left(v_1^2(|s|)+v_2^2(|t|)\right)}>u\right).
\EQNY
Note that $\rho_2^2(t)=o(\rho_1^2(t))$ as $t\rw 0$ and
$$(1\pm
\epsilon)u^2\left(v_1^2(|\overleftarrow{\LL_1}(u^{-1})s+b_{12}\overleftarrow{\LL_1}(u^{-1})t|)+v_2^2(|\overleftarrow{\LL_1}(u^{-1})t|)\right)\rw (1\pm\epsilon)\gamma_1\left(|s+b_{12}t|^{\alpha_1}+\theta|t|^{\alpha_1}\right), \ \ u\rw\IF$$
uniformly with respect to $(s,t)\in [-S,S]\times[-S_1,S_1]$.
It follows from Lemma \ref{PIPI1} that
\BQNY
\pi_{10}^{\pm}(u)\sim \Psi(u)\EE{\exp{\sup_{(s,t)\in [-S,S]\times
[-S_1,S_1]}\left[\sqrt{2}B_{\alpha_1}(s)-|s|^{\alpha_1}-(1\pm\epsilon)\gamma_1\left(|s+b_{12}t|^{\alpha_1}+\theta|t|^{\alpha_1}\right)\right]}}.
\EQNY
Since
$$\lim_{S_1\rw\IF}\EE{\exp{\sup_{(s,t)\in [-S,S]\times
[-S_1,S_1]}\left[\sqrt{2}B_{\alpha_1}(s)-|s|^{\alpha_1}-(1\pm\epsilon)\gamma_1\left(|s+b_{12}t|^{\alpha_1}+\theta|t|^{\alpha_1}\right)\right]}}
=\widehat{\mathcal{P}}_{\alpha_1}^{(1\pm \epsilon)\gamma_1 M_{\alpha_1} }(S),$$
we have
$$\pi_{10}^{\pm}(u)\sim \widehat{\mathcal{P}}_{\alpha_1}^{\gamma_1 M_{\alpha_1} }\Psi(u), \ \ u\rw\IF, S_1\rw\IF, S\rw\IF,
\eMM{\ve \to 0}.$$
\KD{Using that
$\frac{Z(s,t)}{1+\frac{\kappa_1}{2}\left(v_1^2(|s|)+v_2^2(|t|)\right)}$
satisfies the conditions of Case iii) $\gamma_1, \gamma_2\in (0,\IF)$ of Theorem \ref{th.T1},
by the same argument as given in the proof of (\ref{PG4}), we obtain that}
$\pi_{11}(u)=o\left(\Psi(u)\right), u\rw\IF, S_1\rw\IF, S\rw\IF.$
Thus the proof is completed. \\
\underline{Case iii) $\gamma_1=\gamma_2=\IF$.} \ \
It follows from (\ref{eq5}) and (\ref{Y}) with the specific $B$ in this case that
\BQNY
\mathbb{P}\left(X(0,0)>u\right)\leq \pi_1(u)\leq \mathbb{P}\left(\sup_{(s,t)\in
D_u}Y(s,t)>u\right),
\EQNY
where $\kappa_1$ is defined in Lemma \ref{simple1}.
The Gaussian random field involved in the right hand side of the above inequality satisfies the assumption of
Case iii) $\gamma_1=\gamma_2=\IF$ in Theorem \ref{th.T1}
and
therefore it follows that
$$\mathbb{P}\left(\sup_{(s,t)\in D_u}\frac{\overline{X}(s,t)}{1+\frac{\kappa_1}{2}\left(v_1^2(|s|)+v_2^2(|t|)\right)}>u\right)\sim \Psi(u), \ \ u\rw\IF.$$
This completes the proof. \QED
\prooftheo{th.T5}
\eMM{For $\ve>0$ sufficiently small, let} $Z^{\pm \epsilon}$ be a stationary Gaussian process with continuous trajectories, unit variance and correlation function satisfying
$$1-r_{Z^{\pm \epsilon}}(t)\sim (1\mp \epsilon) \rho_1^2(|t|), \ \ t\rw 0.$$
By Slepain's inequality, we have
\BQNY
\pi_{12}^+(u)\leq \pi_1(u)\leq \pi_{12}^-(u),
\EQNY
where
$$\pi_{12}^\pm(u)=\mathbb{P}\left(\sup_{(s,t)\in D_u}\frac{Z^{\pm\epsilon}(s)}{1+(1\pm\epsilon)\left(v_1^2(|s|)+v_2^2(|t|)\right)}>u\right).$$
By the fact that for any $u>0$,
$$\sup_{(s,t)\in D_u}\frac{Z^{\pm\epsilon}(s)}{1+(1\pm\epsilon)\left(v_1^2(|s|)+v_2^2(|t|)\right)}=\sup_{|s|\leq \overleftarrow{\vv_1}(u^{-1}\ln
u)}\frac{Z^{\pm\epsilon}(s)}{1+(1\pm\epsilon)v_1^2(|s|)},$$
we have
$$\pi_{12}^\pm(u)=\mathbb{P}\left(\sup_{|s|\leq \overleftarrow{\vv_1}(u^{-1}\ln u)}\frac{Z^{\pm\epsilon}(s)}{1+(1\pm\epsilon)v_1^2(|s|)}>u\right).$$
Applying Theorem \ref{mainT}, we establish the claims.
\QED
\prooftheo{th.T6}
Set below for $u>0$
$$D_u=\{|s|\leq \overleftarrow{\vv_1}(u^{-1}\ln u), |t|\leq
2\mu\overleftarrow{\vv_1}(u^{-1}\ln u)\}.$$
Using the same $Z^{\pm \epsilon}$ as in the proof of Theorem \ref{th.T5}, by Slepian's inequality, we have
\BQNY
\pi_{13}^+(u)\leq \pi_1(u)\leq \pi_{13}^-(u),
\EQNY
where
$$\pi_{13}^\pm(u)=\mathbb{P}\left(\sup_{(s,t)\in D_u}\frac{Z^{\pm\epsilon}(s)}{1+(1\pm\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)}>u\right).$$
\KD{The same analysis as given between (\ref{EQ14}) and (\ref{E11}) implies that,} for $u$ sufficiently large
$$(1-\epsilon) \Mbb v_1^2(|s|)\leq \inf_{|t|\leq 2\mu\overleftarrow{\vv_1}(u^{-1}\ln u)}v_1^2(|s+b_{12}t|)+v_2^2(|t|)\leq (1+\epsilon)
\Mbb v_1^2(|s|)$$
hold for $|s|\leq \overleftarrow{\vv_1}(u^{-1}\ln u)$. Thus we have
$$
\pi_{13}^-(u)\leq \mathbb{P}\left(\sup_{|s|\leq \overleftarrow{\vv_1}(u^{-1}\ln u)}\frac{Z^{-\epsilon}(s)}{1+(1-\epsilon)^2
\Mbb v_1^2(|s|)}>u\right),$$ and $$\pi_{13}^+(u)\geq \mathbb{P}\left(\sup_{|s|\leq \overleftarrow{\vv_1}(u^{-1}\ln
u)}\frac{Z^{+\epsilon}(s)}{1+(1+\epsilon)^2 \Mbb v_1^2(|s|)}>u\right).$$
Hence the claim follows by Theorem \ref{mainT}. \QED
\section{Appendix A}
In this section \eM{we derive \eMM{some} key uniform expansions of the tail of maximum of Gaussian random fields
over short intervals}. \eMM{For any $\gamma \in (0,\IF), S>0$ we define
$$ \eMM{\mathcal{P}_{\alpha}^{\gamma} = \EE{ \sup_{[0,S]} e^{ \sqrt{2} B_\alpha(t)- (1+ \gamma) \abs{t}^\alpha}}}$$
and we set
$$\mathcal{P}_{\alpha}^{\IF}[0,S]=1, \quad \mathcal{P}_{\alpha}^{0}[0,S]=\mathcal{H}_{\alpha}[0,S] , \quad
\alpha\in (0,2], S>0 .$$
\eM{The claim of the following three lemmas follows by Theorem 2.1 in \cite{KEP2016}};
the detailed proofs are omitted here.
In the following $h_k, k\in \mathcal{K}_u$ with $\mathcal{K}_u$ an index set are positive functions such that $\limit{u}h_k(u)/u=1$ uniformly with respect to $k\in\mathcal{K}_u$.
\BEL\label{Pickands1} Let $X_{u,k}(t), t\in [0,T], k\in \mathcal{K}_u$ be a sequence of \eM{centered} Gaussian processes with continuous trajectories, variance 1 and correlation function $r(\cdot,\cdot)$ satisfying (\ref{eR}) uniformly with respect to $k\in \mathcal{K}_u$.
Suppose that $\rho\in\mathcal{R}_\alpha/2, \vv\in \mathcal{R}_{\beta/2}$ with $0<\alpha\leq 2,
\beta>0$. If $\lim_{t\rw 0}\frac{\vv^2(t)}{\LL^2(t)} =\gamma\in [0,\IF]$, then
\BQNY
\lim_{u\rw\IF}\sup_{k\in \mathcal{K}_u}\left| \frac{1} {\Psi(h_k(u))} \pk{ \sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]} \frac{X_{u,k}(t)}{1+\vv^2(t)}>h_k(u)}-\mathcal{P}_{\alpha}^\gamma[0,S]\right|=0.
\EQNY
\EEL
\iffalse \prooflem{Pickands1} The idea of the proof is similar to \cite{Pit96}. Here we give the main steps of the proof. Conditioning on
$\mathcal{A}_w=\{G(0)=h(u)-\frac{w}{h(u)}\}, \omega\in \R$, we have
\BQNY
\lefteqn{
\mathbb{P}\left(\sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]}\frac{G(t)}{1+\vv^2(t)}>h(u)\right)}\\
&=&\frac{1}{\sqrt{2\pi}h(u)}e^{-\frac{h^2(u)}{2}}\int_{-\IF}^\IF e^{w-\frac{w^2}{h^2(u)}}\mathbb{P}\left(\sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]}\frac{G(t)}{1+\vv^2(t)}>h(u)\Bigl\lvert
\mathcal{A}_w\right)dw\\
&=&\frac{1}{\sqrt{2\pi}h(u)}e^{-\frac{h^2(u)}{2}}\int_{-\IF}^\IF e^{w-\frac{w^2}{h^2(u)}}\mathbb{P}\left(\sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]}\frac{h(u)G_u(t)-(h^2(u)-w)
(1+\vv^2(\overleftarrow{\rho}(u^{-1})t))}{1+\vv^2(\overleftarrow{\rho}(u^{-1})t)}>w\Bigl\lvert
\mathcal{A}_w \right)dw,
\EQNY
where $G_u(t)=G(\overleftarrow{\rho}(u^{-1})t), t\in [0,S]$.
Further, with $r$ the correlation function of $G$
\BQNY
\lefteqn{
h(u)G_u(t)-(h^2(u)-w)
(1+\vv^2(\overleftarrow{\rho}(u^{-1})t))\Bigl\lvert \ehb{\mathcal{A}_w }=} \\
&\stackrel{d}{=}&
G_u^*(t)+r(\overleftarrow{\rho}(u^{-1})t,0)h(u)(h(u)-\frac{w}{h(u)})-(h^2(u)-w)
(1+\vv^2(\overleftarrow{\rho}(u^{-1})t))
\EQNY
for $t\in [0,S]$, where
$
G_u^*(t)=h(u)\left(G_u(t)-r(\overleftarrow{\rho}(u^{-1})t,0)G_u(0)\right), t\in [0,S].
$
Using (\ref{eR}) yields that for $t,t_1\in [0,S]$, as $u\rw\IF$
$$
Cov(G_u^*(t), G_u^*(t_1))\rw 2Cov\left(B_{\alpha}(t), B_{\alpha}(t_1)\right),$$
and for $\gamma\in [0,\IF)$, uniformly with respect to $t\in [0,S]$,
\BQNY
r(\overleftarrow{\rho}(u^{-1})t,0)h(u)(h(u)-\frac{w}{h(u)})-(h^2(u)-w)
(1+\vv^2(\overleftarrow{\rho}(u^{-1})t))
\rw -(1+\gamma)t^{\alpha}.
\EQNY
Moreover,
\BQNY
\mathbb{E}\left\{\left(G_u^*(t)-G_u^*(t_1)\right)^2\right\}
&\leq & 8h^2(u)\rho^2(\overleftarrow{\rho}(u^{-1})|t-t_1|)\notag
\leq \mathbb{C}|t-t_1|^{\alpha/2}.
\EQNY
\ehb{Consequently},
$$\frac{h(u)G_u(t)-(h^2(u)-w)
(1+\vv^2(\overleftarrow{\rho}(u^{-1})t))}{1+\vv^2(\overleftarrow{\rho}(u^{-1})t)}\Bigl\lvert \ehb{\mathcal{A}_w }, t\in [0,S]$$
converges weakly, as $u \to \IF$, to
$\sqrt{2}B_{\alpha}(t)-(1+\gamma)t^{\alpha}, t\in [0,S].$
Hence
\COM{Therefore, it follows from the continuity of the functional $\sup$ over the space $C[0,S]^2$ that
\BQNY
\mathbb{P}\left(\sup_{(s,t)\in[0,S]^2}\frac{h(u)G_u(s,t)-(h^2(u)-w)
(1+v_4^2(\overleftarrow{v_2}(u^{-1})t))}{1+v_4^2(\overleftarrow{v_2}(u^{-1})t)}>w\Bigl\lvert G_u(0,0)=h(u)-\frac{w}{h(u)}\right)\\
&\quad&\quad\rw
\mathbb{P}\left(\sup_{(s,t)\in [0,S]^2}\left(\sqrt{2}B_{\beta_1/2}(s)+\sqrt{2}B_{\beta_2/2}(t)-s^{\beta_1}-(1+r_2)t^{\beta_2}\right)\right).
\EQNY
Besides, \eE{by (\ref{HOLDER}) using \nelem{lemP} for $u$ large enough}
\BQNY
&&\mathbb{P}\left(\sup_{(s,t)\in[0,S]^2}\frac{h(u)G_u(s,t)-(h^2(u)-w)
(1+v_4^2(\overleftarrow{v_2}(u^{-1})t))}{1+v_4^2(\overleftarrow{v_2}(u^{-1})t)}>w\Bigl\lvert G_u(0,0)=h(u)-\frac{w}{h(u)}\right)\\
&\quad&\quad\leq
\mathbb{P}\left(\sup_{(s,t)\in [0,S]^2}G_u^*(s,t)>(1-a_1)w-a_2\right)\leq
\mathbb{Q}S^2((1-a_1)w-a_2)^{4/\beta_1+4/\beta_2}e^{-\mathbb{Q}_1((1-a_1)w-a_2)^2}, w>w_0.
\EQNY
\eE{Consequently, application of dominated convergence theorem yields}
}
\BQNY
\pk{ \sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]} \frac{G(t)}{1+\vv^2(t)}>h(u)}\sim \mathcal{P}_{\alpha}^\gamma[0,S]\Psi(h(u)), \ \ u\rw\IF.
\EQNY
For $\gamma=\IF$, we have, for any $x>0$ and $u$ sufficiently large,
$$\pk{ \frac{G(0)}{1+\vv^2(0)}>h(u)}\leq \pk{ \frac{G(t)}{1+\vv^2(t)}>h(u)}\leq \pk{ \sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]}
\frac{G(t)}{1+x\LL^2(t)}>h(u)}.$$
Further,
$$\pk{ \sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]} \frac{G(t)}{1+x\LL^2(t)}>h(u)}\sim \mathcal{P}_{\alpha}^x[0,S]\Psi(h(u))\sim \pk{
\frac{G(0)}{1+\vv^2(0)}>h(u)},$$
as $u\rw\IF$ and $x\rw\IF$, which completes the proof.\fi
\iffalse In order to apply theorem 1.1 in \cite{KEP2016}, we check conditions {\bf C0-C3} in \cite{KEP2016}. Note that here $K_u$ only has one element and therefore we denote by
$$Z_u(t)=G(\overleftarrow{\rho}(u^{-1})t), \ \ h_u(t)=v^2(\overleftarrow{\rho}(u^{-1})t), t\in [0,S],\ \ g_u=h(u).$$
{\bf C0} holds in a straightforward way. Using UCT yields that for $\gamma \in [0,\IF)$
$$\lim_{u\rw\IF}|h^2(u)v^2(\overleftarrow{\rho}(u^{-1})t)-\gamma t^\alpha|=0,$$
uniformly with respect to $t\in [0,S]$, which confirms that {\bf C1} hold. Let $\theta_u(s,t)=h^2(u)(1-Cov(Z_u(s), Z_u(t)))$. Then for any $t\in [0,S]$,
$$\lim_{u\rw\IF}|\theta_u(s,t)-Var(B_{\alpha}(t)-B_{\alpha}(s))|\leq \lim_{u\rw\IF}\left|\theta_u(s,t)- \frac{\rho^2(\overleftarrow{\rho}(u^{-1})|t-s|)}{\rho^2(\overleftarrow{\rho}(u^{-1}))}\right|+\lim_{u\rw\IF}\left|
\frac{\rho^2(\overleftarrow{\rho}(u^{-1})|t-s|)}{\rho^2(\overleftarrow{\rho}(u^{-1}))}-|t-s|^{\alpha}\right| =0,$$
implying that {\bf C2} is satisfied. Potter's theorem, \eM{see e.g., \cite{BI1989}} leads to for $u$ sufficiently large
\BQNY
|\theta_u(s,t)|\leq 2h^2(u)\rho^2(\overleftarrow{\rho}(u^{-1})|t-s|)\leq 4\frac{\rho^2(\overleftarrow{\rho}(u^{-1})|t-s|)}{\rho^2(\overleftarrow{\rho}(u^{-1}))}\leq \mathbb{Q}|t-s|^{\alpha/2}, \ \ s,t\in[0,S].
\EQNY
Moreover, using UCT again,
\BQNY
\lim_{\epsilon\downarrow 0} \lim_{u\rw\IF}\sup_{s,t \in [0,S],|t-s|<\epsilon}|h^2(u)\EE{[Z_u(t)-Z_u(s)]Z_u(0)}|\leq \lim_{u\rw\IF}\sup_{t\in [0,S]}|\theta_u(0,t)-t^{\alpha}|+\lim_{\epsilon\downarrow 0}\sup_{s,t\in [0,S], |t-s|<\epsilon}|t^{\alpha}-s^{\alpha}|=0.
\EQNY
Thus {\bf C3} is satisfied. We apply theorem 1.1 in \cite{KEP2016} with $\Gamma=\sup$ and $\eta=B_{\alpha}$ to our case with $\gamma\in [0,\IF)$ establishing the claim. For $\gamma=\IF$, we have, for any $x>0$ and $u$ sufficiently large,
$$\pk{ \frac{G(0)}{1+\vv^2(0)}>h(u)}\leq \pk{ \frac{G(t)}{1+\vv^2(t)}>h(u)}\leq \pk{ \sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]}
\frac{G(t)}{1+x\LL^2(t)}>h(u)}.$$
Further,
$$\pk{ \sup_{t\in [0,\overleftarrow{\rho}(u^{-1})S]} \frac{G(t)}{1+x\LL^2(t)}>h(u)}\sim \mathcal{P}_{\alpha}^x[0,S]\Psi(h(u))\sim \pk{
\frac{G(0)}{1+\vv^2(0)}>h(u)},$$
as $u\rw\IF$ and $x\rw\IF$, which completes the proof.
\QED\\ \fi
\iffalse
\BEL\label{PIPI} Suppose that $G(s,t)$ is a \CE{centered} Gaussian random field over $[0, \overleftarrow{\LL_1}(u^{-1})S_1]\times[0,
\overleftarrow{\LL_2}(u^{-1})S_2]$ with variance 1, \CE{continuous trajectories} and correlation function $r_G(s,t,s_1,t_1)$ satisfying (\ref{Cor2}).
If $\limit{u}\frac{\vv_1^2(t)}{\LL_1^2(t)}\EHE{=} \gamma_1\in [0,\IF]$ and
$\limit{u}\frac{\vv_2^2(t)}{\LL_2^2(t)}\rw \gamma_2\in [0,\IF]$, then as $u\rw\IF$,
\BQNY
\mathbb{P}\left(\sup_{(s,t)\in [0, \overleftarrow{\LL_1}(u^{-1})S_1]\times[0,
\overleftarrow{\LL_2}(u^{-1})S_2]}\frac{G(s,t)}{(1+\vv_1^2(s))(1+\vv_2^2(t))}>h(u)\right)\sim
\mathcal{P}_{\alpha_1}^{\gamma_1}[0,S_1]\mathcal{P}_{\alpha_2}^{\gamma_2}[0,S_2]\Psi(h(u)).
\EQNY
\EEL
\prooflem{PIPI} Similarly as in the proof of Lemma \ref{Pickands1}, we can establish the claims.\fi
\bigskip
Let $\rho_i\in \mathcal{R}_{\alpha_i/2}$, $v_i\in \mathcal{R}_{\beta_i/2}, i=1,2$ be non-negative functions with $0<\alpha_i\leq 2, \beta_i>0, i=1,2.$
Let $\eM{X_{u,k}(s,t)}, k\in \mathcal{K}_u$ be \CE{centered} Gaussian random fields over $\mathcal{E}(u):=\{(\overleftarrow{\LL_1}(u^{-1})\eM{s},\overleftarrow{\LL_2}(u^{-1}) \eM{t} ) , (s,t)\in \mathcal{E}\}$ with $\mathcal{E}$ an compact set containing $0$.
Suppose further that $X_{u,k}$ has unit variance, continuous trajectories and correlation function $r_k(s,t,s_1,t_1)$ satisfying (\ref{Cor2}) uniformly with respect to $k\in \mathcal{K}_u$.
\BEL\label{PIPI}
Let $d_u(s,t),u>0$ be continuous functions satisfying
\BQN\label{eq1}
\lim_{u\rw\IF}\sup_{(s,t)\in\mathcal{E}, k\in \mathbb{K}_u}|h_k^2(u)d_u(\overleftarrow{\LL_1}(u^{-1})s, \overleftarrow{\LL_2}(u^{-1})t)-d(s,t)|=0.
\EQN
\eM{If further $d_u(s,t) > -1$ for any $s,t \in \mathcal{E}$ and all $u$ large}, then
\BQNY
\lim_{u\rw\IF}\sup_{k\in \mathcal{K}_u}\left| \frac{1}{\Psi(h_k(u))) }\mathbb{P}\left(\sup_{(s,t)\in \mathcal{E}(u)}\frac{X_{u,k}(s,t)}{1+d_u(s,t)}>h_k(u)\right)-
\EE{e^{\sup_{(s,t)\in \mathcal{E}}\{W_{\alpha_1,\alpha_2}(s,t)-d(s,t)\}}}\right|=0,
\EQNY
where
$ W_{\alpha_1,\alpha_2}(s,t)=\sqrt{2}(B_{\alpha_1}(s)+\tilde{B}_{\alpha_2}(t))-|s|^{\alpha_1}-|t|^{\alpha_2}, s,t\in \mathbb{R}$,
with $B_{\alpha_1}$ and $B_{\alpha_2}$ being two independent fBm's with indices $\alpha_1, \alpha_2$, respectively.
\EEL
\iffalse\prooflem{PIPI} The idea is to apply theorem 1.1 in \cite{KEP2016} by checking {\bf C0-C3}. Let
$$Z_u(s,t)=G(\overleftarrow{\rho_1}(u^{-1})s, \overleftarrow{\rho_2}(u^{-1})t ), \ \ h_u(s,t)=d_u(s,t), (s,t)\in\mathcal{E},\ \ g_u=h(u),$$
$$\theta_u(s,t,s_1,t_1)=h^2(u)(1-Cov(Z_u(s,t), Z_u(s_1,t_1))), (s,t), (s_1,t_1)\in \mathcal{E}.$$
{\bf C0} holds in a straightforward way and {\bf C1} is satisfied due to ({\ref{eq1}}). {\bf C2} holds since for any $(s,t)\in \mathcal{E}$,
\BQNY
&&\lim_{u\rw\IF}|\theta_u(s,t,s_1,t_1)-Var(B_{\alpha_1}(s)+\widetilde{B}_{\alpha_2}(t)-B_{\alpha_1}(s_1)-\widetilde{B}_{\alpha_2}(t_1))|\\
&& \ \ \leq\lim_{u\rw\IF}\left|\theta_u(s,t,s_1,t_1)- \frac{\rho_1^2(\overleftarrow{\rho_1}(u^{-1})|s-s_1|)}{\rho_1^2(\overleftarrow{\rho_1}(u^{-1}))}-
\frac{\rho_2^2(\overleftarrow{\rho_2}(u^{-1})|t-t_1|)}{\rho_2^2(\overleftarrow{\rho_2}(u^{-1}))}\right|+\lim_{u\rw\IF}\left|
\frac{\rho_1^2(\overleftarrow{\rho_1}(u^{-1})|s-s_1|)}{\rho_1^2(\overleftarrow{\rho_1}(u^{-1}))}-|s-s_1|^{\alpha}\right|\\
&& \ \ \ +\lim_{u\rw\IF}\left|
\frac{\rho_2^2(\overleftarrow{\rho_2}(u^{-1})|t-t_1|)}{\rho_2^2(\overleftarrow{\rho_2}(u^{-1}))}-|t-t_1|^{\alpha}\right| =0.
\EQNY
with $B_{\alpha_1}$ and $\widetilde{B}_{\alpha_2}$ being independent fBms with indices $\alpha_1$ and $\alpha_2$ respectively.
Using similar arguments as in the proof of Theorem \ref{Pickands1}, we obtain that {\bf C3} holds. Thus applying theorem 1.1 in \cite{KEP2016} with $\Gamma=\sup$ and $\eta(s,t)=B_{\alpha_1}(s)+\widetilde{B}_{\alpha_2}(t)$, we establish the claim. \QED \fi
\BEL\label{PIPI1}
Suppose that $d_u(s,t),u>0$ are continuous functions satisfying
$$\lim_{u\rw\IF}\sup_{(s,t)\in\mathcal{E}, k\in \mathcal{K}_u}|h_k^2(u)d_u(\overleftarrow{\LL_1}(u^{-1})s, \overleftarrow{\LL_1}(u^{-1})t)-d(s,t)|=0.$$
If $\rho_2^2(t)=o(\rho_1^2(t))$ as $t\rw 0$ and \eM{$d_u(s,t)> -1$ for any $s,t\in $ and all $u$ large}, then
\BQNY
\lim_{u\rw\IF}\sup_{k\in\mathcal{K}_u}\left|\frac{1}{\Psi(h_k(u)))}\mathbb{P}\left(\sup_{(s,t)\in \widetilde{\mathcal{E}}(u)}\frac{\eM{X}_{u,k}(s,t)}{1+d_u(s,t)}>h_k(u)\right)-
\EE{\sup_{(s,t)\in \mathcal{E}}e^{\sqrt{2}B_{\alpha_1}(s)-|s|^{\alpha_1}-d(s,t)}}\right|=0,
\EQNY
with
$B_{\alpha_1}$ an fBm with index $\alpha_1$ and $\widetilde{\mathcal{E}}(u):=\{(\overleftarrow{\LL_1}(u^{-1})\eM{s},\overleftarrow{\LL_1}(u^{-1})\eM{t}), (s,t)\in \mathcal{E}\}$.
\EEL
\iffalse\prooflem{PIPI1} The idea is also to apply Lemma 1.1 in \cite{KEP2016} by checking {\bf C0-C3} in \cite{KEP2016}. Let
$$Z_u(s,t)=G(\overleftarrow{\rho_1}(u^{-1})s, \overleftarrow{\rho_1}(u^{-1})t ), \ \ h_u(s,t)=d_u(s,t), (s,t)\in\mathcal{E},\ \ g_u=h(u),$$
$$\theta_u(s,t,s_1,t_1)=h^2(u)(1-Cov(Z_u(s,t), Z_u(s_1,t_1))), (s,t), (s_1,t_1)\in \mathcal{E}.$$
\BQNY
&&\lim_{u\rw\IF}|\theta_u(s,t,s_1,t_1)-Var(B_{\alpha_1}(s)-B_{\alpha_1}(s_1))|\\
&& \ \ \leq\lim_{u\rw\IF}\left|\theta_u(s,t,s_1,t_1)- \frac{\rho_1^2(\overleftarrow{\rho_1}(u^{-1})|s-s_1|)}{\rho_1^2(\overleftarrow{\rho_1}(u^{-1}))}-
\frac{\rho_2^2(\overleftarrow{\rho_1}(u^{-1})|t-t_1|)}{\rho_2^2(\overleftarrow{\rho_2}(u^{-1}))}\right|+\lim_{u\rw\IF}\left|
\frac{\rho_1^2(\overleftarrow{\rho_1}(u^{-1})|s-s_1|)}{\rho_1^2(\overleftarrow{\rho_1}(u^{-1}))}-|s-s_1|^{\alpha}\right|\\
&& \ \ \ +\lim_{u\rw\IF}\left|
\frac{\rho_2^2(\overleftarrow{\rho_1}(u^{-1})|t-t_1|)}{\rho_2^2(\overleftarrow{\rho_2}(u^{-1}))}\right| =0.
\EQNY
Thus we confirm that {\bf C2} holds with $\eta(s,t)=B_{\alpha_1}(s)$. Similar arguments as in the proof of Lemma {\ref{Pickands1}} and Lemma \ref{PIPI} can show that {\bf C0-C1} and {\bf C3} also hold. Thus the proof is completed by applying Lemma 1.1 in \cite{KEP2016}.\QED\fi
\iffalse\BEL\label{PIPI2}
Suppose that $\theta=0$, $\eta\in (0,\IF)$ and $h$ is a positive function such that $\limit{u}h(u)/u=1$. If $\limit{u}\frac{\vv_1^2(t)}{\LL_1^2(t)}\rw \gamma_1\in [0,\IF]$ and
$\limit{u}\frac{\vv_2^2(t)}{\LL_2^2(t)}\rw \gamma_2\in [0,\IF]$, then as $u\rw\IF$,
\BQNY
\mathbb{P}\left(\sup_{(s,t)\in \mathcal{E}(u)}\frac{G(s,t)}{1+\vv_1^2(|s+d_{12}t|)+\vv_2^2(|t|)}>h(u)\right)\sim
\mathcal{P}_{\alpha_1}^{\gamma_1,\gamma_2, D_{\eta, \alpha_1}}\left([-S_1,S_1]\times[-S_2,S_2]\right)\Psi(h(u)).
\EQNY
\EEL\fi
Assume now that $X(t), t=(t_1 \ldot t_d)\in \mathbb{R}^d$ is a Gaussian field with continuous trajectories, unit variance and covariance function satisfying
\BQN\label{C1}
1-Cov(X(s), X(t))\sim \sum_{i=1}^d\rho_i(|t_i-s_i|), \quad s, t\rw 0,
\EQN
with $\rho_i $ positive regularly varying function with index $\alpha_i/2\in (0,1]$.
Denote by $\overleftarrow{\rho}_{\vk{d}}(u^{-1})=(\overleftarrow{\rho}_1(u^{-1}), \dots, \overleftarrow{\rho}_d(u^{-1}))$ and define
$\overleftarrow{\rho}_{\vk{d}}(u^{-1})t=(\overleftarrow{\rho}_1(u^{-1})t_1, \dots, \overleftarrow{\rho}_d(u^{-1})t_d)$. Moreover,
set
$F(A,B)=\inf_{s\in A, t\in B}\sum_{i=1}^d|s_i-t_i|$ for any $A, B \subset \mathbb{R}^d$ and let
$$
D_u=\prod_{i=1}^d[-\frac{\delta_u}{\overleftarrow{\rho}_i(u^{-1})}, \quad
\frac{\delta_u}{\overleftarrow{\rho}_i(u^{-1})}], \quad
\mathbb{K}=\{(\lambda_1, \lambda_2)\in D_u\times D_u, \lambda_1+\mathcal{E}_1, \lambda_2 +\mathcal{E}_2\subset D_u\}.$$
Further, let $u_\lambda, \NE{\lambda\in D_u},$ with $\delta_u\rw 0, u\rw\IF$ satisfy
$$\lim_{u\rw\IF}\sup_{\lambda\in \NE{D_u}}\left|\frac{u_\lambda}{u}-1\right|=0.$$
\COM{Denote by
$$ D(\lambda_1, \lambda_2, \mathcal{E}_1, \mathcal{E}_2, u):=\pk{ \sup_{t\in \overleftarrow{\rho}_{\vk{d}}(u^{-1})(\lambda_1+\mathcal{E}_1)} X(t)> u_{\lambda_1}, \sup_{s\in \overleftarrow{\rho}_{\vk{d}}(u^{-1})(\lambda_2 +\mathcal{E}_2)} X(s)> u_{\lambda_2}}$$
with $\mathcal{E}_i\subset [0,S]^d, i=1,2$.
}
\HEH{We state next the result of Corollary 3.2 in \cite{KEP2016}, below $\mathcal{E}_1, \mathcal{E}_2$ are assumed to be compact sets.}
\BEL\label{uni} Suppose that $X(t), t=(t_1\ldot t_d)\in \mathbb{R}^d$ is a Gaussian field with continuous trajectories, unit variance and covariance function satisfying (\ref{C1}). Then there exists $\mathcal{C}, \mathcal{C}_1>0$ such that
for
any $S>1$ as $u$ sufficiently large,
\BQNY
\sup_{(\lambda_1,\lambda_2)\in \mathbb{K} , \mathcal{E}_1, \mathcal{E}_2\subset[0,S]^d} e^{\frac{\mathcal{C}_1F^{\beta^*/2}(\lambda_1+\mathcal{E}_1, \lambda_2+\mathcal{E}_2)}{16}} \frac{ \pk{ \sup_{t\in \overleftarrow{\rho}_{\vk{d}}(u^{-1})(\lambda_1+\mathcal{E}_1)} X(t)> u_{\lambda_1}, \sup_{s\in \overleftarrow{\rho}_{\vk{d}}(u^{-1})(\lambda_2 +\mathcal{E}_2)} X(s)> u_{\lambda_2}}}{ \HEH{S}^{2d} \Psi(\HEH{u}_{\lambda_1, \lambda_2}(u)) }
& \le & \mathcal{C}
\EQNY
with $u_{\lambda_1, \lambda_2}=\min(u_{\lambda_1}, u_{\lambda_2})$ and $\beta^*=\min_{i=1,\dots, d}\alpha_i$.
\EEL
\section{Appendix B}
Let in this section $g$ be a positive function \KD{ such that
$$\lim_{u\rw\IF}g(u)=\IF, \quad \lim_{u\rw\IF}\frac{g(u)}{u}=0.$$
Further, let $v \in \mathcal{R}_\beta, \beta>0$ be a non-negative function. }
We set throughout in the following
\BQN \label{tu}
t(u):= \overleftarrow{\vv}\left(\frac{g(u)}{u}\right), \quad u>0.
\EQN
We shall investigate first the asymptotic behaviour of an integral determined by $g$ and $v$.
\BEL\label{P1} i) For any $0<x\leq y<\IF$ and $c>0$, \KD{as $u\to\infty$}
\BQNY
\int_0^{xt(u)}e^{-c u^2v^2(t)}dt\sim \int_0^{yt(u)}e^{-c u^2v^2(t)}dt.
\EQNY
ii) If $a \in \mathcal{R}_\beta$ \KD{is} such that $a(t)\sim \vv(t)$ as $t\rw 0$, \KD{then as $u\to\infty$}
\BQNY
\int_0^{t(u)}e^{-u^2v^2(t)}dt \sim\int_0^{\overleftarrow{a}(u^{-1}g(u))}e^{-u^2a^2(t)}dt.
\EQNY
\EEL
\def\ttu{t(u)}
\prooflem{P1} i)
\KD{Using standard properties of regularly varying functions, see e.g., \cite{EKM97}, }
for $u$ sufficiently large and
and $0<x<y<\IF$, we have
\BQNY
\int_{x \ttu }^{y\ttu }e^{-c u^2\vv^2(t)}dt
&\leq& e^{-cu^2\vv^2((x/2)\ttu )}(y-x)\ttu \\
&\leq& e^{-(x/3)^{2\beta}cu^2\vv^2(\ttu )}(y-x)\ttu \\
&\leq & e^{-(x/4)^{2\beta}c (g(u))^2}(y-x)\ttu
\EQNY
and
\BQNY
\int_{0}^{x\ttu }e^{-c u^2\vv^2(t)}dt&\geq& \int_{0}^{(x/8)\ttu }e^{-c u^2\vv^2(t)}dt\\
&\geq&e^{-c u^2\vv^2((x/7)\ttu )}(x/8)\ttu \\
&\geq& e^{-(x/6)^{2\beta}cu^2\vv^2(\ttu )}(x/8)\ttu \\
&\geq & e^{-(x/5)^{2\beta}c(g(u))^2}(x/8)\ttu ,
\EQNY
which imply that, \KD{as $u\to\infty$},
\BQNY
\int_{0}^{x\ttu }e^{-c u^2\vv^2(t)}dt\sim \int_{0}^{y \ttu }e^{-c u^2\vv^2(t)}dt.
\EQNY
ii) For any $0<\epsilon<1/2$
$$(1-\epsilon) a(t)\leq \vv(t)\leq (1+\epsilon)a(t)$$
holds for $t$ sufficiently small. Consequently, for $u$ sufficiently large
\BQNY
\int_{0}^{\ttu }e^{-u^2\vv^2(t)}dt&\leq& \int_{0}^{\ttu }e^{-(1-\epsilon)^2u^2 a^2(t)}dt\leq \int_{0}^{\ttu }e^{-u^2a^2((1-2\epsilon)^{1/\beta}t)}dt\\
&=& (1-2\epsilon)^{-1/\beta}\int_{0}^{(1-2\epsilon)^{1/\beta}\ttu }e^{-u^2 a^2(t)}dt\\
&\leq& (1-2\epsilon)^{-1/\beta}\int_{0}^{\ttu }e^{-u^2 a^2(t)}dt
\EQNY
and
\BQNY
\int_{0}^{\ttu }e^{-u^2\vv^2(t)}dt&\geq& \int_{0}^{\ttu }e^{-(1+\epsilon)^2u^2 a^2(t)}dt\geq \int_{0}^{t(u)}e^{-u^2
a^2((1+2\epsilon)^{1/\beta}t)}dt\\
&=& (1+2\epsilon)^{-1/\beta}\int_{0}^{(1+2\epsilon)^{1/\beta}\ttu }e^{-u^2 a^2(t)}dt\\
&\geq& (1+2\epsilon)^{-1/\beta}\int_{0}^{\ttu }e^{-u^2 a^2(t)}dt.
\EQNY
Letting $\epsilon\rw 0$ and by the fact that $\overleftarrow{a}(u^{-1}g(u))\sim t_u$, we establish the second claim. \QED
\iffalse The next
\KD{
lemma provides a slight modification of
Proposition 1.15 in \cite{Soulier}, where instead of $g(u)$ one has $\IF$. Its proof uses the same arguments as therein.
}
\BS Suppose that $z\in \mathcal{R}_\alpha$ \KD{is} locally bounded, and let $h$ be a measurable function such that
$$\abs{h(t)} \le c_1 \min(t,1)^\beta +c_2 \max(t,1)^\gamma, \quad \forall t\in (0,\IF), $$
where $\alpha + \beta > -1, \alpha + \gamma < -1$. If additionally $\abs{h(t)}\le c_1 t^\delta$ in a neighbourhood of 0 with $\delta$ such that $\beta +
\delta > -1$, then
$$
\limit{u} \int_{1/g_1(u)}^{g_2(u)} \frac{ z(ux)}{z(u)} h(x)\, dx = \int_0^\IF x^\beta h(x) \, dx,$$
provided that $\limit{u} g_i(u)=\IF.$
\ES
\fi
\BEL \label{integ} We have
\BQN
\int_0^{\ttu }e^{-cu^2v^2(t)}dt \sim c^{-1/(2\beta)}\Gamma(1+1/(2\beta)) \overleftarrow{v}(1/u),
\quad u\to \IF.
\EQN
\EEL
\prooflem{integ} By \nelem{P1}, ii) we can assume that $v(x)= \ell(x) x^\beta$ with $\ell$ normalized slowly varying function at 0. It is well-known that
$\ell(x) x^\beta$ is ultimately monotone for any $\beta\not=0$, $\ell$ is continuously differentiable and
\BQN\label{ldif}
\lim_{x\to0} \frac{x \ell'(x)}{\ell(x)}=0.
\EQN
Since $v$ is ultimately monotone, we have with \eM{$g(u)$ and $t(u)$ defined by \eqref{tu}}
\BQN
\int_0^{\ttu }e^{-cu^2v^2(t)}dt \sim u^{-1} \int_0^{g(u)} \frac{1}{v'(\overleftarrow{v}(y/u))} e^{- cy^2}\, dy, \quad u\to \IF.
\EQN
Further, \eqref{ldif} implies
$$ \frac{1}{v'(\overleftarrow{v}(y/u))} \sim \frac{1}{\beta} \frac {\overleftarrow{v}(y/u)) } {v(\overleftarrow{v}(y/u))}
\sim \frac{1}{\beta}\frac{u}{y} \overleftarrow{v}(y/u) $$
Consequently, as $u\to \IF$
\BQNY
\int_0^{\ttu}e^{-cu^2v^2(t)}dt
&\sim& \frac{1}{\beta} \int_0^{g(u)} \overleftarrow{v}(y/u)y^{-1} e^{-c y^2}\, dy\notag\\
&\sim& \frac{1}{\beta} \overleftarrow{v}(1/u)\int_0^{g(u)} \frac{\overleftarrow{v}(y/u)}{\overleftarrow{v}(1/u)}y^{-1} e^{-c y^2}\, dy\notag.\\
\EQNY
Potter's theorem shows that there exists a constant $C$ such that for $u$ sufficiently large,
$$\frac{\overleftarrow{v}(y/u)}{\overleftarrow{v}(1/u)}\leq C(\max(1,y))^{2/\beta}, \ \ 0\leq y\leq g(u).$$
By the fact that for any $y>0$
$$ \limit{u} \frac{\overleftarrow{v}(y/u)}{\overleftarrow{v}(1/u)}= y^{1/\beta}$$
and the dominated convergence theorem, \eM{since $\limit{u} g(u)=\IF$}, we obtain
\BQNY
\int_0^{\ttu}e^{-u^2v^2(t)}dt
&\sim& \frac{1}{\beta}\overleftarrow{v}(1/u) \int_0^{\IF} y^{1/\beta-1} e^{-c y^2}\, dy \\
&\sim& c^{-1/(2\beta)}\Gamma(1+1/(2\beta)) \overleftarrow{v}(1/u)\notag.
\EQNY
\eM{Note that alternatively, by \cite{Soulier}[Proposition 1.18] it follows that}
$$\int_0^{g(u)} \frac{\overleftarrow{v}(y/u)}{\overleftarrow{v}(1/u)}y^{-1} e^{-c y^2}\, dy
\sim \int_0^{g(u)} y^{1/\beta-1} e^{-c y^2}\, dy, \quad u\to \IF$$
and thus again the claim follows.} \QED
\iffalse
\BEL\label{integral}
Let $\vv_3\in\mathcal{R}_{\beta_3/2}$ and $\vv_4\in\mathcal{R}_{\beta_4/2}$ be two positive functions with $\beta_3, \beta_4>0$.
\footnote{K: Maybe it suffices to give assumptions for $v_3, v_4$ instead of $v^2_3, v^2_4$.}
If
$\frac{\vv^2_3(t)}{\vv^2_4(t)}\rw 0$, then
we have
\BQNY
\sum_{k=0}^{N(u)}e^{-c_1u^2\vv^2_3(k\overleftarrow{\vv_4}(u^{-1})S)}&\sim& \frac{1}{\overleftarrow{v}_4(u^{-1})
c_1^{1/\beta_3}S}\int_0^{
\overleftarrow{v_3}(u^{-1}g(u))}e^{-u^2v^2_3(t)}dt\\
&\sim& c_1^{-1/\beta_3}\Gamma(1/\beta_3+1)S^{-1}\frac{\overleftarrow{v}_3(u^{-1})}{\overleftarrow{v}_4(u^{-1})
}\rw\IF
, \ \ u\rw\IF,
\EQNY
with $c_1>0$, $S>0$ and $N(u)=\left[\frac{\overleftarrow{\vv_3}(u^{-1}g(u))}{\overleftarrow{\vv_4}(u^{-1})S}\right]+1$.
\EEL
\prooflem{integral}.
By UCT
we have that for any
$0<\epsilon<1/2$ there exists $u_0$ such that for $u>u_0$
\BQN\label{REIN}
v_3^2(\overleftarrow{v_4}(u^{-1})kS)\leq (1+\epsilon)v_3^2(t), \ \ v_3^2(\overleftarrow{v_4}(u^{-1})kS)\geq \left(
\left(\frac{k}{k+1}\right)^{\beta_3}-\epsilon/2\right)v_3^2(t)
\EQN
hold for $t\in [k\overleftarrow{\vv_4}(u^{-1})S, (k+1)\overleftarrow{\vv_4}(u^{-1})S]$ and any $0\leq k\leq N(u)$. Further, we can find $k_{\epsilon}$
such that for $u>u_0$ and $ k_{\epsilon_1}\leq k\leq N(u)$,
\BQN\label{REIN1}
v_3^2(\overleftarrow{v_4}(u^{-1})kS)\geq \left(1-\epsilon\right)v_3^2(t),
\EQN
for $t\in [k\overleftarrow{\vv_4}(u^{-1})S, (k+1)\overleftarrow{\vv_4}(u^{-1})S]$.
Consequently, for $u$ sufficiently large
\BQN\label{NEQ1}
\sum_{k=0}^{N(u)}
e^{-c_1u^2v_3^2(\overleftarrow{v_4}(u^{-1})kS)}&\leq& k_{\epsilon}+\sum_{k= k_{\epsilon}}^{N(u)}
e^{-c-1u^2v_3^2(\overleftarrow{v_4}(u^{-1})kS)}\nonumber\\
&\leq&k_{\epsilon}+\frac{1}{\overleftarrow{v_4}(u^{-1})S}\sum_{k= k_{\epsilon}}^{N(u)}
\int_{\overleftarrow{v_4}(u^{-1})kS}^{\overleftarrow{v_4}(u^{-1})(k+1)S}e^{-c_1u^2(1-\epsilon)v_3^2(t)}dt\nonumber\\
&\leq& k_{\epsilon}+\frac{1}{\overleftarrow{v_4}(u^{-1})S}\sum_{k= k_{\epsilon}}^{N(u)}
\int_{\overleftarrow{v_4}(u^{-1})kS}^{\overleftarrow{v_4}(u^{-1})(k+1)S}e^{-u^2v_3^2((1-2\epsilon)^{1/\beta_3}c_1^{1/\beta_3}t)}dt\nonumber\\
&\leq&k_{\epsilon}+\frac{1}{\overleftarrow{v_4}(u^{-1})S}
\int_{0}^{\overleftarrow{v_3}\left(\frac{g(u)}{u}\right)}e^{-u^2v_3^2((1-2\epsilon)^{1/\beta_3}c_1^{1/\beta_3})t)}dt\nonumber\\
&\leq&k_{\epsilon}+\frac{1}{\overleftarrow{v_4}(u^{-1})(1-2\epsilon)^{1/\beta_3}c_1^{1/\beta_3}S}
\int_{0}^{c_1^{1/\beta_3}\overleftarrow{v_3}\left(\frac{g(u)}{u}\right)}e^{-u^2v_3^2(t)}dt.
\EQN
Similarly, by (\ref{REIN}), for $u$ sufficiently large, we have
\BQN\label{NEQ2}
\sum_{k=0}^{N(u)}
e^{-c_1u^2v_3^2(\overleftarrow{v_4}(u^{-1})kS)}&\geq&\frac{1}{\overleftarrow{v_4}(u^{-1})S}\sum_{k= 0}^{N(u)-1}
\int_{\overleftarrow{v_4}(u^{-1})kS}^{\overleftarrow{v_4}(u^{-1})(k+1)S}e^{-c_1u^2(1+\epsilon)v_3^2(t)}dt\nonumber\\
&\geq& \frac{1}{\overleftarrow{v_4}(u^{-1})S}\sum_{k=0}^{N(u)-1}
\int_{\overleftarrow{v_4}(u^{-1})kS}^{\overleftarrow{v_4}(u^{-1})(k+1)S}e^{-u^2v_3^2((1+2\epsilon)^{1/\beta_3}c_1^{1/\beta_3}t)}dt\nonumber\\
&\geq&\frac{1}{\overleftarrow{v_4}(u^{-1})S}
\int_{0}^{\overleftarrow{v_3}\left(\frac{g(u)}{u}\right)}e^{-u^2v_3^2((1+2\epsilon)^{1/\beta_3}c_1^{1/\beta_3}t)}dt\nonumber\\
&\geq&\frac{1}{\overleftarrow{v_4}(u^{-1})(1+2\epsilon)^{1/\beta_3}c_1^{1/\beta_3}S}
\int_{0}^{c_1^{1/\beta_3}\overleftarrow{v_3}\left(\frac{g(u)}{u}\right)}e^{-u^2v_3^2(t)}dt.
\EQN
Moreover, using Potter's bound for the regularly varying function $v_3(t)$ at zero we have, for $u$ large enough and $S>1$
$$\frac{v_3^2(\overleftarrow{v_4}(u^{-1})kS)}{v_3^2(\overleftarrow{v_4}(u^{-1}))}\leq (kS)^{\beta_3'} $$
for $1\leq k\leq N(u)$ and $\beta_3'>\beta_3$.
Therefore, we have
\BQN\label{NEQ3}
\sum_{k=0}^{N(u)}
e^{-c_1u^2v_3^2(\overleftarrow{v_4}(u^{-1})kS)}&\geq& \sum_{k=1}^{N(u)-1}
e^{-c_1u^2v_3^2(\overleftarrow{v_4}(u^{-1}))(kS)^{\beta_3'}}\nonumber\\
&\sim& (u^2v_3^2(\overleftarrow{v_4}(u^{-1})))^{-1/\beta_3'}S^{-1}\int_{0}^\IF e^{-c_1x^{\beta_3'}}dx\rw\IF,
\EQN
as $u\rw\IF$.
Combing (\ref{NEQ1})-(\ref{NEQ3}) with \nelem{P1}, i) and letting $u\rw\IF$ and $\epsilon\rw 0$ leads to
\BQNY
\sum_{k=0}^{N(u)}
e^{-c_1u^2v_3^2(\overleftarrow{v_4}(u^{-1})kS)}\sim \frac{1}{\overleftarrow{v}_4(u^{-1})
c_1^{1/\beta_3}S}\int_0^{
\overleftarrow{v_3}(u^{-1}\ln u)}e^{-u^2v^2_3(t)}dt.
\EQNY
\fi
\BEL\label{simple} Suppose that $\rho^2_1\in \mathcal{R}_{\alpha_1}$ and $\rho_2^2\in \mathcal{R}_{\alpha_2}$ with $\alpha_1, \alpha_2 >0$. If
$\rho_1^2(|t|)=o(\rho_2^2(|t|))$ as $t\rw 0$, then for any $a, b \in \mathbb{R}$,
\BQNY
\rho_1^2(|as +b t|)+\rho_2^2(|t|)\sim \rho_1^2(|as|)+\rho_2^2(|t|), \quad s, t\rw 0.
\EQNY
\EEL
\prooflem{simple} The claim follows \eM{easily if $abst=0$}. Next we suppose that
$abst\neq 0$.
It suffices to prove that
$$\lim_{s,t\rw 0, st\neq 0}\frac{\abs{\rho_1^2(|as +b t|)-\rho_1^2(|as|)}}{\rho_1^2(|as|)+\rho_2^2(|t|)}=0.$$
For any $\epsilon\in (0,\alpha_1)$, if $\abs{\frac{as}{bt}}>\frac{4\alpha_1}{\epsilon}$, then
$$1-\frac{\epsilon}{4\alpha_1}\leq\frac{|as+bt|}{|as|}\leq 1+\frac{\epsilon}{4\alpha_1}.$$
Thus in light of UCT, we have, for $s,t$ sufficiently small
\BQNY
\frac{\abs{\rho_1^2(|as +b t|)-\rho_1^2(|as|)}}{\rho_1^2(|as|)+\rho_2^2(|t|)}&\leq& \frac{\rho_1^2(|as|)\abs{\frac{\rho_1^2(|as +b
t|)}{\rho_1^2(|as|)}-1}}{\rho_1^2(|as|)+\rho_2^2(|t|)}\leq \ABs{\frac{\rho_1^2(|as +b t|)}{\rho_1^2(|as|)}-1}\\
&\leq& \max\left(\left(1+\frac{\epsilon}{2\alpha_1}\right)^{\alpha_1}-1, 1-\left(1-\frac{\epsilon}{2\alpha_1}\right)^{\alpha_1}\right)
=: b_{\epsilon}.
\EQNY
For any $\epsilon \in (0,\alpha_1)$, if $\abs{\frac{as}{bt}}\leq\frac{4\alpha_1}{\epsilon}$, then
$$
\frac{|as+bt|}{|bt|}\leq 1+\frac{4\alpha_1}{\epsilon}.$$
\eM{Applying} again UCT we obtain
\BQNY
\frac{\abs{\rho_1^2(|as +b t|)-\rho_1^2(|as|)}}{\rho_1^2(|as|)+\rho_2^2(|t|)}&\leq&
\frac{\rho_1^2(|bt|)}{\rho_1^2(|as|)+\rho_2^2(|t|)}\ABs{\frac{\rho_1^2(|as +b t|)}{\rho_1^2(|bt|)}-\frac{\rho_1^2(|as|)}{\rho_1^2(|bt|)}}\\
&\leq&
\frac{\rho_1^2(|bt|)}{\rho_2^2(|t|)}\left[\left(1+\frac{8\alpha_1}{\epsilon}\right)^{\alpha_1}+\left(\frac{8\alpha_1}{\epsilon}\right)^{\alpha_1}\right]\\
&\rw& 0, \ \ \text{as}\ \ st\neq 0, s,t\rw 0.
\EQNY
Consequently,
$$\lim_{s,t\rw 0, st\neq 0}\frac{\abs{\rho_1^2(|as +b t|)-\rho_1^2(|as|)}}{\rho_1^2(|as|)+\rho_2^2(|t|)}\leq b_{\epsilon}\rw 0,\ \ \epsilon\rw 0.$$
This completes the proof. \QED
\BEL\label{simple1} Suppose that $v^2_1, v^2_2\in \mathcal{R}_{\beta}$ $\beta >0$. If $a_1 v^2_2(|t|)\leq v_1^2(|t|)\leq a_2 v_2^2(|t|))$ with $a_1,
a_2>0$ for $t$ sufficiently small, then for any reversible matrix
$B=\left(\begin{array}{cc}
b_{11}& b_{12}\\
b_{21}& b_{22}\\
\end{array}
\right),$
there exist two positive constants $\kappa_1$ and $\kappa_2$ such that
$$\kappa_1v^2_1(|s|)+\kappa_1v^2_2(|t|)\leq v^2_1(|b_{11}s+b_{12}t|)+v^2_2(|b_{21}s+b_{22}t|)\leq \kappa_2v^2_1(|s|)+\kappa_2v^2_2(|t|)$$
holds in a neighbourhood of \Kd{0}.
\EEL
\prooflem{simple1} Without loss of generality, we assume that $|t|\geq |s|$ and $|t|>0$.
By UCT, we have
\BQNY
\frac{ v^2_1(|b_{11}s+b_{12}t|)+v^2_2(|b_{21}s+b_{22}t|)}{v^2_1(|s|)+v_2^2(|t|)}&\leq& \frac{
a_2v^2_2(|t|(|b_{12}|+|b_{11}\frac{s}{t}|))+v^2_2(|t|(|b_{22}|+|b_{21}\frac{s}{t}|))}{v^2_2(|t|)}\\
&\leq&2\left(a_2(|b_{11}|+|b_{12}|)^{\beta}+(|b_{21}|+|b_{22}|)^{\beta}\right),
\EQNY
for $t$ sufficiently small. Hence we get the upper bound. For the lower bound, making a linear transformation
$$(s,t)^\top= B^{-1}(s',t')^\top=\left(\begin{array}{cc}
b_{11}'& b_{12}'\\
b_{21}'& b_{22}'\\
\end{array}
\right)(s',t')^\top,$$
and then using the above conclusion, we have
\BQNY
v^2_1(|s|)+v^2_2(|t|)&=& v^2_1(|b_{11}'s'+b_{12}'t'|)+v^2_2(|b_{21}'s'+b_{22}'t'|)\\
&\leq& 2\left(a_2(|b_{11}'|+|b_{12}'|)^{\beta}+(|b_{21}'|+|b_{22}'|)^{\beta}\right)(v^2_1(|s'|)+v^2_2(|t'|))\\
&\leq& 2\left(a_2(|b_{11}'|+|b_{12}'|)^{\beta}+(|b_{21}'|+|b_{22}'|)^{\beta}\right)\left(v^2_1(|b_{11}s+b_{12}t|)+v^2_2(|b_{21}s+b_{22}t|)\right),
\EQNY
provided $|t'|\geq |s'|$ and $|t'|>0$ for $t'$ sufficiently small. This completes the proof.
\QED
{\textbf {Proof of (\ref{EQ2})}}.
Note that
\BQN\label{EQ14}
v_1^2(|s+b_{12}t|)+v_2^2(|t|)=v_1^2(|s|)\frac{v_1^2(|s||1+b_{12}t/s|)+v_2^2(|s||t/s|)}{v_1^2(|s|)}, \ \ (s,t)\in D_u.
\EQN
If $|t/s|\leq M<\IF$, then by UCT
$$\sup_{(s,t)\in D_u, |t/s|\leq M }\left|\frac{v_1^2(|s||1+b_{12}t/s|)+v_2^2(|s||t/s|)}{v_1^2(|s|)}-|1+b_{12}t/s|^{\beta}-\theta|t/s|^{\beta}\right|\rw 0, \ \ u\rw
\IF.$$
If $|t/s|\geq M$, then using Potter's bound, for $u$ sufficiently large
$$\inf_{(s,t)\in D_u, |t/s|\geq M}\frac{v_1^2(|s||1+b_{12}t/s|)+v_2^2(|s||t/s|)}{v_1^2(|s|)}\geq 1/2\left(||b_{12}|M-1|^{\beta/2}+\theta M^{\beta/2}\right).$$
Therefore, the minimum of $v_1^2(|s+b_{12}t|)+v_2^2(|t|)$ \KD{is attained for}
$|t/s|\leq M$ with $M$ sufficiently large. Further, the minimum of
$|1+b_{12}t/s|^{\beta}+\theta|t/s|^{\beta}$ is attained at $\mu :=t/s\in [-1/|b_{12}|, 1/|b_{12}|]$. Thus, for $(s,t)\in D_u$ and $u$ sufficiently large
\BQN\label{E11}
\frac{v_1^2(|s+b_{12}t|)+v_2^2(|t|)}{v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu
s|)}&=&\frac{v_1^2(|s+b_{12}t|)+v_2^2(|t|)}{v_1^2(|s|)}\frac{v_1^2(|s|)}{v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)}\nonumber\\
&\geq& \frac{|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}}{2}\frac{1}{2(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta})}=1/4,
\EQN
Recall that $v(s,t)=v_1^2(|s+b_{12}t|)+v_2^2(|t|)-v_1^2(|(1+b_{12}\mu)s|)-v_2^2(|\mu s|), \ \
(s,t)\in D_u$.
Note that $v(s,t)$ may be negative at some point. It follows that
\BQNY
&&\left[1+(1-2\epsilon)\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)\right]\left[1+(1-2\epsilon)v(s,t)\right]\\
&& \ \ =1+(1-2\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)+(1-2\epsilon)^2\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)v(s,t).
\EQNY
\KD{Moreover} (\ref{E11}) yields that
$$\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)v(s,t)=o\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right), \ \ (s,t)\in D_u, u\rw\IF.$$
Thus, we have for any $0<\epsilon<1/4$ \KD{and} sufficiently large $u$
\BQN\label{eq7}
\left[1+(1-2\epsilon)\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)\right]\left[1+(1-2\epsilon)v(s,t)\right]\leq
1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right),
\EQN
with $(s,t)\in D_u$. Since for
$|s|\in [\overleftarrow{\LL_2}(u^{-1})x/2, \overleftarrow{\LL_2}(u^{-1})2y], |t|\in [M\overleftarrow{\LL_2}(u^{-1}), \overleftarrow{\vv}_2(\ln
u/u)]$
\BQNY
v(s,t)&=&v_1^2(|t|)\frac{v_1^2(|s+b_{12}t|)+v_2^2(|t|)-v_1^2(|(1+b_{12}\mu)s|)-v_2^2(|\mu s|)}{v_1^2(|t|)}\\
&\sim& v_1^2(|t|)\left(\left|b_{12}+\frac{s}{t}\right|^{\beta}+\theta-|1+b_{12}\mu|^{\beta}\left|\frac{s}{t}\right|^{\beta}-\theta\left|\mu
\frac{s}{t}\right|^{\beta}\right), \ \ u\rw\IF
\EQNY
then for $M,u$ sufficiently large
\BQN\label{EQ1}
v(s,t)\geq \frac{1-3\epsilon}{1-2\epsilon}v(s_1,t_1),
\EQN
with $|s|, |s_1|\in [\overleftarrow{\LL_2}(u^{-1})x/2, \overleftarrow{\LL_2}(u^{-1})2y], |t|, |t_1|\in [M\overleftarrow{\LL_2}(u^{-1}),
\overleftarrow{\vv}_2(\ln u/u)]$.
Moreover, for any $\epsilon_1>0$, $|s|\in [\overleftarrow{\LL_2}(u^{-1})x/2, \overleftarrow{\LL_2}(u^{-1})2y]$ and $|t|\in [0,
M\overleftarrow{\LL_2}(u^{-1})]$, by UCT
\BQNY
v(s,t)&\geq&v_1^2(|s|)\left[(1-\epsilon_1)\left(\left|1+b_{12}t/s\right|^{\beta}+\theta\left|t/s\right|^{\beta}\right)-(1+\epsilon_1)
\left(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}\right)\right]\nonumber\\
&\geq& -2\epsilon_1 \left(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}\right)v_1^2(|s|),
\EQNY
and for any $|s|, |s_1|\in [\frac{i-1}{n}\overleftarrow{\LL_2}(u^{-1}), \frac{i+2}{n}\overleftarrow{\LL_2}(u^{-1})]$ with $x/2\leq \frac{i}{n}\leq 2y$
and $|t|\in [0, M\overleftarrow{\LL_2}(u^{-1})]$ and $u$ and $n$ sufficiently large
\BQNY
&&|v(s,t)-v(s_1,t)|
\leq v_1^2(|s|) \sup_{d_1,d_2\in \{\pm \epsilon_1\}}\left|(1+d_1)\left(|1+b_{12}t/s|^{\beta}+|(1+b_{12}\mu)s_1/s|^{\beta}+\theta|\mu
s_1/s|^{\beta}\right)\right.\\
&& \ \ \left.-(1+d_2)\left(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}+|s_1/s+b_{12}t/s|^{\beta}\right)\right|\\
&& \ \ \leq v_1^2(|s|)\mathbb{Q}\epsilon_1\left(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}+|1+2|b_{12}|M/x|^{\beta}\right)+v_1^2(|s|)||s_1/s|^{\beta}-1|
\left(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}\right)\\
&& \ \ \ \ +v_1^2(|s|)\sup_{|z|\in [0, 4M/x] }|h_{s_1/s}( z)-h_1(z)|,
\EQNY
where $h_s(z)=|s+b_{12}z|^{\beta}, s, z\in \mathbb{R}$. Therefore, for $|s|, |s_1|\in [\frac{i-1}{n}\overleftarrow{\LL_2}(u^{-1}), \frac{i+2}{n}\overleftarrow{\LL_2}(u^{-1})]$ with $x/2\leq \frac{i}{n}\leq 2y$
and $|t|\in [0, M\overleftarrow{\LL_2}(u^{-1})]$ when $\epsilon_1$ sufficiently small and $u$ and $n$ sufficiently large
$$v(s,t)\geq -\epsilon/4 \left(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}\right)v_1^2(|s|), \ \ \ |v(s,t)-v(s_1,t)|\leq \epsilon/8
\left(|1+b_{12}\mu|^{\beta}+\theta|\mu|^{\beta}\right)v_1^2(|s|),$$
which implies that (\eM{recall that $\limit{u}\sup_{(s,t)\in D_u}|v(s,t)|= 0$)}
\BQNY
&&\epsilon \left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)+\epsilon v(i\overleftarrow{\LL_2}(u^{-1})/n,t)\\
&& \ \ \ \geq (1-2\epsilon)\left(v(i\overleftarrow{\LL_2}(u^{-1})/n,t)-v(s,t)\right)-(1-2\epsilon)^2\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu
s|)\right)v(s,t)\\
&& \ \ \ \ \ +(1-3\epsilon)^2\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)v(i\overleftarrow{\LL_2}(u^{-1})/n,t).
\EQNY
Hence, \KD{combining the above} with (\ref{eq7}) and (\ref{EQ1}) for any $0<\epsilon<1/4$, we have for $n,u$ sufficiently large,
\BQNY
1+(1-\epsilon)\left(v_1^2(|s+b_{12}t|)+v_2^2(|t|)\right)&\geq& \left[1+(1-2\epsilon)\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu
s|)\right)\right]\left[1+(1-2\epsilon)v(s,t)\right]\nonumber\\
&\geq& \left[1+(1-3\epsilon)\left(v_1^2(|(1+b_{12}\mu)s|)+v_2^2(|\mu s|)\right)\right]\left[1+(1-3\epsilon)v(i\overleftarrow{\LL_2}(u^{-1})/n,t)\right],
\EQNY
holds for $|s|\in [\frac{i-1}{n}\overleftarrow{\LL_2}(u^{-1}), \frac{i+2}{n}\overleftarrow{\LL_2}(u^{-1})]$ with $x/2\leq \frac{i}{n}\leq 2y$ and and
$|t|\in [0, \overleftarrow{\vv}_2(\ln u/u)]$, which completes the proof. \QED\\
{\bf Acknowledgement}: Thanks to Swiss National Science Foundation grant No. 200021-166274. KD acknowledges partial support
by NCN Grant No 2015/17/B/ST1/01102 (2016-2019).
\bibliographystyle{plain}
|
\section{Introduction}
Deep Neural Networks (DNNs) have been shown to be a powerful tool in learning from large datasets in many domains. However, they require a large number of training samples in order to have reasonable generalization power. Indeed, this fact tends to limit their application to the type of images where it is easy to construct a large labeled database, such as natural images where collection can be done from image sharing websites and labels obtained at low cost via crowdsourcing. When only few labeled samples are available, such as in medical imaging, DNNs tend to overfit to the data and may not be competitive with methods for which known tractable methods for stronger regularization are available, e.g.\ kernel methods.
Generally in learning theory, one wants to minimize the statistical risk $
\mathcal{R}(W) = \int \ell(f_W(x),y)dP(x,y)$, where $\ell$ is a loss a function and $P(x,y)$ is the data distribution of inputs and outputs, respectively. However, as $P$ is generally unknown, only the empirical risk $\mathcal{R}_N(W) = \frac{1}{N}\sum_{i=1}^N \ell(f_W(x_i),y_i)$ is accessible.
Statistical learning theory \cite{vapnik2013nature} gives conditions for the empirical risk to be a good approximation of the statistical risk, by giving a bound of the form:
\begin{equation}
\mathcal{R}(W) \le \mathcal{R}_N(W) + g(h) / \sqrt{N},
\end{equation}
where $g(h)$ is an increasing function of $h$, the Vapnik–Chervonenkis dimension (VC dimension). This measure depends on the function set. When $g(h)$ is small, the minimizer of the empirical risk gives a small statistical risk even when $N$ is small. Through their activation function and their weights, DNNs index very efficiently a large function space with few parameters \cite{cohen2015expressive}. However, the VC dimension of this function set tends to be high. Indeed, many researchers have studied the VC-dimension of some simple neural networks \cite{bianchini2014complexity, karpinski1997polynomial, sontag1998vc}. It has been shown that, for example, for networks with only sigmoidal activation functions, the optimal bound on VC dimension is $\mathcal{O}(\rho^2)$, where $\rho$ is the number of weights. The authors of~\cite{bianchini2014complexity} also show that a deep network implement functions of higher complexity than a shallow network with the same number of weights (i.e.\ with the same generalization power according to the VC-dimension bounds), proving one of the many advantages of DNNs. Nevertheless, the DNNs that have become popular recently tend to have a high number of weights (For example, the Oxford VGG networks have more that 110 million parameters~\cite{simonyan2014very}), which makes their generalization power decrease significantly for small databases. For this reason, an appropriate regularization is required. Training a DNN with regularization is solving the following problem:
\begin{equation}
\arg\min_{W} \mathcal{R}_N(W) + \lambda\Omega(f_W).
\label{eq:prob2}
\end{equation}
Much research has approached the idea of regularized DNN training. However, until now, this regularization tends to be poor compared for examples to regularization in kernel methods~\cite{Scholkopf:2001:LKS:559923}. The common idea behind regularization is the application of Occam's razor: if several solutions exist for a given problem, we favor the ``simplest'' solution. Which notion of simplicity is employed typically is determined by prior knowledge or computational considerations. Two categories of regularization can be distinguished in the existing literature:
\begin{inparaenum}[(i)]
\item regularizing by controlling the magnitude of the weights $W$($\ell_1$ and $\ell_2$ regularization \cite{Goodfellow-et-al-2016-Book,moody1995simple});
\item regularizing by controlling the network architecture (DropOut \cite{hinton2012improving} and DropConnect \cite{wan2013regularization}).
\end{inparaenum}
\iffalse
To control the magnitude of the weights, the term $\Omega(f_W)$ in \eqref{eq:prob2} is replaced by another constraint operating on the vector of weights $W$. The most commonly used constraints are the $\ell_1$ and $\ell_2$\footnote{This method is known as weight decay.} norms of $W$ \cite{moody1995simple}. In this type of regularization, the main goal is to fit the training data, while keeping the weights small, which means a reasonably complex network.
In the regularization based on topology control, rather than introducing a regularization term, some researchers have explored the possibility of keeping the problem \eqref{eq:prob1} while changing the architecture during the training process:
\begin{inparaenum}[(i)]
\item The first (and more popular) way to apply this idea is a method known as DropOut \cite{hinton2012improving}: At each step of training, half of the neuron activation functions are randomly set to zero. This makes the training procedure similar to model averaging, and improves the generalization of the resulting network.
\item The second method is inspired from the DropOut method; rather than dropping activation functions, DropConnect sets a random part of the weights to zero \cite{wan2013regularization}.
\end{inparaenum}
Even if these methods have been shown to be efficient on data where a big database can be constructed, they fail when only few samples are available. Indeed, these regularization methods are poor and deal with the complexity of $f_W$ only indirectly.
\fi
In order to make the regularization stronger, we study in this paper the feasibility of using $L_2$ function norm:
\begin{equation}
\Omega(f_W) = \| f_W\|_2^2 =\int \langle f_W(x), f_W(x) \rangle dx.
\end{equation}
As the exact value of this norm, and more importantly of its gradient, are not accessible, a good approximation is needed. Two approximations are considered.
The first idea is to consider the function norm with respect to the probability measure (we suppose for simplicity that all the distributions admit density functions):
\begin{equation}
\Omega(f_W) =\int \langle f_W(x), f_W(x) \rangle dP(x),
\end{equation}
where $P$ is the marginal data distribution:
\begin{equation}
P(x):= \int P(x,y) dy.
\label{eq:proba_x}
\end{equation}
One can easily see that this is a valid function norm~\cite{galambos1995advanced}, and furthermore is just a weighted canonical function norm such that the regularization has a higher effect on the data that has the higher probability to be observed.
The second idea is to use a Markov-chain sampling method for numerical integration~\cite{andrieu2003introduction}. Here, the idea is to generate samples from a probability density proportional to $\| f_W(x)\|_2^2$ (vector $\ell_2$ norm), and to use them to approximate the integral by:
\begin{equation}
\int \langle f_W(x), f_W(x) \rangle dx \approx \|\alpha\|_1^{-1} \sum_{i=1}^n \alpha_{i} \langle f_W(x_i), f_W(x_i) \rangle
\end{equation}
for some positive weights $\alpha_i$ and samples $x_i$.
The generated samples will be concentrated around the higher values of $\| f_W(x)\|_2^2$, and the regularization will have a higher effect on these regions.
In Section~\ref{method}, we detail these two methods, and describe the algorithms we use in order to integrate them into gradient descent. In Section~\ref{experiment}, we describe experiments using MNIST and CIFAR databases, and report the results. Finally, in Section~\ref{discussion}, the results are discussed.
\section{Methods and algorithms : Function norm regularization}
\label{method}
\subsection{Function norm with respect to a probability measure}
For this first method, our function norm, written another way is
\begin{equation}
\| f_W\|_{L_2(\mathbb{R}^d, dP)}^2 = \mathbb{E}_{x \sim P(x)} \left[ \langle f_W(x), f_W(x) \rangle \right]
\end{equation}
where $d$ is the dimension of $x$ the inputs of the network, $\| .\|_{L_2(\mathbb{R}^d, dP)}$ denotes the $L_2$ norm with respect to the probability measure on domain $\mathbb{R}^d$. For some set $\{x_1, \dots ,x_m\}$ drawn i.i.d.\ from $P(x)$
\begin{equation}
\frac{1}{m} \sum_{i=1}^m \langle f_W(x_i), f_W(x_i) \rangle
\end{equation}
is an unbiased estimate. For samples outside the training set, this is a $U$-statistic of order 1, and has an asymptotic Gaussian distribution for which finite sample estimates converge quickly\cite{lee1990u}.
For these samples, only their images by $f_W$ are needed and not their labels, which is particularly interesting for applications such as medical imaging
where labeling the samples require expert intervention and is usually highly expensive.
Therefore, the problem we aim to solve can be written as following (using \eqref{eq:proba_x}):
\begin{align}
\arg\min_{W} \ & \mathbb{E}_{(x,y)\sim P(x,y)}[\ell(f_W(x),y)] + \lambda \mathbb{E}_{x\sim P_x}[ \langle f_W(x), f_W(x) \rangle]\nonumber \\ &\rightarrow \arg\min_{W} \int \int \ell(f_W(x),y) P(x,y) dxdy + \lambda \int \langle f_W(x), f_W(x) \rangle P(x) dx \nonumber \\ &\rightarrow \arg\min_{W}\int \int \ell(f_W(x),y) + \lambda \langle f_W(x), f_W(x) \rangle P(x,y) dxdy \nonumber \\ &\rightarrow \arg\min_{W} \underbrace{\mathbb{E}_{(x,y)\sim P(x,y)} [\ell(f_W(x),y) + \lambda \langle f_W(x), f_W(x) \rangle]}_{=: C_1(W)}
\end{align}
\subsection{Markov-chain sampling}
Our second idea is to use stochastic integration methods to approximate the integral of $\| f_W(x)\|_2^2$. A simple way to apply these methods is to draw samples from the uniform distribution in the domain of the function and then to approximate the integral by the average value of the function in these points. This type of methods is known as Monte-Carlo and Quasi-Monte-Carlo methods~\cite{caflisch1998monte}.
However, as we are working in a very high dimension, this type of numerical integration is not appropriate because it suffers from the \textit{curse of dimensionality} and results in a sample concentration independent of the characteristics of the function we want to integrate, i.e. $\| f_W(x)\|_2^2$. As this function is real and non-negative, we can consider sampling from a distribution which density is proportional to the function itself, using Markov-chain methods such as Gibbs~\cite{casella1992explaining} and Metropolis–Hastings~\cite{hastings1970monte} sampling. As Gibbs sampling requires to approximate the marginal distribution (which is not easily accessible in our case and most importantly not standard), and Metropolis-Hastings sampling is depending on the choice of the proposal distribution, we generate our samples using slice sampling \cite{neal2003}.
This method draw samples uniformly from the hypervolume under the n-dimensional graph of the function by iterating the following steps (to sample from a function $f$):
\begin{inparaenum}[(i)]
\item Choose an initial sample $x_0$;
\item Draw a uniformly distributed sample $y \in (0, f(x_0))$;
\item Find an hyperrectangle, around $x_0$, that is contained in the slice $S =\lbrace x : f(x)>f(x_0)\rbrace$;
\item Draw the new samples $x_1$ uniformly from this hyperrectangle;
\item Go back to the second step and iterate until collecting the wanted number of samples.
\end{inparaenum}
In~\cite{neal2003}, many methods to define the hyperrectangle in the third step are proposed. Here we used the method called "stepping-out and shrinking-in". This method consist in finding a neighborhood of $x_0$, increasing it progressively until being outside the slice, and drawing samples from this increased neighborhood. If the drawn sample is outside the slice, it is used to shrink the neighborhood.
The samples that are drawn using this procedure will have a concentration that depends on the variations of $\| f_W(x)\|_2^2$. They will be concentrated around the regions where its values are high. The effect of the \textit{curse of dimensionality} is then limited. Moreover, this means that the regularization will have a higher effect on the regions where the samples are more concentrated, i.e.\ the regions where the values of the function are high, resulting in a less complex function. The integral will then be approximated by
$\frac{1}{M}\sum_{i=1}^M \| f_W(x_i)\|_2^2$,
which is an unbiased estimate of
$\mathbb{E}_{x\sim \tilde{P}(x)}[\| f_W(x)\|_2^2]$,
where $\tilde{P}(x)$ is proportional to $\| f_W(x)\|_2^2$. The normalization factor is actually equal to the integral we want to approximate.\footnote{With this distribution, we need to prove that the used expression is a proper norm. The proof will be added soon.}
Now, if we consider any distribution $\overline{P}(y)$, and define $\tilde{P}(x,y) = \tilde{P}(x)\overline{P}(y)$, we can rewrite :
\begin{align}
\mathbb{E}_{x\sim \tilde{P}(x)} \ & [\| f_W(x)\|_2^2] =
\mathbb{E}_{(x,y)\sim \tilde{P}(x,y)} [\| f_W(x)\|_2^2] = \nonumber\\&
\mathbb{E}_{(x,y)\sim P(x,y)} \left[\frac{\tilde{P}(x,y)}{P(x,y)}\| f_W(x)\|_2^2\right] =
\mathbb{E}_{(x,y)\sim P(x,y)} [\alpha(x,y)\| f_W(x)\|_2^2]
\label{eqSlice}
\end{align}
Note that $\mathbb{E}_{(x,y)\sim P(x,y)}[\alpha(x,y)] = 1$, with $P(x,y)$ the observed data distribution.\footnote{This expression is needed only for theoretical development. In practice, we will keep sampling from $\| f_W(x)\|_2^2$.}
Finally, the problem to solve can be written as:
\begin{equation}
\arg\min_W \underbrace{\mathbb{E}_{(x,y)\sim P(x,y)} [\ell(f_W(x),y) + \lambda\alpha(x,y)\| f_W(x)\|_2^2]}_{=: C_2(W)}
\end{equation}
\subsection{Stochastic gradient descent and algorithms}
\subsubsection{On the convergence of stochastic gradient descent}
In general, the stochastic gradient descent to minimize an objective of the form~\cite{bottou1998online}:
\begin{equation}
C(W) = \mathbb{E}_{(x,y) \sim P(x,y)}(Q(x,y,W))
\label{obj}
\end{equation}
is operated with an update of the form:
\begin{equation}
W_{t+1} = W_t - \gamma_t H(\mathbf{x_t,y_t},W_t)
\end{equation}
where
\begin{equation}
\mathbb{E}_{(\mathbf{x,y}) \sim P(\mathbf{x,y})}[H(\mathbf{x,y},W)] = \nabla_WC(W).
\label{eqCondH}
\end{equation}
For both of the cases we consider, the objective to minimize has the same form as~\eqref{obj}. In our methods,
\begin{inparaenum}[(i)]
\item for the first case, $H(\mathbf{x_t,y_t},W_t)$ is replaced by $\nabla_W \left(\ell(f(x_{t}),y_{t}) + \lambda \| f(r_{t})\|_2^2)\right)$ where $r_t$ is not the same as $x_t$ but drawn from the same distribution;
\item for the second case, $H(\mathbf{x_t,y_t},W_t)$ is replaced by $\nabla_W \left(\ell(f(x_{t}),y_{t}) + \lambda \| f(r_{t})\|_2^2)\right)$ where $r_t$ is drawn from a density proportional to $\| f(x))\|_2^2$, or equivalently, according to~\eqref{eqSlice}, $\nabla_W \left(\ell(f(x_{t}),y_{t}) + \lambda \alpha(r_t,z_t)\|f(r_{t})\|_2^2)\right)$ where $(r_t,z_t)$ is drawn from the data distribution.
\end{inparaenum}
If we suppose that we can permute integration and derivation, it is straightforward to show that the two updates obey to the condition \eqref{eqCondH}.
Much research considers the question of the convergence of this procedure. Many of them are based on the results of~\cite{robbins1985convergence}. Nemirovsky \textit{et al}~\cite{nemirovski2009robust} showed that if $C(W)$ is convex, and if its gradient is Lipshitz continuous and bounded, then the optimal convergence rate is $\mathcal{O}(\frac{1}{t})$ and is obtained for $\gamma_t = \mathcal{O}(\frac{1}{t})$. In this case, the speed of decrease of $\gamma_t$ has the best compromise between the speed of convergence and the control of the variance introduced by using one sample rather than all samples to approximate the gradient.
Bottou~\cite{bottou1998online, bottou2010large, bottou2012stochastic} has shown this property holds for less constraining conditions on C. In~\cite{bottou1991stochastic}, he has studied in particular the type of objective that one wants to minimize in neural networks. These objectives are not convex and may admit several minimums. He shows that as long as:
\begin{inparaenum}[(i)]
\item $\exists C_{min}, \forall W, C(W)>C_{min}$,
\item $\mathbb{E}_{(x,y)\sim P(x,y)}[H(x,y,W)] = \nabla_WC(W)$,
\item $\mathbb{E}_{(x,y)\sim P(x,y)}[H(x,y,W)^TH(x,y,W)] \le A + BC(W), A,B \geq 0$,
\item $\sum_t \gamma_t = \infty$ and $\sum_t \gamma_t^2 < \infty,$
\end{inparaenum}
$C(W_t)$ converges and $\nabla_WC(W_t)$ converges to 0, using SGD. We know that the second condition holds for both of the cases we study. Thus, for our methods, knowing that the two other conditions on $C$ and $H$ hold for $C_0(W) = \mathbb{E}_{(x,y) \sim P(x,y)}[\ell(f_W(x),y)]$ and $H_0(x_t,y_t,W_t) = \nabla_W\ell(f_{W_t}(x),y_t)$ (which are the classical settings for neural networks), we need to prove that they still hold after adding the regularization term. Let $C_{min,0}, A_0$ and $B_0$ be the parameters that ensures the first and third conditions for $C_0$ and $H_0$.
\textbf{First case:} In this case, we can rewrite the objective as :
\begin{align}
C_1(W) &= C_0(W) + \lambda \mathbb{E}_{(x,y) \sim P(x,y)}[\| f_W(x)\|_2^2]\nonumber\\ H_1(x_t,y_t,W_t) &= H_0(x_t,y_t,W_t) + \lambda\nabla_W\| f_{W_t}(x_t)\|_2^2
\end{align}
Thus :
\begin{equation}
H_1(x_t,y_t,W_t) = H_0(x_t,y_t,W_t) + 2\lambda J_Wf_{W_t}(x_t)^Tf_{W_t}(x_t).
\end{equation}
Then, as $\lambda \ge 0$ and $\mathbb{E}_{(x,y) \sim P(x,y)}[\| f_W(x)\|_2^2] \ge 0$, it follows that $C_1(W) > C_{min,0}$. In addition, we have:
\begin{align}
\mathbb{E}[H_1(x,y,W)^T&H_1(x,y,W)] = \mathbb{E}[H_0(x,y,W)^TH_0(x,y,W)] \nonumber\\&+ 2\lambda \mathbb{E}[H_0(x,y,W)^TJ_Wf_{W}(x)^Tf_{W}(x) + f_{W}(x)^TJ_Wf_{W}(x)H_0(x,y,W)] \nonumber\\&+ 4\lambda^2 \mathbb{E}[f_{W}(x)^TJ_Wf_{W}(x)J_Wf_{W}(x)^Tf_{W}(x)]
\end{align}
We know that the first term is bounded above by $A_0 + B_0C_0(W)$. Let us suppose that $H_0, f_W$ and $J_Wf_W$ are bounded. This is a reasonable hypothesis as we want to control the $L_2$ norm of $f_W$. Moreover, when $J_Wf_W$ is bounded, the condition on $H_0$ is verified for many popular loss functions such as the cross-entropy loss. All our experiments are conducted with this loss. Thus, we can find $A_1$ that bounds the second term, and $A_2$ such that the third term is bounded by $4\lambda^2A_2\mathbb{E}[f_{W}(x)^Tf_{W}(x)]$. Finally, it suffices to find $\lambda$ such that $4\lambda^2A_2 - B_0$ is positive, which will finish the proof of the third condition for $C_1$ and $H_1$.
\textbf{Second case:} In this case, we can rewrite the objective as :
\begin{align}
C_2(W) &= C_0(W) + \lambda \mathbb{E}_{(x,y) \sim P(x,y)}[\alpha(x,y)\| f_W(x)\|_2^2]\nonumber\\ H_2(x_t,y_t,W_t) &= H_0(x_t,y_t,W_t) + \lambda\nabla_W\alpha(x_t,y_t)\| f_{W_t}(x_t)\|_2^2
\end{align}
Note that $\alpha$ depends also on $W$. However, it is positive and its expected value is equal to 1. Thus, the first condition on $C_0$ holds for $C_2$, and the same sketch of proof for the third condition for $C_1$ holds for $C_2$.
Thus, if we consider a learning rate $\gamma_t = \mathcal{O}(\frac{1}{t})$, it obeys to the fourth condition and ensures the convergence of SGD for both of the considered objectives if $\lambda$ is tuned correctly. In practice, this parameter is tuned empirically.
\subsubsection{Algorithm}
During training, the algorithm~\ref{alg1} is used for each epoch (for $N$ training samples and $M$ regularization samples):\footnote{This algorithm is designed such that the network structure includes $\ell$ as a final layer.} Rather than using the SGD with single samples (online learning), we use a mini-batch approach which is most popular and used by the library we use for programming (MatConvNet). In this algorithm, we denote the network $N_{tr}$ and $N_{reg}$ for training and regularization and we note $N_*.L$ their layers. The training network differs from the regularization network only by its last layer that computes the loss function. $B_S$ denotes the batch size, and $B_N$ the number of patches, $\lambda$ is the regularization variable, ``mode" can be either ``train" or ``test", $R_{tr}$ and $R_{reg}$ contain the output of the network with the same subscript and $e$ designs the error. For the regularization step, backpropagation is initialized with the value of the output of the forward pass. The difference between the two proposed methods is in the definition of the regularization data (see~\eqref{eqChoice}).
\section{Experiments and results}
\label{experiment}
To test our algorithms, two series of experiments using two different databases are considered: MNIST and CIFAR10. The experiments are conducted using MatConvNet \cite{vedaldi15matconvnet}. For both of the databases, the convolutional neural network LeNet is used. The training algorithms provided by the cited library are modified according to the described algorithms. Both of the algorithms are used with weight decay. The results of our algorithms are compared with:
\begin{inparaenum}[(i)]
\item Training with only weight decay
\item Training with DropOut (with different rates) + weight decay.
\end{inparaenum}
More extensive experimental results can be found in the supplementary material, but qualitatively follow the results reported here.
\subsection{MNIST}
In this paragraph, we report the results of the tests conducted with our algorithms on MNIST. The MNIST data base is composed of 60,000 samples for training and 10,000 samples for test. These experiments where conducted using only a 2.4 GHz Intel Core 2 Duo processor.
\begin{center}
\begin{minipage}{.7\linewidth}
\begin{algorithm}[H]
\KwData{$N_{tr}, B_S, B_N, \lambda$, mode}
\KwResult{$N_{tr}$}
$N_{reg}.L = N_{tr}.L$(1:end-1);
$R_{tr} = [\ ], R_{reg} = [\ ], e = [\ ]$\;
$B_{S,tr} = B_S\times\frac{n}{n+m}, B_{S,reg} = B_S - B_{S,tr}$\;
\For{i=1:$B_N$}{
\eIf{mode = 'train'}{
$(X,y)\leftarrow B_{S,tr}$ training samples\;
$R_{tr}$ := Forward + Backward ($N_{tr},X$)\;
$e$ := [$e$, ErrorFunction($R_{tr}$, y)]\;
Accumulate gradients\;
$N_{reg}.L = N_{tr}.L$(1:end-1))\;
\begin{equation}
X \leftarrow \begin{cases}
B_{S,reg} \text{regularization samples
(only Data) if using data distribution;}\;\\
B_{S,reg} \text{samples from slice sampling if using slice sampling}\;
\end{cases}
\label{eqChoice}
\end{equation}\;
$R_{reg}$ := Forward + Backward ($N_{reg},X$)\;
Accumulate gradients for $N_{reg}$ using $\lambda$\;
$N_{tr}.L$(1:end-1)) = $N_{reg}.L$\;
}{
$(X,y)\leftarrow B_S$ test samples\;
$R_{tr}$ := Forward ($N_{tr},X$)\;
$e$ := [$e$, ErrorFunction($R_{tr}$, y)]\;
}
}
\caption{Epoch procedure}
\label{alg1}
\end{algorithm}
\end{minipage}
\end{center}
\subsubsection{Sampling from the data distribution}
In this set of experiments, we test our first algorithm and the two reference methods listed above using a decreasing number of samples of MNIST in the training data set. The remaining training samples are used for regularization.
Figure~\ref{reg-im-MNIST} shows the evolution of Top1 error during training (the x-axis represents the number of epochs) for weight decay, weight decay+ DropOut with the rate 50\%, weight decay+ DropOut with the rate 25\% and approximate function norm using data distribution. This error is displayed for the test samples.
\begin{figure}[ht!]
\centering
\includegraphics[trim = 2.5cm 7.5cm 2cm 7.5cm,clip=true,width=0.3\textwidth]{images/mnist-12000.pdf} \hfill
\includegraphics[trim = 3cm 8.3cm 2.5cm 8.5cm,clip=true,width=0.3\textwidth]{images/mnist-600.pdf} \hfill
\includegraphics[trim = 3cm 8.3cm 2.5cm 8.5cm,clip=true,width=0.3\textwidth]{images/mnist-100.pdf}
\caption{MNIST - Left: 12,000 samples - Center: 600 samples - Right: 100 samples}
\label{reg-im-MNIST}
\end{figure}
\subsubsection{Comparison of the two sampling methods}
\begin{wrapfigure}{R}{9cm}
\centering
\includegraphics[trim = 1cm 7.5cm 1cm 8cm,width=0.5\textwidth]{images/mnist-100-sampling-compare.pdf}
\caption{MNIST - Sampling method comparison - Test Top1-error evolution in time}
\label{mnist-compare}
\end{wrapfigure}
In this test, we aim to compare the two proposed algorithms. We consider for this test the problem of classification using 100 samples.In Figure~\ref{mnist-compare}, we display the test Top1 error evolution in function of the training time. The number (1:N) indicates the ratio (Training samples:Regularization samples). Table~\ref{mnist-correct} shows the accuracy obtained on the test set at the end of the training. As a reference, note that the best accuracy obtained when all the training data is used is 99.06\%.
\begin{table}[ht!]
\centering
\begin{tabular}{|c|c||c|c|c|c|c|}
\hline
Training & Regularization & Weight & WD &WD & Function norm & Function norm\\
& & decay & + DO(25\%)& + DO(50\%) &(data dist.)&(slice sampling)\\
\hline
\hline
12,000 & 48,000 & 98.26 & 97.65 & 95.82 & 97.99 & 97.99 (1:4)\\
1,200 & 58,800 & 92.60 & 89.93 & 85.52 & 95.15 & 92.70 (1:4)\\
600 & 59,400 & 88 & 83.12 & 73.97 & 92.84 & 90.47 (1:10)\\
100 & 59,900 & 30.27 & 18.75 & 13.36 & 80.53 & 79.74 (1:40)\\
\hline
\end{tabular}
\caption{MNIST - Correct classification rate (\%) for different methods}
\label{mnist-correct}
\end{table}
\subsection{CIFAR}
In this paragraph, we report the results of our algorithms when applied to CIFAR 10. The used CIFAR database is composed of 50,000 samples for training and 10,000 samples for test. These experiments where conducted using GPU, on a machine equipped with a 4 core CPU (4.00 GHz) and a GPU GeForce GTX 750 Ti. In this set of experiments, we test our two algorithms, weight decay and weight decay + DropOut(10\% and 50\%) with a decreasing number of samples of CIFAR in the training data set. The remaining training samples are used for regularization. Figure~\ref{reg-im-CIFAR} shows the Top1 error for the test set as a function of the number of epochs using respectively 10,000, 500 and 100 training samples. Figure~\ref{cifar500} shows the evolution of the test Top1-error with the training time using 500 samples for training. Table~\ref{cifar-correct} shows the accuracy obtained on the test set at the end of the training. As a reference, note that the best accuracy obtained when all the training data is used is 80.38\%.
\begin{figure}[ht!]
\includegraphics[trim = 1.5cm 7.5cm 1cm 7.5cm,clip=true,width=0.3\textwidth]{images/cifar-10000.pdf} \hfill
\includegraphics[trim = 1.5cm 7.5cm 1cm 7.5cm,clip=true,width=0.3\textwidth]{images/cifar-500.pdf} \hfill
\includegraphics[trim = 1.5cm 7.5cm 0cm 7cm,clip=true,width=0.3\textwidth]{images/cifar-100.pdf}
\caption{CIFAR - Left: 10,000 samples - Center: 500 samples - Right: 100 samples}
\label{reg-im-CIFAR}
\end{figure}
\begin{table}[ht!]
\centering
\begin{tabular}{|c|c||c|c|c|c|c|}
\hline
Training & Regularization & Weight & WD &WD & Function norm & Function norm\\
& & decay & + DO(10\%)& + DO(50\%) &(data dist.)&(slice sampling)\\
\hline
\hline
10,000 & 40,000 & 71.53 & 69.35 & 40.79 & 71.90& 72.02 (1:4)\\
5,000 & 45,000 & 67.27 & 64.59 & 28.08 & 68.26 & 68.21 (1:4)\\
1,000 & 49,000 & 52.57 & 47.99& 12.03 & 50.75 & 54.38 (1:10)\\
500 & 49,500 & 43.64 & 37.16 & 11.57 & 39.96 & 48.47 (1:10)\\
\hline
\end{tabular}
\caption{CIFAR- Correct classification rate (\%) for different methods}
\label{cifar-correct}
\end{table}
\begin{wrapfigure}{R}{7cm}
\centering
\includegraphics[trim = 1cm 7.5cm 1cm 8cm,width=0.5\textwidth]{images/cifar-500-bis.pdf}
\caption{CIFAR - 500 samples - Test Top1-error evolution in time}
\label{cifar500}
\end{wrapfigure}
\section{Discussion}
\label{discussion}
\subsection{MNIST}
The MNIST experiments show the feasibility and the efficiency of a regularization based on function norm. Indeed, Figure~\ref{reg-im-MNIST} shows that the function norm regularization with sampling from the data distribution behaves at least as well as the state-of-the art methods when a high number of samples is used, and outperforms them when we decrease the number of samples. Figure~\ref{mnist-compare} shows that the two proposed methods converge to similar errors, but with different computation time. Indeed, in this case, the data distribution and the network are simple enough to ensure equally high accuracy for the approximation of the function norm using the data distribution or the slice sampling. Both of the proposed methods resulted in an accuracy as high as 80\% with only 100 training samples.
As mentioned above, for the case of MNIST, the main difference between the sampling methods is the computation time. In this point of view, using the data distribution is the fastest, as it needs only to read data from memory while the slice sampling approach require the generation of new samples. However, these methods are preferred to the first algorithm in the situations where the number of available samples is small, such as in medical imaging, where the data collection itself (even without labeling) may be expensive.
\subsection{CIFAR}
The experiments conducted on CIFAR10 show that our idea of regularization is also feasible for more complicated data and a more complicated network. The curves and the accuracy rates show that the performance of our method is always higher than the state-of-the-art methods (at least for the tested sizes of database). Our methods gave an accuracy close to the best accuracy obtained using 50,000 samples for training, with only 10,000 samples. It is also more robust to the decrease of the number of samples. Even if the obtained results with this method 500 and 100 sample are not optimal, the error is still lower than with the other methods.
Another important observation is that the slice sampling seems more robust than using the data distribution for this database. This may be explained by the fact that CIFAR's distribution is more complicated, and the available samples for regularization may be not enough to capture all the variations of the distribution. However, in all the cases, the slice sampling captures well the characteristics of the function variations.
\section{Conclusion}
In this paper, two methods to introduce function norm regularization in neural network training have been developed. We have demonstrated both theoretically and empirically that their integration into stochastic backpropagation converges. Their efficiency have been showed on the MNIST data set and on the more complicated data of CIFAR10. Our experiments suggest that the developed methods are of high quality on small databases that lie in low dimensional manifolds. For more complicated data with a small number of samples, the results of our methods outperform the state-of-the-art regularization methods available in the literature.
\subsection*{Acknowledgements}
This work is partially funded by Internal Funds
KU Leuven, the European Commission through FP7-MC-CIG
334380, and the Research Foundation - Flanders (FWO) through project number G0A2716N.
\bibliographystyle{abbrv}
|
\section{Introduction}
In 1960, Opial \cite{O} established the following integral inequality
\begin{equation}
\int_0^b|f(x)f'(x)|dx \le \frac{b}{4}\int_0^b |f'(x)|^2dx,
\label{1.1}
\end{equation}
where $f$ is absolutely continuous on $[0,b]$ such that $f(0)=f(b)=0.$ In 1962, Beesack \cite{B} showed the following result which implies \eqref{1.1} and is very useful in applications: If $f$ is absolutely continuous on $[0,b]$ and $f(0)=0$, then
\begin{equation}\label{eqO2}
\int_0^b|f(x)f'(x)|dx \le \frac{b}{2}\int_0^b |f'(x)|^2dx.
\end{equation}
Since then, many generalizations of Opial's inequality \eqref{eqO2} in various directions have been given, one of which was given in 1967 by Godunova and Levin \cite{Go}: Let $F$ be a convex and increasing function on $[0, \infty)$ with $F(0) = 0$, and $f$ be a real-valued absolutely continuous function defined on $[a,b]$ with $f(a) = 0$; then
\begin{equation}\label{eq1.2}
\int_a^bF'(|f(x)|)|f'(x)|dx \le F\bigg( \int_a^b|f'(x)|dx\bigg).
\end{equation}
Later, some multidimentional generalizations of \eqref{eq1.2} were given, such as Pe\v{c}ari\'{c} \cite{pe}, Pachpatte \cite{pach}, and Andri\'{c} et al. \cite{AND}.
In 2015, Duc, Nhan and Xuan \cite{NDX} extended and generalized \eqref{eqO2} to several independent variables and demonstrated the usefulness in the field of partial differential equations:
\begin{equation}\label{eq:}
\begin{split}
&\bigg[ \int_{\Omega}\bigg| \partial^{\boldsymbol{\alpha}} \bigg( \prod^m_{j=1}{ G_j(u_j(\boldsymbol{x})) } \bigg) \bigg|^s K_{\boldsymbol{\alpha}}(\boldsymbol{b}, \boldsymbol{x})\sigma(\boldsymbol{x}) d \boldsymbol{x} \bigg]^{1/s}\\
&\leq C \prod^m_{j =1}{ \bigg[ G_j \bigg( \int_{\Omega } | \partial^{\boldsymbol{\alpha}}u_j(\boldsymbol{x}) |^p K_{\boldsymbol{\alpha}}(\boldsymbol{b}, \boldsymbol{x})\rho_j(\boldsymbol{x}) d \boldsymbol{x} \bigg) \bigg]^{1/p} }, \\
\end{split}
\end{equation}
where $C$ is a constant.
In recent years, the theory of time scales which was introduced by Hilger \cite{Hilger} in order to unify the study of differential and difference equations, has received a lot of attention. The readers may find much of time scales calculus in books by Bohner and Peterson \cite{BP}, \cite{BP1}. One of main subjects of the qualitative analysis on time scales is to prove some new dynamic inequalities. Opial-type inequalities on time scales was first proved by Bohner and Kaymak\c{c}alan \cite{BK} in 2001 (see also \cite{ABP}), in which they showed that if $f: [0, b]_{\mathbb{T}}\to \mathbb{R}$ is delta differentiable with $ f(0) = 0$, then
\begin{equation}
\int_0^b|[f(x)+f^{\sigma}(x)]f^{\Delta}(x) | \Delta x \le b\int_0^b |f^{\Delta}(x)|^2\Delta x.
\label{eq1.1}
\end{equation}
Afterwards, numerous authors have studied variants of \eqref{eq1.1} (see, for example, \cite{Ka}, \cite{Lihan}, \cite{SAK}, \cite{SAK1}, \cite{SORA}, \cite{SIVA}, and \cite{YZ}). The best reference here is the book by Agarwal, O'Regan, and Saker \cite[Chapter 3]{AOS}, where the most popular articles on this subject are collected.
However, to the best of the author knowledge nothing is known regarding Opial-type inequalities involving functions of several variables and their partial derivatives on time scales. Thus, the aim of this paper is to study some weighted integral inequalities for delta derivatives acting on compositions of functions on time scales which are in turn applied to establish multidimentional dynamic Opial-type inequalities. As applications, we establish Lyapunov-type inequalities for half-linear dynamic equations and obtain upper bounds of solutions of certain integro-partial dynamic equations.
\section{Preliminaries}
In the most part we assume the readers were so familiar with basic time scales calculus. More information about time scales calculus can be found in \cite{BP} and \cite{BP1}. In this section, we only present some basic definitions and notations about calculus in several variables on time scales.
Let $\mathbb{T}$ be a time scale, and let $\sigma, \rho,$ and $\Delta$ denote, respectively, the forward jump, backward jump, and delta operator on $\mathbb{T}$.
Fix $n \in \mathbb{N}$ and let $\mathbb{T}_j,$ where $ j =1, ..., n,$ be time scales, and
\[\Lambda^n = \mathbb{T}_1\times \cdot \cdot \cdot\times\mathbb{T}_n = \{ \boldsymbol{x}= (x_1, ..., x_n): x_j \in \mathbb{T}_j \,\,\, \text{for all}\,\,\, j\in [1, n]_{\mathbb{N}}\}\]
be the $n$-dimensional time scale.
For $i \in [1, n]_{\mathbb{N}},$ let $\sigma_j, \rho_j,$ and $\Delta_j$ denote, respectively, the forward jump, backward jump, and delta operator on $\mathbb{T}_j$. We define ${\mathbb T}_j^{\kappa} = {\mathbb T}_j$ if ${\mathbb T}_j$ does not have a left-scattered maximum ${x_j}_{\max}$; otherwise ${\mathbb T}_j^{\kappa} = {\mathbb T}_j\setminus\{{x_j}_{\max}\}$.
The graininess functions $\mu_j:{\mathbb T}_j\to [0,\infty)$ are defined by $\mu_j(x_j)=\sigma_j(x_j)-x_j$ for $j \in [1, n]_{\mathbb{N}}$.
For $x_j \in \mathbb{T}_j, j \in [1, n]_{\mathbb{N}},$ we denote $\rho^2_j(x_j) = \rho_j(\rho_j(x_j))$ and $\rho^k_j(x_j) = \rho_j(\rho^{k-1}_j(x_j))$ for $k \in \mathbb{N}.$ For convenience we put $\rho_j^0(x_j) = x_j$, $ j\in [1, n]_{\mathbb{N}}.$
For $\boldsymbol{x} = (x_1, ..., x_n), \boldsymbol{y}= (y_1, ..., y_n) \in \Lambda^n$, we shall write $\boldsymbol{x}\leq \boldsymbol{y}$ instead of $x_j \leq y_j$ for all $j\in [1,n]_{\mathbb{N}} $. Analogously one has to understand $\boldsymbol{x} = \boldsymbol{y}, \boldsymbol{x} > \boldsymbol{y}$ and $\boldsymbol{x} < \boldsymbol{y}$, respectively. We put $\boldsymbol{x} +\boldsymbol{y} = (x_1+y_1, ..., x_n+y_n)$.
We denote by $\boldsymbol{\lambda}= (\lambda_1, ..., \lambda_n)$ the multi-index, i.e. $\lambda_j \in \mathbb{N}_0 = \mathbb{N}\cup \{0\}, j \in [1,n]_{\mathbb{N}}.$ In particular, let $\boldsymbol{1}=(1, ..., 1)$.
Let $\boldsymbol{a}= (a_1, ..., a_n)$, $ \boldsymbol{b} =(b_1, ..., b_n)$, and ${\boldsymbol\rho}^{\boldsymbol{\lambda}-\boldsymbol{1}}(\boldsymbol{b}) = (\rho^{\lambda_1-1}_1(b_1), ..., \rho^{\lambda_n-1}_n(b_n))$ for $\boldsymbol{\lambda}\geq \boldsymbol{1},$ be in $\Lambda^n$ such that $\boldsymbol{a}< {\boldsymbol \rho}^{\boldsymbol{\lambda}-\boldsymbol{1}}(\boldsymbol{b})$. Then we set
\[ \Omega = \{\boldsymbol{x} \in \Lambda^n: \boldsymbol{a}\leq \boldsymbol{x}\leq \boldsymbol{b} \}, \]
\[ \Omega^{\kappa^{\boldsymbol{\lambda-1}}} = \{\boldsymbol{x} \in \Lambda^n: \boldsymbol{a}\leq \boldsymbol{x}\leq {\boldsymbol\rho}^{\boldsymbol{\lambda}-\boldsymbol{1}}(\boldsymbol{b}) \}, \]
\[\Omega_{\boldsymbol{x}} =\{\boldsymbol{t} \in \Lambda^n: \boldsymbol{a}\leq \boldsymbol{t}\leq \boldsymbol{x} \}, \quad \boldsymbol{x} \in \Omega, \]
\[\bar{\Omega}_{\boldsymbol{x}} =\{\boldsymbol{t} \in \Lambda^n: \boldsymbol{x}\leq \boldsymbol{t}\leq {\boldsymbol \rho}^{\boldsymbol{\lambda}-\boldsymbol{1}}(\boldsymbol{b}) \}, \quad \boldsymbol{x} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}, \]
\[ \Omega' =[a_2, b_2]_{\mathbb{T}_2} \times \cdot \cdot \cdot \times [a_n, b_n]_{\mathbb{T}_n}. \]
For any real-valued rd-continuous function $f$ defined on $\Omega$ we denote by $\int_{\Omega}f(\boldsymbol{x}) \Delta \boldsymbol{x}$, $\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}f(\boldsymbol{x}) \Delta \boldsymbol{x}$,
$\int_{\Omega_{\boldsymbol{x}}}f(\boldsymbol{t}) \Delta \boldsymbol{t}$ for any $\boldsymbol{x} \in \Omega,$
$\int_{\bar{\Omega}_{\boldsymbol{x}}}f(\boldsymbol{t}) \Delta \boldsymbol{t}$ for any $\boldsymbol{x} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}},$ and
$\int_{\Omega'}f(\boldsymbol{x'}) \Delta \boldsymbol{x'}$ for $\boldsymbol{x'} \in \Omega'$, are
$n-$fold integrals $\int^{b_1}_{a_1}\cdot \cdot \cdot \int^{b_n}_{a_n} f(x_1, ..., x_n)\Delta x_1 \cdot \cdot \cdot \Delta x_n,$
$\int^{\rho^{\lambda_1-1}_1(b_1) }_{a_1}\cdot \cdot \cdot \int^{\rho^{\lambda_n-1}_n(b_n) }_{a_n} f(x_1, ..., x_n)\Delta x_1 \cdot \cdot \cdot \Delta x_n,$
$ \int^{x_1}_{a_1}\cdot \cdot \cdot\int^{x_n}_{a_n} f(t_1, ..., t_n)\Delta t_1 \cdot \cdot \cdot \Delta t_n,$
$ \int^{ \rho^{\lambda_1-1}_1(b_1) }_{x_1}\cdot \cdot \cdot\int^{\rho^{\lambda_n-1}_n(b_n) }_{x_n} f(t_1, ..., t_n)\Delta t_1 \cdot \cdot \cdot \Delta t_n$, and
$(n-1)-$fold integral $ \int^{b_2}_{a_2}\cdot \cdot \cdot\int^{b_n}_{a_n} f(x_2, ..., x_n)\Delta x_2\cdot \cdot \cdot \Delta x_n$,
respectively.
Let $f: \Lambda^n \to \mathbb{R}$. The $partial$ $delta$ $derivative$ of $f$ with respect to $x_j \in \mathbb{T}^\kappa_j$ is defined as the limit
\[\lim_{\underset{ t_j \neq \sigma_j(x_j)}{t_j \to x_j}}\frac{f(x_1, ..., \sigma_j(x_j), ..., x_n) -f(x_1, ..., t_j, ..., x_n) }{\sigma_j(x_j)-t_j} \]
provided that this limit exists as a finite number, and is denoted by $ \frac{\partial f(\boldsymbol{x})}{\Delta_j x_j}. $
If $f$ has partial derivatives $\frac{\partial f(\boldsymbol{x})}{\Delta_jx_j}, j \in [1,n]_{\mathbb{N}},$ then we can also consider their partial delta derivatives. These are called $second$ $order$ partial delta derivatives. We write
\[
\frac{\partial^2 f(\boldsymbol{x})}{\Delta_j x^2_j}=\frac{\partial }{\Delta_j x_j}\bigg(\frac{\partial f(\boldsymbol{x})}{\Delta_j x_j}\bigg), \quad \frac{\partial^2 f(\boldsymbol{x})}{\Delta_j x_j \Delta_i x_i} =\frac{\partial }{\Delta_j x_j}\bigg(\frac{\partial f(\boldsymbol{x})}{\Delta_i x_i}\bigg).
\]
Higher order partial delta derivatives are similarly defined.
Let $\boldsymbol{\lambda}\geq \boldsymbol{1}$ be a multi-index, we denote by $|\boldsymbol{\lambda}| = \lambda_1+ \cdot \cdot \cdot+ \lambda_n$; then we set
\[ \frac{\partial^{\boldsymbol{\lambda}}f(\boldsymbol{x}) }{\Delta {\boldsymbol{x}}^{\boldsymbol{\lambda}}}=\frac{\partial^{|\boldsymbol{\lambda}|}f(\boldsymbol{x}) }{\Delta_1x^{\lambda_1}_1 \cdot \cdot \cdot\Delta_nx^{\lambda_n}_n }.\]
By $C^{n\boldsymbol{\lambda}}_{\text{rd}}(\Omega),$ we denote the set of all functions $f: \Omega \to \mathbb{R}$ which have rd-continuous derivatives $\frac{\partial ^{k_1+ \cdot \cdot \cdot+ k_j}f(\boldsymbol{x})}{\Delta_1 x^{k_1}_1 \cdot \cdot \cdot \Delta_j x^{k_j}_j }$ for $k_j \in [1, \lambda_j]_{\mathbb{N}}$, $j \in [1, n]_{\mathbb{N}}$.
A function $\tau : \Omega \to \mathbb{R}$ is said to be a weight on $\Omega$ if $\tau$ is positive-valued and rd-continuous on $\Omega.$ Let us denote by $\mathcal{W}(\Omega)$ the set of all weights on $\Omega$.
Let $p\geq1$ and $\tau \in \mathcal{W}(\Omega)$. We represent by $\mathcal{L}^p_{\boldsymbol{a}}(\Omega, \tau, \boldsymbol{\lambda})$
the set of all functions $f: \Omega \to \mathbb{R}$ of class $C^{n\boldsymbol{\lambda}}_{\text{rd}}(\Omega)$ for which
$\frac{\partial^{k_j} f(\boldsymbol{x})}{\Delta_jx^{k_j}_j}|_{x_j =a_j} = 0$ for $k_j \in [0, \lambda_j-1]_{\mathbb{N}}, j \in[1, n]_{\mathbb{N}},$ and that $\int_{\Omega}|\frac{\partial^{\boldsymbol{\lambda}} f(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}|^{p}\tau(\boldsymbol{x})\Delta \boldsymbol{x} <\infty.$
We set \[H_{\boldsymbol{\lambda}}(\boldsymbol{x},\boldsymbol{t}) = \prod^n_{j=1}{h^{(j)}_{\lambda_j-1}(x_j,\sigma_j(t_j))}, \quad \boldsymbol{x}, \boldsymbol{t} \in \Omega,\]
where $h^{(j)}_{k}: \mathbb{T}^2_j \to \mathbb{R},$ $k \in \mathbb{N}_0,$ is such that
$h^{(j)}_{0}(t,s)\equiv 1$ for all $t, s\in \mathbb T_j,$ and $h^{(j)}_{k+1}(t,s)=\int_s^t h^{(j)}_{k}(u,s)\Delta u $ for all $ t, s\in \mathbb T_j, k \in \mathbb{N}_0. $
Let $m$ be a positive integer and $0 <R\leq \infty$. We represent by $\mathcal{H}^m_R$ the set of all functions $F : (-R, R)^m \to \mathbb{R}$ such that
\begin{enumerate}
\item $F\in C^1((-R, R)^m)$,
\item $F(0, ..., 0) = 0$, and
\item $D_iF$ for $i \in [1, m]_{\mathbb{N}}$, are non-negative and increasing in each variable on $(0, R)$,
where $D_i = \partial/\partial{t_i}, i \in [1, m]_{\mathbb{N}}$ for all $(t_1, ..., t_m) \in (-R, R)^m.$
\end{enumerate}
We give a preliminary lemma that we shall use in Section $3$.
\begin{lemma}\label{chainp}
Let $F \in \mathcal{H}^m_R$ and $g_i: \mathbb{T} \to [0, R)$ for $i \in [1, m]_{\mathbb{N}},$ are delta differentiable on $\mathbb{T}^{\kappa}$ such that $g^\Delta_i$ are non-negative on $\mathbb{T}^{\kappa}$; then the composite function $ F(g_1(x), ..., g_m(x))$ is delta differentiable on $ \mathbb{T}^\kappa$ such that
\begin{equation}\label{chainnew}
[F(g_1(x), ..., g_m(x))]^\Delta \geq \sum^m_{i=1}{D_i F(g_1(x), ..., g_m(x))g_i^{\Delta}(x)}, \quad x \in \mathbb{T}^{\kappa}.
\end{equation}
\end{lemma}
\begin{proof}
Fix $x \in \mathbb{T}^\kappa$ and put $ t_i = g_i(x)$, $t^\sigma_i = g_i(\sigma(x))$, we see that $t_i \leq t^\sigma_i$ for $i \in [1, m]_{\mathbb{N}}.$ We have two cases.
\textit{Case 1.} Suppose that $x <\sigma(x).$ Then
\[
[F(t_1, ..., t_m)]^\Delta = \frac{F(t^\sigma_1, ..., t^\sigma_m)-F(t_1, ..., t_m)}{\sigma(x)-x}.\]
For all $i \in [1, m]_{\mathbb{N}}$, we set
\[ A_i=
\begin{cases}
\dfrac{F(t_1, ..., t^\sigma_i, ..., t^\sigma_m)-F(t_1, ..., t_i, t_{i+1}^\sigma, ..., t_m^\sigma)}{t^\sigma_i-t_i} & \text{if}\quad t_i < t^\sigma_i, \\
0& \text{if}\quad t_i = t^\sigma_i.
\end{cases}
\]
Then,
\[ [F(t_1, ..., t_m)]^\Delta=\sum^m_{i=1} A_i \frac{g_i(\sigma(x))-g_i(x)}{\sigma(x) -x} = \sum^m_{i=1} A_i g^\Delta_i(x).
\]
If $t_i = t^\sigma_i$, then $g^\Delta_i(x) = 0$; if $t_j < t^\sigma_j$ with $j \neq i$, then
\[ A_j=\frac{F(t_1, ..., t^\sigma_j, ..., t^\sigma_m)-F(t_1, ..., t_j, t_{j+1}^\sigma, ..., t_m^\sigma)}{t^\sigma_j-t_j} =D_jF(t_1, ..., c_j, t^\sigma_{j+1}, ..., t^\sigma_m), \]
by the mean value theorem, where $c_j \in (t_j, t^\sigma_j)$. Since $D_jF$, $i \in [1, m]_{\mathbb{N}}$, are increasing in each variable on $(0,R)$, we have
\[D_jF(t_1, ..., c_j, t^\sigma_{j+1}, ..., t^\sigma_{m}) \geq D_jF(t_1, ..., t_m) = D_j F(g_1(x), ..., g_m(x)),\]
which yields \eqref{chainnew}.
\textit{ Case 2.} Suppose that $x = \sigma(x)$. We set
\[ F_i(g(s)) :=F(g_1(x), ..., g_{i-1}(x), g_i(s), g_{i+1}(s), ..., g_m(s)), \]
\[ F_i(g(x)) :=F(g_1(x), ..., g_{i-1}(x), g_i(x), g_{i+1}(s), ..., g_m(s)), \quad i \in [1, m]_{\mathbb{N}}. \]
We have
\[ [F(g_1(x), ..., g_m(x))]^\Delta = \lim_{s \to x}\frac{F(g_1(s), ..., g_m(s))-F(g_1(x), ..., g_m(x))}{s-x}, \]
where
\[\frac{F(g_1(s), ..., g_m(s))-F(g_1(x), ..., g_m(x))}{s-x} = \sum^m_{i =1} {\frac{g_i(s)-g_i(x)}{s-x} \frac{F_i(g(s)) -F_i(g(x))}{g_i(s)-g_i(x)} }. \]
By the mean value theorem, there exist $\xi_i(s), i \in [1, m]_{\mathbb{N}}$, which are between $g_i(x)$ and $g_i(s)$, such that
\[\frac{F_i(g(s)) -F_i(g(x))}{g_i(s)-g_i(x)} = D_iF(g_1(x), ..., g_{i-1}(x), \xi_i(s), g_{i+1}(s), ..., g_m(s)).\]
Since $g_i$ for $i \in [1, m]_{\mathbb{N}},$ are delta differentiable on $\mathbb{T}^\kappa$, then $g_i$ for $i \in [1, m]_{\mathbb{N}},$ are continuous at $x$. Therefore, $\lim_{s \to x}{\xi_i(s)} = g_i(x)$ and $\lim_{s \to x}{g_i(s)} = g_i(x)$ for $i \in [1, m]_{\mathbb{N}},$ which gives us the desired result. \qed
\end{proof}
\section{Integral inequalities on time scales}
\begin{theorem}\label{THEO1}
Let $ F \in \mathcal{H}^m_R$ for $0 <R\leq \infty$, and $f_i : \Omega \to (-R, R)$ which satisfies $\int_{\Omega}|\frac{\partial^{ \boldsymbol{1} }f_i(\textbf{x})}{\Delta \textbf{x}^{ \boldsymbol{1} }}|\Delta \textbf{x}<R$ for $i \in [1, m]_{\mathbb{N}}$. If $f_i \in \mathcal{L}^1_{\textbf{a}}(\Omega, 1, \boldsymbol{1})$ for all $i \in [1,m]_{\mathbb{N}},$ then
\begin{equation}\label{ptTHEO1}
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{D_iF(|f_1(\textbf{x})|, ..., |f_m(\textbf{x})|)\bigg|\frac{\partial^{\boldsymbol{1}}f_i(\textbf{x})}{\Delta \textbf{x}^{ \boldsymbol{1} } }\bigg|}\bigg) \Delta \textbf{x} \\
&\leq F\bigg(\int_{\Omega}\bigg|\frac{\partial^{ \boldsymbol{1} }f_1(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1} } }\bigg|\Delta \textbf{x}, ..., \int_{\Omega}\bigg|\frac{\partial^{ \boldsymbol{1} }f_m(\textbf{x})}{\Delta \textbf{x}^{ \boldsymbol{1} }}\bigg|\Delta \textbf{x} \bigg). \\
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
Since $f_i \in \mathcal{L}^1_{\boldsymbol{a}}(\Omega, 1, \boldsymbol{1})$, it follows that
\begin{equation}\label{eqTHEO1}
f_i(\boldsymbol{x}) = \int_{\Omega_{\boldsymbol{x}}}\frac{\partial^{\boldsymbol{1}}f_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}} } \Delta \boldsymbol{t}
\end{equation}
for all $\boldsymbol{x} \in \Omega$ and all $i \in [1, m]_{\mathbb{N}}$.
Let
\[ g_i(x_1): =\int^{x_1}_{a_1}\int_{\Omega'}\bigg|\frac{\partial^{\boldsymbol{1}}f_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}}\bigg| \Delta \boldsymbol{t} \quad \text{for} \quad x_1 \in [a_1, b_1]_{\mathbb{T}}, \quad i \in [1, m]_{\mathbb{N}}. \]
We see that $|f_i(\boldsymbol{x})| \leq g_i(x_1)$ for $\boldsymbol{x} \in \Omega$ and functions $g_i, i \in [1, m]_{\mathbb{N}},$ are increasing on $[a_1, b_1]_{\mathbb{T}}$. By $F \in \mathcal{H}^m_R$, we obtain
\[
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{D_iF(|f_1(\boldsymbol{x})|, ..., |f_m(\boldsymbol{x})|)\bigg|\frac{\partial^{\boldsymbol{1}}f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}\bigg|}\bigg) \Delta \boldsymbol{x}\\
& \leq \int_{\Omega}\bigg(\sum^m_{i=1}{D_iF(g_1(x_1)), ..., g_m(x_1))\bigg|\frac{\partial^{\boldsymbol{1}}f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}\bigg|}\bigg) \Delta \boldsymbol{x}\\
& \leq \int^{b_1}_{a_1}\bigg(\sum^m_{i=1}{D_iF(g_1(x_1), ..., g_m(x_1))\int_{\Omega'}{\bigg|\frac{\partial^{\boldsymbol{1}}f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}\bigg|}\Delta \boldsymbol{x'}}\bigg) \Delta x_1\\
& \leq \int^{b_1}_{a_1}\bigg(\sum^m_{i=1}{D_iF(g_1(x_1), ..., g_m(x_1))\frac{\partial g_i(x_1)}{\Delta_1x_1}\bigg)} \Delta x_1,\\
\end{split}
\]
which, in view of Lemma \ref{chainp}, yields
\[
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{D_iF(|f_1(\boldsymbol{x})|, ..., |f_m(\boldsymbol{x})|)\bigg|\frac{\partial^{\boldsymbol{1}}f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}\bigg|}\bigg) \Delta \boldsymbol{x} \\
& \leq \int^{b_1}_{a_1}F^{\Delta_1}(g_1(x_1), ..., g_m(x_1))\Delta x_1\\
& =F\bigg(\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{1}}f_1(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}\bigg|\Delta \boldsymbol{x}, ..., \int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{1}}f_m(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}\bigg|\Delta \boldsymbol{x} \bigg), \\
\end{split}
\]
which completes the proof. \qed
\end{proof}
\begin{remark} From Theorem \ref{THEO1} we can obtain many known results.
\begin{enumerate}
\item If $\mathbb{T} = \mathbb{R}$, then Theorem \ref{THEO1} becomes \cite[Theorem 1]{BPe} which was established by Brneti\'{c} and Pe\u{c}ari\'{c}.
\item If $\mathbb{T} = \mathbb{R}$, $n = m = 1,$ and $F$ is convex on $[0, \infty)$, then inequality \eqref{ptTHEO1} reduces to \eqref{eq1.2}.
\item Let $\mathbb{T} = \mathbb{R}$, $n=1$, and $F(x_1, ..., x_m) = |x_1 \cdot \cdot \cdot x_m|$; then inequality \eqref{ptTHEO1} becomes \cite[Theorem 1]{pach}.
\end{enumerate}
\end{remark}
The following theorem is a generalization of Theorem \ref{THEO1}.
\begin{theorem}\label{THEO3}
Let $ F \in \mathcal{H}^m_R$ for $0 <R\leq \infty$, and $f_i : \Omega \to (-R, R)$ which satisfies $ H_{\boldsymbol{\lambda}}(\textbf{b},\textbf{a})\int_{\Omega}|\frac{\partial^{ \boldsymbol{\lambda} }f_i(\textbf{x})}{\Delta \textbf{x}^{ \boldsymbol{1} }}|\Delta \textbf{x}<R$ for $i \in [1, m]_{\mathbb{N}}$. If $f_i \in \mathcal{L}^1_{\textbf{a}}(\Omega, 1, \boldsymbol{\lambda})$ for all $i \in [1,m]_{\mathbb{N}},$ then
\begin{equation}\label{ptTHEO3}
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{D_iF(|f_1(\textbf{x})|, ..., |f_m(\textbf{x})|)\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{x})}{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|}\bigg) \Delta \textbf{x} \\
&\leq \frac{1}{H_{\boldsymbol{\lambda}}(\textbf{b},\textbf{a})}F\bigg(H_{\boldsymbol{\lambda}}(\textbf{b},\textbf{a})\int_{\Omega} \bigg|\frac{\partial^{\boldsymbol{\lambda}}f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|\Delta \textbf{x}, ...,H_{\boldsymbol{\lambda}}(\textbf{b}, \textbf{a})\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_m(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|\Delta \textbf{x} \bigg). \\
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
By $f_i \in \mathcal{L}^1_{\boldsymbol{a}}(\Omega, 1, \boldsymbol{\lambda}), i \in [1,m]_{\mathbb{N}},$ and Taylor's formula \cite{Hig}, we have
\[
f_i(\boldsymbol{x}) = \int_{\Omega_{\boldsymbol{x}}} H_{\boldsymbol{\lambda}}(\boldsymbol{x},\boldsymbol{t})\frac{\partial^{\boldsymbol{\lambda}} f_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}} \Delta \boldsymbol{t}, \quad \boldsymbol{x} \in \Omega.
\]
Therefore,
\[
|f_i(\boldsymbol{x})| \leq \int_{\Omega_{\boldsymbol{x}}} H_{\boldsymbol{\lambda}}(\boldsymbol{x},\boldsymbol{t})\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}}\bigg| \Delta \boldsymbol{t} \leq H_{\boldsymbol{\lambda}}(\boldsymbol{b},\boldsymbol{a}) \int_{\Omega_{\boldsymbol{x}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}}\bigg| \Delta \boldsymbol{t}, \quad \boldsymbol{x} \in \Omega.
\]
Now, we define
\[ u_i(\boldsymbol{x}): =H_{\boldsymbol{\lambda}}(\boldsymbol{b},\boldsymbol{a}) \int_{\Omega_{\boldsymbol{x}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}}\bigg| \Delta \boldsymbol{t}, \quad \boldsymbol{x}\in \Omega, \quad i \in [1, m]_{\mathbb{N}}, \]
we see that $ |f_i(\boldsymbol{x})| \leq u_i(\boldsymbol{x})$ for $\boldsymbol{x} \in \Omega$ and $u_i(\boldsymbol{x})$ are increasing in each variable and
\[
\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}\bigg| = \frac{1}{H_{\boldsymbol{\lambda}}(\boldsymbol{b},\boldsymbol{a})} \frac{\partial^{\boldsymbol{1}}u_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1}}}, \quad \boldsymbol{x} \in \Omega, \quad i \in [1, m]_{\mathbb{N}}.
\]
Since $D_iF$ for $ i \in [1, m]_{\mathbb{N}}$, are non-negative, continuous and increasing in each variable on $(0, R)$, we have
\[
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{D_iF(|f_1(\boldsymbol{x})|, ..., |f_m(\boldsymbol{x})|)\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}\bigg|}\bigg) \Delta \boldsymbol{x}\\
&\leq \frac{1}{H_{\boldsymbol{\lambda}}(\boldsymbol{b},\boldsymbol{a})} \int_{\Omega}\bigg(\sum^m_{i=1}{D_iF(u_1(\boldsymbol{x}), ..., u_m(\boldsymbol{x}))\frac{\partial^{\boldsymbol{1}} u_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}}\bigg) \Delta \boldsymbol{x}, \\
\end{split}
\]
which, in view of \eqref{ptTHEO1}, gives \eqref{ptTHEO3}. \qed
\end{proof}
\begin{remark}
Let $\mathbb{T} = \mathbb{R}$; then Theorem \ref{THEO3} becomes \cite[Theorem 2.1]{AND} which was established by Andri\'{c}.
\end{remark}
Next, for $0<R\le \infty$, we denote by $\mathcal{G}^{1, m}_R$ the class of all functions $G: (-R, R)^m \to \mathbb{R}$ satisfying the following conditions:
\begin{enumerate}
\item[(i)] $G \in C^1((-R, R)^m)$,
\item[(ii)] $G(0, ..., 0)=0$, and
\item[(iii)] if $x_i\leq y_i^{1/p}z_i^{1/q}, 0<x_i, y_i, z_i<R$ for $ i \in [1, m]_{\mathbb{N}},$ then
$0\leq D_iG(x_1, ..., x_m)\leq [D_iG(y_1, ..., y_m)]^{1/p}[D_iG(z_1, ..., z_m)]^{1/q}$, where $p$, $q$ are conjugate exponents
$1/p + 1/q = 1$.
\end{enumerate}
\begin{remark}
If $G \in \mathcal{G}^{1, m}_R$, then $G \in \mathcal{H}^m_R.$
\end{remark}
\begin{proof}
Let $G \in \mathcal{G}^{1, m}_R$. For each $i \in [1, m]_{\mathbb{N}}$ and $0 <x_i \leq y_i <R,$ from $(iii)$ we have
\[
\begin{split}
0\leq D_iG(x_1, ..., x_i, ..., x_m)&\leq [D_iG(y_1, ..., y_i, ..., y_m)]^{1/p}[D_iG(y_1, ..., y_i, ..., y_m)]^{1/q}\\
&=D_iG(y_1, ..., y_i, ..., y_m). \\
\end{split}
\]
Therefore, $G \in \mathcal{H}^m_R.$
\end{proof}
\begin{example}
The functions $G(x_1, ..., x_m) = |x_1|^{\gamma_1}\sign(x_1)\cdot \cdot \cdot|x_m|^{\gamma_m}\sign(x_m)$, $H(x_1, ..., x_m) = |x_1|^{\gamma_1}+ \cdot \cdot \cdot +|x_m|^{\gamma_m}$ for $\gamma_i \geq 1$ for all $i \in [1, m]_{\mathbb{N}}$ are in $\mathcal{G}^{1, m}_{\infty}$.
\end{example}
From on now, we always assume that $\alpha, \beta>0$ and $\alpha + \beta >1$ and $G \in \mathcal{G}^{1, m}_R$. We have the following result.
\begin{theorem}\label{THEO4}
Let $\omega_i, \tau_i, \in \mathcal{W}(\Omega)$, and $f_i : \Omega \to (-R, R)$ be such that
\[\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_i(\textbf{x})\Delta \textbf{x}<R \quad \text{for} \quad i \in [1,m]_{\mathbb{N}}. \]
If $f_i \in \mathcal{L}^{\alpha+\beta}_{\textbf{a}}(\Omega, \tau_i, \boldsymbol{\lambda})$ for all $ i \in [1, m]_{\mathbb{N}}$ and
\begin{equation}\label{k2}
K_{\Omega}:=\bigg[\int_{\Omega}\bigg(\sum^m_{i=1}{[D_iG(V_1(\textbf{x}), ..., V_m(\textbf{x}))]^{{\frac{\alpha(\alpha+\beta-1)}{\beta}}}\omega^{\frac{\alpha+\beta}{\beta}}_i(\textbf{x})\tau^{-\frac{\alpha}{\beta}}_i(\textbf{x}) }\bigg)\Delta \textbf{x}\bigg]^{\frac{\beta}{\alpha+\beta}}<\infty,
\end{equation}
where
\[V_i(\textbf{x}): =\int_{\Omega_\textbf{x}} \bigg(H_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t})\bigg)^{\frac{\alpha+\beta}{\alpha+\beta-1}}(\tau_i(\textbf{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \textbf{t} \]
for $\textbf{ x }\in \Omega, i \in [1, m]_{\mathbb{N}}$, then
\begin{equation}\label{ptTHEO4}
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{[D_iG(|f_1(\textbf{x})|, ..., |f_m(\textbf{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol\lambda}}\bigg|^\alpha\omega_i(\textbf{x})}\bigg) \Delta \textbf{x} \\
&\leq K_{\Omega} \bigg[G\bigg(\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_1(\textbf{x})\Delta \textbf{x}, ..., \int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_m(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}. \\
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
As in the proof of Theorem \ref{THEO3}, for any $\boldsymbol{x} \in \Omega$ and all $i \in [1, m]_{\mathbb{N}}$, we have
\begin{equation}\label{eq36}
|f_i(\boldsymbol{x})| \leq \int_{\Omega_{\boldsymbol{x}}} H_{\boldsymbol{\lambda}}(\boldsymbol{x},\boldsymbol{t})\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}}\bigg| \Delta \boldsymbol{t}.
\end{equation}
Applying H\"{o}lder's inequality with indices $(\alpha+\beta)/(\alpha+\beta-1)$ and $(\alpha+\beta)$ to \eqref{eq36}, we get
\begin{align}
\notag|f_i(\boldsymbol{x})| &\leq \bigg(\int_{\Omega_{\boldsymbol{x}}} \bigg(H_{\boldsymbol{\lambda}}(\boldsymbol{x},\boldsymbol{t})\bigg)^{\frac{\alpha+\beta}{\alpha+\beta-1}}(\tau_i(\boldsymbol{t}))^{\frac{1}{1-\alpha-\beta}} \Delta \boldsymbol{t}\bigg)^{\frac{\alpha+\beta-1}{\alpha+\beta}} \\
\notag&\quad \times \bigg(\int_{\Omega_{\boldsymbol{x}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta} \tau_i(\boldsymbol{t}) \Delta t\bigg)^{\frac{1}{\alpha+\beta}}\\
&=: \bigg(V_i(\boldsymbol{x})\bigg)^{\frac{\alpha+\beta-1}{\alpha+\beta}} \bigg(U_i(\boldsymbol{x})\bigg)^{\frac{1}{\alpha+\beta}},
\label{eqfa}
\end{align}
where
\[U_i(\boldsymbol{x}) =\int_{\Omega_{\boldsymbol{x}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_i(\boldsymbol{t}) \Delta \boldsymbol{t}, \quad \boldsymbol{x} \in \Omega, \quad i \in [1, m]_{\mathbb{N}}.\]
Thus, since $G\in \mathcal{G}^{1, m}_R$ from \eqref{eqfa} we obtain
\[
\begin{split}
D_iG(|f_1(\boldsymbol{x})|, ..., |f_m(\boldsymbol{x})|) \leq [&D_iG(V_1(\boldsymbol{x}), ..., V_m(\boldsymbol{x}) )]^{\frac{\alpha+\beta-1}{\alpha+\beta}}\\
& \times [D_iG(U_1(\boldsymbol{x}), ..., U_m(\boldsymbol{x}))]^{\frac{1}{\alpha+\beta}};\\
\end{split}\]
hence
\begin{equation}\label{eq37}
\begin{split}
&\sum^m_{i=1}{[D_iG(|f_1(\boldsymbol{x})|, ..., |f_m(\boldsymbol{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega_i(\boldsymbol{x})}\\
& \leq \sum^m_{i=1}[D_iG(V_1(\boldsymbol{x}), ..., V_m(\boldsymbol{x}))]^{\frac{\alpha(\alpha+\beta-1)}{\alpha+\beta}}\\
& \quad \times [D_iG(U_1(\boldsymbol{x}), ..., U_m(\boldsymbol{x}))]^{\frac{\alpha}{\alpha+\beta}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega_i(\boldsymbol{x}).\\
\end{split}
\end{equation}
By applying H\"{o}lder's inequality for sum with indices $(\alpha+\beta)/\beta$ and $(\alpha+\beta)/\alpha$ to \eqref{eq37}, we get
\begin{equation}\label{eqsum}
\begin{split}
&\sum^m_{i=1}{[D_iG(|f_1(\boldsymbol{x})|, ..., |f_m(\boldsymbol{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha}\omega_i(\boldsymbol{x})\\
&\leq \bigg(\sum^m_{i=1}{[D_iG(V_1(\boldsymbol{x}), ..., V_m(\boldsymbol{x})) ]^{ \frac{\alpha(\alpha+\beta-1)}{\beta} }\omega^{\frac{\alpha+\beta}{\beta}}_i(\boldsymbol{x})\tau^{-\frac{\alpha}{\beta}}_i(\boldsymbol{x})}\bigg)^{\frac{\beta}{\alpha+\beta}}\\
&\quad \times\bigg(\sum^m_{i=1}{D_iG(U_1(\boldsymbol{x}), ..., U_m(\boldsymbol{x}))\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}}\tau_i(\boldsymbol{x}) \bigg)^{{\frac{\alpha}{\alpha+\beta}}}.\\
\end{split}
\end{equation}
Integrating both sides of \eqref{eqsum} with respect to $\boldsymbol{x}$ over $\Omega$ and using H\"{o}lder's inequality with indices $(\alpha+\beta)/\beta$ and $(\alpha+\beta)/\alpha$ give
\begin{equation}
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{[D_iG(|f_1(\boldsymbol{x})|, ..., |f_m(\boldsymbol{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha}\omega_i(\boldsymbol{x})\bigg)\Delta \boldsymbol{x}\\
&\leq \bigg[\int_{\Omega}\bigg(\sum^m_{i=1}{[D_iG(V_1(\boldsymbol{x}), ..., V_m(\boldsymbol{x}))]^{ \frac{\alpha(\alpha+\beta-1)}{\beta} } \omega^{\frac{\alpha+\beta}{\beta}}_i(\boldsymbol{x})\tau^{-\frac{\alpha}{\beta}}_i(\boldsymbol{x}) }\bigg)\Delta \boldsymbol{x}\bigg]^{\frac{\beta}{\alpha+\beta}}\\
&\quad \times \bigg[ \int_{\Omega}\bigg(\sum^m_{i=1}{D_iG(U_1(\boldsymbol{x}), ..., U_m(\boldsymbol{x}))\frac{\partial^{\boldsymbol{1}} U_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1}}}}\bigg) \Delta \boldsymbol{x} \bigg]^{\frac{\alpha}{\alpha+\beta}}, \\
\end{split}
\end{equation}
which, in view of \eqref{ptTHEO1}, yields \eqref{ptTHEO4}. This concludes the proof. \qed
\end{proof}
\begin{remark}\label{Re1}
\begin{enumerate}
\item In the special case when $n=m=1, G(x) =|x|^{(\alpha+\beta)/\alpha}$, $ \omega_1= \tau_1\equiv 1$, $\boldsymbol{\lambda} =\lambda_1,$ then inequality \eqref{ptTHEO4} becomes \cite[Theorem 3.2]{Lihan}.
\item If $n= \boldsymbol{\lambda}=\alpha = 1, m=2, G(x_1, x_2) = |x_1x_2|, \beta = p-1,$ for $1<p\leq 2,$
$\omega_1=\psi^{\sigma}$, $\tau_1=\phi [\psi^{\sigma}]^{p/2}$, where $\psi^{\sigma} = \psi \circ \sigma$, $\phi, \psi \in \mathcal{W}([a,b]_{\mathbb T})$ and $\psi$ is decreasing on $[a,b]_{\mathbb T}$, then inequality \eqref{ptTHEO4} reduces to
\[
\begin{split}
&\int_a^b \psi^{\sigma}(x)(|f_1^{\Delta}(x)f_2(x)|+|f_1(x)f_2^{\Delta}(x)|)\Delta x\\
& \le \frac{K(a, b)}{2^{2/p}}\left[\int_a^b\left(|f_1^{\Delta}(x)|^{p}+|f_2^{\Delta}(x)|^{p}\right)\phi(x) [\psi^{\sigma}(x)]^{p/2}\Delta x \right]^{2/p},\\
\end{split}
\]
where
\begin{align}
\notag K(a, b)&=\left[2\int_a^b\frac{[\psi^{\sigma}(x)]^{q/2}}{\phi^{q/p}(x)}\left(\int_a^x\frac{\Delta t}{\phi^{q/p}(t)[\psi^{\sigma}(t)]^{q/2}}\right)\Delta x \right]^{1/q}\\
\notag& \le \left[2\int_a^b\frac{1}{\phi^{q/p}(x)}\left(\int_a^x\frac{\Delta t}{\phi^{q/p}(t)}\right)\Delta x \right]^{1/q}\\
\label{eqRe1}&\le \left[\int_a^b\frac{\Delta x}{\phi^{q/p}(x)} \right]^{2/q},
\end{align}
where $p, q$ are conjugate exponents. Therefore, Theorem \ref{THEO4} improves and generalizes \cite[Theorem 3.1]{YZ}.
\item Note that when $\mathbb{T} = \mathbb{R}$, $n =m =2, \alpha =s \geq 1, \beta = 2r+s $ for $r \geq 0$, $ G(x_1,x_2) = |x_1x_2|^{(r+s)/s}, \omega_1 = \omega_2 = \tau_1 = \tau_2=\omega,$ where $\omega$ is decreasing in each variables, we see that inequality \eqref{ptTHEO4} improves \cite[Theorem 1]{Che}.
\end{enumerate}
\end{remark}
From Theorem \ref{THEO4} we have the following result.
\begin{corollary}
Assume that conditions in Theorem \ref{THEO4} hold, then
\begin{equation}\label{coro4}
\begin{split}
&\int_{\Omega}\bigg(\sum^m_{i=1}{\bigg[D_iG\bigg(\bigg|\frac{\partial^{\boldsymbol{\xi}_1}f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\xi}_1}}\bigg|, ..., \bigg|\frac{\partial^{\boldsymbol{\xi}_m}f_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\xi}_m}}\bigg| \bigg)\bigg]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol\lambda}}\bigg|^\alpha\omega_i(\textbf{x})}\bigg) \Delta \textbf{x} \\
&\leq \hat{K}_{\Omega} \bigg[G\bigg(\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_1(\textbf{x})\Delta \textbf{x}, ..., \int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_m(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}} \\
\end{split}
\end{equation}
for $\boldsymbol{\xi}_i \leq \boldsymbol{\lambda}, i \in [1, n]_{\mathbb{N}}$,
where $\hat{K}_{\Omega} $ is obtained by substituting
\[ V_i(\textbf{x}): =\int_{\Omega_\textbf{x}}(H_{\boldsymbol{\lambda} - \boldsymbol{\xi}_i}(\textbf{x},\textbf{t}))^{\frac{\alpha+\beta}{\alpha+\beta-1}}(\tau_i(\textbf{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \textbf{t}\] into \eqref{k2}.
\end{corollary}
\begin{remark}
Inequality \eqref{coro4} is the same as \cite[Theorem 2.10]{SAK1}, if we take $n =m=1, G(x) = |x|^\gamma, \gamma \geq 1$.
\end{remark}
The following theorem is similar to Theorem \ref{THEO4}.
\begin{theorem}\label{THEO4'}
Let $\omega_i, \tau_i \in \mathcal{W}(\Omega^{\kappa^{\boldsymbol{\lambda-1}}})$, and $f_i : \Omega \to (-R, R)$ which satisfies
\[\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_i(\textbf{x})\Delta \textbf{x}<R \quad \text{for} \quad i \in [1,m]_{\mathbb{N}}. \]
If $f_i \in \mathcal{L}^{\alpha+\beta}_{{\boldsymbol \rho}^{\boldsymbol{\lambda}- \textbf{1}}(\textbf{b})}(\Omega^{\kappa^{\boldsymbol{\lambda-1}}}, \tau_i, \boldsymbol{\lambda})$ for all $ i \in [1,m]_{\mathbb{N}}$ and
\begin{equation}\label{k2'}
K^*_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}:=\bigg[\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg(\sum^m_{i=1}{[D_iG(V^*_1(\textbf{x}), ..., V^*_m(\textbf{x}))]^{ {\frac{\alpha(\alpha+\beta-1)}{\beta}} }\omega^{\frac{\alpha+\beta}{\beta}}_i(\textbf{x})\tau^{-\frac{\alpha}{\beta}}_i(\textbf{x}) }\bigg)\Delta \textbf{x}\bigg]^{\frac{\beta}{\alpha+\beta}}
\end{equation}
is finite, where
\[V^*_i(\textbf{x}) :=\int_{\bar{\Omega}_\textbf{x}} |H_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t})|^{\frac{\alpha+\beta}{\alpha+\beta-1}}(\tau_i(\textbf{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \textbf{t} \]
for $\textbf{x} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}, i \in [1, m]_{\mathbb{N}},$ then
\begin{equation}\label{eq4'}
\begin{split}
&\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg(\sum^m_{i=1}{[D_iG(|f_1(\textbf{x})|, ..., |f_m(\textbf{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol\lambda}}\bigg|^\alpha\omega_i(\textbf{x})}\bigg) \Delta \textbf{x} \leq K^*_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}} \\
&\quad \times \bigg[G\bigg(\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_1(\textbf{x})\Delta \textbf{x}, ..., \int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_m(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}. \\
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
For all $ i \in[1, m]_{\mathbb{N}},$ and $\boldsymbol{x} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}$, we have
\[
f_i(\boldsymbol{x})=(-1)^n \int_{\bar{\Omega}_{\boldsymbol{x}}} H_{\boldsymbol{\lambda}}(\boldsymbol{x},\boldsymbol{t})\frac{\partial^{\boldsymbol{\lambda}} f_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{\lambda}}}\Delta \boldsymbol{t}. \\
\]
In the proof of Theorem \ref{THEO4}, replacing $\Omega_{\boldsymbol{x}}$ by $\bar{\Omega}_{\boldsymbol{x}}$ we get \eqref{eq4'}. \qed
\end{proof}
\begin{remark}
Since \eqref{eqRe1}, note that when $n= \boldsymbol{\lambda}=\alpha = 1, \beta = p-1,$ for $1<p\leq 2$, $ m=2, G(x_1, x_2) =|x_1x_2|,$
$\omega_1=\psi^{\sigma}$, $\tau_1=\phi [\psi^{\sigma}]^{p/2}$, where $\phi, \psi \in \mathcal{W}([a,b]_{\mathbb T})$ and $\psi$ is decreasing on $[a,b]_{\mathbb T}$, we see that inequality \eqref{eq4'} improves \cite[Theorem 3.2]{YZ}.
\end{remark}
By applying Theorems \ref{THEO4} and \ref{THEO4'} on $\Omega_{\textbf{c}}$ and $\bar{\Omega}_{\textbf{c}}$, respectively, with $\textbf{c} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}$ is such that $\Omega^{\kappa^{\boldsymbol{\lambda-1}}} =\Omega_{\textbf{c}}\cup \bar{\Omega}_{\textbf{c}}$, and summing the resulting inequalities, we have the following result.
\begin{theorem}\label{THEO4''}
Let $\omega_i, \tau_i \in \mathcal{W}(\Omega^{\kappa^{\boldsymbol{\lambda-1}}})$, and $f_i : \Omega \to (-R, R)$ be such that
\[\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_i(\textbf{x})\Delta \textbf{x}<R \quad \text{for} \quad i \in [1,m]_{\mathbb{N}}. \]
If $f_i \in \mathcal{L}^{\alpha+\beta}_{\textbf{a}}(\Omega_{\textbf{c}}, \tau_i, \boldsymbol{\lambda}) \cap \mathcal{L}^{\alpha+\beta}_{{\boldsymbol \rho}^{\boldsymbol{\lambda-1}}(\textbf{b})}({\bar{\Omega}_{\textbf{c}}}, \tau_i, \boldsymbol{\lambda}) $ for all $ i \in [1, m]_{\mathbb{N}}$,
then we have
\begin{equation}\label{eq4''}
\begin{split}
&\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg(\sum^m_{i=1}{[D_iG(|f_1(\textbf{x})|, ..., |f_m(\textbf{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol\lambda}}\bigg|^\alpha\omega_i(\textbf{x})}\bigg) \Delta \textbf{x}\\
& \leq K_{\Omega_{\textbf{c}}}\bigg[G\bigg(\int_{{\Omega}_{\textbf{c}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_1(\textbf{x})\Delta \textbf{x}, ..., \int_{{\Omega}_{\textbf{c}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_m(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}} \\
&\quad + K^*_{{\bar{\Omega}_{\textbf{c}}}} \bigg[G\bigg(\int_{\bar{\Omega}_{\textbf{c}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_1(\textbf{x})\Delta \textbf{x}, ..., \int_{\bar{\Omega}_{\textbf{c}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau_m(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}\\
\end{split}
\end{equation}
for all $\textbf{c} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}$ is such that $\Omega^{\kappa^{\boldsymbol{\lambda-1}}} =\Omega_{\textbf{c}}\cup \bar{\Omega}_{\textbf{c}}$, where $K_{\Omega_{\textbf{c}}}, K^*_{{\bar{\Omega}_{\textbf{c}}}}$ are defined as in \eqref{k2} and \eqref{k2'}, respectively.
\end{theorem}
\begin{remark}
If $n= \boldsymbol{\lambda}=\alpha = 1, \beta = p-1,$ for $1<p\leq 2$, $ m=2, G(x_1, x_2) =|x_1x_2|,$
$\omega_1=\psi^{\sigma}$, $\tau_1=\phi [\psi^{\sigma}]^{p/2}$, where $\phi, \psi \in \mathcal{W}([a,b]_{\mathbb T})$ and $\psi$ is decreasing on $[a,b]_{\mathbb T}$, we see that inequality \eqref{eq4''} improves \cite[Theorem 3.2]{YZ}.
\end{remark}
Let us mention some important consequences of Theorems \ref{THEO4}, \ref{THEO4'}, and \ref{THEO4''}. First, let $m =1$; then we obtain the following corollary.
\begin{corollary}\label{coG}
Let $G\in \mathcal{G}^{1, 1}_R$, $\omega, \tau \in \mathcal{W}(\Omega)$, and
$f : \Omega \to (-R, R)$ be such that
\[\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x})\Delta \textbf{x}<R.\]
If $f \in \mathcal{L}^{\alpha+\beta}_{\textbf{a}}(\Omega, \tau, \boldsymbol{\lambda})$ and
\[N_{\Omega}:=
\bigg[\int_{\Omega}[G'( \vartheta(\textbf{x}))]^{{\frac{\alpha(\alpha+\beta-1)}{\beta}}}\omega^{\frac{\alpha+\beta}{\beta}}(\textbf{x})\tau^{-\frac{\alpha}{\beta}}(\textbf{x}) \Delta \textbf{x}\bigg]^{\frac{\beta}{\alpha+\beta}}<\infty,
\]
where
\[\vartheta(\textbf{x}): =\int_{\Omega_\textbf{x}} \bigg(H_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t})\bigg)^{\frac{\alpha+\beta}{\alpha+\beta-1}}(\tau(\textbf{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \textbf{t}, \quad \textbf{x} \in \Omega,\]
then
\begin{equation}\label{ptG1}
\int_{\Omega}[G'(|f(\textbf{x})|)]^{\alpha}\bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega(\textbf{x}) \Delta \textbf{x}\leq N_{\Omega}\bigg[G\bigg(\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}. \\
\end{equation}
Similarly, if $f \in \mathcal{L}^{\alpha+\beta}_{{\boldsymbol\rho}^{\boldsymbol{\lambda -1} }(\textbf{b})}(\Omega^{\kappa^{\boldsymbol{\lambda-1}}}, \tau, \boldsymbol{\lambda})$ and
\[N^*_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}:=
\bigg[\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}[G'( \vartheta^*(\textbf{x}))]^{ {\frac{\alpha(\alpha+\beta-1)}{\beta}} }\omega^{\frac{\alpha+\beta}{\beta}}(\textbf{x})\tau^{-\frac{\alpha}{\beta}}(\textbf{x}) \Delta \textbf{x}\bigg]^{\frac{\beta}{\alpha+\beta}}<\infty,
\]
where
\[\vartheta^*(\textbf{x}): =\int_{\bar{\Omega}_\textbf{x}} |H_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t})|^{\frac{\alpha+\beta}{\alpha+\beta-1}}(\tau(\textbf{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \textbf{t}, \quad \textbf{x} \in \Omega,\]
then,
\begin{equation}\label{ptG2}
\begin{split}
&\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}[G'(|f(\textbf{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega(\textbf{x}) \Delta \textbf{x}\\
&\leq N^*_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg[G\bigg(\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}.
\end{split}
\end{equation}
If $f \in \mathcal{L}^{\alpha+\beta}_{\textbf{a}}(\Omega_{\textbf{c}}, \tau, \boldsymbol{\lambda}) \cap \mathcal{L}^{\alpha+\beta}_{{{\boldsymbol\rho}^{\boldsymbol{\lambda -1} }(\textbf{b})}}(\bar{\Omega}_{\textbf{c}}, \tau, \boldsymbol{\lambda})$ and $N_{\Omega_{\textbf{c}}}$, $N^*_{\bar{\Omega}_{\textbf{c}}}$ are finite for all $\textbf{c} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}$ is such that $\Omega^{\kappa^{\boldsymbol{\lambda-1}}} =\Omega_{\textbf{c}}\cup \bar{\Omega}_{\textbf{c}}$, then
\begin{equation}\label{ptG3}
\begin{split}
&\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}[G'(|f(\textbf{x})|)]^\alpha\bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega(\textbf{x}) \Delta \textbf{x} \\
& \leq N_{\Omega_{\textbf{c}}}\bigg[G\bigg(\int_{\Omega_{\textbf{c}} }\bigg|\frac{\partial^{\boldsymbol{\lambda}} f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}\\
&\quad + N^*_{\bar{\Omega}_{\textbf{c}}}\bigg[G\bigg(\int_{\bar{\Omega}_{\textbf{c}}}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}. \\
\end{split}
\end{equation}
\end{corollary}
\begin{remark}
Inequality \eqref{ptG1} is the same as inequality given in \cite[Theorem 2.9 ]{SAK1}, if we take
$n=1$ and $G(x) = |x|^{\gamma}$ for $ \gamma>1.$ If $\lambda_1 = 1$, then
inequalities \eqref{ptG1} and \eqref{ptG2} reduce to inequalities given in \cite[Theorems 3.1 and 3.2]{SORA}, respectively.
\end{remark}
We examine Corollary \ref{coG} further in the case when $G(x) = |x|^{\frac{\alpha+\beta}{\alpha}}$. We obtain the following corollary.
\begin{corollary}\label{Co3.16}
Let $\omega, \tau \in \mathcal{W}(\Omega),$ and
$f : \Omega \to (-R, R)$ be such that
\[\int_{\Omega}\bigg|\frac{\partial^{\boldsymbol{\lambda}} f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x})\Delta \textbf{x}<R.\]
If $f \in \mathcal{L}^{\alpha+\beta}_{\textbf{a}}(\Omega, \tau, \boldsymbol{\lambda})$ and
\[
L_{\Omega}:= \bigg[ \int_{\Omega} \bigg( \int_{\Omega_{\textbf{x}}} H^{{\frac{\alpha+\beta}{\alpha+\beta-1}}}_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t}) \tau^{\frac{1}{1-\alpha-\beta}}(\textbf{t}) \Delta \textbf{t} \bigg)^{\alpha+\beta-1} \omega^{\frac{\alpha+\beta}{\beta}}(\textbf{x})\tau^{-\frac{\alpha}{\beta}}(\textbf{x}) \Delta \textbf{x} \bigg]^{\frac{\beta}{\alpha+\beta}} < \infty,
\]
then
\begin{equation}\label{ptmin1}
\int_{\Omega}|f(\textbf{x})|^\beta\bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega(\textbf{x}) \Delta \textbf{x} \leq \bigg(\frac{\alpha}{\alpha+\beta}\bigg)^{\frac{\alpha}{\alpha+\beta}}L_{\Omega} \int_{\Omega} \bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x}) \Delta \textbf{x}.
\end{equation}
Similarly, if $f \in \mathcal{L}^{\alpha+\beta}_{{\boldsymbol\rho}^{\boldsymbol{\lambda -1} }(\textbf{b})}(\Omega^{\kappa^{\boldsymbol{\lambda-1}}}, \tau, \boldsymbol{\lambda})$ and
\[
L^*_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}:= \bigg[ \int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}} } \bigg( \int_{\bar{\Omega}_{\textbf{x}}} |H_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t})|^{{\frac{\alpha+\beta}{\alpha+\beta-1}}} \tau^{\frac{1}{1-\alpha-\beta}}(\textbf{t}) \Delta \textbf{t} \bigg)^{\alpha+\beta-1} \omega^{\frac{\alpha+\beta}{\beta}}(\textbf{x})\tau^{-\frac{\alpha}{\beta}}(\textbf{x}) \Delta \textbf{x} \bigg]^{\frac{\beta}{\alpha+\beta}}
\]
is finite, then
\begin{equation}\label{ptmin2}
\begin{split}
&\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}|f(\textbf{x})|^\beta\bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega(\textbf{x}) \Delta \textbf{x}\\
& \leq \bigg(\frac{\alpha}{\alpha+\beta}\bigg)^{\frac{\alpha}{\alpha+\beta}}L^*_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}} \int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}} \bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x}) \Delta \textbf{x}.
\end{split}
\end{equation}
If $f \in \mathcal{L}^{\alpha+\beta}_{\textbf{a}}(\Omega_{\textbf{c}}, \tau, \boldsymbol{\lambda}) \cap \mathcal{L}^{\alpha+\beta}_{{\boldsymbol\rho}^{\boldsymbol{\lambda -1} }(\textbf{b})}({\bar{\Omega}_{\textbf{c}}}, \tau, \boldsymbol{\lambda})$, $L_{\Omega_{\textbf{c}}}$ and $L^*_{{\bar{\Omega}_{\textbf{c}}}} $ are finite for $\textbf{c} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}$ is such that $\Omega^{\kappa^{\boldsymbol{\lambda-1}}} = \Omega_{\textbf{c}}\cup {\bar{\Omega}_{\textbf{c}}}$, then
\begin{equation}\label{ptmin}
\begin{split}
& \int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}|f(\textbf{x})|^\beta\bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega(\textbf{x}) \Delta \textbf{x}\\
& \leq \bigg(\frac{\alpha}{\alpha+\beta}\bigg)^{\frac{\alpha}{\alpha+\beta}} \bigg[L_{\Omega_{\textbf{c}}} \int_{ \Omega_{\textbf{c}} } \bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x}) \Delta \textbf{x}\\
&\quad +L^*_{{\bar{\Omega}_{\textbf{c}}}} \int_{{\bar{\Omega}_{\textbf{c}}} } \bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x}) \Delta \textbf{x}\bigg].\\
\end{split}
\end{equation}
Furthermore, if we choose $\textbf{c}\in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}$ is such that $ |L_{\Omega_{\textbf{c}}}-L^*_{{\bar{\Omega}_{\textbf{c}}}}| \to \min$, then \eqref{ptmin} reduces to
\begin{equation}\label{ptMIN}
\begin{split}
& \int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}|f(\textbf{x})|^\beta\bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^\alpha\omega(\textbf{x}) \Delta \textbf{x}\\
& \leq \bigg(\frac{\alpha}{\alpha+\beta}\bigg)^{\frac{\alpha}{\alpha+\beta}}\max \{L_{\Omega_{\textbf{c}}}, L^*_{{\bar{\Omega}_{\textbf{c}}}}\}\int_{ \Omega^{\kappa^{\boldsymbol{\lambda-1}}} } \bigg|\frac{\partial^{\boldsymbol{\lambda}}f(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}\bigg|^{\alpha+\beta}\tau(\textbf{x}) \Delta \textbf{x}.
\end{split}
\end{equation}
\end{corollary}
\begin{remark}
The results in Corollary \ref{Co3.16} when $n=1, \boldsymbol{\lambda} = 1, \Omega = [a, b]_{\mathbb{T}}, f(a) = 0$ or/and $f(b) = 0$ yield many known results. Namely, if we set $a =0, \alpha = \beta = 1, \omega= \tau \equiv 1$, then inequality \eqref{ptmin} becomes
\begin{equation}\label{time1}
\begin{split}
&\int_0^b |f(x)| | f^\Delta(x)| \Delta x\\
&\leq \frac{\sqrt{2}}{2} \bigg [\bigg(\int_0^c x \Delta x \bigg)^{1/2} \int_0^c | f^\Delta(x)|^2 \Delta x + \bigg(\int_c^b (b-x) \Delta x \bigg)^{1/2} \int_c ^b | f^\Delta(x)|^2 \Delta x \bigg] \\
\end{split}
\end{equation}
for all $c \in [0, b]_{\mathbb{T}}.$ Since $x \leq \frac{1}{2}(x^2)^\Delta$, then inequality \eqref{time1} reduces to
\begin{equation}\label{time2}
\int_0^b |f(x)| | f^\Delta(x)| \Delta x \leq \frac{c}{2} \int_0^c | f^\Delta(x)|^2 \Delta x + \frac{b-c}{2}\int_c^b | f^\Delta(x)|^2 \Delta x
\end{equation}
for all $c \in [0, b]_{\mathbb{T}}.$ If $\mathbb{T} =\mathbb{R},$ then we can choose $c =b/2$ and inequality \eqref{time2} yields inequality \eqref{1.1}.
Next, inequalities \eqref{ptmin1} and \eqref{ptmin2} are the same as the ones given in \cite[Theorems 2.4 and 2.5]{SAK}, respectively, if we take $ n =1.$
\begin{remark}
When $\mathbb{T} =\mathbb{R}$, from Corollary \ref{Co3.16} we obtain many known results:
\begin{enumerate}
\item If $n=2$ and $\boldsymbol{\lambda} = (1,1)$, then inequalities \eqref{ptmin1} and \eqref{ptmin2} imply the results of Pachpatte given in \cite[Theorems 1 and 4]{pach1}, respectively.
\item Let $n=2, \boldsymbol{\lambda} = (\lambda_1,\lambda_2), \alpha=\beta =1,$ and $\omega = \tau \equiv 1$; then inequality \eqref{ptmin1} reduces to \cite[Theorem 2.1]{Zhao}.
\end{enumerate}
\end{remark}
\end{remark}
The following result which can be proved in view of Corollary \ref{coG} and the well-known inequality of arithmetic and geometric means:
\begin{equation}\label{mean}
\prod^k_{j =1}a_j \leq \bigg( \frac{1}{k}\sum^k_{j=1}a_j\bigg)^k, \quad a_j \geq 0 \quad \text{for}\quad j \in [1, k]_{\mathbb{N}}.
\end{equation}
\begin{corollary}
Let $k \in \mathbb{N}$, $\alpha_j ,\beta_j >0$ be such that $\alpha_j + \beta_j >1/k$, $G_j\in \mathcal{G}^{1,1}_R$, and $\omega_j$, $ \tau_j$ be in $\mathcal{W}(\Omega)$,
and
$f_j : \Omega \to (-R, R)$ be such that
$\int_{\Omega}|\frac{\partial^{\boldsymbol{\lambda}} f_j(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda}}}|^{k(\alpha+\beta)}\tau_j(\textbf{x})\Delta \textbf{x}<R$ for $j \in [1, k]_{\mathbb{N}}.$
If $f_j \in \mathcal{L}^{k(\alpha_j+\beta_j)}_{\textbf{a}}(\Omega, \tau_j, \boldsymbol{\lambda})$ for all $j \in [1, k]_{\mathbb{N}}$, and
\[
{T_{\Omega}}_j:= \bigg[\int_{\Omega}[G'_j( \eta_j(\textbf{x}))]^{\frac{\alpha_j(k\alpha_j+k\beta_j-1)}{\beta_j}}\omega^{\frac{k(\alpha_k+\beta_i)}{\beta_j}}_j(\textbf{x})\tau^{-\frac{\alpha_j}{\beta_j}}_j(\textbf{x}) \Delta \textbf{x}\bigg]^{\frac{\beta_j}{\alpha_j+\beta_j}}<\infty,\]
where
\[ \eta_j (\textbf{x}) := \int_{\Omega_\textbf{x}} \bigg(H_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t})\bigg)^{\frac{k(\alpha_j+\beta_j)}{k(\alpha_j+\beta_j)-1}}(\tau_j(\textbf{t}))^{\frac{1}{1-k(\alpha_j+\beta_i)}}\Delta \textbf{t}, \quad \textbf{x} \in \Omega, \]
then
\begin{equation}\label{phtich1}
\begin{split}
&\int_{\Omega}\prod^{k}_{j=1}{[G'_j(|f_j(\textbf{x})|)]^{\alpha_j}\bigg|\frac{\partial^{ \boldsymbol{\lambda} }f_j(\textbf{x}) }{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}\bigg|^{\alpha_j}\omega_j(\textbf{x})} \Delta \textbf{x} \\
&\leq \frac{1}{k}\sum^k_{j=1}{T_{\Omega j} \bigg[G_j\bigg(\int_{\Omega}\bigg|\frac{\partial^{ \boldsymbol{\lambda} } f_j(\textbf{x}) }{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}\bigg|^{k(\alpha_j+\beta_j)}\tau_j(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha_j}{\alpha_j+\beta_j}}}. \\
\end{split}
\end{equation}
If $f_j \in \mathcal{L}^{k(\alpha+\beta)}_{{\boldsymbol\rho}^{\boldsymbol{\lambda -1} }(\textbf{b})}(\Omega^{\kappa^{\boldsymbol{\lambda-1}}}, \tau_j, \boldsymbol{\lambda})$ for all $j \in [1, k]_{\mathbb{N}}$, and
\[
T^*_{\Omega j}:= \bigg[\int_{\Omega}[G'_j(\eta^*_j(\textbf{x}))]^{\frac{\alpha_j(k\alpha_j+k\beta_j-1)}{\beta_j}}\omega^{\frac{k(\alpha_k+\beta_i)}{\beta_j}}_j(\textbf{x})\tau^{-\frac{\alpha_j}{\beta_j}}_j(\textbf{x}) \Delta \textbf{x}\bigg]^{\frac{\beta_j}{\alpha_j+\beta_j}}<\infty, \]
where
\[ \eta^*_j(\textbf{x}) := \int_{\bar{\Omega}_\textbf{x}} |H_{\boldsymbol{\lambda}}(\textbf{x},\textbf{t})|^{\frac{k(\alpha_j+\beta_j)}{k(\alpha_j+\beta_j)-1}}(\tau_j(\textbf{t}))^{\frac{1}{1-k(\alpha_j+\beta_i)}}\Delta \textbf{t}, \quad \textbf{x} \in \Omega^{\kappa^{\boldsymbol{\lambda-1}}}, \]
then
\begin{equation}\label{phtich2}
\begin{split}
&\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\prod^{k}_{j=1}{[G'_j(|f_j(\textbf{x})|)]^{\alpha_j}\bigg|\frac{\partial^{ \boldsymbol{\lambda} }f_j(\textbf{x}) }{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}\bigg|^{\alpha_j}\omega_j(\textbf{x})} \Delta \textbf{x} \\
&\leq \frac{1}{k}\sum^k_{j=1}{T^*_{\Omega j} \bigg[G_j\bigg(\int_{\Omega^{\kappa^{\boldsymbol{\lambda-1}}}}\bigg|\frac{\partial^{\boldsymbol{\lambda} } f_j(\textbf{x}) }{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}\bigg|^{k(\alpha_j+\beta_j)}\tau_j(\textbf{x})\Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha_j}{\alpha_j+\beta_j}}}. \\
\end{split}
\end{equation}
\end{corollary}
\begin{proof}
By using inequality \eqref{mean}, we obtain
\[
\begin{split}
&\prod^{k}_{j=1}{[G'_j(|f_j(\boldsymbol{x})|)]^{\alpha_j}\bigg|\frac{\partial^{ \boldsymbol{\lambda} }f_j(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{ \boldsymbol{\lambda} }}\bigg|^{\alpha_j}\omega_j(\boldsymbol{x})} \\
&=\bigg(\prod^{k}_{j=1}{[G'_j(|f_j(\boldsymbol{x})|)]^{\alpha_j k}\bigg|\frac{\partial^{ \boldsymbol{\lambda} }f_j(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{ \boldsymbol{\lambda} }}\bigg|^{\alpha_j k}\omega^k_j(\boldsymbol{x})}\bigg)^{\frac{1}{k}} \\
& \leq \frac{1}{k}\sum^k_{j=1}{{[G'_j(|f_j(\boldsymbol{x})|)]^{\alpha_j k}\bigg|\frac{\partial^{ \boldsymbol{\lambda} }f_j(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{ \boldsymbol{\lambda} }}\bigg|^{\alpha_j k}\omega^k_j(\boldsymbol{x})} }. \\
\end{split}
\]
Applying inequalities \eqref{ptG1} and \eqref{ptG2} we arrive at \eqref{phtich1} and \eqref{phtich2}. \qed
\end{proof}
The following theorem generalizes Rozanova's inequality \cite[Theorem 2.5]{AND} to functions of several variables on time scales.
\begin{theorem}\label{THEO5}
Let $F \in \mathcal{H}^m_{\infty}$, $\phi_i$ be convex, non-negative, and increasing on $[0, \infty)$, and $\varphi_i: \Omega \to \mathbb{R}$ be such that $\frac{\partial^{\boldsymbol{1}} \varphi_i(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}$ is non-negative with $\frac{\partial^{k_j} \varphi_i(\textbf{x})}{\Delta_jx^{k_j}_j}|_{x_j = a_j} = 0,$ where $ k_j \in [0, \lambda_j-1]_{\mathbb{N}}$, $ j \in [1, n]_{\mathbb{N}},$ and $i \in [1, m]_{\mathbb{N}}$. For any $ f_i \in \text{C}^{n \boldsymbol{\lambda}}_{\text{rd}}(\Omega) $ is such that $\frac{\partial^{k_j} f_i(\textbf{x})}{\Delta_jx^{k_j}_j}|_{x_j = a_j} = 0$ for all $ k_j \in [0, \lambda_j-1]_{\mathbb{N}}$, $ j \in [1, n]_{\mathbb{N}},$ and $i \in [1, m]_{\mathbb{N}}$, we have
\begin{equation}\label{ptTHEO5}
\begin{split}
&\int_{\Omega}\bigg[\sum^m_{i=1}D_iF\bigg(\varphi_1(\textbf{x})\phi_1\bigg(\frac{|f_1(\textbf{x})|}{\varphi_1(\textbf{x})}\bigg), ..., \varphi_m(\textbf{x})\phi_m\bigg(\frac{|f_m(\textbf{x})|}{\varphi_m(\textbf{x})}\bigg)\bigg) \\
&\quad \times \frac{\partial^{\boldsymbol{1}} \varphi_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{1}} } \phi_i\bigg(H_{ \boldsymbol{\lambda} }(\textbf{b},\textbf{a})\frac{|\frac{\partial^{ \boldsymbol{\lambda} } f_i(\textbf{x}) }{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}} \varphi_i(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}} }} \bigg)\bigg]\Delta \textbf{x} \\
&\leq F\bigg( \int_{\Omega} \frac{\partial^{\boldsymbol{1}} \varphi_1(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}\phi_1\bigg( H_{ \boldsymbol{\lambda} }(\textbf{b},\textbf{a})\frac{ |\frac{\partial^{ \boldsymbol{\lambda} } f_1(\textbf{x}) }{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}} \varphi_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{1}}}}\bigg) \Delta \textbf{x}, ..., \\
&\quad \int_{\Omega} \frac{\partial^{\boldsymbol{1}} \varphi_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{1}}}\phi_m\bigg(H_{\boldsymbol{\lambda}}(\textbf{b},\textbf{a})\frac{| \frac{\partial^{ \boldsymbol{\lambda} }f_m(\textbf{x})}{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}} \varphi_m(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}} \bigg) \Delta \textbf{x} \bigg). \\
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
As in the proof of Theorem \ref{THEO3}, for all $\boldsymbol{x} \in \Omega$, and all $i \in [1, m]_{\mathbb{N}},$ we have
\[
|f_i(\boldsymbol{x})| \leq \int_{\Omega_{\boldsymbol{x}}} H_{ \boldsymbol{\lambda} }(\boldsymbol{x},\boldsymbol{t})\bigg|\frac{\partial^{ \boldsymbol{\lambda} } f_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{ \boldsymbol{\lambda} }}\bigg| \Delta \boldsymbol{t} := y_i(\boldsymbol{x}).
\]
We see that
\[
y_i(\boldsymbol{x}) =\int_{\Omega_{\boldsymbol{x}}} H_{ \boldsymbol{\lambda} }(\boldsymbol{x},\boldsymbol{t})\frac{\partial^{ \boldsymbol{\lambda} } y_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{ \boldsymbol{\lambda} }} \Delta \boldsymbol{t} \leq H_{ \boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a}) \int_{\Omega_{\boldsymbol{x}}}\frac{\partial^{ \boldsymbol{\lambda} } y_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{ \boldsymbol{\lambda} }}\Delta \boldsymbol{t}, \quad \boldsymbol{x} \in \Omega.
\]
By using the fact that $\phi_i, i \in [1, m]_{\mathbb{N}},$ are non-negative, and increasing on $[0, \infty)$, we get
\[
\begin{split}
\phi_i\bigg(\frac{|f_i(\boldsymbol{x})|}{\varphi_i(\boldsymbol{x})}\bigg) &\leq \phi_i\bigg(\frac{y_i(\boldsymbol{x})}{\varphi_i(\boldsymbol{x})}\bigg)\\
&\leq \phi_i\bigg(H_{ \boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a})\frac{\int_{\Omega_{\boldsymbol{x}}}\frac{\partial^{ \boldsymbol{\lambda} } y_i(\boldsymbol{x}) }{\Delta \boldsymbol{t}^{ \boldsymbol{\lambda} }}\Delta \boldsymbol{t}}{\int_{\Omega_{\boldsymbol{x}}}\frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}}\Delta \boldsymbol{t}}\bigg) \\
&= \phi_i\bigg(H_{ \boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a}) \int_{\Omega_{\boldsymbol{x}}}\frac{\frac{\partial^{{\boldsymbol{1}}} \varphi_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{1}}}\frac{\partial^{ \boldsymbol{\lambda} }y_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{ \boldsymbol{\lambda} }} }{\frac{\partial^{{\boldsymbol{1}}}\varphi_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{1}}}} \Delta \boldsymbol{t} \bigg), \\
\end{split}
\]
which, in view of Jensen's inequality \cite{Jen}, we obtain
\begin{equation}\label{ptphi}
\phi_i\bigg(\frac{|f_i(\boldsymbol{x})|}{\varphi_i(\boldsymbol{x})}\bigg) \leq \frac{1}{\varphi_i(\boldsymbol{x})}\int_{\Omega_{\boldsymbol{x}}}\frac{\partial^{\boldsymbol{1}} \varphi_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}}\phi_i\bigg( H_{ \boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a})\frac{ \frac{\partial^{ \boldsymbol{\lambda} } y_i(\boldsymbol{t})}{\Delta \boldsymbol{t}^{ \boldsymbol{\lambda} }}}{\frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{1} }}}\bigg) \Delta \boldsymbol{t}
\end{equation}
for $\boldsymbol{x} \in \Omega$ and $i \in [1, m]_{\mathbb{N}}.$
Setting
\[
W_i(\boldsymbol{t}) :=\frac{\partial^{\boldsymbol{1}} \varphi_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{1} }}\phi_i\bigg(H_{ \boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a})\frac{ \frac{\partial^{ \boldsymbol{\lambda} } y_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{ \boldsymbol{\lambda} }}}{\frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{t}) }{\Delta \boldsymbol{t}^{\boldsymbol{1}}}}\bigg), \quad \boldsymbol{t} \in \Omega_{\boldsymbol{x}}, \quad i \in [1, m]_{\mathbb{N}}.
\]
Thus, \eqref{ptphi} implies that
\begin{equation}\label{eqW}
\varphi_i(\boldsymbol{x})\phi_i\bigg(\frac{|f_i(\boldsymbol{x})|}{\varphi_i(\boldsymbol{x})}\bigg) \leq \int_{\Omega_{\boldsymbol{x}}}W_i(\boldsymbol{t})\Delta \boldsymbol{t}, \quad \boldsymbol{x} \in \Omega, \quad i \in [1, m]_{\mathbb{N}}.
\end{equation}
Since $D_iF$ for $i \in [1, m]_{\mathbb{N}}$, are non-negative, increasing in each variable on $(0, \infty)$ and \eqref{eqW}, we get
\[
\begin{split}
&\int_{\Omega}\bigg[\sum^m_{i=1}D_iF\bigg(\varphi_1(\boldsymbol{x})\phi_1\bigg(\frac{|f_1(\boldsymbol{x})|}{\varphi_1(\boldsymbol{x})}\bigg), ..., \varphi_m(\boldsymbol{x})\phi_m\bigg(\frac{|f_m(\boldsymbol{x})|}{\varphi_m(\boldsymbol{x})}\bigg)\bigg)\\
&\quad \times \frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \phi_i\bigg(H_{ \boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a}) \frac{|\frac{\partial^{ \boldsymbol{\lambda} } f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{ \boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1}}}} \bigg)\bigg]\Delta \boldsymbol{x} \\
&\leq \int_{\Omega}\bigg[\sum^m_{i=1}D_iF\bigg(\int_{\Omega_{\boldsymbol{x}}}W_1(\boldsymbol{t})\Delta \boldsymbol{t}, ...,\int_{\Omega_{\boldsymbol{x}}}W_m(\boldsymbol{t})\Delta \boldsymbol{t} \bigg)W_i(\boldsymbol{x})\bigg]\Delta \boldsymbol{x}. \\
\end{split}
\]
Using \eqref{ptTHEO1}, we obtain \eqref{ptTHEO5}. \qed
\end{proof}
\begin{remark}
\cite[Theorem 2.5]{AND} is a special case of Theorem \ref{THEO5} when $\mathbb{T} = \mathbb{R}$.
\end{remark}
The following theorem is an interesting generalization of Theorem \ref{THEO4}.
\begin{theorem}\label{THEO6}
Let $ G \in \mathcal{G}^{1, m}_R$ and $\omega_i, \tau_i \in \mathcal{W}(\Omega)$ for $\in [1, m]_{\mathbb{N}} $ be such that
\[ P:=\bigg[ \int_{\Omega}{\bigg( \sum^m_{i=1}{D_iG(\nu_1(\textbf{x}), ..., \nu_m(\textbf{x}))\omega^{\frac{\alpha+\beta}{\beta}}_i(\textbf{x})\tau^{-\frac{\alpha}{\beta}}_i(\textbf{x})} \bigg)\Delta \textbf{x}} \bigg]^{\frac{\beta}{\alpha+\beta}}<\infty, \]
where
\[ \nu_i(\textbf{x}) = \bigg(\int_{\Omega_\textbf{x}}(\tau_i(\textbf{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \textbf{t} \bigg)^{\frac{\alpha(\alpha+\beta-1)}{\beta}}, \quad \textbf{x} \in \Omega. \]
Further, for $i \in [1, m]_{\mathbb{N}}$, let $\phi_i$ be convex, non-negative, and increasing on $[0, \infty)$, and $\varphi_i: \Omega \to \mathbb{R}$ be such that $\frac{\partial^{\boldsymbol{1}} \varphi_i (\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}$ is non-negative with $\frac{\partial^{k_j} \varphi_i(\textbf{x})}{\Delta_jx^{k_j}_j}|_{x_j = a_j} = 0,$ where $ k_j \in [0, \lambda_j-1]_{\mathbb{N}}$, $ j \in [1, n]_{\mathbb{N}}.$ If $ f_i \in \text{C}^{n \boldsymbol{\lambda}}_{\text{rd}}(\Omega) $ is such that $\frac{\partial^{k_j} f_i(\textbf{x})}{\Delta_jx^{k_j}_j}|_{x_j = a_j} = 0$ for all $ k_j \in [0, \lambda_j-1]_{\mathbb{N}}$, $ j \in [1, n]_{\mathbb{N}},$ and $i \in [1, m]_{\mathbb{N}}$, and
\[ \int_{\Omega} \bigg| \frac{\partial^{\boldsymbol{1}} \varphi_i(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}\phi_i\bigg( H_{ \boldsymbol{\lambda} }(\textbf{b},\textbf{a})\frac{|\frac{\partial^{ \boldsymbol{\lambda} } f_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}} \varphi_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{1}} }}\bigg)\bigg|^{\alpha+\beta}\tau_i(\textbf{x}) \Delta \textbf{x}<R \]
for all $i \in [1, m]_{\mathbb{N}}$, then we have
\begin{equation}\label{pt37}
\begin{split}
&\int_{\Omega}\bigg[\sum^m_{i=1}D_iG\bigg(\varphi^\alpha_1(\textbf{x})\phi^\alpha_1\bigg(\frac{|f_1(\textbf{x})|}{\varphi_1(\textbf{x})}\bigg), ..., \varphi^\alpha_m(\textbf{x})\phi^\alpha_m\bigg(\frac{|f_m(\textbf{x})|}{\varphi_m(\textbf{x})}\bigg)\bigg)\\
&\quad \times \bigg|\frac{\partial^{\boldsymbol{1}} \varphi_i(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{1}} } \phi_i\bigg(H_{ \boldsymbol{\lambda} }(\textbf{b},\textbf{a})\frac{|\frac{\partial^{ \boldsymbol{\lambda} }f_i(\textbf{x}) }{\Delta \textbf{x}^{ \boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}} \varphi_i(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}} }} \bigg)\bigg|^\alpha\omega_i(\textbf{x})\bigg]\Delta \textbf{x} \\
&\leq P\bigg[G\bigg( \int_{\Omega} \bigg| \frac{\partial^{\boldsymbol{1}} \varphi_1(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}\phi_1\bigg( H_{ \boldsymbol{\lambda} }(\textbf{b},\textbf{a})\frac{ |\frac{\partial^{ \boldsymbol{\lambda} } f_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}} \varphi_1(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{1}} }}\bigg)\bigg|^{\alpha+\beta}\tau_1(\textbf{x}) \Delta \textbf{x}, ...,\\
&\quad \int_{\Omega} \bigg| \frac{\partial^{\boldsymbol{1}} \varphi_m(\textbf{x}) }{\Delta \textbf{x}^{\boldsymbol{1}}}\phi_m\bigg(H_{ \boldsymbol{\lambda} }(\textbf{b},\textbf{a})\frac{| \frac{\partial^{ \boldsymbol{\lambda} }f_m(\textbf{x})}{\Delta \textbf{x}^{ \boldsymbol{\lambda} }} |}{\frac{\partial^{\boldsymbol{1}} \varphi_m(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}} \bigg)\bigg|^{\alpha+\beta}\tau_m(\textbf{x}) \Delta \textbf{x} \bigg)\bigg]^{\frac{\alpha}{\alpha+\beta}}. \\
\end{split}
\end{equation}
\end{theorem}
\begin{proof}
As in the proof of Theorem \ref{THEO5}, using \eqref{eqW} and H\"{o}lder's inequality with indices $(\alpha+\beta)/(\alpha+\beta-1)$ and $(\alpha+\beta)$, we obtain
\[
\begin{split}
\varphi_i(\boldsymbol{x})\phi_i\bigg(\frac{|f_i(\boldsymbol{x})|}{\varphi_i(\boldsymbol{x})}\bigg) &\leq \int_{\Omega_{\boldsymbol{x}}}W_i(\boldsymbol{t})\Delta \boldsymbol{t}\\
& \leq \bigg(\int_{\Omega_{\boldsymbol{x}}}(\tau_i(\boldsymbol{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \boldsymbol{t} \bigg)^{\frac{\alpha+\beta-1}{\alpha+\beta}} \bigg( \int_{\Omega_{\boldsymbol{x}}}(W_i(\boldsymbol{t}))^{\alpha+\beta} \tau_i(\boldsymbol{t})\Delta \boldsymbol{t}\bigg)^{\frac{1}{\alpha+\beta}}\\
\end{split}
\]
for $\boldsymbol{x} \in \Omega$ and $i \in [1, m]_{\mathbb{N}}.$ Hence
\begin{align}
\notag&\varphi^\alpha_i(\boldsymbol{x})\phi^\alpha_i\bigg(\frac{|f_i(\boldsymbol{x})|}{\varphi_i(\boldsymbol{x})}\bigg)\\
\notag&\leq \bigg(\int_{\Omega_{\boldsymbol{x}}}(\tau_i(\boldsymbol{t}))^{\frac{1}{1-\alpha-\beta}}\Delta \boldsymbol{t} \bigg)^{\frac{\alpha(\alpha+\beta-1)}{\alpha+\beta}} \bigg( \int_{\Omega_{\boldsymbol{x}}}(W_i(\boldsymbol{t}))^{\alpha+\beta}\tau_i(\boldsymbol{t})\Delta \boldsymbol{t}\bigg)^{\frac{\alpha}{\alpha+\beta}}\\
& = (\nu_i(\boldsymbol{x}))^{\frac{\beta}{\alpha+\beta}} \bigg( \int_{\Omega_{\boldsymbol{x}}}(W_i(\boldsymbol{t}))^{\alpha+\beta}\tau_i(\boldsymbol{t})\Delta \boldsymbol{t}\bigg)^{\frac{\alpha}{\alpha+\beta}}, \quad \boldsymbol{x} \in \Omega, \quad i \in [1, m]_{\mathbb{N}}.
\label{ww}
\end{align}
Thus, by $ G \in \mathcal{G}^{1, m}_R$ and \eqref{ww}, we get
\[
\begin{split}
&\int_{\Omega}\bigg[\sum^m_{i=1}D_iG\bigg(\varphi^\alpha_1(\boldsymbol{x})\phi^\alpha_1\bigg(\frac{|f_1(\boldsymbol{x})|}{\varphi_1(\boldsymbol{x})}\bigg), ..., \varphi^\alpha_m(\boldsymbol{x})\phi^\alpha_m\bigg(\frac{|f_m(\boldsymbol{x})|}{\varphi_m(\boldsymbol{x})}\bigg)\bigg)\\
&\quad \times \bigg| \frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1} }} \phi_i\bigg(H_{ \boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a}) \frac{|\frac{\partial^{ \boldsymbol{\lambda} } f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{ \boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1}}}} \bigg)\bigg|^\alpha \omega_i(\boldsymbol{x}) \bigg]\Delta \boldsymbol{x}, \\
&\leq \int_{\Omega}\bigg[\sum^m_{i=1}(D_iG(\nu_1(\boldsymbol{x}), ..., \nu_m(\boldsymbol{x})))^{\frac{\beta}{\alpha+\beta}}\bigg(D_iG\bigg(\int_{\Omega_{\boldsymbol{x}}}(W_1(\boldsymbol{t}))^{\alpha+\beta}\tau_1(\boldsymbol{t})\Delta \boldsymbol{t}, ...,\\
&\quad \int_{\Omega_{\boldsymbol{x}}}(W_m(\boldsymbol{t}))^{\alpha+\beta}\tau_m(\boldsymbol{t})\Delta \boldsymbol{t} \bigg)\bigg)^{\frac{\alpha}{\alpha+\beta}} \bigg| \frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1}} } \phi_i\bigg(H_{\boldsymbol{\lambda} }(\boldsymbol{b},\boldsymbol{a}) \frac{|\frac{\partial^{ \boldsymbol{\lambda} } f_i(\boldsymbol{x})}{\Delta \boldsymbol{x}^{ \boldsymbol{\lambda} }}|}{\frac{\partial^{\boldsymbol{1}}\varphi_i(\boldsymbol{x}) }{\Delta \boldsymbol{x}^{\boldsymbol{1}}}} \bigg)\bigg|^\alpha \omega_i(\boldsymbol{x}) \bigg]\Delta \boldsymbol{x}, \\
\end{split}
\]
which, applying H\"{o}lder's inequality with indices $(\alpha+\beta)/\beta$ and $(\alpha+\beta)/\alpha$, yields
\[
\begin{split}
& \leq \bigg[ \int_{\Omega}{\bigg( \sum^m_{i=1}{D_iG(\nu_1(\boldsymbol{x}), ..., \nu_m(\boldsymbol{x}))\omega^{\frac{\alpha+\beta}{\beta}}_i(\boldsymbol{x})\tau^{-\frac{\alpha}{\beta}}_i(\boldsymbol{x})} \bigg)\Delta \boldsymbol{x}} \bigg]^{\frac{\beta}{\alpha+\beta}} \\
&\quad \times \bigg[ \int_{\Omega}\bigg(\sum^m_{i=1}D_iG\bigg(\int_{\Omega_{\boldsymbol{x}}}(W_1(\boldsymbol{t}))^{\alpha+\beta}\tau_1(\boldsymbol{t})\Delta \boldsymbol{t}, ...,\\
&\qquad \int_{\Omega_{\boldsymbol{x}}}(W_m(\boldsymbol{t}))^{\alpha+\beta}\tau_m(\boldsymbol{t})\Delta \boldsymbol{t} \bigg)(W_i(\boldsymbol{x}))^{\alpha+\beta}\tau_i(\boldsymbol{x})\bigg)\Delta \boldsymbol{x} \bigg]^{\frac{\alpha}{\alpha+\beta}}, \\
\end{split}
\]
which completes the proof by applying \eqref{ptTHEO1}. \qed
\end{proof}
\section{Applications}
In this section, we use Opial-type inequalities to establish Lyapunov-type inequalities for the following half-linear dynamic equation
\begin{equation}\label{Lya}
\frac{\partial^{\boldsymbol{1}}}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \bigg(r(\boldsymbol{x})\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg) + s(\boldsymbol{x})y^{\boldsymbol\sigma}(\boldsymbol{x}) =0 \quad \text{on } \quad D = [a_1,b_1]_{\mathbb{T}_1}\times[a_2,b_2]_{\mathbb{T}_2},
\end{equation}
where $\mathbb{T}_1, \mathbb{T}_2$ be arbitrary time scales, $r(\boldsymbol{x}), s(\boldsymbol{x})$ be weights on $D$, and $y^{\boldsymbol\sigma}(\boldsymbol{x}) :=y(\sigma_1(x_1),\sigma_2(x_2))$ for all $\boldsymbol{x} = (x_1, x_2)\in D.$
We define
\[ D_{\boldsymbol{x}} = \{\boldsymbol{t} \in D: \boldsymbol{a} \leq \boldsymbol{t} \leq \boldsymbol{x}\}, \quad \boldsymbol{a }=(a_1, a_2), \]
\[ \bar{D}_{\boldsymbol{x}} = \{\boldsymbol{t} \in D: \boldsymbol{x} \leq \boldsymbol{t} \leq \boldsymbol{b}\}, \quad \boldsymbol{b }=(b_1, b_2), \]
\[S(\boldsymbol{x}) = \int_{\bar{D}_{\boldsymbol{x}}}s(\boldsymbol{t}) \Delta \boldsymbol{t},\]
and
\[S^*(x_1) = (b_2-a_2)[\sup_{\boldsymbol{x} \in D}S(\boldsymbol{x})](b_1-x_1) \quad \text{for }\quad \boldsymbol{x} = (x_1, x_2)\in D. \]
For $\textbf{c} =(c_1, c_2)\in D$, we set
\[
K_1(\boldsymbol{c}) =\bigg[ \int_{D_{\boldsymbol{c}}} \bigg( \int_{D_{\boldsymbol{x}}}r^{-1}(\boldsymbol{x}) \Delta \boldsymbol{t} \bigg)(S^*(x_1)+S(\boldsymbol{x}))^2r^{-1}(\boldsymbol{x}) \Delta \boldsymbol{x}\bigg]^{\frac{1}{2}},
\]
\[
L_1(\boldsymbol{c}) =\bigg[ \int_{\bar{D}_{\boldsymbol{c}}} \bigg( \int_{\bar{D}_{\boldsymbol{x}}}r^{-1}(\boldsymbol{x}) \Delta \boldsymbol{t} \bigg)(S^*(x_1)+S(\boldsymbol{x}))^2r^{-1}(\boldsymbol{x}) \Delta \boldsymbol{x}\bigg]^{\frac{1}{2}},
\]
and
\[ M = \frac{\sqrt{2}}{2}\max\{K_1(\boldsymbol{c}), L_1(\boldsymbol{c}) \}, \]
where $\boldsymbol{c} \in D$ is such that $D =D_{\boldsymbol{c}} \cup \bar{D}_{\boldsymbol{c}} $ and $|K_1(\boldsymbol{c}) -L_1(\boldsymbol{c})| \to \min$.
Now, fix $x_1 \in [a_1,b_1]_{\mathbb{T}_1},$ and put
\[
N_1(x_1,c'_2) =\bigg[ \int^{c'_2}_{a_2} \bigg( \int^{x_2}_{a_2}r^{-1}(\boldsymbol{x}) \Delta t_2 \bigg)(S^*(x_1)+S(\boldsymbol{x}))^2r^{-1}(\boldsymbol{x}) \Delta x_2\bigg]^{\frac{1}{2}},
\]
\[
N_2(x_1,c'_2) =\bigg[ \int^{b_2}_{c'_2} \bigg( \int^{b_2}_{x_2}r^{-1}(\boldsymbol{x}) \Delta t_2 \bigg)(S^*(x_1)+S(\boldsymbol{x}))^2r^{-1}(\boldsymbol{x}) \Delta x_2\bigg]^{\frac{1}{2}},
\]
and
\[ N(x_1) = \frac{\sqrt{2}}{2}\max\{N_1(x_1,c'_2), N_2(x_1,c'_2) \}, \]
where $c'_2 \in [a_2,b_2]_{\mathbb{T}}$ is such that $|N_1(x_1,c'_2) -N_2(x_1,c'_2)| \to \min$.
\begin{theorem}\label{ap}
Suppose that $y$ is a nontrivial solution of \eqref{Lya} such that
\begin{equation}\label{boundary}
y(\textbf{x})|_{x_j =a_j}=y(\textbf{x})|_{x_j =b_j} = \frac{\partial y(\textbf{x})}{\Delta_i x_i}\bigg|_{x_j = a_j} = \frac{\partial y(\textbf{x})}{\Delta_i x_i}\bigg|_{x_j = b_j}=0, \quad i, j=1, 2,
\end{equation}
then
\begin{equation}\label{eq:}
2M+\sup_{x_1 \in [a_1,b_1]_{\mathbb{T}_1}} [\mu_1(x_1)N(x_1)] \geq 1.
\end{equation}
\end{theorem}
\begin{proof}
Multiplying \eqref{Lya} by $y^{\boldsymbol\sigma}(\boldsymbol{x})$ and integrating both sides of \eqref{Lya} with respect to $\boldsymbol{x}$ over $D$ we get
\begin{equation}\label{Mul}
\int_{D}\frac{\partial^{\boldsymbol{1}}}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \bigg(r(\boldsymbol{x})\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg) y^{\boldsymbol\sigma}(\boldsymbol{x}) \Delta \boldsymbol{x} = -\int_{D} s(\boldsymbol{x})(y^{\boldsymbol\sigma}(\boldsymbol{x}))^{2}\Delta \boldsymbol{x}.
\end{equation}
By integrating by parts each side of \eqref{Mul}, we have
\[
\begin{split}
&\int_{D}\frac{\partial^{\boldsymbol{1}}}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \bigg(r(\boldsymbol{x})\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg) y^{\boldsymbol\sigma}(\boldsymbol{x}) \Delta \boldsymbol{x}\\
& = - \int^{b_1}_{a_1} \int^{b_2}_{a_2}
\frac{\partial}{\Delta_1 x_1} \bigg(r(\boldsymbol{x})\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg) \frac{\partial y(\sigma_1(x_1),x_2)}{\Delta_2x_2} \Delta x_2 \Delta x_1\\
&=\int_{D}r(\boldsymbol{x})\bigg(\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg)^2 \Delta \boldsymbol{x}, \\
\end{split}
\]
where, we have used boundary conditions \eqref{boundary}, and
\[
\int_{D} s(\boldsymbol{x})(y^{\boldsymbol\sigma}(\boldsymbol{x}))^{2}\Delta \boldsymbol{x} = - \int_{D}\frac{\partial^{\boldsymbol{1}}S(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}(y^{\boldsymbol\sigma}(\boldsymbol{x}))^{2}\Delta \boldsymbol{x}= -\int_{D} S(\boldsymbol{x}) \frac{\partial^{\boldsymbol{1}} y^2(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \Delta \boldsymbol{x},
\]
where, we have used $$y^2(\sigma_1(x_1),a_2) = y^2(\sigma_1(x_1),b_2) =y(\sigma_1(x_1),b_2)= \frac{\partial y^2(a_1,x_2)}{\Delta_2x_2}=\frac{\partial y^2(b_1,x_2)}{\Delta_2x_2}= 0.$$
Therefore, we obtain
\begin{equation}\label{A0}
\int_{D}r(\boldsymbol{x})\bigg(\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg)^2 \Delta \boldsymbol{x}=\bigg|\int_{D} S(\boldsymbol{x}) \frac{\partial^{\boldsymbol{1}} y^2(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \Delta \boldsymbol{x}\bigg|.
\end{equation}
Since
\[
\frac{\partial^{\boldsymbol{1}} y^2(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \Delta \boldsymbol{x} = \bigg(y^{\boldsymbol\sigma}(\boldsymbol{x})+ y(x_1, \sigma_2(x_2) \bigg) \frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1} +\bigg( \mu_1(x_1)\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1} + 2y(\boldsymbol{x}) \bigg) \frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}
\]
it follows that
\begin{equation}\label{A1}
\begin{split}
\bigg|\int_{D} S(\boldsymbol{x}) \frac{\partial^{\boldsymbol{1}} y^2(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \Delta \boldsymbol{x}\bigg|& \leq
\bigg|\int_{D} S(\boldsymbol{x})y^{\boldsymbol\sigma}(\boldsymbol{x})\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\Delta \boldsymbol{x}\bigg|\\
&\quad +\bigg|\int_{D} S(\boldsymbol{x})y(x_1, \sigma_2(x_2))\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\Delta \boldsymbol{x}\bigg| \\
&\quad + \int_{D} S(\boldsymbol{x})\mu_1(x_1)\bigg|\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\bigg|\bigg|\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg|\Delta \boldsymbol{x} \\
&\quad + 2\int_{D} S(\boldsymbol{x})|y(\boldsymbol{x})|\bigg|\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg|\Delta \boldsymbol{x}. \\
\end{split}
\end{equation}
We have
\begin{align}
\notag& \bigg|\int_{D} S(\boldsymbol{x})y(x_1, \sigma_2(x_2))\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\Delta \boldsymbol{x}\bigg|\\
\notag&\leq \int^{b_1}_{a_1}S^*(x_1) \bigg| \int^{b_2}_{a_2} y(x_1, \sigma_2(x_2))\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\Delta x_2\bigg| \Delta x_1 \\
\notag& =\int^{b_1}_{a_1} S^*(x_1) \bigg| \int^{b_2}_{a_2} y(x_1, x_2)\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \Delta x_2\bigg| \Delta x_1 \\
& =\int_{D}S^*(x_1)| y(\boldsymbol{x})|\bigg|\frac{\partial^{\boldsymbol{1}} (\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg|\Delta \boldsymbol{x},
\label{A2}
\end{align}
where, we have used
\[
\begin{split}
\int^{b_2}_{a_2} y(x_1, \sigma_2(x_2))\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\Delta x_2& =y(x_1, x_2)\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\bigg|^{x_2 = b_2}_{x_2=a_2} - \int^{b_2}_{a_2} y(x_1, x_2)\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \Delta x_2\\
&= - \int^{b_2}_{a_2} y(x_1, x_2)\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \Delta x_2. \\
\end{split}
\]
Similarly, we have
\begin{align}
\notag& \bigg|\int_{D} S(\boldsymbol{x})y^{\boldsymbol\sigma}(\boldsymbol{x})\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\Delta \boldsymbol{x}\bigg| \\
\notag&\leq \int_{D}S^*(x_1) |y(\sigma_1(x_1), x_2)|\bigg| \frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg|\Delta \boldsymbol{x} \\
\notag& \leq \int_{D}S^*(x_1) \bigg| y(\boldsymbol{x}+ \mu_1(x_1) \frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\bigg|\bigg|\frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta \boldsymbol{x} \\
\label{A3}& \leq\int_{D} S^*(x_1) |y(\boldsymbol{x})|\bigg|\frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta \boldsymbol{x}\\
\notag&+ \int_{D}S^*(x_1)\mu_1(x_1)\bigg| \frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\bigg|\bigg| \frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta \boldsymbol{x}.
\end{align}
From \eqref{A1}, \eqref{A2}, and \eqref{A3}, we get
\[
\begin{split}
&\bigg| \int_{D} S(\boldsymbol{x}) \frac{\partial^{\boldsymbol{1}} y^2(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \Delta \boldsymbol{x}\bigg|\\
& \leq 2 \int_{D}( S^*(x_1)+S(\boldsymbol{x}))|y(\boldsymbol{x})|\bigg|\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta \boldsymbol{x}\\
&\quad + \int_{D}(S^*(x_1)+S(\boldsymbol{x}))\mu_1(x_1) \bigg|\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\bigg|\bigg| \frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta \boldsymbol{x}. \\
\end{split}
\]
Applying inequality \eqref{ptMIN} with $n=2, \alpha =\beta =1, \boldsymbol{\lambda} =\boldsymbol{1}$, $\omega(\boldsymbol{x}) = S^*(x_1)+ S(\boldsymbol{x})$, and $\tau (\boldsymbol{x}) = r(\boldsymbol{x})$, we get
\begin{equation}\label{B1}
\int_{D}( S^*(x_1)+S(\boldsymbol{x}))|y(\boldsymbol{x})|\bigg|\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta \boldsymbol{x}\leq M \int_{D}r(\boldsymbol{x})\bigg(\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg)^2 \Delta \boldsymbol{x}.
\end{equation}
Applying inequality \eqref{ptMIN} with $n=1, \alpha =\beta =1, \boldsymbol{\lambda} =1$, $\omega(\boldsymbol{x}) = S^*(x_1)+ S(\boldsymbol{x})$, and $\tau (\boldsymbol{x}) = r(\boldsymbol{x})$, we have
\[
\int^{b_2}_{a_2}(S^*(x_1)+S(\boldsymbol{x})) \bigg|\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\bigg|\bigg| \frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta x_2 \leq N(x_1)\int^{b_2}_{a_2}r(\boldsymbol{x}) \bigg( \frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg)^2\Delta x_2.
\]
This implies that
\begin{align}
\notag&\int^{b_1}_{a_1} \int^{b_2}_{a_2}(S^*(x_1)+S(\boldsymbol{x}))\mu_1(x_1) \bigg|\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}\bigg|\bigg| \frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg| \Delta x_2 \Delta x_1 \\
\notag& \leq \int^{b_1}_{a_1} \mu_1(x_1)N(x_1)\int^{b_2}_{a_2}r(\boldsymbol{x}) \bigg( \frac{\partial^{\boldsymbol{1}}y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg)^2\Delta x_2 \Delta x_1\\
\label{B2}& \leq \bigg(\sup_{x_1 \in [a_1,b_1]_{\mathbb{T}_1}} [\mu_1(x_1)N(x_1)] \bigg)\int_{D}r(\boldsymbol{x})\bigg(\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}}\bigg)^2 \Delta \boldsymbol{x}.
\end{align}
Substituting \eqref{B1} and \eqref{B2} into \eqref{A0}, we have
\[
\int_{D}r(\boldsymbol{x})\bigg(\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg)^2 \Delta \boldsymbol{x} \leq \bigg(2M+\sup_{x_1 \in [a_1,b_1]_{\mathbb{T}_1}} [\mu_1(x_1)N(x_1)] \bigg)
\int_{D}r(\boldsymbol{x})\bigg(\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg)^2 \Delta \boldsymbol{x}.
\]
Then, we get
\[2M + \sup_{x_1 \in [a_1,b_1]_{\mathbb{T}_1}} [\mu_1(x_1)N(x_1)] \geq 1. \]
\qed
\end{proof}
In Theorem \ref{ap} if $r(\boldsymbol{x}) \equiv 1$, then we have the following result.
\begin{corollary}\label{Co1}
Suppose that $y$ is a nontrivial solution of
\[
\frac{\partial^{\boldsymbol{1}}}{\Delta \textbf{x}^{\boldsymbol{1}}} \bigg(\dfrac{\partial^{\boldsymbol{1}} y(\textbf{x})}{\Delta\textbf{x}^{\boldsymbol{1}}} \bigg) +s(\textbf{x}) y^{\boldsymbol\sigma}(\textbf{x}) =0, \quad \boldsymbol{x} \in D
\]
which satisfies boundary conditions \eqref{boundary}, then
\[
2 \max\{K_2(\textbf{c}), L_2(\textbf{c}) \} +\sup_{x_1 \in [a_1,b_1]_{\mathbb{T}_1}}\bigg( \mu_1(x_1) \max\{N_3(x_1,c'_2), N_4(x_1,c'_2) \} \bigg) \geq \sqrt{2},
\]
where
\[
K_2(\textbf{c}) =\bigg[ \int_{D_{\textbf{c}}} (x_1-a_1)(x_2-a_2)[S^*(x_1)+S(\textbf{x})]^2 \Delta \textbf{x}\bigg]^{\frac{1}{2}},
\]
\[
L_2(\textbf{c}) =\bigg[ \int_{\bar{D}_{\textbf{c}}} (b_1-x_1)(b_2-x_2)[S^*(x_1)+ S(\textbf{x})]^2 \Delta \textbf{x}\bigg]^{\frac{1}{2}},
\]
with $\textbf{c} \in D$ is such that $D = D_{\textbf{c}} \cup \bar{D}_{\textbf{c}}$ and $|K_2(\textbf{c}) -L_2(\textbf{c})| \to \min,$ and
\[
N_3(x_1,c'_2) =\bigg[ \int^{c'_2}_{a_2} (x_2-a_2) [S^*(x_1)+ S(\textbf{x})]^2 \Delta x_2\bigg]^{\frac{1}{2}},
\]
\[
N_4(x_1,c'_2) =\bigg[ \int^{b_2}_{c'_2} (b_2-x_2) [S^*(x_1)+ S(\textbf{x})]^2 \Delta x_2\bigg]^{\frac{1}{2}},
\]
for $x_1 \in [a_1,b_1]_{\mathbb{T}_1}$, $c'_2 \in [a_2,b_2]_{\mathbb{T}}$ is such that $|N_3(x_1,c'_2) -N_4(x_1,c'_2)| \to \min$.
\end{corollary}
As a special case of Corollary \ref{Co1}, when $\mathbb{T}_1 = \mathbb{T}_2 = \mathbb{R}$, we see that if $\boldsymbol{c}= (\frac{a_1+b_1}{2}, b_2) $ and $\boldsymbol{c}= (a_1, \frac{a_2+b_2}{2})$ are in $ \in [a_1,b_1]\times[a_2,b_2]$, then $|K_2(\boldsymbol{c}) -L_2(\boldsymbol{c})| = 0$, and $\mu_1(x_1) = 0$, we have the following result.
\begin{corollary}
Suppose that $y$ is a nontrivial solution of
\[
\frac{\partial^{\boldsymbol{1}}}{\partial \textbf{x}^{\boldsymbol{1}}} \bigg(\dfrac{\partial^{\boldsymbol{1}} y(\textbf{x})}{\partial\textbf{x}^{\boldsymbol{1}}} \bigg) + s(\textbf{x})y(\textbf{x}) =0, \quad \boldsymbol{x} \in [a_1,b_1]\times[a_2,b_2]
\]
which satisfies \eqref{boundary}, then
\[
\int_{D} (x_1-a_1)(x_2-a_2)[S^*(x_1)+S(\textbf{x})]^2 d\textbf{x} \geq 1.
\]
\end{corollary}
\begin{remark}
We see that the sufficient condition that equation \eqref{Lya} does not have a nontrivial solution $y$ such that
\[
y(\boldsymbol{x})|_{x_j =a_j}=y(\boldsymbol{x})|_{x_j =b_j} = \frac{\partial y(\boldsymbol{x})}{\Delta_i x_i}\bigg|_{x_j = a_j} = \frac{\partial y(\boldsymbol{x})}{\Delta_i x_i}\bigg|_{x_j= b_j}=0, \quad i, j=1, 2,
\]
can be obtained from Theorem \ref{ap}.
\end{remark}
Now, we consider the following half-linear delay dynamic equation
\begin{equation}\label{delay}
\frac{\partial^{\boldsymbol{1}}} {\Delta \boldsymbol{x}^{\boldsymbol{1}}} \bigg(r(\boldsymbol{x})\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta\boldsymbol{x}^{\boldsymbol{1}}} \bigg) + s(\boldsymbol{x})y(\theta(\boldsymbol{x})) =0 \quad \text{on } \quad D,
\end{equation}
where $\theta(\boldsymbol{x}) = (\theta_1(x_1), \theta_2(x_2))$ for all $\boldsymbol{x} = (x_1,x_2) \in D, $ $\theta_j : \mathbb{T}_j \to \mathbb{T}_j$, $\theta_j(x_j) \leq x_j$, and $\lim_{x_j \to \infty}\theta_j(x_j) = \infty$ for $j =1, 2.$
\begin{corollary}
Suppose that $y$ is a nontrivial solution of equation \eqref{delay} which satisfies boundary conditions \eqref{boundary} and $\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}, $ $ \frac{\partial y(\boldsymbol{x})}{\Delta_2 x_2},$ do not change sign in $(a_1,b_1)_{\mathbb{T}_1}\times(a_2,b_2)_{\mathbb{T}_2}$, then
\[
2M+\sup_{x_1 \in [a_1,b_1]_{\mathbb{T}_1}}[ \mu_1(x_1)N(x_1)] \geq 1.
\]
\end{corollary}
\begin{proof}
Since $\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}, \frac{\partial y(\boldsymbol{x})}{\Delta_2 x_2}$ do not change sign in $(a_1,b_1)_{\mathbb{T}_1}\times(a_2,b_2)_{\mathbb{T}_2}$ we can assume that \eqref{delay} has a solution $y$ satisfying $\frac{\partial y(\boldsymbol{x})}{\Delta_1 x_1}>0, \frac{\partial y(\boldsymbol{x})}{\Delta_2 x_2}>0$. Therefore, since $\theta(\boldsymbol{x}) \leq \boldsymbol{x}$, that $y(\theta(\boldsymbol{x})) \leq y^{\boldsymbol\sigma}(\boldsymbol{x})$.
The remainder of the proof is similar to the proof of Theorem \ref{ap} and hence omitted.\qed
\end{proof}
Finally, we give an upper bound of the solutions of the following integro-partial dynamic equation
\begin{equation}\label{par}
\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} = \zeta\bigg(\boldsymbol{x}, \frac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}, I(y(\boldsymbol{x}))\bigg), \quad \boldsymbol{x} \in \Omega,
\end{equation}
with the initial conditions $y(\boldsymbol{x})|_{x_j =a_j} = 0$ for all $j \in [1, n]_{\mathbb{N}}$, where
\[I(y(\boldsymbol{x})) = \int_{\Omega_{\boldsymbol{x}}}\Phi \bigg(\boldsymbol{t}, y(\boldsymbol{t}),\frac{\partial^{\boldsymbol{1}} y(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \bigg) \Delta \boldsymbol{t}. \]
\begin{theorem}
Suppose that $\Phi (\textbf{x}, y(\textbf{x}),\frac{\partial^{\boldsymbol{1}} y(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}) \leq \omega(\textbf{x}) |y(\textbf{x})|^\beta | \frac{\partial^{\boldsymbol{1}} y(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}|^\alpha$,
where $\omega \in \mathcal{W}(\Omega)$, and
\begin{equation}\label{ptzeta}
\bigg|\zeta\bigg(\textbf{x}, \frac{\partial^{\boldsymbol{1} } y(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}}, I(y(\textbf{x}))\bigg)\bigg| \leq w_1(\textbf{x})+ w_2(\textbf{x})\bigg| \frac{\partial^{ \boldsymbol{1} } y(\textbf{x})}{\Delta \textbf{x}^{\boldsymbol{1}}} \bigg|^\gamma + w_3(\textbf{x})|I(y(\textbf{x}))|,
\end{equation}
where $\gamma \in (0; 1)$, $w_1, w_2,$ and $w_3$ are non-negative on $\Omega$, $w_2(\textbf{x}) < 1$ for all $\textbf{x}\in \Omega$. If equation \eqref{par} has a
solution $y \in \mathcal{L}^{\alpha+\beta}_{\textbf{a}}(\Omega, 1, \boldsymbol{1})$, then
\begin{equation}\label{dgy}
y(\textbf{x}) \leq \int_{\Omega_{\textbf{x}}}[A^{1-\alpha-\beta}(\textbf{t}) +(1-\alpha-\beta)B(\textbf{t}) \Vol(\Omega_{\textbf{t}}) ]^{\frac{1}{1-\alpha-\beta}} \Delta \textbf{t}
\end{equation}
as long as the right-hand side integral exists, where
\[B(\textbf{x}) = \sup_{\textbf{t }\in\Omega_{\textbf{x}}}{\dfrac{w_3(\textbf{t})K(\textbf{t})}{1-w_2(\textbf{t})} },\quad A(\textbf{x}) = \sup_{\textbf{t }\in \Omega_{\textbf{x}}}{\dfrac{w_1(\textbf{t}) + (1-\gamma)\gamma^{\frac{\gamma}{1-\gamma}}}{1-w_2(\textbf{t})} },\]
$\Vol(\Omega_{\textbf{t}})$ is the volume of a rectangular region $\Omega_{\textbf{t}},$ and
\[
K(\textbf{x}) =\bigg(\frac{\alpha}{\alpha+\beta}\bigg)^{\frac{\alpha}{\alpha+\beta}} \bigg(\int_{\Omega_{\textbf{x}}} (\Vol(\Omega_{\textbf{t}}))^{\alpha+\beta -1}\omega^{\frac{\alpha+\beta}{\beta}}(\textbf{t}) \Delta \textbf{t}\bigg)^{\frac{\beta}{\alpha+\beta}}\quad \text{for} \quad \textbf{x} \in \Omega.
\]
\end{theorem}
\begin{proof}
By applying inequality \eqref{ptmin1} with $\boldsymbol{\lambda} = \boldsymbol{1}, \tau \equiv 1$, we have
\begin{equation}\label{ptO1}
\int_{\Omega_{\boldsymbol{x}}} \omega(\boldsymbol{t})|y(\boldsymbol{t})|^\beta\bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \bigg|^\alpha \Delta \boldsymbol{t} \leq K(\boldsymbol{x}) \int_{\Omega_{\boldsymbol{x}}}\bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \bigg|^{\alpha+\beta} \Delta \boldsymbol{t}
\end{equation}
for $ \boldsymbol{x} \in \Omega.$
We consider the function $f(x) = x^\gamma- x -(1-\gamma)\gamma^{\frac{\gamma}{1-\gamma}}$ for $\gamma \in (0,1)$ and $x\geq 0$. We see that $f(x)$ obtains its maximum at $x =\gamma^{\frac{1}{1-\gamma}}$ and $f_{\text{max}} = 0.$ Then, we have
\begin{equation}\label{ptO2}
\bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \bigg|^{\gamma} \leq \bigg|\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}}\bigg| + (1-\gamma)\gamma^{\frac{\gamma}{1-\gamma}}, \quad \boldsymbol{x} \in \Omega.
\end{equation}
Substituting \eqref{ptO1} and \eqref{ptO2} into \eqref{ptzeta}, we obtain
\[(1-w_2(\boldsymbol{x})) \bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \bigg| \leq w_1(\boldsymbol{x}) + (1-\gamma)\gamma^{\frac{\gamma}{1-\gamma}} + w_3(\boldsymbol{x})K(\boldsymbol{x}) \int_{\Omega_{\boldsymbol{x}}}\bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \bigg|^{\alpha+\beta} \Delta \boldsymbol{t} \]
for $ \boldsymbol{x} \in \Omega.$
This implies that
\begin{equation}\label{ptss}
\bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{x})}{\Delta \boldsymbol{x}^{\boldsymbol{1}}} \bigg| \leq \frac{w_1(\boldsymbol{x}) + (1-\gamma)\gamma^{\frac{\gamma}{1-\gamma}}}{1-w_2(\boldsymbol{x})}+ \frac{w_3(\boldsymbol{x})K(\boldsymbol{x})}{1-w_2(\boldsymbol{x})}\int_{\Omega_{\boldsymbol{x}}}\bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \bigg|^{\alpha+\beta} \Delta \boldsymbol{t}
\end{equation}
for $ \boldsymbol{x} \in \Omega.$
Let $\boldsymbol{s} \in \Omega$ be arbitrary, but fixed; then inequality \eqref{ptss} gives
\begin{equation}\label{ptRR}
\bigg| \dfrac{\partial^{\boldsymbol{1} }y(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \bigg| \leq A(\boldsymbol{s}) + B(\boldsymbol{s})\int_{\Omega_{\boldsymbol{t}}}\bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{u})}{\Delta \boldsymbol{u}}^{\boldsymbol{1}} \bigg|^{\alpha+\beta} \Delta \boldsymbol{u}, \quad \boldsymbol{t} \in \Omega_{\boldsymbol{s}}.
\end{equation}
Next, let $R(\boldsymbol{t})$ be the right-hand side of \eqref{ptRR}; then
\[ \dfrac{\partial^{\boldsymbol{1}} R(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} = B(\boldsymbol{s}) \bigg| \dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \bigg|^{\alpha+\beta} \leq B(\boldsymbol{s}) R^{\alpha+\beta}(\boldsymbol{t}), \quad \boldsymbol{t} \in \Omega_{\boldsymbol{s}}, \]
where $ R|_{t_j = a_j} = A(\boldsymbol{s})$ for all $j \in [1, n]_{\mathbb{N}}$. Since
\[ \int_{\Omega_{\boldsymbol{z}}} \frac{1}{ {R^{\alpha+\beta}(\boldsymbol{t})} }\frac{\partial^{\boldsymbol{1}} R(\boldsymbol{t})}{\Delta \boldsymbol{t}^{\boldsymbol{1}}} \Delta \boldsymbol{t} \geq \frac{1}{1-\alpha-\beta}[ R^{1-\alpha -\beta}(\boldsymbol{z})- A^{1-\alpha-\beta}(\boldsymbol{s}) ], \quad \boldsymbol{z} \in \Omega_{\boldsymbol{s}}, \]
then
\[ \frac{1}{1-\alpha-\beta}[ R^{1-\alpha -\beta}(\boldsymbol{z})- A^{1-\alpha-\beta}(\boldsymbol{s})] \leq B(\boldsymbol{s}) \Vol(\Omega_{\boldsymbol{z}}), \quad \boldsymbol{z} \in \Omega_{\boldsymbol{s}}. \]
Therefore,
\[ \bigg|\dfrac{\partial^{\boldsymbol{1}} y(\boldsymbol{z})}{\Delta \boldsymbol{z}^{\boldsymbol{1}}}\bigg| \leq R(\boldsymbol{z}) \leq [A^{1-\alpha-\beta}(\boldsymbol{s}) +(1-\alpha-\beta)B(\boldsymbol{s}) \Vol(\Omega_{\boldsymbol{z}}) ]^{\frac{1}{1-\alpha-\beta}}, \quad \boldsymbol{z} \in \Omega_{\boldsymbol{s}}. \]
In the above inequality replacing $\boldsymbol{z}$ by $\boldsymbol{s}$ and integrating both sides with respect
to $\boldsymbol{s}$ over $\Omega_{\boldsymbol{x}}$ for $\boldsymbol{x} \in \Omega$ we obtain \eqref{dgy}.\qed
\end{proof}
\begin{acknowledgements}
The author would like to express his deepest gratitude to Assoc. Prof. Dinh Thanh Duc, Prof. Vu Kim Tuan and Nguyen Du Vi Nhan for their comments and suggestions to improve this paper.
\end{acknowledgements}
|
\section{Introduction}
Mordell initiated the study of the integral $\displaystyle\int_{-\infty}^{\infty}\frac{e^{ax^2+bx}}{e^{cx}+d}\, dx$, Re$(a)<0$, in his two influential papers \cite{mordell1, mordell2}. Prior to his work, special cases of this integral had been studied, for example, by Riemann in his Nachlass \cite{siegel} to derive the approximate functional equation for the Riemann zeta function, by Kronecker \cite{kronecker1, kronecker2} to derive the reciprocity for Gauss sums, and by Lerch \cite{lerch1, lerch2, lerch3}. Mordell showed that the above integral can be reduced to two standard forms, namely,
\begin{align}
\varphi(z,\tau)&:=\tau\int_{-\infty}^{\infty}\frac{e^{\pi i\tau x^2-2\pi zx}}{e^{2\pi x}-1}\, dx,\label{mint1}\\
\sigma(z,\tau)&:=\int_{-\infty}^{\infty}\frac{e^{\pi i\tau x^2-2\pi zx}}{e^{2\pi \tau x}-1}\, dx,\nonumber
\end{align}
for Im$(\tau)>0$, and was the first to study the properties of these integrals with respect to modular transformations. Bellmann \cite[p.~52]{bellman} coined the terminology `Mordell integrals' for these types of integrals.
Mordell integrals play a very important role in the groundbreaking Ph.D. thesis of Zwegers \cite{zwegers} which sheds a clear light on Ramanujan's mock theta functions. The definition of a Mordell integral $h(z;\tau)$, Im$(\tau)>0$, employed by Zwegers \cite[p.~6]{zwegers}, and now standard in the contemporary literature is,
\begin{equation*}
h(z;\tau):=\int_{-\infty}^{\infty}\frac{e^{\pi i\tau x^2-2\pi zx}}{\cosh\pi x}\, dx.
\end{equation*}
As remarked by Zwegers himself \cite[p.~5]{zwegers}, $h(z;\tau)$ is essentially the function $\varphi(z;\tau)$ defined in \eqref{mint1}, i.e.,
\begin{equation}\label{relvph}
h(z;\tau)=\tfrac{-2i}{\tau}e^{-\left(\frac{\pi i\tau}{4}+\pi iz\right)}\varphi\left(z+\tfrac{\tau-1}{2},\tau\right).
\end{equation}
Kuznetsov \cite{kuznetsov} has recently used $h(z;\tau)$ to simplify Hiary's algorithm \cite{hiary2} for computing the truncated theta function $\sum_{k=0}^{n}\text{exp}(2\pi i(zk+\tau k^2))$, which in turn is used to compute $\zeta\left(\frac{1}{2}+it\right)$ to within $\pm t^{-\lambda}$ in $O_{\lambda}(t^{\frac{1}{3}}(\ln t)^{\kappa})$ arithmetic operations \cite{hiary1}. We refer the reader to a more recent article \cite{cr} and the references therein for further applications of Mordell integrals.
In \cite{ram1915} and \cite{ram1919}, Ramanujan studied the integrals
\begin{align*}
\phi_{\omega}(z):=\int_{0}^{\infty}\frac{\cos(\pi xz)}{\cosh(\pi x)}e^{-\pi\omega x^2}\, dx,\\
\psi_{\omega}(z):=\int_{0}^{\infty}\frac{\sin(\pi xz)}{\sinh(\pi x)}e^{-\pi\omega x^2}\, dx.
\end{align*}
Of course, we require Re $\omega>0$ for the integrals to converge. If we replace $\omega$ by $-i\tau$ with Im$(\tau)>0$ and $z$ by $2iz$, then the integral $\phi$ is nothing but the Mordell integral. That is,
\begin{equation*}
h(z,\tau)=2\phi_{-i\tau}(2iz).
\end{equation*}
Later, Ramanujan briefly worked on these two integrals in a two-page fragment transcribed by G.~N.~Watson from Ramanujan's loose papers and published along with Ramanujan's Lost Notebook \cite[p.~221-222]{lnb}. See also \cite[pp.~307-328]{geabcbrln4} for details.
A third integral of this kind studied by Ramanujan on page 198 of the Lost Notebook is
\begin{equation*}
F_{\omega}(z):=\int_{0}^{\infty}\frac{\sin(\pi xz)}{\tanh{(\pi x)}}e^{-\pi \omega x^2}\, dx.
\end{equation*}
As before, one needs Re $\omega>0$ for convergence. One can easily rephrase this integral as
\begin{equation}\label{ftzrep}
F_{\omega}(z)=\int_{-\infty}^{\infty}\frac{e^{-\pi\omega x^2}\sin(\pi xz)}{e^{2\pi x}-1}\, dx=\int_{-\infty}^{\infty}\frac{e^{-\pi\omega x^2+2\pi x}\sin(\pi xz)}{e^{2\pi x}-1}\, dx.
\end{equation}
For Im$(\tau)>0$, this integral $F$ is connected to the integral $\varphi(z,\tau)$ in \eqref{mint1} (and hence to the Mordell integral $h(z;\tau)$) via
\begin{equation*}
F_{-i\tau}(2iz)=\frac{1}{2i\tau}\left(\varphi(z,\tau)-\varphi(-z,\tau)\right).
\end{equation*}
Thus Mordell integrals pervade Ramanujan's papers and his Lost Notebook. In a further support to this claim, we refer the readers to an interesting paper of Andrews \cite{andmor}.
Berndt and Xu \cite{berndtxu} have proved all of the properties of $F_{\omega}(z)$ claimed by Ramanujan in the Lost Notebook. Following Ramanujan, we assume $\omega>0$. Suppose a certain result holds for $\omega>0$, it is clear that by analytic continuation, one may be able to extend it to complex values of $\omega$ in a certain region containing the positive real line. Among the various properties claimed by Ramanujan is the transformation
\begin{equation}\label{fwttrans}
F_{\omega}(z)=\frac{-i}{\sqrt{\omega}}e^{-\frac{\pi z^2}{4\omega}}F_{1/\omega}\left(\frac{iz}{\omega}\right).
\end{equation}
In view of this property, Berndt and Xu call $F_{\omega}(z)$ as an integral analogue of a theta function.
Assume $\alpha>0$, let $\omega=\alpha^2$ and replace $z$ by $\a z/\sqrt{\pi}$ in \eqref{fwttrans}. Using \eqref{ftzrep}, one sees that the above transformation translates into the following identity.
\textit{For $\alpha, \beta>0$ such that $\a\b=1$,
\begin{align}\label{rt}
\sqrt{\alpha}e^{\frac{z^2}{8}}\int_{-\infty}^{\infty}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx=\sqrt{\beta}e^{\frac{-z^2}{8}}\int_{-\infty}^{\infty}\frac{e^{-\pi\beta^2 x^2}\sinh(\sqrt{\pi}\beta xz)}{e^{2\pi x}-1}\, dx.
\end{align}}
Now consider the integral
\begin{equation}\label{intanpar}
\int_{0}^{\infty}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx.
\end{equation}
In analogy with the partial theta function which is the same as a theta function but summed over only half of the integer lattice, we call the above integral an integral analogue of partial theta function. Note that unlike $F_{\omega}(z)$, this integral cannot be expressed solely in terms of Mordell integrals. Also unlike how the transformation formula for the Jacobi theta function trivially gives the transformation formula for the corresponding partial theta function \cite[p.~22, Equation (2.6.3)]{titch}, the transformation in \eqref{rt} does not give rise to a corresponding transformation for the integral in \eqref{intanpar}. (It is easy to check that the integrands in \eqref{rt} are not even functions of $x$). Nevertheless, the primary goal of this paper is to prove a new and interesting modular transformation for the integral in \eqref{intanpar} which involves error functions.
The error function $\textup{erf}(z)$ and the complementary error function $\textup{erfc}(z)$, defined by \cite[p.~275]{temme}
\begin{equation}\label{erfz}
\textup{erf}(z)=\frac{2}{\sqrt{\pi}}\int_{0}^{z}e^{-t^2}\, dt
\end{equation}
and
\begin{equation*}
\textup{erfc}(z)=1-\textup{erf}(z)=\frac{2}{\sqrt{\pi}}\int_{z}^{\infty}e^{-t^2}\, dt
\end{equation*}
respectively, are two important special functions having a number of applications in probability theory, statistics, physics and partial differential equations. In probability, they are related to the Gaussian normal distribution. Glashier \cite{glashier1} was the first person to coin the term \emph{Error-function} and then the term \emph{Error-function complement} in the sequel \cite{glashier2}. However, his definitions are exactly opposite to the standard definitions given above and do not involve the normalization factor $2/\sqrt{\pi}$.
The imaginary error function $\textup{erfi}(z)$ is defined by \cite[p.~32]{jl}\footnote{This definition differs from a factor of $\frac{2}{\sqrt{\pi}}$ from the definition in \cite{jl}.}
\begin{equation}\label{erfiz}
\textup{erfi}(z)=\frac{2}{\sqrt{\pi}}\int_{0}^{z}e^{t^2}\, dt.
\end{equation}
From (\ref{erfz}) and (\ref{erfiz}), it is straightforward that
\begin{equation}\label{errrel}
\textup{erf}(iz)=i\textup{erfi}(z).
\end{equation}
We now give below the transformation linking the integrals of the type in \eqref{intanpar} with the error functions $\textup{erf}(z)$ and $\textup{erfi}(z)$. This transformation is of the form $G(z,\alpha)=G(iz,\beta)$ for $\alpha,\beta >0$, $\alpha\beta=1$ and $z\in\mathbb{C}$. It is also related to an integral involving the Riemann $\Xi$-function, which is defined by
\begin{equation}\label{xif}
\Xi(t):=\xi(\tf{1}{2}+it),
\end{equation}
where $\xi(s)$ is Riemann's $\xi$-function defined by \cite[p.~60]{dav}
\begin{equation}\label{rixi}
\xi(s):=\frac{1}{2}s(s-1)\pi^{-s/2}\Gamma\left(\frac{s}{2}\right)\zeta(s),
\end{equation}
$\Gamma(s)$ and $\zeta(s)$ being the Gamma function and the Riemann zeta function respectively.
\begin{theorem}\label{genmrramg}
Let $z\in\mathbb{C}, \alpha>0$ and let $\Delta\left(\alpha,z,\frac{1+it}{2}\right)$ be defined by
\begin{equation}\label{delta}
\Delta(x,z,s):=\omega(x,z,s)+\omega(x,z,1-s),
\end{equation}
with
\begin{align}\label{omega}
\omega(x,z,s)&:=x^{\frac{1}{2}-s}e^{-\frac{z^2}{8}}{}_1F_{1}\left(1-\frac{s}{2};\frac{3}{2};\frac{z^2}{4}\right),
\end{align}
where ${}_1F_{1}(a;c;z)$ is Kummer's confluent hypergeometric function. Let $\Xi(t)$ be defined in \eqref{xif}. Then for $\alpha\beta=1$,
\begin{align}\label{mrramg}
&\sqrt{\alpha}e^{\frac{z^2}{8}}\left(\textup{erf}\left(\frac{z}{2}\right)-4\int_{0}^{\infty}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx\right)\nonumber\\
&=\sqrt{\beta}e^{\frac{-z^2}{8}}\left(\textup{erfi}\left(\frac{z}{2}\right)-4\int_{0}^{\infty}\frac{e^{-\pi\beta^2 x^2}\sinh(\sqrt{\pi}\beta xz)}{e^{2\pi x}-1}\, dx\right)\nonumber\\
&=\frac{z}{8\pi^2}\int_{0}^{\infty}\Gamma\left(\frac{-1+it}{4}\right)\Gamma\left(\frac{-1-it}{4}\right)\Xi\left(\frac{t}{2}\right)\Delta\left(\alpha,z,\frac{1+it}{2}\right)\, dt,
\end{align}
where $\textup{erf}(z)$ and $\textup{erfi}(z)$ are defined in \eqref{erfz} and \eqref{erfiz} respectively.
\end{theorem}
We prove Theorem \ref{genmrramg} by evaluating the integral on the extreme right of \eqref{mrramg} to be equal to the extreme left, and by exploiting the fact that it is invariant under the simultaneous replacement of $\a$ by $\b$ and $z$ by $iz$.
This transformation generalizes a formula of Ramanujan which he wrote in his first letter to Hardy \cite[p.~XXVI]{ramcol} and which also appears in \cite[Equation (13)]{sdi}. This formula is equivalent to the first equality in the following identity, valid for $\alpha\beta=1$, and which is also due to Ramanujan \cite{riemann}:
\begin{align}\label{mrram}
&\alpha^{\frac{1}{2}}-4\pi\alpha^{\frac{3}{2}}\int_{0}^{\infty}\frac{xe^{-\pi\alpha^2 x^2}}{e^{2\pi x}-1}\, dx=\beta^{\frac{1}{2}}-4\pi\beta^{\frac{3}{2}}\int_{0}^{\infty}\frac{xe^{-\pi\beta^2 x^2}}{e^{2\pi x}-1}\, dx\nonumber\\
&=\frac{1}{4\pi\sqrt{\pi}}\int_{0}^{\infty}\Gamma\left(\frac{-1+it}{4}\right)\Gamma\left(\frac{-1-it}{4}\right)\Xi\left(\frac{t}{2}\right)\cos \left(\frac{1}{2}t\log\alpha\right)\, dt.
\end{align}
Mordell \cite[p.~331]{mordell2} rephrased the first equality in the above formula in the form \footnote{There is a misprint in Mordell's formulation of Equation \eqref{morpara}, namely, there is an extra minus sign in front of the right-hand side which should not be present.}
\begin{equation}\label{morpara}
\int_{-\infty}^{\infty}\frac{xe^{-\pi ix^2/\tau}}{e^{2\pi x}-1}\, dx=(-i\tau)^{3/2}\int_{-\infty}^{\infty}\frac{xe^{-\pi i\tau x^2}}{e^{2\pi x}-1}\, dx.
\end{equation}
That (\ref{mrram}) is a special case of Theorem \ref{genmrramg} is not difficult to derive: for $z\neq 0$, divide both sides by $z$, let $z\to 0$ and note that
\begin{equation*}
\lim_{z\to 0}\frac{\textup{erf}(z)}{z}=\frac{2}{\sqrt{\pi}}=\lim_{z\to 0}\frac{\textup{erfi}(z)}{z}.
\end{equation*}
A one-variable generalization of the integral on the extreme right-hand side in \eqref{mrram} was given in \cite[Theorem 1.5]{dixthet}, which in turn gave a generalization of the extreme left side. However, this general integral is not invariant under the simultaneous application of $\alpha\to\beta$ and $z\to iz$, and so a transformation formula generalizing the first equality in \eqref{mrram} could not be obtained there. This shortcoming is overcome in Theorem \ref{genmrramg}.
We also obtain another transformation involving error functions that is complementary to the one in Theorem \ref{genmrramg}.
\begin{theorem}\label{thmseft}
Let $z\in\mathbb{C}$. For $\a, \b>0, \a\b=1$,
\begin{align}\label{seft2}
&\sqrt{\alpha}e^{\frac{z^2}{8}}\left(\textup{erf}\left(\frac{z}{2}\right)+4\int_{-\infty}^{0}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx\right)\nonumber\\
&=\sqrt{\beta}e^{\frac{-z^2}{8}}\left(\textup{erfi}\left(\frac{z}{2}\right)+4\int_{-\infty}^{0}\frac{e^{-\pi\beta^2 x^2}\sinh(\sqrt{\pi}\beta xz)}{e^{2\pi x}-1}\, dx\right).\nonumber\\
\end{align}
\end{theorem}
It is important to observe here that subtracting the corresponding sides of the first equality in \eqref{mrramg} from those of \eqref{seft2} results in \eqref{rt}, thus providing a new proof of \eqref{rt}, and hence of \eqref{fwttrans}.
Let $\chi$ denote a primitive Dirichlet character modulo $q$. The character analogue $\Xi(t,\chi)$ of $\Xi(t)$ is given by
\begin{equation*}
\Xi(t,\chi):=\xi\left(\frac{1}{2}+it,\chi\right),
\end{equation*}
where $\xi(s,\chi):=\left(\pi/q\right)^{-(s+a)/2}\Gamma\left(\frac{s+a}{2}\right)L(s,\chi)$, and $a=0$ if $\chi(-1)=1$ and $a=1$ if $\chi(-1)=-1$. The functional equation of $\xi(s,\chi)$ is given by $\xi(1-s,\overline{\chi})=\epsilon (\chi)\xi(s,\chi)$, where $\epsilon (\chi)=i^{a}q^{1/2}/G(\chi)$ and $G(\chi)=\sum_{m=1}^{q}\chi(m)e^{2\pi im/q}$ is the Gauss sum. See \cite[p.~69-72]{dav}. For real primitive characters, we have
\begin{equation*}
G(\chi)=
\begin{cases}
\sqrt{q}, \quad\mbox{for}\hspace{1mm}\chi\hspace{1mm}\mbox{even},\\
i\sqrt{q}, \quad\mbox{for}\hspace{1mm}\chi\hspace{1mm}\mbox{odd}.\\
\end{cases}
\end{equation*}
Hence the functional equation in this case reduces to $\xi(1-s,\chi)=\xi(s,\chi)$, which also gives $\Xi(-t,\chi)=\Xi(t,\chi)$.
We now give below the analogues of Theorem \ref{genmrramg} for real primitive characters.
\begin{theorem}\label{thmrgenrchi}
Let $z\in\mathbb{C}$ and let $\alpha$ and $\beta$ be positive numbers such that $\alpha\beta=1$. Let $\chi$ be a real primitive Dirichlet character modulo $q$.\\
\textup{(i)} If $\chi$ is even,
\begin{align}\label{rgenrchi}
&\sqrt{\alpha}e^{\frac{z^2}{8}}\int_{0}^{\infty}e^{-\frac{\pi\alpha^2 x^2}{q}}\sin\left(\frac{\sqrt{\pi}\alpha xz}{\sqrt{q}}\right)\frac{\sum_{r=1}^{q-1}\chi(r)e^{-\frac{2\pi r x}{q}}}{1-e^{-2\pi x}}\, dx\nonumber\\
&=\sqrt{\beta}e^{\frac{-z^2}{8}}\int_{0}^{\infty}e^{-\frac{\pi\beta^2 x^2}{q}}\sinh\left(\frac{\sqrt{\pi}\beta xz}{\sqrt{q}}\right)\frac{\sum_{r=1}^{q-1}\chi(r)e^{-\frac{2\pi r x}{q}}}{1-e^{-2\pi x}}\, dx\nonumber\\
&=\frac{z\sqrt{q}}{16\pi^{2}}\int_{0}^{\infty}\Gamma\left(\frac{3+it}{4}\right)\Gamma\left(\frac{3-it}{4}\right)\Xi\left(\frac{t}{2},\chi\right)\Delta\left(\alpha,z,\frac{1+it}{2}\right)\, dt.
\end{align}
\textup{(ii)} If $\chi$ is odd,
\begin{align}\label{rgenrchio}
&\sqrt{\alpha}e^{\frac{z^2}{8}}\int_{0}^{\infty}e^{-\frac{\pi\alpha^2 x^2}{q}}\cos\left(\frac{\sqrt{\pi}\alpha xz}{\sqrt{q}}\right)\frac{\sum_{r=1}^{q-1}\chi(r)e^{-\frac{2\pi r x}{q}}}{1-e^{-2\pi x}}\, dx\nonumber\\
&=\sqrt{\beta}e^{\frac{-z^2}{8}}\int_{0}^{\infty}e^{-\frac{\pi\beta^2 x^2}{q}}\cosh\left(\frac{\sqrt{\pi}\beta xz}{\sqrt{q}}\right)\frac{\sum_{r=1}^{q-1}\chi(r)e^{-\frac{2\pi r x}{q}}}{1-e^{-2\pi x}}\, dx\nonumber\\
&=\frac{1}{16\pi^{\frac{3}{2}}}\int_{0}^{\infty}\Gamma\left(\frac{1+it}{4}\right)\Gamma\left(\frac{1-it}{4}\right)\Xi\left(\frac{t}{2},\chi\right)\nabla\left(\alpha,z,\frac{1+it}{2}\right)\, dt.
\end{align}
\end{theorem}
Note that the sums inside the integrals in the above theorem are Gauss sums of purely imaginary arguments.\\
The first equality in \eqref{mrram} can be rewritten for $\alpha\beta=\pi^2$ as
\begin{equation}\label{mtt}
\alpha^{-\frac{1}{4}}\left(1+4\alpha\int_{0}^{\infty}\frac{xe^{-\alpha x^2}}{e^{2\pi x}-1}\, dx\right)=\beta^{-\frac{1}{4}}\left(1+4\beta\int_{0}^{\infty}\frac{xe^{-\beta x^2}}{e^{2\pi x}-1}\, dx\right).
\end{equation}
In \cite[Volume 2, p.~268]{ramnot}, Ramanujan gives an elegant approximation to the above expressions.\\
\textit{Let $\alpha>0$, $\beta>0$, $\alpha\beta=\pi^2$. Define
\begin{equation*}
I(\alpha):=\alpha^{-\frac{1}{4}}\left(1+4\alpha\int_{0}^{\infty}\frac{xe^{-\alpha x^2}}{e^{2\pi x}-1}\, dx\right).
\end{equation*}
Then
\begin{equation}\label{near}
I(\alpha)=\left(\frac{1}{\alpha}+\frac{1}{\beta}+\frac{2}{3}\right)^{1/4}, \text{``nearly''}.
\end{equation}}\\
As mentioned by Berndt and Evans in \cite{br}, Ramanujan frequently used the words ``nearly'' or ``very nearly'' at the end of his asymptotic expansions and approximations. The above approximation is very good for values of $\alpha$ that are either very small or very large. A proof of the above fact was given in \cite{br}, where as an intermediate result, the following asymptotic expansion for $I(\a)$ as $\a\to 0$ was first obtained:
\begin{equation*}
I(\a)\sim\frac{1}{\a^{1/4}}+\frac{\a^{3/4}}{6}-\frac{\a^{7/4}}{60}+\cdots.
\end{equation*}
Observe that for $\alpha\beta=\pi^2$ and $z\neq 0$, the first equality in Theorem \ref{genmrramg} can be rephrased as follows:
\begin{align}\label{gen}
I(z,\a)&:=\frac{\sqrt{\pi}}{z}\alpha^{-\frac{1}{4}}e^{\frac{z^2}{8}}\textup{erf}\left(\frac{z}{2}\right)+\frac{4}{z}\alpha^{\frac{1}{4}}e^{-\frac{z^2}{8}}\int_{0}^{\infty}\frac{e^{-\alpha x^2}\sinh(\sqrt{\alpha} xz)}{e^{2\pi x}-1}\, dx\nonumber\\
&=\frac{\sqrt{\pi}}{z}\beta^{-\frac{1}{4}}e^{-\frac{z^2}{8}}\textup{erfi}\left(\frac{z}{2}\right)+\frac{4}{z}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\int_{0}^{\infty}\frac{e^{-\beta x^2}\sin(\sqrt{\beta} xz)}{e^{2\pi x}-1}\, dx=:I(iz,\b),
\end{align}
of which \eqref{mtt} is the special case when $z\to 0$. The following general asymptotic expansion holds for the two sides in the above identity as $\a\to 0$, or equivalently as $\beta\to\infty$.
\begin{theorem}\label{1asy}
Fix $z\in\mathbb{C}$.
As $\a\to 0$,
\begin{align}\label{1asyeq}
I(z,\a)&\sim\frac{-2}{\sqrt{\pi}}\a^{-1/4}e^{z^2/8}\sum_{m=0}^{\infty}\left(\frac{-\a}{\pi^2}\right)^{m}\zeta(2m)\G\left(m+\frac{1}{2}\right){}_1F_{1}\left(m+\frac{1}{2};\frac{3}{2};\frac{-z^2}{4}\right).
\end{align}
That is,
\begin{align*}
I(z,\a)&\sim\frac{\sqrt{\pi}}{z}e^{z^2/8}\textup{erf}\left(\frac{z}{2}\right)\alpha^{-1/4}+\frac{e^{-z^2/8}}{6}\alpha^{3/4}+\frac{(z^2-6)e^{-z^2/8}}{360}\alpha^{7/4}\nonumber\\
&\quad+\frac{(60-20z^2+z^4)e^{-z^2/8}}{15120}\alpha^{11/4}+\cdots.
\end{align*}
\end{theorem}
Note that both sides of \eqref{gen} are even functions of $z$. If we successively differentiate \eqref{gen} $n$ times with respect to $z$ and then let $z\to 0$, we do not get anything interesting for odd $n$. However for $n$ even, two different behaviors are noted.
First, $n\equiv0\hspace{1mm}(\text{mod}\hspace{1mm} 4)$, i.e., $n=4k, k\in\mathbb{N}\cup\{0\}$, gives the following transformation of the form $H_{k}(\a)=H_{k}(\b)$.
\begin{theorem}\label{thmtrn0m4}
Let $\a\b=\pi^2$. Then for a non-negative integer $k$,
\begin{align}\label{trn0m4}
&\a^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)+4\a^{3/4}\int_{0}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&=\b^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)+4\b^{3/4}\int_{0}^{\infty}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\b x^2\right)\, dx.
\end{align}
\end{theorem}
Ramanujan's approximation in \eqref{near} is a special case, when $k=0$, of the following result:
\begin{theorem}\label{thmgennear}
Let $k$ be a non-negative integer. Both sides of \eqref{trn0m4} are approximated by
\begin{align}\label{gennear}
&\a^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)+4\a^{3/4}\int_{0}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&={}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)\left(\frac{1}{\a}+\frac{1}{\b}+\frac{2}{3\cdot{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)}\right)^{1/4}, \text{``nearly''}.
\end{align}
\end{theorem}
Again, the above right side is a very good approximation of the left side for the values of $\a$ that are either very small or very large.
When $n\equiv2\hspace{1mm}(\text{mod}\hspace{1mm} 4)$, i.e., $n=4k+2, k\in\mathbb{N}\cup\{0\}$, we get a transformation of the form $J_{k}(\a)=-J_{k}(\b)$ given below.
\begin{theorem}\label{thmtrn2m4}
Let $\a\b=\pi^2$ and let $k$ be a non-negative integer. Then,
\begin{align}\label{trn2m4}
&J_{k}(\a):=\a^{-1/4}{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right)+4\a^{3/4}\int_{0}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&=-\b^{-1/4}{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right)-4\b^{3/4}\int_{0}^{\infty}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\b x^2\right)dx=:-J_{k}(\b).
\end{align}
\end{theorem}
In particular, $J_{k}(\pi)=0$, which results in a beautiful exact evaluation of the integral in \eqref{trn2m4}.
\begin{corollary}\label{corexact}
For any non-negative integer $k$,
\begin{align}\label{exact}
\int_{0}^{\infty}\frac{xe^{-\pi x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\pi x^2\right)\, dx=-\frac{1}{4\pi}{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right).
\end{align}
\end{corollary}
Results corresponding to the ones in Theorem \ref{thmtrn0m4} - Corollary \ref{corexact} that can be obtained by writing Theorem \ref{thmseft} in an alternate form (see \eqref{sefty} below) are collectively put in Theorem \ref{appseft} at the end of Section 5. When we combine the results from Theorem \ref{thmtrn0m4}-Corollary \ref{corexact} with those in Theorem \ref{appseft}, we obtain the following interesting theorem.
\begin{theorem}\label{ramtran}
Let $\a, \b$ be two positive numbers such that $\a\b=\pi^2$ and let $k$ be any non-negative integer. Then
{\allowdisplaybreaks\begin{align}
&\textup{(i)}\hspace{2mm} \a^{3/4}\int_{-\infty}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right)\, dx=\b^{3/4}\int_{-\infty}^{\infty}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\b x^2\right)\, dx.\nonumber\\
&\textup{(ii)}\hspace{2mm} \a^{3/4}\int_{-\infty}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&\quad\quad=\frac{1}{2}\cdot{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)\left(\frac{1}{\a}+\frac{1}{\b}+\frac{2}{3\cdot{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)}\right)^{1/4}, \text{``nearly''}.\label{gennearg}\\
&\textup{(iii)}\hspace{2mm} \a^{3/4}\int_{-\infty}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad=-\b^{3/4}\int_{-\infty}^{\infty}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\b x^2\right)\, dx.\label{dires}
\end{align}}
In particular when $\a=\b=\pi$, we have
\begin{equation}\label{exact0}
\int_{-\infty}^{\infty}\frac{xe^{-\pi x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\pi x^2\right)\, dx=0.
\end{equation}
\end{theorem}
In \cite{riemann}, Ramanujan considered two integrals, one being that on the extreme right of \eqref{mrram}, and the second one given by
\begin{multline}
\int_{0}^{\infty}
\Gamma \left( \frac{z-1+ i t }{4} \right)
\Gamma \left( \frac{z-1- i t }{4} \right)\Xi \left( \frac{t + i z}{2} \right)
\Xi \left( \frac{t - i z}{2} \right)
\frac{\cos( \tfrac{1}{2} t \log \alpha) \, dt}{t^{2} + (z+1)^{2}}.
\label{ramanujan-98}
\end{multline}
Oloa \cite[Equation 1.5]{oloa} found the asymptotic expansion\footnote{There is a misprint in this asymptotic expansion given in Oloa's paper. The minus sign in front of the second expression on the right should be a plus.} of the special case of this integral when $z=0$, namely, as $\a\to\infty$,
\textit{\begin{align*}
&\frac{1}{\pi^{3/2}}\int_0^{\infty}\Xi^{2}\left(\frac{1}{2}t\right)\left|\Gamma\left(\frac{-1+it}{4}\right)\right|^2
\frac{\cos\left(\frac{1}{2}t\log\alpha\right)}{1+t^2}\, dt\\
&\sim\frac{1}{2}\frac{\log\alpha}{\sqrt{\alpha}}+\frac{1}{2\sqrt{\alpha}}\left(\log 2\pi-\gamma\right)+\frac{\pi^{2}}{72\alpha^{3/2}}-\frac{\pi^4}{10800\alpha^{7/2}}+\cdots.
\end{align*}}
In the following theorem, we obtain the asymptotic expansion of the general integral \eqref{ramanujan-98} as $\a\to\infty$.
\begin{theorem}\label{2asy}
Fix $z$ such that $-1 <$ \textup{Re} $z < 1$. As $\alpha\to\infty$,
\begin{align}\label{oloag}
&\frac{1}{\pi^{\frac{z+3}{2}}}\int_{0}^{\infty}
\Gamma \left( \frac{z-1+it}{4} \right)
\Gamma \left( \frac{z-1-it}{4} \right)
\Xi \left( \frac{t + iz}{2} \right)
\Xi \left( \frac{t - iz}{2} \right)\frac{\cos( \tfrac{1}{2} t \log \alpha) \, dt}{t^{2} + (z+1)^{2}}\nonumber\\
&\sim-\frac{\Gamma(z)\zeta(z)\a^{\frac{z-1}{2}}}{(2\pi)^z}-\frac{\G(z+1)\zeta(z+1)}{2\a^{\frac{z+1}{2}}(2\pi)^z}\nonumber\\
&\quad+2\a^{\frac{z+1}{2}}\sum_{m=0}^{\infty}\frac{(-1)^m}{(2\pi\a)^{2m+z+2}}\G(2m+2+z)\zeta(2m+2)\zeta(2m+z+2).
\end{align}
\end{theorem}
This paper is organized as follows. In Section 2, we state preliminary theorems and lemmas that are subsequently used. Section 3 contains proofs of Theorems \ref{genmrramg} and \ref{thmrgenrchi}. In Section 4, we prove Theorems \ref{thmseft} and \ref{1asy}. The analogues of Theorem \ref{1asy} corresponding to the second error function transformation and to Ramanujan's transformation \eqref{rt} are also given in this section. Section 5 is devoted to proofs of Theorem \ref{thmtrn0m4} - Corollary \ref{corexact} and their analogues. We prove Theorem \ref{2asy} in Section 6. Finally, Section 7 is reserved for some concluding remarks and open problems.
\section{Nuts and bolts}
Let $f$ be an even function of $t$ of the form $f(t)=\phi(it)\phi(-it)$, where $\phi$ is analytic in $t$ as a function of a real variable. Using the functional equation for $\zeta(s)$ in the form $\xi(s)=\xi(1-s)$, where $\xi(s)$ is defined in \eqref{rixi}, it is easy to obtain the following line integral representation for the integral on the left side below, of which the integral on the extreme right of \eqref{mrramg} is a special case
\begin{equation}\label{sp4}
\int_{0}^{\infty}f\left(\frac{t}{2}\right)\Xi\left(\frac{t}{2}\right)\Delta\left(\alpha,z,\frac{1+it}{2}\right)\, dt
=\frac{2}{i}\int_{\frac{1}{2}-i\infty}^{\frac{1}{2}+i\infty}\phi\left(s-\frac{1}{2}\right)\phi\left(\frac{1}{2}-s\right)\xi(s)\omega(\alpha,z,s)\, ds,
\end{equation}
whenever the integral on the left converges. Here $\Delta(x, z, s)$ and $\omega(x, z, s)$ are the same as defined in \eqref{delta} and \eqref{omega}. Analogous to these, define
\begin{equation}\label{nabla}
\nabla(x,z,s):=\rho(x,z,s)+\rho(x,z,1-s),
\end{equation}
where
\begin{align*}
\rho(x,z,s):=x^{\frac{1}{2}-s}e^{-\frac{z^2}{8}}{}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{z^2}{4}\right).
\end{align*}
Then for $\chi$, a real primitive character modulo $q$, the following formulas can be similarly proved.
\begin{align}\label{sp4chi}
\int_{0}^{\infty}f\left(\frac{t}{2}\right)\Xi\left(\frac{t}{2},\chi\right)\nabla\left(\alpha,z,\frac{1+it}{2}\right)\, dt
&=\frac{2}{i}\int_{\frac{1}{2}-i\infty}^{\frac{1}{2}+i\infty}\phi\left(s-\frac{1}{2}\right)\phi\left(\frac{1}{2}-s\right)\xi(s,\chi)\rho(\alpha,z,s)\, ds,\nonumber\\
\int_{0}^{\infty}f\left(\frac{t}{2}\right)\Xi\left(\frac{t}{2},\chi\right)\Delta\left(\alpha,z,\frac{1+it}{2}\right)\, dt
&=\frac{2}{i}\int_{\frac{1}{2}-i\infty}^{\frac{1}{2}+i\infty}\phi\left(s-\frac{1}{2}\right)\phi\left(\frac{1}{2}-s\right)\xi(s,\chi)\omega(\alpha,z,s)\, ds,
\end{align}
whenever the integrals on the left-hand sides converge. Note that for $\alpha\beta=1$,
\begin{equation}\label{rel}
\Delta\left(\alpha,z,\frac{1+it}{2}\right)=\Delta\left(\beta,iz,\frac{1+it}{2}\right), \nabla\left(\alpha,z,\frac{1+it}{2}\right)=\nabla\left(\beta,iz,\frac{1+it}{2}\right),
\end{equation}
both of which can be proved using Kummer's first transformation for ${}_1F_{1}(a;c;w)$ \cite[p.~125, Equation (2)]{rain} given by
\begin{equation}\label{kft}
{}_1F_{1}(a;c;w)=e^w{}_1F_{1}(c-a;c;-w).
\end{equation}
The formulas in \eqref{rel} render the integrals on the left-hand sides of \eqref{sp4} and \eqref{sp4chi} invariant under the simultaneous replacement of $\alpha$ by $\beta$ and $z$ by $iz$, and hence, as a by-product of the evaluation of these integrals, we obtain identities of the form $G(z,\alpha)=G(iz,\beta)$ and $G(z,\alpha, \chi)=G(iz,\beta,\chi)$.
In proving Theorem \ref{thmrgenrchi}, we make use of the following special case \cite[Theorem 2.1]{bds} of a result due to Berndt \cite[Theorem 10.1]{bcbspfunc}:
\begin{theorem}\label{wat}
Let $x>0$. If $\chi$ is even with period $q$ and \textup{Re} $\nu\geq 0$, then
\begin{equation*}
\sum_{n=1}^{\infty}\chi(n) n^{\nu} K_{\nu}\left(\frac{2\pi nx}{q}\right)
=\frac{\pi^{\frac{1}{2}}}{2xG(\overline{\chi})}\left(\frac{qx}{\pi}\right)^{\nu+1}\Gamma\left(\nu+\frac{1}{2}\right)
\sum_{n=1}^{\infty}\overline{\chi}(n)(n^2+x^2)^{-\nu-\frac{1}{2}};
\end{equation*}
if $\chi$ is odd with period $k$ and \textup{Re} $\nu>-1$, then
\begin{equation}\label{watodd}
\sum_{n=1}^{\infty}\chi(n) n^{\nu+1} K_{\nu}\left(\frac{2\pi
nx}{q}\right)=\frac{i\pi^{\frac{1}{2}}}{2x^2{G(\overline{\chi})}}\left(\frac{qx}{\pi}\right)^{\nu+2}
\Gamma\left(\nu+\frac{3}{2}\right)\sum_{n=1}^{\infty}\overline{\chi}(n)n(n^2+x^2)^{-\nu-\frac{3}{2}}.
\end{equation}
\end{theorem}
The following two lemmas, given in \cite[p.~503, Formula (3.952.7)]{grn} and \cite[pp.~318, 320, Formulas (10), (30)]{tti} respectively, will also be employed in the proof of Theorem \ref{thmrgenrchi}.
\begin{lemma}\label{corinvsinl}
For $c=$ Re $s>-1$ and Re $a>0$, we have
\begin{equation*}
\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\frac{b}{2}a^{-\frac{1}{2}-\frac{s}{2}}e^{-\frac{b^2}{4a}}\Gamma\left(\frac{s+1}{2}\right){}_1F_{1}\left(1-\frac{s}{2};\frac{3}{2};\frac{b^2}{4a}\right)x^{-s}\, ds=e^{-ax^2}\sin bx.
\end{equation*}
\end{lemma}
\begin{lemma}\label{corinvcosl}
For $c=$ Re $s>0$ and Re $a>0$, we have
\begin{equation}\label{invmel}
\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\frac{1}{2}a^{-\tfrac{s}{2}}\Gamma\left(\frac{s}{2}\right)e^{-\frac{b^2}{4a}}{}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{b^2}{4a}\right)x^{-s}\, ds=e^{-ax^2}\cos bx.
\end{equation}
\end{lemma}
\noindent
We note that \cite[Equation (2.10)]{dixthet}
\begin{equation*}
{}_1F_{1}\left(\tfrac{1}{4}-\lambda;\tfrac{1}{2};\tfrac{z^{2}}{4}\right)\sim e^{z^2/8}\cos\left(\sqrt{\lambda}z\right),
\end{equation*}
as $|\lambda|\to\infty$ and $|\arg(\lambda z)|<2\pi$, and the Stirling's formula for $\Gamma(s)$, $s=\sigma+it$, in a vertical strip $\alpha\leq\sigma\leq\beta$ given by
\begin{equation}\label{strivert}
|\Gamma(s)|=(2\pi)^{\tf{1}{2}}|t|^{\sigma-\tf{1}{2}}e^{-\tf{1}{2}\pi |t|}\left(1+O\left(\frac{1}{|t|}\right)\right),
\end{equation}
as $|t|\to\infty$ give convergence of the integrals on the extreme right-hand sides of \eqref{mrramg}, \eqref{rgenrchi} and \eqref{rgenrchio}. If $F(s)$ and $G(s)$ denote the Mellin transforms of $f(x)$ and $g(x)$ respectively and $s$ with Re $s=c$ lies in a common strip where both $F$ and $G$ are analytic, then a variant of Parseval's formula \cite[p.83, Equation (3.1.13)]{kp} gives
\begin{equation}\label{melconv}
\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}F(s)G(s)w^{-s}\, ds=\int_{0}^{\infty}f(x)g\left(\frac{w}{x}\right)\frac{dx}{x}.
\end{equation}
Watson's lemma \cite[p.~71]{olver} is given by
\begin{lemma}\label{watlem}
If $q(t)$ is a function of the positive real variable $t$ such that
\begin{equation*}
q(t)\sim\sum_{s=0}^{\infty}a_st^{(s+\lambda-\mu)/\mu}\hspace{3mm} (t\to 0)
\end{equation*}
for positive constants $\lambda$ and $\mu$, then
\begin{equation}\label{watlem1}
\int_{0}^{\infty}e^{-xt}q(t)\, dt\sim\sum_{s=0}^{\infty}\Gamma\left(\frac{s+\lambda}{\mu}\right)\frac{a_s}{x^{(s+\lambda)/\mu}}\hspace{3mm} (x\to\infty),
\end{equation}
provided that this integral converges throughout its range for all sufficiently large $x$.
\end{lemma}
The above result also holds \cite[p.~32]{temme} for complex $\lambda$ with Re $\lambda>0$, and for $x\in\mathbb{C}$ with the integral being convergent for all sufficiently large values of Re $x$.
\section{The first error function transformation and its character analogues}
We begin by proving the first error function transformation given in Theorem \ref{genmrramg} and then proceed to a proof of its character analogues given in Theorem \ref{thmrgenrchi}.
\begin{proof}[Theorem \textup{\ref{genmrramg}}][]
Let $\phi(s)=\Gamma\left(\frac{-1}{4}+\frac{s}{2}\right)$ and let
\begin{equation*}
J(z,\alpha)=\int_{0}^{\infty}\Gamma\left(\frac{-1+it}{4}\right)\Gamma\left(\frac{-1-it}{4}\right)\Xi\left(\frac{t}{2}\right)\Delta\left(\alpha,z,\frac{1+it}{2}\right)\, dt.
\end{equation*}
Use \eqref{xif}, \eqref{rixi}, \eqref{sp4}, the functional equation for $\Gamma(s)$ and the reflection formula to see that
\begin{align*}
J(z,\alpha)&=\frac{2\sqrt{\alpha}e^{-\frac{z^2}{8}}}{i}\int_{\frac{1}{2}-i\infty}^{\frac{1}{2}+i\infty}\Gamma\left(\frac{s+1}{2}\right)\Gamma\left(-\frac{s}{2}\right)\Gamma\left(1+\frac{s}{2}\right)\zeta(s){}_1F_{1}\left(1-\frac{s}{2};\frac{3}{2};\frac{z^2}{4}\right)\left(\sqrt{\pi}\alpha\right)^{-s}\, ds\nonumber\\
&=-\frac{4\sqrt{\alpha}e^{-\frac{z^2}{8}}}{i}\int_{\frac{1}{2}-i\infty}^{\frac{1}{2}+i\infty}\frac{\pi}{\sin\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s+1}{2}\right)\zeta(s){}_1F_{1}\left(1-\frac{s}{2};\frac{3}{2};\frac{z^2}{4}\right)\left(\sqrt{\pi}\alpha\right)^{-s}\, ds.
\end{align*}
Now in order to use the series representation for $\zeta(s)$, we shift the line of integration from Re $s=\frac{1}{2}$ to Re $s=1+\delta$, where $0<\delta<1$. Consider a positively oriented rectangular contour with sides $[\frac{1}{2}+iT, \frac{1}{2}-iT], [\frac{1}{2}-iT, 1+\delta-iT], [1+\delta-iT,1+\delta+iT]$ and $[1+\delta+iT,\frac{1}{2}+iT]$, where $T$ is any positive real number. We have to consider the contribution of the pole of order $1$ of the integrand (due to $\zeta(s)$). Using the residue theorem, noting that by (\ref{strivert}) the integrals along the horizontal line segments tend to zero as $T\to\infty$, and then interchanging the order of summation and integration we have
\begin{align*}
J(z,\alpha)&=-\frac{4\sqrt{\alpha}e^{-\frac{z^2}{8}}}{i}\bigg(\sum_{n=1}^{\infty}\int_{1+\delta-i\infty}^{1+\delta+i\infty}\frac{\pi}{\sin\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s+1}{2}\right)\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\times{}_1F_{1}\left(1-\frac{s}{2};\frac{3}{2};\frac{z^2}{4}\right)\left(\sqrt{\pi}\alpha n\right)^{-s}\, ds-2\pi iL\bigg),
\end{align*}
where
\begin{equation*}
L=\lim_{s\to 1}(s-1)\zeta(s)\frac{\pi}{\sin\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s+1}{2}\right){}_1F_{1}\left(1-\frac{s}{2};\frac{3}{2};\frac{z^2}{4}\right)\left(\sqrt{\pi}\alpha\right)^{-s}.
\end{equation*}
It is easy to see that
\begin{equation*}
L=\frac{\sqrt{\pi}}{\alpha}{}_1F_{1}\left(\frac{1}{2};\frac{3}{2};\frac{z^2}{4}\right).
\end{equation*}
Now using the fact \cite[p.~98]{kp} that for $0<c=$ Re $s<2$,
\begin{equation*}
\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\frac{\pi}{\sin\left(\frac{1}{2}\pi s\right)}x^{-s}\, ds=\frac{2}{(1+x^2)},
\end{equation*}
combined with the special case when $b=z\neq 0$, $a=1$ of Lemma \ref{corinvsinl}, and (\ref{melconv}), we see that
\begin{align*}
J(z,\alpha)=-8\pi\sqrt{\alpha}e^{-\frac{z^2}{8}}\bigg(\frac{4e^{\frac{z^2}{4}}}{z}\sum_{n=1}^{\infty}\int_{0}^{\infty}\frac{e^{-x^2}\sin xz}{1+\left(\frac{\sqrt{\pi}\alpha n}{x}\right)^{2}}\, \frac{dx}{x}-\frac{\sqrt{\pi}}{\alpha}{}_1F_{1}\left(\frac{1}{2};\frac{3}{2};\frac{z^2}{4}\right)\bigg).
\end{align*}
Employ the change of variable $x\to\sqrt{\pi}\alpha x$ and \eqref{kft} to see that
\begin{align}\label{jzabc}
J(z,\alpha)=-8\pi\sqrt{\alpha}e^{\frac{z^2}{8}}\bigg(\frac{4e^{\frac{z^2}{4}}}{z}\sum_{n=1}^{\infty}\int_{0}^{\infty}\frac{xe^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)}{n^2+x^2}\, dx-\frac{\sqrt{\pi}}{\alpha}{}_1F_{1}\left(1;\frac{3}{2};\frac{-z^2}{4}\right)\bigg).
\end{align}
Now for $t\neq 0$ \cite[p.~191]{con},
\begin{equation}\label{cotid1}
\sum_{n=1}^{\infty}\frac{1}{t^2+n^2}=\frac{\pi}{t}\left(\frac{1}{e^{2\pi t}-1}-\frac{1}{2\pi t}+\frac{1}{2}\right).
\end{equation}
Thus, interchanging the order of summation and integration in (\ref{jzabc}) and then substituting (\ref{cotid1}) and simplifying, we observe that
\begin{align}\label{jzabc1}
J(z,\alpha)&=-8\pi\sqrt{\alpha}e^{\frac{z^2}{8}}\bigg(\frac{4\pi}{z}\int_{0}^{\infty}\frac{e^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)}{e^{2\pi x}-1}\, dx-\frac{2}{z}\int_{0}^{\infty}\frac{e^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)}{x}\, dx\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad+\frac{2\pi}{z}\int_{0}^{\infty}e^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)\, dx-\frac{\sqrt{\pi}}{\alpha}{}_1F_{1}\left(1;\frac{3}{2};\frac{-z^2}{4}\right)\bigg).
\end{align}
However from \cite[p.~488, formula 3.896, no. 3]{grn},
\begin{equation}\label{dawson}
\int_{0}^{\infty}e^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)\, dx=\frac{z}{2\sqrt{\pi}\alpha}{}_1F_{1}\left(1;\frac{3}{2};\frac{-z^2}{4}\right)
\end{equation}
and by \cite[p.~503, formula 3.952, no. 7]{grn} and \cite[p.~889, formula 8.253, no. 1]{grn} (see also \cite[p.~421]{glashier2}),
\begin{equation}\label{errapp}
\int_{0}^{\infty}\frac{e^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)}{x}\, dx=\frac{\pi}{2}\textup{erf}\left(\frac{z}{2}\right).
\end{equation}
Thus, substituting (\ref{dawson}) and (\ref{errapp}) in (\ref{jzabc1}) and simplifying, we finally arrive at
\begin{align}\label{mrramg1}
\frac{1}{8\pi^2}J(z,\alpha)=\frac{\sqrt{\alpha}e^{\frac{z^2}{8}}}{z}\left(\textup{erf}\left(\frac{z}{2}\right)-4\int_{0}^{\infty}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx\right).
\end{align}
Using \eqref{errrel}, it is clear that simultaneously replacing $\alpha$ by $\beta$ and $z$ by $iz$ in (\ref{mrramg1}) and employing \eqref{rel} and (\ref{errrel}) give (\ref{mrramg}) since $J(z,\alpha)$ is invariant.
\end{proof}
\begin{proof}[Theorem \textup{\ref{thmrgenrchi}}][]
We prove the theorem only for odd real $\chi$. The case when $\chi$ is even and real can be similarly proved. Let $\phi(s)=\Gamma\left(\frac{1}{4}+\frac{s}{2}\right)$ and let
\begin{equation*}
P(z,\alpha,\chi)=\int_{0}^{\infty}\Gamma\left(\frac{1+it}{4}\right)\Gamma\left(\frac{1-it}{4}\right)\Xi\left(\frac{t}{2},\chi\right)\nabla\left(\alpha,z,\frac{1+it}{2}\right)\, dt.
\end{equation*}
Using the first equality in (\ref{sp4chi}), we see that
{\allowdisplaybreaks\begin{align*}
P(z,\alpha,\chi)&=\frac{2\sqrt{\alpha q}e^{-\frac{z^2}{8}}}{i\sqrt{\pi}}\int_{\frac{1}{2}-i\infty}^{\frac{1}{2}+i\infty}\Gamma\left(\frac{s}{2}\right)\Gamma\left(\frac{1-s}{2}\right)\Gamma\left(\frac{s+1}{2}\right)L(s,\chi)\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\times {}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{z^2}{4}\right)\left(\frac{\sqrt{\pi}\alpha}{\sqrt{q}}\right)^{-s}\, ds\nonumber\\
&=\frac{2\sqrt{\alpha q}e^{-\frac{z^2}{8}}}{i\sqrt{\pi}}\int_{\frac{1}{2}-i\infty}^{\frac{1}{2}+i\infty}\frac{\pi}{\cos\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s}{2}\right)L(s,\chi){}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{z^2}{4}\right)\left(\frac{\sqrt{\pi}\alpha}{\sqrt{q}}\right)^{-s}\, ds,
\end{align*}
where in the last step, we used a different version of the reflection formula, namely, $\Gamma\left(\frac{1}{2}+z\right)\Gamma\left(\frac{1}{2}-z\right)=\frac{\pi}{\cos \pi z}$ for $z-\frac{1}{2}\notin\mathbb{Z}$. As before, shift the line of integration from Re $s=\frac{1}{2}$ to Re $s=1+\delta$, $0<\delta<2$, employ the residue theorem and take into account the contribution from the pole of order $1$ at $s=1$ of the integrand (due to $\cos\frac{1}{2}\pi s$). This gives,
\begin{align}\label{hzacbc}
P(z,\alpha,\chi)&=\frac{2\sqrt{\alpha q}e^{-\frac{z^2}{8}}}{i\sqrt{\pi}}\bigg(\sum_{n=1}^{\infty}\chi(n)\int_{1+\delta-i\infty}^{1+\delta+i\infty}\frac{\pi}{\cos\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s}{2}\right)\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\times{}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{z^2}{4}\right)\left(\frac{\sqrt{\pi}\alpha n}{\sqrt{q}}\right)^{-s}\, ds-2\pi iL_1\bigg),
\end{align}
where
\begin{equation}\label{l1}
L_1=\lim_{s\to 1}\frac{(s-1)\pi}{\cos\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s}{2}\right)L(s,\chi){}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{z^2}{4}\right)\left(\frac{\sqrt{\pi}\alpha}{\sqrt{q}}\right)^{-s}=-\frac{2\sqrt{q}}{\alpha}L(1,\chi).
\end{equation}
Also replacing $s$ by $(s+1)/2$, $x$ by $x^2$ in the formula \cite[p.~91, Equation (3.3.10)]{kp}
\begin{equation*}
\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\frac{x^{-s}}{\sin\pi s}\, ds=\frac{1}{\pi (1+x)},
\end{equation*}
and simplifying, we see that for $-1<$ Re $s<1$,
\begin{equation*}
\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\frac{\pi}{\cos\left(\frac{1}{2}\pi s\right)}x^{-s}\, ds=\frac{2x}{1+x^2}.
\end{equation*}
Another application of the residue theorem yields for $0<c<1$,
\begin{align}\label{anotheres}
&\int_{1+\delta-i\infty}^{1+\delta+i\infty}\frac{\pi}{\cos\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s}{2}\right){}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{z^2}{4}\right)\left(\frac{\sqrt{\pi}\alpha n}{\sqrt{q}}\right)^{-s}\, ds\nonumber\\
&=\int_{c-i\infty}^{c+i\infty}\frac{\pi}{\cos\left(\frac{1}{2}\pi s\right)}\Gamma\left(\frac{s}{2}\right){}_1F_{1}\left(\frac{1-s}{2};\frac{1}{2};\frac{z^2}{4}\right)\left(\frac{\sqrt{\pi}\alpha n}{\sqrt{q}}\right)^{-s}\, ds-\frac{4\pi i\sqrt{q}}{\alpha n}\nonumber\\
&=2\pi i\left(4e^{\frac{z^2}{4}}\int_{0}^{\infty}\frac{n}{x^2+n^2}e^{-\frac{\pi\alpha^2x^2}{q}}\cos\left(\frac{\sqrt{\pi}\alpha xz}{\sqrt{q}}\right)\, dx-\frac{2\sqrt{q}}{\alpha n}\right),
\end{align}
where in the last step we used Lemma \ref{corinvcosl} with $a=1$, $x=\frac{\sqrt{\pi}\alpha n}{\sqrt{q}}$ and $b=z$, and (\ref{melconv}), followed by a change of variable $x\to\frac{\sqrt{\pi}\alpha x}{\sqrt{q}}$. Now substitute (\ref{anotheres}) and (\ref{l1}) in (\ref{hzacbc}) and simplify to obtain
\begin{equation}\label{hzacbc1}
P(z,\alpha,\chi)=16\sqrt{\pi\alpha q}e^{\frac{z^2}{8}}\int_{0}^{\infty}\left(\sum_{n=1}^{\infty}\frac{n\chi(n)}{x^2+n^2}\right)e^{-\frac{\pi\alpha^2x^2}{q}}\cos\left(\frac{\sqrt{\pi}\alpha xz}{\sqrt{q}}\right)\, dx.
\end{equation}
Now use (\ref{watodd}) with $\chi$ real and $\nu=-1/2$ to see that
\begin{equation}\label{watodd2}
\sum_{n=1}^{\infty}\frac{n\chi(n)}{x^2+n^2}=\frac{\pi}{\sqrt{q}}\sum_{n=1}^{\infty}\chi(n)e^{-\frac{2\pi nx}{q}}.
\end{equation}
Employing (\ref{watodd2}) in (\ref{hzacbc1}), we have
\begin{equation}\label{hzacbc2}
P(z,\alpha,\chi)=16\sqrt{\pi^{3}\alpha}e^{\frac{z^2}{8}}\int_{0}^{\infty}e^{-\frac{\pi\alpha^2x^2}{q}}\cos\left(\frac{\sqrt{\pi}\alpha xz}{\sqrt{q}}\right)\sum_{n=1}^{\infty}\chi(n)e^{-\frac{2\pi nx}{q}}\, dx.
\end{equation}
Writing $n=mq+r, 0\leq m<\infty, 0\leq r\leq q-1$, and noting that $\chi$ is periodic with period $q$, we have
\begin{equation}\label{rc}
\sum_{n=1}^{\infty}\chi(n)e^{-\frac{2\pi nx}{q}}=\frac{\sum_{r=0}^{q-1}\chi(r)e^{-\frac{2\pi r x}{q}}}{1-e^{-2\pi x}}.
\end{equation}
Finally, (\ref{rc}) along with (\ref{hzacbc2}) gives
\begin{equation*}
\frac{1}{16\pi^{\frac{3}{2}}}P(z,\alpha,\chi)=\sqrt{\alpha}e^{\frac{z^2}{8}}\int_{0}^{\infty}e^{-\frac{\pi\alpha^2x^2}{q}}\cos\left(\frac{\sqrt{\pi}\alpha xz}{\sqrt{q}}\right)\frac{\sum_{r=1}^{q-1}\chi(r)e^{-\frac{2\pi r x}{q}}}{1-e^{-2\pi x}}\, dx.
\end{equation*}
This gives (\ref{rgenrchio}) as $P(z,\alpha,\chi)$ is invariant under the simultaneous application of the maps $\alpha\to\beta$ and $z\to iz$, which can be seen from \eqref{rel}.
\end{proof}
\section{The second error function transformation and an asymptotic expansion}
We first establish the second error function transformation given in Theorem \ref{thmseft} and then the asymptotic expansion from Theorem \ref{1asy}.
\begin{proof}[Theorem \textup{\ref{thmseft}}][]
Note that from \eqref{dawson} and \cite[p.~889, formula 8.253, no. 1]{grn},
\begin{equation}\label{fre}
\int_{0}^{\infty}e^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)\, dx=\frac{1}{2\a}e^{-\frac{z^2}{4}}\textup{erfi}\left(\frac{z}{2}\right).
\end{equation}
Also,
\begin{align*}
&\sqrt{\alpha}e^{\frac{z^2}{8}}\left(\textup{erf}\left(\frac{z}{2}\right)+4\int_{-\infty}^{0}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx\right)\nonumber\\
&=\sqrt{\alpha}e^{\frac{z^2}{8}}\left(\textup{erf}\left(\frac{z}{2}\right)+4\int_{0}^{\infty}e^{-\pi\alpha^2x^2}\sin\left(\sqrt{\pi}\alpha xz\right)\, dx+4\int_{0}^{\infty}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx\right),
\end{align*}
where in the first step, we replaced $x$ by $-x$ in the integral, and then simplified it using $e^{2\pi x}/(e^{2\pi x}-1)=1+1/(e^{2\pi x}-1)$. Now use \eqref{fre} and the first error function transformation in \eqref{mrramg} to replace the second integral in the above equation to obtain
\begin{align}\label{seft22}
&\sqrt{\alpha}e^{\frac{z^2}{8}}\left(\textup{erf}\left(\frac{z}{2}\right)+4\int_{-\infty}^{0}\frac{e^{-\pi\alpha^2 x^2}\sin(\sqrt{\pi}\alpha xz)}{e^{2\pi x}-1}\, dx\right)\nonumber\\
&=\sqrt{\alpha}e^{\frac{z^2}{8}}\left(2\textup{erf}\left(\frac{z}{2}\right)+\frac{1}{\a}e^{-\frac{z^2}{4}}\textup{erfi}\left(\frac{z}{2}\right)+\frac{4}{\a}e^{-\frac{z^2}{4}}\int_{0}^{\infty}\frac{e^{-\pi\b^2 x^2}\sinh(\sqrt{\pi}\b xz)}{e^{2\pi x}-1}\, dx\right)\nonumber\\
&=\sqrt{\b}e^{\frac{-z^2}{8}}\left(\textup{erfi}\left(\frac{z}{2}\right)+\frac{2}{\b}e^{\frac{z^2}{4}}\textup{erf}\left(\frac{z}{2}\right)+4\int_{0}^{\infty}\frac{e^{-\pi\b^2 x^2}\sinh(\sqrt{\pi}\b xz)}{e^{2\pi x}-1}\, dx\right),
\end{align}
where we used the fact $\a\b=1$ to simplify the last step. Now replace $\a$ by $\b$ and $z$ by $iz$ in \eqref{fre}, and use \eqref{errrel} to obtain
\begin{equation}\label{fre2}
\int_{0}^{\infty}e^{-\pi\b^2x^2}\sinh\left(\sqrt{\pi}\b xz\right)\, dx=\frac{1}{2\b}e^{\frac{z^2}{4}}\textup{erf}\left(\frac{z}{2}\right).
\end{equation}
Finally, use \eqref{fre2} to simplify the extreme right of \eqref{seft22}, thereby obtaining \eqref{seft2} and thus completing the proof.
\end{proof}
\begin{proof}[Theorem \textup{\ref{1asy}}][]
By a change of variable $x^2=t$,
\begin{equation*}
\int_{0}^{\infty}\frac{e^{-\beta x^2}\sin(\sqrt{\beta} xz)}{e^{2\pi x}-1}\, dx=\int_{0}^{\infty}\frac{e^{-\beta t}\sin(z\sqrt{\beta t})}{2\sqrt{t}(e^{2\pi\sqrt{t}}-1)}\, dt.
\end{equation*}
Let
\begin{equation*}
f(t, z)=\frac{\sin(z\sqrt{\beta t})}{2\sqrt{t}(e^{2\pi\sqrt{t}}-1)}.
\end{equation*}
First consider, for $|t|<1$,
{\allowdisplaybreaks\begin{align*}
\frac{2\pi t\sin(at)}{e^{2\pi t}-1}&=\sum_{m=0}^{\infty}\frac{B_m(2\pi)^mt^m}{m!}\sum_{n=0}^{\infty}\frac{(-1)^na^{2n+1}t^{2n+1}}{(2n+1)!}\nonumber\\
&=\sum_{j=1}^{\infty}\left(\sum_{m+2n+1=j}\frac{B_m(2\pi)^m}{m!}\frac{(-1)^na^{2n+1}}{(2n+1)!}\right)t^j\nonumber\\
&=\sum_{j=1}^{\infty}\left(\sum_{k=0\atop{k\hspace{0.5mm}\text{even}}}^{j-1}\frac{B_{j-1-k}(2\pi)^{j-1-k}}{(j-1-k)!}\frac{(-1)^{k/2}a^{k+1}}{(k+1)!}\right)t^j\nonumber\\
&=t\sum_{j=0}^{\infty}\left(\sum_{k=0\atop{k\hspace{0.5mm}\text{even}}}^{j}\frac{B_{j-k}(2\pi)^{j-k}}{(j-k)!}\frac{(-1)^{k/2}a^{k+1}}{(k+1)!}\right)t^j,
\end{align*}}
Replacing $t$ by $\sqrt{t}$ and $a=z\sqrt{\beta}$, we have for $|t|<1$,
\begin{equation*}
\frac{\sin(z\sqrt{\beta t})}{2\sqrt{t}(e^{2\pi\sqrt{t}}-1)}=\frac{1}{4\pi}\sum_{j=0}^{\infty}\left(\sum_{k=0\atop{k\hspace{0.5mm}\text{even}}}^{j}\frac{B_{j-k}(2\pi)^{j-k}}{(j-k)!}\frac{(-1)^{k/2}(z\sqrt{\beta})^{k+1}}{(k+1)!}\right)t^{\frac{j-1}{2}}
\end{equation*}
Thus, as $t\to 0+$,
\begin{equation*}
f(t, z)\sim\frac{1}{4\pi}\sum_{j=0}^{\infty}\left(\sum_{k=0\atop{k\hspace{0.5mm}\text{even}}}^{j}\frac{B_{j-k}(2\pi)^{j-k}}{(j-k)!}\frac{(-1)^{k/2}(z\sqrt{\beta})^{k+1}}{(k+1)!}\right)t^{\frac{j-1}{2}}
\end{equation*}
Hence by Watson's lemma, as $\beta\to\infty$,
\begin{equation}\label{watappl}
\int_{0}^{\infty}\frac{e^{-\beta t}\sin(z\sqrt{\beta t})}{2\sqrt{t}(e^{2\pi\sqrt{t}}-1)}\, dt\sim \sum_{j=0}^{\infty}\frac{\Gamma\left(\frac{j+1}{2}\right)}{\beta^{\frac{j+1}{2}}}\sum_{k=0\atop{k\hspace{0.5mm}\text{even}}}^{j}\frac{B_{j-k}(2\pi)^{j-k-1}}{2(j-k)!}\frac{(-1)^{k/2}(z\sqrt{\beta})^{k+1}}{(k+1)!}.
\end{equation}
From \eqref{watappl} and the notation in \eqref{gen}, we find that
\begin{equation}\label{ibetaz}
I(iz, \beta)\sim\frac{\sqrt{\pi}}{z}\beta^{-\frac{1}{4}}e^{-\frac{z^2}{8}}\textup{erfi}\left(\frac{z}{2}\right)+\sum_{j=0}^{\infty}I_{j,z},
\end{equation}
where
\begin{align*}
I_{j, z}=\frac{4}{z}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\frac{\Gamma\left(\frac{j+1}{2}\right)}{\beta^{\frac{j+1}{2}}}\sum_{k=0\atop{k\hspace{0.5mm}\text{even}}}^{j}\frac{B_{j-k}(2\pi)^{j-k-1}}{2(j-k)!}\frac{(-1)^{k/2}(z\sqrt{\beta})^{k+1}}{(k+1)!}.
\end{align*}
We first evaluate $I_{j, z}$ when $j$ is odd, say $j=2n+1$. Since all of the odd-indexed Bernoulli numbers except $B_1$ are equal to zero and $B_1=-1/2$, only the last term, namely $j=2n+1$ and $k=2n$, contributes to the sum giving
\begin{equation}\label{a14}
I_{2n+1,z}=\frac{(-1)^{n+1}n!}{(2n+1)!}z^{2n}e^{z^2/8}\beta^{-1/4}.
\end{equation}
Now let $j$ be even, say $j=2n$. Then using the fact \cite[p.~5, Equation (1.14)]{temme} that
\begin{equation*}
\frac{(-1)^{m-1}2^{2m-1}\pi^{2m}}{(2m)!}B_{2m}=\zeta(2m)
\end{equation*}
in the second step below, we see that
{\allowdisplaybreaks\begin{align}\label{evenc}
\sum_{n=0}^{\infty}I_{2n,z}&=\frac{2}{z}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\sum_{n=0}^{\infty}\frac{\Gamma\left(n+\frac{1}{2}\right)}{\beta^{n+\frac{1}{2}}}\sum_{m=0}^{n}\frac{B_{2n-2m}(2\pi)^{2n-2m-1}}{(2n-2m)!}\frac{(-1)^{m}(z\sqrt{\beta})^{2m+1}}{(2m+1)!}\nonumber\\
&=\frac{2}{z}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\sum_{n=0}^{\infty}\frac{(-1)^n(z\sqrt{\b})^{2n+1}}{\beta^{n+\frac{1}{2}}}\Gamma\left(n+\frac{1}{2}\right)\sum_{m=0}^{n}\frac{B_{2m}(2\pi)^{2m-1}}{(2m)!}\frac{(-1)^{m}(z\sqrt{\beta})^{-2m}}{(2n-2m+1)!}\nonumber\\
&=\frac{-2}{\pi z}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\sum_{n=0}^{\infty}(-1)^nz^{2n+1}\Gamma\left(n+\frac{1}{2}\right)\sum_{m=0}^{n}\frac{\zeta(2m)}{(z\sqrt{\beta})^{2m}(2n-2m+1)!}\nonumber\\
&=\frac{-2}{\pi}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\sum_{m=0}^{\infty}\frac{\zeta(2m)}{(z\sqrt{\beta})^{2m}}\sum_{n=m}^{\infty}\frac{(-1)^nz^{2n}}{(2n-2m+1)!}\G\left(n+\frac{1}{2}\right)\nonumber\\
&=\frac{-2}{\pi}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\sum_{m=0}^{\infty}\frac{(-1)^m\zeta(2m)}{\b^{m}}\G\left(m+\frac{1}{2}\right){}_1F_{1}\left(m+\frac{1}{2};\frac{3}{2};\frac{-z^2}{4}\right).\nonumber\\
\end{align}}
From \eqref{ibetaz}, \eqref{a14} and \eqref{evenc}, we obtain the asymptotic expansion of $I(iz,\b)$ as $\b\to\infty$ as
\begin{align}\label{ibetaz1}
I(iz, \beta)&\sim\left(\frac{\sqrt{\pi}}{z}e^{-\frac{z^2}{8}}\textup{erfi}\left(\frac{z}{2}\right)-e^{\frac{z^2}{8}}\sum_{m=0}^{\infty}\frac{(-1)^mm!z^{2m}}{(2m+1)!}\right)\beta^{-\frac{1}{4}}\nonumber\\
&\quad-\frac{2}{\pi}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\sum_{m=0}^{\infty}\frac{(-1)^m\zeta(2m)}{\b^{m}}\G\left(m+\frac{1}{2}\right){}_1F_{1}\left(m+\frac{1}{2};\frac{3}{2};\frac{-z^2}{4}\right).
\end{align}
Note that $\textup{erf}(z)$ has the following Taylor series expansion, which is valid for all $z\in\mathbb{C}$ \cite[p.~162, 7.6.1]{nist}:
\begin{equation}\label{erfseries}
\textup{erf}(z)=\frac{2}{\sqrt{\pi}}\sum_{n=0}^{\infty}\frac{(-1)^nz^{2n+1}}{n!(2n+1)}.
\end{equation}
From \eqref{errrel},
\begin{equation}\label{erfiseries}
\textup{erfi}(z)=\frac{2}{\sqrt{\pi}}\sum_{n=0}^{\infty}\frac{z^{2n+1}}{n!(2n+1)}.
\end{equation}
It is now easy to see that
\begin{equation}\label{giv0}
\sum_{m=0}^{\infty}\frac{(-1)^mm!z^{2m}}{(2m+1)!}=\frac{\sqrt{\pi}}{z}e^{-z^2/4}\textup{erfi}\left(\frac{z}{2}\right).
\end{equation}
Substituting \eqref{giv0} in \eqref{ibetaz1}, we arrive at
\begin{align*}
I(iz, \beta)&\sim-\frac{2}{\pi}\beta^{\frac{1}{4}}e^{\frac{z^2}{8}}\sum_{m=0}^{\infty}\frac{(-1)^m\zeta(2m)}{\b^{m}}\G\left(m+\frac{1}{2}\right){}_1F_{1}\left(m+\frac{1}{2};\frac{3}{2};\frac{-z^2}{4}\right).
\end{align*}
Since $\a\b=\pi^2$, using \eqref{gen}, this also gives the asymptotic expansion of $I(z,\a)$ as $\a\to 0$ as claimed in \eqref{1asyeq}.
\end{proof}
The second error function transformation in Theorem \ref{thmseft} can be rephrased for $\a\b=\pi^2$ and $z\neq 0$ as follows:
\begin{align}\label{sefty}
K(z,\a)&:=\frac{\sqrt{\pi}}{z}\alpha^{-\frac{1}{4}}e^{\frac{z^2}{8}}\textup{erf}\left(\frac{z}{2}\right)-\frac{4}{z}\alpha^{\frac{1}{4}}e^{-\frac{z^2}{8}}\int_{-\infty}^{0}\frac{e^{-\alpha x^2}\sinh(\sqrt{\a} xz)}{e^{2\pi x}-1}\, dx\nonumber\\
&=\frac{\sqrt{\pi}}{z}\b^{-\frac{1}{4}}e^{-\frac{z^2}{8}}\textup{erfi}\left(\frac{z}{2}\right)-\frac{4}{z}\b^{\frac{1}{4}}e^{\frac{z^2}{8}}\int_{-\infty}^{0}\frac{e^{-\b x^2}\sin(\sqrt{\b} xz)}{e^{2\pi x}-1}\, dx=:K(iz,\b).
\end{align}
The analogue of Theorem \ref{1asy} for the above identity is given below. The details are similar to those in the proof of Theorem \ref{1asy} and hence not provided.
\begin{theorem}\label{1asy2}
Fix $z\in\mathbb{C}$. As $\a\to 0$,
\begin{align*}
K(z,\a)&\sim\frac{2}{\sqrt{\pi}}\a^{-1/4}e^{z^2/8}\sum_{m=0}^{\infty}\left(\frac{-\a}{\pi^2}\right)^{m}\zeta(2m)\G\left(m+\frac{1}{2}\right){}_1F_{1}\left(m+\frac{1}{2};\frac{3}{2};\frac{-z^2}{4}\right).
\end{align*}
\end{theorem}
Combining Theorems \ref{1asy} and \ref{1asy2} leads us to the following asymptotic expansion of the integral analogue of theta function.
\begin{theorem}\label{ysa}
Fix $z\in\mathbb{C}$. As $\a\to 0$,
\begin{align*}
&\int_{-\infty}^{\infty}\frac{e^{-\a x^2}\sinh(\sqrt{\a}xz)}{e^{2\pi x}-1}\, dx\nonumber\\
&\sim\frac{-z}{\sqrt{\pi\a}}e^{z^2/4}\sum_{m=0}^{\infty}\left(\frac{-\a}{\pi^2}\right)^{m}\zeta(2m)\G\left(m+\frac{1}{2}\right){}_1F_{1}\left(m+\frac{1}{2};\frac{3}{2};\frac{-z^2}{4}\right).
\end{align*}
\end{theorem}
\section{Generalization of Ramanujan's approximation and integral identities involving hypergeometric functions}
The results from this section follow from successively differentiating the first error function transformation in the form \eqref{gen} with respect to $z$. The presence of $-x^2$ in the exponential term in the numerator of either sides justifies differentiation under the integral sign. As mentioned in the introduction, differentiating \eqref{gen} $n$ times where $n$ is odd just gives $0=0$. However when $n$ is even, two different behaviors are observed accordingly as $n\equiv0\hspace{1mm}(\text{mod}\hspace{1mm} 4)$ and as $n\equiv2\hspace{1mm}(\text{mod}\hspace{1mm} 4)$.
\subsection{The case $n\equiv0\hspace{1mm}(\text{mod}\hspace{1mm} 4)$}
\begin{proof}[Theorem \textup{\ref{thmtrn0m4}}][]
Let $n=4k$ for a non-negative integer $k$. Let
\begin{equation}\label{ga}
G(z,\a):=\frac{\sqrt{\pi}}{z}\alpha^{-\frac{1}{4}}e^{\frac{z^2}{8}}\textup{erf}\left(\frac{z}{2}\right)+\frac{4}{z}\alpha^{\frac{1}{4}}e^{-\frac{z^2}{8}}\int_{0}^{\infty}\frac{e^{-\alpha x^2}\sinh(\sqrt{\alpha} xz)}{e^{2\pi x}-1}\, dx
\end{equation}
and let
\begin{equation}\label{ika}
H_{k}(\a):=\left.\frac{d^{4k}}{dz^{4k}}G(z,\a)\right|_{z=0}.
\end{equation}
Using \eqref{gen} leads us to a transformation of the form $H_{k}(\a)=H_{k}(\b)$ since
\begin{equation}\label{ikab}
H_{k}(\a)=\left.\frac{d^{4k}}{dz^{4k}}G(z,\a)\right|_{z=0}=\left.\frac{d^{4k}}{d(iz)^{4k}}G(iz,\b)\right|_{z=0}=H_{k}(\b).
\end{equation}
Multiply the power series expansion of $e^{z^2/8}$ with that of $\textup{erf}(z)$ given in \eqref{erfseries} and extract the coefficient of $z^{4k}$ in the product. This gives
\begin{align*}
\left.\frac{d^{4k}}{dz^{4k}}\frac{e^{\frac{z^2}{8}}}{z}\textup{erf}\left(\frac{z}{2}\right)\right|_{z=0}&=\frac{2(4k)!}{\sqrt{\pi}}\sum_{m=0}^{2k}\frac{(-1)^m}{2^{4k+2m+1}m!(2k-m)!(4k-2m+1)}\nonumber\\
&=\frac{(4k)!}{2^{4k}\sqrt{\pi}(2k)!(4k+1)}{}_2F_{1}\left(-\frac{1}{2}-2k,-2k;\frac{1}{2}-2k;\frac{1}{2}\right).
\end{align*}
In \cite[p.~113, Equation (5.12)]{temme}, we find the following hypergeometric transformation, valid for $|\arg(1-z)|<\pi$:
\begin{align}\label{transformation}
{}_2F_{1}(a,b;c;z)&=\frac{\G(c)\G(b-a)}{\G(b)\G(c-a)}(1-z)^{-a}{}_2F_{1}\left(a,c-b;a-b+1;\frac{1}{1-z}\right)\nonumber\\
&\quad+\frac{\G(c)\G(a-b)}{\G(a)\G(c-b)}(1-z)^{-b}{}_2F_{1}\left(b,c-a;b-a+1;\frac{1}{1-z}\right).
\end{align}
Use this transformation with $z=1/2, a=-\frac{1}{2}-2k, b=-2k, c=\frac{1}{2}-2k$ to obtain
\begin{equation*}
{}_2F_{1}\left(-\frac{1}{2}-2k,-2k;\frac{1}{2}-2k;\frac{1}{2}\right)=\frac{(4k+1)}{2^{2k}}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right).
\end{equation*}
This gives
\begin{align}\label{ika1}
\left.\frac{d^{4k}}{dz^{4k}}\frac{e^{\frac{z^2}{8}}}{z}\textup{erf}\left(\frac{z}{2}\right)\right|_{z=0}=\frac{(4k)!}{2^{6k}\sqrt{\pi}(2k)!}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right).
\end{align}
Similarly multiplying the power series expansion of $e^{-z^2/8}$ with that of $\sinh(\sqrt{\a}xz)$ and extracting the coefficient of $z^{4k}$ in the product, we get
\begin{align*}
\left.\frac{d^{4k}}{dz^{4k}}\frac{e^{-\frac{z^2}{8}}}{z}\sinh(\sqrt{\a}xz)\right|_{z=0}&=(4k)!\sum_{m=0}^{2k}\frac{(-1)^m(\sqrt{\a}x)^{4k-2m+1}}{8^mm!(4k-2m+1)!}\nonumber\\
&=\frac{(4k)!\sqrt{\a}x}{2^{6k}}\sum_{m=0}^{2k}\frac{(-8)^{m}(\sqrt{\a}x)^{2m}}{(2k-m)!(2m+1)!},
\end{align*}
where we replaced $m$ by $2k-m$ in the last step. It can be seen that
\begin{equation*}
\sum_{m=0}^{2k}\frac{(-8)^{m}(\sqrt{\a}x)^{2m}}{(2k-m)!(2m+1)!}=\frac{1}{(2k)!}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right).
\end{equation*}
Thus
\begin{align}\label{ika2}
\left.\frac{d^{4k}}{dz^{4k}}\frac{e^{-\frac{z^2}{8}}}{z}\sinh(\sqrt{\a}xz)\right|_{z=0}=\frac{(4k)!\sqrt{\a}x}{2^{6k}(2k)!}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right).
\end{align}
Hence from \eqref{ga}, \eqref{ika}, \eqref{ika1} and \eqref{ika2}, we obtain
\begin{equation*}
H_{k}(\a)=\frac{(4k)!}{2^{6k}(2k)!}\left\{\a^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)+4\a^{3/4}\int_{0}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right)\, dx\right\},
\end{equation*}
which when combined with \eqref{ikab}, proves Theorem \ref{thmtrn0m4}.
\end{proof}
Next, we prove a generalization of Ramanujan's approximation in \eqref{near}.
\begin{proof}[Theorem \textup{\ref{thmgennear}}][]
Let $H_{k}(\b)$ be as defined in \eqref{ika} and consider the integral
\begin{equation}
\int_{0}^{\infty}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\b x^2\right)\, dx.
\end{equation}
Employing the change of variable $x=\sqrt{t}$, using the series definition of ${}_1F_{1}$ and interchanging the order of summation and integration, it is seen that
\begin{equation*}
\int_{0}^{\infty}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\b x^2\right)\, dx=\frac{1}{2}\sum_{m=0}^{2k}\frac{(-2k)_{m}(2\b)^{m}}{\left(\frac{3}{2}\right)_{m}m!}\int_{0}^{\infty}\frac{e^{-\b t}t^m}{e^{2\pi\sqrt{t}}-1}\, dt.
\end{equation*}
Now use the generating function for Bernoulli numbers to obtain, for $|t|<1$,
\begin{equation*}
\frac{t^m}{e^{2\pi\sqrt{t}}-1}=\sum_{j=0}^{\infty}\frac{B_{j}(2\pi)^{j-1}}{j!}t^{\frac{j-1+2m}{2}}.
\end{equation*}
Using Watson's lemma from \eqref{watlem1}, we find that as $\b\to\infty$,
\begin{align*}
H_{k}(\b)\sim\b^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)+2\b^{3/4}\sum_{m=0}^{2k}\frac{(-2k)_{m}(2\b)^{m}}{\left(\frac{3}{2}\right)_{m}m!}\sum_{j=0}^{\infty}\frac{B_{j}(2\pi)^{j-1}}{j!\b^{\frac{j+2m+1}{2}}}\G\left(\frac{j+2m+1}{2}\right).
\end{align*}
Since $H_{k}(\a)=H_{k}(\b)$, using the fact $\b=\pi^2/\a$ yields for $\a\to 0$,
\begin{align*}
H_{k}(\a)&\sim\frac{\a^{1/4}}{\sqrt{\pi}}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)+\frac{2}{\sqrt{\pi}}\a^{-1/4}\sum_{m=0}^{2k}\frac{(-2k)_{m}2^{m}}{\left(\frac{3}{2}\right)_{m}m!}\sum_{j=0}^{\infty}\frac{B_{j}2^{j-1}\a^{j/2}}{j!}\G\left(\frac{j+2m+1}{2}\right)\nonumber\\
&=\frac{\a^{-1/4}}{\sqrt{\pi}}\sum_{j=0\atop{j\neq 1}}^{\infty}\frac{B_{j}2^{j}\a^{j/2}}{j!}\G\left(\frac{j+1}{2}\right){}_2F_{1}\left(-2k,\frac{1+j}{2};\frac{3}{2};2\right)\nonumber\\
&=\a^{-1/4}{}_2F_{1}\left(\frac{1}{2},-2k;\frac{3}{2};2\right)+\frac{\a^{3/4}}{6}+\frac{\a^{-1/4}}{\sqrt{\pi}}\sum_{j=3}^{\infty}\frac{B_{j}2^{j}\a^{j/2}}{j!}\G\left(\frac{j+1}{2}\right){}_2F_{1}\left(-2k,\frac{1+j}{2};\frac{3}{2};2\right).
\end{align*}
We now find a simpler function, namely the one claimed on the right-hand side of \eqref{gennear}, that is ``nearly'' equal to $H_{k}(\a)$ when $\a$ is very small in the sense that the asymptotic expansion of this simpler function as $\a\to 0$ matches the first two terms in those of $H_{k}(\a)$. Note that such a function should preserve the invariance under replacing $\a$ by $\b$ and vice-versa. In order to match the leading term in the asymptotic expansion, we raise $1/\a$ to the power $1/4$ and have its coefficient as $\displaystyle {}_2F_{1}\left(\tfrac{1}{2},-2k;\tfrac{3}{2};2\right)$, which is equal to ${}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)$ by Pfaff's transformation \cite[p.~110, Equation (5.5)]{temme}
\begin{equation*}\label{pfaff}
{}_2F_{1}(a,b;c;z)=(1-z)^{-b}{}_2F_{1}\left(c-a,b;c;\frac{z}{z-1}\right).
\end{equation*}
Since the approximating function has to be symmetric, we need to raise $1/\b$ along with $1/\a$ to the power $1/4$. So the function we are seeking assumes the form
\begin{equation*}
{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)\left(\frac{1}{\a}+\frac{1}{\b}+c(k)\right)^{1/4},
\end{equation*}
where $c(k)$ is some constant depending on only $k$. Since $\a$ is very small, the main contribution in the asymptotic expansion comes from $1/\a$ and $c(k)$ but not from $1/\b$. Thus, the next term in the Taylor series of this function is
\begin{equation*}
\frac{\a^{3/4}}{4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)c(k).
\end{equation*}
As we want this to be equal to $\frac{\a^{3/4}}{6}$, it is clear that $c(k)=\frac{2}{3\cdot{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)}$. Thus the required function is
\begin{equation*}
{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)\left(\frac{1}{\a}+\frac{1}{\b}+\frac{2}{3\cdot{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)}\right)^{1/4}.
\end{equation*}
This completes the proof of \eqref{gennear}.
\end{proof}
\subsection{The case $n\equiv2\hspace{1mm}(\text{mod}\hspace{1mm} 4)$}
\begin{proof}[Theorem \textup{\ref{thmtrn2m4}}][]
Now let $n=4k+2$, where $k$ is again any non-negative integer. Let
\begin{equation*}
J_{k}(\a):=\left.\frac{d^{4k+2}}{dz^{4k+2}}G(z,\a)\right|_{z=0},
\end{equation*}
where $G(z,\a)$ is defined in \eqref{ga}. This time \eqref{gen} gives us a transformation of the form $J_{k}(\a)=-J_{k}(\b)$ since
\begin{equation*}
J_{k}(\a)=\left.\frac{d^{4k+2}}{dz^{4k+2}}G(z,\a)\right|_{z=0}=\left.\frac{d^{4k+2}}{dz^{4k+2}}G(iz,\b)\right|_{z=0}=-\left.\frac{d^{4k+2}}{d(iz)^{4k+2}}G(iz,\b)\right|_{z=0}=-J_{k}(\b).
\end{equation*}
The proof is now similar to that of Theorem \ref{thmtrn0m4} and so we state only the important steps. Let us start with the fact that
\begin{equation}\label{jka0}
-J_{k}(\b)=\left.\frac{d^{4k+2}}{dz^{4k+2}}G(iz,\b)\right|_{z=0}.
\end{equation}
Multiply the power series expansion of $e^{-z^2/8}$ with that of $\textup{erfi}\left(\frac{z}{2}\right)$ given in \eqref{erfiseries} and extract the coefficient of $z^{4k+2}$ in the product. Then identifying the coefficient as a hypergeometric function and using the transformation \eqref{transformation} with $z=1/2, a=-\frac{3}{2}-2k, b=-1-2k, c=-\frac{1}{2}-2k$ to simplify this hypergeometric function results in
\begin{align*}
\left.\frac{d^{4k+2}}{dz^{4k+2}}\frac{e^{\frac{-z^2}{8}}}{z}\textup{erfi}\left(\frac{z}{2}\right)\right|_{z=0}=\frac{(4k+2)!}{2^{6k+3}\sqrt{\pi}(2k+1)!}{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right).
\end{align*}
Similarly,
{\allowdisplaybreaks\begin{align}\label{jka2}
\left.\frac{d^{4k+2}}{dz^{4k+2}}\frac{e^{\frac{z^2}{8}}}{z}\sin(\sqrt{\b}xz)\right|_{z=0}&=-(4k+2)!\sum_{m=0}^{2k+1}\frac{(-1)^m(\sqrt{\b}x)^{4k-2m+3}}{8^mm!(4k-2m+3)!}\nonumber\\
&=\frac{(4k+2)!\sqrt{\b}x}{2^{6k+3}}\sum_{m=0}^{2k+1}\frac{(-8)^{m}(\sqrt{\b}x)^{2m}}{(2k+1-m)!(2m+1)!}\nonumber\\
&=\frac{(4k+2)!\sqrt{\b}x}{2^{6k+3}(2k+1)!}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\b x^2\right).
\end{align}}
From \eqref{jka0} and \eqref{jka2},
\begin{align*}
-J_{k}(\b)&=\frac{(4k+2)!}{2^{6k+3}(2k+1)!}\bigg\{\b^{-1/4}{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right)\nonumber\\
&\quad\quad\quad\quad\quad\quad\quad\quad+4\b^{3/4}\int_{0}^{\infty}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\b x^2\right)\, dx\bigg\}.
\end{align*}
This proves Theorem \ref{thmtrn2m4}.
\end{proof}
As remarked in the introduction, results corresponding to the ones in Theorem \ref{thmtrn0m4} - Corollary \ref{corexact} can be obtained using similar techniques through \eqref{sefty}. These results are collectively put in the following theorem. We refrain from giving the proof since the details are similar to those of Theorem \ref{thmtrn0m4} - Corollary \ref{corexact}.
\begin{theorem}\label{appseft}
Let $\a, \b$ be two positive numbers such that $\a\b=\pi^2$. Then
{\allowdisplaybreaks\begin{align}
&\textup{(i)}\hspace{2mm} \a^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)-4\a^{3/4}\int_{-\infty}^{0}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&\quad\quad=\b^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)-4\b^{3/4}\int_{-\infty}^{0}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\b x^2\right)\, dx.\nonumber\\
&\textup{(ii)}\hspace{2mm} \a^{-1/4}{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)-4\a^{3/4}\int_{-\infty}^{0}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&\quad\quad=-{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)\left(\frac{1}{\a}+\frac{1}{\b}+\frac{2}{3\cdot{}_2F_{1}\left(-2k,1;\frac{3}{2};2\right)}\right)^{1/4}, \text{``nearly''}.\nonumber\\
&\textup{(iii)}\hspace{2mm} -\a^{-1/4}{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right)+4\a^{3/4}\int_{-\infty}^{0}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\a x^2\right)\, dx\nonumber\\
&=\b^{-1/4}{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right)-4\b^{3/4}\int_{-\infty}^{0}\frac{xe^{-\b x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\b x^2\right)\, dx.\nonumber
\end{align}}
In particular when $\a=\b=\pi$, we have
\begin{equation}\label{exact1}
\int_{-\infty}^{0}\frac{xe^{-\pi x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\pi x^2\right)\, dx=\frac{1}{4\pi}\cdot{}_2F_{1}\left(-2k-1,1;\frac{3}{2};2\right).
\end{equation}
\end{theorem}
\section{Generalization of an asymptotic expansion of Oloa}
We first explain how the integral in \eqref{ramanujan-98} is related to the one on the extreme right side of \eqref{mrram}. Write the latter integral as
\begin{equation}\label{amrram}
\int_{0}^{\infty}(1+t^2)\Gamma\left(\frac{-1+it}{4}\right)\Gamma\left(\frac{-1-it}{4}\right)\frac{\Xi\left(\frac{t}{2}\right)}{1+t^2}\cos \left(\frac{1}{2}t\log\alpha\right)\, dt.
\end{equation}
If we now square the expression $\frac{\Xi\left(\frac{t}{2}\right)}{1+t^2}$ in \eqref{amrram}, then as discussed in \cite{dixitmoll}, this amounts to squaring the functional equation of the Riemann zeta function, and moreover the squared expression admits a generalization
\begin{equation*}
\frac{
\Xi \left( \frac{t + i z}{2} \right)
\Xi \left( \frac{t - i z}{2} \right) }
{ (t^{2} + (z+1)^{2})(t^{2}+ (z-1)^{2})}.
\end{equation*}
This is what Ramanujan may have had at the back of his mind when he worked \cite[Section 5]{riemann} with the generalization
\begin{equation*}
\int_{0}^{\infty}(t^2+(z-1)^2)\Gamma\left(\frac{z-1+it}{4}\right)\Gamma\left(\frac{z-1-it}{4}\right)\frac{\Xi \left( \frac{t + i z}{2} \right)
\Xi \left( \frac{t - i z}{2} \right) }
{ (t^{2} + (z+1)^{2})(t^{2}+ (z-1)^{2})}\cos \left(\frac{1}{2}t\log\alpha\right)\, dt,
\end{equation*}
of \eqref{amrram}, which upon simplification gives \eqref{ramanujan-98}. Ramanujan \cite{riemann} obtains a transformation formula associated with this integral. In \cite{dixitmoll}, Moll and one of the authors found the following new representation of this transformation, which generalizes a transformation of Koshliakov \cite[Equation (6)]{koshli} \cite[Equations (21), (27)]{kosh1937}.\\
\textit{Assume $-1 <$ \textup{Re} $z < 1$. Let $\Omega(x, z)$ be defined by
\begin{equation*}
\Omega(x,z) = 2 \sum_{n=1}^{\infty} \sigma_{-z}(n) n^{z/2}
\left( e^{\pi i z/4} K_{z}( 4 \pi e^{\pi i/4} \sqrt{nx} ) +
e^{-\pi i z/4} K_{z}( 4 \pi e^{-\pi i/4} \sqrt{nx} ) \right),
\end{equation*}
where $\sigma_{-z}(n)=\sum_{d|n}d^{-z}$ and $K_\nu(z)$ denotes the modified Bessel function of order $\nu$. Then for $\alpha, \beta>0, \alpha\beta=1$,
\begin{align}\label{genelkosh}
&\alpha^{(z+1)/2}
\int_{0}^{\infty} e^{-2 \pi \alpha x} x^{z/2}
\left( \Omega(x,z) - \frac{1}{2 \pi} \zeta(z) x^{z/2-1} \right) dx\nonumber \\
&=\beta^{(z+1)/2}
\int_{0}^{\infty} e^{-2 \pi \beta x} x^{z/2}
\left( \Omega(x,z) - \frac{1}{2 \pi} \zeta(z) x^{z/2-1} \right) dx\nonumber\\
&=\frac{1}{2\pi^{(z+5)/2}}
\int_{0}^{\infty}
\Gamma \left( \frac{z-1+it}{4} \right)
\Gamma \left( \frac{z-1-it}{4} \right)
\Xi \left( \frac{t + iz}{2} \right)
\Xi \left( \frac{t - iz}{2} \right)\frac{\cos( \tfrac{1}{2} t \log \alpha) \, dt}{(t^{2} + (z+1)^{2})}.
\end{align}}
Theorem \ref{2asy} is now proved using \eqref{genelkosh}.
\begin{proof}[Theorem \textup{\ref{2asy}}][]
We obtain the asymptotic expansion of the integral indirectly by obtaining the same for the integral on the extreme left of \eqref{genelkosh}. Let
\begin{equation*}
g(t, z):=t^{z/2}
\left( \Omega(t,z) - \frac{1}{2 \pi} \zeta(z) t^{z/2-1} \right).
\end{equation*}
We use the following identity established in \cite[Proposition 6.1]{dixitmoll}.
\begin{equation*}
\Omega(t,z) = - \frac{\Gamma(z) \zeta(z)}{(2 \pi \sqrt{t})^{z}} +
\frac{t^{z/2-1}}{2 \pi} \zeta(z) -
\frac{t^{z/2}}{2} \zeta(z+1) +
\frac{t^{z/2+1}}{\pi} \sum_{n=1}^{\infty} \frac{\sigma_{-z}(n)}{n^{2}+t^{2}}.
\end{equation*}
Since for $|t|<1$,
{\allowdisplaybreaks\begin{align*}
\sum_{n=1}^{\infty} \frac{\sigma_{-z}(n)}{n^{2}+t^{2}}&=\sum_{n=1}^{\infty} \frac{\sigma_{-z}(n)}{n^2}\frac{1}{\left(1+\left(\frac{t}{n}\right)^{2}\right)}\nonumber\\
&=\sum_{n=1}^{\infty}\frac{\sigma_{-z}(n)}{n^2}\sum_{m=0}^{\infty}(-1)^m\left(\frac{t}{n}\right)^{2m}\nonumber\\
&=\sum_{m=0}^{\infty}(-1)^mt^{2m}\sum_{n=1}^{\infty}\frac{\sigma_{-z}(n)}{n^{2m+2}}\nonumber\\
&=\sum_{m=0}^{\infty}(-1)^m\zeta(2m+2)\zeta(2m+2+z)t^{2m},
\end{align*}
we see that if
\begin{equation*}
h(t,z):=g(t,z)+\frac{\Gamma(z) \zeta(z)}{(2 \pi)^{z}}+\frac{t^{z}}{2} \zeta(z+1),
\end{equation*}
then
\begin{equation*}
h(t,z)=\sum_{m=0}^{\infty}\frac{(-1)^m}{\pi}\zeta(2m+2)\zeta(2m+2+z)t^{2m+z+1},
\end{equation*}
so also as $t\to 0$,
\begin{equation*}
h(t,z)\sim\sum_{m=0}^{\infty}\frac{(-1)^m}{\pi}\zeta(2m+2)\zeta(2m+2+z)t^{2m+z+1}.
\end{equation*}
We now apply Lemma \ref{watlem} with $\lambda=(z+2)/2$ and $\mu=1/2$. The condition $-1<$ Re $z<1$ guarantees that Re $\lambda>0$ which is necessary as remarked after Lemma \ref{watlem}. Then as $\a\to\infty$,
\begin{equation}\label{aftwat1}
\int_{0}^{\infty}e^{-2\pi\a t}h(t, z)\, dt\sim\sum_{m=0}^{\infty}\frac{(-1)^m}{\pi(2\pi\a)^{2m+z+2}}\G(2m+2+z)\zeta(2m+2)\zeta(2m+2+z).
\end{equation}
Note that
\begin{align}\label{aftwat2}
\int_{0}^{\infty}e^{-2\pi\a t}h(t, z)\, dt&=\int_{0}^{\infty}e^{-2\pi\a t}g(t, z)\, dt+\frac{\Gamma(z) \zeta(z)}{(2 \pi)^{z}}\int_{0}^{\infty}e^{-2\pi\a t}\, dt\nonumber\\
&\quad+\frac{\zeta(z+1)}{2}\int_{0}^{\infty}e^{-2\pi\a t}t^{z}\, dt\nonumber\\
&=\int_{0}^{\infty}e^{-2\pi\a t}g(t, z)\, dt+\frac{\Gamma(z) \zeta(z)}{\a(2 \pi )^{z+1}}+\frac{\zeta(z+1)\G(z+1)}{2(2 \pi\a )^{z+1}}.
\end{align}
From \eqref{aftwat1} and \eqref{aftwat2} and \eqref{genelkosh}, one now obtains \eqref{oloag} after simplification.
\end{proof}
\section{Concluding remarks and some open questions}
1. In this paper, we found two new transformations involving error functions, namely the ones in Theorems \ref{genmrramg} and \ref{thmseft}, which when combined give Ramanujan's transformation \eqref{rt} (or equivalently \eqref{fwttrans}) for an integral analogue of theta functions, thus giving a better understanding of Ramanujan's transformation. Also, the results in Theorem \ref{ramtran} could have been obtained directly without having to resort to Theorems \ref{thmtrn0m4} - Corollary \ref{corexact} and Theorem \ref{appseft}. However, obtaining Theorem \ref{ramtran} from these theorems is useful since they give us many interesting results which otherwise would not have been revealed. For example, one could have proved \eqref{exact0} directly through \eqref{dires}. However, proving it through \eqref{exact} and \eqref{exact1} gives those non-trivial integral evaluations in addition.
In light of the existence of the integral on the extreme right side of \eqref{mrramg} which equals two sides of the first error function transformation, it is natural to ask if a similar such integral exists for the second error function transformation in \eqref{thmseft}. We were unable to find such an integral and so we leave it as an open problem. However, it is important to state here the difficulty in finding this integral, if it exists.
If we reverse the steps used in proving the equality of the extreme sides of the transformation \eqref{mrramg} in Theorem \ref{genmrramg}, we notice that a crucial step was to use the integral representation for the error function given in \eqref{errapp}. However, employing the same method to the left-hand side of \eqref{thmseft} does not help as the error function there does not cancel with the integral $\displaystyle\frac{2}{\pi}\int_{0}^{\infty}\frac{e^{-\pi\alpha^2x^2+2\pi x}\sin\left(\sqrt{\pi}\alpha xz\right)}{x}\, dx$. Another reason why this is a difficult problem is that while the Mellin transform of $e^{-ax^2}\sin bx$ is essentially just a ${}_1F_{1}$ (see Lemma \ref{corinvsinl}), that of $e^{-ax^2-cx}\sin bx$ involves parabolic cylinder functions \cite[p.~503, formula 3.953, no. 1]{grn}.
We now explain the significance of this integral, provided it exists. As remarked in the introduction, subtracting the first error function transformation in \eqref{mrramg} from the second given in \eqref{seft2} leads to Ramanujan's transformation \eqref{rt} for, what is called, an integral analogue of the Jacobi theta function. The corresponding transformation for the Jacobi theta function, which has an integral involving $\Xi(t)$ equal to it \cite[Theorem 1.2]{dixthet} as well, is
\begin{align}\label{eqsym0}
&\sqrt{\alpha}\bigg(\frac{e^{-\frac{z^2}{8}}}{2\alpha}-e^{\frac{z^2}{8}}\sum_{n=1}^{\infty}e^{-\pi\alpha^2n^2}\cos(\sqrt{\pi}\alpha nz)\bigg)&=\sqrt{\beta}\bigg(\frac{e^{\frac{z^2}{8}}}{2\beta}-e^{-\frac{z^2}{8}}\sum_{n=1}^{\infty}e^{-\pi\beta^2n^2}\cos(i\sqrt{\pi}\beta nz)\bigg)\nonumber\\
&=\frac{1}{\pi}\int_{0}^{\infty}\frac{\Xi(t/2)}{1+t^2}\nabla\left(\alpha,z,\frac{1+it}{2}\right)\, dt,
\end{align}
where the $\nabla$ function is defined in \eqref{nabla}.
The equality of the extreme left and right sides of the special case $z=0$ of the above identity was used by Hardy \cite[Eqn.(2)]{hardyfrench} to prove that infinitely many non-trivial zeros of $\zeta(s)$ lie on the critical line Re $s=\frac{1}{2}$. Thus if an integral involving $\Xi(t)$ equal to both sides of \eqref{rt} is found, then this integral analogue of Hardy's formula may be used to obtain more information on the non-trivial zeros of $\zeta(s)$. However, this requires us to first obtain an integral involving $\Xi(t)$ equal to the two sides of \eqref{seft2}.
Further, since our results involve an extra parameter $z$, it may be important to see what else about $\zeta(s)$, or some generalization of it, could be extracted from them. It would also be worth further studying these two error function transformations from the point of view of further applications in analytic number theory.
\textbf{Remark.} Hardy \cite{ghh} conjectured that Ramanujan's formula \eqref{mrram} may also be used for proving the infinitude of the zeros of $\zeta(s)$ on the critical line. However, we believe that it is not this formula but rather the special case $z=0$ of the identity which has an integral involving $\Xi(t)$ equal to both sides of \eqref{rt} which leads to the existence of infinitely many zeros.\\
2. Consider the transformation in \eqref{mrram} and its equivalent version \eqref{morpara} given by Mordell. Let $q=e^{i\pi w}$, Im $w>0$, and let
\begin{equation}
\Lambda(w):=\sum_{n=1}^{\infty}F(n)q^n,
\end{equation}
where $F(D)$ denotes the number of classes of positive definite binary quadratic forms $ax^2+2hxy+by^2$ with $a, b$ not both even, and determinant $-D$. Then Mordell \cite[Equation (2.18)]{mordell2} proved that
\begin{align}
\int_{0}^{\infty}\frac{xe^{\pi iwx^2}}{e^{2\pi x}-1}\, dx=-\frac{i}{4\pi w}-\Lambda(w)+\frac{\sqrt{-iw}}{w^2}\Lambda\left(-\frac{1}{w}\right)+\frac{1}{8}\left(\sum_{n=-\infty}^{\infty}e^{i\pi n^2 w}\right)^3,
\end{align}
so that with $\alpha^2=-iw$, we have for Re $\alpha^2>0$,
\begin{align}
\int_{0}^{\infty}\frac{xe^{-\pi\alpha^2x^2}}{e^{2\pi x}-1}\, dx=\frac{-1}{4\pi\alpha^2}-\sum_{n=1}^{\infty}F(n)e^{-\pi n\alpha^2}-\frac{1}{\alpha^3}\sum_{n=1}^{\infty}F(n)e^{-\pi n/\alpha^2}+\frac{1}{8}\left(\sum_{n=-\infty}^{\infty}e^{-\pi\alpha^2n^2 }\right)^3.
\end{align}
It will be interesting whether the above result admits a generalization when we work with the integral in \eqref{mrramg}.\\
3. For a fixed $z\in\mathbb{C}$, consider the integral
\begin{equation}\label{genexact}
\int_{-\infty}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(z;\frac{3}{2};2\a x^2\right)\, dx.
\end{equation}
Using the asymptotic expansion of the confluent hypergeometric function \cite[p.~508, Equation 13.5.1]{stab}, it can be seen that as $|x|\to\infty$,
\begin{equation*}
{}_1F_{1}\left(z;\frac{3}{2};2\a x^2\right)\sim\frac{\sqrt{\pi}}{2}\left(\frac{e^{i\pi z}(2\a x^2)^{-z}}{\G\left(\frac{3}{2}-z\right)}+\frac{e^{2\a x^2}(2\a x^2)^{z-\frac{3}{2}}}{\G(z)}\right).
\end{equation*}
Note that because of the presence of $e^{2\a x^2}$ in the second expression of the asymptotic expansion, and since $\a>0$, the only way the integral in \eqref{genexact} can converge is if this expression is annihilated by $\G(z)$. This happens only when $z$ is a non-positive integer. This leads us to consider two cases based on the parity of such $z$.\\
\textbf{Case 1: $z$ is a non-positive even integer.} Note that for $\a$ either very small or very large, the integral in \eqref{gennearg} is nicely approximated by the expression on its right side, as in this case $\b$ is respectively very large or very small. However, the case $\a=\b=\pi$ is the worst in terms of approximating this integral, i.e., the integral
\begin{equation*}
\int_{0}^{\infty}\frac{xe^{-\pi x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\pi x^2\right)\, dx,
\end{equation*}
since then $\a$ and $\b$ are not only of the same order of magnitude but in fact equal. Table 1 below shows how the integral in \eqref{gennearg} is approximated by the right side of \eqref{gennearg} for some small values of $\a$. (The calculations in this table are done for the identity obtained by dividing both sides of \eqref{gennearg} by $\a^{3/4}$. They are performed in \emph{Mathematica}.)\\
\textbf{Case 2: $z$ is a non-positive odd integer.} When $\a=\pi$ in the integral \eqref{genexact}, we have shown in \eqref{exact0} that it is equal to zero.\\
Thus there is a trade-off in that \eqref{gennearg} has no restriction on $\a$ (except $\a>0$) but is an approximation, where as we can exactly evaluate the integral \eqref{genexact}, but only for a specific value of $\a$, i.e., when $\a=\pi$. This leads us to two open questions:
\textbf{Question 1:} Find an exact evaluation of $\displaystyle\int_{0}^{\infty}\frac{xe^{-\pi x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k;\frac{3}{2};2\pi x^2\right)\, dx$ for $k\in\mathbb{Z^{+}}\cup\{0\}$.\\
\textbf{Question 2:} Find an exact evaluation of, or at least an approximation to, the integral $\displaystyle\int_{0}^{\infty}\frac{xe^{-\a x^2}}{e^{2\pi x}-1}{}_1F_{1}\left(-2k-1;\frac{3}{2};2\a x^2\right)\, dx$ when $\a\neq\pi$ is a positive real number and $k\in\mathbb{Z^{+}}\cup\{0\}$.\\
It is interesting to note that in Theorem \ref{thmrgenrchi}, the integral involving $\Xi(t,\chi)$ involves the $\Delta$ function when $\chi$ is even and the $\nabla$ function when $\chi$ is odd, which is exactly opposite of what happens in Theorems 1.3, 1.4 and 1.5 in \cite{drz3}. Besides the fact that, in doing so, one can explicitly evaluate the Mellin transforms and that one \emph{does} get what one is looking for, is there some intrinsic reason behind this reversal?\\
\begin{table}[ht]
\caption{ Both side of \eqref{gennearg} (after dividing throughout by $\a^{3/4}$)}
\begin{center}
\renewcommand{\arraystretch}{1.5}
{\tiny
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c}
\hline
\multicolumn{1}{|c|}{$\alpha$} & \multicolumn{2}{|c|}{$.0000009$} &\multicolumn{2}{|c|}{$.000007$} & \multicolumn{2}{|c|}{$1.5$} & \multicolumn{2}{|c|}{$2.378$}& \multicolumn{2}{|c|}{$9361.79$}\\
\hline
\multicolumn{1}{|c|}{k} & \multicolumn{1}{|c|}{LHS} & \multicolumn{1}{|c|}{RHS} & \multicolumn{1}{|c|}{LHS} & \multicolumn{1}{|c|}{RHS} & \multicolumn{1}{|c|}{LHS} & \multicolumn{1}{|c|}{RHS} & \multicolumn{1}{|c|}{LHS} & \multicolumn{1}{|c|}{RHS} & \multicolumn{1}{|c|}{LHS} & \multicolumn{1}{|c|}{RHS} \\
\hline
\multicolumn{1}{|c|}{1} & \multicolumn{1}{|c|}{$259259$} & \multicolumn{1}{|c|}{$259259$} & \multicolumn{1}{|c|}{$33333.4$} & \multicolumn{1}{|c|}{$33333.4$} & \multicolumn{1}{|c|}{$.212975$} & \multicolumn{1}{|c|}{$.210775$}& \multicolumn{1}{|c|}{$0.1483410$} & \multicolumn{1}{|c|}{$0.1465060$} & \multicolumn{1}{|c|}{$0.00136109$} & \multicolumn{1}{|c|}{$0.001361096$}\\
\hline
\multicolumn{1}{|c|}{2} & \multicolumn{1}{|c|}{$188713$} & \multicolumn{1}{|c|}{$188713$} & \multicolumn{1}{|c|}{$24263.1$} & \multicolumn{1}{|c|}{$24263.1$} & \multicolumn{1}{|c|}{$.162014$} & \multicolumn{1}{|c|}{$.161821$}& \multicolumn{1}{|c|}{$0.112982$} & \multicolumn{1}{|c|}{$0.112883$}& \multicolumn{1}{|c|}{$0.000990862$} & \multicolumn{1}{|c|}{$0.000990862$}\\
\hline
\multicolumn{1}{|c|}{3} & \multicolumn{1}{|c|}{$154475$} & \multicolumn{1}{|c|}{$154475$} & \multicolumn{1}{|c|}{$19861.2$} & \multicolumn{1}{|c|}{$19861.2$} & \multicolumn{1}{|c|}{$.135921$} & \multicolumn{1}{|c|}{$.137363$}& \multicolumn{1}{|c|}{$0.0948065$} & \multicolumn{1}{|c|}{$0.0960151$}& \multicolumn{1}{|c|}{$0.000811187$} & \multicolumn{1}{|c|}{$0.000811187$}\\
\hline
\multicolumn{1}{|c|}{4} & \multicolumn{1}{|c|}{$133517$} & \multicolumn{1}{|c|}{$133517$} & \multicolumn{1}{|c|}{$17166.6$} & \multicolumn{1}{|c|}{$17166.6$} & \multicolumn{1}{|c|}{$.11939$} & \multicolumn{1}{|c|}{$.122057$} & \multicolumn{1}{|c|}{$0.0832805$} & \multicolumn{1}{|c|}{$0.085431$} & \multicolumn{1}{|c|}{$0.000701201$} & \multicolumn{1}{|c|}{$0.000701201$} \\
\hline
\multicolumn{1}{|c|}{5} & \multicolumn{1}{|c|}{$119074$} & \multicolumn{1}{|c|}{$119074$} & \multicolumn{1}{|c|}{$15309.6$} & \multicolumn{1}{|c|}{$15309.6$} & \multicolumn{1}{|c|}{$.107718$} & \multicolumn{1}{|c|}{$.111318$}& \multicolumn{1}{|c|}{$0.0751402$} & \multicolumn{1}{|c|}{$0.07799044$} & \multicolumn{1}{|c|}{$0.000625405$} & \multicolumn{1}{|c|}{$0.000625405$}\\
\hline
\multicolumn{1}{|c|}{6} & \multicolumn{1}{|c|}{$108375$} & \multicolumn{1}{|c|}{$108375$} & \multicolumn{1}{|c|}{$13934$} & \multicolumn{1}{|c|}{$13934$} & \multicolumn{1}{|c|}{$.0989131$} & \multicolumn{1}{|c|}{$.103239$}& \multicolumn{1}{|c|}{$0.0689983$} & \multicolumn{1}{|c|}{$0.0723852$}& \multicolumn{1}{|c|}{$0.000569256$} & \multicolumn{1}{|c|}{$0.000569256$}\\
\hline
\multicolumn{1}{|c|}{7} & \multicolumn{1}{|c|}{$100053$} & \multicolumn{1}{|c|}{$100053$} & \multicolumn{1}{|c|}{$12864$} & \multicolumn{1}{|c|}{$12864$} & \multicolumn{1}{|c|}{$.091965$} & \multicolumn{1}{|c|}{$.096872$}& \multicolumn{1}{|c|}{$0.0641517$} & \multicolumn{1}{|c|}{$0.0679618$}& \multicolumn{1}{|c|}{$0.000525582$} & \multicolumn{1}{|c|}{$0.000525583$}\\
\hline
\multicolumn{1}{|c|}{8} & \multicolumn{1}{|c|}{$93348.4$} & \multicolumn{1}{|c|}{$93348.4$} & \multicolumn{1}{|c|}{$12002$} & \multicolumn{1}{|c|}{$12002$} & \multicolumn{1}{|c|}{$.0863014$} & \multicolumn{1}{|c|}{$.0916811$}& \multicolumn{1}{|c|}{$0.060201$} & \multicolumn{1}{|c|}{$0.0643522$}& \multicolumn{1}{|c|}{$0.000490396$} & \multicolumn{1}{|c|}{$0.000490397$}\\
\hline
\multicolumn{1}{|c|}{9} & \multicolumn{1}{|c|}{$87801.4$} & \multicolumn{1}{|c|}{$87801.4$} & \multicolumn{1}{|c|}{$11288.8$} & \multicolumn{1}{|c|}{$11288.8$} & \multicolumn{1}{|c|}{$.0815698$} & \multicolumn{1}{|c|}{$.0873407$}& \multicolumn{1}{|c|}{$0.0569004$} & \multicolumn{1}{|c|}{$0.0613316$} & \multicolumn{1}{|c|}{$0.000461286$} & \multicolumn{1}{|c|}{$0.000461287$}\\
\hline
\multicolumn{1}{|c|}{10} & \multicolumn{1}{|c|}{$83116.1$} & \multicolumn{1}{|c|}{$83116.1$} & \multicolumn{1}{|c|}{$10686.4$} & \multicolumn{1}{|c|}{$10686.4$} & \multicolumn{1}{|c|}{$.0775398$} & \multicolumn{1}{|c|}{$.0836389$} & \multicolumn{1}{|c|}{$0.0540892$} & \multicolumn{1}{|c|}{$0.0587534$}& \multicolumn{1}{|c|}{$0.000436698$} & \multicolumn{1}{|c|}{$0.000436698$}\\
\hline
\end{tabular}}
\end{center}
\label{tab:multicol}
\end{table}
\begin{center}
\textbf{Acknowledgements}
\end{center}
The first author is funded in part by the grant NSF-DMS 1112656 of Professor Victor H. Moll of Tulane University and sincerely thanks him for the support.
|
\section{Introduction}
Cosmic strings can naturally arise as a result of phase transitions
followed by spontaneously symmetry breaking in the very early universe
\cite{Kibble:1976sj,Vilenkin}. It has also been argued that strings
of cosmological size can be formed in scenarios of the early universe
based on superstring theory, where they play the role of cosmic
strings \cite{Sarangi:2002yt,Jones:2003da,Dvali:2003zj}. They are
among the few direct possible features of the very early universe and
could be employed to test high-energy theories.
Cosmic strings continuously generate gravitational waves throughout
the history of the universe after their formation. Gravitational wave
bursts from different epochs and different directions overlap one
another and form a gravitational wave background over a wide range of
frequencies. Thus, gravitational wave experiments are expected to be
a powerful tool to test the existence of cosmic strings. Various types
of experiments can be used to probe the gravitational wave background
at different frequencies: pulsar timing experiments
\cite{Verbiest:2016vem,Janssen:2014dka} measure gravitational waves
at $\sim 10^{-8}$Hz; space missions such as eLISA
\cite{AmaroSeoane:2012km,AmaroSeoane:2012je} and DECIGO
\cite{Seto:2001qf,Kawamura:2011zz} explore $10^{-3}$Hz and $0.1$Hz,
respectively; ground-based experiments such as Advanced-LIGO
\cite{Harry:2010zz}, Advanced-VIRGO \cite{Accadia:2011zzc} and KAGRA
\cite{Somiya:2011np} focus on $\sim 100$Hz.
Gravitational wave signatures from cosmic strings have been
extensively investigated in the literature
\cite{Vilenkin:1981bx,Hogan:1984is,Sakellariadou:1990ne,Caldwell:1991jj,Caldwell:1996en}.
It has been widely accepted that the string network evolves towards
the scaling regime, where infinite strings continuously decay into
loops and the string network keeps ${\cal O}(1)$ infinite strings per
Hubble volume. Thus, the network consists of infinite strings and
loops, both of which can be sources of gravitational waves. In
refs. \cite{Damour:2000wa,Damour:2001bk}, it has been suggested that
non-smooth structures in strings, such as cusps and kinks, emit strong
gravitational wave bursts. Cosmic string loops generically have cusps
and kinks, and various works have shown that they generate a large
gravitational wave background at high frequencies
\cite{Damour:2001bk,Damour:2004kw,Siemens:2006yp,DePies:2007bm,Olmez:2010bi,Binetruy:2010cc,Sanidas:2012ee,Sanidas:2012tf,Binetruy:2012ze,Kuroyanagi:2012wm,Kuroyanagi:2012jf,Sousa:2013aaa}.
While loops generate gravitational waves of wavelength shorter than
the loop size, gravitational waves from infinite strings become
important for long wavelength. The spectrum of the gravitational wave
background originating from kinks on infinite strings are calculated
in ref. \cite{Kawasaki:2010yi}.
In this paper, we reexamine the spectrum of the gravitational wave
background from kinks on infinite strings. Since the strength of
gravitational wave bursts depends on the sharpness of kinks, we need
to obtain the distribution function of the sharpness to calculate the
spectrum. The evolution equation for the sharpness distribution is
modeled in ref. \cite{Copeland:2009dk}, and
ref. \cite{Kawasaki:2010yi} calculated the spectrum by using analytic
solutions of the differential equation for the distribution function.
The analytic solutions are obtained separately for radiation-dominated
(RD) and matter-dominated (MD) eras and the normalization for the RD
era is chosen to have the same amplitude with the MD era at
radiation-matter equality. Instead of using analytic solutions, we
numerically solve the differential equation to obtain the sharpness
distribution function, which enables us to smoothly connect the RD and
MD eras. In fact, since the string network evolves differently in these eras
\cite{Bennett:1987vf,Bennett:1989ak,Bennett:1989yp,Allen:1990tv,Vincent:1996rb,Vanchurin:2005pa,Ringeval:2005kr,Martins:2005es},
the parameters in the differential equation differ for MD and RD.
They should determine the normalization of the distribution function
and our numerical method correctly takes into account these effects.
The change of the parameters at radiation-matter equality is taken into account in two different ways. First, we interpolate the values using the tangent hyperbolic function. Second, we calculate the time evolution of the parameters by using the velocity-dependent one-scale (VOS) model \cite{Martins:1996jp}. In the first case, the values of the numerical parameters are set to be the same as the previous work, which makes the comparison easier and enables us to show the effect of their change at radiation-matter equality clearly. The second case enables us to follow the scaling law of the string network and provides more realistic time evolution of the parameters.
The outline of this paper is as follows. In section~\ref{sec:II}, we
briefly describe the methods to calculate the distribution function
for the kink sharpness and gravitational wave background spectrum. In
section~\ref{sec:III}, we perform the numerical calculation to
evaluate the distribution function of kinks. Then, using the kink
distribution, we calculate the spectrum of the gravitational wave
background. In section~\ref{sec:IV}, we make a comparison with
previous works. Section~\ref{sec:V} is devoted to conclusions.
\section{Gravitational wave from kinks on the infinite strings}\label{sec:II}
First, we review the dynamics of cosmic strings and describe the
method to calculate the distribution function of kink sharpness and
the power spectrum of the gravitational wave background.
\subsection{Dynamics of cosmic strings}
We consider cosmic strings in a spatially flat Friedmann-Lema{\^\i}tre-Robertson-Walker (FLRW)
metric,
\begin{equation}
{\rm d}s^2 = a^2(\tau) \left (-{\rm d} \tau^2 +{\rm d} {\bm x}^2 \right ) = g_{\mu \nu} {\rm d} x^\mu {\rm d} x^\nu \, ,
\end{equation}
where $a(\tau)$ is the scale factor of the universe. A cosmic string is
represented as
a two-dimensional worldsheet in the four-dimensional spacetime. We choose the coordinates on the worldsheet as $\zeta^1 = \tau$ (conformal time), $\zeta^2 = \sigma$ (a direction along a cosmic string),
and $\frac{\partial x^\mu}{\partial \tau}\frac{\partial x_\mu}{\partial \sigma}=0$,
then the action of the Nambu-Goto string is given by
\begin{equation}
S[x^\mu] = -\mu \int {\rm d}^2 \zeta \sqrt{-{\rm det} (\gamma_{ab})} \, , \label{eq:string_action}
\end{equation}
where $\mu$ is the tension of the string,
$\gamma_{ab} = \frac{\partial x^\mu}{\partial \zeta^a}\frac{\partial x^\nu}{\partial \zeta^b}g_{\mu\nu}$
is the induced metric on the string worldsheet.
Taking the variation of the action with respect to $x^\mu$, we obtain the equation of motion for a cosmic string,
\begin{equation}
\frac{\partial^2 {\bm x}}{\partial \tau^2} +\frac{2}{a} \frac{{\rm d} a}{{\rm d} \tau} \frac{\partial {\bm x}}{\partial \tau} \Biggl\{1 -\left (\frac{\partial {\bm x}}{\partial \tau} \right )^2 \Biggr\} = \frac{1}{\epsilon} \frac{\partial}{\partial \sigma} \left (\frac{1}{\epsilon} \frac{\partial {\bm x}}{\partial \sigma} \right ) \, , \label{eq:string_EoM}
\end{equation}
where
\begin{equation}
\epsilon \equiv \sqrt{\frac{(\partial {\bm x}/\partial \sigma)^2}{1-(\partial {\bm x}/\partial \tau)^2}} \, ,
\end{equation}
is interpreted as energy per unit $\sigma$, and we set $\epsilon=1$ at the present time.
When the Hubble friction is negligible, the equation has solutions of
left and right propagating waves. Accordingly, we define the new variable ${\bm
p}_{\pm}$ which corresponds to the left and right moving modes,
\begin{equation}
{\bm p}_{\pm} \equiv \frac{\partial {\bm x}}{\partial \tau} \mp \frac{1}{\epsilon} \frac{\partial {\bm x}}{\partial \sigma} \, . \label{eq:p_pm}
\end{equation}
\subsection{Cosmic string network}
Cosmic strings follow ``scaling law" where the number of infinite strings conserves in the horizon.
In the VOS model \cite{Martins:1996jp}, the network evolution is characterized by the correlation length $L$.
The total energy of a cosmic string and the average velocity are defined by
\begin{eqnarray}
E & = & \mu a \int {\rm d} \sigma \epsilon \label{eq:total_energy} \\
v^2 & \equiv & \frac{\int {\rm d} \sigma \left (\frac{\partial {\bm x}}{\partial \tau} \right )^2 \epsilon}{\int {\rm d} \sigma \epsilon}, \label{eq: string_v}
\end{eqnarray}
Then, the energy density $\rho_{\rm inf}$ of infinite strings is defined as
\begin{equation}
\rho_{\rm inf} = \frac{\mu}{L^2}. \label{eq:infinite_string_E}
\end{equation}
Using the physical time $t$, which relates to the conformal time as ${\rm d} t = a {\rm d} \tau$, the evolution equations of the correlation length and velocity are
\begin{eqnarray}
\frac{{\rm d} L}{{\rm d} t} & = & HL(1+v^2) +\frac{1}{2}cpv, \label{eq:Leq} \\
\frac{{\rm d} v}{{\rm d} t} & = & (1-v^2) \left (\frac{k}{L} -2Hv \right ), \label{eq:veq}
\end{eqnarray}
where
$k(v) \equiv \frac{1}{v (1 -v^2)}\frac{\int {\rm d} \sigma \bigl\{1-({\rm d} {\bm x}/{\rm d} \tau)^2 \bigr\} ({\rm d} {\bm x}/{\rm d} \tau) \cdot {\bm u} \epsilon}{\int {\rm d} \sigma \epsilon} \simeq \frac{2 \sqrt{2}}{\pi} \frac{1-8v^6}{1+8v^6}$ and ${\bm u}$ is a unit vector parallel to the curvature radius vector, and $H$ is the Hubble parameter $H = \frac{{\rm d} a/{\rm d} t}{a}$. The second term of the right hand of \eqref{eq:Leq} is the energy transmitted to loops per unit time, $p$ is a probability of reconnection and $c$ is the loop chopping efficiency parameter which is set $c \simeq 0.23$ \cite{Martins:2000cs}.
With $\gamma \equiv L/t$, the first equation is rewritten as
\begin{equation}
\frac{{\rm d} \gamma}{{\rm d} t} = \frac{1}{t} \Biggl\{-\gamma +H \gamma t(1+v^2) +\frac{1}{2}cpv \Biggr\} \, . \label{eq:gammaeq} \\
\end{equation}
By setting ${\rm d} \gamma/{\rm d} t$ and ${\rm d} v/{\rm d} t$ to be zero in
\eqref{eq:veq} and
\eqref{eq:gammaeq}, we obtain the asymptotic solutions
\begin{equation}
\gamma = {\rm Const.}, \, v = {\rm Const.}\,
\end{equation}
As we find the correlation length $L = \gamma t$ grows in proportion to $t$, the number of infinite strings is conserved in the horizon. The velocity keeps constant value for a fixed cosmic expansion rate.
\subsection{Distribution function of kinks on infinite strings}
Kinks are defined as discontinuities in the string tangent vector ${\bm x}$. They are produced by reconnection between cosmic strings and propagate along strings.
The sharpness of the kink is defined by
\begin{equation}
\psi \equiv \frac{1}{2} (1- {\bm p}_{\pm, \, 1} \cdot {\bm p}_{\pm, \, 2}) \, . \label{eq:psi}
\end{equation}
The subscript $\pm$ denotes the left and right moving modes, and $1$/$2$ represent the left/right side of the discontinuity, respectively. The range of sharpness is $0 \leq \psi \leq 1$ and
a large value of $\psi$ corresponds to a sharp kink.
Let us define $-\alpha \equiv \langle{\bm p}_+ \cdot {\bm p}_-\rangle=-(1-2v^2)$,
where the bracket means ensemble average in the string network, $v^2$ is the mean square velocity of strings.
Rewriting \eqref{eq:p_pm} and \eqref{eq:psi} in terms of $\psi$, we have \cite{Copeland:2009dk}
\begin{equation}
\psi \propto t^{-2 \zeta} \, ,
\label{eq:blunting}
\end{equation}
where $t$ is the proper time $t=\int a d\tau$ and $\zeta = \alpha \nu$. The parameter $\nu$ characterizes the evolution of the scale factor as $a \propto t^\nu$.
The value of $\zeta$ in the MD era differs from the one in the RD era, as we provide
in table~\ref{tab:const}.
Intersections in the cosmic string network continuously generate kinks on infinite strings. We define
the distribution function of kinks
as a function of the sharpness and proper time, $N(\psi, \, t)$, so that $N(\psi, \, t)d\psi$ is the number of kinks between $\psi$ and $\psi +{\rm d} \psi$ within the volume $V$ at proper time $t$. Then its time evolution is given by \cite{Copeland:2009dk}
\begin{equation}
\frac{\partial N}{\partial t}(\psi, \, t) -\frac{2 \zeta}{t} \frac{\partial}{\partial \psi} (\psi N(\psi, \, t))
=\frac{\bar{\Delta} V}{\gamma^4 t^4} g(\psi) -\frac{\eta}{\gamma t} N(\psi, \, t), \label{eq:dNdtwithoutY}
\end{equation}
where $\bar{\Delta}$ is the probability of the intersection\cite{Austin:1993rg},
$\gamma$ characterizes the correlation length of the string network $L$ as $L = \gamma t$,
and $\eta$ is the decrease rate of kinks due to the loop production which is determined from simulations \cite{Kibble:1990ym}.
The function $g(\psi)$ in \eqref{eq:dNdtwithoutY} is the initial sharpness distribution, and given by
\begin{equation}
g(\psi) = \frac{35}{256} \sqrt{\psi} (15 -6 \psi -\psi^2) \, ,
\end{equation}
where we set $g(\psi) = 0$ for $\psi<0$ or $1<\psi$.
When the left hand side of \eqref{eq:dNdtwithoutY} equals to zero, the equation demonstrates that
the number of kinks is conserved while the sharpness decreases as in \eqref{eq:blunting}.
In the right hand side of \eqref{eq:dNdtwithoutY}, the first term denotes a production of kinks by intersection of strings,
the second term denotes decreasing of the number of kinks by the loop production. This term can be obtained by considering the length of cosmic strings $d$ transferred from infinite strings to loops,
\begin{equation}
\left. \frac{\dot{d}}{d} \right |_{\rm loop} = -\frac{\eta}{\gamma t} \, ,
\end{equation}
and we have assumed that the fraction of kinks taken away on loops is proportional to the loss of length, $\dot{d}/d \propto \dot{N}/{N}$.
\begin{table}[htb]
\begin{minipage}[t]{.45\textwidth}
\centering
\begin{tabular}{|c||c|c|} \hline
& RD & MD \\ \hline \hline
$\gamma$ & 0.31 & 0.50 \\ \hline
$\zeta$ & 0.09 & 0.2 \\ \hline
$\bar{\Delta}$ & 0.20 & 0.21 \\ \hline
$\eta$ & 0.18 & 0.1 \\ \hline
\end{tabular}
\caption{The values of the constant adopted in ref. \cite{Kawasaki:2010yi} are summarized for RD and MD eras.}
\label{tab:const}
\end{minipage}
\hfill
\begin{minipage}[t]{.45\textwidth}
\centering
\begin{tabular}{|c||c|c|} \hline
& RD & MD \\ \hline \hline
$\gamma$ & 0.27 & 0.56 \\ \hline
$\zeta$ & 0.062 & 0.16 \\ \hline
$\bar{\Delta}$ & 0.19 & 0.21 \\ \hline
$\eta$ & 0.076 & 0.068 \\ \hline
\end{tabular}
\caption{The values of the constant for RD and MD eras obtained by solving the VOS equations.}
\label{tab:const_scaling_solve}
\end{minipage}
\end{table}
To obtain the kink distribution using (\ref{eq:dNdtwithoutY}), we need the time evolution of $\gamma, \, \zeta, \, \bar{\Delta},$ and $\eta$. In this paper, we show results by using two different methods to obtain them. In the first case, we use the parameter values used in ref. \cite{Kawasaki:2010yi} and we smoothly change them from RD to MD at radiation-matter equality $t_{\rm eq} \simeq 2.0 \times 10^{12} {\rm s}$ using
\begin{equation}
\chi(t) = \chi_m \frac{1+{\rm tanh}(100{\rm ln}(t/t_{\rm eq}))}{2} + \chi_r \frac{1-{\rm tanh}(100{\rm ln}(t/t_{\rm eq}))}{2} \, , \label{eq:tanh}
\end{equation}
where $\chi_m$ and $\chi_r$ describe values for MD and RD. The values for RD and MD eras are listed in table~\ref{tab:const}. Using the same values with the previous work makes easier to see the effect of the parameter transitions at radiation-matter equality, which was not taken into account in the previous work.
In the second case, we calculate the time evolution of $\gamma, \, \zeta, \, \bar{\Delta},$ and $\eta$ by solving the VOS equations \eqref{eq:veq} and \eqref{eq:gammaeq}.
The parameter values $\zeta, \eta, \bar{\Delta}$ are obtained from $v$ as
\begin{eqnarray}
\zeta & = & \alpha \nu = (1-2v^2) \left (\frac{{\rm ln} \, (a/a_{\rm ini})}{{\rm ln} \, (t/t_{\rm ini})} \right ) \, , \label{eq:zetaeq}\\
\eta & = & \frac{1}{2} cpv \, , \label{eq:etaeq}\\
\bar{\Delta} & = & \frac{2 \pi}{35} \Biggl\{1+\frac{2}{3}(1-2v^2)-\frac{1}{11}(1-2v^2)^2 \Biggr\} \, . \label{eq:Deltaeq}
\end{eqnarray}
Table.~\ref{tab:const_scaling_solve} shows the asymptotic values of the parameters for RD and MD eras obtained by solving the VOS equations. As we can find by comparing the two tables, some of the parameter values are different from the previous work, and they affect the kink distribution as well as the amplitude of the gravitational wave background.
\subsection{Gravitational waves from kinks}\label{sec:GWfromkink}
It has been shown in ref. \cite{Kawasaki:2010yi} that the kinks which contribute the most to the power of gravitational waves with angular frequency $\omega$ satisfy the following condition:
\begin{equation}
\left (\psi \frac{N(\psi, \, t)}{V(t)/(\gamma t)^2} \right )^{-1} \sim \omega^{-1} . \label{eq:kink_N_omega_relation}
\end{equation}
We define the sharpness of kinks which satisfies \eqref{eq:kink_N_omega_relation} for a given frequency $\omega$ as $\psi_{\rm max}(\omega, \, t)$. This condition means that the main contribution on the gravitational wave background at physical frequency $\omega$ comes from kinks with sharpness $\psi_{\rm max}$ whose average\\ interval $(\psi N(\psi, \, t)/(V(t)/(\gamma t)^2))^{-1}$ is comparable with the wavelength of the gravitational waves $\omega^{-1}$.
The strength of a gravitational wave burst from one kink on loops has been formalized in ref. \cite{Damour:2001bk}. Including the dependence on the sharpness $\psi$, the strain amplitude is given by
\begin{equation}
h(f,z)=\frac{G\mu [\psi_{\rm max}(\omega,z)]^{1/2}l}{[(1+z)fl]^{2/3}}\frac{1}{r(z)} \Theta(1-\theta_m), \label{h}
\end{equation}
where $\theta_m = [(1+z)fl]^{-1/3}$, $f = a \omega/(2 \pi a_0)$ is the gravitational wave frequency today with $a_0=1$ being the present scale factor, $r$ is the distance to the source $r(z) = \int^z_{0} dz/H(z)$, and $l$ is twice the fundamental period $T_l=l/2$ of string loops. Since we consider infinite strings and their typical curvature is given by $\gamma t$, $l$ is replaced by the correlation length $\gamma t$ in our calculation. The step function $\Theta (1 -\theta_m)$ is introduced to set a low-frequency cutoff, which reflects the fact that kinks do not emit gravitational waves larger than the horizon size. We calculate the Hubble parameter using $H=H_0[\Omega_r (a/a_0)^{-4}+\Omega_m (a/a_0)^{-3}+\Omega_\Lambda]^{1/2}$, where $\Omega_r$, $\Omega_m$ and $\Omega_\Lambda$ are the density parameters for radiation, matter, and the cosmological constant, respectively. We use $\Omega_r h^2=4.31\times 10^{-5}$ where $h$ is the reduced Hubble constant. In this paper, we assume a flat universe and use the values obtained from Planck satellite \cite{Ade:2015xua}: $h=0.692$, $\Omega_m=0.308$ and $\Omega_\Lambda=0.692$.
The power of the gravitational wave background is usually characterized by $\Omega_{\rm gw}\equiv ({\rm d}\rho_{\rm gw}/{\rm d}{\rm ln}f)/\rho_c $, where $\rho_{\rm gw}$ is the energy density of gravitational waves and $\rho_c$ is the critical density of the universe. The gravitational wave spectrum generated from kinks on infinite strings is given by
\begin{equation}
\Omega_{\rm gw}(f) = \frac{2\pi^2f^2}{3H_0^2}
\int \frac{dz}{z}\Theta(n(f,z)-1) n(f,z)h^2(f,z) \, ,
\label{eq:Omega_gw}
\end{equation}
where
\begin{equation}
n(f,z)=\frac{1}{f}\frac{d\dot{N}}{d\ln z} \label{n} = \frac{1}{f}\cdot\frac{1}{2}\theta_m(f,z)\frac{z}{1+z}
\frac{\psi_{\rm max}(\omega,z)N(\psi_{\rm max}(\omega,z),z)}{V}
l^{-1}\frac{dV}{dz} \, ,
\end{equation}
and $dV/dz=4\pi a^3r^2(z)/H(z)$ is the volume between the redshift $z$ and $z + dz$. The step function $\Theta(n(f,z)-1)$ is introduced to exclude rare bursts, whose intervals are longer than $\sim 1/f$ and cannot form a continuous background of gravitational waves. Note the difference in the notation: $\psi\tilde{N}$ (number of kinks with sharpness $\ln\psi \sim \ln\psi + d\ln\psi$ per volume) in ref. \cite{Kawasaki:2010yi} is identical to $\psi N/V$ in our paper. In summary, the differences with respect to ref. \cite{Kawasaki:2010yi} are
\begin{itemize}
\item We replace the typical curvature of infinite string as $l \sim \gamma t$ instead of $l\sim t$.
\item The probability of observing the gravitational wave burst from a kink is $\theta_m/2$ \cite{Olmez:2010bi} instead of $\theta_m/4$.
\item The distance $r$ and the volume $dV/dz$ are calculated numerically instead of using approximated analytic expressions.
\item $\Omega_\Lambda$ is included in the calculation of the Hubble parameter.
\end{itemize}
These changes increase the overall spectral amplitude by $9.6$ in RD era and $2.7$ in MD era compared to the one calculated in ref. \cite{Kawasaki:2010yi}.
\section{Results}\label{sec:III}
\subsection{Result with the tanh interpolation}\label{sec:III_1}
We first solve the differential equation \eqref{eq:dNdtwithoutY} using the tanh interpolation \eqref{eq:tanh} with the values in table \ref{tab:const}.
The result is shown in figure~\ref{fig:Npsit0}.
\begin{figure}[htbp]
\centering
\includegraphics[width=12cm,clip]{Npsit0_Sep_1_tanh_Miya_Sep_5_old_up.eps}
\caption{The distribution function of kinks obtained using the tanh interpolation.
The vertical axis is the number of kinks on infinite strings per length. The horizontal axis is the sharpness of kinks. The light-blue broken line is the analytic estimation in the previous work \cite{Kawasaki:2010yi} and the red solid line is our numerical result.}
\label{fig:Npsit0}
\end{figure}
As mentioned in the previous section, the sharpness of kinks decreases with time. The number of old kinks with small sharpness is larger than new ones, because ${\cal O}(1-10)$ of kinks are produced per horizon and the number of newly produced kinks per comoving length decreases as the horizon grows. In figure~\ref{fig:Npsit0}, we find that the distribution function of kinks has two regions with different slopes. The left part ($\psi \lesssim 10^{-2}$) corresponds to kinks generated during the RD era, and the right part ($\psi \gtrsim 10^{-2}$) corresponds to kinks generated in the MD era. Note that our result has a step at radiation-matter equality ($\psi \sim 7.4 \times 10^{-3}$), which is not seen in the result of the previous work. The reason will be discussed in the next section.
Figure \ref{fig:Omega_gw} is the numerical results for the power spectrum of the gravitational wave background $\Omega_{\rm gw}$.
\begin{figure}[htbp]
\centering
\includegraphics[width=12cm,clip]{Omega_gw_50_Sep_1_tanh_Sep_5_old.eps}
\caption{Power spectrum of the gravitational wave background for $G \mu = 10^{-7}$. The red solid line is the result obtained excluding rare bursts and the green broken line is the result including rare bursts.}
\label{fig:Omega_gw}
\end{figure}
To calculate the gravitational wave background, we first look for the value which satisfies \eqref{eq:kink_N_omega_relation} each time in the calculation of $N(\psi,t)$ for each gravitational wave frequency $\omega$, and define it as $\psi_{\rm max}(\omega , t)$. Then, using the values of $\psi_{\rm max}$, we numerically integrate \eqref{eq:Omega_gw} to obtain the power spectrum. Note that the vertical axis of figure~\ref{fig:Npsit0} is identical to the inverse of the left hand side of \eqref{eq:kink_N_omega_relation}. So gravitational wave frequency $\omega$ is corresponded to the vertical axis of figure~\ref{fig:Npsit0}.
As seen in figure~\ref{fig:Npsit0}, old kinks are numerous and the gravitational wave emission has a short interval, while new kinks are few and the interval is large. Thus, the high frequency gravitational waves are emitted from old kinks and low frequency gravitational waves are from new kinks. The gravitational waves in the range of $10^{-13} \, {\rm Hz} \leq f$ are generated from kinks with small sharpness produced in the RD era. The middle frequency $10^{-15} \, {\rm Hz} \leq f \leq 10^{-13} \, {\rm Hz}$ corresponds to kinks produced during the transition from the RD era to the MD era. The low frequency gravitational waves $f \leq 10^{-15} \, {\rm Hz}$ are emitted from kinks produced in the MD era.
\subsection{Result with the VOS model} \label{sec:III_2}
In this section, we solve the differential equation \eqref{eq:dNdtwithoutY} by simultaneously solving the VOS equations \eqref{eq:veq} and \eqref{eq:gammaeq}. The VOS model provides time evolution of $\gamma$ and $v$, which can be converted to $\zeta, \, \eta$ and $\bar{\Delta}$ by \eqref{eq:zetaeq}, \eqref{eq:etaeq} and \eqref{eq:Deltaeq}. The time evolution of the parameters is shown in figure \ref{fig:const}.
\begin{figure}[htbp]
\centering
\begin{tabular}{c}
\begin{minipage}{0.5\hsize}
\centering
\includegraphics[width=7cm,clip]{zeta_Sep_5_old.eps}
\end{minipage}
\begin{minipage}{0.5\hsize}
\centering
\includegraphics[width=7cm,clip]{gamma_eta_delta_Sep_5_old.eps}
\end{minipage}
\end{tabular}
\caption{The time evolution of the parameters which characterize the production of kinks and their evolution. In the left figure, the time evolution of $\zeta$ is shown. The red solid line is obtained by solving VOS equations and the green broken line is the one interpolated by the tanh function. In the right figure, we show time evolution of $\gamma, \bar{\Delta}$ and $\eta$ obtained by solving the VOS equations. The red solid line is $\gamma$, the green broken line is $\bar{\Delta}$ and the blue broken line is $\eta$.}
\label{fig:const}
\end{figure}
We find their evolution is very different from the tanh interpolation. First, the VOS equations with $c=0.23$ provide different asymptotic values of the parameters as seen by comparing tables \ref{tab:const} and \ref{tab:const_scaling_solve}. Second, the transition from the RD era to the MD era is not instant and it takes time to approach the asymptotic value. In addition, the parameter values change near the present time, since we include cosmological constant.
The distribution of kinks obtained by the VOS model is shown in figure \ref{fig:Npsit_scaling}.
\begin{figure}[htbp]
\centering
\includegraphics[width=12cm,clip]{Npsit0_Aug_15_scaling_Sep_5_old.eps}
\caption{The distribution function of kinks obtained by solving the VOS equations (the red solid line). The axises is same as figure \ref{fig:Npsit0}. For comparison, we also show the analytic estimation by the previous work \cite{Kawasaki:2010yi} (light-blue broken line) and our result of the tanh interpolation (magenta broken line).}
\label{fig:Npsit_scaling}
\end{figure}
From the figure, we find two differences between the results with the tanh interpolation and the VOS model. First, the number of kinks increases considerably because the slope of the distribution function becomes steeper both for the RD and MD eras. Second, the position corresponding to radiation-matter equality has moved toward large $\psi$. The reason is discussed in the next section.
Figure \ref{fig:Omega_gw_scaling} shows the power spectrum of the gravitational wave background $\Omega_{\rm gw}$ calculated using the kink distribution obtained by the VOS model.
\begin{figure}[htbp]
\centering
\includegraphics[width=12cm,clip]{Omega_gw_50_Aug_15_scaling_Sep_5_old.eps}
\caption{Power spectrum of the gravitational wave background for $G \mu = 10^{-7}$. The axises are same as figure \ref{fig:Omega_gw}.}
\label{fig:Omega_gw_scaling}
\end{figure}
We see that the amplitude is larger than the case of the tanh interpolation, because of the increase in the number of kinks. Since the number increases more at small $\psi$, which corresponds kinks generated during the RD era, the power of the gravitational wave spectrum is enhanced in high frequencies.
Figure \ref{fig:Omega_gw_scaling_iroiro} is the comparison between sensitivity curves of future gravitational wave observations and the power spectra of the gravitational wave background for different values of string tension. The SKA \cite{Janssen:2014dka} is a radio interferometer, which can detect gravitational waves by pulsar timing arrays. The eLISA \cite{AmaroSeoane:2012km,AmaroSeoane:2012je} and DECIGO \cite{Seto:2001qf,Kawamura:2011zz} missions will observe gravitational waves using laser interferometers at space. Advanced-LIGO \cite{Harry:2010zz} is a laser interferometer constructed on the ground and will construct observation network with other ground-based detectors such as Advanced-VIRGO \cite{Accadia:2011zzc} and KAGRA in near future \cite{Somiya:2011np}.
\begin{figure}[htbp]
\centering
\includegraphics[width=15cm,clip]{Omega_gw_scaling_Sep_5_old.eps}
\caption{The power spectrum of the gravitational wave background for different string tensions. The spectrum shown here do not include rare bursts. The black solid and broken lines are sensitivity curves of gravitational wave experiments.}
\label{fig:Omega_gw_scaling_iroiro}
\end{figure}
\section{Discussion}\label{sec:IV}
First, let us compare our numerical result of the tanh interpolation with the previous work \cite{Kawasaki:2010yi}. The major difference is that our result has a step-like feature in the distribution function of kinks at radiation-matter equality as seen in figure~\ref{fig:Npsit0}. This step arises because of the changes in the value of $\gamma$ in the evolution equation of the distribution function.
The source term in \eqref{eq:dNdtwithoutY} (the first term in the right hand side) has a factor of $\bar\Delta/\gamma^4$, and it becomes smaller in the MD era. Thus, the number of newly produced kinks is smaller in the MD era.
The gravitational wave spectrum reflects the existence of this transition phase, and has three regions of different spectral slopes: the MD era $f \leq 10^{-15} \, {\rm Hz}$, the transition phase $10^{-15} \, {\rm Hz} \leq f \leq 10^{-13} \, {\rm Hz}$, and the RD era $10^{-13} \, {\rm Hz}\leq f$. Let us analytically estimate the frequency dependence of $\Omega_{\rm gw}$.
The distribution function of kinks is related to the frequency of the gravitational wave background by \eqref{eq:kink_N_omega_relation}. The low-frequency gravitational waves $f \leq 10^{-15} \, {\rm Hz}$ are generated by kinks produced in the MD era, and we can read $\psi N/(V/(\gamma t)^2) \sim \psi^{-2}$ from figure~\ref{fig:Npsit0}. Then we can derive $\Omega_{\rm gw} \propto f^{-1/6}$ using \eqref{eq:kink_N_omega_relation} and \eqref{h}. This frequency dependence of the spectrum coincides with the analytic result in the previous work \cite{Kawasaki:2010yi} and is also consistent with the numerical result shown in figure~\ref{fig:Omega_gw}. The high-frequency region $10^{-13} \, {\rm Hz}\leq f$ corresponds to the gravitational wave from kinks produced in the RD era, where the result in figure~\ref{fig:Npsit0} gives $ \psi N/(V/(\gamma t)^2) \sim \psi^{-5.1}$ and we get $\Omega_{\rm gw} \propto f^{7/51}$. This also coincides with the analytic result of ref. \cite{Kawasaki:2010yi}, and is consistent with the spectrum with rare bursts in figure~\ref{fig:Omega_gw}. The difference from the previous work arises in $10^{-15} \, {\rm Hz} \leq f \leq 10^{-13} \, {\rm Hz}$. In the transition phase, the value of $\psi_{\rm max}$ is the same for all the given frequency, so $\psi_{\rm max}\simeq {\rm Const.}$, and we get $\Omega_{\rm gw} \propto f^{1/3}$ from \eqref{eq:kink_N_omega_relation} and \eqref{h}. In fact, this frequency dependence can be seen in the numerical result of figure~\ref{fig:Omega_gw}.
Let us compare the spectral amplitude with the previous work \cite{Kawasaki:2010yi}. We compare the results using the case with rare bursts, since the condition of excluding rare bursts depends on the distribution function of kinks and the comparison cannot be made simply. First of all, the overall amplitude is $9.6$ and $2.7$ times larger than the previous work in RD and MD eras respectively, because of the modifications listed in section \ref{sec:GWfromkink}. In addition, the amplitude of the high frequency region increases because of the larger number of kinks produced during the RD era. By extracting the dependence on $\psi_{\rm max}$ and $\gamma$ from \eqref{eq:Omega_gw}, we obtain
\begin{eqnarray}
\Omega_{\rm gw}(f) & \propto & f^2 \int \frac{{\rm d}z}{z} \psi_{\rm max}(\omega,z) \left(\psi_{\rm max}(\omega,z) \frac{N(\psi_{\rm max}(\omega,z),z)}{V/(\gamma t)^2} \right ) \gamma^{-8/3} \nonumber \\
& \propto & f^2 \int \frac{{\rm d}z}{z} \psi_{\rm max}(\omega,z) \gamma^{-8/3} .
\label{eq:Omega_gw_kantan}
\end{eqnarray}
As seen in figure \ref{fig:Npsit0}, the number of kinks produced during RD era is larger than the previous work and the value of $\psi_{\rm max}$ becomes larger about twice for a fixed $\omega$. Therefore, for every frequency of the gravitational wave background corresponding to the RD era, the amplitude $\propto \psi_{\rm max}$ becomes twice larger than in the previous work. In total, the spectral amplitude is a few times larger in the low frequency and ${\mathcal O}(10)$ larger in the high frequency.
\vspace{5pt}
Next, let us discuss the case where we solve the parameter evolution with the VOS model.
First, we explain why the distribution of kinks has different shape compared to the case of the tanh interpolation. In ref. \cite{Kawasaki:2010yi}, the solution of the distribution function is provided as
\begin{equation}
\psi \frac{N}{V/(\gamma t)^2} \propto \psi^{\frac{-3+3\nu+\eta/\gamma}{2 \zeta}} .
\label{eq:kink_bunpu_analysis}
\end{equation}
As seen in tables \ref{tab:const} and \ref{tab:const_scaling_solve}, the parameter values from the VOS equations are different from the ones used in the tanh case. They largely affect the $\psi$ dependence of the distribution function as seen in \eqref{eq:kink_bunpu_analysis}.
In particular, the small value of $\zeta$ increases the power of $\psi$, and makes the slope steeper.
This increases the kink number considerably at small $\psi$. In addition, the number of kink production is determined by the coeficient of $\bar{\Delta}/\gamma^4$ in the first term of the right hand side of \eqref{eq:dNdtwithoutY}. The difference in this factor also increases the overall amplitude slightly.
In figure \ref{fig:Npsit_scaling}, we also see the position of radiation-matter equality shifts to larger $\psi$.
As provided in ref. \cite{Kawasaki:2010yi}, the value of $\psi$ corresponding to radiation-matter equality $\psi_*$ is given by
\begin{equation}
\psi_*(t_0) = \left (\frac{t_{\rm eq}}{t_0} \right )^{2\zeta_m},
\label{eq:rm-eq_position}
\end{equation}
where the suffix $m$ is the value during MD era. Since the value of $\zeta$ is smaller than the one used in the tanh case, $\psi_*$ becomes larger in the VOS case.
Then, let us describe the reason of the large increase of the spectral amplitude in figure \ref{fig:Omega_gw_scaling}. The reason is the same as described in the tanh case, that is the increase of $\psi_{\rm max}$. For example, $\psi_{\rm max}$ increases 100 times at $10^2$[Hz].
Then, from \eqref{eq:Omega_gw_kantan}, $\Omega_{\rm gw}$ increases ${\mathcal O}(10^2)$.
Taking account the enhancement of the overall amplitude $9.6$ for RD era, we find that the power spectrum has increased ${\mathcal O}(10^3)$ at high frequencies compared to the previous work.
From figure~\ref{fig:Omega_gw_scaling_iroiro}, we find that the gravitational wave background from kinks on infinite strings is testable by future experiments depending on the tension of strings. The SKA would probe $G \mu \gtrsim 10^{-9}$, eLISA and Advanced-LIGO can test $G \mu \gtrsim 10^{-7}$, and DECIGO has the strongest sensitivity to reach $G \mu \gtrsim 10^{-10}$. Note that the spectrum shown in figure~\ref{fig:Omega_gw_scaling_iroiro} do not include rare bursts. Since rare bursts with large amplitude exists at high frequencies as seen in figure~\ref{fig:Omega_gw}, laser-interferometer experiments could be also used to search for rare burst signals.
\section{Conclusions}\label{sec:V}
In this work, we have calculated the power spectrum of the gravitational wave background from kinks on infinite strings. First, we have solved the differential equation \eqref{eq:dNdtwithoutY} to obtain the distribution function of kinks numerically in two ways. First, unlike the analytic estimation of
ref. \cite{Kawasaki:2010yi}, we have smoothly connected the parameters using a tangent hyperboric function, which are related to the evolution of the cosmic string network, at radiation-matter equality. As a result, we have found a step in the distribution function of kinks at the transition from the RD era to the MD era, which was overlooked in the previous work \cite{Kawasaki:2010yi}.
At the same time, we have found an increase in the number of kinks generated in the RD era. Second, we have calculated the distribution of kinks by following the time evolution of the parameters with the VOS equations. We have found a steeper slope of the distribution function, which gives a large increase of the kink number at small sharpness, and the shift of the position of radiation-matter equality.
Next, using the numerical result of the distribution function of kinks, we have calculated the power spectrum of the gravitational wave background.
In the case where we use the tanh interpolation, due to the step in the distribution function of kinks, we have found that the power spectrum behaves as $\Omega_{\rm gw} \propto f^{1/3}$ at $10^{-15} \, {\rm Hz} \lesssim f \lesssim 10^{-13} \, {\rm Hz}$. The power spectrum has increased more in the case where we solve the the VOS equations. In addition to the precise estimation of the kink distribution, we have also carefully evaluated all the factors involved in the calculation of the spectrum. This allows us to offer a rather precise prediction on the spectral amplitude.
By comparing the results with sensitivities of future experiments,
we have shown that gravitational waves from kinks on infinite strings can be probed at different frequencies.
Finally, let us comment on the gravitational wave background from cusps and kinks on cosmic string loops. Loops emit gravitational waves whose wavelength is shorter than the loop size, and usually the number of loops is more than that of infinite strings. Thus, it is more likely that the gravitational wave background generated by infinite strings is sub-dominant at high frequencies compared to the one from loops. However, since the typical loop size is not known yet, gravitational waves from kinks on infinite strings could be more important than that from loops, especially for the SKA which probes low frequencies. It is important to consider both origins of gravitational wave background to provide constraints on cosmic strings, and thus our careful estimation of the gravitational wave background from kinks on infinite strings would help to constrain cosmic strings by observation at low frequencies.
Moreover, there are cosmic superstrings predicted in superstring theory, and they also form the network consisting of loops and infinite strings. They have cusps and kinks which emit gravitational waves.
Recently, the power spectrum of the gravitational wave background produced by loops of cosmic superstrings has been investigated \cite{Sousa:2016ggw}, but the one from kinks on infinite superstrings is not clear yet.
It is being examined in a work in progress.
If a new era of multi-wavelength gravitational wave observations is successful and a detection was made, we might even be able to get insight in the physics of the very early universe.
\acknowledgments
KH is supported by a Grant-in-Aid for JSPS Research under Grant No.15J05029,
DN is supported by MEXT Grant-in-Aid for Scientific Research on Innovative Areas,~No.15H05890,
SK is supported by Career Development Project for Researchers of Allied Universities.
|
\section{introduction}
The curved $N$-body problem is a natural extension of the Newtonian $N$-body problem in $\R^3$ to isotropic, complete, simply connected
spaces of constant nonzero curvature, $\S^3$ and $\H^3$. For its history, we refer the readers to \cite{Dia12a}, where the equations of motion are written in extrinsic coordinates in $\R^4$ for $\S^3$, and the Minkowski space $\R^{3,1}$ for $\H^3$. This approach, different form more traditional ones like \cite{Kil98}, led to fruitful results, especially in the study of relative equilibria, which are rigid motions that become fixed points in some rotating coordinates system, \cite{DPS05,Dia12a,Dia12b,Dia13,Dia-A,DSZ-A}.
Based on the work of Diacu, especially \cite{Dia12a,Dia13}, the authors of \cite{DSZ-B} proposed to study central configurations.
Roughly speaking, central configurations are special arrangements of the point particles and the exact definition will be given later. The central configurations of the Newtonian $N$-body problem, first formulated by Laplace \cite{Lap}, are quite important in the study of the Newtonian $N$-body problem. In \cite{DSZ-B}, the authors have also showed the importance of central configurations for the curved $N$-body problem. For instance, each central configuration gives rise to a one-parameter family of relative equilibria, and central configurations are the bifurcation points in the topological classification of the curved $N$-body problem.
Some questions about these configurations were also raised in \cite{DSZ-B}. For example, find all central configurations for $N$ point particles when $N$ is small (the three-particles case has been recently solved and will appear in a forthcoming paper). Another interesting problem is to prove (or disprove) that for generic $N$ point particles, the number of equivalent classes of central configurations is finite. Though these questions are similar to those of the Newtonian $N$-body problem \cite{Moe-A, Sma98}, the answers are quite different in general. For example, Moulton's theorem concerning the collinear central configurations has been generalized to $\H^3$, \cite{DSZ-B}, but it can not be directly generalized to $\S^3$.
In this paper, we put into the evidence another difference: each central configuration on $\H^3$ is equivalent to one central configuration on $\H^2_{xyw}$, which will be defined later, whereas in $\S^3$ there are central configurations that are not confined to any two-dimensional sphere. In some sense, the number of central configurations in $\H^3$ is smaller than that of $\S^3$. When we consider the Wintner-Smale conjecture in $\H^3$ raised in \cite{DSZ-B} asking whether the number of classes of central configurations in $\H^3$ is finite or not for generic $N$ point particles, we only need to study the problem on $\H^2_{xyw}$.
The paper is organized as follows: in Section \ref{basic}, we recall the basic setting of the curved $N$-body problem and the corresponding facts about central configurations; in Section \ref{h3}, we prove the result about central configurations in $\H^3$; in Section \ref{s3}, we construct a two-parameter family of three-dimensional central configurations in $\S^3$.
\section{the curved $N$-body problem and central configurations}\label{basic}
\subsection{Equations of motion}
As done in \cite{Dia12a,Dia13}, the equations will be written in $\R^4$ for $\S^3$ and in the Minkowski space $\R^{3,1}$, for $\H^3$. For convenience, we understand the two linear spaces as $\R^4$ endowed with two inner products: for two vectors, $\q_1=(x_1,y_1,z_1,w_1)^T$ and $\q_2=(x_2,y_2,z_2,w_2)^T$, the inner products are given by
\[ \q_1\cdot \q_2 = x_1x_2+y_1y_2+z_1z_2 +\sigma w_1w_2, \]
where $\sigma=1$ for the Euclidean space and $\sigma=-1$ for the Minkowski space. We define the unit sphere $\S^3$ and the unit hyperbolic sphere $\H^3$ as
\begin{equation*}
\begin{split}
\S^3&:=
\{(x,y,z,w)^T\in\mathbb R^4\ \! |\ \!
x^2+y^2+z^2+w^2=1\}\,\ \ {\rm and }\\
\H^3&:=
\{(x,y,z,w)^T\in\mathbb R^{4}\ \! |\ \!
x^2+y^2+z^2-w^2=-1, \ w>0\},
\end{split}
\end{equation*}
respectively. We can merge these two manifolds into
\[ \M^3:=
\{(x,y,z,w)^T\in\mathbb R^{4}\ \! |\ \!
x^2+y^2+z^2+\sigma w^2=\sigma, \ {\rm with}\ w>0\ {\rm for}\ \sigma=-1\}. \]
Given the positive masses $m_1,\dots, m_N$, whose positions are described by the
configuration $\q=(\q_1,\dots,\q_N)\in(\M^3)^N$,
$\q_i=(x_i,y_i,z_i,w_i)^T,\ i=\overline{1,N}$, we define the
singularity set
\begin{equation}
\D=\cup_{1\le i<j\le N}\{\q\in (\M^3)^N\ \! ; \ \! \q_i\cdot\q_j=\pm 1\}.
\notag\end{equation}
Let $d_{ij}$ be the geodesic distance between the point masses $m_i$ and $m_j$, which is computed by
\[ d_{ij}(\q)= \arccos (\q_i \cdot \q_j) \ {\rm for}\ \S^3, \ \ d_{ij}(\q)= {\rm arccosh} (-\q_i \cdot \q_j) \ {\rm for}\ \H^3. \]
The force function $U$ ($-U$ being the potential function) in $(\M^3)^N\setminus\D$ is
$$
U(\q):=\sum_{1\le i<j\le N}m_im_j\ctn d_{ij}(\q),
$$
where $\ctn (x)$ stands for $\cot (x)$ in $\S^3$ and $\coth (x)$ in $\H^3$. We also introduce two more notations, which unify the trigonometric and hyperbolic functions,
\[ \sn (x)= \sin(x) \ {\rm or}\ \sinh(x), \
\ \csn (x)= \cos(x) \ {\rm or}\ \cosh(x). \]
Define the kinetic energy as $T(\dot \q)= \sum_{1\le i\le N}m_i \dot \q_i \cdot \dot \q_i$, $\dot \q=(\dot \q_1,...,\dot\q_N)$. Then the curved $N$-body problem is given by the Lagrange system on $T((\M^3)^N\setminus\D)$, with
\[ \ L(\q,\dot\q):=T(\dot\q)+U(\q). \]
Using variational methods, we obtain the equations of motion in $\S^3$ and in $\H^3$, \cite{DSZ-B}. Merged into one, they are
\begin{equation*}
\begin{cases}
\ddot\q_i=\sum_{j=1,j\ne i}^N\frac{m_im_j [\q_j-\csn d_{ij}\q_i]}{\sn^3 d_{ij}}-\sigma m_i(\dot{\q}_i\cdot \dot{\q}_i)\q_i\cr
\q_i\cdot \q_i=\sigma, \ \ \ \ i=\overline{1,N}.
\end{cases}
\end{equation*}
The first part of the acceleration access from the gradient of the force function, $\nabla_{\q_i}U(\q)$, and we will denote it by $\F_i$. It is the sum of $\F_{ij}=\frac{m_im_j [\q_j-\csn d_{ij}\q_i]}{\sn^3 d_{ij}}$ for $j\ne i$.
\subsection{Central configurations }
\begin{definition}
A configuration $\q\in (\M^3)^N\setminus \D$ is called a \emph{central configuration} if there is some constant $\lambda$ such that
\begin{equation}\label{CCE}
\nabla_{\q_i}U(\q)=\lambda\nabla_{\q_i} I(\q),\ i=\overline{1,N},
\end{equation}
where $\nabla$ is the gradient operator in $\M^3$, $I(\q)= \sum _{i=1}^N m_i (x_i^2+y_i^2)$, and the explicit form of $\nabla_{\q_i} I$ is
\begin{equation}\label{I}
2m_i
\begin{bmatrix}
x_i(w_i^2+ z_i^2)\\
y_i(w_i^2+ z_i^2)\\
- z_i(x_i^2+ y_i^2)\\
- w_i(x_i^2+ y_i^2)
\end{bmatrix}{\rm in}\ \ \! T(\S^3)^N\ \ {\rm and}\ \ 2m_i
\begin{bmatrix}
x_i(w_i^2- z_i^2)\\
y_i(w_i^2- z_i^2)\\
z_i(x_i^2+ y_i^2)\\
w_i(x_i^2+ y_i^2)
\end{bmatrix} {\rm in}\ \ \! T(\H^3)^N.
\end{equation}
\end{definition}
Since the two functions $U$ and $I$ are both invariant under the group action of $SO(2)\times SO(2)$ (in the case of $\S^3$) and $SO(2)\times SO(1,1)$ (in the case of $\H^3$), it is easy to check that a central configuration remains a central configuration after an $SO(2)\times SO(2)$ action (or an $SO(2)\times SO(1,1)$ action), \cite{DSZ-B}. Two central configurations are said to be \emph{equivalent} if one can be transformed to the other by these group actions. When we say a central configuration, we mean a class of central configurations as defined by the equivalence relation.
A central configuration with $\lambda=0$ is called a \emph{special central configuration}, which only occurs in $\S^3$, \cite{Dia12a}. Otherwise, it is called an \emph{ordinary central configuration}. A central configuration lying on a geodesic is called a \emph{geodesic central configuration}. A central configuration lying on a two-dimensional sphere is called an \emph{$\S^2$ central configuration}, a central configuration lying on a two-dimensional hyperbolic sphere is called an \emph{$\H^2$ central configuration}. All the other central configurations are called \emph{three-dimensional central configurations}.
Here, a two-dimensional sphere (hyperbolic sphere) means a sphere (hyperbolic sphere) isometric to the unit sphere (hyperbolic sphere) in $\R^3$ ($\R^{2,1}$). It is the non-empty intersection of $\M^3$ with a three-dimensional linear subspace: $\{ (x,y,z,w)^T\in \R^4| ax+by+cz+dw=0 \}$, \cite{BH99}.
We begin with the following result.
\begin{proposition}\label{gradient_of_I} Let $V=\{ (x,y,z,w)^T\in \R^4| cz+\sigma dw=0 \}$.
If a configuration $\q=(\q_1, \cdots, \q_N)$ lies on the two-dimensional sphere (hyperbolic sphere): $ V \cap \M^3$, then $\nabla_{\q_i} I$ is in $V$ for $i=\overline{1,N}$.
\end{proposition}
\begin{proof}
By the explicit form of $\nabla_{\q_i} I$, equation \eqref{I}, we get
\[ \nabla_{\q_i} I \cdot (0,0,c,d)^T= -\sigma 2m_i (x_i^2+y_i^2)(cz_i+\sigma dw_i)=0. \]
This equation completes the proof.
\end{proof}
\section{Central configurations on $\H^3$}\label{h3}
Let us define $\H_{xyw}^2:=\{(x,y,z,w)^T\in \R^4| z=0\} \cap \H^3$. We can prove the following result.
\begin{theorem}\label{theorem_on_h3}
Each central configuration in $\H^3$ is equivalent to some central configuration on $\H^2_{xyw}$.
\end{theorem}
\begin{proof}
We first show that all central configurations in $\H^3$ must lie on a two-dimensional hyperbolic sphere. Then we show that there is some action $\chi\in SO(2)\times SO(1,1) $ which transforms that hyperbolic sphere to $\H^2_{xyw}$. Thus by the definition of equivalent central configurations, each central configuration in $\H^3$ is equivalent to some central configuration on $\H^2_{xyw}$.
Consider the two-dimensional hyperbolic sphere: $\H^2_{\phi}:=\{ (x,y,z,w)^T\in \R^4| \cosh \phi z- \sinh \phi w=0 \}\cap \H^3 $. The intersection is not empty, since the linear subspace and $\H^3$ share the point $(0,0, \sinh \phi,\cosh \phi )^T$. We show that each central configuration will be confined to only one such two-dimensional hyperbolic sphere.
Assume that this is not the case. Suppose that there is a central configuration $\q=(\q_1, \cdots, \q_N)$ with $\q_i \in \H^2_{\phi_i}$, $\phi_1\ge \phi_i$ for $i\ne 1$ and there is at least one $i$ such that $\phi_1> \phi_i$. Then $\q_i$ can be written as $(x_i,y_i,\rho_i\sinh \phi_i,\rho_i\cosh \phi_i )^T$ with $\rho_i>0$ since $w_i=\rho_i\cosh \phi_i>0$. By Proposition \ref{gradient_of_I}, $\nabla_{\q_1}I$ is in the linear subspace $\{ (x,y,z,w)^T\in \R^4| \cosh \phi_1 z- \sinh \phi_1 w=0 \}$. In order to have a central configuration, $\nabla_{\q_1}U$ must be in the linear subspace, i.e.,
\[ \nabla_{\q_1} U \cdot (0,0, \cosh \phi_1,\sinh \phi_1)^T=\F_{1z} \cosh \phi_1 - \F_{1w} \sinh\phi_1 =0, \]
where $\F_{1z}$ and $\F_{1w} $ stand for the $z$-coordinate and $w$-coordinate of $\F_1$, respectively. However, using the explicit form of $\F_1$, we get
\begin{equation}
\begin{split}
&\ \ \ \ \F_{1z} \cosh \phi_1 - \F_{1w} \sinh\phi_1\\
&= \sum _{i=2}^{N} m_im_1\left(\frac{z_i-\cosh d_{i1} z_1}{\sinh^3 d_{i1}}\cosh \phi_1 - \frac{w_i-\cosh d_{i1} w_1}{\sinh^3 d_{i1}}\sinh \phi_1\right)\\
&= \sum _{i=2}^{N} m_im_1\frac{\rho_i\sinh \phi_i\cosh \phi_1-\rho_i\cosh \phi_i\sinh \phi_1-\cosh d_{i1} (z_1\cosh \phi_1 - w_1\sinh \phi_1) }{\sinh^3 d_{i1}}\\
&= \sum _{i=2}^{N} m_im_1\frac{ \rho_i\sinh (\phi_i - \phi_1)}{\sinh^3 d_{i1}}<0,
\end{split}
\notag
\end{equation}
since $\phi_i \le \phi_1$ for $i\ne 1$ and there is at least one $i$ such that $\phi_i<\phi_1$.
Thus any central configuration must lie on only one such hyperbolic sphere, say $\H^2_{\phi}$. Let $$\chi=(\begin{bmatrix}
1 & 0 \\
0 &1 \end{bmatrix}, \begin{bmatrix}
\cosh \phi & -\sinh \phi \\
-\sinh\phi &\cosh \phi \end{bmatrix})\in SO(2)\times SO(1,1).$$
Since $\begin{bmatrix}
\cosh \phi & -\sinh \phi \\
-\sinh\phi &\cosh \phi \end{bmatrix} \begin{bmatrix}
\rho_i\sinh \phi \\
\rho_i\cosh \phi \end{bmatrix} =\begin{bmatrix} 0 \\
\rho_i\end{bmatrix},$
$\chi (\H^2_\phi)= \H^2_{xyw}$. This calculation completes the proof.
\end{proof}
To offer more insight into this result, we provide a heuristic argument.
Recall that the Poincar\'{e} ball model of $\H^3$ is
\[ \left( \bar{x}^2+\bar{y}^2+\bar{z}^2<1, \ ds^2=\frac{4(d\bar{x}^2+d\bar{y}^2+d\bar{z}^2)}{(1-\bar{x}^2-\bar{y}^2-\bar{z}^2)^2} \right). \]
In this model, a two-dimensional hyperbolic sphere is the intersection of the three-dimensional ball with a two-dimensional Euclidean sphere that orthogonally intersects the boundary of the ball.
The hyperbolic spheres $\H^2_\phi$ defined above are those that intersect the $\bar{z}$-axis orthogonally, \cite{BH99}.
For example, $\H^2_{xyw}$ in this model is the disk in the plane $\bar{z}=0$. Now suppose that $\q_i \in \H^2_{\phi_i}$ and $\phi_1> \phi_2$. Proposition \ref{gradient_of_I} implies that $\nabla_{\q_1}I \in T_{\q_1} \H^2_{\phi_1}$, as showed in Figure \ref{fig:ball}. However, $\F_{12}$ points towards the lower hyperbolic sphere $\H^2_{\phi_2}$. Thus $\nabla_{\q_1}I$ and $\nabla_{\q_1}U$ cannot be collinear.
\begin{figure}[!h]
\centering
\begin{tikzpicture}
\draw (0, 1.3) circle (2);
\draw [dashed] (0, -0.7)--(0, 3.3);
\draw [dashed] (-2, 1.3)--(2, 1.3);
\draw [->] (0, 3.3)--(0, 3.5) node (zaxis) [left] {$\bar{z}$};
\draw [->] (2, 1.3)--(2.3, 1.3) node (zaxis) [below right] {$\bar{x}$};
\draw [-] (-2.3, 1.3)--(-2, 1.3);
\draw [-] (0,-1)--(0,-0.7);
\draw [dashed][domain=-2:2] plot (\x, {1.3+sqrt(4-\x*\x)/4});
\draw [domain=-1.95:1.95] plot (\x, {1.3-sqrt(4-\x*\x)/4});
\draw[dashed] (-1.61,2.5) to [out=325,in=215] (1.61,2.5);
\draw [dashed] (0,2.5) ellipse (1.55 and .25);
\fill (-0.7,2) circle (2.5pt) node[left] {$m_1$};
\fill (1,1.3) circle (2.5pt) node[right] {$m_2$};
\draw [->] (-0.7,2)--(-0.3, 1.5) node [left] {$\F_{12}$};
\draw [->] (-0.7,2)--(-0.1, 1.85) node [above right] {$\nabla_{\q_1}I $};
\end{tikzpicture}
\caption{ A configurations in the Poincar\'e ball model.} \label{fig:ball}
\end{figure}
\begin{remark}
Recall that there are central configurations not in a plane (called spatial central configurations) in the Newtonian $N$-body problem, such as the regular tetrahedron for any given four masses.
However, those spatial configurations do not lead to rigid motions. Thus if we defined central configurations in $\R^3$ as those that lead to rigid motions, there would be no spatial ones.
\end{remark}
\section{Central configurations in $\S^3$}\label{s3}
Apparently, the compactness of $\S^3$ makes the set of central configuration in it richer than in $\H^3$.
With computations similar to the ones we performed in $\H^3$, we can get the following necessary conditions for central configurations in $\S^3$,
\[\sum _{j=1, j\ne i}^{N} m_im_j\frac{ \rho_j\sin (\phi_j - \phi_i)}{\sin^3 d_{ij}}=0,\ i=\overline{1,N}.\]
These equations, however, do not rule out the existence of three-dimensional central configurations. For example, we have the pentatope special central configuration of five equal masses, \cite{DSZ-B},
\begin{align*}
x_1&=1,& y_1&=0,& z_1&=0,& w_1&=0, \displaybreak[0]\\
x_2&=-1/4,& y_2&=\sqrt{15}/4,& z_2&=0,& w_2&=0, \displaybreak[0]\\
x_3&=-1/4,& y_3&=-\sqrt{5}/(4\sqrt{3}),& z_3&=\sqrt{5}/\sqrt{6},& w_3&=0,\displaybreak[0]\\
x_4&=-1/4,& y_4&=-\sqrt{5}/(4\sqrt{3}),& z_4&=-\sqrt{5}/(2\sqrt{6}),& w_4&=\sqrt{5}/(2\sqrt{2}), \displaybreak[0]\\
x_5&=-1/4,& y_5&=-\sqrt{5}/(4\sqrt{3}),& z_5&=-\sqrt{5}/(2\sqrt{6}),& w_5&=-\sqrt{5}/(2\sqrt{2}).
\end{align*}
However, all known three-dimensional central configurations are special central configurations (i.e., $\lambda=0$). We will further construct a two-parameter family of ordinary three-dimensional central configurations of five masses. Suppose that the masses are $m_1=m_2=m, m_3=m_4=m_5=1$, and their positions are given by
\begin{align*}
x_1&=0,& y_1&=0,& z_1&=\cos \theta,& w_1&=\sin \theta, \displaybreak[0]\\
x_2&=0,& y_2&=0,& z_2&=\cos \theta,& w_2&=-\sin \theta, \displaybreak[0]\\
x_3&=r,& y_3&=0, & z_3&=c,& w_3&=0,\displaybreak[0]\\
x_4&=r\cos \frac{2\pi}{3},& y_4&=r\sin \frac{2\pi}{3}, & z_4&=c,& w_4&=0,\displaybreak[0]\\
x_5&=r\cos \frac{4\pi}{3},& y_5&=r\sin \frac{4\pi}{3}, & z_5&=c,& w_5&=0,
\end{align*}
where $c\in (-1,1)\setminus\{0\}$, $r>0$, $r^2+c^2=1$ and $\th\in (0, \pi)\setminus \{\frac{\pi}{2}\}$. Such configurations depend on two parameters, $c$ and $\th$, and we denote them by $\q(c, \th)$. It is easy to see that these configurations are not confined to any two-dimensional sphere. In Figure \ref{fig:s3}, we illustrate such a configuration in a $\R^3$ hyperplane by the stereographic projection of $\S^3$ from $(0,0,1,0)$ onto the corresponding equatorial $\R^3$ hyperplane, i.e.,
\[ \bar{x}=\frac{x}{1-z}, \ \ \bar{y}=\frac{y}{1-z}, \ \bar{w}=\frac{w}{1-z}. \]
\begin{figure}[!h]
\centering
\begin{tikzpicture}
\draw [->] (0, -0.8)--(0, 3.5) node (zaxis) [left] {$\bar{w}$};
\draw [->] (-2, 1.3)--(2, 1.3) node (xaxis) [below right] {$\bar{x}$};
\draw [->] (-1.7, 0.1)--(1.7, 2.5) node (yaxis) [right] {$\bar{y}$};
\draw [-] (0,1.3) ellipse (1.5 and .45);
\draw (0, 1.3) circle (1.5);
\draw [->] (-2,2.5)--(-1.5, 1.85) node at (-2,2.5) [above] {$\bar{x}^2 +\bar{y}^2 + \bar{w}^2 =1 $};
\fill (0,3.2) circle (2.5pt) node[right] {$m_1$};
\fill (0,-0.6) circle (2.5pt) node[right] {$m_2$};
\fill (1,1.3) circle (2.5pt) node[right] {$m_3$};
\fill (-0.3,1.6) circle (2.5pt) node[above left] {$m_4$};
\fill (-0.8,1.1) circle (2.5pt) node[below left] {$m_5$};
\end{tikzpicture}
\caption{ A configuration $\q(c, \th)$ with $(c,\th) \in (-1,0)\times ( 0, \frac{\pi}{2})$.} \label{fig:s3}
\end{figure}
\begin{proposition}
For any $(c,\th) \in (-1,0)\times ( 0, \frac{\pi}{2})$, and $(c,\th) \in (0,1)\times ( \frac{\pi}{2}, \pi)$, the configurations $\q(c, \th)$ constructed above are central configurations if
\begin{equation} \label{condition}
m= -\frac{3c |\sin^3 2\th|}{2\cos \th (1-c^2 \cos^2 \th)^{3/2}}.
\end{equation}
Generally, they are ordinary central configurations.
\end{proposition}
\begin{proof}
We check that the central configuration equations $\nabla _{\q_i} U= \lambda \nabla _{\q_i} I, i=1, \cdots, 5$, are satisfied.
The function $U$ can be written as $U=U_1+U_2$, where
\[ U_1= \cot d_{34} + \cot d_{45}+\cot d_{35} ,\
U_2= m^2 \cot d_{12} +m\sum_{i=3}^5 (\cot d_{1i} + \cot d_{2i}).\]
Note that the three equal masses $m_3$, $m_4$, and $m_5$ form an ordinary central configuration themselves, i.e., $\nabla _{\q_i} U_1= \lambda_1 \nabla _{\q_i} I$, for $i= 3,4,5$, $\lambda_1= \frac{-3}{2\sin^3 d_{34}}$, \cite{DSZ-B}. Note that $\nabla_{\q_1}I =\nabla_{\q_2}I=0$ by equation \eqref{I}. Thus $\nabla _{\q_i} U= \lambda \nabla _{\q_i} I$ is satisfied if and only if there is some constant $\lambda_2$ such that
\[ \nabla _{\q_i} U_2= \lambda_2 \nabla _{\q_i} I, i=3,4,5, \ \ {\rm and }\ \ \F_1=\F_2=0. \]
By symmetry, we only need to check $\nabla _{\q_3} U_2= \lambda_2 \nabla _{\q_3} I$, and $\F_1=0$.
Note that $d_{13}=d_{23}=d_{14}=d_{24}=d_{15}=d_{25}$, $d_{34}=d_{45}=d_{35}$,
and
\[ \cos d_{12} = \cos 2\th, \ \cos d_{13} =c\cos \theta, \ \ \cos d_{34} =\frac{3}{2}c^2-\frac{1}{2}. \]
Some straightforward computation shows
\begin{align*}
\nabla _{\q_3} U_2&= \F_{31}+\F_{32}= \frac{m(\q_1-\cos d_{13}\q_3)}{\sin^3 d_{13}} + \frac{m(\q_2-\cos d_{23}\q_3)}{\sin^3 d_{23}}\\
&= \frac{m}{\sin^3 d_{13}}(\q_1+\q_2- 2\cos d_{13} \q_3)=\frac{m}{\sin^3 d_{13}} \left( (0,0,2\cos \th,0)^T-2c\cos \th ( r,0,c,0)^T\right) \\
&=\frac{-2mr \cos \th}{\sin^3 d_{13}} (c, 0, -r, 0)^T
\end{align*}
Using equation \eqref{I}, we obtain $\nabla _{\q_3}I=2rc(c, 0, -r, 0)^T$. Thus we can write that
\[ \nabla _{\q_3} U_2= \lambda_2 \nabla _{\q_3} I, \ \lambda_2=\frac{-m \cos\th}{ c \sin^3 d_{13}}. \]
By direct computation, we obtain
\begin{equation}
\begin{split}
\F_1&= \F_{12}+ \sum _{j=3}^5\F_{1j}
= \frac{m^2}{|\sin ^3 2\th|}(\q_2-\cos 2\th \q_1) + \sum _{i=3}^5\frac{m}{\sin ^3 d_{13}}(\q_i- \cos d_{13}\q_1)\\
&= \frac{m^2}{|\sin ^3 2\th|}(\q_2-\cos 2\th \q_1) + \frac{m}{\sin ^3 d_{13}}(\sum _{i=3}^5\q_i- 3c \cos \th \q_1)\\
&= m \sin \th \left( \frac{2m \cos \th }{|\sin ^3 2\th|} +
\frac{3c}{\sin ^3 d_{13}} \right) (0,0, \sin \th, -\cos \th)^T.
\end{split}
\notag
\end{equation}
Thus $\F_1=0$ if and only if $m= -\frac{3c |\sin^3 2\th|}{2\cos \th (1-c^2 \cos^2 \th)^{3/2}}$. Since we need positive masses, $c \cos \th$ needs to be negative.
We have thus obtained a two-parameter family of central configurations $\q(c, \th)$ for any $(c,\th) \in (-1,0)\times ( 0, \frac{\pi}{2})$, and $(c,\th) \in (0,1)\times ( \frac{\pi}{2}, \pi)$. The central configuration equations $\nabla _{\q_i} U= \lambda(c, \th) \nabla _{\q_i} I, i=1, \cdots, 5$, are satisfied, and the constant is
\begin{align*}
\lambda (c, \th)&= \lambda_1+\lambda_2=\frac{-3}{2\sin^3 d_{34}} - \frac{m \cos \th}{c \sin^3 d_{13}}= \frac{-3}{2\sin^3 d_{34}} + \frac{3|\sin^3 2\th|} {2 \sin^6 d_{13}} \\
&= \frac{3}{2}\left( \frac{-8}{3\sqrt{3}(1+3c^2)^{3/2} (1-c^2)^{3/2} } + \frac{|\sin^3 2\th|}{ (1-c^2 \cos^2 \th)^{3}} \right),
\end{align*}
which is zero on a one-dimensional manifold. Factually, it is homeomorphic to two open unit intervals.
Thus generally, $\q(c, \th)$ are ordinary central configurations. This remark completes the proof.
\end{proof}
Moreover, if $(c,\th) \in (-1,0)\times ( 0, \frac{\pi}{2})$, then the masses $m_3, m_4, m_5$ are contained in the unit ball, $\bar{x}^2 +\bar{y}^2 + \bar{w}^2 \le1$, and the masses $m_1, m_2$ are outside, see Figure \ref{fig:s3}. This happens because
\[ \bar{w}_1 = \frac{w_1}{1-z_1}= \frac{\sin \th}{1-\cos \th}>1,\
\bar{x}_3^2 +\bar{y}_3^2= (\frac{x_3}{1-z_3})^2+ (\frac{y_3}{1-z_3})^2=\frac{1+c}{(1-c)}<1. \]
Similarly, if $(c,\th) \in (0,1)\times ( \frac{\pi}{2}, \pi)$, then masses $m_3, m_4, m_5$ are outside, but the masses $m_1, m_2$ are inside the ball.
Obviously, we can still obtain central configurations if we substitute the equilateral triangle by a regular $n$-gon with equal masses, and generally they are ordinary ones. \\
\textbf{Acknowledgements.} The authors thank F. Diacu for reading the original manuscript and making many useful suggestions. Suo Zhao is supported by NSFC 11501530 and NSFC 11571242. Shuqiang Zhu is funded by a University of Victoria Scholarship and a David and Geoffery Fox Graduate Fellowship.
|
\section{Introduction}
\IEEEPARstart{N}{owadays}, both carrier frequency and direction of arrival (DOA) are needed in some applications, such as Cognitive Radio (CR) aiming at solving the spectral congestion \cite{Haykin2005, Yucek2009, Mishali2011a, Sun2013, Cohen2014}. The most important function of CRs is to autonomously exploit locally unused spectrum to provide new paths to spectrum access. Therefore, spectrum sensing is an essential part of CRs. The conventional spectrum opportunity only contains three dimensions of the spectrum space: frequency, time, and space. However, with the advancement in array processing technologies \cite{Krim1996, Schmidt1986, Roy1986}, the new dimension, DOA, also creates new spectrum opportunities. Joint frequency spectrum and spatial spectrum would enhance the performance of CRs.
Recently, significant effort have been made towards jointly estimation of carrier frequencies and their DOAs \cite{Lemma1998, Lemma2003}. An obvious drawback is that they require additional pairing between the carrier frequencies and the DOAs. Besides, both works assume that the signal is sampled at least at its Nyquist rate. The main challenge of CRs lies in wideband signal procesing for their costly or even unreachable Nyquist rate sampling. The distribution range of the spectrum under monitoring is from 300 MHz to several GHz \cite{Haykin2005, Yucek2009, Mishali2011a, Sun2013, Cohen2014}. It leads to high Nyquist sampling rate and a large number of sampling data to process.
Fortunately, sub-Nyquist sampling technology can reconstruct a multiband signal from its sub-Nyquist samples \cite{Mishali2011,Mishali2010,Eldar2009,Mishali2009}. Latterly, some joint DOA and carrier frequency estimation methods are proposed at sub-Nyquist sampling rates. \cite{Ariananda2013} proposes a structure, i.e. a linear array by employing a multi-coset sampling at the output of every sensor. This method compresses the wide-sense stationary signal in both the time domain and spatial domain. To simplify the hardware complexity, \cite{Kumar2014} uses an additional identical delayed channel at the output of every sensor. But there are ambiguities during pairing with their corresponding DOAs in an underlying uniform linear array (ULA) scenario. To solve the pairing issue, \cite{ Kumar2015} proposes a structure with the hardware complexity identical to that of \cite{Kumar2014}. However, those papers do not give a unified signal reception model. \cite{Stein2015} presents two joint DOA and carrier frequency recovery approaches for an L-shaped ULA scenario. In \cite{Liu2016}, we propose a new array receiver architecture associated with two sub-Nyquist sampling based methods for simultaneously estimate the frequencies and DOAs of multiple narrowband far-field signals impinging on a ULA, where signals’ carrier frequencies spread around the whole wide spectrum. The architecture is complex due to every sensor following a multi-channel sub-Nyquist sampling receiver.
We consider a scene as \cite{Liu2016} in this paper. For reducing the complexity of receiver, we propose a simplified array receiver architecture. For this model, we propose a unified formula and methods for joint estimation of DOA and carrier frequency.
The following notations are used in the paper. ${\left( \cdot \right)^{\rm T}}$ and ${\left( \cdot \right)^{\rm H}}$ denote the transpose and Hermitian transpose, respectively. $E\left( \cdot \right)$ stands for the expectation operator. ${x_j}$ is the $j$th entry of a vector ${\bf{x}}$. ${{\bf{A}}_i}$ and ${A_{ij}}$ are the $i$th column and $(i,j)$th entry of a matrix ${\bf{A}}$, respectively. $ \otimes$ denotes the Hadamard product. ${{\bf{I}}_M}$ stands for an $M \times M$ identity matrix.
\section{Array Signal Model with Sub-Nyquist Sampling}
In \cite{Liu2016}, we proposed an array signal receiver architecture and the corresponding signal reception model, which introduces sub-Nyquist sampling technology. In this letter, on one hand, the proposed architecture is the simplified form of the previous architecture, on the other hand, we will take advantage of the previous model when estimation algorithm deducing. Therefore, we review the main conclusions of \cite{Liu2016} in this section.
Consider $K$ narrowband far-field signals impinging on a ULA composed of $M$ $(M>K)$ sensors. Our previous receiver architecture applies multi-coset sampling \cite{Mishali2009}. And every array sensor is followed by same $P$ delay branches. All the ADCs are synchronized and samples at a sub-Nyquist sampling rate of ${f_s} = {{{f_N}} \mathord{\left/
{\vphantom {{{f_N}} L}} \right.
\kern-\nulldelimiterspace} L}$, where ${f_N} = {1 \mathord{\left/
{\vphantom {1 {{T_N}}}} \right.
\kern-\nulldelimiterspace} {{T_N}}}$ is the Nyquist sampling rate. The constant set $C=[c_1,c_2,\cdots,c_P]$ is the sampling pattern where
$0 \le {c_1} < {c_2} < \cdots < {c_P} \le L - 1$. ${y_{mp}}\left[ n \right]$ denotes the sampled signal corresponding to the $m$th sensor, $p$th branch.
The matrix output of all branches of all sensors is given by
\begin{align}\label{Yf}
{\bf{Y}}\left( f \right) &= \left( {{\bf{A}} \otimes {\bf{B}}} \right)\overline {\bf{S}} \left( f \right) + \left( {{{\bf{I}}_M} \otimes {\bf{B}}} \right)\widehat {\bf{N}}\left( f \right)\\
&\buildrel \Delta \over= {\bf{G}}\overline {\bf{S}} \left( f \right)+{{\bf{I}}_{\bf{B}}}\widehat {\bf{N}}\left( f \right), f \in \mathcal{F} \buildrel \Delta \over = \left[ {0,\frac{1}{{LT}}} \right),\label{Yf2}
\end{align}
where ${B_{il}} = \frac{1}{{\sqrt L }}\exp \left( {j\frac{{2\pi }}{L}{c_i}l} \right)$, ${{{A}}_{mk}} = \exp \left( { - j{\phi _k}\left( {m - 1} \right)} \right)$ is the $mk$th element of the steer array ${\bf{A}}$, where spatial phase
\begin{align}\label{Phi}
{\phi _k} = \frac{{2\pi d\sin \left( {{\theta _k}} \right)}}{{{c \mathord{\left/
{\vphantom {c {{f_k}}}} \right.
\kern-\nulldelimiterspace} {{f_k}}}}},
\end{align}
where ${\theta _k}$ and ${f_k}$ are the DOA and the center frequency of ${s_k}\left( t \right)$, respectively.
$\overline {\bf{S}} \left( f \right) = {\left[ {\overline {\bf{S}} _1^{\rm{T}}\left( f \right),\overline {\bf{S}} _2^{\rm{T}}\left( f \right), \cdots ,\overline {\bf{S}} _K^{\rm{T}}\left( f \right)} \right]^{\rm{T}}}$, ${\overline {\bf{S}} _k}\left( f \right) = {\left[ {{S_{k1}}\left( f \right),{S_{k2}}\left( f \right), \cdots ,{S_{kL}}\left( f \right)} \right]^{\rm{T}}}$, ${S_{kl}}\left( f \right) = {S_k}\left( {f + \frac{{l - 1}}{{LT}}} \right)$, ${S_k}\left( f \right)$ is the Fourier transform of ${s_k}\left( t \right)$. ${\bf{s}}\left( t \right){\rm{ = }}{\left[ {{s_1}\left( t \right),{s_2}\left( t \right), \cdots ,{s_K}\left( t \right)} \right]^{\rm{T}}}$ is the vector of all signal values.
Because ${s_k}\left( t \right)$ is a narrowband signal, there is one, and only one frequency band which is occupied in ${\overline{{\bf{S}}}_k}\left( f \right)$. Further, ${\overline{{\bf{S}}}_{k}}\left( f \right)$ is a sparse vector of length $L$ when $k$ is fixed and there is one, and only one index ( marked as ${l_k}$), which is activated.
${\bf{Y}}\left( f \right) = {\left[ {{\bf{Y}}_1^{\rm{T}}\left( f \right),{\bf{Y}}_2^{\rm{T}}\left( f \right), \cdots ,{\bf{Y}}_M^{\rm{T}}\left( f \right)} \right]^{\rm{T}}}$. The $p$th element of ${Y_{mp}}\left( f \right) = \sqrt L T_N{Y_{mp}}\left( {{e^{j2\pi fT}}} \right)$, which is the discrete-time Fourier transform of the signal ${y_{mp}}\left[ n \right]$ except a coefficient difference $\sqrt L T_N$.
$\widehat {\bf{N}}\left( f \right) = {\left[ {\widehat {\bf{N}}_1^{\rm{T}}\left( f \right), \cdots ,\widehat {\bf{N}}_M^{\rm{T}}\left( f \right)} \right]^{\rm{T}}}$, ${\widehat {\bf{N}} }_m\left( f \right) = \left[ {{N_{m1}}\left( f \right), \cdots ,{N_{ml}}\left( f \right)} \right]^{\rm T}$, ${N_{kl}}\left( f \right) = {N_k}{f + \frac{{l - 1}}{{LT}}}$ is the Fourier transform of ${n_k}\left( t \right)$. ${\bf{n}}\left( t \right)= {\left[ {{n_1}\left( t \right), \cdots ,{n_M}\left( t \right)} \right]^{\rm{T}}} $ is the noise vector, which subjects to the zero-mean circular complex Gaussian distribution with covariance matrix ${\sigma}^2 \bf{I}_M$.
\section{Proposed receiver architecture and joint DOA and frequency estimation algorithm}
\subsection{Proposed receiver architecture}
To largely decrease hardware complexity, we design the simplified receiver architecture when achieving joint frequency and DOA estimate. This architecture is set up based on the previous architecture. The main difference between the two architectures is that the former only reserves all branches of one array sensor and whole same branch of all sensors. The proposed receiver architecture is shown in Fig.\ref{figArcPart}. Without loss of generality, we select all branches of the first sensor and whole first branch of all sensors in the Fig.\ref{figArcPart}. Namely, our output is ${\bf{W}}\left( f \right) = {\left[ {{Y_{11}}\left( f \right),{Y_{12}}\left( f \right), \cdots ,{Y_{1P}}\left( f \right),{Y_{21}}\left( f \right), \cdots ,{Y_{M1}}\left( f \right)} \right]^{\rm T}}$. We define a $(M + P - 1)\; \times MP$ matrix ${\bf{J}}$, where ${J_{ij}} = 1$ for $\;i = 1, \cdots ,P$, and $ j=i$; or $i = P + 1, \cdots ,M + P - 1$, and $j={1 + iP - {P^2}}$; else ${J_{i,j}} = 0$ for else.
We have ${\bf{W}}\left( f \right) = {\bf{JY}}\left( f \right)$. According to (\ref{Yf}), we have
\begin{align}\label{Ypf}
{\bf{W}}\left( f \right) = {\bf{H}}\overline {\bf{S}} \left( f \right) + {\bf{J}}{{\bf{I}}_{\bf{B}}}\widehat {\bf{N}}\left( f \right),f \in\mathcal{F},
\end{align}
where ${\bf{H}} = {\bf{J}}\left( {{\bf{A}} \otimes {\bf{B}}} \right) = {\bf{JG}}$. Combing ${{\bf{I}}_{MP}} = {{\bf{I}}_{\bf{B}}}{\bf{I}}_{\bf{B}}^{\rm H}$ in \cite{Liu2016}, we have
\begin{align}\label{Noise}
{\bf{J}}{{\bf{I}}_{\bf{B}}}\widehat {\bf{N}}\left( f \right){\left( {{\bf{J}}{{\bf{I}}_{\bf{B}}}\widehat {\bf{N}}\left( f \right)} \right)^{\rm H}}{\rm{ = }}{\sigma ^2}{{\bf{I}}_{M + P - 1}}.
\end{align}
\begin{figure}[!t]
\centering
\includegraphics[width=2.0in]{A2.eps}
\caption{Proposed receiver architecture.}
\label{figArcPart}
\end{figure}
\subsection{Algorithm Based on Individual Estimate}
\subsubsection{ Spatial Phase Estimate}
We denote the outputs of the 1st branch of all sensors as ${ {\bf{Q}} }\left( f \right) = {\left[ {\begin{array}{*{20}{c}}
{{{ {\bf{Y}} }_{11}}\left( f \right)}& \cdots &{{{ {\bf{Y}} }_{M1}}\left( f \right)}
\end{array}} \right]^{\rm T}}$. According to \cite{Liu2016}, we have the following equation.
\begin{align}\label{ASub}
{{\bf{Q}}}\left( f \right) = {\bf{A}}{{\bf{Z}}}\left( f \right) + \widehat {\bf{N}}_1\left( f \right),
\end{align}
where ${{\bf{Z}}}\left( f \right) = {\left[ {\begin{array}{*{20}{c}}
{\sum\limits_{l = 1}^L {{{{B}}_{1l}}{S_{1l}}\left( f \right)} }& \cdots &{\sum\limits_{l = 1}^L {{{{B}}_{1l}}{S_{Kl}}\left( f \right)} }
\end{array}} \right]^{\rm{T}}}$.
Because ${{\bf{S}}_{k}}\left( f \right)$ is a $1$-sparse vector of length $L$, and the activated index is $l_k$. We can simplify ${{\bf{Z}}}\left( f \right)$ as ${\bf{Z}}\left( f \right) = {\left[ {\begin{array}{*{20}{c}}
{{B_{1{l_1}}}{S_{1{l_1}}}\left( f \right)}& \cdots &{{B_{1{l_K}}}{S_{K{l_K}}}\left( f \right)}
\end{array}} \right]^{\rm{T}}}$. (\ref{ASub}) is a standard array reception model, there are many existing method to get $\phi$, such as MUSIC, ESPRIT, and so on. Further, we can get the least square solution of ${ {{ {\bf{Z}} }} }\left( f \right)$,
\begin{align}\label{ALS}
{ {{ {\bf{Z}} }} }\left( f \right) = {\bf{A}} ^\dag { {\bf{Q}} }\left( f \right).
\end{align}
\subsubsection{ Frequency Estimate}
According to \cite{Liu2016} section III part B, the output of all branches of the 1st sensor is
\begin{align}\label{BSub}
{{\bf{Y}}_1}\left( f \right) = {\bf{B}}{\overline {\bf{X}} _1}\left( f \right),
\end{align}
where ${\overline {\bf{X}} _1}\left( f \right) = {\left[ {\begin{array}{*{20}{c}}
{\sum\limits_{k = 1}^K {{{{A}}_{1k}}{S_{k1}}\left( f \right)} }& \cdots &{\sum\limits_{k = 1}^K {{{{A}}_{1k}}{S_{kL}}\left( f \right)} }
\end{array}} \right]^{\rm{T}}}$. Since ${{\bf{S}}_{k}}\left( f \right)$ is a $1$-sparse vector of length $L$, ${\overline {\bf{X}} _1}\left( f \right)$ is $K$-sparse vector of length $L$. We denote the support set of ${\overline {\bf{X}} _1}\left( f \right)$ as $\Omega$. We can use the CTF algorithm to solve (\ref{BSub}) to obtain $\Omega $. Then, we hold
\begin{align}
{\overline {\bf{Y}} _1}\left( f \right) = {\bf{B}}{\overline {\bf{X}} _1}\left( f \right) = {{\bf{B}}_\Omega }{\left( {{{\overline {\bf{X}} }_1}} \right)_\Omega }\left( f \right).
\end{align}
Further, we can get the least square solution of ${\left( {{\overline {\bf{X}} _1}} \right)_\Omega }\left( f \right)$,
\begin{align}\label{BLS}
{\left( {{\overline {\bf{X}} _1}} \right)_\Omega }\left( f \right) = {\bf{B}}_\Omega ^\dag {\overline {\bf{Y}} _1}\left( f \right).
\end{align}
\subsubsection{ Spatial Phase and Frequency matching algorithm}
We calculate the cross-correlation function of signal estimates ${{\bf{Z}}\left( f \right)}$ and ${\left( {{{\bf{X}}_1}} \right)_\Omega }\left( f \right)$. The absolute value of the cross-correlation matrix element has the following expression
\begin{align}\label{Rpm}
\left| {{R_{ij}}} \right|
&= \left| {E\left\{ {{B_{1{l_i}}}{S_{i{l_i}}}\left( f \right){{\left( {\sum\limits_{k = 1}^K {{A_{1k}}{S_{k{\Omega _j}}}\left( f \right)} } \right)}^{\rm H}}} \right\}} \right|\\\label{mid}
&= \left| {E\left\{ {{S_{k{l_k}}}\left( f \right){{\left( {{S_{k{\Omega _j}}}\left( f \right)} \right)}^{\rm H}}} \right\}} \right|\\
&= \left\{ {\begin{array}{*{20}{c}}
{ > 0,when\;{l_i} = {\Omega _j}}\\
{ = 0,when\;{l_i} \ne {\Omega _j}}
\end{array}} \right.,1 \le i \le K,1 \le j \le c.
\end{align}
The conditions for the establishment of (\ref{mid}) have the signals are uncorrelated, the magnitudes of both ${B_{1{l_i}}}$ and ${A_{1k}}$ are 1.
If any of the two signal frequencies are in different frequency bands, we have $c = K$, or $c < K$. According to (\ref{Rpm}), we know that there is one absolute value of element is dominant in each row of ${\bf{R}}$. Further, the support index $\mathcal{S}$ of $\bf{H}$ is determined as following:
\begin{align}\label{Supp}
\mathcal{S}_i = \left( {i - 1} \right)L + {\Omega_j}, j = \mathop {\arg \max }\limits_j \left| {{R_{ij}}} \right|,1\leq i \leq K.
\end{align}
With known the support index $\mathcal{S}$, we obtain
\begin{align}\label{CS}
{\bf{W}}\left( f \right){\rm{ = }}{\bf{H}}\overline {\bf{S}} \left( f \right) = {{\bf{H}}_\mathcal{S} }{\overline {\bf{S}} _\mathcal{S} }\left( f \right) + {\bf{J}}{{\bf{I}}_{\bf{B}}}\widehat {\bf{N}}\left( f \right), f \in\mathcal{F}.
\end{align}
Then we have the least square solution of ${\overline {\bf{S}} _\mathcal{S} }\left( f \right)$
\begin{align}\label{MLES}
\overline{\bf{S}}_\mathcal{S}\left( f \right){\rm{ = }}{\bf{H}}_\mathcal{S} ^\dag {\bf{W}}\left( f \right).
\end{align}
We can gain the received signal's frequency ${\overline f _k}$ through $\overline{\bf{S}}_\mathcal{S}\left( f \right)$. Besides, there is a relationship between ${\overline f _k}$ and the original signal's frequency ${ f _k}$,
\begin{align}\label{MatchFreDir}
{f_k} = \left( {{{{\cal S}_k}\% L} - 1} \right)\frac{{{f_N}}}{L} + \overline {{f}}_k.
\end{align}
We can calculate $\theta_k$ through (\ref{Phi}). We outline the main steps of this individual estimate method for partial channels named algorithm JDFPI in table \ref{Alg4}.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.0}
\caption{\textbf{Algorithm JDFPI}}\label{Alg4}
\label{table_example}
\centering
\begin{tabularx}{8.4cm}{lX}
\toprule
1)&According to (\ref{ASub}),obtain $\phi$ applying the MUSIC, ESPRIT algorithm, and so on;\\
2)&Compute ${ {{ {\bf{Z}} }} }\left( f \right)$ according to (\ref{ALS});\\
3)&Apply the CTF algorithm to solve (\ref{BSub}) to obtain $\Omega $;\\
4)&Compute ${\left( {{\overline {\bf{X}} _1}} \right)_\Omega }\left( f \right)$ according to (\ref{BLS});\\
5)&Determine the support index $\mathcal{S}$ according to (\ref{Supp}); \\
6)&Compute ${\overline {\bf{S}} _\mathcal{S} }\left( f \right)$ according to (\ref{CS}); \\
7)&Determine ${\overline f _k}$ through $\overline{\bf{S}}_\mathcal{S}\left( f \right)$ applying the MUSIC, ESPRIT algorithm, and so on; \\
8)&Acquire ${ f _k}$ according to (\ref{MatchFreDir}); \\
9)&Calculate $\theta_k$ through (\ref{Phi});\\
\bottomrule
\end{tabularx}
\end{table}
\subsection{Algorithm Based on subspace decomposition}
If we calculate the covariance matrix of ${\bf{W}}\left( f \right)$, and take advantage of the subspace decomposition theory as \cite{Schmidt1986}, we will have similar conclusion:
\begin{align}\label{Orth}
{{\bf{a}}_l}\left( \phi \right) \bot {{\bf{U}}_N},
\end{align}
where ${{\bf{U}}_N}$ is the noise subspace. The difference is ${{\bf{a}}_l}\left( \phi \right) = {\bf{J}}\left( {{\bf{a}}\left( \phi \right) \otimes {{\bf{B}}_l}} \right)$. Similarly, we can execute the steps of Algorithm JDFSD \cite{Liu2016}. It is worth pointing out that ${\bf{G}}$ and ${\bf{Y}}\left( f \right)$ in \cite{Liu2016} need to be replaced by ${\bf{H}}$ and ${\bf{W}}\left( f \right)$, respectively. We name this method as algorithm based on subspace decomposition for partial channels (JDFSDPJ).
\subsection{Performance Analysis: Cram\'{e}r\text{-}Rao Bound}
Comparing the model (\ref{CS}) and model (11) in \cite{Liu2016} and noticing that (\ref{Noise}) holds, and making use of the conclusion of Section V equation (29) in \cite{Liu2016}, we have
\begin{align}\label{CRB4Sim}
{\rm{CRB}}_{sub}(sim) &= \frac{\sigma ^2}{{2T/L}}{\left( {\Re \left( {\left( {{{\bf{E}}^{\rm{H}}}{{\bf{P}}_{{{\bf{H}}_{\mathcal{S}}}}}{\bf{E}}} \right) \odot {\bf{R}}_{\overline {\bf{S}} }^{\rm{H}}} \right)} \right)^{ - 1}}\nonumber \\
&= \frac{{{\sigma ^2}}}{{2T}}{\left( {\Re \left( {\left( {{{\bf{E}}^{\rm{H}}}{{\bf{P}}_{{{\bf{H}}_{\mathcal{S}}}}}{\bf{E}}} \right) \odot {\bf{R}_{\bf{S}}^{\rm{H}}}} \right)} \right)^{ - 1}},
\end{align}
where ${{\bf{P}}_{{{\bf{H}}_{\mathcal{S}}}}} = {\bf{I}} - {{\bf{H}}_{\mathcal{S}}}{\bf{H}}_{\mathcal{S}}^\dag $, where ${\bf{H}}_{\mathcal{S}}^\dag = {\left( {{\bf{H}}_{\mathcal{S}}^{\rm{H}}{{\bf{H}}_{\mathcal{S}}}} \right)^{ - 1}}{\bf{H}}_{\mathcal{S}}^{\rm{H}}$, ${\bf{E}} = \left[ {{{\bf{E}}_1}, \cdots ,{{\bf{E}}_K}} \right]$, ${{\bf{E}}_i} = \frac{{d{{{\bf{H}}_{{\mathcal{S}}_i}}}}}{{d{\phi _i}}}$.
\section{Simulation}
In this section, we present the numerical simulation results to illustrate the performance of the proposed algorithms. For the sake of comparison, we take JDFSD in \cite{Liu2016} as a representative of full structure, as JDFSD and JDFTD have the same performance. We set the receiver structure as \cite{Liu2016}, and we take the all branches of the 1st sensor and the 1st branch of all sensors as our simplified structure. For the same reason mentioned in \cite{Liu2016}, we will only give the phase estimation simulation result rather than the DOA estimation simulation result in those simulations.
\subsection{Performance with noise}
Firstly, we will show our model can be solved by the proposed algorithm in different noise levels. In this subsection, the simulation scenario is the same as section VI-A in \cite{Liu2016}.
Fig.\ref{figPS}-Fig.\ref{figFS} depict the RMSE versus SNR in terms of spatial phase and frequency estimation, respectively. Fig.\ref{figPS} shows that the phase estimation performance of algorithms JDFSDPJ and JDFPI improves with SNR, where JDFSDPJ achieves the $\textrm{CRB}_{sub}(Sim)$. The phase estimation performance of JDFSDPJ is better than that of JDFPI is because of jointly using the information in frequency domain and spatial domain. And we observe that $\textrm{CRB}_{sub}(Sim)$ lies between $\textrm{CRB}_{sub}$ and $\textrm{CRB}_{Ny}$. $\textrm{CRB}_{sub}(Sim)$ is higher than $\textrm{CRB}_{sub}$ is obvious. The simplified structure use the jointly information from frequency domain and spatial domain. It leads to a big improvement although the simplified structure have much less samplings comparing with Nyquist sampling. In Fig.\ref{figFS} demonstrates that the frequency estimation performances of JDFSDPJ and JDFPI can achieve the $\textrm{CRB}_{sub}(Sim)$, which is certainly higher than $\textrm{CRB}_{sub}(Sim)$ because of using less branches.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{PhiVsSNR.eps}
\caption{RMSE of phase estimates versus SNR.}
\label{figPS}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{FreVsSNR.eps}
\caption{RMSE of frequency estimates versus SNR.}
\label{figFS}
\end{figure}
\subsection{Performance with various signal number}
In this subsection, we will investigate the estimation performance when the signal number changes as section VI-C in \cite{Liu2016}.
Fig.\ref{figPK} shows that the phase (DOA) estimation performance of algorithm JDFSDPJ is slightly influenced by the signal number and achieves $\textrm{CRB}_{sub}(Sim)$, however JDFPI is influenced by the signal number. This is due to the former jointly using the information from frequency domain and spatial domain. To some degree, it maintains good robustness in terms of the number of signals as JDFSD. Without doubt the performance of JDFSDPJ is still worse than $\textrm{CRB}_{sub}$. Fig.\ref{figFK} shows that the frequency estimation performances of algorithms JDFSDPJ and JDFPI are not influenced by the signal number and can reach $\textrm{CRB}_{sub}(Sim)$.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{PhiVsK.eps}
\caption{RMSE of phase estimates versus number of source.}
\label{figPK}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{FreVsK.eps}
\caption{RMSE of frequency estimates versus number of source.}
\label{figFK}
\end{figure}
\section{Conclusions}
In this paper, we designed an simplified array receiver architecture by introducing sub-Nyquist sampling technology. We realized the joint DOA and frequency estimation under lower sampling rate. Although the estimate precision of using partial channels is worse than that of using full channels, the former has lower equivalent sampling rate and hardware complexity. And increase time of sensing will enhance its estimation performance. The simulations demonstrated that the joint algorithm can closely match the CRB according to noise levels and source number as well.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section*{Supplemental Material}
\newcommand{\beginsupplement}
\setcounter{table}{0}
\renewcommand{\thetable}{S\Roman{table}
\setcounter{figure}{0}
\renewcommand{\thefigure}{S\arabic{figure}
\setcounter{equation}{0}
\renewcommand{\theequation}{S\arabic{equation}}%
}
\beginsupplement
This supplemental material provides additional discussions and further
(theoretical and experimental) details.
\textbf{Work probability distribution}. In the feedback control protocol
depicted in Fig.~\ref{figS1}, the mean work done on the system is
given by the averaged work of each possible history of the feedback
process weighted by its corresponding probability
\begin{equation}
\left\langle W\right\rangle =\sum_{k,l}p\left(k,l\right)\text{U}\left(\rho_{\tau_{2}}^{\left(k,l\right)}\right)-\text{U}(\rho_{0}^{eq}),\label{eq:average work}
\end{equation}
where $p\left(k,l\right)=p\left(k|l\right)p\left(l\right)$ is the
joint probability for the $l$-th measurement outcome and $k$-th
feedback operation, $\text{U}\left(\rho_{0}^{eq}\right)=\mbox{tr}\left[\mathcal{H}_{0}\rho_{0}^{eq}\right]$
and $\text{U}\left(\rho_{\tau_{2}}^{\left(k,l\right)}\right)=\mbox{tr}\left[\mathcal{H}_{\tau_{2}}^{\left(k\right)}\mathcal{F}^{\left(k\right)}\left(\rho_{\tau_{1}}^{\left(l\right)}\right)\right]$
are the initial and final internal energy, respectively, $\rho_{\tau_{2}}^{\left(k,l\right)}=\mathcal{F}^{\left(k\right)}\left(\rho_{\tau_{1}}^{\left(l\right)}\right)$
are the possible system's final states. The operator sum decomposition
of the feedback operation is $\mathcal{F}^{\left(k\right)}\left(\cdot\right)=\sum_{j}\Gamma_{j}^{\left(k\right)}\left(\cdot\right)\Gamma_{j}^{\left(k\right)\dagger}$,
whereas the post-measurement state of the $l$-th projective measurement
is $\rho_{\tau_{1}}^{\left(l\right)}=\mathcal{M}_{l}\mathcal{U}\rho_{0}^{eq}\mathcal{U}^{\dagger}\mathcal{M}_{l}/p\left(l\right)$.
Since the unital processes considered here do not involve energy exchange
with the reservoir, the change in the internal energy is regarded
as work. Using the spectral decomposition of both Hamiltonians, $\mathcal{H}_{0}=\sum_{n}\varepsilon_{n}^{(0)}\Pi_{n}^{0}$
and $\mathcal{H}_{\tau_{2}}^{\left(k\right)}=\sum_{m}\varepsilon_{m}^{\left(\tau_{2},k\right)}\Pi_{m}^{\left(\tau_{2},k\right)}$,
one can write Eq.~(\ref{eq:average work}) as
\begin{eqnarray}
\left\langle W\right\rangle & = & \sum_{m,j,k,l,n}p\left(k|l\right)p\left(m,j,l,n\right)\Delta\varepsilon_{m,n}^{\left(k\right)}\nonumber \\
& = & \sum_{m,j,k,l,n}p\left(m,j,k,l,n\right)\Delta\varepsilon_{m,n}^{\left(k\right)},\label{eq:average}
\end{eqnarray}
with $p\left(m,j,l,n\right)\equiv\mbox{tr}\left(\Pi_{m}^{\left(\tau_{2},k\right)}\Gamma_{j}^{\left(k\right)}\mathcal{M}_{l}\mathcal{U}\Pi_{n}^{(0)}\rho_{0}^{eq}\mathcal{U}^{\dagger}\mathcal{M}_{l}\Gamma_{j}^{\left(k\right)\dagger}\right)$,
$p\left(m,j,k,l,n\right)\equiv p\left(k|l\right)p\left(m,j,l,n\right)$,
and $\Delta\varepsilon_{m,n}^{\left(k\right)}=\varepsilon_{m}^{\left(\tau_{2},k\right)}-\varepsilon_{n}^{(0)}$.
We can express the work distribution in the presence of feedback as
$P\left(W\right)=\sum_{m,j,k,l,n}p\left(m,j,k,l,n\right)\delta\left(W-\Delta\varepsilon_{m,n}^{\left(k\right)}\right)$
and the average work as $\left\langle W\right\rangle =\int dW\,P\left(W\right)W$.
This work distribution can also be related to the two-point energy
measurement paradigm~\cite{Talkner2007}, considering a measurement
on the energy basis of $\mathcal{H}_{0}$ at the beginning of the
protocol described in Fig.~\ref{figS1} (before the unitary driven
$\mathcal{U}$) and another measurement at the end of the protocol
in the energy basis of $\mathcal{H}_{\tau_{2}}^{\left(k\right)}$
in the $k$-th history. Notice that in the presence of the feedback,
the second measurement of the two-point paradigm depends on the feedback
operation implemented since it can drive the Hamiltonian to $\mathcal{H}_{\tau_{2}}^{\left(k\right)}$
as illustrated in Fig.~\ref{figS1}.
\begin{figure}[h]
\includegraphics[scale=0.25]{fig4.png}\caption{\textbf{Two-point measurement paradigm for work distribution in the
presence of feedback}. The energy basis measurements, $\Pi_{n}^{0}$,
for initial system Hamiltonian $\mathcal{H}_{0}$ and, $\Pi_{m}^{\left(\tau_{2},k\right)}$,
for $k$-th history of the feedback control (corresponding the final
Hamiltonian $\mathcal{H}_{\tau_{2}}^{\left(k\right)}$), are regarded
here as mathematical tools for the definition of the work probability
distribution. }
\label{figS1}
\end{figure}
\textbf{Fluctuation relation in the presence of a unital feedback}.
For the sake of completeness, we will verify the validity of the fluctuation
relation in Eq.~(\ref{eq:fluctuation relation}) of the main text, which was previously discussed
in Refs.~\cite{Rastegin2013,Albash2013}. Consider the following
average
\begin{align}
\left\langle e^{-\beta\left(W-\Delta\text{F}^{\left(k\right)}\right)-I^{\left(k,l\right)}}\right\rangle & =\sum_{m,j,k,l,n}p\left(m,j,k,l,n\right)\nonumber \\
& \times e^{-\beta\left(\Delta\varepsilon_{m,n}^{\left(k\right)}-\Delta F^{\left(k\right)}\right)-I^{\left(k,l\right)}.}\nonumber \\
& =\sum_{m,j,k,l,n}p\left(m,j,l,n\right)\nonumber \\
& \times Z_{0}e^{+\beta\varepsilon_{n}^{(0)}}\frac{e^{-\beta\varepsilon_{m}^{\left(\tau_{2},k\right)}}}{Z_{\tau_{2}}^{\left(k\right)}}p\left(k\right).\label{eq:aux2}
\end{align}
Remembering the definition of $p\left(m,j,l,n\right)$ introduced
in the previous section and identifying $\Pi_{n}^{(0)}\rho_{0}^{eq}=Z_{0}^{-1}e^{-\beta\varepsilon_{n}^{0}}\Pi_{n}^{(0)}$
and $\rho_{\tau_{2}}^{\left(k,eq\right)}=\frac{1}{Z_{\tau_{2}}^{\left(k\right)}}\sum_{m}e^{-\beta\varepsilon_{m}^{\left(\tau_{2},k\right)}}\Pi_{m}^{\left(\tau_{2},k\right)}$;
the right hand side (rhs) of Eq.~(\ref{eq:aux2}) can be simplified
to $\sum_{k,j,l}p\left(k\right)\,\mbox{tr}\left(\rho_{\tau_{2}}^{\left(k,eq\right)}\Gamma_{j}^{\left(k\right)}\mathcal{M}_{l}\Gamma_{j}^{\left(k\right)\dagger}\right)$.
Using the completeness of the measurement and the unitality of the
map, $\mathcal{F}^{(k)}(\mathds{1})=\sum_{k}\Gamma_{j}^{\left(k\right)}\Gamma_{j}^{\left(k\right)\dagger}=\mathds{1}$,
Eq.~(\ref{eq:aux2}) turns out to be
\[
\left\langle e^{-\beta\left(W-\Delta\text{F}^{\left(k\right)}\right)-I^{\left(k,l\right)}}\right\rangle =\sum_{k}p\left(k\right)\,\mbox{tr}\left(\rho_{\tau_{2}}^{\left(k,eq\right)}\right).
\]
Since $\mbox{tr}\left(\rho_{\tau_{2}}^{\left(k,eq\right)}\right)=1$
and $p\left(k\right)$ is a normalized distribution, we obtain the
fluctuation relation in Eq.~(\ref{eq:fluctuation relation}) of the main text.
\textbf{Derivation of the trade-off relation}. Let us start from KL
relative entropy between an arbitrary state, $\rho$, and the equilibrium
state, $\rho^{eq}$, associated with the Hamiltonian $\mathcal{H}$
at inverse temperature $\beta$,
\begin{eqnarray}
S_{KL}\left(\rho||\rho^{eq}\right) & = & \mathrm{tr}\left[\rho\left(\ln\rho-\ln\rho^{eq}\right)\right]\nonumber \\
& = & \mathrm{-tr}\left(\rho\ln\frac{e^{-\beta\mathcal{H}}}{Z}\right)-S\left(\rho\right)\nonumber \\
& = & \mathrm{\beta tr}\left(\rho\mathcal{H}\right)-\ln Z-S\left(\rho\right)\nonumber \\
& = & \beta\left[\text{U\ensuremath{\left(\rho\right)}}-\text{F}\right]-S\left(\rho\right).\label{eq:identityutil}
\end{eqnarray}
From the above identity we can write $\beta U\left(\rho_{\tau_{2}}^{\left(k,l\right)}\right)=S_{KL}\left(\rho_{\tau_{2}}^{\left(k,l\right)}||\rho_{\tau_{2}}^{\left(k,eq\right)}\right)+\beta\text{F}_{\tau_{2}}^{(k)}+S\left(\rho_{\tau_{2}}^{\left(k,l\right)}\right)$
and $\beta U\left(\rho_{0}^{eq}\right)=\beta\text{F}_{0}+S\left(\rho_{0}^{eq}\right)$,
with $\text{F}_{\tau_{2}}^{(k)}=-\beta^{-1}\ln Z_{\tau_{2}}^{(k)}$
and $\text{F}_{0}=-\beta^{-1}\ln Z_{0}$ . These results combined
with Eq.~(\ref{eq:average work}) lead to the following expression
for the mean nonequilibrium entropy production in the presence of
feedback:
\begin{eqnarray}
\left\langle \Sigma\right\rangle & = & \beta\left\langle W-\Delta\text{F}^{(k)}\right\rangle \nonumber \\
& = & -S\left(\rho_{0}^{eq}\right)+\sum_{l}p\left(l\right)S\left(\rho_{\tau_{1}}^{\left(l\right)}\right)\nonumber \\
& & +\sum_{k,l}p\left(k,l\right)S_{KL}\left(\rho_{\tau_{2}}^{\left(k,l\right)}||\rho_{\tau_{2}}^{\left(k,eq\right)}\right)\nonumber \\
& & +\sum_{k,l}p\left(k,l\right)\left(S\left(\rho_{\tau_{2}}^{\left(k,l\right)}\right)-S\left(\rho_{\tau_{1}}^{\left(l\right)}\right)\right),\label{eq:auxfinal}
\end{eqnarray}
where we have added and subtracted the averaged entropy $\sum_{l}p\left(l\right)S\left(\rho_{\tau_{1}}^{\left(l\right)}\right)$.
The rhs of Eq.~(\ref{eq:auxfinal}) can be identified with the
information-theoretic quantities (i.e. information gain, KL relative
entropy, and the von Neumann entropy variation, respectively) resulting
in the trade-off relation in Eq. ~(\ref{eq: trade-off equality}) of the main text.
\textbf{Information gain.} When performing a measurement on a quantum system, the observer acquires
information about the system. The description of such a process plays
a central role in quantum measurement theory \cite{Jacobs2014}. The
information gain was first proposed by Groenewold in 1971 \cite{Groenewold1971}
whose aim was to quantify how much information is obtained by a quantum
measurement. Groenewold proposed that the information gained by the
observer is given by the average uncertainty reduction of the quantum
state, and its given by
\begin{equation}
\mathcal{I}_{gain}=S\left(\rho\right)-\sum_{l}p\left(l\right)S\left(\rho^{(l)}\right),\label{eq:igain-sup}
\end{equation}
where $\rho$ is the state immediately before the measurement and
$\rho^{(l)}$ the post-measurement states, which occur with outcome
probability $p(l)$. A year after its introduction, Lindblad showed
that the information gain, as expressed in Eq.~(\ref{eq:igain-sup}),
is non-negative for von Neumann projective measurements \cite{Lindblad1972}.
Over a decade later, Ozawa generalized the demonstration of Lindblad
showing that the information gain is non-negative for any positive-operator
valued measure (POVM) \cite{Ozawa1986}. POVMs and von Neumann measurements
are called efficient measurements because each measurement operator
is associated with a single outcome (each measurement operator is
described by a single Kraus operator). On the other hand, measurements
which possess an intrinsic classical uncertainty are called inefficient.
Inefficient measurements induce more back-action to the system than
quantum mechanics would allow for the same amount of information extracted
\cite{Jacobs2014}. Equation~(\ref{eq:igain-sup}) may becomes negative
for such inefficient measurements and therefore may not be a good
quantifier in this particular case. Recently, Buscemi, Hayashi,
and Horodeck \cite{Buscemi2008a} generalized the expression
for the information gain providing a new framework which gives positive
results for any kind of measurement (efficient or inefficient) and
recovers Eq.~(\ref{eq:igain-sup}) for efficient measurements. Furthermore,
they endorsed the interpretation of the information gain by providing
an operational meaning to it. In order to approach this very general
class of measurements, in Ref.~\cite{Buscemi2008a} the measuring apparatus
is composed by three ancillary systems apart from the system to be
measured (so-called quantum instruments). One ancilla is employed
as a purification system and the other two to emulate the desired
measurement, similar to the role of the quantum memory in the measurement
stage of our feedback protocol. This newly proposed information gain
is based on the mutual information of these auxiliary systems, although
ultimately it depends only on quantities of the system.
In the present article, we have considered projective measurements
(which are experimentally implemented with a very high accuracy assessed
by quantum tomography), therefore the Groenewold definition of information
gain, in Eq.~(\ref{eq:igain-sup}), is suitable to be applied throughout
all of our analyzes. We also note that Eq.~(\ref{eq: trade-off equality}) of the main text holds
for a general projective measurement feedback control based on unital
operations. An interesting point for future investigation is the possibility
to derive an information thermodynamics trade-off relation for a feedback
mechanism involving inefficient measurements.
For the sake of clarity, we provide below a short demonstration
(based on Refs. \cite{Jacobs2014,Fuchs2001}) that the information
gain as defined in Eq.~(\ref{eq:igain-sup}) is non-negative for POVM
measurements. Any operator $A$ may be decomposed as $A=PU$,
where $P$ is a positive operator and $U$ is a unitary operator.
This decomposition is called the polar decomposition of $A$. Using
this decomposition we can show that $AA^{\dagger}=P^{2}$. Therefore,
$A^{\dagger}A=U^{\dagger}P^{2}U=U^{\dagger}AA^{\dagger}U$, which
means that the operators $A^{\dagger}A$ and $AA^{\dagger}$ possess
the same set of eigenvalues (since unitary transformation preserve
eigenvalues). For a POVM, $\left\{ M_{l}^{\dagger}M_{l}\right\} $,
the $l$-th post-measurement state, $\rho^{(l)}=M_{l}\rho M_{l}^{\dagger}/p(l)$,
is obtained with probability $p(l)=\mbox{tr}\left(M_{l}^{\dagger}M_{l}\rho\right)$.
The measurement operators of the POVM satisfy the completeness relation $\sum_{l}M_{l}^{\dagger}M_{l}=\mathds{1}$.
We can write the state before the measurement, $\rho,$ in the following
convenient form
\begin{equation}
\rho=\sqrt{\rho}\mathds{1}\sqrt{\rho}=\sum_{l}\sqrt{\rho}M_{l}^{\dagger}M_{l}\sqrt{\rho}=\sum_{l}p(l)\mu^{(l)},
\end{equation}
where $\mu^{(l)}=\sqrt{\rho}M_{l}^{\dagger}M_{l}\sqrt{\rho}/p(l)$.
The von Neumann entropy is a concave function, which means $S\left(\sum_{n}p(n)\rho^{(n)}\right)\geq\sum_{n}p(n)S\left(\rho^{(n)}\right)$
with $\sum_{n}p(n)=1$. From the concavity of the von Neumann entropy
we obtain
\begin{equation}
S\left(\rho\right)\geq\sum_{l}p(l)S\left(\mu^{(l)}\right).
\end{equation}
Defining $X_{l}=\sqrt{\rho}M_{l}/\sqrt{p(l)}$ we write $\mu^{(l)}=X_{l}X_{l}^{\dagger}$.
So $\mu^{(l)}$ has the same eigenvalues as $\rho^{(l)}=X_{l}^{\dagger}X_{l}=M_{l}\rho M_{l}^{\dagger}/p(l)$.
Since the von Neumann entropy is a function of only the eigenvalues
of the density operator $S\left(\mu^{(l)}\right)=S\left(\rho^{(l)}\right)$,
which implies the non-negativity of the information gain defined in
Eq.~(\ref{eq:igain-sup}) for any POVM.
\textbf{Experimental set-up.} The liquid sample consist of $50$~mg
of $99$\% $^{13}\text{C}$-labeled $\text{CHCl}_{3}$ (Chloroform)
diluted in $0.7$ ml of $99.9$\% deutered Acetone-d6, in a flame
sealed Wildmad LabGlass 5 mm tube. All experiments are carried out
in a Varian $500$~MHz Spectrometer employing a double-resonance
probe-head equipped with a magnetic field gradient coil. Chloroform
sample is very diluted so that the intermolecular interaction can
be neglected, and the sample can be regarded as a set of identically
prepared pairs of spin-1/2 systems. The sample is placed in the presence
of a longitudinal static magnetic field (whose direction is taken
to be along the positive $z$ axes) with strong intensity, $B_{0}\approx11.75$
T. The nuclear magnetization of $^{1}\text{H}$ and $^{13}\text{C}$
precess around $B_{0}$ with Larmor frequencies about $500$~MHz
and $125$~MHz, respectively. Magnetization of the nuclear spins
are controlled by time-modulated rf-field pulses in the transverse
($x$ and $y$) direction and longitudinal field gradients.
Spin-lattice relaxation times, measured by the inversion recovery
pulse sequence, are \foreignlanguage{english}{$\left(\mathcal{T}_{1}^{H},\mathcal{T}_{1}^{C}\right)=\left(7.42,\:11.31\right)$~}s.
Transverse relaxations, obtained by the Carr-Purcell-Meiboom-Gill
(CPMG) pulse sequence, have characteristic times \foreignlanguage{english}{$\left(\mathcal{T}_{2}^{\text{*H}},\mathcal{T}_{2}^{*\text{C}}\right)=\left(1.11,\:0.30\right)$}~s.
The total experimental running time, to implement the entropy rectification
protocol, is about $22.4$\textbf{\footnotesize{}~}ms, which is considerably
smaller than the spin-lattice relaxation and therefore decoherence
can be disregard. The data for the process tomography, showed in Fig.~\ref{fig2}
of the main text, also endorses this consideration, since the experimentally
implemented process does not exhibit significant decoherence effects.
The initial state of the nuclear spins is prepared by spatial average
techniques \cite{Oliveira2007,Jones2011,Batalhao2014a,Batalhao2015a},
being $^{1}\text{H}$ nucleus prepared in the ground state and the
$^{13}\text{C}$ nucleus in a pseudo-thermal state with the populations
(in the energy basis of $\mathcal{H}_{0}^{\text{C}}$) and corresponding
pseudo-temperatures displayed in Tab.~\ref{tabSI}.
\begin{table}[H]
\caption{\textbf{Population and pseudo-temperature of the Carbon initial states. }}
{
\begin{center}
\footnotesize
$\begin{array}{cccc}
\hline
p_0^{(0)} & p_1^{(0)} & k_{B}T \, \text{(peV)}\\
\hline
\hline
0.96\pm0.01 & 0.04\pm0.01 & 2.6\pm0.2\\
0.92\pm0.01 & 0.08\pm0.01 & 3.4\pm0.2\\
0.88\pm0.01 & 0.12\pm0.01 & 4.2\pm0.2\\
0.84\pm0.01 & 0.16\pm0.01 & 4.9\pm0.2\\
0.81\pm0.01 & 0.19\pm0.01 & 5.9\pm0.3\\
0.76\pm0.01 & 0.24\pm0.01 & 7.0\pm0.3\\
0.73\pm0.01 & 0.27\pm0.01 & 8.6\pm0.4\\
0.69\pm0.01 & 0.31\pm0.01 & 10.7\pm0.6\\
0.65\pm0.01 & 0.35\pm0.01 & 13.8\pm1.0\\
\hline
\end{array}$
\label{tabSI}
\end{center}
}
\end{table}
\textbf{Data acquisition.} Quantum state tomography (QST) is employed to
obtain the relevant information quantities in the controlled feedback
process. We have performed QST along the protocol implementation
of the demon and we have employed an auxiliary circuit as depicted in
Figs.~\ref{figS2}(a) and \ref{figS2}(b), respectively. QST~1 is
used to verify the effective temperature of the initial state. From
QST~2 and QST~3 we obtain the information gain, $\mathcal{I}_{gain}$.
The mutual information, $\left\langle I^{(k,l)}\right\rangle $ is
obtained from QST~3. The remaining information quantities ($\left\langle S_{KL}\left(\rho_{\tau_{2}}^{(k,l)}||\rho_{\tau_{2}}^{(k,eq)}\right)\right\rangle $
and $\left\langle \Delta S^{(k,l)}\right\rangle _{\mathcal{F}}$)
are obtained from the aforementioned tomographic data combined with
QST~4 (obtained from the auxiliary circuit in Fig.~\ref{figS2}(b)). Due to the fact that our implementation is based on nonselective measurements, an auxiliary circuit (Fig.~\ref{figS2}(b)) is employed to obtain the states $\rho_{\tau_{2}}^{(k,l)}$ by QST. This enables us to characterize the information quantities in rhs of Eq.~(\ref{eq: trade-off equality}) in the main text. The optimized pulse sequence used to implement the Maxwell's demon is displayed in Fig.~\ref{figS2}(c).
\begin{figure}[h]
\includegraphics[scale=0.22]{fig5.png}\caption{\textbf{Characterization of the Maxwell's demon operation protocol.
}(a) and (b) Quantum state tomography strategy to obtain the relevant
information quantities. The initial states, as displayed, are prepared
and the feedback control is applied. This enables the full information
characterization of the feedback controlled evolution. QST~1 and
QST~2 are single spin-1/2 tomography realized on the Carbon nucleus,
whereas QST~3 and QST~4 are joint tomographies implemented in both
Hydrogen and Carbon nuclei. (c) Optimized pulse sequence used to implement
the measurement-based feedback control operation. Blue (red) circles
represent one spin transverse rf-pulses producing rotations on the
$x$ ($y$) by the displayed angle. Free evolutions under the natural
$J$ coupling ($\frac{1}{2}\pi J\hbar\sigma_{z}^{\text{H}}\sigma_{z}^{\text{C}}$)
are represented by two-spins connections (in orange), with the time-length
displayed above the connections. The grey regions represent the two
longitudinal field gradients.}
\label{figS2}
\end{figure}
The quantum process tomography, as illustrated in Fig.~\ref{fig2}(a) of the
main text, is carried out by preparing a set of mutually unbiased
basis (MUB) states \cite{Nielsen2000}, implementing the nonselective
projective measurement operation (the full controlled feedback protocol)
in QPT~1 (in QPT~2) and then a full quantum state tomography at
the end. From this data it is possible to obtain the Choi-Jamiolkowski
matrix, $\chi$, of the process. The error in the Maxwell's demon
realization is probed by the process trace distance, $\delta=\frac{1}{2}\text{tr}\left|\mathcal{\chi}^{exp}-\mathcal{\chi}^{id}\right|$,
between the ideal ($id$) and the experimentally ($exp$) implemented
processes, as displayed in Fig.~\ref{fig2}(c) of the main text. Its operational
interpretation is related with the bias for the distinguishability
between the ideal and experimental processes. The average success
probability when distinguishing the two processes is $\frac{1}{2}+\frac{1}{2}\delta$,
when both processes are performed with equal a priori probability.
\textbf{Information bounds for entropy production. }For the projective
nonselective measurement implemented in our protocol, the information
gain reduces to $\mathcal{I}_{gain}=S\left(\rho_{\tau_{1}}\right)$,
in the ideal case. Since the driving process implemented, $\mathcal{U}$,
is a unitary sudden quench, the von Neumann entropy after the quench,
$S\left(\rho_{\tau_{1}}\right)$, is the same as for the initial equilibrium
state, $S\left(\rho_{0}^{eq}\right)$. This latter entropy is also
equal to the classical Shannon entropy $\textsf{H}_{Sh}(p_{0}^{(0)},p_{1}^{(0)})=-\sum_{i=0,1}p_{i}^{(0)}\ln p_{i}^{(0)}$,
of the populations in the initial Hamiltonian ($\mathcal{H}_{0}$)
energy basis (with $p_{i}^{(0)}=\text{tr}\left(\Pi_{i}\rho_{0}^{eq}\right)$).
So $\mathcal{I}_{gain}=\textsf{H}_{Sh}(p_{0}^{(0)},p_{1}^{(0)})$.
On the other hand, the probability for the $l$-th measurement outcome
after the sudden quench is equally weighted in the ideal case, $p(l)=\frac{1}{2}$,
for $l=0,1$. In the absence of basis mismatch, the correlation generated,
by the coherent implementation of the measurement-based feedback,
leads ideally to the joint probability distribution [$p(k,l)$] of
the controlled operation ($k$) and measurement outcome ($l$), $p(0,0)=p(1,1)=\frac{1}{2}$
and $p(0,1)=p(1,0)=0$. This implies that the marginal distribution
for the control operation is \textbf{$p\left(k\right)=\sum_{l}p(k,l)=1/2$}
for $k=0,1$.\textbf{ }Accordingly, $\left\langle I^{(k,l)}\right\rangle =\sum_{k,l}p(k,l)\ln\frac{p(k,l)}{p(k)p(l)}=\textsf{H}_{Sh}(\frac{1}{2},\frac{1}{2})=\ln2$~nats
meaning maximum correlation between system and memory. In fact, the
measurement-based feedback protocol was designed to achieve this maximum.
Since the Shannon entropy, $\textsf{H}_{Sh}(p_{0}^{(0)},p_{1}^{(0)})$,
is upper bounded by $\ln2$~nats, we conclude that $\mathcal{I}_{gain}\leq\left\langle I^{(k,l)}\right\rangle $,
where the inequality is saturated in the limit $\beta\rightarrow0$
as can be noted in the experimental data displayed in Fig.~\ref{fig3}(b) of
the main text.
\textbf{Error analysis. }The main sources of error in the experiments
are small non-homogeneities of the transverse rf-field, non-idealities
in its time modulation, and non-idealities in the longitudinal field
gradient. In order to estimate the error propagation, we have used
a Monte Carlo method, to sampling deviations of the tomographic data
with a Gaussian distribution having widths determined by the variances
corresponding to such data. The standard deviation of the distribution
of values for the relevant information quantities is estimated from
this sampling. The variances of the tomographic data are obtained
preparing the same state ten times, taking the full state tomography
and comparing it with the theoretical expectation. These variances
include random and systematic errors in both state preparation and
data acquisition by QST. The error in each element of the density
matrix estimated from this analysis is about 1\%. All parameters in
the experimental implementation, such as pulses intensity and its
time duration, are optimized in order to minimize errors.
|
\section{Introduction}
In this paper, we consider optimizing the $\ell_2$-norm regularized empirical loss minimization problem which is arising ubiquitously in supervised machine learning:
\begin{eqnarray}
\label{primal}
\min\limits_{w \in \mathbb{R}^d} P(w ) := \min\limits_{w \in \mathbb{R}^d} {\frac{1}{n} \sum\limits_{i=1}^n \phi_i( w) }+ \frac{\lambda}{2} \|w\|^2_2.
\end{eqnarray}
We let $f(w) = {\frac{1}{n} \sum_{i=1}^n \phi_i( w) }$ and $w \in \mathbb{R}^d$ be the linear predictor to be optimized. There are many applications falling into this formulation, such as classification, regression, and principal component analysis (PCA). In classification, given features $x_i \in \mathbb{R}^d$ and labels $y_i \in \{1, -1\}$, we obtain Support Vector Machine (SVM) when we let $\phi_i( w) = \max \{ 0, 1- y_i x_i^T w\}$. In regression, given features $x_i \in \mathbb{R}^d$ and response $y_i \in \mathbb{R} $, we have Ridge Regression problem if $\phi_i( w) = (y_i - x_i^T w)^2$.
Recently, \cite{garber2015fast,allen2016improved} showed that the problem of PCA can be solved through convex optimization. Supposing $C = \frac{1}{n} \sum_{i=1}^n x_i x_i^T$ be normalized covariance matrix, \cite{garber2015fast} showed that approximating the principle component of $A$ is equivalent to minimizing $f(w) = \frac{1}{2} w^T(\mu I - C)w- b^Tw$ given $\mu >0$ and $b\in \mathbb{R}^d$. Defining $\phi_i(w) =\frac{1}{2} w^T((\mu -\frac{\lambda}{2}) I - x_ix_i^T)w - b^Tw$ and $\mu > \sigma_1(C) + \frac{\lambda}{2}$ where $\sigma_1(C)$ denotes the largest singular value of $C$, it also falls into problem (\ref{primal}). In this case, $f(w)$ is convex while each $\phi_i(w)$ is probably non-convex.
Distributed machine learning methods are required to optimize problem (\ref{primal}) when the dataset is distributed over multiple machines.
In \cite{jaggi2014communication}, the authors proposed communication-efficient distributed dual coordinate ascent (CoCoA) for primal-dual distributed optimization. In each iteration, the CoCoA framework allows workers to optimize subproblems independently at first. After that, it calls ``Reduce" operation to collect local solution from all workers, and updates global variable and broadcasts the up-to-date global variable to workers in the end. It uses stochastic dual coordinate ascent (SDCA) as the local solver which is one of the most successful methods proposed for solving the problem (\ref{primal}) \cite{hsieh2008dual,shalev2013accelerated}. In \cite{shalev2013stochastic}, the authors proved that SDCA has linear convergence if the convex function $\phi_i(w)$ is smooth, which is much faster than stochastic gradient descent (SGD). \cite{yang2013trading, takavc2015distributed} also proposed distributed SDCA and analyzed the tradeoff between computation and communication.
\cite{ma2015adding,ma2017distributed} accelerated the CoCoA by allowing for more aggressive updates, and proved that CoCoA has linear primal-dual convergence for the smooth convex problem and sublinear convergence for the non-smooth convex problem. However, there are two issues for these primal-dual distributed methods. Firstly, all of them use SDCA as the local solver. SDCA is not applicable when the dual problem is unknown, \emph{e.g.} $\phi_i$ is non-convex. Therefore, the applications of these primal-dual distributed methods are limited. Secondly, all of these methods assume that the workers have similar computing speed, which is not true in practice. Straggler problem is an unavoidable practical issue in the distributed optimization. Thus, the computing time of CoCoA and distributed SDCA is dependent on the slowest worker. Even if there is only one bad worker, they will work far slower than expectation.
In \cite{shalev2015sdca, shalev2016sdca}, the authors proposed dual free stochastic dual coordinate ascent (dfSDCA). It was proved to admit similar convergence rate to SDCA while it did not rely on duality at all. However there is no distributed optimization method using dfSDCA, and its convergence analysis is still unknown yet.
In this paper, we solve the above two challenging issues in previous primal-dual distributed optimization methods by proposing Distributed Dual Free Stochastic Dual Coordinate Ascent (Dis-dfSDCA). We use dfSDCA as the local solver such that Dis-dfSDCA can be applied to the non-convex problem easily. We alleviate the effect of straggler problem by allowing asynchronous communication between server and workers. As shown in Figure \ref{intro}, the server does not wait and workers may store the stale global variable in the local. We also analyze the convergence rate of our method and prove that it admits linear convergence rate even if the individual losses ($\phi_i$) are non-convex, as long as the sum of losses $f$ is convex. Finally, we conduct simulation on the distributed system with straggler problem. Experimental results verify our theoretical conclusions and show that our method works well in practice.
\begin{figure}[t]
\centering
\includegraphics[width=3.2in]{fig}
\caption{Distributed asynchronous dual free stochastic dual coordinate ascent for parameter server framework. In iteration $t$, the server receives gradient message $v_k$ from worker $k$, and sends the up-to-date $w^{t}$ back to the worker $k$. Global variables in other workers are stale. For example worker $1$ and $K$ store stale global variables $w^{t-2}$ and $w^{t-5}$ respectively. }
\label{intro}
\end{figure}
\section{Preliminary}
To optimize the primal problem (\ref{primal}), we often derive and optimize its dual problem alternatively:
\begin{eqnarray}
\max\limits_{\alpha \in \mathbb{R}^n} D(\alpha):= \max \limits_{\alpha \in \mathbb{R}^n} \frac{1}{n} \sum\limits_{i=1}^n - \phi_i^*(-\alpha_i) - \frac{\lambda}{2} \|\frac{1}{\lambda n} A\alpha \|^2_2\,,
\label{dual}
\end{eqnarray}
where $\phi_i^*$ is the convex conjugate function to $\phi_i$, $A = [x_1,x_2,...x_n] \in \mathbb{R}^{d\times n}$ denotes data matrix and $\alpha \in \mathbb{R}^n$ denotes dual variable. We can use stochastic gradient descent (SGD) to optimize primal problem (\ref{primal}), however, there are always two issues: (1) SGD is too aggressive at the beginning of the optimization; (2) it does not have a clear stopping criterion. One of the biggest advantages of optimizing the dual problem is that we can keep tracking the duality gap $G(\alpha)$ to monitor the progress of optimization. Duality gap is defined as: $G(\alpha) = P(w(\alpha)) - D(\alpha)$, where $P(w(\alpha))$ and $D(\alpha)$ denote objective values of primal problem and dual problem respectively. If $w^*$ is the optimal solution of primal problem (\ref{primal}) and $\alpha^*$ is the optimal solution of dual problem (\ref{dual}), the primal-dual relation always holds that:
\begin{eqnarray}
\label{primal_dual}
w^* = w(\alpha^*) = \frac{1}{\lambda n} A \alpha^*\,.
\end{eqnarray}
\subsection{Stochastic Dual Coordinate Ascent}
In \cite{shalev2013stochastic}, the authors proposed stochastic dual coordinate ascent (SDCA) to optimize the dual problem (\ref{dual}). The pseudocode of SDCA is presented in Algorithm \ref{sdca}. In iteration $t$, given sample $i$ and other dual variables $\alpha_{j\neq i}$ fixed, we maximize the following subproblem:
\begin{eqnarray}
\label{dual_sub}
\max \limits_{\Delta \alpha_i \in \mathbb{R}} -\frac{1}{n} \phi_i^*(-(\alpha_i^{t} + \Delta \alpha_i )) - \frac{\lambda}{2} \| w^{t} + \frac{1}{\lambda n}\Delta \alpha_i x_i \|^2_2
\end{eqnarray}
$e_i$ denotes coordinate vector of size $n$, where element $i$ is $1$ and other elements are $0$. In their paper, the authors proved that SDCA admits linear convergence rate for smooth loss, which is much faster than stochastic gradient descent (SGD). An accelerated SDCA was also proposed in \cite{shalev2013accelerated}. However, SDCA is not applicable when it is difficult to derive the dual problem, \emph{e.g.} $\phi_i$ are non-convex.
\begin{algorithm}[h]
\caption{SDCA}
\begin{algorithmic}[1]
\STATE Initialize $\alpha^0$ and $w^0 = w(\alpha^0)$;
\FOR{$t=0,1,2,\dots, T-1$}
\STATE Randomly sample $i$ from $\{1,2,...,n\}$;
\STATE Find $\Delta \alpha_i $ to maximize the subproblem (\ref{dual_sub});
\STATE Update dual variable $\alpha$ through: \\
\hspace{1.5cm} $\alpha^{t+1} \leftarrow \alpha^{t} + \Delta \alpha_i e_i $;
\STATE Update primal variable $w$ through:\\
\hspace{1.5cm} $w^{t+1} \leftarrow w^{t} + \frac{1}{\lambda n} \Delta \alpha_i x_i$;
\ENDFOR
\end{algorithmic}
\label{sdca}
\end{algorithm}
\subsection{Dual Free Stochastic Dual Coordinate Ascent}
To address the limitation of SDCA, \cite{shalev2015sdca} proposes Dual Free Stochastic Coordinate Ascent (dfSDCA) which has similar convergence property to SDCA. The pseudocode of dfSDCA is presented in Algorithm \ref{dfsdca}.
Although we keep vector $\alpha \in \mathbb{R}^n$ in the optimization, the derivation of dual problem is not necessary for dfSDCA. According to the update rule of $\alpha$ and $w$ in the algorithm, the primal-dual relation (\ref{primal_dual}) also holds for dfSDCA. The drawback of dfSDCA is that it is space-consuming to store $\alpha$, whose space complexity $O(nd)$. We can reduce it to $O(n)$ if $\nabla \phi_i(w)$ can be written as $\nabla \phi_i(x_i^Tw)x_i$. In \cite{he2015dual}, the authors accelerated dfSDCA by using non-uniform sampling strategy in each iteration and proved that it admits faster convergence.
\begin{algorithm}[h]
\caption{Dual Free SDCA}
\begin{algorithmic}[1]
\STATE Initialize dual variable $\alpha^0 = (\alpha_0^0, ..., \alpha_n^0)$ where $\forall i, \alpha_i^0 \in \mathbb{R}^d$, primal variable $w^0 = w(\alpha^0)$;
\FOR{$t=0,1,2,\dots, T-1$}
\STATE Randomly sample $i$ from $\{1,2,...,n\}$;
\STATE Compute dual residue $\kappa$ through: \\
\hspace{1.5cm}$ \kappa \leftarrow \nabla \phi_i(w^{ t}) + \alpha_i^t$;
\STATE Update dual variable $ \alpha_i$ through:\\
\hspace{1.5cm} $\alpha_i^{t+1} \leftarrow \alpha_i^t - \eta \lambda n \kappa$;
\STATE Update primal variable $w$ through:\\
\hspace{1.5cm} $w^{ t+1} \leftarrow w^{t} - \eta \kappa$;
\ENDFOR
\end{algorithmic}
\label{dfsdca}
\end{algorithm}
\section{Distributed Asynchronous Dual Free Stochastic Dual Coordinate Ascent}
In this section, we propose Distributed Asynchronous Dual Free Stochastic Coordinate Ascent (Dis-dfSDCA) for distributed optimization. Dis-dfSDCA fits for any parameter server framework, where the star-shape network is used. We assume that there are $n$ samples in the dataset, and they are evenly distributed over $K$ workers. In worker $k$, there are $n_k$ samples. It is satisfied that $n = \sum_{k=1}^K n_k $. Different from sequential dfSDCA, we split the update of dual variable and primal variable into different nodes. The pseudocodes of Dis-dfSDCA for server node and worker nodes are presented in Algorithm \ref{alg2} and Algorithm \ref{alg1} respectively.
\subsection{Update Global Variable $w$ on Server}
The up-to-date global variable $w \in \mathbb{R}^d$ is stored and updated on the server. Initially, $w$ is set to be vector zero. At the beginning of each iteration, the server receives gradient message $v_k$ from arbitrary worker $k$ and let $v^t = v_k$. Then it updates the global variable through:
\begin{eqnarray}
w^{s, t+1} = w^{s,t} - \eta v^t
\end{eqnarray}
Finally, it sends the up-to-date global variable back to the worker $k$ for further computation. Asynchronous method is robust to straggler problem because it allows for updating the global variable when receiving from only one worker. However, if the $w$ in the worker is too stale, it may lead the algorithm to diverge. Therefore, we induce two loops in our algorithm. Server broadcasts the latest global variable $w$ to all workers after every $T$ iterations. In this way, we prevent the problem of divergence and keep the advantage of asynchronous communication at the same time.
Algorithm \ref{alg2} summarizes the pseudocode on the server.
\begin{algorithm}[h]
\caption{Dis-dfSDCA (Server) }
\label{alg2}
\begin{algorithmic}
\STATE Initialize $w \in \mathbb{R}^d$, $ \eta$
\FOR{$s= 0, 1,..., S-1 $ }
\FOR{$t = 0, 1,...,T-1$}
\STATE Receive gradient message $v^{s,t} = v_k$ from worker $k$;
\STATE Update global variable $w^{s+1, t+1}$ through:\\
\hspace{1.5cm} $w^{s, t+1} \leftarrow w^{s,t} - \eta v^{s,t }$;
\STATE Send $w^{s, t+1}$ back to worker $k$ ;
\ENDFOR
\STATE $w^{s+1, 0} = w^{s, T}$
\STATE Broadcast the up-to-date global variable $w^{s+1,0} $ to all workers.
\ENDFOR
\end{algorithmic}
\end{algorithm}
In Algorithm \ref{alg2}, we use the update of vanilla dfSDCA in the server. \cite{shalev2016sdca} proposed accelerated dfSDCA by using ``Catalyst" algorithm of \cite{lin2015universal}. It is proved to admit faster convergence rate by a constant factor. Our Algorithm \ref{alg2} can also be extended to the accelerated version easily. In our paper, we only consider the vanilla version and analyze the convergence rate of our algorithm.
\subsection{Update Local Variable $\alpha$ on Worker}
In the distributed optimization, workers are responsible for the gradient computation which is the main workload during the optimization. We take arbitrary worker $k$ as an example.
Dual variable $\alpha_{[k]} \in \mathbb{R}^{n_k}$ is only stored and updated in the worker $k$, each $\alpha_i$ is corresponding to sample $i$. Initially, local variable $\alpha_{[k]}$ is set to be vector zero. After receiving stale global variable $w^{s, d(t)} \in \mathbb{R}^{d}$ from the server, worker $k$ computes the dual residue $\kappa$ and updates local variable $\alpha_i$ and gradient message $v_k$ for $H$ iterations.
Samples $I_t$ are randomly selected in the local dataset, and we set $|I_t|=H$.
\begin{algorithm}[h]
\caption{Dis-dfSDCA (Worker $k$)}
\label{alg1}
\begin{algorithmic}
\STATE Initialize $\alpha_{[k]} \in \mathbb{R}^{d \times n_k }$, $\eta$, $H$
\REPEAT
\STATE Receive global variable $w^{s, d(t)}$ from server;
\STATE Initialize gradient message: $v_k \leftarrow {0}$;
\STATE Randomly select samples $I_t$ from $\{1,\cdots,n_k \}$ where $|I_t|=H$;
\FOR{sample $i$ in $I_t$}
\STATE Compute dual residue $\kappa$ through: \\
\hspace{1.5cm}$ \kappa \leftarrow \nabla \phi_i(w^{s, d(t)}) + \alpha_i$;
\STATE Update local dual variable $ \alpha_i$ through:\\
\hspace{1.5cm} $\alpha_i \leftarrow \alpha_i - \eta \lambda n \kappa $;
\STATE Update gradient message $v_k$ through:\\
\hspace{1.5cm} $ v_k \leftarrow v_k + \kappa$;
\ENDFOR
\STATE Send gradient message $v_k$ to server;
\UNTIL{Termination}
\end{algorithmic}
\end{algorithm}
In each iteration, worker $k$ selects a sample $i$ randomly and computes the dual residue $\kappa$ for coordinate $i$ of the dual variable through:
\begin{eqnarray}
\kappa = \nabla \phi_i(w^{s, d(t)}) + \alpha_i
\end{eqnarray}
Dual residue can also be viewed as the gradient in Stochastic Gradient Descent. When we obtain optimal dual variable $\alpha^*$ and primal variable $w^*$, $\kappa$ should be $0$. Therefore, it is satisfied that $\alpha_i^*=-\nabla \phi_i(w^*)$.
Then worker $k$ updates local dual variable $\alpha_i$ and gradient message $v_k$ separately through:
\begin{eqnarray}
\alpha_i& = &\alpha_i- \eta \lambda n \kappa, \hspace{0.3cm} i \in I_t \\
v_k & =& v_k + \kappa
\end{eqnarray}
Because there is only one $\alpha_i$ in the cluster, it is always up-to-date.
After $H$ iterations, the worker $k$ sends gradient message $v_k$ to the server. From the update rule in our algorithm, it is easy to know that the well-known primal-dual relation in the equation (\ref{primal_dual}) is always satisfied.
The pseudocode of Dis-dfSDCA in worker node $k$ is described in Algorithm \ref{alg1}.
In Algorithm \ref{alg1}, we use vanilla dfSDCA in the worker which samples with uniform distribution. There are also other sampling techniques proposed to accelerate dfSDCA. As per the sampling strategy in \cite{shalev2015sdca,he2015dual,shalev2016sdca,qu2015stochastic}, there are three options: uniform sampling, importance sampling, and adaptive sampling. In importance sampling strategy \cite{shalev2016sdca}, it first computes the fixed probability distribution $p_i$ using smoothness parameter of each function $\phi_i$, then selects samples following this probability. In adaptive sampling strategy \cite{he2015dual}, it computes the adaptive probability distribution $p_i$ using dual residue $\kappa$ for each sample every iteration, then selects samples following this probability. Both of them are proved to admit faster convergence than vanilla dfSDCA with uniform sampling. We only consider the uniform sampling strategy, and analyze its corresponding convergence rate in our paper. However, other sampling techniques are straightforward to be applied to our distributed method.
\section{Convergence Analysis}
In this section, we provide the theoretical convergence analysis of Dis-dfSDCA. For the case of convex losses $\phi_i$, we prove that Dis-dfSDCA admits linear convergence rate.
If losses $\phi_i$ are non-convex, we also prove linear convergence rate as long as the sum-of-non-convex objectives $f$ is convex.
We make the following assumptions for the primal problem (\ref{primal}) for further analysis. All of them are common assumptions in the theoretical analysis for the asynchronous stochastic methods.
\begin{assumption} [Lipschitz Constant]
We assume $\nabla \phi_i$ is Lipschitz continuous, and there is Lipschitz constant $L$ such that $\forall x, y \in \mathbb{R}^d$:
\begin{eqnarray}
\|\nabla \phi_i(x) - \nabla \phi_i(y)\|_2 \leq L \|x-y \|_2
\end{eqnarray}
We can also know that $P$ is $(L+\lambda)$-smooth:
\begin{eqnarray}
\|\nabla P(x) - \nabla P(y)\|_2 \leq (L+\lambda) \|x-y \|_2
\end{eqnarray}
\end{assumption}
\begin{assumption} [Maximum Time Delay]
We assume that the maximum time delay of the global variable in each worker is upper bounded by $\tau$, such that:
\begin{eqnarray}
d(t) \geq t - \tau
\end{eqnarray}
$\tau$ is relevant to the number of workers $K$ in the system. We can also control $\tau$ through inner iteration $T$ in our algorithm.
\label{time_delay}
\end{assumption}
\subsection{Convex Case}
In this section, we assume that the losses $\phi_i$ are convex, and prove that our method admits linear convergence.
\begin{assumption} [Convexity]
We assume losses $\phi_i$ are convex, such that $\forall x , y \in \mathbb{R}^d$:
\begin{eqnarray}
\phi_i(x) \geq \phi_i(y) + \nabla \phi_i(y)^T(x-y)\,.
\end{eqnarray}
\end{assumption}
In our algorithm, dual variables $\alpha_{[1]},..., \alpha_{[K]}$ are stored in local workers. For worker $k$, there is no update of $\alpha_{[k]}$ from $d(t)$ to $t$. Therefore, it is always true that $\alpha_{[k]}^{s,t}= \alpha_{[k]}^{s,d(t)}$.
For brevity, we write $v^{s,t}$, $w^{s,t}$ and $\alpha^{s,t}$ as $v^t$, $w^t$ and $\alpha^t$. According to our algorithm, we know that:
\begin{eqnarray}
v^{t} = \sum\limits_{i\in I_t} \left(\nabla \phi_i(w^{d(t)}) + \alpha_i^{d(t)} \right) =
\sum\limits_{i\in I_t} v_i^t
\end{eqnarray}
where $|I_t|=H$ and $\mathbb{E}[v_i^{t}] = \nabla P(w^{d(t)})$. In our analysis, we also assume that there are no duplicate samples in $I_t$.
\iffalse
Thus, in each iteration, the update of local variable $\alpha_i$ and global variable $w$ can be written as follows:
\begin{eqnarray}
\alpha_i^{t+1} &=& \alpha_i^{t} - \eta \lambda n (\nabla \phi_i(w^{d(t)}) + \alpha_i^{t-1}), i \in I_t \\
w^{t+1} &=& w^{t} - \eta v^t
\end{eqnarray}
\fi
To analyze the convergence rate of our method, we need to prove the following Lemma \ref{lem2} at first.
\begin{lemma}
\label{lem2}
Let $w^*$ be the global solution of $P(w)$, and $\alpha_i^*=-\nabla \phi_i(w^*)$. Following the proof in \cite{shalev2015sdca}, we define $A_t$ and $B_t$ as follows:
\begin{eqnarray}
A_t &= &\mathbb{E} \| \alpha_i^{t} - \alpha_i^* \|^2\\
B_t& = & \mathbb{E}\|w^{t} - w^* \|^2
\end{eqnarray}
According to our algorithm, we can prove that $A_{t+1}$ and $B_{t+1}$ are upper bounded:
\begin{eqnarray}
\label{ineq_a}
\mathbb{E}[A_{t+1} - A_{t} ]
&\leq& - \eta \lambda H \mathbb{E}\|\alpha_i^{t} - \alpha_i^*\|^2 - 2\eta HL \lambda^2 \mathbb{E}\|w^t - w^*\|^2 + 4\eta \lambda H L \left( P(x^t) - P(w^*)\right)\nonumber \\
&& - \eta \lambda(1-\eta \lambda n) \mathbb{E}\|v^t\|^2 + 2 \lambda \tau H L^2 \eta^3 \sum\limits_{j=d(t)}^{t-1}\mathbb{E}\|v^j\|^2
\end{eqnarray}
\begin{eqnarray}
\label{ineq_b}
\mathbb{E} [B_{t+1} - B_{t}]
& \leq & -2\eta \left( P(w^{d(t)}) - P(w^*)\right) + \eta^2 \mathbb{E} \|v^t\|^2
-2\eta \left< w^t - w^{d(t)}, \nabla P(x^{d(t)})\right>
\end{eqnarray}
\end{lemma}
\begin{theorem}
\label{them_convex}
Suppose losses $\phi_i$ are convex and $\nabla \phi_i$ are Lipschitz continuous. Let $w^*$ be the optimal solution to $P(w)$, and $\alpha_i^*=-\nabla \phi_i(w^*)$. Define $C_t = \frac{1}{2\lambda L} A_t+ B_t$. We can prove that as long as:
\begin{eqnarray}
\eta \leq \frac{1}{4HL\tau^2 + \lambda n + 2L }
\end{eqnarray}
the following inequality holds:
\begin{eqnarray}
\label{theom_iq1}
\mathbb{E}[C_T] \leq (1-\eta \lambda H) \mathbb{E}[C_0]
\end{eqnarray}
\end{theorem}
\begin{proof}\footnote{We provide the proof sketch here, please check the supplementary material for details.}
Substituting $A_{t+1}$ and $B_{t+1}$ according to Lemma \ref{lem2}, the following inequality holds that:
\begin{eqnarray}
\mathbb{E}[C_{t+1}] &=& \frac{1}{2\lambda L} A_{t+1} + B_{t+1} \nonumber \\
&\leq & (1-\eta \lambda H) \mathbb{E}[C_t] + 2\tau HL\eta^3 \sum\limits_{j=d(t)}^{t-1} \mathbb{E}\|v^j\|^2 +\left( \frac{\eta^2 \lambda n}{2L} + \eta^2 - \frac{\eta}{2L} \right) \mathbb{E}\|v^t\|^2
\end{eqnarray}
Adding the above inequality from $t=0$ to $t=T-1$, we have:
\begin{eqnarray}
\sum\limits_{t=0}^{T-1} \mathbb{E} [C_{t+1}] & \leq& \sum\limits_{t=0}^{T-1} (1-\eta \lambda H)\mathbb{E}[C_t]+ \left(2H\tau^2 \eta^2 + \frac{\eta^2 \lambda n}{2L} + \eta^2 - \frac{\eta}{2L} \right) \sum\limits_{t=0}^{T-1} \mathbb{E}\|v^t\|^2
\end{eqnarray}
where the inequality follows from Assumption \ref{time_delay} and $\eta L \leq 1$.
If $2H\eta^2\tau^2 + \frac{\eta^2 \lambda n}{2L} + \eta^2 - \frac{\eta}{2L} \leq 0$, such that:
\begin{eqnarray}
\eta \leq \frac{1}{4HL\tau^2 + \lambda n + 2L },
\end{eqnarray}
we have the following inequality:
\begin{eqnarray}
\sum\limits_{t=0}^{T-1} \mathbb{E} [C_{t+1}] &\leq& \sum\limits_{t=0}^{T-1} (1-\eta \lambda H) \mathbb{E}[C_t] \nonumber \\
&\leq & \sum\limits_{t=1}^{T-1} \mathbb{E}[C_t] + (1-\eta \lambda H) C_0
\end{eqnarray}
Because $C_t \geq 0$, then we complete the proof that
$\mathbb{E}[C_T] \leq (1- \eta \lambda H) \mathbb{E} [C_0].$ \QEDB \\
\end{proof}
Because $\nabla P(w)$ is Lipschitz continuous, we know that:
\begin{eqnarray}
P(w^t) - P(w^*) \leq \frac{L+\lambda}{2} \|w^t - w^*\|^2 \leq \frac{L+\lambda}{2} C_t
\end{eqnarray}
\begin{theorem}
\label{them_convex_2}
We consider the outer iteration $s$, and write $C^{t}$ as $C^{s,t}$. According to Algorithm \ref{alg2}, we know $C^{s+1, 0} = C^{s, T}$. Following Theorem \ref{them_convex} and applying (\ref{theom_iq1}) for $S$ iterations, it is satisfied that:
\begin{eqnarray}
\mathbb{E}[C_{S, 0}] \leq (1- \eta \lambda H)^S \mathbb{E}[C_{0, 0}]
\end{eqnarray}
In particular, to achieve $\mathbb{E} [P(w^{S, 0}) - P(w^*)] \leq \varepsilon$, it suffices to set
$\eta = \frac{1}{4HL\tau^2 + \lambda n + 2L }$ and
\begin{eqnarray}
\label{con_s}
S \geq O\left( \left( \frac{L}{\lambda} \left(\tau^2 + \frac{1}{H} \right) + \frac{n}{H} \right) \log \left(\frac{1}{\varepsilon}\right) \right)
\end{eqnarray}
\end{theorem}
From Theorem \ref{them_convex} and \ref{them_convex_2}, we know that our Dis-dfSDCA admits linear convergence if losses $\phi_i$ are convex. According to Theorem \ref{them_convex_2}, we observe that $\tau$ affects the speed of our convergence, if $\tau \rightarrow \infty$, it may lead our algorithm to diverge. Therefore, it is important to keep $\tau$ within a reasonable bound. In our algorithm, $\tau$ is relevant to the number of workers and less than $T$. When we let $H=1$ and $\tau=0$, $S$ is relevant to $O(\frac{L}{\lambda}+n)$. It is compatible with the convergence analysis of sequential dfSDCA in \cite{shalev2015sdca}.
\iffalse
From \cite{shalev2015sdca}, dis-dfSDCA is also an instance of variance reduced SGD method. Right now, we can derive the upper bound of the variance of $v_i^t$.
\begin{corollary}
Let $w^*$ be the optimal solution to problem (\ref{primal}) and $\alpha_i^* = -\nabla \phi_i(x_i^Tw^*)$, as $t$ goes to infinity, $\mathbb{E}\|v_i^t\|^2$ goes to zero:
\begin{eqnarray}
&&\mathbb{E}\| v_i^t - \nabla P(w^{d(t)}) \|^2 = \mathbb{E} \|v_i^t\|^2 - \|\nabla P(w^{d(t)})\|^2 \nonumber \\
&\leq & \mathbb{E} \|\alpha_i^{t}x_i - \alpha_i^* x_i + \alpha_i^* x_i + \nabla \phi_i(x_i^Tw^{d(t)})x_i \|^2 \nonumber \\
&\leq & 2\mathbb{E} \|\alpha_i^{t-1} x_i - \alpha_i^* x_i \|^2 + 2\mathbb{E} \|\alpha_i^{*}x_i+ \nabla \phi_i(x_i^Tw^{t-\tau})x_i \|^2 \nonumber \\
&\leq & 2\mathbb{E} \|\alpha_i^{t-1} x_i - \alpha_i^* x_i \|^2 + 2L^2\mathbb{E} \|w^{t-\tau} - w^* \|^2 \nonumber \\
\end{eqnarray}
\end{corollary}
where the second inequality follows from the Lipschitz continuity of $\nabla \phi_i$. From Theorem \ref{them_convex}, both of $\mathbb{E} \|\alpha_i^{t-1} x_i - \alpha_i^* x_i \|^2$ and $\mathbb{E} \|w^{t-\tau} - w^* \|^2$ go to zero linearly. Therefore, the variance of $v_i^t$ is close to zero finally.
\fi
\subsection{Non-convex Case}
In this section, we assume that the losses $\phi_i$ are non-convex, while the sum-of-non-convex objectives $f$ is convex. We also prove that Dis-dfSDCA admits linear convergence rate for this case. Firstly, we get the following Lemma \ref{non_lem2}.
\begin{lemma}
\label{non_lem2}
Let $w^*$ be optimal solution to $P(w)$, and $\alpha_i^*=-\nabla \phi_i(w^*)$.
Following the definition of $A_t$ and $B_t$ in Lemma \ref{lem2}, we prove that $A_{t+1}$ and $B_{t+1}$ are upper bounded:
\begin{eqnarray}
\label{non_ineq_a}
\mathbb{E}[A_{t+1} - A_{t} ]
&\leq& - \eta \lambda H \mathbb{E}\|\alpha_i^{t}- \alpha_i^*\|^2 + 2\eta \lambda H L^2\mathbb{E}\|w^t - w^*\|^2 \nonumber \\
&&- \eta \lambda(1-\eta \lambda n) \mathbb{E}\|v^t\|^2 + 2\lambda \tau HL^2 \eta^3 \sum\limits_{j=d(t)}^{t-1}\mathbb{E}\|v^j\|^2
\end{eqnarray}
\begin{eqnarray}
\label{non_ineq_b}
\mathbb{E} [B_{t+1} - B_{t}]
& \leq & - \frac{3\eta \lambda H}{4} \mathbb{E}\|w^t - w^* \|^2 + \eta^2 \mathbb{E} \|v^t\|^2
\frac{2H \tau H^2 (L+\lambda)^2 \eta^3}{\lambda} \sum\limits_{j=d(t)}^{t-1}\mathbb{E}\|v^j\|^2
\end{eqnarray}
\end{lemma}
\begin{theorem}
\label{them_nonconvex}
Suppose $f$ is convex and $\nabla \phi_i$ is Lipschitz continuous. Let $w^*$ be the optimal solution to $P(w)$, and $\alpha_i^*=-\nabla \phi_i(w^*)$. Define $C_t = \frac{1}{4 L^2} A_t+ B_t$. We can prove that as long as:
\begin{eqnarray}
\eta \leq \frac{\lambda^2}{2HL\tau^2\lambda^2 + 8 HL\tau^2(L+\lambda)^2 + 4\lambda L^2 + n \lambda^3 }
\end{eqnarray}
the following inequality holds:
\begin{eqnarray}
\label{non_theom_iq1}
\mathbb{E}[C_T] \leq (1-\eta \lambda H) \mathbb{E}[C_0]
\end{eqnarray}
\end{theorem}
\begin{proof}
Substituting $A_{t+1}$ and $B_{t+1}$ according to Lemma \ref{non_lem2}, the following inequality holds that:
\begin{eqnarray}
\mathbb{E}[C_{t+1}] &=& \frac{1}{4 L^2} A_{t+1} + B_{t+1} \nonumber \\
&\leq & (1-\eta \lambda H) \mathbb{E}[C_t] + \left(\frac{\lambda H\tau \eta^3}{2} + \frac{2H\tau (L+\lambda)^2 \eta^3}{\lambda} \right) \sum\limits_{j=d(t)}^{t-1} \mathbb{E}\|v^j\|^2 \nonumber \\
&&+\left( \eta^2 + \frac{n \eta^2 \lambda^2}{4L^2} - \frac{\eta \lambda }{4L^2} \right) \mathbb{E}\|v^t\|^2
\end{eqnarray}
Adding the above inequality from $t=0$ to $t=T-1$, we have:
\begin{eqnarray}
\sum\limits_{t=0}^{T-1} \mathbb{E} [C_{t+1}] & \leq& \sum\limits_{t=0}^{T-1} (1-\eta \lambda H)\mathbb{E}[C_t] \nonumber\\
&&+ \biggl(\eta^2 + \frac{n \eta^2 \lambda^2}{4L^2}
+ \frac{\lambda H\tau^2 \eta^2}{2L} + \frac{2H\tau^2 (L+\lambda)^2 \eta^2}{\lambda L}- \frac{\eta \lambda }{4L^2} \biggr) \sum\limits_{t=0}^{T-1} \mathbb{E}\|v^t\|^2
\end{eqnarray}
where the inequality follows from Assumption \ref{time_delay} and $\eta L \leq 1$.
If $\eta^2 + \frac{n \eta^2 \lambda^2}{4L^2} +
\frac{\lambda H\tau^2 \eta^2}{2L} + \frac{2H\tau^2 (L+\lambda)^2 \eta^2}{\lambda L}- \frac{\eta \lambda }{4L^2} \leq 0$, such that:
\begin{eqnarray}
\eta \leq \frac{\lambda^2}{2HL\tau^2\lambda^2 + 8 HL\tau^2 (L+\lambda)^2 + 4\lambda L^2 + n \lambda^3 }
\end{eqnarray}
we have the following inequality:
\begin{eqnarray}
\sum\limits_{t=0}^{T-1} \mathbb{E} [C_{t+1}] &\leq& \sum\limits_{t=0}^{T-1} (1-\eta \lambda H) \mathbb{E}[C_t] \nonumber \\
&\leq & \sum\limits_{t=1}^{T-1} \mathbb{E}[C_t] + (1-\eta \lambda H) C_0
\end{eqnarray}
Because $C_t \geq 0$, then we complete the proof that
$\mathbb{E}[C_T] \leq (1- \eta \lambda H) \mathbb{E} [C_0]$.\QEDB \\
\end{proof}
\begin{theorem}
\label{them_nonconvex_2}
We consider the outer iteration $s$, and write $C^{t}$ as $C^{s,t}$. According to Algorithm \ref{alg2}, we know $C^{s+1, 0} = C^{s, T}$. Following Theorem \ref{them_nonconvex} and applying (\ref{non_theom_iq1}) for $S$ iterations, it is satisfied that:
\begin{eqnarray}
\mathbb{E}[C_{S, 0}] \leq (1- \eta \lambda H)^S \mathbb{E}[C_{0, 0}]
\end{eqnarray}
In particular, to achieve $\mathbb{E} [P(w^{S, 0}) - P(w^*)] \leq \varepsilon$, it suffices to set
$\eta = \frac{\lambda^2}{2HL\tau^2\lambda^2 + 8 HL\tau^2 (L+\lambda)^2 + 4\lambda L^2 + n \lambda^3 }$ and
\small
\begin{eqnarray}
\label{non_s}
S \geq O\left( \left( \frac{\left( \tau^2+ 1/H \right) L^2}{\lambda^2} + \frac{\tau^2L^3}{\lambda^3} + \frac{n}{H} \right) \log \left(\frac{1}{\varepsilon}\right) \right)
\end{eqnarray}
\end{theorem}
From Theorem \ref{them_nonconvex} and \ref{them_nonconvex_2}, we know that our Dis-dfSDCA admits linear convergence even if losses $\phi_i$ are non-convex, as long as the sum-of-non-convex objectives is convex. Comparing Theorem \ref{them_convex_2} with \ref{them_nonconvex_2}, we can observe that our method needs more iterations to converge to the similar accuracy when $\phi_i$ are non-convex. It is reasonable because non-convex problem is known to be harder to be optimized than convex problem. When we let $H=1$ and $\tau=0$, $S$ is relevant to $O(\frac{L^2}{\lambda^2}+n)$. It is also compatible with the convergence analysis of sequential dfSDCA in \cite{shalev2015sdca}.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.31\textwidth}
\centering
\includegraphics[width=2.2in]{ijcnn_gap_time}
\end{subfigure}
\begin{subfigure}[b]{0.31\textwidth}
\centering
\includegraphics[width=2.2in]{covtype_gap_time}
\end{subfigure}
\begin{subfigure}[b]{0.31\textwidth}
\centering
\includegraphics[width=2.2in]{rcv1_gap_time}
\end{subfigure}
\begin{subfigure}[b]{0.31\textwidth}
\centering
\includegraphics[width=2.2in]{ijcnn_gap_epoch}
\caption{IJCNN1}
\label{ijcnn1_gap_epoch}
\end{subfigure}
\begin{subfigure}[b]{0.31\textwidth}
\centering
\includegraphics[width=2.2in]{covtype_gap_epoch}
\caption{COVTYPE}
\end{subfigure}
\begin{subfigure}[b]{0.31\textwidth}
\centering
\includegraphics[width=2.2in]{rcv1_gap_epoch}
\caption{RCV1}
\end{subfigure}
\iffalse
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=2.45in]{ijcnn_gradient_time}
\caption{IJCNN1}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=2.45in]{covtype_gradient_time}
\caption{COVTYPE}
\end{subfigure}
\begin{subfigure}[b]{0.33\textwidth}
\centering
\includegraphics[width=2.45in]{rcv1_gradient}
\caption{RCV1}
\end{subfigure}
\fi
\caption{Figures (a) - (c) present the convergence of duality gap of compared methods in terms of time. Figures (d) - (f) present the convergence of duality gap of compared methods in terms of epoch number.
We train IJCNN1 dataset with $4$ workers, COVTYPE dataset with $8$ workers and RCV1 dataset with $16$ workers.}
\label{convergence}
\end{figure}
\section{Experiments }
In this section, we conduct two simulated experiments on the distributed system with straggler problem. There are mainly three goals, firstly, we want to verify that our Dis-dfSDCA has linear convergence rate for the convex and smooth problem; secondly, we would like to make sure that our method has better speedup property than other primal-dual methods; thirdly, we would like to show that our method is also fit for non-convex losses.
Our algorithm is implemented using C++, and the point-to-point communication between worker and server is handled by openMPI \cite{gabriel04:_open_mpi}. We use Armadillo library \cite{sanderson2016armadillo} for efficient matrix computation. Experiments are performed on Amazon Web Services, and each node is a t2.medium instance which has two virtual CPUs. In our distributed system, we simulate the straggler problem by forcing one selected worker node to the delaying state for $m$ times as long as the normal computing time of other normal workers with probability $p$. In our experiments, we set $p=0.2$ and $m$ is selected from $[0,10]$ randomly. In practice, all nodes have a tiny possibility of being delayed. The setting in our experiments is to verify that our algorithm is robust to straggler problem, even in the extreme situation.
\subsection{Convex Case}
In our experiment, we optimize quadratic loss with $\ell_2 $ regularization term to solve binary classification problem:
\begin{eqnarray}
\min_{w \in \mathbb{R}^d} \frac{1}{n} \sum\limits_{i=1}^n \frac{1}{2}(x_i^T w - y_i)^2 + \frac{\lambda}{2} \| w\|^2
\label{problem}
\end{eqnarray}
where $\lambda=0.1$. Datasets in our experiments are from LIBSVM \cite{CC01a}.
Table \ref{table_data} shows brief details of each dataset. In this problem, because $\nabla \phi_i(w)$ can be written as $\nabla \phi_i(x_i^Tw)$, we just need to store $\hat \alpha \in \mathbb{R}^n$, and recover $\alpha \in \mathbb{R}^{d \times n} $ through $a_i = x_i \hat \alpha_i$. Therefore the space complexity is $O(n)$.
We compare our method with CoCoA+ \cite{ma2015adding}, which is the state-of-the-art distributed primal-dual optimization framework. We reimplement CoCoA+ framework using C++, and use SDCA as the local solver. Learning rate $\eta$ in our method is selected from $\eta = \{1,0.1,0.001,0.0001\}$.
\begin{table}[h]
\center
\begin{tabular}{c|c|c|c}
\hline
\hline
Dataset & $\#$ of samples & Dimension & Sparsity \\
\hline
IJCNN1 & 49,990 & 22 & 41 \% \\
\hline
COVTYPE & 581,012 & 54 & 22 \% \\
\hline
RCV1 & 677,399 & 47,236 & 0.16\% \\
\hline
\hline
\end{tabular}
\caption{Experimental datasets from LIBSVM. }
\label{table_data}
\end{table}
\subsubsection{Convergence of Duality Gap}
We compare the duality gap convergence of compared methods in terms of time and epoch number respectively, where duality gap is well defined in \cite{shalev2013stochastic}. Experimental results are presented in Figure \ref{convergence}.
We distribute IJCNN1 dataset over $4$ workers. Figures \ref{ijcnn1_gap_epoch} in the first column show the duality gap convergence in terms of time and epoch on IJCNN1 dataset. From the second figure, it is easy to know that Dis-dfSDCA and CoCoA+ have similar convergence rate. Since CoCoA+ has linear convergence if the problem is convex and smooth, it is verified that Dis-dfSDCA has linear convergence rate as well. In the experiment, we evaluate Dis-dfSDCA when we set different amount of local computations, $H=10^2$ and $H=10^3$. Results show that our method is faster than CoCoA+ method in both two cases. The reason is that CoCoA+ is affected by the straggler problem in the distributed system. We also optimize problem (\ref{problem}) with COVTYPE dataset using $8$ workers, and RCV1 dataset using $16$ workers. We can draw the similar conclusion from the results of other two datasets.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[width=3in]{ijcnn1_speedup}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[width=3in]{covtype_speedup}
\end{subfigure}
\begin{subfigure}[b]{0.8\textwidth}
\includegraphics[width=5.5in]{rcv1_workers}
\end{subfigure}
\caption{Time speedup in terms of the number of workers. Row $1$ left: IJCNN1; Row $1$ right: COVTYPE; Row $2$: RCV1. }
\label{workers}
\end{figure}
\subsubsection{Speedup}
In this section, we evaluate the scaling up ability of compared methods. The first row of Figure \ref{workers} presents the speedup of compared methods on IJCNN1 and COVTYPE datasets. Speedup is defined as follows:
\begin{eqnarray}
\text{ Time speedup } = \frac{\text{Running time for serial computation}}{\text{Running time of using } K \text{ workers}}
\end{eqnarray}
Figure in the second row shows the convergence of duality gap on RCV1 on multiple machines. It is obvious that Dis-dfSDCA always converges faster than CoCoA+ when they have the same number of workers. Experimental results verify that Dis-dfSDCA has better speedup property than CoCoA+ when there is straggler problem.
\subsection{Non-convex Case}
In this experiment, we optimize the following convex objective, which is an essential step for principal component analysis in \cite{garber2015fast}:
\begin{eqnarray}
\label{exp_pca}
\min\limits_{w \in \mathbb{R}^d} \frac{1}{n}\sum\limits_{i=1}^n \frac{1}{2}w^T \left((\mu - \lambda) - x_ix_i^T\right)w - b^Tw + \frac{\lambda}{2} \|w\|^2
\end{eqnarray}
We conduct the experiment on synthetic data and generate $n = 500,000$ random vectors $\{x_1, ..., x_{500,000} \}\in \mathbb{R}^{500}$ which are mean subtracted and normalized to have Euclidean norm $1$. $C= \frac{1}{n} \sum_{i=1}^n x_ix_i^T $ denotes covariance matrix, $b\in \mathbb{R}^d$ denotes a random vector and we let $\mu = 100$, $\lambda = 10^{-4}$ in the experiment.
\begin{figure}[t]
\centering
\includegraphics[width=4in]{pca_conv}
\caption{Suboptimum ($P(w) - P(w^*)$) convergence of compared methods in terms of time. $w^*$ denotes the optimal solution to problem (\ref{exp_pca}) , and it is obtained by running Dis-dfSDCA until convergence. }
\label{pca}
\end{figure}
Because each $\phi_i$ is probably non-convex, CoCoA is not able to solve this problem. In this experiment, we compare with Distributed asynchronous SVRG \cite{huo2017asynchronous}.
In Figure \ref{pca}, it is obvious that Dis-dfSDCA runs faster than Distributed SVRG when there are $4$ workers. We can observe the similar phenomenon when there are $8$ workers. This observation is reasonable because Distributed SVRG needs to compute two gradients in each inner iteration and full gradient in each outer iteration. Dis-dfSDCA is faster because it only needs to compute one gradient in each iteration. However, Dis-dfSDCA needs $O(nd)$ space for storing $\alpha$ , because $\nabla \phi_i(w)$ cannot be written as $\nabla \phi_i(x_i^Tw)x_i$ in this problem.
\vspace*{-3pt}
\section{Conclusion}
In this paper, we proposed Distributed Asynchronous Dual Free Coordinate Ascent (Dis-dfSDCA) method for distributed machine learning. We addressed two challenging issues in previous primal-dual distributed optimization methods: firstly, Dis-dfSDCA does not rely on the dual formulation, and can be used to solve the non-convex problem; secondly, Dis-dfSDCA uses asynchronous communication and can be applied on the complicated distributed system where there is straggler problem. We also analyze the convergence rate of Dis-dfSDCA and prove linear convergence even if the loss functions are non-convex, as long as the sum of non-convex objectives is convex. We conduct experiments on the simulated distributed system with straggler problem, and all experimental results consistently verify our theoretical analysis.
\bibliographystyle{apalike}
|
\section{Introduction and Setup}
\label{Section_Introduction}
Ideal magnetohydrodynamics (iMHD) is the electromagnetic system consisting of a perfect fluid plasma coupled to Maxwell's electromagnetic fields. Its ``force-free" limit, to be explained shortly, is believed to be relevant for describing the highly magnetized atmospheres of rapidly rotating neutron stars (aka pulsars) \cite{Pulsar,Goldreich:1969sb}. This ``Force-Free-Electrodynamics" (FFE) may also be applied to analyze the magnetized environment around supermassive spinning black holes \cite{Blandford:1977ds}, which through the Penrose process could explain astrophysical observations of energetic phenomenon such as quasars and highly relativistic jets from active galactic nuclei.
In this paper, we embark on a program to seek iMHD solutions in curved geometries. As a starting strategy, we shall do so by extending the recently discovered FFE solutions in \cite{Brennan:2013kea,Gralla:2014yja} to the iMHD case. We also hope the code in \cite{MHDSystem}, which has aided us in obtaining the solutions reported here, will provide a useful tool for other researchers.
{\bf Setup} \qquad It is possible to encode the dynamics of iMHD in a 4D curved metric $g_{\mu\nu}$ through an action principle. The basic degrees of freedom are 3 scalar fields, which we will denote as
\begin{align}
\label{MHD_3Scalars}
\Phi^\text{I} = \Phi_\text{I} , \qquad\qquad \text{I} \in \left\{ 1,2,3 \right\} .
\end{align}
From the first two scalar fields we may build the Maxwell field strength
\begin{align}
F_{\mu\nu} \equiv \partial_{[\mu} \Phi^1 \partial_{\nu]} \Phi^2 ,
\end{align}
where the anti-symmetrization symbol is defined as $F_{[\mu\nu]} \equiv F_{\mu\nu} - F_{\nu\mu}$. From all three $\{ \Phi^\text{I} \}$ we may then form what we will call the plasma current\footnote{The $\widetilde{\epsilon}$ is the covariant Levi-Civita tensor; specifically, $\widetilde{\epsilon}_{\mu\nu\alpha\beta} = \sqrt{|g|} \epsilon_{\mu\nu\alpha\beta}$, where $\epsilon_{\mu\nu\alpha\beta}$ is the Levi-Civita symbol, which in turn returns the sign of the permutation that brings $\{ 0,1,2,3 \}$ to $\{ \mu, \nu, \alpha, \beta \}$ or is zero otherwise.}
\begin{align}
\label{PlasmaCurrent}
n^\mu
&\equiv \widetilde{\epsilon}^{\mu \alpha\beta\gamma} \nabla_\alpha \Phi^1 \nabla_\beta \Phi^2 \nabla_\gamma \Phi^3 \\
\label{PlasmaCurrent_ContainsFmunu}
&= \frac{1}{2} \widetilde{\epsilon}^{\mu \alpha\beta\gamma} F_{\alpha\beta} \nabla_\gamma \Phi^3 ,
\end{align}
This plasma current (or the mass density current) $n^\mu$ describes the mass flow of both positively and negative charged particles. This is to be distinguished from the electromagnetic current that we will introduce below. The physical feature specific to magnetohydrodynamics is that the plasma current itself is orthogonal to the Maxwell tensor,
\begin{align}
n^\mu F_{\mu\nu} = 0 ;
\end{align}
this is an identity because $\nabla_\alpha \Phi^\text{I} n^\alpha = 0$ for all $\text{I} \in \{ 1,2,3 \}$. At this point, we have the necessary ingredients to write down the iMHD action principle. It reads
\begin{align}
\label{MHD_Lagrangian}
S_\text{iMHD} = S_\text{Plasma} + S_\text{Maxwell} ,
\end{align}
with
\begin{align}
\label{MHD_Lagrangian_Plasma}
S_\text{Plasma} &\equiv -\int \text{d}^4 x \sqrt{|g|} \rho_0\left[\sigma_g n^2/2\right] , \qquad\qquad n^2 \equiv n_\mu n^\mu , \\
\label{MHD_Lagrangian_EM}
S_\text{Maxwell} &\equiv - \frac{1}{2} \int \text{d}^4 x \sqrt{|g|} \varphi^2 ;
\end{align}
where
\begin{align}
\sigma_g &= +1, \qquad\qquad \text{ if } \qquad\qquad \eta_{\mu\nu} = \text{diag}[+1,-1,-1,-1] \qquad \text{(``Mostly minus")} , \\
\sigma_g &= -1, \qquad\qquad \text{ if } \qquad\qquad \eta_{\mu\nu} = \text{diag}[-1,+1,+1,+1] \qquad \text{(``Mostly plus")} .
\end{align}
and\footnote{The plasma action in eq. \eqref{MHD_Lagrangian_Plasma} does not take into account additional conserved charges that might be present; for a field theoretic discussion, see \cite{ParticleFluid}.}$^{,}$\footnote{Throughout this paper we will write our results in a metric sign convention independent manner, in order to make the results here and in the accompanying {\sf Mathematica} code accessible to readers from ``both sides of the aisle". For instance, $\sigma_g n^2 > 0$ if $n^\mu$ is timelike in either sign convention.}
\begin{align}
\varphi^2 \equiv \frac{1}{2} F_{\mu\nu} F^{\mu\nu} .
\end{align}
We also remark that the action principle in eq. \eqref{MHD_Lagrangian} is in fact a special case of the Schubring-Vanchurin class of theories \cite{Schubring:2014ena},
\begin{align}
\label{SchubringVanchurin_Lagrangian}
S_{\text{Schubring-Vanchurin}} \equiv \int\text{d}^4 x \sqrt{|g|} \mathcal{L}\left[ \sigma_g n^2/2, \varphi^2/2 \right]
\end{align}
describing the so-called {\it string fluids} which generalize both perfect particle fluids \cite{ParticleFluid} and pressure-less string fluids \cite{StringFluid}. In Ref. \cite{Schubring:2014iwa, Schubring:2014ena} it was shown that iMHD is an example of the string fluid whose equations of motion can be obtained by varying the action of Eqs. \eqref{MHD_Lagrangian}. Our work here can thus be viewed as the initiation of an effort to explore the solution space of theories described by action \eqref{SchubringVanchurin_Lagrangian}.
{\bf Stress Tensors} \qquad By varying the actions \eqref{MHD_Lagrangian_Plasma} and \eqref{MHD_Lagrangian_EM} with respect to the metric, i.e., $g_{\alpha\beta} \to g_{\alpha\beta} + \delta g_{\alpha\beta}$, the stress-energy tensors $T^{\alpha\beta}$ of the plasma and that of the electromagnetic fields can be read off as the coefficients of $-(\sigma_g/2) \sqrt{|g|} \delta g_{\alpha\beta}$. They are
\begin{align}
\label{StressTensor_Plasma}
T[\text{Plasma}]^{\alpha\beta}
&= \sigma_g g^{\alpha\beta} \left( \rho_0\left[ \sigma_g n^2/2\right] - \rho_0'\left[ \sigma_g n^2/2\right] \sigma_g n^2 \right)
+ \rho_0'\left[ \sigma_g n^2/2\right] n^\alpha n^\beta , \\
\label{StressTensor_EM}
T[\text{EM}]^{\alpha\beta}
&= \sigma_g \left( - F^{\alpha\sigma} F^{\beta}_{\phantom{\beta}\sigma} + \frac{1}{4} g^{\alpha\beta} F^{\sigma\kappa} F_{\sigma\kappa} \right) .
\end{align}
The total iMHD stress tensor is, of course, their sum
\begin{align}
T[\text{Total}]^{\mu\nu} = T[\text{Plasma}]^{\mu\nu} + T[\text{EM}]^{\mu\nu} .
\end{align}
When the plasma current $n^\mu$ is timelike, i.e., $\sigma_g n^2 > 0$, its stress tensor in eq. \eqref{StressTensor_Plasma} takes on the perfect fluid form
\begin{align}
T[\text{Plasma}]^{\alpha\beta} = (\rho + p) U^\alpha U^\beta - p \ \sigma_g g^{\alpha\beta}, \qquad\qquad U^\alpha \equiv n^\alpha/\sqrt{|n^2|} ,
\end{align}
where the energy and pressure densities are
\begin{align}
\rho &= \rho_0\left[|n^2|/2\right], \\
p &= -\rho_0\left[|n^2|/2\right] + \rho_0'\left[|n^2|/2\right] |n^2| .
\end{align}
This is the form of the stress tensor derived in \cite{Schubring:2014ena}. However, for the solutions obtained in this paper $n^\mu$ will be null. In such a case, and assuming
\begin{align}
\lim_{n^2 \to 0} \rho_0'\left[ \sigma_g n^2/2\right] \sigma_g n^2 = 0,
\end{align}
the stress tensor becomes
\begin{align}
\label{StressTensor_Plasma_NullCase}
T[\text{Plasma}]^{\alpha\beta} = \sigma_g \rho_0\left[0\right] g^{\alpha\beta} + \rho_0'\left[0\right] n^\alpha n^\beta .
\end{align}
Furthermore, if we choose a Lagrangian density for the plasma such that its first derivative vanishes at the origin, i.e., $\rho_0 '[0]=0$, then eq. \eqref{StressTensor_Plasma_NullCase} informs us the plasma stress-energy tensor would then take a cosmological-constant-like form:
\begin{align}
\label{StressTensor_Plasma_CCForm}
T[\text{Plasma}]^{\alpha}_{\phantom{\alpha}\beta} = (\sigma_g \rho_0\left[0\right]) \delta^{\alpha}_{\phantom{\alpha}\beta} .
\end{align}
{\bf iMHD equations} \qquad The iMHD action principle in eq. \eqref{MHD_Lagrangian} leads to Maxwell's equations
\begin{align}
\label{EOM1}
\nabla_\sigma \Phi^1 \nabla_\mu F^{\mu\sigma} &= \nabla_\beta \Phi^1 \nabla_\gamma \Phi^3 P^{\beta\gamma}, \\
\label{EOM2}
\nabla_\sigma \Phi^2 \nabla_\mu F^{\mu\sigma} &= \nabla_\beta \Phi^2 \nabla_\gamma \Phi^3 P^{\beta\gamma},
\end{align}
and the plasma equation
\begin{align}
\label{EOM3}
0 &= F_{\beta\gamma} P^{\beta\gamma} ,
\end{align}
with
\begin{align}
\label{Rank2P}
P^{\beta\gamma}
&\equiv - \nabla_\alpha \left( \sigma_g \cdot \rho_0'\left[\sigma_g n^2/2\right]\cdot n_\mu \widetilde{\epsilon}^{\mu \alpha\beta\gamma} \right) \nonumber\\
&= \nabla_\alpha \left( \sigma_g \cdot \rho_0'\left[\sigma_g n^2/2\right] \cdot \nabla^{[\alpha} \Phi_1 \nabla^\beta \Phi_2 \nabla^{\gamma]} \Phi_3 \right) .
\end{align}
The prime indicates a derivative with respect to the argument. We remark in passing that, because the Schubring-Vanchurin class of theories in eq. \eqref{SchubringVanchurin_Lagrangian} are invariant under additive shifts, $\Phi^\text{I} \to \Phi^\text{I} + $ constant, equations \eqref{EOM1}--\eqref{EOM3} can be expressed as the conservation of three Noether currents constructed from the Lagrangian densities in eq. \eqref{MHD_Lagrangian},
\begin{align}
\nabla_\sigma \mathcal{J}^\sigma_{\text{I}} = 0, \qquad\qquad
\mathcal{J}^\sigma_{\text{I}} \equiv \frac{\partial \left( - \rho_0\left[ \sigma_g n^2/2 \right] - \varphi^2/2 \right)}{\partial\left( \nabla_\sigma \Phi^\text{I} \right)},
\qquad\qquad \text{I} \in \{ 1,2,3 \} .
\end{align}
{\bf Force-Free Limit} \qquad By using Maxwell's equations
\begin{align}
\label{EMCurrent}
\nabla_\mu F^{\mu\nu} &= \frac{\partial_\mu \left( \sqrt{|g|} g^{\mu\alpha} g^{\nu\beta} F_{\alpha\beta} \right)}{\sqrt{|g|}} = J^\nu, \\
\nabla_{[\alpha} F_{\beta\gamma]} &= 0 ,
\end{align}
where $J$ is the electromagnetic current\footnote{Within iMHD, where there is no externally prescribed electromagnetic current $J^\nu$, eq. \eqref{EMCurrent} is to be viewed as the definition of $J^\nu$.}, the divergence of the electromagnetic stress tensor in eq. \eqref{StressTensor_EM} is
\begin{align}
\label{StressTensor_EM_div}
\nabla_\mu T[\text{EM}]^{\mu\nu} = \sigma_g \ J_\mu F^{\mu\nu} .
\end{align}
The force-free condition holds whenever the electromagnetic current becomes orthogonal to the Maxwell field strength,
\begin{align}
\label{ForceFreeCondition}
J_\mu F^{\mu\nu} = 0.
\end{align}
Referring to eq. \eqref{StressTensor_EM_div}, we see that the force-free condition of eq. \eqref{ForceFreeCondition} is thus equivalent to the conservation of the electromagnetic stress-energy tensor $\nabla_\mu T$[EM]$^{\mu\nu} = 0$. When electromagnetic fields are not the only matter present in the system, such as the case for iMHD, this is not a trivial requirement, since it is usually the total stress tensor that is conserved $\nabla_\mu T$[Total]$^{\mu\nu} = 0$. On the other hand, even though $\rho_0$ in eq. \eqref{MHD_Lagrangian} is usually set to zero when ``Force-Free Electrodynamics" (FFE) is discussed in the literature -- as we shall proceed to show explicitly, when the force-free condition in eq. \eqref{ForceFreeCondition} is obeyed, this does not necessarily imply there is no other matter present.
In \S \eqref{Section_Solutions}, we will delineate two families of analytic iMHD solutions that, despite the presence of a non-trivial plasma, continues to obey the force-free condition. To this end, we seek a null plasma current so that the plasma stress-energy tensor (and thus that of the electromagnetic fields) becomes separately conserved because $T$[Plasma]$^{\mu\nu}$ takes a cosmological-constant-like form in eq. \eqref{StressTensor_Plasma_CCForm}, due to the choice in eq. \eqref{rho0prime}. In \S \eqref{Section_SelfConsistent} we highlight a special case that allows us to obtain self-consistent solutions to the Einstein-iMHD equations. We then close in \S \eqref{Section_Outlook} with comments on possible future research directions.
\section{Two Analytic iMHD Solutions}
\label{Section_Solutions}
{\bf Spacetime geometry} \qquad We will work with the class of 4D curved spacetimes containing a free function $f$,
\begin{align}
\label{g}
\text{d} s^2 = \sigma_g \left\{ f[u^\pm,r,\theta,\phi] (\text{d} u^\pm)^2 \pm 2 \text{d} u^\pm \text{d} r - g_{\mathfrak{A}\mathfrak{B}} \text{d} \psi^\mathfrak{A} \text{d} \psi^\mathfrak{B} \right\} ,
\end{align}
where\footnote{The $\pm$ in eq. \eqref{g} refers to either the ingoing $u^-$ or outgoing $u^+$ null coordinate; for the rest of the paper, every time there is a choice of signs, the upper one would apply if $u^+$ is being used and the lower one if $u^-$ is employed instead.} the metric on the $2-$sphere of radius $r$ is
\begin{align}
\label{g_2Sphere}
g_{\mathfrak{A}\mathfrak{B}} \text{d} \psi^{\mathfrak{A}} \text{d} \psi^{\mathfrak{B}} = r^2 \left( \text{d}\theta^2 + \sin^2 \theta \text{d}\phi^2 \right), \qquad\qquad
\psi^\mathfrak{A} \equiv (\theta,\phi), \qquad\qquad
\mathfrak{A}, \mathfrak{B} \in \left\{ 2,3 \right\} .
\end{align}
Because $f$ is arbitrary, the metric in eq. \eqref{g} includes the Vaidya-(anti-)de Sitter class of spacetimes describing a dark energy dominated cosmology with null matter accreting onto a central black hole, where
\begin{align}
\label{Vaidya}
f_\text{V(A)dS}[u^\pm,r] = 1 - \frac{\Lambda}{3} r^2 - \frac{2 G_{\rm N} M[u^\pm]}{r} .
\end{align}
This can be reduced to the Schwarzschild black hole, by setting the cosmological constant $\Lambda$ to zero and the mass $M$ to a constant.
{\bf Vanishing of plasma energy density's first derivative} \qquad Notice from eq. \eqref{Rank2P} that the right hand side of equations \eqref{EOM1}--\eqref{EOM3} will vanish if we can arrange for the plasma current to be null ($n^2=0$) and for the first derivative of the plasma energy density to vanish there, namely
\begin{align}
\label{rho0prime}
\rho_0'[0] = 0 .
\end{align}
If we can also find $\Phi^{1,2}$ such that they obey the force-free condition of eq. \eqref{ForceFreeCondition}, we would then have solved the full set of iMHD equations \eqref{EOM1}--\eqref{EOM3}. This scenario will play out for the following two families of iMHD solutions.
{\bf Brannan-Gralla-Jacobson-iMHD solution} \qquad The first set of solutions we have found is a direct generalization of that found by Brennan, Gralla and Jacobson \cite{Brennan:2013kea} (see also \cite{Gralla:2014yja}) for FFE, with $\rho_0=0$ in eq. \eqref{MHD_Lagrangian}, to those of the full iMHD system in equations \eqref{EOM1} through \eqref{EOM3}. Their solution, involving only our $\Phi^{1,2}$, consists of
\begin{align}
\label{BGJ_IofII}
\Phi^1 = \zeta[u^\pm,\theta,\phi], \qquad\qquad \Phi^2 = u^\pm .
\end{align}
In addition, we will assume eq. \eqref{rho0prime} holds, and proceed to make $n^2=0$ -- thereby satisfying all three equations \eqref{EOM1} through \eqref{EOM3} -- by putting
\begin{align}
\label{BGJ_IIofII}
\Phi^3 = Z[u^\pm,\theta,\phi] .
\end{align}
{\bf Axis-symmetric Maxwell tensor} \qquad The second set of solutions involve assuming cylindrical symmetry of $\Phi^1$ but now allow $\Phi^2$ to depend on the altitude angle $\theta$,
\begin{align}
\label{AxiallySymmetric}
\Phi^1 = X[u^\pm,\theta], \qquad\qquad \Phi^2 = Y[u^\pm,\theta], \qquad\qquad \Phi^3 = Z[u^\pm,\theta,\phi] .
\end{align}
We continue to assume eq. \eqref{rho0prime}.
For both solutions in equations \eqref{BGJ_IofII}--\eqref{BGJ_IIofII} and \eqref{AxiallySymmetric},
\begin{align}
F_{\alpha\beta} F^{\alpha\beta} = 0 = \widetilde{\epsilon}^{\mu\nu \alpha\beta} F_{\mu\nu} F_{\alpha\beta} .
\end{align}
In the following subsections, we shall describe the derivation of these solutions. Following that, in \S \eqref{Section_SelfConsistent}, we will find a special case of eq. \eqref{AxiallySymmetric} such that the full Einstein-iMHD system may be solved simultaneously.
\subsection{Brennan-Gralla-Jacobson solution generalized to iMHD}
\label{Section_BGJ}
We start with the solutions in eq. \eqref{BGJ_IofII} and \eqref{BGJ_IIofII}; when $\rho_0 \to 0$ in eq. \eqref{MHD_Lagrangian} and $\Phi^3$ is neglected, the $\Phi^{1,2}$ are the solutions in \cite{Brennan:2013kea}. Suppose for the moment that the $Z$ in eq. \eqref{BGJ_IIofII} depended on $r$ as well,
\begin{align}
\Phi^3 = Z[u^\pm,r,\theta,\phi] .
\end{align}
One would find that the plasma current is
\begin{align}
\label{BGJ_PlasmaCurrent}
n^\mu = \frac{\csc[\theta]}{r^2} \left(
0, \partial_{[\phi} \zeta \partial_{\theta]} Z, -\partial_{\phi} \zeta \partial_r Z, \partial_\theta \zeta \partial_r Z
\right) ,
\end{align}
which indicates it is either null or spacelike
\begin{align}
n^2 = -\sigma_g (\partial_r Z)^2 g^{\mathfrak{A}\mathfrak{B}} \partial_\mathfrak{A} \zeta \partial_\mathfrak{B} \zeta .
\end{align}
To make $n^2=0$, one can choose $Z$ to be $r$-independent; or,
\begin{align}
0
= g^{\mathfrak{A}\mathfrak{B}} \partial_{\mathfrak{A}} \zeta \partial_{\mathfrak{B}} \zeta
= \frac{(\partial_\theta \zeta)^2 + \csc^2[\theta] (\partial_\phi\zeta)^2}{r^2} .
\end{align}
But since this latter choice would be setting to zero the sum of the two squares $(\partial_\theta \zeta)^2$ and $\csc^2[\theta] (\partial_\phi\zeta)^2$, for it to hold for all angles, $\zeta$ must therefore be independent of both angular coordinates -- this would mean the plasma current in eq. \eqref{BGJ_PlasmaCurrent} becomes trivial. Hence, to have a non-zero null plasma current while keeping the Brennan-Gralla-Jacobson solutions in eq. \eqref{BGJ_IofII}, we must let $Z$ be independent of $r$. A direct calculation will show that at this point, out of the 3 iMHD equations, only eq. \eqref{EOM1} is non-trivial:
\begin{align}
0 = \pm \frac{2 \csc ^2[\theta] \rho_0'[0]}{r^5} \left( \partial_{[\phi} \zeta \partial_{\theta]} Z \right)^2 .
\end{align}
We see that this equation is satisfied if $\rho_0'[0]$ is chosen to be zero. ($Z$ cannot depend on spacetime through $u^\pm$ alone, because this would set to zero the plasma current in eq. \eqref{BGJ_PlasmaCurrent}.)
The electromagnetic current $J^\nu$ is non-zero only in its $r$-component, which in turn can be expressed as ($\mp$ times) the covariant Laplacian, with respect to the $2-$sphere metric, acting on $\zeta$:
\begin{align}
J^\nu \equiv \nabla_\mu F^{\mu\nu} = \mp \left( 0,g^{\mathfrak{A}\mathfrak{B}} \nabla_{\mathfrak{A}} \nabla_{\mathfrak{B}} \zeta ,0,0 \right) .
\end{align}
Since $g_{rr} = 0$, observe that $J^\nu$ is null. The stress-energy tensor of the plasma takes the form in eq. \eqref{StressTensor_Plasma_CCForm}; whereas, the only non-zero component of the Maxwell stress-energy tensor \eqref{StressTensor_EM} is
\begin{align}
T[\text{EM}]_{\pm \pm} = g^{\mathfrak{A}\mathfrak{B}} \partial_{\mathfrak{A}} \zeta \partial_{\mathfrak{B}} \zeta .
\end{align}
Consistency checks can be made by ensuring that the FFE condition in eq. \eqref{ForceFreeCondition} is satisfied; and the divergence of the plasma and electromagnetic stress-energy tensors are separately zero ($\nabla_\mu T[\text{Plasma}]^{\mu\nu} = \nabla_\mu T[\text{EM}]^{\mu\nu} = 0$).
\subsection{iMHD Solution With Axially Symmetric Maxwell Tensor}
\label{Section_AxiallySymmetric}
Now, we will allow $\Phi^2 = Y[u^\pm,\theta]$ to pick up a $\theta$ dependence at the cost of assuming cylindrical symmetry for $\Phi^1 = X[u^\pm,\theta]$. This ansatz for $\Phi^{1,2}$ maintains the force-free condition in eq. \eqref{ForceFreeCondition}. Following the previous subsection, if we suppose for the moment that the $Z$ in eq. \eqref{AxiallySymmetric} depended on $r$ as well,
\begin{align}
\Phi^3 = Z[u^\pm,r,\theta,\phi] .
\end{align}
The corresponding plasma current is
\begin{align}
\label{AxiallySymmetric_PlasmaCurrent}
n^\mu = \frac{\csc[\theta]}{r^2} \partial_{[\pm} X \partial_{\theta]} Y
\left( 0, \partial_\phi Z, 0, -\partial_r Z \right) ,
\end{align}
such that one would find that it is either null or spacelike
\begin{align}
n^2 = -\sigma_g \frac{(\partial_r Z)^2}{r^2} \left( \partial_{[\pm} X \partial_{\theta]} Y \right)^2 .
\end{align}
To make $n^2=0$, one can choose $Z$ to be $r$-independent; or,
\begin{align}
0 = \partial_{[\pm} X \partial_{\theta]} Y .
\end{align}
However, this latter choice would render the entire plasma current in eq. \eqref{AxiallySymmetric_PlasmaCurrent} trivial. Hence, to have a non-zero null plasma current while maintaining the form of the $\Phi^{1,2}$ solutions in eq. \eqref{AxiallySymmetric}, we must let $Z$ be independent of $r$. At this point, a direct calculation would show, out of the 3 iMHD equations, only eq. \eqref{EOM1} and \eqref{EOM2} are non-trivial:
\begin{align}
0 &= \pm \frac{2 \csc ^2[\theta] \rho_0'[0]}{r^5} \partial_\theta X \left( \partial_{[\theta} X \partial_{\pm]} Y \right) (\partial_\phi Z)^2 , \\
0 &= \pm \frac{2 \csc ^2[\theta] \rho_0'[0]}{r^5} \partial_\theta Y \left( \partial_{[\theta} X \partial_{\pm]} Y \right) (\partial_\phi Z)^2 .
\end{align}
We see that this pair of equations are satisfied if $\rho_0'[0]$ is chosen to be zero. ($Z$ cannot be $\phi$-independent because that would render the plasma current in eq. \eqref{AxiallySymmetric_PlasmaCurrent} trivial.)
The only non-zero component of the electromagnetic current is
\begin{align}
J^r = \nabla_\mu F^{\mu r}
&= \pm r^{-2} \Big(
\partial_u X \partial_\theta^2 Y - \partial_\theta X \partial_u \partial_\theta Y \\
&\qquad\qquad
- \partial_u Y \left ( \partial_\theta^2 X + \cot[\theta] \partial_\theta X \right )
+ \partial_\theta Y \left ( \partial_u \partial_\theta X + \cot[\theta] \partial_u X \right)
\Big) . \nonumber
\end{align}
Because $g_{rr} = 0$, $J^\nu$ is null. The stress-energy tensor of the plasma takes the form in eq. \eqref{StressTensor_Plasma_CCForm}; whereas, the only non-zero component of the Maxwell stress-energy tensor \eqref{StressTensor_EM} is
\begin{align}
T[\text{EM}]_{\pm \pm} = \left( \frac{\partial_{[\pm} X \partial_{\theta]} Y}{r} \right)^2 .
\end{align}
For consistency checks, we have ensured the divergence of the plasma and electromagnetic stress-energy tensors are separately zero ($\nabla_\mu T[\text{Plasma}]^{\mu\nu} = \nabla_\mu T[\text{EM}]^{\mu\nu} = 0$).
\section{A Self-Consistent Einstein-iMHD Solution}
\label{Section_SelfConsistent}
In this section we wish to present a self-consistent solution to the Einstein-iMHD equations. We start with the axially-symmetric-Maxwell ansatz of eq. \eqref{AxiallySymmetric} and the metric in eq. \eqref{g}. Since the iMHD equations are already satisfied, we shall focus on Einstein's:\footnote{To set conventions, we record here that our Ricci tensor is $R_{\beta\nu} \equiv \partial_{[\mu} \Gamma^\mu_{\phantom{\alpha}\nu]\beta} + \Gamma^\mu_{\phantom{\mu}\sigma [\mu} \Gamma^\sigma_{\phantom{\sigma}\nu]\beta}$, while the Christoffel symbols themselves are $\Gamma^\mu_{\phantom{\mu}\alpha \beta} \equiv (1/2) g^{\mu\sigma} ( \partial_\alpha g_{\beta\sigma} + \partial_\beta g_{\alpha\sigma} - \partial_\sigma g_{\alpha\beta} )$. The Einstein tensor is $G_{\beta\nu} \equiv R_{\beta\nu} - (g_{\beta\nu}/2) \mathcal{R}$, where the Ricci scalar is $\mathcal{R} \equiv g^{\beta\nu} R_{\beta\nu}$.}
\begin{align}
\label{Einstein}
G[g]^\mu_{\phantom{\mu}\nu} - \Lambda_\text{cc} \cdot \sigma_g \delta^\mu_{\phantom{\mu}\nu} = 8\pi G_{\rm N} T[\text{Total}]^\mu_{\phantom{\mu}\nu} .
\end{align}
The $00$ component hands us a first-order-in-$r$ equation for $f$,
\begin{align}
\frac{1-\partial_r(rf)}{r^2} = \Lambda_\text{cc} + 8 \pi G_{\rm N} \rho_0[0],
\end{align}
which may be readily integrated to yield
\begin{align}
f[u^\pm,r,\theta,\phi] = 1 - \frac{\Lambda_\text{cc} + 8\pi G_{\rm N} \rho_0[0]}{3} r^2 + \frac{\chi[u^\pm,\theta,\phi]}{r} .
\end{align}
The $r\theta$ and $r\phi$ components of eq. \eqref{Einstein} then tell us $\chi$ needs, in fact, to be independent of the angular coordinates.
\begin{align}
G[g]^r_{\phantom{r}\theta} &= - \sigma_g \frac{\partial_\theta \chi}{2 r^2} = 8 \pi G_{\rm N} T[\text{Total}]^r_{\phantom{r}\theta} = 0 , \\
G[g]^r_{\phantom{r}\phi} &= - \sigma_g \frac{\partial_\phi \chi}{2 r^2} = 8 \pi G_{\rm N} T[\text{Total}]^r_{\phantom{r}\phi} = 0 .
\end{align}
At this point, we re-define $\chi \equiv -2G_{\rm N} M[u^\pm]$, i.e.,
\begin{align}
\label{SelfConsistent_f}
f[u^\pm,r,\theta,\phi] = 1 - \frac{\Lambda_\text{cc} + 8\pi G_{\rm N} \rho_0[0]}{3} r^2 - \frac{2 G_{\rm N} M[u^\pm]}{r} .
\end{align}
Comparison with eq. \eqref{Vaidya} shows we may identify the ``effective cosmological constant" as
\begin{align}
\label{SelfConsistent_Lambda_eff}
\Lambda = \Lambda_\text{cc} + 8 \pi G_{\rm N} \rho_0[0] .
\end{align}
Finally, by examining the $r \pm$ components of eq. \eqref{Einstein}, we find that Einstein's equations have been reduced to
\begin{align}
\label{VaidyaMass_Eqn}
M'[u^\pm] = \mp 4\pi \left( \partial_{[\pm} X \partial_{\theta]} Y \right)^2 .
\end{align}
Eq. \eqref{VaidyaMass_Eqn} can be integrated to solve for $M$ if and only if $\partial_{[\pm} X \partial_{\theta]} Y$ on the right hand side can be made $\theta$-independent. Let us view $\partial_{[\pm} X \partial_{\theta]} Y$ is the non-trivial component $\mathcal{F}_{\pm \theta} = -\mathcal{F}_{\theta\pm}$ of an antisymmetric tensor in the 2-dimensional space parametrized by $(u^\pm,\theta)$,\footnote{The following argument was inspired by Appendix D of \cite{Gralla:2014yja}, which in turn was based on \cite{Uchida}.} i.e.,
\begin{align}
\mathcal{F}_{\sf AB}
\equiv \partial_{\sf [ A} X \partial_{\sf B ]} Y, \qquad\qquad
\partial_{\sf A} \equiv (\partial_\pm,\partial_\theta) .
\end{align}
(This $\mathcal{F}_{\sf AB}$ is of course the 4D Maxwell tensor restricted to the 2D $(u^\pm,\theta)$--plane.) Furthermore, let $\xi^{\sf A} \partial_{\sf A} = \partial_\theta$. Then the requirement that $\mathcal{F}_{\pm \theta}$ be $\theta$-independent can be phrased in the following covariant form:
\begin{align}
\label{d(Fxi)IsZero}
\partial_{\sf [A} \left( \mathcal{F}_{\sf B]C} \xi^{\sf C} \right) = 0 .
\end{align}
To see this we note that the only non-trivial components are ${\sf AB} = \pm \theta$ and ${\sf AB} = \theta \pm$; moreover, since $\mathcal{F}_{\theta\theta} = 0$, by setting say ${\sf AB} = \theta \pm$,
\begin{align}
\partial_{\theta} \mathcal{F}_{\pm \theta} = 0 .
\end{align}
In form notation, eq. \eqref{d(Fxi)IsZero} is denoted as
\begin{align}
\text{d} \left( \mathcal{F} \cdot \xi \right) = 0 ;
\end{align}
this in turn implies $\mathcal{F} \cdot \xi$ is a gradient of some function $H[u^\pm,\theta]$:
\begin{align}
\mathcal{F}_{\sf BC} \xi^{\sf C} = \mathcal{F}_{{\sf B}\theta} = \partial_{\sf B} H .
\end{align}
When ${\sf B}=\theta$, this equation tells us $H$ is actually independent of $\theta$: $0 = \partial_\theta H$. When ${\sf B}=\pm$,
\begin{align}
\label{SelfConsistent_dXdY}
\partial_{\pm} X \partial_\theta Y - \partial_{\pm} Y \partial_\theta X = H'[u^\pm] .
\end{align}
If we proceed to perform a change-of-variables $\partial_\pm = H' \partial_H$,
\begin{align}
\label{SelfConsistent_XYEqn}
\partial_{[H} \left( X\left[ H,\theta \right] \partial_{\theta]} Y\left[H,\theta\right] \right) = 1.
\end{align}
Again, if we view the left hand side as the non-trivial component of the 2D antisymmetric tensor
\begin{align}
\mathcal{F}_{{\sf A}'{\sf B}'} \equiv \partial_{\sf [ A} X \partial_{\sf B ]} Y, \qquad\qquad \partial_{\sf A} = (\partial_H,\partial_\theta) ,
\end{align}
the $X \text{d} Y$ can be viewed as the corresponding gauge potential. Eq. \eqref{SelfConsistent_XYEqn} translates to
\begin{align}
\text{d} \left( X \text{d} Y \right) = \text{d} \left( H \text{d} \theta \right),
\end{align}
which implies the gauge potential is
\begin{align}
\label{SelfConsistent_GaugePotential}
X \text{d} Y &= H \text{d} \theta + \text{d} K ,
\end{align}
where $K$ is some arbitrary function. Referring to the iMHD equations \eqref{EOM1}--\eqref{EOM3} (and \eqref{Rank2P}), as well as the definition of the plasma current in eq. \eqref{PlasmaCurrent}, let us keep in mind that it is only the components of $F = \text{d} (X \text{d} Y)$ that are relevant, not the individual $X$ and $Y$ themselves. Since $\text{d}^2 K = 0$ anyway, we ``choose a gauge" and set $K=$ constant. Following that, we proceed to expand the left hand side of eq. \eqref{SelfConsistent_GaugePotential} in the basis forms $\text{d} H$ and $\text{d} \theta$,
\begin{align}
\label{SelfConsistent_GaugePotential_BasisForms}
X (\partial_H Y \text{d} H + \partial_\theta Y \text{d} \theta) &= H \text{d} \theta .
\end{align}
Equating the coefficient of $\text{d} H$ on both sides of eq. \eqref{SelfConsistent_GaugePotential_BasisForms} then informs us that $Y$ itself must be $H$-independent:
\begin{align}
\label{SelfConsistent_YSoln}
\Phi^2 =Y[\theta] .
\end{align}
Equating the coefficient of $\text{d} \theta$ on both sides of eq. \eqref{SelfConsistent_GaugePotential_BasisForms}, and taking eq. \eqref{SelfConsistent_YSoln} into account, directs us to
\begin{align}
\label{SelfConsistent_XSoln}
\Phi^1 = X[u^\pm] = \frac{H[u^\pm]}{Y'[\theta]} .
\end{align}
Our arguments have determined the form of $X$ and $Y$ that would give us the most general $\theta$-independent expression for $\partial_{[\pm} X \partial_{\theta]} Y$. From equations \eqref{SelfConsistent_YSoln} and \eqref{SelfConsistent_XSoln}, we may check explicitly that eq. \eqref{SelfConsistent_dXdY} is recovered; whereas, the third scalar field remains as
\begin{align}
\label{SelfConsistent_ZSoln}
\Phi^3 = Z[u^\pm,\theta,\phi] .
\end{align}
We have therefore determined that
\begin{align}
\label{VaidyaMass_Eqn_HPrime}
M'[u^\pm] = \mp 4\pi H'[u^\pm]^2 .
\end{align}
To sum: the scalar fields in eq. \eqref{SelfConsistent_YSoln}--\eqref{SelfConsistent_ZSoln} not only satisfy the iMHD equations \eqref{EOM1}--\eqref{EOM3} (as well as the force-free condition eq. \eqref{ForceFreeCondition}); they also satisfy Einstein's equations in eq. \eqref{Einstein}, sourcing the metric in eq. \eqref{g} with the particular Vaidya-(anti-)de Sitter $f$ in eq. \eqref{SelfConsistent_f} -- provided eq. \eqref{VaidyaMass_Eqn_HPrime} holds for the ``mass function" $M$. Notice the effective cosmological constant in eq. \eqref{SelfConsistent_Lambda_eff} receives contributions only from the plasma; whereas the mass $M$, through eq. \eqref{VaidyaMass_Eqn_HPrime}, does so only from electromagnetism.
To be sure our solution here is not trivial, we record here the various physical tensors of the setup. The only non-zero component of the plasma current is
\begin{align}
n^r = \frac{\csc[\theta] H'[u^\pm] \partial_\phi Z}{r^2} .
\end{align}
The only non-zero component of the electromagnetic current is
\begin{align}
J^r = \pm \frac{\cot[\theta] H'[u^\pm]}{r^2} .
\end{align}
The only non-zero component of the Maxwell tensor is
\begin{align}
F_{\pm \theta} = -F_{\theta \pm} = H'[u^\pm] .
\end{align}
Next, the plasma stress-tensor is
\begin{align}
\label{SelfConsistent_Plasma_StressTensor}
T[\text{Plasma}]_{\mu\nu} = (\sigma_g \rho_0[0]) \cdot g_{\mu\nu} ;
\end{align}
while the only non-zero component of the electromagnetic one is
\begin{align}
\label{SelfConsistent_EM_StressTensor}
T[\text{EM}]_{\pm \pm} = \left( \frac{H'[u^\pm]}{r} \right)^2 .
\end{align}
\section{Summary and Outlook}
\label{Section_Outlook}
We have found two families of iMHD solutions, one in eq. \eqref{BGJ_IofII}--\eqref{BGJ_IIofII} and another in eq. \eqref{AxiallySymmetric}, in the curved background geometry of eq. \eqref{g}. Additionally, we discovered that the scalar fields in equations \eqref{SelfConsistent_YSoln}, \eqref{SelfConsistent_XSoln}, and \eqref{SelfConsistent_ZSoln} together with the Vaidya-(anti-)de Sitter metric for which $f$ in eq. \eqref{SelfConsistent_f} is employed in eq. \eqref{g}, simultaneously solve the equations of iMHD \eqref{EOM1}--\eqref{EOM3} and of Einstein's \eqref{Einstein}, provided the mass function is subject to eq. \eqref{VaidyaMass_Eqn}. Because all these solutions assume eq. \eqref{rho0prime}, they still maintain the force-free condition of eq. \eqref{ForceFreeCondition} despite the presence of a plasma.
We close with thoughts on possible next steps to take. For astrophysical applications, it would be important to seek iMHD solutions in axially symmetric geometries or Kerr black hole backgrounds, by perhaps once again extending the known FFE solutions. We also have not studied the stability of the iMHD solutions in this paper. Finally, we believe it is of physical interest to move away from the ideal MHD limit, and include effects from dissipation \cite{Schubring:2014iwa}.
\section{Acknowledgments}
\label{Section_Acknowledgments}
We thank Daniel Schubring for discussions. Much of the analytic work in this paper was done with {\sf Mathematica} \cite{Mathematica}.
|
\section{Introduction}
Robust predictions via computer simulation necessitate accounting for the prevailing uncertainties in the parameters of the computational model. Uncertainty quantification provides the mathematically rigorous framework for propagating the uncertainties surrounding the model input to a response quantity of interest. It comprises three fundamental steps \citep{Sudret2007, Derocquigny2012}: First, the model representing the physical system under consideration is defined; the model maps a given set of input parameters to a unique value of the response quantity of interest. The second step involves the probabilistic description of the input parameters by incorporating available data, expert judgment or a combination of both. In the third step, the uncertainty in the input is propagated upon the response quantity of interest through repeated evaluations of the model for appropriate combinations of the input parameters. In cases when the uncertainty in the response proves excessive or when it is of interest to reduce the dimensionality of the model, sensitivity analysis may be employed to rank the input parameters with respect to their significance for the response variability. Accordingly, important parameters may be described in further detail, whereas unimportant ones may be fixed to nominal values.
Methods of sensitivity analysis can be generally classified into local and global methods. Local methods are limited to examining effects of variations of the input parameters in the vicinity of nominal values. Global methods provide more complete information by accounting for variations of the input parameters in their entire domain. Under the simplifying assumption of linear or nearly linear behavior of the model, global sensitivity measures can be computed by fitting a linear-regression model to a set of input samples and the respective responses (see \eg \citet{Iooss2014, Wei2015} for definitions of such measures). The same methods can be employed in cases with models that behave nonlinearly but monotonically, after applying a rank transformation on the available data. Variance-based methods represent a more powerful and versatile approach, also applicable to nonlinear and non-monotonic models. These methods, known as functional ANOVA (denoting ANalysis Of VAriance) techniques in statistics, rely upon the decomposition of the response variance as a sum of contributions from each input parameter or their combinations \citep{Efron:Stein:1981}. The Sobol' indices, originally proposed in \citet{Sobol1993}, constitute the most popular tool thereof. Although these indices have proven powerful in a wide range of applications, their definition is ambiguous in cases with dependent input variables, which has led to different extensions of the original framework \citep{DaVeiga2009, LiRabitz2010, Kucherenko2012, Mara2012, Zhang2015}. An alternative perspective is offered by the distribution-based indices, which are well-defined regardless of the dependence structure of the input \citep{Borgonovo2007, Liu2010, Borgonovo2014, Zhou2015, Greegar2016}. The key idea is to use an appropriate distance measure to evaluate the effect of suppressing the uncertainty in selected variables on the Probability Density Function (PDF) or the Cumulative Distribution Function (CDF) of the response. These indices are especially useful when consideration of the variance only is not deemed sufficient to characterize the response uncertainty. However, contrary to the Sobol' indices, they do not sum up to unity, which may hinder interpretation. For further information on global sensitivity analysis methods, the interested reader is referred to the review papers \citep{Saltelli2008, Iooss2014, Wei2015, Borgonovo2016}.
It should be mentioned that different classifications of sensitivity analysis techniques can be found in the literature. In cases when one needs to perform a fast exploration of the model behavior with respect to a possibly large number of uncertain input parameters, the so-called screening methods may be employed. These methods can provide a preliminary ranking of the importance of the various input parameters at low computational cost before more precise and costly methods are applied. The Cotter method \citep{Cotter1979} and the Morris method \citep{Morris1991} are widely used screening methods, with the latter covering the input space in a more exhaustive manner. The more recently proposed derivative-based global sensitivity measures can also be classified into this category, while they also provide upper bounds for the Sobol' indices \citep{Sobol2009, Lamboni2013, Sudret2015a}.
The focus of the present paper is on sensitivity analysis by means of Sobol' indices. We limit our attention to cases with independent input and address the computation of these indices for high-dimensional expensive-to-evaluate models, which are increasingly used across engineering and sciences. Various methods have been investigated for computing the Sobol' indices based on Monte Carlo simulation \citep{Archer1997, Sobol2001, Saltelli2002, Sobol2005, Saltelli2010}; because of the large number of model evaluations required, these methods are not affordable for computationally costly models. To overcome this limitation, more efficient estimators have recently been proposed \citep{Sobol2007, Janon2013, Kucherenko2015, Owen2013}. A different approach is to substitute a complex model by a \emph{meta-model}, which has similar statistical properties while maintaining a simple functional form (see \eg \citet{Oakley2004, Marrel2009, Storlie2009, Zuniga2013, Zhang2014, SudretHandbookUQ} for global sensitivity analysis with various types of meta-models). \citet{Sudret2008c} proposed to compute the Sobol' indices by post-processing the coefficients of Polynomial Chaos Expansion (PCE) meta-models. The key concept in PCE is to expand the model response onto a basis made of orthogonal multivariate polynomials in the input variables. The computational cost of the associated Sobol' indices essentially reduces to the cost of estimating the PCE coefficients, which can be curtailed by using sparse PCE \citep{Blatman2010c}. The PCE-based approach for computing the Sobol' indices is employed by a growing number of researchers in various fields including hydrogeology \citep{Fajraoui2011, Formaggia2013, Deman2015}, geotechnics \citep{Albittar2012}, ocean engineering \citep{Alexanderian2012}, biomedical engineering \citep{Huberts2014}, hybrid dynamic simulation \citep{Abbiati2015} and electromagnetism \citep{Kersaudy2014, Yuzugullu2015}. Unfortunately, the PCE approach faces the so-called \emph{curse of dimensionality}, meaning the exploding size of the candidate basis with increasing dimension.
The goal of this paper is to derive a novel approach for solving global sensitivity analysis problems in high dimensions. To this end, we make use of a recently emerged technique for building meta-models with polynomial functions based on Low-Rank Approximations (LRA) \citep{Nouy2010b, Chevreuil2013, Doostan2013, Hadigol2014partitioned, Validi2014low, Konakli2015UNCECOMP}. LRA express the model response as a sum of a small number or rank-one tensors, which are products of univariate functions in each of the input parameters. Because the number of unknown coefficients in LRA grows only linearly with the input dimension, this technique is particularly promising for dealing with cases of high dimensionality. We herein derive analytical expressions for the Sobol' sensitivity indices based on the general functional form of LRA with polynomial bases. As in the case of PCE, the computational cost of the LRA-based Sobol' indices reduces to the cost of estimating the coefficients of the meta-model. Once a LRA meta-model of the response quantity of interest is available, the Sobol' indices can be calculated with elementary operations at nearly zero additional computational cost.
The paper is organized as follows: In Section~2, we review the basic concepts of Sobol' sensitivity analysis and define the corresponding sensitivity indices. In Section~3, we describe the mathematical setup of non-intrusive meta-modeling and define error measures that characterize the meta-model accuracy. After reviewing the computation of Sobol' indices using PCE meta-models in Section~4, we introduce the LRA-based approach in Section~5. In this, we first detail a greedy algorithm for the construction of LRA in a non-intrusive manner and then, use their general functional form to derive analytical expressions for the Sobol' indices. In Section~6, we demonstrate the accuracy of the proposed method by comparing the LRA-based indices to reference ones, with the latter representing the exact solution or Monte-Carlo estimates relying on a large sample of responses of the actual model. Furthermore, we examine the comparative performance of the LRA- and PCE-based approaches in example applications that involve analytical rank-one functions and finite-element models pertinent to structural mechanics and heat conduction. The main findings are summarized in Section 7.
\section{Sobol' Sensitivity analysis}
\label{sec:Sobol}
We consider a computational model $\cm$ describing the behavior of a physical or engineered system of interest. Let $\ve X=\{X_1 \enum X_M\}$ denote the $M$-dimensional input random vector of the model with prescribed joint PDF $f_{\ve X}$. Due to the input uncertainties embodied in $\ve X$, the model response becomes random. By limiting our focus to a scalar response quantity $Y$, the computational model represents the map:
\begin{equation}
\label{eq:model}
\ve X \in \cd_{\ve X} \subset \Rr^M\longmapsto Y=\cm (\ve X) \in \Rr,
\end{equation}
where $\cd_{\ve X}$ denotes the support of $\ve X$.
Sobol' sensitivity analysis aims at apportioning the uncertainty in $Y$, described by its variance, to contributions arising from the uncertainty in individual input variables and their interactions. As explained in the Introduction, the theoretical framework described in the sequel is confined to the case when the components of $\ve X$ are independent. Under this assumption, the joint PDF $f_{\ve X}$ is the product of the marginal PDF $f_{X_i}$ of each input parameter.
\subsection{Sobol' decomposition}
Assuming that $\cm$ is a square-integrable function with respect to the probability measure associated with $f_{\ve X}$, its Sobol' decomposition in summands of increasing dimension is given by \citet{Sobol1993}:
\begin{align}
\label{eq:Sobol_decomp}
\begin{split}
\cm(\ve X)
& = \cm_0+\sum_{i=1}^M\cm_i(X_i)+\sum_{1\leq i<j\leq M}\cm_{i,j}(X_i,X_j)+\ldots+\cm_{1,2\enum M}(\ve X) \\
& = \cm_0+\sum_{\substack{\iu\subset\{1 \enum M\} \\ \iu\neq\emptyset}}\cm_{\iu}(\ve X_{\iu}),
\end{split}
\end{align}
where $\iu=\{i_1 \enum i_s\}$, $1\leq s \leq M$, denotes a subset of $\{1 \enum M\}$ and $\ve X_{\iu}=\{X_{i_1} \enum X_{i_s}\}$ is the subvector of $\ve X$ containing the variables of which the indices comprise $\iu$.
The uniqueness of the decomposition is ensured by choosing summands that satisfy the conditions:
\begin{equation}
\label{eq:M0}
\cm_0 = \Esp{\cm(\ve X)}
\end{equation}
and
\begin{equation}
\label{eq:Sobol_orth}
\Esp{\cm_{\iu}(\ve X_{\iu})~\cm_{\iv}(\ve X_{\iv})}=0 \hspace{5mm}
\forall~\iu,\iv\subset\{1 \enum M\},~\iu \neq \iv.
\end{equation}
Note that the above condition implies that all summands $\{\cm_{\iu},~\iu\neq\emptyset\}$, in Eq.~(\ref{eq:Sobol_decomp}) have zero mean values. A recursive construction of summands satisfying the above conditions is obtained as:
\begin{align}
\begin{split}
\label{eq:Sobol_recur}
\cm_i(X_i) = & \Esp{\cm(\ve X)|X_i}-\cm_0 \\
\cm_{i,j}(X_i,X_j) = & \Esp{\cm(\ve X)|X_i,X_j}-\cm_i(X_i)-\cm_j(X_j)-\cm_0
\end{split}
\end{align}
and so on. By introducing:
\begin{equation}
\label{eq:Mtild_u}
\widetilde{\cm}_{\iu}(\ve X_{\iu})=\Esp{\cm(\ve X)|\ve X_{\iu}},
\end{equation}
the above recursive relationship is written in the general form:
\begin{equation}
\label{eq:Sobol_recur_2}
\cm_{\iu}(\ve X_{\iu})=\widetilde{\cm}_{\iu}(\ve X_{\iu})-\sum_{\substack{\iv\subset\iu \\ \iv\neq\iu}} \cm_{\iv}(\ve X_{\iv}).
\end{equation}
\subsection{Sobol' sensitivity indices}
\label{sec:Sobol_ind}
The uniqueness and orthogonality properties of the Sobol' decomposition allow the following decomposition of the variance $D$ of $\cm(\ve X)$:
\begin{equation}
\label{eq:D}
D=\Var{\cm(\ve X)} = \sum_{i=1}^M D_i+\sum_{1\leq i<j\leq M}D_{i,j}+ \ldots+D_{1,2\enum M} =
\sum_{\substack{\iu\subset\{1 \enum M\}\\ \iu\neq\emptyset}} D_{\iu},
\end{equation}
where $D_{\iu}$ denotes the partial variance:
\begin{equation}
\label{eq:Du}
D_{\iu}=\Var{\cm_{\iu}(\ve X_{\iu})}=\Esp{\left(\cm_{\iu}(\ve X_{\iu})\right)^2}.
\end{equation}
The Sobol' index $S_{\iu}$, defined as:
\begin{equation}
\label{eq:Su}
S_{\iu}=\frac{D_{\iu}}{D}=\frac{\Var{\cm_{\iu}(\ve X_{\iu})}}{\Var{\cm(\ve X)}},
\end{equation}
represents the fraction of the total variance that is due to the interaction between the components of $\ve X_{\iu}$, \ie $S_{i_1 \enum i_s}$ describes the influence from the interaction between variables $\{X_{i_1} \enum X_{i_s}\}$. By definition, the Sobol' indices satisfy:
\begin{equation}
\label{eq:S_u_sum}
\sum_{i=1}^M S_i+\sum_{1\leq i<j\leq M}S_{i,j}+ \ldots+S_{1,2\enum M} =
\sum_{\substack{\iu\subset\{1 \enum M\}\\ \iu\neq\emptyset}} S_{\iu} = 1.
\end{equation}
Accordingly, the first-order index for a single variable $X_i$ is defined as:
\begin{equation}
\label{eq:Si}
S_i=\frac{D_i}{D}=\frac{\Var{\cm_i(X_i)}}{\Var{\cm(\ve X)}}
\end{equation}
and describes the influence of $X_i$ considered separately, also called \textit{main effect}. It is noted that the first-order Sobol' indices are equivalent to the first-order indices obtained by the Fourier amplitude sensitivity test {FAST} method \citep{Cukier1978, Saltelli1999}. The total Sobol' index, denoted by $S^T_i$, represents the \textit{total effect} of $X_i$, accounting for its main effect and all interactions with other input variables. It is derived from the sum of all partial indices $S_{\iu}$ that involve the variable $X_i$:
\begin{equation}
\label{eq:SiT}
S^T_i=\sum_{\substack{\iu\subset\{1 \enum M\} \\ i \in \iu}} S_{\iu}.
\end{equation}
First-order and total Sobol' indices are also defined for groups of variables \citep{Sobol1993,Kucherenko2012,Owen2013}. We herein respectively denote by $\Stu$ and $\StuT$ the first-order and total Sobl' indices of the subvector $\ve X_{\iu}$ of $\ve X$. The first-order index $\Stu$ describes the influence of the elements of $\ve X_{\iu}$ \emph{only}, including their main effects and all interactions with other components of $\ve X_{\iu}$:
\begin{equation}
\label{eq:Stu}
\Stu=\sum_{\substack{\iv\subset\iu \\ \iv \neq \emptyset}} S_{\iv}.
\end{equation}
The total index $\StuT$ describes the total effect of the components of $\ve X_{\iu}$, including their main effects and all interactions with other variables in $\ve X$:
\begin{equation}
\label{eq:StuT}
\StuT=\sum_{\substack{\iv\subset\{1 \enum M\}\\ \iv\cap\iu\neq\emptyset}} S_{\iv}.
\end{equation}
The total index $\StuT$ can equivalently be obtained as:
\begin{equation}
\label{eq:StuT_2}
\StuT=1-\widetilde{S}_{\smallsetminus\iu},
\end{equation}
where $\smallsetminus\iu$ denotes the complementary set of $\iu$ so that $\{X_{\iu},X_{\smallsetminus\iu}\}=\ve X$. For $\iu=\{i\}$, Eq.~(\ref{eq:Stu}) and Eq.~(\ref{eq:StuT}) respectively reduce to Eq.~(\ref{eq:Si}) and Eq.~(\ref{eq:SiT}).
To elaborate on the above definitions of the Sobol' indices, let us consider a model with three input variables \ie $\ve X=\{X_1,X_2,X_3\}$. Setting $\iu=\{1,2\}$ for instance, the first-order and total Sobol' indices of $X_{\iu}=\{X_1,X_2\}$ are respectively given by:
\begin{equation}
\label{eq:S12}
\widetilde{S}_{1,2}=S_1+S_2+S_{1,2}
\end{equation}
and
\begin{equation}
\label{eq:S12T}
\widetilde{S}^T_{1,2}=S_1+S_2+S_{1,2}+S_{1,3}+S_{2,3}+S_{1,2,3}=1-S_3.
\end{equation}
Eq.~(\ref{eq:S12}) follows from the specialization of Eq.~(\ref{eq:Stu}) to the considered case. In Eq.~(\ref{eq:S12T}), the first part is consistent with the definition in Eq.~(\ref{eq:StuT}), whereas the second part follows from Eq.~(\ref{eq:StuT_2}).
Note that the first-order index of the subvector $\ve X_{\iu}$ can alternatively be obtained as:
\begin{equation}
\label{eq:Stu_Esp}
\Stu = \frac{\Var{\Esp{\cm(\ve X)|\ve X_{\iu}}}}{\Var{\cm(\ve X)}}=\frac{\Var{\widetilde{\cm}_{\iu}(\ve X_{\iu})}}{\Var{\cm(\ve X)}}.
\end{equation}
The equivalence between the above equation and Eq.~(\ref{eq:Stu}) can be verified easily by considering the recursive relationship in Eq.~(\ref{eq:Sobol_recur_2}) and the orthogonality property in Eq.~(\ref{eq:Sobol_orth}).
For the total index of the subvector $\ve X_{\iu}$, Eq.~(\ref{eq:StuT_2}) in conjunction with Eq.~(\ref{eq:Stu_Esp}) leads to:
\begin{equation}
\label{eq:StuT_Esp}
\StuT =
1-\frac{\Var{\Esp{\cm(\ve X)|\ve X_{\smallsetminus\iu}}}}{\Var{\cm(\ve X)}} =
1-\frac{\Var{\widetilde{\cm}_{\smallsetminus\iu}(\ve X_{\smallsetminus\iu})}}{\Var{\cm(\ve X)}}.
\end{equation}
The Sobol' indices can be computed by estimating the variance of the summands in Eq.~(\ref{eq:Sobol_recur_2}) using sampling-based formulas \citep{Sobol1993, Saltelli2002}. However, this computation often requires a large number of model evaluations (say more than $10^4$ per index) to obtain accurate estimates; it thus becomes cumbersome or even intractable in cases when a single evaluation of the model is time consuming. In Sections \ref{sec:Sobol_PCE} and \ref{sec:Sobol_LRA}, we demonstrate that the Sobol' indices can be evaluated \emph{analytically} in terms of the coefficients of LRA or PCE meta-models. Consequently, once a PCE or LRA representation of $\cm(\ve X)$ is available, the Sobol' indices indices can be obtained at nearly zero additional computational cost. In the case of PCE meta-models, the derivation is based on Eq.~(\ref{eq:Stu}) and Eq.~(\ref{eq:StuT}), following the original idea presented in \citet{Sudret2008c}. In the case of LRA meta-models, analytical expressions for the Sobol' indices are herein derived by relying on Eq.~(\ref{eq:Stu_Esp}) and Eq.~(\ref{eq:StuT_Esp}).
\section{Non-intrusive meta-modeling and error estimation}
\label{sec:err}
A meta-model $\widehat{\cm}$ is an analytical function that mimics the behavior of $\cm$ in Eq.~(\ref{eq:model}). In non-intrusive approaches, which are of interest herein, the meta-model is developed using a set of realizations of the input vector $\ce=\{\ve \chi^{(1)} \enum \ve \chi^{(N)}\}$, called \emph{Experimental Design} (ED), and the corresponding responses of the original model $\cy=\{\cm(\ve \chi^{(1)}) \enum \cm(\ve \chi^{(N)})\}$. Thus, non-intrusive meta-modeling treats the original model as a \textit{black box}.
In the following, we describe measures of accuracy of the meta-model response $\widehat{Y}=\widehat{\cm}(\ve X)$. To this end, we introduce the discrete $L_2$ semi-norm of a function $\ve x~\in~\cd_{\ve X}~\longmapsto~a(\ve x)~\in~\Rr$, given by:
\begin{equation}
\label{eq:norm}
\norme a {\cx}=\left(\frac{1}{n}\sum_{i=1}^n a^2(\ve x_i)\right)^{1/2},
\end{equation}
where $\cx=\{\ve x_1 \enum \ve x_n\}\subset \cd_{\ve X}$ denotes a set of realizations of the input vector.
A good measure of the accuracy of the meta-model is the \emph{generalization error} $Err_G$, which represents the mean-square of the difference $(Y-\widehat Y)$. $Err_G$ can be estimated by:
\begin{equation}
\label{eq:hatErrG}
\widehat{Err}_G =\left\|\cm-\widehat {\cm}\right\|_{\cx_{\mathrm{val}}}^2,
\end{equation}
where $\cx_{\mathrm{val}}=\{{\ve x}_1 \enum {\ve x}_{n_{\mathrm{val}}}\}$ is a sufficiently large set of realizations of the input vector, called \textit{validation set}. The estimate of the relative generalization error, denoted by $\widehat{err}_G$, is obtained by normalizing $\widehat{Err}_G$ with the empirical variance of $\cy_{\mathrm{val}}=\{\cm({\ve x}_1)\enum \cm({\ve x}_{n_{\mathrm{val}}})\}$.
However, meta-models are typically used in cases when only a limited number of model evaluations is affordable. It is thus desirable to obtain an estimate of ${Err}_G$ by relying solely on the ED. One such error estimate is the \emph{empirical error} $\widehat{Err}_E$, given by:
\begin{equation}
\label{eq:ErrE}
\widehat{Err}_E =\left\|\cm-\widehat {\cm}\right\|_{\ce}^2,
\end{equation}
in which the subscript $\ce$ indicates that the semi-norm is evaluated at the points of the ED. The relative empirical error, denoted by $\widehat{err}_E$, is obtained by normalizing $\widehat{Err}_E$ with the empirical variance of $\cy=\{\cm({\ve \chi}^{(1)})\enum \cm({\ve \chi}^{(N)})\}$, the latter representing the set of model responses at the ED. Unfortunately, $\widehat{err}_E$ tends to underestimate the actual generalization error, which might be severe in cases of overfitting. Indeed, it can even decrease down to zero if the obtained meta-model interpolates the data at the ED, while it is not necessarily accurate at other points.
By using the information contained in the ED only, a fair approximation of the generalization error can be obtained with cross-validation techniques. In $k$-fold cross-validation, (i) the ED is randomly partitioned into $k$ sets of approximately equal size, (ii) for $i=1 \enum k$, a meta-model is built considering all but the $i$-th partition and the excluded set is used to evaluate the respective generalization error, (iii) the generalization error of the meta-model built with the full ED is estimated as the average of the errors of the $k$ meta-models obtained in (ii). Leave-one-out cross-validation corresponds to the case when $k=N$, \ie one point of the ED being set apart in turn \citep{Allen1971}.
\section{Sobol' sensitivity analysis using polynomial chaos expansions}
\label{sec:Sobol_PCE}
\subsection{Formulation and construction of polynomial chaos expansions}
\label{sec:PCE_form}
A meta-model of $Y=\cm(\ve X)$ in Eq.~(\ref{eq:model}) belonging to the class of PCE has the form \citep{Xiu2002wiener}:
\begin{equation}
\label{eq:PCE}
Y^{\rm PCE}=\cm^{\rm{PCE}}(\ve X)=\sum_{\ua \in \ca}{y_{\ua}} \Psi_{\ua}(\ve {X}),
\end{equation}
where $\ca$ denotes a set of multi-indices $\ua=(\alpha_1 \enum \alpha_M)$, $\{\Psi_{\ua},~\ua \in \ca\}$ is a set of multivariate polynomials that are orthonormal with respect to $f_{\ve X}$ and $\{y_{\ua},~\ua \in \ca\}$ is the set of polynomial coefficients. The orthonormality condition reads:
\begin{equation}
\label{eq:PCE_orthonorm_multi}
\Esp{\Psi_{\ua} (\ve X) \hspace{0.3mm} \Psi_{\ve \beta} (\ve X)}=\delta_{\ua \ve \beta},
\end{equation}
where $\delta_{\ua \ve \beta}$ is the Kronecker delta, equal to one if $\ua=\ve \beta$ and zero otherwise.
The multivariate polynomials that comprise the PCE basis are obtained by tensorization of appropriate univariate polynomials:
\begin{equation}
\label{eq:multi_pol}
\Psi_{\ua}(\ve X)=\prod_{i=1}^M P^{(i)}_{\alpha_i}(X_i).
\end{equation}
In the above equation, $P^{(i)}_{\alpha_i}(X_i)$ is a polynomial of degree ${\alpha_i}$ in the $i$-th input variable belonging to a family of polynomials that are orthonormal with respect to $f_{X_i}$, thus satisfying:
\begin{equation}
\label{eq:PCE_orthonorm_uni}
\Esp{P_j^{(i)}(X_i) \hspace{0.3mm} P_k^{(i)}(X_i)}=\delta_{jk}.
\end{equation}
For standard distributions, the associated family of orthonormal polynomials is well-known, \eg a standard normal variable is associated with the family of Hermite polynomials, whereas a uniform variable over $[-1,1]$ is associated with the family of Legendre polynomials. A general case can be treated through an isoprobabilistic transformation of the input random vector $\ve X$ to a basic random vector, \eg a vector with independent standard normal components or independent uniform components over $[-1,1]$.
The set of multi-indices $\ca$ in Eq.~(\ref{eq:PCE}) is determined by an appropriate truncation scheme. A typical scheme consists in selecting multivariate polynomials up to a total degree $p^t$, \ie $\ca=\{\ua\in\Nn^M~|~\|\ua\|_1 \leq p^t\}$, with $\|\ua\|_1=\sum_{i=1}^M{\alpha_i}$. The corresponding number of terms in the truncated series is:
\begin{equation}
\label{eq:card}
{\rm{card}} \ca={M+p^t \choose p^t} =\frac{(M+p^t)!}{M!p^t!},
\end{equation}
which grows exponentially with $M$, thus giving rise to the so-called \emph{curse of dimensionality}. Because the terms that include interactions between input variables are usually less significant, \citet{BlatmanPEM2010} proposed to use a hyperbolic truncation scheme, where the set of retained multi-indices is defined as $\ca=\{\ua\in\Nn^M~|~\|\ua\|_q\leq p^t\}$, with the $q$-norm given by:
\begin{equation}
\label{eq:qnorm}
\|\ua\|_q=\left(\sum_{i=1}^M{\alpha_i}^q\right)^{1/q}, \hspace{5mm} 0<q<1.
\end{equation}
According to Eq.~(\ref{eq:qnorm}), the lower the value of $q$, the smaller is the number of interaction terms in the PCE basis. The case $q=1$ corresponds to the common truncation scheme using a maximal total degree $p^t$, whereas the case $q=0$ corresponds to an additive model.
Once the basis has been specified, the set of coefficients $\ve y=\{y_{\ua},~\ua \in \ca\}$ may be computed by minimizing the mean-square error of the approximation over the ED. More efficient solution schemes can be devised by considering the respective regularized problem:
\begin{equation}
\label{eq:PCE_coef_reg}
\ve y= \mathrm{arg} \underset{\ve {\upsilon}\in\Rr^{\rm{card}\ca}}{\mathrm{min}}\left\|\cm-\sum_{\ua \in \ca}\upsilon_{\ua}\Psi_{\ua}\right\|_{\ce}^2+\lambda \cp(\ve \upsilon),
\end{equation}
in which $\cp(\ve \upsilon)$ is an appropriate regularization functional of $\ve \upsilon=\{\upsilon_1 \enum \upsilon_{\rm{card}\ca}\}$. If $\cp(\ve \upsilon)$ is selected as the $L_1$ norm of $\ve \upsilon$, \ie $\cp(\ve \upsilon)=\sum_{i=1}^{\rm{card}\ca} \abs {\upsilon_i}$, insignificant terms may be disregarded from the set of predictors, leading to \emph{sparse} solutions. \citet{BlatmanJCP2011} proposed to use the hybrid Least Angle Regression (LAR) method for building sparse PCE. This method employs the LAR algorithm \citep{Efron2004} to select the best set of predictors and subsequently, estimates the coefficients using Ordinary Least Squares (OLS). Other techniques to derive sparse expansions can be found in \eg \citet{Doostan2011, Yang2013}.
A good measure of the PCE accuracy is the leave-one out error. As mentioned in Section~\ref{sec:err}, this corresponds to the cross-validation error for the case $k=N$. Using algebraic manipulations, the leave-one-out error can be computed based on a \emph{single} PCE that is built with the full ED. Let $h(\ve \chi^{(i)})$ denote the $i$-th diagonal term of matrix $\ve \Psi(\ve \Psi^{\rm{T}}\ve \Psi)^{-1}\ve \Psi^{\rm{T}}$, with $\ve \Psi=\{\ve \Psi_{ij}=\Psi_j(\ve \chi^{(i)}),\hspace{2mm} i=1\enum N;\hspace{1mm} j=1\enum \rm{card}\ca\}$. The leave-one-out error can then be computed as \citep{BlatmanThesis}:
\begin{equation}
\label{eq:errLOO}
\widehat{Err}_{\rm LOO}=\left\|\frac{\cm-\cm^{\rm PCE}}{1-h}\right\|_{\ce}^2.
\end{equation}
The relative leave-one-out error, denoted by $\widehat{err}_{\rm LOO}$, is obtained by normalizing $\widehat{Err}_{\rm LOO}$ with the empirical variance of $\cy=\{\cm(\ve \chi^{(1)}) \enum \cm(\ve \chi^{(N)})\}$, the latter representing the set of model responses at the ED. Because $\widehat{err}_{\rm LOO}$ may be too optimistic, the following corrected estimate may be used instead \citep{Chapelle2002}:
\begin{equation}
\label{eq:errLOO_corr}
\widehat{err}^*_{\rm LOO}=\widehat{err}_{\rm LOO}\left(1-\frac{\rm{card}\ca}{N}\right)^{-1}\left(1+\rm{tr}((\ve\Psi^{\rm{T}}\ve\Psi)^{-1})\right).
\end{equation}
This corrected leave-one-out error is a good compromise between fair error estimation and affordable computational cost.
\subsection{Computation of Sobol' sensitivity indices using polynomial chaos expansions}
\label{sec:PCE_indices}
Let us consider the PCE representation $Y^{\rm PCE}=\cm^{\rm{PCE}}(\ve X)$ in Eq.~(\ref{eq:PCE}). It is straightforward to obtain the Sobol' decomposition of $\cm^{\rm{PCE}}$ in an analytical form by observing that the summands in Eq.~(\ref{eq:Sobol_decomp}) can be written as:
\begin{equation}
\label{eq:summands_PCE}
\cm^{\rm PCE}_{\iu}(\ve X_{\iu}) = \sum_{\ua\in\ca_{\iu}}y_{\ua}\Psi_{\ua}(\ve X), \hspace{5mm}
\ca_{\iu}=\{\ua\in\ca~|~\alpha_k\neq 0 \Leftrightarrow k \in \iu\}
\end{equation}
and
\begin{equation}
\label{eq:M0_PCE}
\cm_0^{\rm PCE}=y_{\ve 0}.
\end{equation}
Because of the orthonormality of the PCE basis, the conditions in Eq.~(\ref{eq:M0}) and Eq.~(\ref{eq:Sobol_orth}) are satisfied, ensuring the uniqueness of the decomposition. Note that Eq.~(\ref{eq:M0}) in conjunction with Eq.~(\ref{eq:M0_PCE}) lead to:
\begin{equation}
\label{eq:mean_PCE}
\Esp{\cm^{\rm PCE}(\ve X)}=y_{\ve 0}.
\end{equation}
Furthermore, the orthonormality of the PCE basis allows to obtain the total variance $D^{\rm PCE}$ and any partial variance $D_{\iu}^{\rm PCE}$ of $\cm^{\rm{PCE}}(\ve X)$ by a mere combination of the squares of the PCE coefficients. These quantities are respectively given by:
\begin{equation}
\label{eq:D_PCE}
D^{\rm PCE}=\Var{\cm^{\rm PCE}(\ve X)}=\sum_{\ua\in\ca \backslash \{\ve 0\}}{y_{\ua}}^2
\end{equation}
and
\begin{equation}
\label{eq:Du_PCE}
D^{\rm PCE}_{\iu}=\Var{\cm_{\iu}^{\rm PCE}(\ve X_{\iu})}=\sum_{\ua\in\ca_{\iu}}{y_{\ua}}^2.
\end{equation}
By utilizing Eq.~(\ref{eq:D_PCE}) and Eq.~(\ref{eq:Du_PCE}), the first-order and total Sobol' indices of a subvector $\ve X_{\iu}$ of $\ve X$ are respectively obtained as:
\begin{equation}
\label{eq:Stu_PCE}
\widetilde{S}_{\iu}^{\rm PCE} = \sum_{\ua\in\widetilde{\ca}_{\iu}}{y_{\ua}}^2/D^{\rm PCE},
\hspace{5mm}
\widetilde{\ca}_{\iu}=\{\ua\in\ca\backslash\{\ve 0\}~|~\alpha_{i}=0~\forall~i\notin\iu\}
\end{equation}
and
\begin{equation}
\label{eq:StuT_PCE}
\widetilde{S}_{\iu}^{T,\hspace{0.3mm} \rm PCE} = \sum_{\ua\in\widetilde{\ca}_{\iu}^T}{y_{\ua}}^2/D^{\rm PCE},
\hspace{5mm}
\widetilde{\ca}_{\iu}^T=\{\ua\in\ca~|~\exists~i\in\iu : \alpha_i>0 \}.
\end{equation}
To elaborate the above equations, let us consider a model with three input variables \ie $\ve X=\{X_1,X_2,X_3\}$ and focus on the case $\iu=\{1,2\}$. Then, the set $\widetilde{A}_{\iu}$ includes the multi-indices of the form $(\alpha_1, 0, 0)$, $(0,\alpha_2, 0)$ or $(\alpha_1, \alpha_2, 0)$, where $\alpha_i$, $i=1,2,3$, denotes a non-zero element. The set $\widetilde{A}_{\iu}^T$ is a superset of $\widetilde{A}_{\iu}$, additionally including the multi-indices of the form $(\alpha_1, 0, \alpha_3)$, $(0,\alpha_2, \alpha_3)$ or $(\alpha_1, \alpha_2, \alpha_3)$. Note that the set ${A}_{\iu}$ in Eq.~(\ref{eq:summands_PCE}) includes only the multi-indices of the form $(\alpha_1, \alpha_2, 0)$.
By specializing Eq.~(\ref{eq:Stu_PCE}) and Eq.~(\ref{eq:StuT_PCE}) to the case $\iu=\{i\}$, the first-order and total Sobol' indices of a single variable $X_i$ are respectively obtained as:
\begin{equation}
\label{eq:Si_PCE}
S_i^{\rm PCE}=\sum_{\ua\in\ca_i}{y_{\ua}}^2/D^{\rm PCE},
\hspace{5mm}
\ca_i=\{\ua\in\ca~|~\alpha_i>0,~\alpha_{j\neq i}=0\}
\end{equation}
and
\begin{equation}
\label{eq:SiT_PCE}
S_i^{T, \hspace{0.3mm} \rm PCE}=\sum_{\ua\in\ca_i^T}{y_{\ua}}^2/D^{\rm PCE},
\hspace{5mm}
\ca_i^T=\{\ua\in\ca~|~\alpha_i>0\}.
\end{equation}
\section{Sobol' sensitivity analysis using low-rank tensor approximations}
\label{sec:Sobol_LRA}
\subsection{Formulation and construction of low-rank tensor approximations}
\label{sec:LRA_form}
Let us consider again the mapping in Eq.~(\ref{eq:model}). A rank-one function of the input vector $\ve X$ has the form:
\begin{equation}
\label{eq:rank1}
w(\ve X)= \prod_{i=1}^M {v^{(i)}(X_i)},
\end{equation}
where $v^{(i)}$ denotes a univariate function of $X_i$. A representation of $\cm$ as a sum of a finite number of rank-one functions constitutes a canonical decomposition with rank equal to the number of rank-one components. Naturally, of interest are representations where the exact response is approximated with sufficient accuracy by using a relatively small number of terms, leading to the name \emph{low-rank approximations}.
A meta-model of $Y=\cm(\ve X)$ belonging to the class of LRA can be written as:
\begin{equation}
\label{eq:LRA}
Y^{\rm LRA} = \cm^{\rm LRA}(\ve X)=\sum_{l=1}^R b_l\left(\prod_{i=1}^M {v_l^{(i)}(X_i)}\right),
\end{equation}
where $R$ is the rank of the approximation, $v_l^{(i)}$ is a univariate function of $X_i$ in the $l$-th rank-one component and $\{b_l,~l=1 \enum R\}$ are scalars that can be viewed as normalizing constants. In order to obtain a representation in terms of polynomial functions, we expand $v_l^{(i)}$ onto a polynomial basis that is orthonormal with respect to $f_{X_i}$, \ie satisfies the condition in Eq.~(\ref{eq:PCE_orthonorm_uni}). This leads to the expression:
\begin{equation}
\label{eq:vli}
v_l^{(i)}(X_i)=\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} P_k^{(i)} (X_i),
\end{equation}
where $P_k^{(i)}$ is the $k$-th degree univariate polynomial in the $i$-th input variable, $p_i$ is the maximum degree of $P_k^{(i)}$ and $z_{k,l}^{(i)}$ is the coefficient of $P_k^{(i)}$ in the $l$-th rank-one term. Appropriate families of univariate polynomials satisfying the orthonormality condition can be determined as discussed in Section \ref{sec:PCE_form}. By substituting Eq.~(\ref{eq:vli}) into Eq.~(\ref{eq:LRA}), we obtain:
\begin{equation}
\label{eq:LRA_pol}
Y^{\rm LRA} = \cm^{\rm LRA}(\ve X)= \sum_{l=1}^R b_l \left(\prod_{i=1}^M\left(\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} P_k^{(i)} (X_i)\right)\right).
\end{equation}
Disregarding the redundant parameterization arising from the normalizing constants, the number of unknowns in the above equation is $R \cdot \sum_{i=1}^M (p_i+1)$. Note that this number grows only linearly with the input dimension $M$. A representation of $Y=\cm (\ve X)$ in the form of LRA drastically reduces the number of unknowns compared to PCE. In order to emphasize this, we consider PCE with the candidate basis determined by the truncation scheme $\ca=\{\ua \in \Nn^M~|~\alpha_i\leq p_i,~i=1 \enum M\}$, so that the expansion relies on the same polynomial functions as those used in Eq.~(\ref{eq:LRA_pol}). For the case when $p_i=p$, $i=1 \enum M$, the resulting number of unknowns is $(p+1)^M$ in PCE versus $R\cdot M\cdot (p+1)$ in LRA. Considering a typical engineering problem with $M=10$ and low-degree polynomials with $p=3$, the aforementioned formulas yield $1,048,576$ unknowns in PCE versus $40 R$ unknowns in LRA; for a low rank, say $R\leq 10$, the number of unknowns in LRA does not exceed a mere $400$. In other words, the LRA meta-model constitutes a compressed representation of a PCE meta-model, with the basis of the latter defined by constraining the maximum polynomial degree in each dimension. As will be seen in the following, the compressed formulation leads to a fundamentally different construction algorithm.
Non-intrusive algorithms proposed in the literature for building LRA in the form of Eq.~(\ref{eq:LRA_pol}) rely on Alternated Least-Squares (ALS) minimization for the computation of the polynomial coefficients \citep{Chevreuil2013, Chevreuil2013least, Doostan2013, RaiThesis}. The ALS approach consists in sequentially solving a least-squares minimization problem along each dimension ${1\enum M}$ separately, while ``freezing'' the coefficients in all remaining dimensions. We herein employ the \emph{greedy} algorithm proposed in \citet{Chevreuil2013} and further investigated in \citet{Konakli2015arxiv}, which involves progressive increase of the rank by successively adding rank-one components.
Let $Y_r^{\rm LRA} = \cm_r^{\rm LRA}(\ve X)$ denote the rank-$r$ approximation of $Y=\cm(\ve X)$:
\begin{equation}
\label{eq:Yr}
Y_r^{\rm LRA} = \cm_r^{\rm LRA}(\ve X)=\sum_{l=1}^r b_l w_l (\ve X),
\end{equation}
with:
\begin{equation}
\label{eq:wl}
w_l(\ve X)=\prod_{i=1}^M {v_l^{(i)}(X_i)}=\prod_{i=1}^M\left(\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} P_k^{(i)} (X_i)\right).
\end{equation}
The employed algorithm comprises a sequence of pairs of a \textit{correction step} and an \textit{updating step}, so that the $r$-th correction step yields the rank-one term $w_r$ and the $r$-th updating step yields the set of coefficients $\{b_1 \enum b_r\}$. These steps are detailed in the sequel.
\textbf{Correction step}: Let ${\mathcal R}_r(\ve X)$ denote the residual after the completion of the $r$-th iteration:
\begin{equation}
\label{eq:res}
{\mathcal R}_r(\ve X)=\cm(\ve X)-\cm^{\rm LRA}_r(\ve X).
\end{equation}
The sequence is initiated by setting $\cm^{\rm LRA}_0(\ve X)=0$ leading to ${\mathcal R}_0(\ve X)=\cm(\ve X)$. Based on the available experimental design, $\ce$, the rank-one tensor $w_r$ is obtained as the solution of:
\begin{equation}
\label{eq:solve_wr}
w_r=\mathrm{arg} \underset{\omega \in \cw}{\hspace{0.3mm}\mathrm{\min}} \left\|{\mathcal R}_{r-1}-\omega\right\|_{\ce}^2,
\end{equation}
where $\cw$ represents the space of rank-one tensors. By employing an ALS scheme, the minimization problem in Eq.~(\ref{eq:solve_wr}) is substituted by a series of smaller minimization problems, each involving the coefficients along one dimension only:
\begin{equation}
\label{eq:solve_zr}
\ve z_r^{(j)}= \mathrm{arg}\underset{\ve \zeta \in\Rr^{p_j}}{\mathrm{min}}\left\|{\mathcal R}_{r-1}-\left(\prod_{i\neq j}v_r^{(i)} \right)\left(\sum_{k=0}^{p_j} \zeta_k \hspace{0.3mm} P_k^{(j)}\right)\right\|_{\ce}^2,
\hspace{5mm} j=1 \enum M.
\end{equation}
The correction step is initiated by assigning arbitrary values to $v_r^{(i)}$, $i=1\enum M$, \eg unity values, and may involve several iterations over the set of dimensions $\{1 \enum M\}$. The stopping criterion proposed in \citet{Konakli2015arxiv} combines the number of iterations, denoted by $I_r$, with the decrease in the relative empirical error in two successive iterations, denoted by $\Delta \widehat{err}_r$, where the relative empirical error is given by:
\begin{equation}
\label{eq:err_r}
\widehat{err}_r=\frac{\left\|{\mathcal R}_{r-1}-w_{r}\right\|_{\ce}^2}{\Varhat{\cy}}.
\end{equation}
In the above equation, $\Varhat{\cy}$ represents the empirical variance of the set comprising the model responses at the ED. Accordingly, the algorithm exits the $r$-th correction step if either $I_r$ reaches a maximum allowable value, denoted by $I_{\rm max}$, or $\Delta \widehat{err}_r$ becomes smaller than a prescribed threshold, denoted by $\Delta\widehat{err}_{\rm min}$. \citet{Konakli2015arxiv} showed that the choice of $I_{\rm max}$ and $\Delta\widehat{err}_{\rm min}$ may have a significant effect on the accuracy of LRA; based on a number of example investigations, they proposed to use $I_{\rm max}=50$ and $\Delta\widehat{err}_{\rm min} \leq 10^{-6}$.
\textbf{Updating step}: After the completion of a correction step, the algorithm moves to an updating step, in which the set of coefficients $\ve b=\{b_1 \ldots b_r\}$ is obtained as the solution of the minimization problem:
\begin{equation}
\label{eq:solve_b}
\ve b =\mathrm{arg} \underset{\ve \beta \in\Rr^{r}}{\mathrm{min}}\left\|\cm-\sum_{l=1}^r \beta_l w_l\right\|_{\ce}^2.
\end{equation}
Note that in each updating step, the size of vector $\ve b$ is increased by one. In the $r$-th updating step, the value of the new element $b_r$ is determined for the first time, whereas the values of the existing elements $\{b_1 \enum b_{r-1}\}$ are recomputed (updated).
Construction of a rank-$R$ approximation in the form of Eq.~(\ref{eq:LRA_pol}) requires repeating pairs of a correction and an updating step for $r=1\enum R$. The algorithm is summarized below.\vspace{2mm}
\textbf{Algorithm 1}: Non-intrusive construction of a rank-$R$ approximation of $\cm$ with polynomial bases:
\begin{enumerate}
\item Set $\cm^{\rm LRA}_0(\ve x)=0$.
\item For $r=1\enum R$, repeat steps (a)-(d):
\begin{enumerate}
\item Initialize: $v_r^{(i)}(x_i)=1$, $i=1 \enum M$; $I_r=0$; $\Delta \widehat{err}_r=\epsilon>\Delta\widehat{err}_{\rm min}$ .
\item While $\Delta \widehat{err}_r>\Delta\widehat{err}_{\rm min}$ and $I_r<I_{\rm max}$, repeat steps i-iv:
\begin{enumerate}
\item Set $I_r \leftarrow I_r+1$.
\item For $i=1 \enum M$, repeat steps A-B (correction step):
\begin{enumerate}
\item Determine ${\ve z}_r^{(i)}=\{z_{1,r}^{(i)} \ldots z_{p_i,r}^{(i)}\}$ (Eq.~(\ref{eq:solve_zr})).
\item Use the current values of ${\ve z}_r^{(i)}=\{z_{1,r}^{(i)} \ldots z_{p_i,r}^{(i)}\}$ to update $v_r^{(i)}$ (Eq.~(\ref{eq:vli})).
\end{enumerate}
\item Use the current functional forms of $v_r^{(i)}$, $i=1 \enum M$, to update $w_r$ (Eq.~(\ref{eq:wl})).
\item Compute $\widehat{err}_r$ (Eq.~(\ref{eq:err_r})) and update $\Delta \widehat{err}_r$.
\end{enumerate}
\item Determine $\ve b=\{b_1 \ldots b_r\}$ (Eq.~(\ref{eq:solve_b}), updating step).
\item Evaluate $\cm^{\rm LRA}_r$ at the points of the ED (Eq.~(\ref{eq:Yr})).
\item Evaluate ${\mathcal R}_r$ at the points of the ED (Eq.~(\ref{eq:res})).
\end{enumerate}
\end{enumerate}
Note that the above algorithm relies on the solution of several small-size minimization problems: each iteration in a correction step involves $M$ minimization problems of size $\{p_i+1, \hspace{0.3mm} i=1 \enum M\}$ (usually $p_i<20$), whereas the $r$-th updating step involves one minimization problem of size $r$ (recall that small ranks are of interest in LRA). Because of the small number of unknowns, Eq.~(\ref{eq:solve_zr}) and Eq.~(\ref{eq:solve_b}) can be efficiently solved with OLS, as shown in \citet{Konakli2015arxiv}. An alternative approach employed in \citet{Chevreuil2013} is to substitute these equations with respective regularized problems.
In a typical application, the optimal rank $R$ is not known \textit{a priori}. As noted earlier, the progressive construction of LRA results in a set of decompositions of increasing rank. Thus, one may set $r=1 \enum r_{\rm max}$ in Step 2 of Algorithm 1, where $r_{\rm max}$ is a maximum allowable candidate rank, and at the end, select the optimal-rank approximation using error-based criteria. In the present work, we select the optimal rank $R\in\{1 \enum r_{\rm max}\}$ based on the 3-fold cross-validation error, as proposed by \citet{Chevreuil2013least} (see Section~\ref{sec:err} for details on $k$-fold cross-validation). \citet{Konakli2015arxiv} investigated the accuracy of the approach in a number of applications and showed that it leads to optimal or nearly optimal rank selection in terms of the relative generalization error estimated with a large validation set.
\subsection{Computation of Sobol' sensitivity indices using low-rank tensor approximations}
\label{sec:LRA_indices}
In this section, we consider the LRA meta-model $Y^{\rm LRA} = \cm^{\rm LRA}(\ve X)$ in Eq.~(\ref{eq:LRA_pol}) and derive analytical expressions for the Sobol' indices in terms of the polynomial coefficients $z_{k,l}^{(i)}$ and the normalizing constants $b_l$. To this end, we rely on the definitions of the first-order and total Sobol' indices for a subset $\ve X_{\iu}$ of $\ve X$ given in Eq.~(\ref{eq:Stu_Esp}) and Eq.~(\ref{eq:StuT_Esp}), respectively. In the specific case $\iu=\{i\}$, these equations yield the first-order and total Sobol' indices of a single variable $X_i$.
Employing the definition of variance and the property:
\begin{equation}
\label{eq:Mtu_mean}
\Esp{\widetilde{\cm}_{\iu}(\ve X_{\iu})}=\Esp{\Esp{\cm(\ve X)|\ve X_{\iu}}}=\Esp{\cm(\ve X)},
\end{equation}
Eq.~(\ref{eq:Stu_Esp}) is equivalently written as:
\begin{equation}
\label{eq:Stu_LRA}
\Stu = \frac{\Esp{\left(\widetilde{\cm}_{\iu}(\ve X_{\iu})\right)^2}-\left(\Esp{\cm (\ve X)}\right)^2}{\Var{\cm(\ve X)}}.
\end{equation}
We next derive analytical expressions for each of the quantities in the right-hand side of Eq.~(\ref{eq:Stu_LRA}) considering $\cm^{\rm LRA}$ in place of $\cm$. In these derivations, we make use of the equalities:
\begin{equation}
\label{eq:mean_vli}
\Esp{v_l^{(i)}(X_i)} = z_{0,l}^{(i)}
\end{equation}
and
\begin{equation}
\label{eq:mean_vli_vlip}
\Esp{v_l^{(i)}(X_i) \hspace{0.3mm} v_{l'}^{(i)}(X_i)}=
\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} z_{k,l'}^{(i)},
\end{equation}
which follow from Eq.~(\ref{eq:vli}) in conjunction with the orthonormality condition in Eq.~(\ref{eq:PCE_orthonorm_uni}).
Employing the expression for the LRA response in Eq.~(\ref{eq:LRA}), its mean value can be computed as:
\begin{align}
\begin{split}
\label{eq:mean_LRA}
\Esp{\cm^{\rm LRA}(\ve X)} = &
\sum_{l=1}^R b_l \left(\prod_{i=1}^M \Esp{{v_l^{(i)}(X_i)}}\right) \\ = &
\sum_{l=1}^R b_l \left(\prod_{i=1}^M z_{0,l}^{(i)} \right).
\end{split}
\end{align}
In the first part of the above equation, we have utilized the property $\Esp{A\cdot B}=\Esp{A}\cdot \Esp{B}$, which holds under the condition that $A$ and $B$ are independent; note that because the components of $\ve X$ are assumed independent, the quantities $\{v_l^{(i)}(X_i),~i=1\enum M\}$ are also independent. The second part of Eq.~(\ref{eq:mean_LRA}) follows directly from Eq.~(\ref{eq:mean_vli}).
The mean-square of the LRA response in Eq.~(\ref{eq:LRA}) can be computed as:
\begin{align}
\begin{split}
\label{eq:ms_LRA}
\Esp{\left(\cm^{\rm LRA}(\ve X)\right)^2} = &
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'} \left( \prod_{i=1}^M \Esp{v_l^{(i)}(X_i) \hspace{0.3mm} v_{l'}^{(i)}(X_i)} \right) \\ = &
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'} \left( \prod_{i=1}^M \left( {\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} z_{k,l'}^{(i)}} \right) \right).
\end{split}
\end{align}
The first part of the above equation relies on the independence of the quantities $\{v_l^{(i)}(X_i) \hspace{0.3mm} v_{l'}^{(i)}(X_i),~i=1\enum M\}$, which allows the use of the property $\Esp{A\cdot B}=\Esp{A}\cdot \Esp{B}$, whereas the second part follows directly from Eq.~(\ref{eq:mean_vli_vlip}).
Substituting Eq.~(\ref{eq:mean_LRA}) and Eq.~(\ref{eq:ms_LRA}) into the definition of the variance, we have:
\begin{align}
\begin{split}
\label{eq:var_LRA}
\Var{\cm^{\rm LRA}(\ve X)} = &
\Esp{\left(\cm^{\rm LRA}(\ve X)\right)^2}-\left(\Esp{\cm^{\rm LRA}(\ve X)}\right)^2 \\ = &
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'} \left( \left( \prod_{i=1}^M \left( {\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} z_{k,l'}^{(i)}} \right) \right)- \left(\prod_{i=1}^M z_{0,l}^{(i)} \hspace{0.3mm} z_{0,l'}^{(i)} \right) \right).
\end{split}
\end{align}
In the following, we derive an analytical expression for $\Esp{\left(\widetilde{\cm}_{\iu}^{\rm LRA}(\ve X_{\iu})\right)^2}$, beginning with the case $\iu=\{i\}$. Noting that $v_l^{(i)}(X_i)$ represents a constant in the quantity $(\cm^{\rm LRA}(\ve X)|X_i)$, substitution of Eq.~(\ref{eq:LRA}) into the definition of $\widetilde{\cm}_i^{\rm LRA}(X_i)$ (Eq.~(\ref{eq:Mtild_u})) yields:
\begin{align}
\begin{split}
\label{eq:Mti_LRA}
\widetilde{\cm}_i^{\rm LRA}(X_i) = &
\sum_{l=1}^R b_l \hspace{0.3mm} v_l^{(i)}(X_i) \left(\prod_{j\neq i}^M \Esp{{v_l^{(j)}(X_j)}}\right) \\ = &
\sum_{l=1}^R b_l \left(\prod_{j\neq i}^M z_{0,l}^{(j)} \right) v_l^{(i)}(X_i).
\end{split}
\end{align}
Using the above expression, we have:
\begin{align}
\begin{split}
\label{eq:ms_Mti_LRA}
\Esp{\left(\widetilde{\cm}_i^{\rm LRA}(X_i)\right)^2} = &
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'} \left(\prod_{j\neq i}^M z_{0,l}^{(j)} \hspace{0.3mm} z_{0,l'}^{(j)} \right) \Esp{v_l^{(i)}(X_i) \hspace{0.3mm} v_{l'}^{(i)}(X_i)} \\ = &
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'} \left(\prod_{j\neq i}^M z_{0,l}^{(j)} \hspace{0.3mm} z_{0,l'}^{(j)} \right) \left(\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} z_{k,l'}^{(i)} \right).
\end{split}
\end{align}
It is straightforward to extend Eq.~(\ref{eq:Mti_LRA}) to $\widetilde{\cm}_{\iu}^{\rm LRA}(\ve X_{\iu})$ considering that the components of $\ve X_{\iu}$ represent constants in the quantity $(\cm^{\rm LRA}(\ve X)|\ve X_{\iu})$. We thereby obtain:
\begin{align}
\begin{split}
\label{eq:Mtu_LRA}
\widetilde{\cm}_{\iu}^{\rm LRA}(\ve X_{\iu}) = &
\sum_{l=1}^R b_l \left(\prod_{i\in \iu} v_l^{(i)}(X_i) \right) \left(\prod_{j\notin \iu} \Esp{v_l^{(j)}(X_j)} \right) \\ = &
\sum_{l=1}^R b_l \left(\prod_{j\notin \iu} z_{0,l}^{(j)} \right) \left(\prod_{i\in \iu} v_l^{(i)}(X_i) \right)
\end{split}
\end{align}
and finally:
\begin{align}
\begin{split}
\label{eq:ms_Mtu_LRA}
\Esp{\left(\widetilde{\cm}_{\iu}^{\rm LRA}(\ve X_{\iu})\right)^2} = &
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'}
\left(\prod_{j\notin \iu} z_{0,l}^{(j)} \hspace{0.3mm} z_{0,l'}^{(j)} \right)
\left(\prod_{i\in \iu} \Esp{v_l^{(i)}(X_i) \hspace{0.3mm} v_{l'}^{(i)}(X_i)} \right) \\ = &
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'}
\left(\prod_{j\notin \iu} z_{0,l}^{(j)} \hspace{0.3mm} z_{0,l'}^{(j)} \right)
\left(\prod_{i\in \iu} \left(\sum_{k=0}^{p_i} z_{k,l}^{(i)} \hspace{0.3mm} z_{k,l'}^{(i)} \right) \right).
\end{split}
\end{align}
Eq.~(\ref{eq:mean_LRA}), Eq.~(\ref{eq:var_LRA}) and Eq.~(\ref{eq:ms_Mtu_LRA}) provide all the elements required to compute the first-order index for any subvector $\ve X_{\iu}$ according to Eq.~(\ref{eq:Stu_LRA}). As explained above, Eq.~(\ref{eq:ms_Mtu_LRA}) simplifies to Eq.~(\ref{eq:ms_Mti_LRA}) for the case of a single variable $X_i$.
Respective analytical expressions for the total Sobol' indices can be derived using Eq.~(\ref{eq:StuT_Esp}), which is equivalently written as:
\begin{equation}
\label{eq:StuT_LRA}
\StuT = 1-\frac{\Esp{\left(\widetilde{\cm}_{\smallsetminus\iu}(\ve X_{\smallsetminus\iu})\right)^2}-\left(\Esp{\cm (\ve X)}\right)^2}{\Var{\cm(\ve X)}}.
\end{equation}
By considering that $\smallsetminus \iu$ is the complementary set of $\iu$ with respect to $\{1 \enum M\}$, we can obtain an expression for $\Esp{\left(\widetilde{\cm}_{\smallsetminus\iu}^{\rm LRA}(\ve X_{\smallsetminus\iu})\right)^2}$ by simply interchanging the indices $i$ and $j$ in Eq.~(\ref{eq:ms_Mtu_LRA}):
\begin{equation}
\label{eq:ms_Mtun_LRA}
\Esp{\left(\widetilde{\cm}_{\smallsetminus\iu}^{\rm LRA}(\ve X_{\smallsetminus\iu})\right)^2} =
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'}
\left(\prod_{i\in \iu} z_{0,l}^{(i)} \hspace{0.3mm} z_{0,l'}^{(i)} \right)
\left(\prod_{j\notin \iu} \left(\sum_{k=0}^{p_j} z_{k,l}^{(j)} \hspace{0.3mm} z_{k,l'}^{(j)} \right) \right).
\end{equation}
For the special case $\iu=\{i\}$, Eq.~(\ref{eq:ms_Mtun_LRA}) reduces to:
\begin{equation}
\label{eq:ms_Mtin_LRA}
\Esp{\left(\widetilde{\cm}_{\smallsetminus i}^{\rm LRA}(X_{\smallsetminus i})\right)^2} =
\sum_{l=1}^R \sum_{l'=1}^R b_l \hspace{0.3mm} b_{l'}
\hspace{0.3mm} z_{0,l}^{(i)} \hspace{0.3mm} z_{0,l'}^{(i)}
\left(\prod_{j\neq i} \left(\sum_{k=0}^{p_j} z_{k,l}^{(j)} \hspace{0.3mm} z_{k,l'}^{(j)} \right) \right).
\end{equation}
Note that by appropriate combinations of first-order indices for single variables and groups of variables, we can obtain any higher-order index representing the effect from the interaction between a set of variables. For instance, the second-order index representing the effect from the interaction between $X_i$ and $X_j$ can be obtained as $S_{i,j}=\widetilde{S}_{i,j}-S_i-S_j$. A general expression can be derived by dividing each part of Eq.~(\ref{eq:Sobol_recur_2}) by $\Var{\cm(\ve X)}$, leading to:
\begin{equation}
\label{eq:Su_LRA}
S_{\iu}=\Stu-\sum_{\substack{\iv\subset\iu \\ \iv\neq\iu}} S_{\iv}.
\end{equation}
\section{Example applications}
\label{sec:examples}
In this section, we perform global sensitivity analysis for the responses of four models with different characteristics and dimensionality. The first two models correspond to analytical functions, namely a common benchmark function in uncertainty quantification, of dimension $M=20$, and a structural-mechanics model of dimension $M=5$; the aforementioned analytical models have a rank-one structure. The subsequent applications involve finite-element models, namely a structural-mechanics model of dimension $M=10$ and a heat-conduction model, with thermal conductivity described by a random field, of dimension $M=53$. The LRA-based Sobol' indices are compared to respective PCE-based indices and reference indices based on the actual model. The latter are computed either analytically or by using a Monte Carlo Simulation (MCS) approach with large samples according to \citet{Janon2013}. The computations of the PCE- and MCS-based indices are performed with the software UQLab \citep{MarelliICVRAM2014, UQdoc_09_106}.
The meta-models are built using two types of EDs, based on Sobol pseudo-random sequences \citep{Niederreiter1992} and the so-called \emph{maximin} Latin Hypercube Sampling (LHS). Each ED of the latter type represents the best among 5 random LHS-based designs, where the selection criterion is the maximum of the minimum distance between the points. The LRA meta-models are built by implementing Algorithm 1 in Section~\ref{sec:LRA_form}. A common polynomial degree $p_1=\ldots=p_M=p$ is considered with its optimal value selected as the one leading to the minimum 3-fold cross-validation error, denoted by $\widehat{err}_{\rm CV3}$ (see \citet{Konakli2015arxiv} for an investigation of the accuracy of the approach). The involved minimization problems are solved using the OLS method. In building the PCE meta-models, a candidate basis is first determined by employing a hyperbolic truncation scheme and then, a sparse expansion is obtained by evaluating the PCE coefficients with the hybrid LAR method. The optimal combination of the maximum total polynomial degree $p^t$ and the parameter $q$ controlling the truncation, with $q\in\{0.25,0.50,0.75,1.0\}$, is selected as the one leading to the minimum corrected leave-one-out error $\widehat{err}_{\rm LOO}^*$ (see Section \ref{sec:PCE_form} for details).
\subsection{Analytical functions}
\subsubsection{Sobol function}
\label{sec:Sob_LRA}
In the first example, we consider the Sobol function:
\begin{equation}
\label{eq:Sobol}
Y = \prod_{i=1}^M\frac{|4X_i-2|+c_i}{1+c_i},
\end{equation}
where $\ve X=\{X_1,\ldots, X_M\}$ are independent random variables uniformly distributed over $[0,1]$ and $\ve c=\{c_1,\ldots, c_M\}$ are non-negative constants. We examine the case with $M=20$ and the constants given by \citet{Kersaudy2015new}:
\begin{equation}
\label{eq:Sobol_c}
\ve c=\{1,2,5,10,20,50,100,500,500,500,500,500,500,500,500,500,500,500,500,500\}^\mathrm{T}.
\end{equation}
This function is a commonly used benchmark for global sensitivity analysis with well-known analytical solutions for the Sobol' sensitivity indices. In particular, the index $S_{i_1 \enum i_s}$ is obtained as \citet{Sudret2008c}:
\begin{equation}
\label{eq:Sobol_S}
S_{i_1 \enum i_s} = \dfrac{1}{D}\prod_{i=i_1}^{i_s} D_i,
\end{equation}
with the partial and total variances respectively given by:
\begin{equation}
\label{eq:Sobol_Di}
D_i=\dfrac{1}{3(1+c_i)^2}
\end{equation}
and
\begin{equation}
\label{eq:Sobol_D}
D =\prod_{i=1}^{M} (D_i+1) -1.
\end{equation}
Then, the first-order or total indices for any single variable or groups of variables can be computed using Eq.~(\ref{eq:Si})-(\ref{eq:StuT}). Because the input variables follow a uniform distribution, we build LRA and PCE meta-models of the function using Legendre polynomials after an isoprobabilistic transformation into standard variables distributed uniformly over $[-1,1]$.
We first assess the comparative accuracy of LRA and PCE in estimating the mean and standard deviation of the response, respectively denoted by $\mu_Y$ and $\sigma_Y$. The LRA- and PCE-based estimates of $\mu_Y$ and $\sigma_Y$ are obtained in terms of the coefficients of the meta-models, as described in Sections~\ref{sec:PCE_indices} and \ref{sec:LRA_indices}. Table~\ref{tab:Sobol_moments} lists the analytical solutions for $\mu_Y$ and $\sigma_Y$ based on the actual model together with their LRA- and PCE-based estimates for two EDs of size $N=200$ and $N=500$ obtained with Sobol sequences (see the Appendix for the parameters of the respective meta-models). The relative errors $\vare$ of the estimates are given in parentheses. We observe that the mean is estimated with high accuracy in all cases, but LRA outperform PCE in the estimation of the standard deviation.
\begin{table} [!ht]
\centering
\caption{Sobol function: Mean and standard deviation of response for experimental designs obtained with Sobol sequences.}
\vspace{2mm}
\label{tab:Sobol_moments}
\begin{tabular}{c c c c c c}
\hline
{} & {} & \multicolumn{2}{c} {$N = 200$} & \multicolumn{2}{c} {$N = 500$} \\
& Analytical & LRA ($\vare \%$) & PCE ($\vare \%$) & LRA ($\vare \%$) & PCE ($\vare \%$) \\
\hline
$\mu_{Y}$ & $1.000$ & $ 1.005 $ ($0.5$) & $0.998$ ($-0.2$) & $1.000$ ($0.0$) & $0.995$ ($-0.5$) \\
$\sigma_{Y}$ & $0.3715$ & $0.3820$ ($2.8$) & $0.3424$ ($-7.8$) & $0.3715$ ($0.0$) & $ 0.3536$ ($-4.8$) \\
\hline
\end{tabular}
\end{table}
In the sequel, we compare LRA- and PCE-based sensitivity indices, obtained by post-processing the coefficients of the meta-models according to Sections~\ref{sec:PCE_indices} and \ref{sec:LRA_indices}, with the corresponding analytical solutions for the actual model. Figures~\ref{fig:Sobol_S1} and \ref{fig:Sobol_Stot} respectively show the first-order and total Sobol' indices of the variables $X_1 \enum X_5$ considering the same EDs of size $N=200$ and $N=500$ as above. The indices of the remaining variables $X_6 \enum X_{20}$ are practically zero. The small differences between the first-order and total indices indicate only minor effects from interactions between the various variables. For $N=200$, the sensitivity indices based on the meta-models are fairly close to the reference ones, whereas for $N=500$, the agreement is nearly perfect. The values of the indices depicted in the two figures are listed in the Appendix.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_S1_N200.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_S1_N500.pdf}
\caption{Sobol function: Comparison of LRA- and PCE-based first-order Sobol' indices to their analytical values for experimental designs obtained with Sobol sequences.}
\label{fig:Sobol_S1}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_Stot_N200.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_Stot_N500.pdf}
\caption{Sobol function: Comparison of LRA- and PCE-based total Sobol' indices to their analytical values for experimental designs obtained with Sobol sequences.}
\label{fig:Sobol_Stot}
\end{figure}
To further assess the effect of the ED size, we examine the convergence of the LRA- and PCE-based indices of the two most important variables, \ie $X_1$ and $X_2$, while $N$ varies from 100 to 2,000; as in the above investigations, the considered EDs are obtained with Sobol sequences. Figure~\ref{fig:Sobol_S1_N} shows the differences between the first-order indices obtained from the meta-model coefficients and their exact values; Figure~\ref{fig:Sobol_Stot_N} shows similar differences for the total indices. For $N\leq 200$, the PCE-based indices are overall more accurate; however, the LRA-based indices converge faster, practically reaching the exact values for $N=500$. Table~\ref{tab:Sobol_errG} lists the relative generalization errors of the considered meta-models, estimated with a MCS-based validation set comprising $10^6$ points. These errors decrease faster for LRA, which is consistent with the faster convergence of the respective sensitivity indices observed above.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_S1_N_LRA_diff.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_S1_N_PCE_diff.pdf}
\caption{Sobol function: Errors of LRA- and PCE-based first-order Sobol' indices for experimental designs obtained with Sobol sequences.}
\label{fig:Sobol_S1_N}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_Stot_N_LRA_diff.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_Stot_N_PCE_diff.pdf}
\caption{Sobol function: Errors of LRA- and PCE-based total Sobol' indices for experimental designs obtained with Sobol sequences.}
\label{fig:Sobol_Stot_N}
\end{figure}
\begin{table} [!ht]
\centering
\caption{Sobol function: Relative generalization errors of meta-models based on Sobol sequences.}
\vspace{2mm}
\label{tab:Sobol_errG}
\begin{tabular}{c c c}
\hline
N & $\widehat{err}_G^{\rm LRA}$ & $\widehat{err}_G^{\rm PCE}$\\
\hline
100 & $8.08\cdot 10^{-2}$ & $5.46\cdot 10^{-2}$ \\
200 & $2.57\cdot 10^{-2}$ & $3.64\cdot 10^{-2}$ \\
500 & $2.32\cdot 10^{-3}$ & $1.45\cdot 10^{-2}$ \\
1,000 & $4.68\cdot 10^{-4}$ & $6.34\cdot 10^{-3}$ \\
2,000 & $2.03\cdot 10^{-4}$ & $2.48\cdot 10^{-3}$ \\
\hline
\end{tabular}
\end{table}
We conclude this example by investigating the accuracy of the LRA- and PCE-based Sobol' indices when the meta-models are built with maximin LHS designs. Figure~\ref{fig:Sobol_Stot_LHS} shows boxplots of the differences between the total Sobol' indices obtained with the meta-models and their exact values, considering 20 EDs of size $N=200$ and $N=500$, \ie for the same ED sizes examined in Figure~\ref{fig:Sobol_Stot}. Note that in all cases, the absolute median errors do not exceed $0.01$. As expected, the error spread decreases with increasing ED size.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_Stot_N200_LHS.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Sobol_Stot_N500_LHS.pdf}
\caption{Sobol function: Errors of LRA- and PCE-based total Sobol' indices for experimental designs obtained with maximin LHS (20 replications).}
\label{fig:Sobol_Stot_LHS}
\end{figure}
\subsubsection{Beam deflection}
In this example, we consider a simply supported beam with a uniform rectangular cross-section subjected to a concentrated load at the mid-span. The response quantity of interest is the midspan deflection, which is obtained through basic structural mechanics as:
\begin{equation}
\label{eq:beam_u}
U=\dfrac{P L^3}{4 E b h^3},
\end{equation}
where $b$ and $h$ respectively denote the width and height of the cross section, $L$ is the length of the beam, $E$ is the Young's modulus and $P$ is the load. The aforementioned parameters are modeled as independent random variables with their distributions, mean and coefficient of variation (CoV) values listed in Table~\ref{tab:beam_input}. The dimensionality of the problem is thus $M=5$. LRA and PCE meta-models are built using Hermite polynomials after an isoprobabilistic transformation of the input variables into standard normal variables.
\begin{table} [!ht]
\centering
\caption{Beam deflection: Distributions of input variables.}
\vspace{2mm}
\label{tab:beam_input}
\begin{tabular}{c c c c}
\hline Variable & Distribution & mean & CoV \\
\hline $b$ (m) & Lognormal & 0.15 & 0.05 \\
$h$ (m) & Lognormal & 0.3 & 0.05 \\
$L$ (m) & Lognormal & 5 & 0.01 \\
$E$ (MPa) & Lognormal & 30,000 & 0.15 \\
$P$ (KN) & Lognormal & 10 & 0.20 \\
\hline
\end{tabular}
\end{table}
In Table~\ref{tab:beam_moments}, we assess the accuracy of LRA and PCE meta-models in estimating the mean $\mu_U$ and standard deviation $\sigma_U$ of the response by post-processing the meta-model coefficients. To this end, we consider two EDs of size $N=30$ and $N=50$ obtained with Sobol sequences (see the Appendix for the parameters of the respective meta-models). Because $U$ in Eq.~(\ref{eq:beam_u}) is a product of lognormal random variables, it easy to obtain the mean and standard deviation of the actual model response analytically. The relative errors $\vare$ of the LRA- and PCE-based estimates with respect to the analytical values are given in parentheses. $N=30$ is sufficient to obtain highly accurate estimates of both $\mu_U$ and $\sigma_U$ with the LRA approach; for both EDs, the PCE approach appears slightly inferior in the estimation of $\sigma_U$.
\begin{table} [!ht]
\centering
\caption{Beam deflection: Mean and standard deviation of response for experimental designs obtained with Sobol sequences.}
\vspace{2mm}
\label{tab:beam_moments}
\begin{tabular}{c c c c c c}
\hline
{} & {} & \multicolumn{2}{c} {$N = 30$} & \multicolumn{2}{c} {$N = 50$} \\
& Analytical & LRA ($\vare \%$) & PCE ($\vare \%$) & LRA ($\vare \%$) & PCE ($\vare \%$) \\
\hline
$\mu_U$ (mm) & $2.677$ & $2.678$ ($0.0$) & $2.675$ ($-0.1$) & $2.677$ ($0.0$) & $2.673$ ($-0.2$) \\
$\sigma_U$ (mm) & $0.8088$ & $0.8047$ ($-0.5$) & $0.7743$ ($-4.3$) & $0.8085$ ($0.0$) & $0.7917$ ($-2.1$)\\
\hline
\end{tabular}
\end{table}
Next, we examine the first-order and total Sobol' indices of the five input variables. The LRA- and PCE-based indices, obtained by post-processing the coefficients of the meta-models, are compared to respective reference values, obtained with a MCS approach using $n=10^6$ samples for each index. For the same EDs considered above, Figures~\ref{fig:beam_S1} and \ref{fig:beam_Stot} depict the first-order and total Sobol' indices, respectively, in order of total importance. It is noteworthy that the LRA-based indices are in nearly perfect agreement with the references ones even for the smaller ED, comprising only $N=30$ points. The PCE-based indices demonstrate a similar accuracy only for $N=50$. Obviously, the uncertainty in the beam deflection is driven by the uncertainty in the load, the cross-section height and the the Young's modulus, with the load having the dominant contribution. Comparisons between the first-order and the total indices indicate relatively small contributions from higher-order effects. The values of the indices depicted in the two figures are listed in the Appendix.
Table~\ref{tab:beam_errG} lists the relative generalization errors of the LRA and PCE meta-models built with the two EDs. These errors are estimated using a MCS-based validation set comprising $10^6$ points. The superior performance of LRA as compared to PCE in the estimation of the response statistics and the Sobol' indices is consistent with the higher accuracy of the former manifested in the lower generalization errors.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Beam_S1_N30.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Beam_S1_N50.pdf}
\caption{Beam deflection: Comparison of LRA- and PCE-based first-order Sobol' indices to their reference values for experimental designs obtained with Sobol sequences.}
\label{fig:beam_S1}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Beam_Stot_N30.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Beam_Stot_N50.pdf}
\caption{Beam deflection: Comparison of LRA- and PCE-based total Sobol' indices to their reference values for experimental designs obtained with Sobol sequences.}
\label{fig:beam_Stot}
\end{figure}
\begin{table} [!ht]
\centering
\caption{Beam deflection: Relative generalization errors of meta-models based on Sobol sequences.}
\vspace{2mm}
\label{tab:beam_errG}
\begin{tabular}{c c c}
\hline
N & $\widehat{err}_G^{\rm LRA}$ & $\widehat{err}_G^{\rm PCE}$\\
\hline
30 & $2.32\cdot 10^{-4}$ & $1.47\cdot 10^{-2}$ \\
50 & $2.63\cdot 10^{-6}$ & $1.81\cdot 10^{-3}$ \\
\hline
\end{tabular}
\end{table}
Finally, we investigate the accuracy of the LRA- and PCE-based Sobol' indices when the meta-models are built with maximin LHS designs. Figure~\ref{fig:beam_Stot_LHS} depicts boxplots of the differences between the total indices based on the meta-models and their reference values for 20 replications with ED sizes $N=30$ and $N=50$. Although both the LRA- and the PCE-based estimates are practically unbiased, the former exhibit a significantly smaller spread, which is nearly zero for $N=50$. The notably superior performance of LRA in the present problem can be explained by the rank-one structure of the underlying model.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Beam_Stot_N30_LHS.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Beam_Stot_N50_LHS.pdf}
\caption{Beam deflection: Errors of LRA- and PCE-based total Sobol' indices for experimental designs obtained with maximin LHS (20 replications).}
\label{fig:beam_Stot_LHS}
\end{figure}
\subsection{Finite element models}
\subsubsection{Truss deflection}
We consider the truss structure shown in Figure~\ref{fig:truss} (also studied in \citet{Sudret2007}), which is subjected to six vertical loads. The mid-span deflection, denoted by $u$, represents the response quantity of interest and is computed with an in-house finite-element analysis code developed in Matlab environment. The random input includes $M=10$ independent random variables: the vertical loads, denoted by $P_1 \enum P_6$, the cross-sectional area and Young's modulus of the horizontal bars, respectively denoted by $A_1$ and $E_1$, and the cross-sectional area and Young's modulus of the vertical bars, respectively denoted by $A_2$ and $E_2$. The distributions of the input random variables are listed in Table~\ref{tab:truss_input}. LRA and PCE meta-models are built using Hermite polynomials after an isoprobabilistic transformation of the input variables into standard normal variables.
\begin{figure}[!ht]
\centering
\includegraphics[trim = 15mm 70mm 15mm 70mm, width=0.6\textwidth]{Figures/Truss.pdf}
\caption{Truss structure.}
\label{fig:truss}
\end{figure}
\begin{table} [!ht]
\centering
\caption{Truss deflection: Distributions of input random variables.}
\vspace{2mm}
\label{tab:truss_input}
\begin{tabular}{c c c c}
\hline Variable & Distribution & mean & CoV \\
\hline
$A_1$ (m) & Lognormal & 0.002 & 0.10 \\
$A_2$ (m) & Lognormal & 0.001 & 0.10 \\
$E_1, E_2$ (MPa) & Lognormal & 210,000 & 0.10\\
$P_1\enum P_6$ (KN) & Gumbel & 50 & 0.15 \\
\hline
\end{tabular}
\end{table}
For two EDs of size $N=50$ and $N=200$ obtained with Sobol sequences, Table~\ref{tab:truss_moments} compares the estimates of the response mean and standard deviation obtained from the coefficients of LRA and PCE meta-models (see the Appendix for the meta-model parameters) with their respective values based on the actual model. The latter, which represent the reference values, are computed with a MCS sample comprising $n=10^6$ points. The relative errors $\vare$ of the LRA- and PCE-based estimates with respect to the reference values are given in parentheses. $N=50$ is sufficient to obtain excellent estimates of both quantities with the LRA approach; for the same ED, the PCE approach is slightly inferior in the estimation of the standard deviation.
\begin{table} [!ht]
\centering
\caption{Truss deflection: Mean and standard deviation of response for experimental designs obtained with Sobol sequences.}
\vspace{2mm}
\label{tab:truss_moments}
\begin{tabular}{c c c c c c}
\hline
{} & {} & \multicolumn{2}{c} {$N = 50$} & \multicolumn{2}{c} {$N = 200$} \\
& Reference & LRA ($\vare \%$) & PCE ($\vare \%$) & LRA ($\vare \%$) & PCE ($\vare \%$) \\
\hline
$\mu_U$ (cm) & $7.941$ & $7.945$ ($0.0$) & $7.931$ ($-0.1$) & $7.942$ ($0.0$) & $7.939$ ($0.0$) \\
$\sigma_U$ (cm) & $1.110$ & $1.109$ ($-0.1$) & $1.074$ ($-3.3$) & $1.114$ ($0.4$) & $1.106$ ($-0.4$) \\
\hline
\end{tabular}
\end{table}
For the same EDs of size $N=50$ and $N=200$, we next compare the Sobol' indices obtained by post-processing the LRA and PCE coefficients to respective reference values. The latter are computed with a MCS approach using $n=10^6$ samples for each index. Figure~\ref{fig:truss_Stot} shows the total Sobol' indices in order of total importance. For $N=50$, the indices obtained with the meta-models are overall in fair agreement with the reference ones. In particular, the dominant contributions of $A_1$ and $E_1$ are estimated with good accuracy with the LRA approach, but with slightly lesser accuracy with the PCE approach. For $N=200$, both meta-modeling approaches yield practically perfect estimates of the reference indices. Because contributions from higher-order effects are negligible in this problem, the first-order indices are nearly equal to the total indices and are thus not shown. The values of both the total and the first-order indices for the cases examined in Figure~\ref{fig:truss_Stot} are listed in the Appendix. Note that in the reference solution, total indices of less significant variables may be marginally smaller than the corresponding first-order indices. This contradiction results from a small bias in the employed Monte-Carlo estimator, which is addressed in \citet{Owen2013}.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\textwidth] {Figures/SA_Truss_Stot_N50.pdf}
\includegraphics[width=0.9\textwidth] {Figures/SA_Truss_Stot_N200.pdf}
\caption{Truss deflection: Comparison of LRA- and PCE-based total Sobol' indices to their reference values for experimental designs obtained with Sobol sequences.}
\label{fig:truss_Stot}
\end{figure}
We next examine the errors (differences) of the LRA- and PCE-based indices with respect to their reference values for the four most important variables, \ie $A_1$, $E_1$, $P_3$ and $P_4$, considering EDs of varying sizes obtained with Sobol sequences. Figure~\ref{fig:truss_Stot_N} shows these errors for the total indices, while $N$ varies from 30 to 500. The PCE approach demonstrates a slightly slower convergence to the reference solution with increasing ED size. Similar results are obtained for the first-order indices (not shown herein). Table~\ref{tab:truss_errG} shows the relative generalization errors of the considered meta-models, estimated with a MCS validation set comprising $10^6$ points. For $N\geq200$, the PCE meta-models are characterized by smaller generalization errors; however, as shown above, smaller values of $N$ are sufficient to perform highly accurate sensitivity analysis with LRA.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Truss_Stot_N_LRA_diff.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Truss_Stot_N_PCE_diff.pdf}
\caption{Truss deflection: Errors of LRA- and PCE-based total Sobol' indices for experimental designs obtained with Sobol sequences.}
\label{fig:truss_Stot_N}
\end{figure}
\begin{table} [!ht]
\centering
\caption{Truss deflection: Relative generalization errors of meta-models based on Sobol sequences..}
\vspace{2mm}
\label{tab:truss_errG}
\begin{tabular}{c c c}
\hline
N & $\widehat{err}_G^{\rm LRA}$ & $\widehat{err}_G^{\rm PCE}$\\
\hline
30 & $2.67\cdot 10^{-2}$ & $3.60\cdot 10^{-2}$ \\
50 & $2.83\cdot 10^{-3}$ & $1.25\cdot 10^{-2}$ \\
100 & $2.07\cdot 10^{-3}$ & $2.57\cdot 10^{-3}$ \\
200 & $1.22\cdot 10^{-3}$ & $1.23\cdot 10^{-4}$ \\
500 & $1.17\cdot 10^{-3}$ & $9.80\cdot 10^{-6}$ \\
\hline
\end{tabular}
\end{table}
As in the previous examples, we also investigate the accuracy of the LRA- and PCE-based Sobol' indices when the meta-models are built with maximin LHS designs. Figure~\ref{fig:truss_Stot_LHS} depicts boxplots of the differences between the total indices based on the meta-models and their reference values for 20 maximin LHS designs of size $N=50$. The LRA-based indices appear slightly superior to the PCE ones in terms of both the median value and the spread. For the two most influential variables, \ie $A_1$ and $E_1$, Figure~\ref{fig:truss_Stot_LHS_N} shows similar boxplots considering EDs of size varying from $N=30$ to $N=500$. For $N\leq 100$, the LRA-based indices exhibit a smaller spread, while for larger $N$, the spread of the estimates is nearly zero for both types of meta-models.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\textwidth] {Figures/SA_Truss_Stot_N50_LHS.pdf}
\caption{Truss deflection: Errors of LRA- and PCE-based total Sobol' indices for experimental designs obtained with maximin LHS (20 replications).}
\label{fig:truss_Stot_LHS}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_Truss_Stot_A1_diff_LHS.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_Truss_Stot_E1_diff_LHS.pdf}
\caption{Truss deflection: Errors of LRA- and PCE-based total Sobol' indices of the two dominant variables for varying sizes of experimental designs obtained with maximin LHS (20 replications).}
\label{fig:truss_Stot_LHS_N}
\end{figure}
\subsubsection{Heat conduction with spatially varying diffusion coefficient}
We consider stationary heat conduction over the two-dimensional domain $D=(-0.5,0.5)~\rm m\times(-0.5,0.5)~\rm m$ shown in Figure~\ref{fig:RF_domain}, with the temperature field $T(\ve z)$, $\ve z \in D$, described by the partial differential equation:
\begin{equation}
\label{eq:diffusion_eq}
-\nabla\cdot(\kappa(\ve z)\hspace{0.3mm} \nabla T(\ve z)) =Q \hspace{0.3mm} I_A(\ve z)
\end{equation}
and boundary conditions: $T=0$ on the top boundary; $\nabla T \cdot \ve n=0$ on the left, right and bottom boundaries, where $\ve n$ is the vector normal to the boundary. In Eq.~(\ref{eq:diffusion_eq}), $Q=500~\rm W/m^3$, $A=(0.2,0.3)~\rm m\times(0.2,0.3)~\rm m$ is a square domain within $D$ (see Figure~\ref{fig:RF_domain}) and $I_A$ is the indicator function equal to $1$ if $\ve z\in A$ and $0$ otherwise. The diffusion coefficient, $\kappa(\ve z)$, is a lognormal random field defined as:
\begin{equation}
\label{eq:diffusion_coef}
\kappa(\ve z)=\exp[a_{\kappa}+b_{\kappa} \hspace{0.3mm} g(\ve z)],
\end{equation}
where $g(\ve z)$ denotes a standard Gaussian random field with autocorrelation function:
\begin{equation}
\label{eq:autocorr}
\rho(\ve z,\ve z')=\exp{(-\|\ve z-\ve z'\|^2/\ell^2)}.
\end{equation}
In Eq.~(\ref{eq:diffusion_coef}), the parameters $a_{\kappa}$ and $b_{\kappa}$ are such that the mean and standard deviation of $\kappa$ are $\mu_{\kappa}=1~\rm W/°C\cdot m$ and $\sigma_{\kappa}=0.3~\rm W/°C\cdot m$, respectively, while in Eq.~(\ref{eq:autocorr}), $\ell=0.2~\rm m$. The response quantity of interest is the average temperature $\overline{T}$ in the square domain $B=(-0.3,-0.2)~\rm m\times(-0.3,-0.2)~\rm m$ (see Figure \ref{fig:RF_domain}):
\begin{equation}
\label{eq:diffusion_Y}
\overline{T}=\frac{1}{|B|}\int_{\ve z \in B}T(\ve z) \hspace{0.3mm} d\ve z.
\end{equation}
Using the Expansion Optimal Linear Estimation (EOLE) method \citep{Li1993optimal}, the random field $g(\ve z)$ is approximated by:
\begin{equation}
\label{eq:EOLE}
\widehat{g}(\ve z) = \sum_{i=1}^M \frac{\xi_i}{\sqrt{l_i}}\ve {\phi}_i^{\rm{T}} \ve C_{\ve z \ve \zeta}.
\end{equation}
In the above equation, $\{\xi_1 \enum \xi_M\}$ are independent standard normal variables; $\ve C_{\ve z \ve \zeta}$ is a vector with elements $\ve C_{\ve z \ve \zeta}^{(k)}=\rho(\ve z,\ve \zeta_k)$, $k=1 \enum n$, where $\{\ve{\zeta}_1 \enum \ve{\zeta}_n\}$ are the points of an appropriately defined grid in $D$; $(l_i,\ve{\phi}_i)$ are the eigenvalues and eigenvectors of the correlation matrix $\ve C_{\ve \zeta \ve \zeta}$ with elements $\ve C_{\ve \zeta \ve \zeta}^{(k,l)}=\rho(\ve{\zeta}_k,\ve{\zeta}_l)$, $k,l=1 \enum n$. It is recommended in \citet{Sudret2000stochastic} that for a square-exponential autocorrelation function, the size of the element in the EOLE grid must be $1/2-1/3$ of $\ell$. Accordingly, in the present numerical application, we use a square grid with element size $0.01~\rm m$, thus comprising $n=121$ points. The number of terms in the EOLE series is determined according to the rule:
\begin{equation}
\label{eq:EOLE_M}
\sum_{i=1}^{M}l_i/\sum_{i=1}^{n} l_i \geq 0.99,
\end{equation}
which herein leads to $M=53$. The shapes of the first 20 basis functions $\{{\phi}_i^{\rm{T}} \ve C_{\ve z \ve \zeta}(\ve z), \hspace{0.3mm} i=1 \enum 20\}$, are shown in Figure~\ref{fig:KL_modes}. Because the random input in this problem comprises independent standard normal variables, LRA and PCE meta-models are built using Hermite polynomials.
For a given realization of $\{\xi_1 \enum \xi_M\}$, the model response considered ``exact'' in the meta-modeling application is obtained with an in-house finite-element analysis code developed in Matlab environment. The left graph of Figure~\ref{fig:RF_domain} depicts the discretization of the domain in $16,000$ triangular T3 elements, obtained with software \emph{Gmsh} \citep{Geuzaine2009gmsh}. The temperature field $T(\ve z)$ for two example realizations of the conductivity random field is shown in In Figure~\ref{fig:maps}.
\begin{figure}[!ht]
\centering
\includegraphics[trim = 0mm 15mm 0mm 10mm, width=0.49\textwidth] {Figures/RF_domain}
\includegraphics[trim = 0mm 15mm 0mm 10mm,width=0.49\textwidth] {Figures/RF_mesh}
\caption{Heat conduction: Domain and boundary conditions (left); finite-element mesh (right).}
\label{fig:RF_domain}
\end{figure}
\vspace{8mm}
\begin{figure}[!ht]
\centering
\includegraphics[trim = 25mm 25mm 25mm 25mm, width=0.60\textwidth] {Figures/RF_modes.png}
\caption{Heat conduction: Shapes of the first 20 basis functions in the EOLE discretization (from left-top to bottom-right row-wise).}
\label{fig:KL_modes}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.49\textwidth] {Figures/map1.png}
\includegraphics[width=0.49\textwidth] {Figures/map2.png}
\caption{Heat conduction: Example realizations of the temperature field.}
\label{fig:maps}
\end{figure}
Because a single run of the considered ``exact'' model is computationally expensive (requires approximately $16~\rm sec$ with an Intel(R) Xeon(R) CPU E3-1225 v3 processor), evaluation of the Sobol' indices using MCS is impractical in this problem. Therefore, the following variance-based sensitivity analysis relies solely on the meta-modeling approach. Figure~\ref{fig:RF_Stot} shows the ten highest total Sobol' indices, obtained by post-processing the coefficients of LRA and PCE meta-models built with an ED of size $N=2,000$. The shown LRA- and PCE-based indices are practically identical and are thus, considered to represent the reference solution. (In the preceding applications, we showed that the indices based on LRA and PCE meta-models converge to the reference solution obtained with a MCS approach.) For the considered ED, a similar agreement is observed between the LRA-and PCE-based first-order indices. Because contributions from interaction effects are insignificant, the differences between the total and the first-order indices are negligible; the latter are thus not shown.
In the following, we examine the convergence of LRA- and PCE-based indices to their reference values with increasing ED size. For the four most significant variables, \ie $\xi_2$, $\xi_6$, $\xi_1$ and $\xi_7$, Figure~\ref{fig:RF_Stot_N} shows the differences between the total indices and their reference values, while $N$ varies between 100 and 2,000. With the LRA approach, the indices of $\xi_6$ and $\xi_1$ practically reach their reference values at $N=200$, while the indices of $\xi_2$ and $\xi_7$ reach their reference values at $N=500$. The PCE-based indices demonstrate a slightly slower convergence. Similar results are obtained for the first-order indices (not shown herein). Table~\ref{tab:RF_errG} shows the relative generalization errors of the considered meta-models, estimated with a validation set comprising $10^4$ points. For $N\geq500$, the LRA meta-models are characterized by larger generalization errors than PCE, but as shown in Figure~\ref{fig:RF_Stot_N}, this does not essentially affect the results of the sensitivity analysis. The values of the first-order and the total indices for the example case with $N=200$ as well as for the reference case with $N=2,000$ are listed in the Appendix.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\textwidth] {Figures/SA_RF_Stot_N2000.pdf}
\caption{Heat conduction: Comparison between LRA- and PCE-based total Sobol' indices for an experimental design of size $N=2,000$ obtained with Sobol sequences.}
\label{fig:RF_Stot}
\end{figure}
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth] {Figures/SA_RF_Stot_N_LRA_diff.pdf}
\includegraphics[width=0.45\textwidth] {Figures/SA_RF_Stot_N_PCE_diff.pdf}
\caption{Heat conduction: Errors of LRA- and PCE-based total Sobol' indices for experimental designs obtained with Sobol sequences.}
\label{fig:RF_Stot_N}
\end{figure}
\begin{table} [!ht]
\centering
\caption{Heat conduction: Relative generalization errors of meta-models based on Sobol sequences.}
\vspace{2mm}
\label{tab:RF_errG}
\begin{tabular}{c c c}
\hline
N & $\widehat{err}_G^{\rm LRA}$ & $\widehat{err}_G^{\rm PCE}$\\
\hline
100 & $2.46\cdot 10^{-2}$ & $3.89\cdot 10^{-2}$ \\
200 & $1.38\cdot 10^{-2}$ & $2.88\cdot 10^{-2}$ \\
500 & $9.68\cdot 10^{-3}$ & $9.15\cdot 10^{-3}$ \\
1,000 & $8.19\cdot 10^{-3}$ & $2.19\cdot 10^{-3}$ \\
2,000 & $7.72\cdot 10^{-3}$ & $9.58\cdot 10^{-4}$ \\
\hline
\end{tabular}
\end{table}
Although no reference solution for the Sobol' indices is available in the present example, one may obtain a reference ranking of the input variables in terms of simpler sensitivity measures, by relying on the validation set used above to compute the meta-model errors. One such sensitivity measure is the Spearman rank correlation $\rho^S_i$ between an input variable $X_i$ and the model response of interest $Y$, describing the linear correlation between the indices (ranks) of the \emph{ordered} samples of $X_i$ and $Y$. For the considered heat-conduction model, Figure~\ref{fig:RF_rhoS} depicts the ten most significant variables according to the absolute values of $\rho^S_i$, computed with the validation set. The ranking of the input variables in terms of the Spearman correlation based on the actual model is identical to their ranking in terms of the total Sobol' indices based on the meta-models. This observation provides a validation of the sensitivity-analysis results obtained with the meta-models under the premise that the model under consideration behaves nearly monotonically.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\textwidth] {Figures/SA_RF_rhoS.pdf}
\caption{Heat conduction: Sensitivity measures based on the Spearman rank correlation.}
\label{fig:RF_rhoS}
\end{figure}
\section{Conclusions}
In this paper, we introduce a novel approach for computing the Sobol' sensitivity indices by post-processing the coefficients of meta-models belonging to the class of low-rank approximations (LRA). The derived analytical expressions concern the particular case when LRA meta-models are built with polynomial functions due to the simplicity and versatility these offer. The proposed method can be particularly efficient in problems with high-dimensional input because: (i) the number of unknowns in LRA grows only linearly with the input dimension and (ii) the construction of LRA relies on a series of least-square minimization problems of small size, independent of the input dimension. The novel approach is presented in parallel to revisiting Sobol' sensitivity analysis by means of polynomial chaos expansions (PCE), which also relies on post-processing coefficients of meta-models with polynomial bases.
Following the detailed description of the construction of LRA in a non-intrusive manner and the derivation of the associated sensitivity indices, the accuracy and efficiency of the novel approach is investigated in example applications involving analytical models with a rank-one structure and finite-element models deriving from structural mechanics and heat conduction. The LRA-based indices are compared to PCE-based ones and to respective reference indices, representing the exact solution or Monte-Carlo estimates based on the actual model response. It is first shown that the indices obtained with the meta-models, using only a few tens or hundreds evaluations of the original model, provide excellent approximations of the exact indices or their Monte-Carlo estimates using a large number of evaluations of the original model. The heat-conduction example then demonstrates that the meta-modeling approach can be used to estimate the Sobol' indices in a case when their computation with the original model is not affordable; such situations are often encountered by practitioners in real-life problems. In the examined applications, the LRA-based indices tend to outperform the PCE-based ones by converging faster to the reference solution.
Despite the strong promise of the proposed approach for the global sensitivity analysis of computationally costly high-dimensional models, further investigations are needed in order to establish the capacities and limitations of LRA as a general meta-modeling tool. Because evaluation of the LRA coefficients relies on a series of alternated minimizations along separate dimensions, issues of convergence may be encountered in certain cases. Investigations into these issues are currently underway.
\bibliographystyle{chicago}
|
\section{Introduction}
The purpose of this paper is to prove the orbital stability
of solitary-wave solutions to a non-linear Schr\"odinger equation
\[
\tag{NLS}
\label{eq.NLS}
i\partial_t \phi(t,x) + \Delta_x \phi(t,x) - f(\phi(t,x)) = 0,\quad
\phi\colon\mathbb{R}_t\times\mathbb{R}\sp n_x\rightarrow\mathbb{C}
\]
in dimension $ n = 1 $ for general class of nonlinear functions
$ f $ such that $ f\colon\mathbb{C}\rightarrow\mathbb{C} $ is $ C\sp 1 $ and
\begin{equation}\label{eq.h1}
f(\overline{s}) = \overline{f(s)}, \quad f(zs) = zf(s),\quad
\forall z\in \mathbb{C}\text{ such that } |z| = 1.
\end{equation}
From \eqref{eq.h1}, if $ s $ is a real number, then $ f(s) $ is
a real number. We denote by $ g\colon\mathbb{R}\to\mathbb{R} $ the restriction of $ f $ to
$ \mathbb{R} $. From the second equality of \eqref{eq.h1}, $ g $ is an
even function. Let $ G $ be the primitive of $ g $ such that $ G(0) = 0 $.
We define for every complex number $ s $
\[
F(s) := G(|s|).
\]
A solitary-wave is a solution to \eqref{eq.NLS} of
the form
\begin{equation}\label{eq.I1}
\phi(t,x) = e\sp{i\omega t} R(x),\quad
(t,x)\in [0,\infty)\times\mathbb{R}\spn
\end{equation}
so that the gauge invariance condition \eqref{eq.h1} implies that
\[
\tag{SW}
\label{eq.SW}
\phi(t,x) = z e\sp{i(v\cdot x - t|v|\sp 2)}
e\sp{i\omega t} R(x - 2tv),\quad
(t,x)\in [0,\infty)\times\mathbb{R}\spn
\]
is also a solution to \eqref{eq.NLS}
for any $ v $ in $ \mathbb{R}\spn $ and any complex number $ z $ such that $ |z| = 1 $.
The profile $ R $ is a real-valued function. If $ \phi $ in \eqref{eq.SW}
is a solution to \eqref{eq.NLS}, then $ R $
satisfies the differential equation
\begin{equation}
\label{eq.E}
-R''(x) +f(R(x)) + \omega R(x) = 0.
\end{equation}
In this paper we address solutions to the equation above
which can be obtained as minima of the energy functional
\[
E\colon H\sp 1 (\RN;\C)\rightarrow\mathbb{R},\quad
E(u) := \frac{1}{2}\int_{\mathbb{R}\spn} |u'(x)|\sp 2 dx + \int_{\mathbb{R}\spn} F(u(x))dx
\]
under the constraint
\[
S(\lambda) := \{u\in H\sp 1\mid M(u) = \lambda\},
\]
where $ M(u) := \int_{\mathbb{R}\spn} |u(x)|^2dx $.
We will assume that the non-linearity satisfies conditions which
guarantee the global well-posedness of the initial value problem
of \eqref{eq.NLS} in $ H^1 $; that is, given an initial datum
$ u_0\inH\sp 1 (\RN;\C) $, there exists a unique solution
$ \phi(t,x)\in C([0,+\infty);H\sp 1 (\RN;\C)) $
to the Schr\"odinger equation
such that $ \phi(0,x) = u_0 (x) $.
The global well-posedness determines a one-parameter family of operators
$ U_t $ on $ H\sp 1 (\RN;\C) $.
To a real number $ \lambda > 0 $, we associate two
subsets of $ H^1 (\mathbb{R}\spn;\mathbb{C}) $:
\begin{gather*}
\mathcal{G}_\lambda :=
\{u\inH\sp 1 (\RN;\C)\mid E(u) = \inf_{S(\lambda)} E\}\\
\mathcal{G}_\lambda (u) := \{zu(\cdot + y)\mid (z,y)\in S^1\times\mathbb{R}\}.
\end{gather*}
The first one is called \textsl{ground state}.
The second set is a subset of the ground state; if $ u(x) = R(x) $,
then $ \mathcal{G}_\lambda (R) $ contains
the orbit of $ R $, that is, the point $ U_t (R) $ for every $ t\geq 0 $.
On $ H\sp 1 (\RN;\C) $, we consider the distance given by the scalar product
\[
(u,w)_{H\sp 1 (\RN;\C)} :=
\text{Re}\int_{\mathbb{R}\spn} u\overline{w}(x) dx +
\text{Re}\int_{\mathbb{R}\spn}\nabla u\cdot\nabla\overline{w}(x) dx.
\]
A set $ S\subseteqH\sp 1 (\RN;\C) $ is said \textsl{stable}
if for every $ \varepsilon > 0 $, there exists $ \delta > 0 $ such that
\[
\mathrm{dist}(u,S) < \delta\implies\mathrm{dist}(U_t (u),S)
\]
for every $ t\geq 0 $.
One of the first result of stability is the work of
T.~Cazenave and P.~L.~Lions in 1982, \cite{CL82},
where $ f $ is a pure power function.
Extensions
to more general non-linearities have been obtained in
\cite{Wei86} and \cite{BBGM07,Shi14}. However, while
in \cite{CL82} the stability of both $ \mathcal{G}_\lambda $ and
$ \mathcal{G}_\lambda (u) $ has been proved, in \cite{BBGM07,Shi14}
only the stability of the ground state is proved.
The pure power case
\[
f(s) = -|s|\sp{p - 2} s
\]
is very special as it exhibits
the rescaling invariance $ f(ts) = t\sp{p - 1} f(s) $ for every
$ t\geq 0 $.
As a consequence, $ \mathcal{G}_\lambda = \mathcal{G}_\lambda (u) $
for every $ u $ in $ \mathcal{G}_\lambda $.
In fact, it is possible to give a precise description of
an element $ u $ of the ground state:
\[
u(x) = \pm z\omega\sp{\frac{1}{p - 1}} R_1 (\omega\sp{1/2} (x + y)),
\quad \omega\sp{\frac{5 - p}{2(p - 1)}} \no{R_1}_{L\sp 2}\sp 2 = \lambda
\]
where $ R_1 $ is the unique positive solution to
\eqref{eq.E} when $ \omega = 1 $,
$ y $ is in $ \mathbb{R}\spn $, and $ z $ is complex with $ |z| = 1 $.
Therefore, the set $ \mathcal{G}_\lambda (u) $
is stable because the ground
state is stable. When more general non-linearities are considered,
the rescaling property fails, and it is not clear anymore whether
the equality $ \mathcal{G}_\lambda = \mathcal{G}_\lambda (u) $ holds. We list our assumptions:
\[
\tag{G1}
\label{G1}
\text{ there exists } s_0\in\mathbb{R}\text{ such that } G(s_0) < 0
\]
there exist $ C $, $ p,q,p_* $ and $ s_* $ such that
\[
\tag{G2}
\label{G2}
\begin{split}
|G'(s)|\leq C(|s|\sp{p - 1} + |s|\sp{q - 1}),\quad\forall s\in\mathbb{R}\\
-C|s|\sp{p_* - 1}\leq G'(s),\quad\forall s\geq s_*
\end{split}
\]
where $ 2 < p_* < 6 $ and $ 2 < p\leq q $.
\begin{theo}[Orbital stability of $ \mathcal{G}_\lambda $]
\label{thm.ground-state}
Suppose that \eqref{G1} and \eqref{G2} hold. Then,
there exists $ \lambda_* \geq 0 $ such that for every
$ \lambda > \lambda_* $, the functional $ E $ has a minimum, and
$ \mathcal{G}_\lambda $ is orbitally stable.
\end{theo}
A proof of the stability of $ \mathcal{G}_\lambda $ has been made
in \cite{BBGM07} in dimension $ n\geq 3 $ and in \cite{Shi14,BS11}.
Here we present a few improvements with respect to the assumptions made
in the quoted references. Firstly, we do not use (at this point)
the growth condition \eqref{G4}, required \cite[$ \mathrm{F}_p ' $]{BBGM07}
to obtain the splitting property \eqref{prop.1.3} of
Proposition~\ref{prop.1},
which follows directly from \cite{BL83}; secondly, \eqref{G2} weakens
\cite[F4]{Shi14}, where $ s_* $ is set to zero. Instead, we allow
the non-linearity to be critical nearby the origin.
We prove the orbital stability of $ \mathcal{G}_\lambda $ with a version of the
Concentration-Compactness Lemma of P.~L.~Lions,
\cite{Lio84a}, introduced by V.~Benci and D.~Fortunato in \cite{BF14} where the
classic definitions of Concentration, Compactness and Vanishing are expressed in
terms of weak convergence, instead of the Concentration Function used in
\cite{Lio84a}.
Concentration. There exists a subsequence $ (u_{n_k}) $, a sequence
$ (y_k) $ and $ u $ such that
\begin{equation}
\label{eq.C}
\tag{C}
u_{n_k} (\cdot + y_{n_k})\rightarrow u\text{ in } H\sp 1 (\mathbb{R};\mathbb{C}).
\end{equation}
Dichotomy. There exists a subsequence $ (u_{n_k}) $ and $ (y_k) $ such that
\begin{equation}
\label{eq.D}
\tag{D}
u_{n_k} (\cdot + y_k)\rightharpoonup u\text{ in } H\sp 1 (\mathbb{R};\mathbb{C})
\end{equation}
for some $ u $ such that
$ 0 < \no{u}_{H\sp 1} < \text{liminf}_{k\rightarrow +\infty} \no{u_{n_k}}_{H\sp 1} $.
Vanishing
\begin{equation}
\label{eq.V}
\tag{V}
u_{n_k}(\cdot + y_k)\rightharpoonup u\implies u = 0.
\end{equation}
The functional $ E $ is defined on $ H\sp 1 (\RN;\C) $ instead of real-valued functions.
This perspective of the minimization problem has the value of highlighting
features of the minima which are essential in the proof of the stability
of the other set, $ \mathcal{G}_\lambda(u) $: if $ u $ is a minimum, then $ u = zR $, where
$ R $ is a real-valued minimum and $ z $ a complex number with $ |z| = 1 $,
\eqref{lem.minima.4} of Lemma~\ref{lem.minima}. This fact relies on the
Convex Inequality for the Gradient, \cite[Lemma~7.8]{LL01}.
In the next assumption $ G $ is $ C^2 $ and
\[
\tag{G3}
\label{G3}
12G(R(x)) - 7R(x)G'(R(x)) + R(x)^2 G''(R(x))\geq 0
\]
for every solution $ R $ of \eqref{eq.E} and $ x\in\mathbb{R} $.
\[
\label{G4}
\tag{G4}
|G''(s)|\leq C(|s|\sp{p - 2} + |s|\sp{q - 2})\text{ for every } s\in\mathbb{R}
\]
and $ p,q $ as in \eqref{G2}.
We denote with $ H^1 _r (\mathbb{R}) $ the space of real-valued $ H^1 $ functions
which are even on $ \mathbb{R} $. Let $ \lambda_* $ be as in
Theorem~\ref{thm.ground-state}.
\begin{theo}
\label{thm.non-degeneracy}
Suppose that \eqref{G1}, \eqref{G2}, \eqref{G3} and \eqref{G4} hold.
Then, for $ \lambda > \lambda_* $, minima of $ E $ on
$ S(\lambda)\cap H^1 _r $ are non-degenerate.
\end{theo}
Our work presents some changes with respect to the one of M.~Weinstein,
\cite{Wei86} for the one-dimensional case. As we mentioned earlier in the
introduction, in order to have stability the condition
\[
\tag{B3}
\label{B3}
\int_{\mathbb{R}\spn}
\left[f(R^2) + 2R^2 f'(R^2) |R'(x)|^2\right]dx\neq 0
\]
was required (in his notation $ f(s)s = G'(s) $). In \eqref{G3}
we offer a different approach, as we prescribe a condition on
the non-linearity rather than on a solution to \eqref{eq.E}.
The non-degeneracy implies that the set $ \mathcal{G}_\lambda\cap H^1 _r $ is finite.
This should be compared with the pure-power case, where $ \mathcal{G}_\lambda\cap H^1 _r $
consists of exactly two functions. Consequently, under the same
assumptions as the theorem above and adding the assumption
\begin{theo}
\label{thm.sw}
Then the set $ \mathcal{G}_\lambda(u) $ is stable for every $ u\in\mathcal{G}_\lambda $.
\end{theo}
Finally, we show that under an additional assumption, a uniqueness
condition holds, just like the pure-power case. Then
$ \mathcal{G}_\lambda = \mathcal{G}_\lambda(u) $, Corollary~\ref{cor.uniqueness}.
For $ \omega > 0 $, we define
\begin{equation}
\label{eq.H}
V(s) := -\frac{2G(s)}{s^2}
\end{equation}
and
\begin{equation}
\label{eq.R}
R_* (\omega) := \inf\{s > 0\mid \omega = V(s)\}
\end{equation}
whenever the set on the right is non-empty.
\begin{enumerate}[(G1)]
\setcounter{enumi}{4}
\item\label{G5}
The set $ \mathscr{A} = \{\omega\mid V'(R_* (\omega)) > 0\} $ is an interval.
\end{enumerate}
\begin{theo}[Uniqueness]
\label{thm.uniqueness}
If (G1-5) hold, then $ \mathcal{G}_\lambda\cap H^1 _r $ consists of exactly two functions,
$ R_+ $ and $ R_- $. The first is positive while $ R_- = -R_+ $.
\end{theo}
Both the proofs of the uniqueness and the stability of $ \mathcal{G}_\lambda $
rely on the function $ d(\omega) $ defined by W.~Strauss in \cite{SS85}
and \cite{GSS87}. Condition which guarantees the
stability $ d''(\omega)\geq 0 $ follows from \eqref{G3}, and
Lemma~\ref{lem.non-degeneracy}. Here, we deduce the stability
of $ \mathcal{G}_\lambda(u) $ from the fact that the set $ H^1 _r \cap\mathcal{G}_\lambda $
is finite, rather than checking the assumptions of \cite{GSS87} directly.
We define
\[
L(s) = 12G(s) - 7sG'(s) + s^2 G''(s).
\]
When $ L = 0 $ on $ (0,+\infty) $, then $ F $ satisfies a Euler
equation whose solutions are linear combinations of $ s^2 $ and $ s^6 $.
If $ G $ is a pure-power $ -as^p $ with $ a > 0 $, then
\[
L(s) = a(p - 2)(6 - p)s^p
\]
which is strictly positive if $ p $ is sub-critical. Therefore
\eqref{G3} can be interpreted as a sub-critical assumption.
However, this interpretation fails as we consider sub-critical pure-power
combined non-linearities as
\begin{equation}
\label{eq.cp}
F(s) = -a|s|^p + b|s|^q,\quad p < q
\end{equation}
where
\[
L(s) = a(p - 2)(6 - p)s^p - b(q - 2)(6 - q)s^q
\]
and can be negative on $ (0,+\infty) $.
However, \eqref{G3} prescribes the behaviour
of $ L $ \textsl{only} on the union of the images of the solutions of
\eqref{eq.E} (for arbitrary $ \omega $). In fact, it turns out that
\eqref{eq.cp} does satisfy \eqref{G3}.
The paper is organized as follows. In \S\ref{sec.cc} we show
the Concentration-Compactness behaviour of minimizing sequences;
in \S\ref{sec.non-degenerate} we discuss the non-degeneracy of minima,
in \S\ref{sec.stability} the stability of the two sets $ \mathcal{G}_\lambda $ and
$ \mathcal{G}_\lambda(u) $, in \S\ref{sec.uniqueness} the uniqueness of positive, even
solutions in $ \mathcal{G}_\lambda $. In \S\ref{sec.cp} we show that \eqref{eq.cp}
satisfies all the assumptions mentioned above.
\section{Properties of the functional $ E $}
\label{sec.cc}
In Lemma~\ref{prop.split} we show that $ G $ can be written as sum of two
terms $ G_1 $ and $ G_2 $ which satisfy estimates \eqref{G2} and \eqref{G4}
having only a single power on the right term.
Since all the properties we will prove are well-behaved with respect to
linear combination, we will assume that in \eqref{G2} and \eqref{G4} there
is only the power $ p $.
Some of the properties listed in the next proposition have been
already thoroughly proved in \cite{BBGM07} in dimension
$ n\geq 3 $. We fill the details of the proof in the dimension
$ n = 1 $. Throughout this section we will assume that \eqref{G2}
holds.
\begin{proposition}
\label{prop1}
The functional $ E $ satisfies the following properties:
\begin{enumerate}[(i)]
\item
\label{prop1:2}
given $ e,\lambda > 0 $, there exists $ C(e,\lambda) $ such that
\[
E(u)\leq e\text{ and } M(u)\geq\lambda\implies\no{u}_{H^1} \leq C.
\]
Then minimizing sequences of $ E $ over $ S(\lambda) $ are bounded
\item
\label{prop.1.3}
if \eqref{G2} holds, given a weakly converging sequence
$ u_n\rightharpoonup u $
\begin{align*}
E(u_n) &= E(u_n - u) + E(u) + o(1)\\
M(u_n) &= M(u_n - u) + M(u) + o(1)
\end{align*}
\item
\label{prop1:8}
given a bounded sequence $ (u_n) $ in $ H\sp 1 $ and
a sequence $ (\lambda_n) $ such that $ \lambda_n\to\lambda $, then
\[
\lim_{n\rightarrow +\infty} \big(E(\lambda_n u_n) - E(\lambda u_n)\big) = 0
\]
\item
\label{prop1:4}
$ E $ is bounded from below on $ S(\lambda) $.
\end{enumerate}
\end{proposition}
\begin{proof}
\,
\eqref{prop1:2} and \eqref{prop1:4}. From the Sobolev-Gagliardo-Nirenberg
inequality there exists a $ S\in\mathbb{R} $ such that
\begin{equation}
\label{eq.SGN}
\no{u}_{L\sp d}\sp d\leq S\sp d
\no{u}_{L\sp 2}\sp{\frac{d + 2}{2}}
\no{u'}_{L\sp 2}\sp{\frac{d - 2}{2}}
\end{equation}
for every $ d\geq 2 $. We set $ A := \{|u|\geq s_*\} $. On $ A $,
$ F(u(x)) $ is bounded from below by $ -C|u(x)|^p $ or
$ -Cs_* ^{p - p_*} |u(x)|^{p_*} $, depending on whether
$ p\leq p_* $. On $ A\sp c $, we use $ -C|u(x)|\sp{p_*} $.
In any case, from \eqref{G2}
\[
\int_{\mathbb{R}\spn} F(u(x))dx\geq -C\max\{1,s_* ^{p - p_*}\} \no{u}_{L^d} ^d
\]
where $ d < 6 $. Then, from \eqref{eq.SGN}
\begin{equation}
\label{eq.14}
\begin{split}
2e\geq 2E(u)\geq\no{u'}_{L\sp 2}\sp 2 -
C'\lambda\sp{\frac{d + 2}{4}}
\no{u'}_{L\sp 2}\sp{\frac{d - 2}{2}}.
\end{split}
\end{equation}
Then $ C(e,\lambda) $ exists because $ (d - 2)/2 < 2 $. Then minimizing
sequences are bounded
\eqref{prop.1.3}. We refer to the paper of H.~Brezis and E.~Lieb~\cite{BL83}.
\eqref{prop1:8}. We write the proof only for the non-linear part
$ \int F(u) dx $, for which we use the same notation $ E $. We define
\[
\begin{split}
k_n (x) &:= F(\lambda_n u_n(x)) - F (\lambda u_n(x)) \\
&= \int_0\sp 1 F'(t\lambda u_n(x) + (1 - t)\lambda_n u_n(x))
(\lambda_n - \lambda)u_n(x) dt.
\end{split}
\]
Then
\[
\begin{split}
|k_n (x)|&\leq C |\lambda_n - \lambda|
\int_0\sp 1 |t\lambda u_n(x) + (1 - t)\lambda_n u_n(x)|\sp{p - 1} |u_n(x)| dt\\
&\leq 2\sp{p - 2} C |\lambda_n - \lambda|
\int_0\sp 1 \big(|t\lambda u_n(x)|\sp{p - 1} +
|(1 - t)\lambda_n u_n(x)|\sp{p - 1}\big) |u_n(x)|dt\\
&\leq 2\sp{p - 2} C|\lambda_n - \lambda|
(|\lambda|\sp{p - 1} + |\lambda_n|\sp{p - 1}) |u_n(x)|\sp{p - 1}.
\end{split}
\]
By integrating on $ \mathbb{R} $, we obtain
\[
\begin{split}
|E(\lambda_n u_n) - E(u_n)|&\leq\int_{\mathbb{R}\spn} |k_n(x)| dx\\
&\leq 2\sp{p - 2} C|\lambda_n - \lambda|
(|\lambda|\sp{p - 1} + |\lambda_n|\sp{p - 1}) \no{u_n}_{L\sp p}\sp p.
\end{split}
\]
Since $ (u_n) $ is bounded in $ H^1 $, as we take the limit
as $ n\to +\infty $, we obtain the conclusion.
\end{proof}
We are then allowed to define
\[
I(\lambda) := \inf_{S(\lambda)} E.
\]
\begin{proposition}
\label{prop.1}
The function $ I $ satisfies the following properties:
\begin{enumerate}[(i)]
\item
\label{prop1:5}
the function $ I $ is non-positive
\item
\label{prop1:6}
for every $ \vartheta\geq 1 $ and $ \lambda > 0 $, there holds
\[
I(\vartheta\lambda)\leq\vartheta I(\lambda).
\]
If equality holds, then either $ \vartheta = 1 $ or $ \vartheta > 1 $
and $ I(\lambda) = 0 $
\item
\label{prop.1.7}
there exists $ \lambda_* > 0 $ such that
\[
I < 0\text{ on } (\lambda_*,+\infty),\quad I = 0\text{ on } (0,\lambda_*].
\]
If $ \lambda\leq\lambda_* $, then $ \mathcal{G}_\lambda $ is empty.
\end{enumerate}
\end{proposition}
\begin{proof}
\,
\eqref{prop1:5}. The proof of this fact follows from
\cite[Lemma~2.3]{Shi14}.
\eqref{prop1:6}.
Such property of $ I $ has been proved in
\cite[Lemma~3.2]{Shi14} and \cite[Proposition~15]{BBGM07} using
rescalings. However, in both references it is assumed
that $ E $ achieves its infimum on $ S(\lambda) $.
Here, we just apply the same rescaling to a minimizing
sequence $ (u_n) $ over $ S(\lambda) $
\[
u_{n,\vartheta} (x) = u_n (\vartheta\sp{-1} x).
\]
Clearly, $ u_n\in S(\vartheta\lambda) $. Then
\begin{equation}
\label{prop.1.eq.1}
\begin{split}
I(\vartheta\lambda)\leq E(u_{n,\vartheta}) &= \vartheta\left(
\frac{\vartheta\sp{-2}}{2} \int_{\mathbb{R}\spn} |u_n'|\sp 2 dx +
\int_{\mathbb{R}\spn} F(u_n)dx\right) \\
&= \vartheta\left(E(u_n) -
\frac{1 - \vartheta\sp{-2}}{2}\no{u_n'}_{L\sp 2}\sp 2\right)
\\
&\leq \vartheta I(\lambda) -
\frac{\vartheta (1 - \vartheta\sp{-2})}{2}
\cdot\no{u_n'}_{L\sp 2}\sp 2\leq\vartheta I(\lambda).
\end{split}
\end{equation}
Clearly, if equality holds and $ \vartheta > 1 $, then the
sequence of gradients converges to zero. From \eqref{eq.f4} and
\eqref{eq.SGN}, we obtain
\[
\int_{\mathbb{R}\spn} F(u_n)dx\to 0.
\]
Therefore, $ I(\lambda) = 0 $.
\eqref{prop.1.7}.
From \cite[Lemma~5]{BBGM07} it follows that there exists
$ \lambda_1 $ such that $ I(\lambda_1) < 0 $. Here we use the assumption
\eqref{G1}. Now, suppose that there exists $ \lambda_0 $ such that
$ I(\lambda_0) = 0 $. If $ \lambda' \leq \lambda_0 $, then
\[
0 = I(\lambda') =
I(\lambda'/\lambda_0\cdot\lambda_0)\leq (\lambda'/\lambda_0)
I(\lambda_0)\leq 0.
\]
The first inequality follows from (ii), while the last inequality follows
from \eqref{prop1:5}. Therefore, the set
$ \{\lambda\mid I(\lambda) = 0\} $ is bounded or is equal to
$ (0,+\infty) $. The second case is ruled about by $ \lambda_1 $.
On the first case, we define
\[
\lambda_* := \sup\{\lambda\mid I(\lambda) = 0\}.
\]
Since $ I $ is continuous (from \cite[Lemma~2.3]{Shi14}),
$ I(\lambda_*) = 0 $. Now, we consider the case $ \lambda \leq\lambda_1 $.
Let $ u\in\mathcal{G}_\lambda $ be a minimum. We apply to $ u $ the same rescaling
as in \eqref{prop.1.eq.1}. For every $ \vartheta $ the endpoints of
the inequalities are zero, then $ \no{u'}_2\sp 2 = 0 $
which gives $ u = 0 $, and obtain a contradiction with $ \lambda > 0 $.
\end{proof}
We define $ J_k $ the open interval $ (k,k + 1) $.
The result of the next lemma is well-known from
\cite[Lemma~I.1]{Lio84a}: if a sequence $ (u_n) $ vanishes, then
all the $ L\sp d $ norms converge to zero. In \cite{Lio84a} they show that if
\begin{equation}
\label{eq.28}
\lim_{n\to +\infty}\sup_{k\in\mathbb{Z}} \no{u_n}_{L\sp 2 (J_k)} = 0,
\end{equation}
then the sequence of $ L\sp d $ norms of $ (u_n) $
also converges to zero. Here we write a proof which provides
an estimate of the $ L^d $ norm by a product of the $ H^1 $ norm and
\eqref{eq.28}. We need the Sobolev-Gagliardo-Nirenberg inequality for the
interval $ J_k $
\begin{equation}
\label{eq.36}
\no{u}_{L\sp d (J_k)}\sp d\leq
s^d\no{u}_{L\sp 2 (J_k)}\sp{\frac{d + 2}{2}}
\no{u}_{H\sp 1 (J_k)}\sp{\frac{d - 2}{2}}.
\end{equation}
\begin{lemma}
\label{lem:vanishing}
For every $ u\inH\sp 1 (\RN;\C) $ there holds
\begin{equation}
\label{eq.v1}
\no{u}_{L\sp d}\sp d\leq s^d
\Big(\sup_{k\in\mathbb{Z}} \no{u}_{L\sp 2 (J_k)}\Big)\sp{d - 2}
\no{u}_{H\sp 1}\sp 2
\end{equation}
if $ d\geq 6 $ and
\begin{equation}
\label{eq.v2}
\no{u}_{L\sp d}\sp d \leq s\sp d
\Big(\sup_{k\in\mathbb{Z}} \no{u}_{L\sp 2 (J_k)}\Big)\sp{\frac{d + 2}{2}}
\no{u}_{H\sp 1}\sp{\frac{d - 2}{2}}
\end{equation}
if $ d\leq 6 $.
\end{lemma}
\begin{proof}
First case: $ d\geq 6 $, that is $ (d - 2)/2 \geq 2 $.
\[
\no{u}_{H\sp 1 (J_k)}\sp{\frac{d - 2}{2}} =
\no{u}_{H\sp 1 (J_k)}\sp 2 \no{u}_{H\sp 1 (J_k)}\sp{\frac{d - 6}{2}}.
\]
Then, in \eqref{eq.36} we have the product
of the bounded sequence
\[
a_k := \no{u}_{L\sp 2 (J_k)}\sp{\frac{d + 2}{2}}
\no{u}_{H\sp 1 (J_k)}\sp{\frac{d - 6}{2}},\quad \no{a}_{\infty} =
\Big(\sup_{k\in\mathbb{Z}} \no{u}_{L\sp 2 (J_k)}\Big)\sp{\frac{d + 2}{2}}
\no{u}_{H\sp 1}\sp{\frac{d - 6}{2}}
\]
and the summable sequence
\[
b_k := \no{u}_{H\sp 1 (J_k)}\sp 2,\quad \no{b}_1 = \no{u}_{H\sp 1}\sp 2.
\]
Therefore,
\[
\no{u}_{L\sp d}\sp d\leq
s\sp d
\Big(\sup_{k\in\mathbb{Z}} \no{u}_{L\sp 2 (J_k)}\Big)\sp{\frac{d + 2}{2}}
\no{u}_{H\sp 1}\sp{\frac{d - 2}{2}}.
\]
Second case: $ d\leq 6 $. Then $ (d - 2)/2 \leq 2 $, and we have
\[
\frac{6 - d}{2} = 2 - \frac{d - 2}{2} < \frac{d + 2}{2}.
\]
Then,
\[
\no{u}_{L\sp 2 (J_k)}\sp{\frac{d + 2}{2}}
\no{u}_{H\sp 1 (J_k)}\sp{\frac{d - 2}{2}}\leq
\no{u}_{L\sp 2 (J_k)}\sp{\frac{d + 2}{2} - \frac{6 - d}{2}}
\no{u}_{H\sp 1 (J_k)}\sp{\frac{d - 2}{2} + \frac{6 - d}{2}} =
\no{u}_{L\sp 2 (J_k)}\sp{d - 2}
\no{u}_{H\sp 1 (J_k)}\sp 2.
\]
Again, taking the sum over $ \mathbb{Z} $, we obtain
\[
\no{u}_{L\sp d}\sp d\leq s^d
\Big(\sup_{k\in\mathbb{Z}} \no{u}_{L\sp 2 (J_k)}\Big)\sp{d - 2}
\no{u}_{H\sp 1}\sp 2.
\]
\end{proof}
\begin{lemma}
\label{lem.L2-H1}
Let $ (w_n)\subseteqH\sp 1 (\RN;\C) $ be a sequence such that
\[
E(w_n)\rightarrow I(\lambda)\text{ and } M(w_n)\rightarrow\lambda.
\]
Suppose that $ (w_n) $ converges to $ w $ in $ L\sp 2 $.
Then there exists a subsequence of $ (w_n) $ which converges strongly
in $ H\sp 1 (\RN;\C) $.
\end{lemma}
\begin{proof}
From \eqref{eq.14}, $ (w_n) $ is bounded in $ H^1 $ (in the inequality
$ \lambda = M(w_n) $). Then, there exists a subsequence $ (w_{n_k}) $
which converges weakly to $ w $ in $ H\sp 1 $, and pointwise a.e.
\[
\begin{split}
o(1) + I(\lambda) &= E(w_{n_k}) =
\frac{1}{2}\int_{\mathbb{R}\spn} |w_{n_k}' (x)|\sp 2 dx + \int_{\mathbb{R}\spn} F(w_{n_k}(x)) dx \\
&\geq \frac{1}{2}\int_{\mathbb{R}\spn} |w'(x)|\sp 2 dx +
\int_{\mathbb{R}\spn} F(|w(x)|) dx \geq I(\lambda) + o(1).
\end{split}
\]
Since $ (w_{n_k}) $ converges pointwise a.e. and $ L^2 $, by \eqref{eq.SGN},
$ \int F(w_{n_k}) dx $ converges to $ \int F(w)dx $. From this fact and
the weak lower-semicontinuity of $ \int |u'|^2 dx $, we obtained
the first inequality. The second inequality follows from the strong
convergence in $ L\sp 2 $
which implies that $ w $ is in $ S(\lambda) $.
Then, taking the limit,
\[
\lim_{k\rightarrow +\infty}\int_{\mathbb{R}\spn} |w_{n_k}'(x)|\sp 2 dx =
\int_{\mathbb{R}\spn} |w'(x)|\sp 2 dx
\]
implying the convergence of $ w_{n_k}' $ to $ w' $ in $ L\sp 2 $.
In the next lemma, $ \lambda_* $ is as in Proposition~\ref{prop.1}.
\end{proof}
\begin{lemma}
\label{lem.concentration-compactness}
Let $ \lambda $ be such that $ \lambda > \lambda_* $. A subsequence of
a sequence $ (u_n) $ such that
\[
M(u_n) \to\lambda,\quad E(u_n)\rightarrow I(\lambda)
\]
satisfies the concentration behaviour \eqref{eq.C}.
\end{lemma}
\begin{proof}
We show that $ (u_n) $ does not vanish and does not have a dichotomy.
If $ (u_n) $ vanishes, from (V), up to extract a subsequence
\begin{equation}
\label{eq.18}
\lim_{n\to +\infty}\sup_{k\in\mathbb{Z}} \no{u_n}_{L\sp 2 (J_k)} = 0.
\end{equation}
Otherwise, there exists $ \varepsilon_0 > 0 $ and a sequence $ (k_n) $ such that
\[
\no{u_{k_n}(\cdot -k_n)}_{L\sp 2 ((0,1))} =
\no{u_{k_n}}_{L\sp 2 (J_{k_n})}\geq\varepsilon_0
\]
However, a subsequence
of $ u_{k_n} (\cdot + y_k) $ converges weakly to zero and,
since $ (0,1) $ is bounded, the $ L\sp 2 $-norm converges to zero,
giving a contradiction.
From \eqref{prop.1.7}, $ I(\lambda) > 0 $.
From the inequalities \eqref{eq.v1} and \eqref{eq.v2}, and \eqref{eq.18},
a subsequence of $ (u_{k_n}) $ converges
to zero in $ L\sp p \cap L\sp q $. Therefore, $ I(\lambda) = 0 $,
and we have a contradiction.
Let $ (u_{n_k}) $, $ (y_k) $ and $ u $ be as in \eqref{eq.D}.
Firstly, we observe that the inequality
\[
0 < \no{u}_{L\sp 2} < \liminf_{k\rightarrow +\infty} \no{u_{n_k}}_{L\sp 2}
\]
holds too. Otherwise, we had strong convergence in $ L\sp 2 $ and
thus, strong convergence in $ H\sp 1 $, by Lemma~\ref{lem.L2-H1}
and a contradiction with the dichotomy assumption.
We define
\[
\lambda_1 \sp k:= \no{u_{n_k} (\cdot + y_k) - u}_{L\sp 2}\sp 2.
\]
Up to extract a subsequence, we can suppose that $ \lambda_1 \sp k $
converges. We use the notation $ \lambda_1 $ for its limit and we have
\[
0 < \lambda_1 < \lambda.
\]
We set
\[
w_k := \frac{\lambda_1}{\lambda_1\sp k}\cdot (u_{n_k} (\cdot + y_k) - u).
\]
By \eqref{prop1:2} of Proposition~\ref{prop1},
the sequence $ (w_k) $ is bounded. Then, we can apply \eqref{prop1:8}
and \eqref{prop.1.3}
\begin{equation}
\label{eq.31}
E(u_{n_k}) = E(u_{n_k} (\cdot + y_k) - u) + E(u) + o(1) =
E(w_k) + E(u) + o(1).
\end{equation}
Here, we use the argument of \cite[Lemma~20,p.\,5]{BF14}.
We define
\[
\Lambda(u) := \frac{E(u)}{M(u)}.
\]
By (v) of Proposition~\ref{prop1} and \eqref{eq.31} we have
\[
o(1) + \frac{I(\lambda)}{\lambda} =
\frac{E(w_k) + E(u)}{M(w_k) + M(u)} + o(1)
\geq
\min\{\Lambda(w_k),\Lambda(u)\} + o(1)
\]
Let us suppose that the term of $ w_k $ is the smaller (on the
other case, the argument is the same). Then
\[
o(1) + \frac{I(\lambda)}{\lambda}\geq
\frac{I(\lambda_1)}{\lambda_1} + o(1)\geq
\frac{I(\lambda)}{\lambda} + o(1).
\]
The last inequality is a consequence of \eqref{prop1:6} of
Proposition~\ref{prop1}: the function $ I(\lambda)/\lambda $ is
decreasing. Then, all the inequalities are equalities.
\[
\frac{I(\lambda)}{\lambda} = \frac{I(\lambda_1)}{\lambda_1}.
\]
From \eqref{prop1:6} of Proposition~\ref{prop1}, either $ \vartheta := \lambda/\lambda_1 = 1 $, which is ruled out by the dichotomy assumption,
or $ I(\lambda) = 0 $, which contradicts the assumptions on $ \lambda $.
Thus, the sequence is not dichotomy.
\end{proof}
\begin{proposition}
\label{prop.minimum}
$ \mathcal{G}_\lambda\neq\emptyset $ for every $ \lambda > \lambda_* $, and
the Lagrange multiplier is negative. If $ \lambda\leq\lambda_* $,
then $ \mathcal{G}_\lambda $ is empty.
\end{proposition}
\begin{proof}
The second part is just \eqref{prop.1.7} of Proposition~\ref{prop.1}.
Let $ (u_n) $ be a minimizing sequence. Since $ \lambda > \lambda_* $
the assumptions of Lemma~\ref{lem.concentration-compactness} are
satisfied and there exists $ (y_n)\subseteq\mathbb{R} $ and $ u\in S(\lambda) $
such that
\[
u_n (\cdot + y_n)\to u.
\]
From Proposition~\ref{prop.regular}, $ E $ is continuous. Then, taking the
limit as $ n\to +\infty $ in
\[
o(1) + I(\lambda) = E(u_n) = E(u_n (\cdot + y_n)) = o(1) + E(u)
\]
we obtain $ E(u)\in S(\lambda) $. Now, we do not set any restriction
on $ \lambda $ and just assume that $ u\in\mathcal{G}_\lambda $. By (i) of
Proposition~\ref{prop.regular}, there exists $ \omega\in\mathbb{R} $ such that
\[
u''(x) - f(u(x)) - \omega u(x) = 0.
\]
Taking the scalar product in $ \mathbb{C} $ with $ u' $ and obtain
\begin{equation}
\label{eq.24}
u'(x)\sp 2 - 2F(u(x)) - \omega u(x)\sp 2\equiv d
\end{equation}
for some constant $ d $. On the left side we have a sum of $ L^1 $
functions. Therefore, $ d = 0 $. Integrating
on $ \mathbb{R} $, we obtain
\[
\int_{\mathbb{R}\spn} |u'(x)|\sp 2 dx - 2\int_{\mathbb{R}\spn} F(u(x))dx - \omega\lambda = 0.
\]
Since $ u $ is a minimum, the equality above becomes
\begin{equation}
\label{thm.existence.1}
2\left(\int_{\mathbb{R}\spn} |u'(x)|\sp 2 dx - I(\lambda)\right) = \omega\lambda.
\end{equation}
From \eqref{prop1:5} of Proposition~\ref{prop.1}, the left term
is non-negative. Then $ \omega > 0 $.
\end{proof}
\begin{remark}
\label{rem.critical}
The critical case $ G(s) = as^6 $ has been already ruled out by the
assumption \eqref{G2}. In this case,
a minimum does not exist. On the contrary,
the rescaling
\[
u_\eta = \eta^{1/2} u(\eta x)
\]
gives $ E(u_\eta) = \eta^2 E(u) = \eta^2 I(\lambda) $. Therefore,
$ I(\lambda) = 0 $ unless $ E $ is unbounded from below. By \eqref{prop.1.7}
of Proposition~\ref{prop.1}, a minimum does not exist.
\end{remark}
We conclude this section by showing general properties satisfied
by minima of $ E $ over $ S(\lambda) $.
\begin{lemma}
\label{lem.minima}
Let $ u $ be a minimum of $ E $ over $ S(\lambda) $.
Then $ R(x) := |u(x)| $ satisfies the following properties:
\begin{enumerate}[(i)]
\item
\label{lem.minima.1}
$ \lim_{|x|\to +\infty} R(x) = 0 $
\item
\label{lem.minima.2}
$ R $ is symmetrically decreasing with respect to a point of
$ \mathbb{R}\spn $
\item
\label{lem.minima.3}
$ R $ is positive
\item there exists $ z $ such that $ |z| = 1 $ and $ u(x) = zR(x) $
for every $ x\in\mathbb{R} $.
\label{lem.minima.4}
\end{enumerate}
\end{lemma}
\begin{proof}
Clearly, $ R $ is in $ S(\lambda) $.
From the equality $ F(s) = F(|s|) $ and the Convex Inequality for the
Gradient,
\cite[Theorem~7.8]{LL01},
there holds $ E(u)\geq E(R) $. Since $ u $ is a minimum,
necessarily
\begin{equation}
\label{eq.21}
E(u) = E(R).
\end{equation}
Thus, $ R $ is solution to \eqref{eq.E} for some $ \omega $.
Since $ R $ is $ H\sp 1 $ it is also $ L\sp{\infty} $.
From \eqref{eq.24} and the continuity of $ F $, the function $ |R'| $
is bounded. Since $ R\in L\sp 2 $, we obtain
\begin{equation}
\label{eq.23}
R (x)\rightarrow 0\text{ as } |x|\rightarrow +\infty
\end{equation}
which is the condition \cite[6.1]{BL83a} (following
their notation $ G'(s) $ should be replaced with $ - G'(s) - \omega s $).
By \cite[(ii), Proof~of~Theorem~5]{BL83a}, $ R $ is positive;
by \cite[(i,iv), Proof~of~Theorem~5]{BL83a}, $ R $ is a symmetrically
decreasing function with respect to a point in $ y $ in $ \mathbb{R} $.
Then, $ R' $ converges to zero as well.
So, we proved (\ref{lem.minima.1},\ref{lem.minima.2}) and
\eqref{lem.minima.3}.
(iv). From \eqref{eq.21} and $ F(s) = F(|s|) $,
there also holds
$ \no{u'}_{L\sp 2}\sp 2 = \no{|u|'}_{L\sp 2}\sp 2 $.
Since $ R > 0 $, by \cite[Lemma~5.1]{Gar12}, there exists a
complex number $ z $ such that $ |z| = 1 $ and $ u(x) = z|u(x)| $
for every $ x $.
\end{proof}
\section{Non-degeneracy of the minima on $ H^1 _r (\mathbb{R}) $}
\label{sec.non-degenerate}
In this section we prove the non-degeneracy of the functional
$ E $ when restricted to the sub-manifold $ S(\lambda)\cap H^1 _r (\mathbb{R}) $
on minima.
We need the notation
\begin{equation}
\label{eq.ND24}
Q(\omega,s) := \omega s^2 + 2G(s).
\end{equation}
We have
\begin{equation}
\label{eq.ND25}
R_* (\omega) = \inf\{s > 0\mid Q(\omega,s) = 0\}.
\end{equation}
where $ R_* (\omega) $ has been defined in \eqref{eq.R}.
\begin{remark}
\label{rem.ND1}
If \eqref{G2} holds, then $ R_* $ is a positive non-decreasing
function defined on $ (0,+\infty) $.
\end{remark}
Let $ R_0 $ be an element of $ \mathcal{G}_\lambda\cap H^1 _r (\mathbb{R}) $.
Then, there exists $ \omega_0 $ such that
\begin{equation}
\label{eq.ND8}
R_0 ''(x) = G'(R_0 (x)) + \omega_0 R_0 (x).
\end{equation}
By \eqref{lem.minima.1} of Lemma~\ref{lem.minima} $ R_0 $
converges to zero and then satisfies the condition \cite[6.1]{BL83a}.
Then, by (iii) of \cite[Theorem~5]{BL83a},
\begin{equation}
\label{eq.ND26}
R_0 (0) = R_* (\omega_0),\quad\partial_s Q(\omega_0,R_*(\omega_0)) < 0.
\end{equation}
By the Implicit Function Theorem, there exists $ \varepsilon_0 > 0 $
such that $ R_* $ is continuously differentiable on
$ (\omega_0 - \varepsilon_0,\omega_0 + \varepsilon_0) $ and
\[
\partial_s Q(\omega,R_*(\omega)) < 0.
\]
Also, since $ \omega_0 > 0 $, by Proposition~\ref{prop.minimum} and
\eqref{eq.21}, on this
interval $ \omega > 0 $.
We consider the solution of the initial value problem
\begin{equation}
\label{eq.IVP}
R_\omega ''(x) = G'(R_\omega (x)) + \omega R_\omega(x),
\quad R_\omega '(0) = 0,\quad R_\omega(0) = R_* (\omega).
\end{equation}
From \cite[Theorem~5]{BL83a}, $ R_\omega $ converges to
zero and, by \cite[Remark~6.3]{BL83a} and the fact that
$ \omega > 0 $, we obtain $ R_\omega\in H^1 $.
Since $ Q(\omega,R_* (\omega)) = 0 $,
differentiating with respect to $ \omega $, we obtain
\begin{equation}
\label{eq.NDS9k}
(2G'(R_*(\omega)) + 2\omega R_*(\omega)) R_* '(\omega) +
R_* ^2(\omega)= 0.
\end{equation}
We define
\[
\lambda\colon (\omega_0 - \varepsilon_0,\omega_0 + \varepsilon_0)\to\mathbb{R},\quad
\lambda(\omega) := \no{R_\omega}_{L^2} ^2.
\]
\begin{equation}
\label{eq.NDS9a}
\omega R_* (\omega)^2 + 2G(R_* (\omega)) = 0.
\end{equation}
\begin{lemma}
\label{lem.non-degeneracy}
$ \lambda'(\omega)\geq 0 $ and $ \lambda'(\omega_0) > 0 $,
provided \eqref{G4} holds.
\end{lemma}
\begin{proof}
From (iv) of \cite[Theorem~5]{BL83a}, $ R_\omega $ is a strictly
decreasing function on $ |x| $. Then, since $ R_\omega $ is
real valued, from \eqref{eq.24} we have
\[
R_\omega '(x)^2 = \omega R_\omega (x) ^2 + 2G(R_\omega(x))
\]
and then
\[
R_\omega '(x) = -\sqrt{\omega R_\omega (x)^2 + 2G(R_\omega (x))}.
\]
We can write
\[
\begin{split}
\lambda &= 2 \int_0^\infty R_\omega (x)^2 dx = 2 \int_0^\infty
\frac{R_\omega (x)^2}{-\sqrt{\omega R_\omega (x)^2 + 2G(R_\omega(x))}}
R_\omega ^\prime (x) dx \\
&= 2 \int_{0}^{R_*(\omega)} \frac{\rho^2 d\rho}%
{\sqrt{\omega \rho^2 + 2G(\rho)}}
= 2 \int_{0}^{1}
\frac{\theta^2 d\theta}{\sqrt{\Psi(\theta,R_*(\omega),\omega)}}.
\end{split}
\]
The third and the fourth equalities follow from the substitutions
$ \rho = R(x) $ and $ \rho = R_* (\omega)\theta $, and
\[
\Psi(\theta,s,\omega) = \omega \theta^2 s^{-4} + 2 s^{-6}G(s\theta).
\]
We prove that the function
\[
\omega \to \Psi(\theta, R_*(\omega), \omega)
\]
is non-increasing in $ \omega $. Then, we have to check that
\begin{equation}\label{eq.mp3}
\partial_\omega \Psi(\theta, R_*(\omega), \omega)\leq 0.
\end{equation}
In turn, the derivative above is equal to
\[
\begin{split}
(R_*(\omega))^{-4}\theta^2
&+ \left[-4\omega\theta^2(R_*(\omega))^{-5} -
12 (R_*(\omega))^{-7} G(R_*(\omega)\theta)\right.\\
&+ \left. 2 (R_*(\omega))^{-6} \theta
G'(R_*(\omega)\theta) \right] R_*^\prime(\omega).
\end{split}
\]
From Remark~\ref{rem.ND1}, the term
\[
R_* (\omega)^7 (R_* ' (\omega))^{-1}
\]
is positive. Then, dividing $ \partial_\omega\Psi $ by that term
and using the relation \eqref{eq.NDS9k}, we obtain
\[
\begin{split}
I(\omega,\theta) = &-R_*(\omega)\theta^2\left[ 2 \omega R_*(\omega)
- 2G'(R_*(\omega)\right] - 4\omega \theta^2 (R_*(\omega))^2\\
&- 12 G(R_*(\omega)\theta) + 2 R_*(\omega) \theta G'(R_*(\omega)\theta)
\end{split}
\]
so using \eqref{eq.NDS9a} we see that
\[
\begin{split}
I(\omega,\theta) &= 12 \theta^2 G(R_*(\omega)) -
2\theta^2 R_*(\omega) G'(R_*(\omega)) \\
&+ 12 G(R_*(\omega)\theta) -
2R_*(\omega)\theta G'(R_*(\omega)\theta).
\end{split}
\]
Setting
\[
H(s) := - 6G(s) + sG'(s),
\]
we obtain
\[
\begin{split}
I(\omega,\theta) &= 2H(R_*(\omega)\theta) - 2\theta^2 H(R_*(\omega)) \\
&= 2\theta^2 R_* (\omega) ^2\left(\frac{H(R_*(\omega)\theta)}%
{R_* (\omega)^2 \theta^2} -
\frac{H(R_* (\omega)}{R_* (\omega)^2}\right).
\end{split}
\]
Now we prove that the function $ H(s)/s^2 $ is monotonically non-decreasing
on the interval $ (0,R_* (\omega)) $.
Equivalently, we need to check that
\[
H' s - 2 H = 12G(s) - 7sG'(s) + s^2 G''(s)\geq 0.
\]
If we require \eqref{G3}, the inequality holds.
Moreover, $ I(\omega,1) = 0 $.
Then, for every $ 0\leq\theta\leq 1 $, we have $ I(\omega,\theta)\leq 0 $.
In conclusion,
\[
\frac{d\lambda}{d\omega} = -\int_0\sp 1
\frac{\theta^2 \partial_\omega (\Psi(\theta,R_* (\omega),\omega)) d\theta}%
{(\Psi(\theta,R_*(\omega),\omega))^{3/2}}\geq 0.
\]
We are now able to prove that $ \lambda'(\omega_0) > 0 $.
On the contrary,
\[
\partial_\omega (\Psi(\theta,R_*(\omega_0),\omega_0)) = 0
\]
for every $ 0 < \theta < 1 $, and the same applies to $ I $.
Therefore,
\[
12G(s) - 7sG'(s) + s^2 G''(s) = 0\text{ for every } s\in (0,R_* (\omega_0)).
\]
Then $ G(s) = as^6 $ on $ (0,R_* (\omega_0)) $.
By \cite[Theorem~5]{BL83a}, there is only one solution to \eqref{eq.ND8}
which is positive and converges to zero at infinity. Then
$ R_{\omega_0}(0) = R_0(0) $, so the image of $ R_0 $ is contained
in a set of $ \mathbb{R} $ where $ G $ is a pure-power critical non-linearity.
From Remark~\ref{rem.critical}, $ \mathcal{G}_\lambda $ is empty, giving a contradiction.
\end{proof}
We wish to evaluate the Hessian operator of $ E $ at the critical point
$ R_0 $, in a vector of the tangent space of $ R_0 $
\begin{equation}
\label{eq.ND7}
T_{R_0} (S_r (\lambda)) = \{v\in H^1 _r (\mathbb{R})\mid (v,R)_{L^2} = 0\}.
\end{equation}
We consider a curve in $ S_r (\lambda) $ as in
\begin{equation}
\label{eq.ND2}
u(t) = R + t v + \alpha(t)R.
\end{equation}
By the Implicit Function Theorem, there exists $ \delta > 0 $
and $ \alpha\colon (-\delta,\delta)\to\mathbb{R} $ such that
\[
M(R + t v + \alpha(t)R) = \lambda.
\]
From the Taylor expansion of $ M $ we get
\[
\alpha(t) = \alpha_0 t^2 + o(t^2),\quad
\alpha_0 = - \frac{\| v \|^2_{L^2}}{2\lambda},
\]
and from the expansion of $ E(u(t)) $ we get
\[
2E(u(t)) = \|R_0 ^\prime\|_{L^2}^2 + 2 t (R_0 ^\prime,
v^\prime)_{L^2} + t^2 \left(\|v^\prime \|^2 + 2\alpha_0
(R_0 ^\prime, v^\prime)_{L^2} \right) + o(t^2)
\]
so using \eqref{eq.E} and \eqref{eq.ND7} we find
\[
2 E(u(t)) = 2E(R_0)
t^2 \left(\|v^\prime \|^2 + ( (G''(R_0)+ \omega R_0) v,v)_{L^2}
\right) + o(t^2).
\]
Therefore,
\[
\begin{split}
D^2 E(R_0)[v,v] &= (E\circ u)''(0) \\
&=
\int_{\mathbb{R}\spn}\big(|v'(x)|^2 + (G''(R_0 (x)) + \omega_0) v(x)^2\big)dx =:
\xi(v).
\end{split}
\]
In order to show that $ R_0 $ is non-degenerate, we have to
prove that the infimum of $ \xi $ is positive
$ T_{R_0} (S_r (\lambda))\cap S(1) $.
If $ v $ is $ H^2 (\mathbb{R}) $, then
\[
\xi(v) = (L_+ v,v)_{L^2},\quad
L_+ (v) := v'' - G''(R_0) v - \omega_0 v.
\]
\begin{proof}[Proof~of~Theorem~\ref{thm.non-degeneracy}]
Since $ R_0 $ is a minimum, $ \xi(v)\geq 0 $.
The infimum of $ \xi $ achieved. A proof of this can be found in
\cite[Proposition~2.10]{Wei85}.
Suppose that the infimum is achieved and that $ \xi(v) = 0 $.
Then, $ v $ is $ H^2 $ and satisfies
\[
L_+ (v) = \beta R_0
\]
for some $ \beta\in\mathbb{R} $ with $ \beta\neq 0 $.
Taking the derivative with respect to $ \omega $ of \eqref{eq.IVP},
and evaluating at $ \omega = \omega_0 $, we obtain
\begin{equation}
\label{eq.ND27}
L_+ (\partial_\omega R(\omega_0,\cdot)) = R_0.
\end{equation}
Then $ y := \beta v_0 + v $ solves the differential equation
$ L_+ (y) = 0 $.
The kernel of the operator $ L_+ $ is generated by $ R_0 ' $,
which is an odd function. Since $ y $ is even, we obtain $ y = 0 $.
Since $ \beta\neq 0 $,
\[
(L_+ (\partial_\omega R(\omega_0,\cdot)),
\partial_\omega R(\omega_0,\cdot))_{L^2} = 0
\]
However, from \eqref{eq.ND27} and the definition $ \lambda $
given in Lemma~\ref{lem.non-degeneracy}, we have
\[
0 = (R_0,\partial_\omega R(\omega_0,\cdot))_{L^2} = \lambda'(\omega_0)
\]
which gives a contradiction with the lemma.
\end{proof}
\defProof{Proof}
\begin{corollary}
The set $ \mathcal{G}_\lambda\cap H^1 _r (\mathbb{R}) $ is finite.
\end{corollary}
\begin{proof}
Let $ (R_n)\subseteq\mathcal{G}_\lambda\cap H\sp 1 _r (\mathbb{R};\mathbb{R}) $ be a sequence
of minima. By Lemma~\ref{lem.concentration-compactness}, up
to extract a subsequence, we can suppose that
$ R_n (\cdot + y_n) $ converges in $ H\sp 1 $, for
some sequence $ (y_n)\subseteq\mathbb{R} $.
By \cite[Theorem~5]{BL83a}, $ R_n $ is symmetric
and radially decreasing with respect to the origin.
Therefore,
\[
\no{R_n - R_m}_{L\sp 2}\sp 2\leq
\no{R_n (\cdot + y_n) - R_m (\cdot + y_m)}_{L\sp 2}\sp 2.
\]
Then $ (R_n) $ is a Cauchy sequence and there exists $ R_0 $ such that
$ R_n\to R_0 $ in $ L^2 $. By Lemma~\ref{lem.L2-H1}, the convergence
is strong in $ H^1 $, which contradicts the fact that
$ R_0 $ is non-degenerate, thus isolated, minimum.
\end{proof}
\begin{proposition}
\label{prop.finite}
As $ u $ varies in $ \mathcal{G}_\lambda $, there are finitely many $ \mathcal{G}_\lambda (u) $.
\end{proposition}
\begin{proof}
By \eqref{lem.minima.2} and \eqref{lem.minima.4} of
Lemma~\ref{lem.minima}, there
exists $ y\in\mathbb{R} $ and a complex number $ |z| = 1 $ such that
$ u(x) = zR(x + y) $, where $ R\in\mathcal{G}_\lambda\cap H_{r,+} ^1 (\mathbb{R}) $.
Therefore, there are as many different $ \mathcal{G}_\lambda (u) $ as
$ \#\mathcal{G}_\lambda\cap H_{r,+} ^1 (\mathbb{R}) $.
\end{proof}
\section{Stability of $ \mathcal{G}_\lambda $ and $ \mathcal{G}_\lambda (u) $}
\label{sec.stability}
According to \cite[Theorem~3.5.1,\,p.\,77]{Caz03} the
The equation \eqref{eq.NLS} is locally well posed in
$ H\sp 1 (\mathbb{R};\mathbb{C}) $. That is, given $ u $
in $ H\sp 1 (\RN;\C) $, there exists a map
\begin{equation}
\label{eq.32}
U\in C\sp 1 \big([0,+\infty);L\sp 2 (\mathbb{R};\mathbb{C})\big)\cap
C\big([0,+\infty);H\sp 1 (\mathbb{R};\mathbb{C})\big)
\end{equation}
such that $ U_0 = u $ and $ \phi(t,\cdot) := U_t (u) $
is a solution to \eqref{eq.NLS}. We briefly check that $ G' $ satisfies the
the condition of \cite[Example~3.2.4,\,p.\,59]{Caz03}: since
$ G'' $ is continuous,
\[
|G'(s_1) - G'(s_2)|\leq L(K)|s_1 - s_2|,\quad L(K) := \sup_{[0,K]} |G''|
\]
for every $ s_1,s_2\in [0,K] $. And the function $ L $ is continuous
because $ G'' $ is continuous. The global well-posedness
follows from the apriori estimates that one can derive from
\eqref{prop1:2} of Proposition~\ref{prop1}.
\begin{proof}[Proof of Theorem~\ref{thm.ground-state}]
The proof of the stability is made with a contradiction argument:
let $ (u_n) $, $ \varepsilon_0 > 0 $ and $ (t_n) $ be such that
\begin{equation}
\label{eq.contradiction}
\mathrm{dist}(u_n,\mathcal{G}_\lambda)\to 0,\quad\mathrm{dist}(U_{t_n} (u_n),\mathcal{G}_\lambda)\geq\varepsilon_0.
\end{equation}
Since $ E $ and $ M $ are continuous functions, and constant
on the orbits $ U_t (u_n) $,
\[
E(U_{t_n}(u_n))\to I(\lambda),\quad M(u_n)\to\lambda.
\]
We set $ v_n := U_{t_n} (u_n) $.
From Lemma~\ref{lem.concentration-compactness}, there exists a
subsequence $ u_{n_k} $, a sequence $ (y_k) $ and
$ u\in S(\lambda) $ such that
\[
\no{u_{n_k} (\cdot + y_k) - u}_{H\sp 1 (\mathbb{R};\mathbb{C})}\to 0
\]
implying
\[
\no{u_{n_k} - u(\cdot - y_k)}_{H\sp 1 (\mathbb{R};\mathbb{C})}\to 0
\]
and giving a contradiction with \eqref{eq.contradiction}.
\end{proof}
\begin{proof}[Proof~of~Theorem~\ref{thm.sw}]
\textsl{Stability of $ \mathcal{G}_\lambda (u) $}. By Proposition~\ref{prop.finite},
\[
\mathcal{G}_\lambda = \bigcup_{i = 1}\sp n \mathcal{G}_\lambda (R_i).
\]
We prove that
\[
\mathrm{dist}(\mathcal{G}_\lambda(R_h),\mathcal{G}_\lambda(R_k)) > 0\text{ for } h\neq k.
\]
In fact, the distance between two arbitrary points in the two
sets is
\[
\begin{split}
\mathrm{dist}(z_1 R_h (\cdot + y_1),z_2 R_k(\cdot + y_2)) &=
\mathrm{dist}(R_h, zR_k (\cdot + y))\\
&\geq\mathrm{dist}(R_h,R_k) > 0,
\end{split}
\]
where $ z = \overline{z}_1 z_2 $ and $ y := y_2 - y_1 $.
The first inequality follows the fact that both $ R_i $ are symmetrically
decreasing with respect to the origin, from \eqref{lem.minima.2} of
Lemma~\ref{lem.minima}. Then
\begin{equation}
\label{eq.distance}
d := \inf_{h\neq k} \mathrm{dist}(\mathcal{G}_\lambda(R_h),\mathcal{G}_\lambda(R_k)) > 0.
\end{equation}
Now we prove that $ \mathcal{G}_{\lambda} (R_i) $ is stable.
Let $ \delta > 0 $ be such that
\begin{equation}
\label{eq.isolated}
B(R_i,\delta)\cap\mathcal{G}_\lambda\cap H^1 _r = \{R_i\},\quad \delta < \frac{d}{3}.
\end{equation}
We define
\[
E_\delta\sp i := \inf_{B(\mathcal{G}_\lambda(R_i),\delta)} E
\]
where the metric restricted on $ S(\lambda) $.
We claim that
\begin{equation}
\label{eq.11}
E_\delta\sp i > I(\lambda).
\end{equation}
Otherwise, we would have a sequence $ (u_n) $ such that
\[
E(u_n)\to I(\lambda),\quad M(u_n) = \lambda.
\]
By Lemma~\ref{lem.concentration-compactness}, there exists
a subsequence $ (u_{n_k}) $, $ u $ in $ S(\lambda) $ and $ (y_k) $ such that
\begin{equation}
\label{eq.17}
u_{n_k} (\cdot + y_k)\rightarrow u\in\mathcal{G}_\lambda.
\end{equation}
From \eqref{eq.isolated} and the choice of $ \delta $, it follows
that $ u $ is in $ \mathcal{G}_\lambda(R_i) $. However, since
\[
\mathrm{dist}(u_{n_k} (\cdot + y_k),\mathcal{G}_\lambda(R_i)) = \delta
\]
there also hold $ \mathrm{dist}(u,\mathcal{G}_\lambda(R_i)) = \delta $, giving a
contradiction with \eqref{eq.17}. We are now able to
prove that $ \mathcal{G}_\lambda(R_i) $ is stable; again,
we use a contradiction argument. Let $ (u_n) $,
$ (t_n) $ and $ \varepsilon_0 > 0 $ be such that
\[
\mathrm{dist}(u_n,\mathcal{G}_\lambda(R_i))\to 0,\quad
\mathrm{dist}(U_{t_n} (u_n),\mathcal{G}_\lambda(R_i))\geq\varepsilon_0.
\]
We set $ v_n := U_{t_n} (u_n) $. Since $ \mathcal{G}_\lambda $ is stable, there exists $ k $
such that
\[
\mathrm{dist}(v_n,\mathcal{G}_\lambda(R_k))\to 0,\quad
\mathcal{G}_\lambda(R_i)\neq\mathcal{G}_\lambda(R_k).
\]
Let $ n_0 $ be such that
\begin{equation}
\label{eq.10}
\max\{\mathrm{dist}(u_n,\mathcal{G}_\lambda(R_i)),\mathrm{dist}(v_n,\mathcal{G}_\lambda(R_k)\}
< \delta
\end{equation}
and
\begin{equation}
\label{eq.12}
E(u_n) < \min\{E_\delta\sp i\mid 1\leq i\}
\end{equation}
for every $ n\geq n_0 $.
Along the curve
\[
\alpha\colon [0,t_{n_0}]\rightarrow H\sp 1 (\mathbb{R};\mathbb{C}),\quad
\alpha(t) = U_t (u_{n_0})
\]
the quantities $ E $ and $ M $ are constant, while the function
\[
\beta\colon\mathbb{R}\rightarrow\mathbb{R},\quad \beta(t) := \mathrm{dist}(U_t (u_{n_0}),\mathcal{G}_\lambda(R_i))
\]
is continuous, from \eqref{eq.32}. From \eqref{eq.10} and \eqref{eq.isolated},
we have
\[
\beta(0) < \delta,\quad \beta(t_{n_0})\geq 2\delta.
\]
Therefore, there exists $ t_* $ such that
\[
\beta(t_*) = \delta.
\]
Then,
\[
E(\alpha(t_*))\geq E_\delta\sp i.
\]
However from the conservation of $ E $ and \eqref{eq.12}
\[
E(\alpha(t_*)) = E(\alpha(0)) < E_\delta\sp i.
\]
And from \eqref{eq.11}, we obtain a contradiction.
\end{proof}
\begin{TAKEOUT}
Now, we wish to prove that solitary-waves are stable, in the following
sense: given $ v $ in $ \mathbb{R} $, the set
\[
G_\lambda (u,v) :=
\big\{ze\sp{ixv} u(\cdot + y)\mid (z,y)\in S\sp 1\times\mathbb{R}\big\}.
\]
is stable. The results relies on \eqref{eq.33}, the fact that
$ G_\lambda (u,0) = G_\lambda (u) $ is stable, and the next lemma.
We define $ S(x) := e\sp{ixv} $. From the local uniqueness of solutions
to \eqref{eq.NLS}, if $ \phi $ is a solution to \eqref{eq.NLS}, then
\[
\psi(t,x) := e\sp{i(v\cdot x - t|v|\sp 2)}\phi(t,x - 2tv)
\]
is a solution to \eqref{eq.NLS}.
Therefore,
\begin{equation}
\label{eq.33}
U_t (S\Psi)(x) = S(x) e\sp{-it|v|\sp 2} U_t (\Psi)(x - 2tv).
\end{equation}
It is convenient to define the following linear operators
\[
L_v \colonH\sp 1 (\RN;\C)\raH\sp 1 (\RN;\C),\quad L(\Phi) = e\sp{ixv}\Phi,\quad
B_t (\Phi) := e\sp{-it|v|\sp 2} \Phi(x - 2tv).
\]
\begin{lemma}
\label{lem:semigroups}
For every $ v,w,t $ in $ \mathbb{R} $
\begin{enumerate}[(i)]
\item
\label{lem:semigroups-1}
$ L_{v + w} = L_v \circ L_w $
\item
\label{lem:semigroups-2}
$ L_v $ is bounded and
$ \no{L_v}_{\mathcal{B}(H\sp 1 (\RN;\C))}\leq\sqrt{2(1 + |v|\sp 2)} $
\item
\label{lem:semigroups-3}
$ L_v (G_\lambda (u,w)) = G_\lambda (u,v + w) $
\item
\label{lem:semigroups-4}
$ B_t $ is an isometry of $ H\sp 1 (\RN;\C) $
\item
\label{lem:semigroups-5}
$ B_t (G_\lambda (u,v)) = G_\lambda (u,v) $.
\end{enumerate}
\end{lemma}
\begin{proof}
We check only \eqref{lem:semigroups-2}, where some computation is
required, while all the other statements can be safely checked by
the reader:
\begin{align*}
\no{L_v (\Phi)}_{L\sp 2}\sp 2 &=
\text{Re}\int_{\mathbb{R}\spn} e\sp{ixv} \Phi(x) e\sp{-ixv} \overline{\Phi}(x)dx =
\no{\Phi}_{L\sp 2}\sp 2\leq 2\no{\Phi}_{H\sp 1}\sp 2\\
\no{L_v (\Phi)'}_{L\sp 2}\sp 2 &=
\no{e\sp{ixv} (\Phi + iv\Phi')}_{L\sp 2}\sp 2\leq
2(\no{\Phi}_{L\sp 2}\sp 2 + |v|\sp 2 \no{\Phi'}_{L\sp 2}\sp 2)\leq 2|v|\sp 2
\no{\Phi}_{H\sp 1}\sp 2.
\end{align*}
Taking the sum, we obtain the desired estimate.
\end{proof}
It is useful to write the relation in \eqref{eq.33} using the operators
$ B_t $ and $ U $:
\begin{equation}
\label{eq.35}
L_v\sp {-1}\circ U_t = B_t\circ U_t\circ L_v\sp{-1}.
\end{equation}
\begin{proof}[Proof~of~Corollary~\ref{cor:1}]
In order to simplify the notation we set
\[
G := G_\lambda (u),\quad G_* := G_\lambda (u,v),\quad L := L_v.
\]
Then
\[
\begin{split}
\mathrm{dist}(U_t (\Psi),G_*) &=
\mathrm{dist}(U_t (\Psi),L(G))\leq\sqrt{2(1 + |v|\sp 2)}
\mathrm{dist}(L\sp{-1} U_t (\Psi),G) \\
&=
\sqrt{2(1 + |v|\sp 2)}\mathrm{dist}(B_t (U_t (L\sp{-1} (\Psi))),G) \\
&=
\sqrt{2(1 + |v|\sp 2)}\mathrm{dist}(U_t (L\sp{-1} (\Psi),G)).
\end{split}
\]
Since $ G $ is stable, given $ \varepsilon > 0 $, there exists $ \delta > 0 $
such that
\[
\mathrm{dist}(\Phi,G) < \delta\implies\mathrm{dist}(U_t(\Phi),G) <
\frac{\varepsilon}{\sqrt{2(1 + |v|\sp 2)}}.
\]
We choose
\[
\delta_v := \frac{\delta}{\sqrt{2(1 + |v|\sp 2)}}
\]
If $ \mathrm{dist}(\Psi,G_*) < \delta_v $, then
$ \mathrm{dist}(L\sp{-1}(\Psi),G) < \delta $. Then
\[
\mathrm{dist}(L\sp{-1}(\Psi),G) < \frac{\varepsilon}{\sqrt{2(1 + |v|\sp 2)}}.
\]
From the chain of inequalities in \eqref{eq.}, we obtain
\[
\mathrm{dist}(U_t(\Psi),G_*) < \varepsilon.
\]
\end{proof}
\begin{proof}[Proof of Corollary~\ref{cor:2}]
Let $ M $ be a stable and invariant subset of $ G_\lambda $,
that is $ U_t (M)\subseteq M $ for every $ t\geq 0 $.
We show that $ M $ is closed under multiplication by complex numbers in
$ S\sp 1 $ and space translations. That is, given $ \Phi $ in $ M $,
there holds
\[
z\Phi(\cdot + y)\in M
\]
for every $ (z,y)\in S\sp 1\times\mathbb{R} $. We set
$ u(x) := |\Phi(x)|\in G_\lambda $. From Lemma~\ref{lem:minima},
there exists $ z $ in $ S\sp 1 $ such that $ \Phi(x) = zu(x) $.
There also exists $ \omega > 0 $ such that
\[
U_t (\Phi)(x) = ze\sp{i\omega t} u(x) = e\sp{i\omega t}\Phi(x).
\]
Since $ M $ is invariant, for every $ t\geq 0 $, the function above
belongs to $ M $. Since $ \omega > 0 $, therefore non-zero,
the image of $ e\sp{i\omega t} $
is $ S\sp 1 $.
Now, we prove that $ M $ is invariant under space translations.
Let $ y $ be a point of $ \mathbb{R} $.
We will show that $ u(\cdot + y) $ is in $ M $.
We consider the example of T.~Cazenave
and P.~L.~Lions, \cite{CL82} and define
\[
\Phi_n (x) := e\sp{\frac{ix}{n}} u(x).
\]
There holds
\[
\mathrm{dist}(\Phi_n,M)\rightarrow 0.
\]
From \eqref{eq.33}, we have
\[
U_t (\Phi_n)(x) = e\sp{i\left(\frac{nx - t}{n\sp 2}\right)} u(x - 2t/n).
\]
We set $ t_n := -yn/2 $ and obtain
\[
U_{t_n}(\Phi_n)(x) = e\sp{-iy/2n} e\sp{ix/n} u(x + y).
\]
Since $ M $ is invariant for the action of $ S\sp 1 $,
\[
\mathrm{dist}(U_{t_n}(\Phi_n),M) = \mathrm{dist}(e\sp{ix/n} u(\cdot + y),M)
\]
Since
\[
\mathrm{dist}(e\sp{ix/n} u(\cdot + y),u(\cdot + y))\rightarrow 0
\]
we have
\[
\mathrm{dist}(e\sp{ix/n} u(\cdot + y),M) =
\mathrm{dist}(u(\cdot + y),M) + o(1).
\]
Since $ M $ is stable and closed, $ u(\cdot + y)\in M $.
Therefore, whenever $ u_i\in M $, we have $ G_\lambda (u_i)\subseteq M $.
Then
\[
M = \bigcup_{u\in M} G_\lambda (u).
\]
In particular, among the closed, invariant and stable subsets of
$ G_\lambda $, $ G_\lambda (u) $ is minimal.
\end{proof}
\end{TAKEOUT}
\section{Uniqueness}
\label{sec.uniqueness}
We assume that (G1-5) hold. We fix $ \lambda > 0 $.
\defProof{Proof~of~Theorem~\ref{thm.uniqueness}}
\begin{proof}
Let $ R_0 $ and $ R_1 $ be two positive functions in $ \mathcal{G}_\lambda\cap H^1 _r $.
The set $ \mathscr{A} $ introduced in \eqref{G5} is the set of $ \omega $
such that a solution to \eqref{eq.E} exists. If $ \mathscr{A} $ is
connected, then
the function $ R_* $ defined on $ (\omega_0 - \varepsilon_0,\omega_0 + \varepsilon_0) $,
in \eqref{eq.ND26}, can be extended to $ \mathscr{A} $, so the function
$ \lambda $.
Let $ \omega_1 $ be the Lagrange multiplier associated to $ R_1 $.
Then $ \omega_1\in A $. Since $ R_0 $ and $ R_1 $ belong to the same
constraint,
\[
\lambda(\omega_0) = \lambda(\omega_1).
\]
By Lemma~\ref{lem.non-degeneracy}, $ \lambda'\geq 0 $ on $ [\omega_0,\omega_1] $.
Then $ \lambda'(\omega) = 0 $ on the whole interval. Then
$ \lambda'(\omega_0) = 0 $ giving a contradiction with
Lemma~\ref{lem.non-degeneracy}. Hence $ \omega_0 = \omega_1 $ and $ R_0 $
and $ R_1 $ solve the same initial value problem \eqref{eq.IVP}. Then
$ R_0 = R_1 =: R_+ $. The other solution is $ R_- := -R_+ $.
\end{proof}
\defProof{Proof}
\begin{corollary}
\label{cor.uniqueness}
If (G1-5) hold, then $ \mathcal{G}_\lambda = \mathcal{G}_\lambda (u) $ for every $ u\in\mathcal{G}_\lambda $.
\end{corollary}
\begin{proof}
We prove that an arbitrary $ v\in\mathcal{G}_\lambda $ belongs to $ \mathcal{G}_\lambda $. In fact,
by \eqref{lem.minima.4} of Lemma~\ref{lem.minima},
there are two complex numbers $ z,w\in\mathbb{C} $ such that
$ |z| = |w| = 1 $ and
\[
v(x) = zR_1 (x),\quad u(x) = w R_2 (x)
\]
where $ R_1,R_2\in\mathcal{G}_\lambda\cap H^1 $ and symmetric with respect to two points
$ y_1 $ and $ y_2 $, respectively, by \eqref{lem.minima.2} of
Lemma~\ref{lem.minima}. Then
$ R_1 (\cdot - y_1) $ and $ R_2 (\cdot - y_2) $ are two positive solutions
in $ \mathcal{G}_\lambda\cap H^1 _r $. By Theorem~\ref{thm.uniqueness},
\[
R_1 (\cdot - y_1) = R_2 (\cdot - y_2).
\]
Then
\[
\begin{split}
v(x) &= zR_1 (x) = zR_2 (x - y_2 + y_1) = w^{-1} z w R_2 (x - y_2 + y_1) \\
&= w^{-1} z u(x - y)
\end{split}
\]
where $ y := y_1 - y_2 $. Then $ v\in\mathcal{G}_\lambda (u) $.
\end{proof}
\section{The combined power-type case}
\label{sec.cp}
An example of non-linearity $ G $ satisfying all the assumptions
(G1-G5) is
\[
G(s) := - a |s|^p + b |s|^q,\quad 2 < p < 6,\quad 2 < q
\]
with $ c_1,c_2 > 0 $. Regularity and power-type estimate assumptions contained
in \eqref{G1}, \eqref{G2} and \eqref{G4}.
\eqref{G3} is satisfied.
Let $ s_0 > 0 $ be the zero of the function $ Q_\omega $ such that
$ Q_\omega (s_0) = 0 $ and $ Q_\omega '(s_0) < 0 $.
First, we prove that $ L(s_0)\geq 0 $. In fact,
the two conditions on $ s_0 $ give
\[
\omega - c_1 s_0^{p - 2} + c_2 s_0 ^{q - 2} = 0,\quad
2\omega - p c_1 s_0^{p - 2} + q c_2 s_0 ^{q - 2} < 0.
\]
A substitution yields
\[
c_1 (p - 2) s_0 ^{p - 2} - c_2 (q - 2) s_0 ^{q - 2} > 0.
\]
Then
\[
\begin{split}
L(s_0) &= c_1 (p - 2) (6 - p) s_0 ^p - c_2 (q - 2)(6 - q) s_0 ^q \\
&> c_2 (q - 2)(6 - p) s_0 ^q - c_2 (q - 2)(6 - q) s_0 ^q \\
&= c_2 (q - 2) (q - p) s_0 ^{q - 2} > 0.
\end{split}
\]
We can show that $ L $ is non-negative on the interval
$ (0,R_* (\omega)] $. Let $ s < R_* (\omega) $ be such that
$ L(s) < 0 $. Then
$ L(R_* (\omega)) < 0 $, because $ L $ has only one zero on $ (0,+\infty) $.
By definition of $ R_* (\omega) $, we have $ Q_\omega (R_*(\omega)) = 0 $ and
$ Q_\omega ' (R_* (\omega)) < 0 $,
which implies $ L(R_* (\omega)) > 0 $ and gives a contradiction.
\eqref{G5} is satisfied. Let $ s_1 $ be the unique local maximum of $ G $.
Then $ A = H((0,s_1)) $, thus connected. Therefore, from
Theorem~\ref{thm.uniqueness}, when $ G $ is a combined power pure-power
non-linearity, there exists only one positive function
$ R\in\mathcal{G}_\lambda\cap H^1 _r $.
\section*{Appendix}
We show that a function satisfied a combined power-type estimate
can be written as sum of two functions satisfying a power-type
estimate. As a consequence, we can suppose that $ G $ satisfies
\begin{equation}
\label{eq.g2}
|G'(s)|\leq c|s|\sp{p - 1},\quad G(0) = 0
\end{equation}
in place of \eqref{G2}, and
\begin{equation}
\label{eq.g4}
|G''(s)|\leq c|s|\sp{p - 2}
\end{equation}
in place of \eqref{G4}.
\begin{proposition}
\label{prop.split}
Let $ G $ be a function satisfying \eqref{G2}.
Then, there are two functions $ C^1 $ functions $ G_1 $ and $ G_2 $
and $ c\in\mathbb{R} $ such that
\[
|G_1 '(s)|\leq c |s|\sp{p - 1},\quad |G_2 '(s)|\leq c|s|\sp{q - 1},\quad
G = G_1 + G_2.
\]
If $ G $ satisfies \eqref{G4} as well, then $ G_1 $, $ G_2 $ and $ c $
can be chosen in such a way that the inequalities
\[
|G_1 '' (s)|\leq c |s|\sp{p - 2},\quad |G_2 ''(s)|\leq c|s|\sp{q - 2}.
\]
are also satisfied.
\end{proposition}
\begin{proof}
In both cases, the function can be obtained as follows: we consider a
non-negative continuous function $ \sigma $ such that
\[
\sigma =
\begin{cases}
1 & \text{on } [-1,1]\\
0 & \text{on } (-\infty,-2]\cup [2,+\infty)
\end{cases}
\]
and $ 0\leq\sigma\leq 1 $, $ |\sigma'|\leq 2 $. Then we choose
$ G_1 $ and $ G_2 $ as the unique functions such that
$ G_1 ' = \sigma G' $, $ G_2 ' = (1 - \sigma) G' $ and
$ c = 2\sp{q - p + 1} C $. If \eqref{G4} holds,
$ G_1 '' = \sigma G'' $ and $ G_2 '' = (1 - \sigma) G'' $, while $ c $
does not change.
\end{proof}
The next proposition is about the regularity of $ E $.
The gradient part of $ E $ is smooth; the regularity
of the non-linear part it is obtained with the same techniques
used by A.~Ambrosetti and G.~Prodi in \cite[Theorem~2.2]{AP93}.
We include the details of this proof in view of slight differences
with the quoted reference, where $ \mathbb{R} $ is replaced
by a bounded domain $ \Omega $, and the class of regularity
$ C^2(H\sp 1 (\RN;\C),\mathbb{R}) $ is replaced by $ C(L^p (\Omega),L^q (\Omega)) $.
The regularity of $ E $ depends on the regularity of $ F $ and the
power-type estimates which, in turn, are related to the estimates
of $ G $ \eqref{eq.g2} and \eqref{eq.g4}.
We identify $ \mathbb{C} $ with $ \mathbb{R}\sp 2 $.
If $ G $ is derivable, then
\begin{gather*}
F(s) = G(|s|)\\
F'(s) = G'(|s|)\frac{s}{|s|},\quad F'(0) = 0
\end{gather*}
for every $ s\in\mathbb{C} - \{0\} $. If $ G $ is two-times derivable,
\begin{gather*}
\partial_{ij} ^2 F(s) = G''(|s|)\frac{s_i s_j}{|s|^2} + G'(|s|)
\frac{\delta_{ij} |s|^2 - s_i s_j}{|s|^3},\quad
\partial^2 _{ij} F(0) = 0
\end{gather*}
Moreover,
\[
\begin{split}
(\partial_{ij} ^2 F)(s)^2 &= G''(s)^2 \frac{(s_i s_j)^2}{|s|^4}
+ G'(s)^2 \left(\frac{\delta_{ij} |s|^2 - s_i s_j}{|s|^3}\right)^2\\
&+ 2G''(s) G'(s) \frac{s_i s_j(\delta_{ij} |s|^2 - s_j s_j)}{|s|^5}.
\end{split}
\]
By inspection,
\begin{gather*}
\sum_{i,j} \frac{(s_i s_j)^2}{|s|^4} = 1,\quad
\sum_{i,j} \left(\frac{\delta_{ij} |s|^2 - s_i s_j}{|s|^3}\right)^2
= \frac{1}{|s|^2}\\
\sum_{i,j} \frac{s_i s_j(\delta_{ij} |s|^2 - s_j s_j)}{|s|^5}
= 0.
\end{gather*}
Therefore
\begin{gather}
\label{eq.f2}
\no{F''(s)}_{\mathcal{L}(\mathbb{R}^2)} ^2\leq G''(s)^2 +
\frac{G'(s)^2}{|s|^2}\leq 2C^2 |s|\sp{2(p - 2)}\\
\label{eq.f4}
|F'(s)|\leq C|s|^{p - 1}
\end{gather}
In the next proposition, the inequality
\begin{equation}
\label{eq.2}
(a + b)\sp c\leq 2\sp{c - 1}(a^c + b^c),\quad a,b\geq 0
\end{equation}
will be applied several times with different positive
values of $ c $.
\begin{proposition}
\label{prop.regular}
Let $ G $ be such that \eqref{eq.g2} is satisfied. Then $ E $ is differentiable.
If \eqref{eq.g4} holds too, then $ E $ is two-times differentiable.
\end{proposition}
\begin{proof}
(i). We will use the notation $ E $ for the non-linear part
$ \int_{\mathbb{R}\spn} F(u)dx $.
We have to prove that
\[
E(u_0 + h) - E(u_0) - \int_{\mathbb{R}\spn} F'(u_0)\cdot h = o(\no{h}_{H^1}).
\]
In fact, the term on the left is
\[
A := \int_{\mathbb{R}\spn} \int_0\sp 1 (F'(u_0 + th) - F'(u_0))\cdot h\,dt dx.
\]
We set
\[
h^*(x) := \int_0\sp 1 |F'(u_0(x) + th(x)) - F'(u_0(x))|dt
\]
From \eqref{eq.f2} and \eqref{eq.2}
there exists $ C_1 = C_1(C,p) $ such that
\[
|F'(u_0 + th) - F'(u_0)|\leq
C_1(|u_0|\sp{p - 1} + t\sp{p - 1} |h|\sp{p - 1}).
\]
After the integration on the interval $ [0,1] $, we obtain
\[
h^* (x)\sp{p/(p - 1)}\leq C_2\left(|u_0(x)|\sp p +
|h(x)|\sp p\right).
\]
We prove that $ A = o(\no{h}) $ with a contradiction
argument. If it is false, there exists $ \varepsilon_0 > 0 $
and a sequence $ (h_n) $ converging to zero in $ H^1 $
such that
\begin{equation}
\label{eq.1}
|A|\geq\varepsilon_0 \no{h_n}.
\end{equation}
Up to extract a subsequence, we can suppose that $ |h_n| $
converges to zero pointwise and it is dominated by an $ H^1 $
function $ h_0 $. Then
\[
h^* _n (x)\sp{p/(p - 1)}\leq C_2 (|u_0 (x)|\sp p +
h_0 (x)\sp p).
\]
We have a dominated sequence in $ L^1 (\mathbb{R}) $ converging
pointwise a.e. to zero. Therefore,
\[
\int_{\mathbb{R}\spn} |h^* _n (x)|\sp{p/(p - 1)} dx\to 0.
\]
From the H\"older inequality (with exponents $ p $ and
$ p/(p - 1) $ and functions $ h_n $ and $ h_n * $), and
\eqref{eq.SGN},
\[
|A|\leq C_2\no{h_n}_p \int_{\mathbb{R}\spn} |h^* _n (x)|\sp{p/(p - 1)}\leq
C_3 \no{h_n}_{H^1} \int_{\mathbb{R}\spn} |h^* _n (x)|\sp{p/(p - 1)}.
\]
Then $ |A|\leq \no{h_n} o(1) $, giving a contradiction
with \eqref{eq.1}. Therefore, $ E $ is differentiable and
\[
\langle dE(u_0),k\rangle = \int_{\mathbb{R}\spn} F'(u_0 (x))\cdot k(x)dx.
\]
(ii). If \eqref{eq.g2} and \eqref{eq.g4} hold, then $ E $ is two-times
differentiable. We set
\[
B := dE(u_0 + h) - dE(u_0) - \int_{\mathbb{R}\spn} F''(u_0(x)) h(x) dx
\]
and prove that $ B = o(\no{h}_{H^1}) $. In fact,
\[
B =
\int_{\mathbb{R}\spn}\int_0 ^1 \big(F''(u_0 (x) + th(x)) - F''(u_0(x))\big)
h(x) dt.
\]
Let $ k $ be an arbitrary vector of $ H^1 $. Then
\[
|\langle B,k\rangle|\leq \int_{\mathbb{R}\spn} |h(x)| |k(x)| h^{**} (x) dx
\]
where
\[
h^{**} (x) := \int_0 ^1 \no{F''(u_0 (x) + th(x)) - F''(u_0(x))}%
_{\mathcal{L}(\mathbb{R}^2)} dt.
\]
From \eqref{eq.f4} and \eqref{eq.2}, there
exists $ C_1 ' = C_1' (C,p) $ such that
\[
\no{F''(u_0 + th) - F''(u_0)}_{\mathcal{L}(\mathbb{R}^2)}
\leq C_1' (|u_0(x)|^{p - 2} + t^{p - 2} |h(x)|\sp{p - 2}).
\]
After the integration on $ [0,1] $, from \eqref{eq.2} we obtain
\[
h^{**} (x)\sp{p/(p - 2)}\leq C_2 '
(|u_0(x)|^p + |h(x)|\sp p).
\]
Now, suppose that there exists $ \varepsilon_0 > 0 $ and
a dominated sequence $ (h_n) $ converging in $ H^1 $,
pointwise a.e. to zero, such that
\begin{equation}
\label{eq.3}
\no{B}_{\mathcal{L}(H^1,\mathbb{R})}\geq\varepsilon_0\no{h_n}_{H^1}.
\end{equation}
By the H\"older inequality and \eqref{eq.SGN}
\[
|\langle B,k\rangle|\leq C_2'
\no{h_n}_p \no{k}_p \no{h_n ^{**}}_{p/(p - 2)}\leq
C_3' \no{h_n}\no{k}\no{h_n ^{**}}_{p/(p - 2)}
\]
for every $ k\in H^1 $. Then,
\[
\no{B}_{\mathcal{L}(H^1,\mathbb{R})}\leq C_3'
\no{h_n}\no{h_n ^{**}}_{p/(p - 2)} = \no{h_n} o(1)
\]
which gives a contradiction with \eqref{eq.3}.
\end{proof}
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'${$'$} \def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Glueballs in top-down holographic QCD}
Glueballs, color-neutral bound states of gluons only, are a cornerstone prediction of QCD.
They appear as additional states in the meson spectrum usually made of quark-antiquark pairs, but
because of possible mixing of flavorless $q\bar q$ states and glueballs, the experimental
status of the latter is still very unclear.\cite{Crede:2008vw}
In quenched lattice QCD, the glueball spectrum is fairly well established,\cite{Morningstar:1999rf}
with a rather mild dependence on the number of colors~\cite{Lucini:2012gg}. Moreover,
a recent study~\cite{Gregory:2012hu} points to only comparatively small changes
when dynamical quarks are included. This seems to suggest that glueballs in real QCD may be approached
from the large-$N_c$ limit, where glueballs are parametrically narrow and mixing is suppressed.
Gauge/gravity duality offers a possible avenue towards QCD in the large-$N_c$ limit. In particular,
the Witten-Sakai-Sugimoto model~\cite{Witten:1998zw,Sakai:2004cn,Sakai:2005yt}
is an almost parameter-free string-theoretic construction that
is connected to the low-energy limit of large-$N_c$ QCD with massless quarks, albeit the limitations
of the supergravity approximation make it prohibitively difficult to systematically improve this model.
Nevertheless, the Witten-Sakai-Sugimoto model has not only been found to successfully reproduce
several qualitative features of low-energy QCD, but frequently gives quantitative predictions within
10-30\% of experimental results.\cite{Rebhan:2014rxa}
In this model, the confinement scale is related to a Kaluza-Klein scale $M_{\rm KK}$ where a five-dimensional
supersymmetric Yang-Mills theory is dimensionally reduced to non-supersymmetric pure Yang-Mills theory.
Flavor D8 branes which accommodate chiral quarks in the fundamental representation are asymptotically
separated in the extra spatial dimension, but they connect in the bulk, thereby realizing chiral symmetry
breaking. Vector mesons appear as flavor gauge field modes on the D8 branes, relating $M_{\rm KK}$ to the mass of the $\rho$ meson,
while the 't Hooft coupling at this scale is related to the pion decay constant. This gives~\cite{Sakai:2004cn,Sakai:2005yt} $M_{\rm KK}=949$ MeV
and $\lambda\approx 16.63$. Matching alternatively~\cite{Rebhan:2014rxa}
to the large-$N_c$ lattice result of the ratio of string tension to $m_\rho^2$
gives a lower value of $\lambda\approx 12.55$.
Encouragingly, with this range of parameters, the experimental decay rate of $\rho$ and $\omega$ mesons is
covered by the result from the Witten-Sakai-Sugimoto\ model, which is of order $\lambda^{-1}N_c^{-1}$ and $\lambda^{-4}N_c^{-2}$, respectively.\cite{Sakai:2005yt}
The model also provides a Witten-Veneziano mass for the would-be U(1)$_\mathrm{A}$ Goldstone boson,\cite{Sakai:2004cn} which
combined with explicit quark masses yields masses for $\eta$ and $\eta'$ mesons that deviate only by $\sim 10\%$
from experiment.\cite{Brunner:2015oga}
The spectrum of glueballs in this model has been worked out already
2000 by Brower, Mathur, and Tan\,\cite{Brower:2000rp}. The lightest (scalar) state turns out to
correspond to an ``exotic'' graviton polarization\,\cite{Constable:1999gb}, followed
by a (mostly) dilatonic scalar and tensor mode which are degenerate in mass.
This appears to resemble the spectrum observed in lattice QCD, where the lightest glueball is
significantly lighter than the tensor glueball. However, with $M_{\rm KK}$ fixed by the $\rho$ meson
of the Sakai-Sugimoto construction, the lightest (``exotic'') scalar glueball comes out at a mere
$M_E\approx 855$ MeV, whereas the mostly dilatonic scalar (and the tensor) mode has $M_D=M_T=1487$ MeV, which
is close to lattice results for the scalar glueball, but too light for the tensor.
Decay rates for the holographic glueballs have been first considered by Hashimoto, Tan, and Terashima\,\cite{Hashimoto:2007ze}.
Brünner, Parganlija, and myself~\cite{Brunner:2015oqa} have revisited these calculations and extended them beyond the lowest
(exotic) mode. Surprisingly enough, the latter turns out to be much broader than the heavier dilatonic scalar,
which adds to the suspicion\,\cite{Constable:1999gb} that the former might have to be discarded from the model.\footnote{At
best, the exotic scalar glueball could play a role as a broad glueball component of the $\sigma$ meson
as proposed by Narison\,\cite{Narison:1996fm} (albeit the $\sigma$ meson itself is absent in the Witten-Sakai-Sugimoto model).}
The predominantly dilatonic scalar is indeed more similar to how a scalar glueball is usually represented
in gauge/gravity duality. The result for its decay rate into (massless) pions reads
\begin{equation}\label{GDpipi}
\Gamma(G_D\to\pi\pi)/M_D\approx{1.359}/{\lambda N_c^2}\approx0.009\ldots0.012.
\end{equation}
This is smaller than the experimental value $0.025(3)$ for the glueball candidates $f_0(1500)$ whose mass would
match almost perfectly with $M_D$. Moreover, this isoscalar $0^{++}$ meson decays predominantly into four pions,
whereas decay into four pions is very strongly suppressed in the holographic model.
Indeed, phenomenological models that favor $f_0(1500)$ as a glueball candidate\,\cite{Amsler:1995td}
do so with rather large mixing with $q\bar q$ states, contrary to the working hypothesis of the holographic approach.
In fact, recently attention has turned increasingly to the meson $f_0(1710)$
as a glueball candidate.\cite{Lee:1999kv,Janowski:2014ppa,Cheng:2015iaa}
Its (not officially established) decay rate into pions indeed seems to match the result (\ref{GDpipi}).
The main difficulty in accepting $f_0(1710)$ as glueball candidate is however the strong flavor asymmetry in its decay
pattern where decays into kaons dominate. To explain this, ``chiral suppression'' of scalar glueball decay
has been proposed\,\cite{Chanowitz:2005du}, but the underlying perturbative reasoning appears questionable\,\cite{Frere:2015xxa}.
In the holographic setup, explicit quark mass terms almost unavoidably lead to additional vertices between
glueballs and pseudoscalar mesons. Brünner and myself~\cite{Brunner:2015yha,Brunner:2015oga} have
studied this question in the Witten-Sakai-Sugimoto model when it is deformed by explicit mass terms
and we have found that when the explicit mass terms couple in the same manner as the Witten-Veneziano mass term
the resulting decay pattern indeed reproduces that of $f_0(1710)$ surprisingly well, see Table \ref{tabrates}.
Here we have extrapolated the mass of the glueball to the observed mass of $f_0(1710)$
in a manner which keeps the dimensionless chiral result (\ref{GDpipi}) fixed. However, because the mass of $f_0(1710)$
is above the threshold of $2\rho$ and $2\omega$ mesons, the decay into four and six pions is no longer suppressed.
Together with limits on the rate of decays into $\eta\eta'$ pairs~\cite{Brunner:2015oga}, this represents
a falsifiable prediction of the holographic model.
\def\,\cite{PDG15}{\,\cite{PDG15}}
\def\,\cite{Brunner:2015oqa}{\,\cite{Brunner:2015oqa}}
\def\,\cite{Brunner:2015yha}{\,\cite{Brunner:2015yha}}
\def\,\cite{Brunner:2015oga}{\,\cite{Brunner:2015oga}}
\def\,\cite{Brunner:2015yha,Brunner:2015oga}{\,\cite{Brunner:2015yha,Brunner:2015oga}}
\begin{table}[p]
\caption[]{Experimental data from the Particle Data Group\,\cite{PDG15}\ for the decay rates of the glueball candidates $f_0(1500)$ and
$f_0(1710)$ [in gray color: the latter combined
with the branching ratio\,\cite{Albaladejo:2008qa} ${\rm Br}(f_0(1710)\to KK)=0.36(12)$]
compared to the results obtained for a pure glueball $G_D$ of same mass in the chiral Witten-Sakai-Sugimoto (WSS) model\,\cite{Brunner:2015oqa}\
and in a deformed model with finite quark masses\,\cite{Brunner:2015yha}.
Red color indicates a significant discrepancy with a pure-glueball interpretation of the respective $f_0$ meson. \newline
}
\label{tabrates}
\begin{center}
\begin{tabular}{|l|r|r|r|}
\hline
decay & $\Gamma/M$ {(exp.\protect\,\cite{PDG15})} & (WSS chiral\,\protect\cite{Brunner:2015oqa}) & (WSS massive\protect\,\cite{Brunner:2015yha})\\
\hline
$f_0(1500)$ (total) & 0.072(5) & \color{red} 0.027\ldots0.03
& 0.057\ldots0.07
\\
$f_0(1500)\to4\pi$ & 0.036(3) & \color{red} 0.003\ldots 0.005 &
\color{red} 0.003\ldots 0.005 \\
$f_0(1500)\to2\pi$ & 0.025(2) & \color{red} 0.009\ldots0.012 & \color{red} 0.010\ldots0.014\\
$f_0(1500)\to 2K$ & 0.006(1) & \color{red} 0.012\ldots0.016 & \color{red} 0.034\ldots0.045\\
$f_0(1500)\to 2\eta$ & 0.004(1) & 0.003\ldots0.004 & \color{red} 0.010\ldots0.013\\
$f_0(1500)\to \eta\eta'$ & 0.0014(6) & \color{red} 0 & $(^{*})$ 0\\
\hline
$f_0(1710)$ (total) & 0.081(5)
& 0.059\ldots0.076 & 0.083\ldots0.106 \\
$f_0(1710)\to 2K$ &
\color{gray} 0.029(10)
& \color{red} 0.012\ldots0.016 & 0.029\ldots0.038 \\[8pt]
$f_0(1710)\to 2\eta$ & \color{gray} 0.014(6)
& \color{red} 0.003\ldots0.004 & 0.009\ldots0.011 \\[8pt]
$f_0(1710)\to2\pi$ & \color{gray} 0.012($+5\atop-6$)
& 0.009\ldots0.012 & 0.010\ldots0.013 \\[8pt]
$f_0(1710)\to2\rho,\rho\pi\pi\to4\pi$ & ? & 0.024\ldots 0.030 & 0.024\ldots 0.030 \\
$f_0(1710)\to2\omega$ & \color{gray} 0.010($+6\atop-7$)
& 0.011\ldots 0.014 & 0.011\ldots 0.014 \\
$f_0(1710)\to \eta\eta'$ & ? & 0 & $(^{*})$ 0\\
\hline
\end{tabular}
\end{center
\footnotesize $(^{*})$ If one relaxes\,\cite{Brunner:2015oga}\ the assumption\,\cite{Brunner:2015yha}\ of a universal coupling of all pseudoscalar mass terms, nonzero $\eta\eta'$ rates are possible in the WSS model with finite quark masses. In the case of $f_0(1710)$ an upper limit of $\Gamma(\eta\eta')/\Gamma(\pi\pi)\lesssim 0.04$ (i.e.\
$\Gamma(\eta\eta')/M \lesssim 0.0005$) is obtained, if one requires that
prediction for $\Gamma(\pi\pi)/\Gamma(KK)$ remains within the current experimental error bar.\,\cite{Brunner:2015oga}
\end{table}
\begin{table}[p]
\caption[]{Extrapolation of tensor glueball decay\,\cite{Brunner:2015oqa}\ with glueball mass
$M=M_T=M_D$ and when the latter is raised to the mass of the tensor glueball candidate $f_2(1950)$, whose experimental width
is 472(18) MeV [$\Gamma/M=0.24(1)$], or
a typical lattice prediction $\sim 2.4 \,{\rm GeV}$, for which the extrapolation from the WSS model predicts an extremely broad tensor glueball.
}
\label{tabtensorextrapolmpi}
\vspace{0.4cm}
\begin{center}
\begin{tabular}{|l|c|c|}
\hline
decay & $M$ & $\Gamma/M$ \\
\hline
$T\to 2\pi$ & 1487 & 0.013\ldots0.018 \\
$T\to 2K$ & 1487 & 0.004\ldots0.006 \\
$T\to 2\eta$ & 1487 & 0.0005\ldots0.0007 \\
$T$ (total) & 1487 & $\approx 0.02\ldots0.03$\\
\hline
$T\to 2\rho\to 4\pi$ & 1944 & 0.129\ldots 0.171 \\ %
$T\to 2K^*\to 2(K\pi)$ & 1944 & 0.105\ldots 0.157 \\
$T\to 2\omega\to 6\pi$ & 1944 & 0.043\ldots 0.057 \\
$T\to 2\pi$ & 1944 & 0.014\ldots0.018 \\
$T\to 2K$ & 1944 & 0.009\ldots0.012 \\
$T\to 2\eta$ & 1944 & 0.0017\ldots0.0022 \\
$T$ (total) & 1944 & $\approx 0.30\ldots0.42$\\
\hline
$T\to 2K^*\to 2(K\pi)$ & 2400 & 0.173\ldots 0.250 \\
$T\to 2\rho\to 4\pi$ & 2400 & 0.159\ldots 0.211 \\ %
$T\to 2\omega\to 6\pi$ & 2400 & 0.053\ldots 0.070 \\
$T\to 2\phi$ & 2400 & 0.032\ldots 0.051 \\
$T\to 2\pi$ & 2400 & 0.014\ldots0.019 \\
$T\to 2K$ & 2400 & 0.012\ldots0.016 \\
$T\to 2\eta$ & 2400 & 0.0025\ldots0.0034 \\
$T\to 2\eta'$ & 2400 & 0.0004\ldots0.0005 \\
$T$ (total) & 2400 & $\approx 0.45\ldots0.62$\\
\hline
\end{tabular}
\end{center}
\end{table}
The Witten-Sakai-Sugimoto model also predicts the decay pattern of tensor glueballs,\cite{Brunner:2015oqa}
which should in fact be
less sensitive to explicit quark mass terms.
However, the lattice prediction of the tensor glueball mass at around 2.4 GeV is very much above
the holographic result. Extrapolating the latter to such high masses, again by keeping
the dimensionless chiral result of decay into pairs of pseudoscalar mesons fixed,
now gives a very large contribution from decays into $2K^*$, $2\rho$, $2\omega$, and $2\phi$ vector mesons
as shown in Table \ref{tabtensorextrapolmpi}. This suggests that tensor glueballs in this mass
range may be too broad to be observable at all.
A comparatively broad glueball candidate is provided by the isoscalar $f_2(1950)$ which has $\Gamma/M\approx 0.24(1)$.
With the corresponding mass parameter, the holographic result of 0.3\ldots0.42 appears to be at least marginally comparable.
A more promising case seems to be that of pseudoscalar glueballs. According to the Witten-Sakai-Sugimoto model,
these glueballs should be rather narrow states.\cite{BRinprep}
\section*{Acknowledgments}
The research reported here has been obtained in collaboration with
F.\ Br\"unner and D.\ Parganlija, and has been funded by the Austrian Science Funds through
FWF project no.\ P26366, and the FWF doctoral program no.\ W1252.
\section*{References}
|
\section*{Introduction}
The classical isoperimetric property of a circle in the two-dimensional space of constant curvature equal to $c$ claims that among all simple closed curves of a fixed length the maximal area is enclosed only by a circle. This property can be reformulated in an equivalent way in the form of an isoperimetric inequality. For an arbitrary simple closed curve of the length $L$ that encloses the domain of the area $A$ the following inequality holds (see, for example,~\cite{BZ})
\begin{equation}
\label{isoineq}
L^2 - 4\pi A + c A^2 \geqslant 0,
\end{equation}
and the equality is attained only by circles.
Inequality~(\ref{isoineq}) gives a sharp upper bound on the area of the domain bounded by a curve provided that its length is fixed. At the same time, there exist simple closed curves that bound domains whose areas are arbitrary close to zero.
Therefore, in order to obtain some meaningful estimates from below of the bounded area we need to further restrict the class of curves under consideration. One of the natural ways to do this is to consider curves of bounded curvature. Such class appeared in a number of extremal problems (see, for example,~\cite{AB,BorDr1,BorDr2,HTr,H,M}, and also~\cite{PZh}). In particular, in~\cite{HTr} the authors gave a bound on the area of domains enclosed by closed embedded plane curves of the fixed lengths whose curvatures $k$ satisfy the inequality $|k| \leqslant 1$ and the lengths satisfy some additional restrictions. Another type of bounds can be obtained if we consider the class of curves of curvatures bounded from below. In~\cite{BorDr1} for closed embedded plane curves whose curvatures $k$ in the generalized sense (see the definition below) satisfy the inequality $k \geqslant \lambda$ for some constant number $\lambda$ it was proved the following
\begin{theopargself}
\begin{theorem}[\cite{BorDr1}]
\label{ec2dth1}
Let $\gamma$ be a closed curve embedded in the Euclidean plane $\mathbb E^2$. If the curvature $k$ of $\gamma$ is bounded from below by some positive constant $\lambda$, then for the length $L$ of $\gamma$ and the area $A$ of the domain enclosed by $\gamma$ the following inequality holds:
\begin{equation}
\label{ec2dth1eq1}
A \geqslant \frac{L}{2\lambda} - \frac{1}{\lambda^2}\sin\frac{\lambda L}{2}.
\end{equation}
Moreover, the equality case holds only for a \textmd{lune}, that is a boundary curve of the intersection of two domains enclosed by circles of curvature equal to $\lambda$.
\end{theorem}
\end{theopargself}
In the present paper we generalize Theorem~\ref{ec2dth1} for curves lying on a two-dimensional sphere.
\section{Preliminaries and the main result}
Let $\mathbb S^2 (k_1^2)$ be a standard two-dimensional sphere of Gaussian curvature equal to $k_1^2$ with $k_1 > 0$. In order to state the main result we need the following
\begin{definition}
A locally convex curve $\gamma \subset \mathbb S^2(k_1^2)$ is called \textit{$\lambda$-convex} with $\lambda \geqslant 0$ if for every point $P \in \gamma$ there exists a curve $\mu_P$ of constant geodesic curvature equal to $\lambda$ passing through $P$ in such a way, that in a neighborhood of $P$ the curve $\gamma$ lies from the convex side of $\mu_P$.
\end{definition}
By the definition above, $0$-convex curves are just locally convex. If $\lambda>0$, then for every point on a curve there exists a locally supporting circle of curvature equal to $\lambda$.
We should also note that at $C^k$-regular points of $\gamma$ with $k\geqslant 2$ the condition of being $\lambda$-convex is equivalent to the condition that at such points the geodesic curvature $\kappa_g$ of $\gamma$ satisfies the inequality $\kappa_g \geqslant \lambda$. Hence the class of $\lambda$-convex curves is a non-regular extension for the class containing smooth curves of geodesic curvature bounded from below by $\lambda$.
It is known that a convex curve is twice continuously differentiable almost everywhere and thus its geodesic curvature is almost everywhere well-defined. Therefore, a convex curve is $\lambda$-convex if and only if the inequality $\kappa_g \geqslant \lambda$ is satisfied at all points where geodesic curvature is defined. Even more, the set of a curve's \textit{vertexes}, that is points at which left and right semi-tangents do not coincide, is no-more-than countable.
We recall that $\gamma \subset \mathbb S^2(k_1^2)$ is a \textit{closed embedded curve} if it is homeomorphic to a circle and has no self-intersections. Embedded curves may not be smooth in general.
Note that a closed embedded $\lambda$-convex curve on a sphere is globally convex. Furthermore, if $\lambda > 0$, then at any point of this curve there exist a locally supporting circle, and thus the curve is globally strictly convex.
Observe also that a geodesic digon being a $0$-convex lune can enclose a domain with area arbitrary close to zero. Thus in order to get non-trivial estimates for the enclosed area, everywhere below we will assume $\lambda > 0$.
\begin{definition}
\textit{$\lambda$-convex polygon} is a closed embedded $\lambda$-convex curve composed of circular arcs of geodesic curvature equal to $\lambda$.
\end{definition}
\begin{definition}
A $\lambda$-convex polygon composed of two circular arcs we will call a \textit{$\lambda$-convex lune} or simply a \textit{lune}.
\end{definition}
We are now ready to formulate the main result of the paper that is summarized in the following theorem.
\begin{theorem}
\label{mainth}
Let $\gamma$ be a closed embedded $\lambda$-convex curve lying on a two-dimensional sphere $\mathbb S^2 (k_1^2)$ of Gaussian curvature equal to $k_1^2$. If $L(\gamma)$ is the length of $\gamma$ and $A(\gamma)$ is the area of the domain enclosed by $\gamma$, then
\begin{equation}
\label{maintheq1}
A(\gamma) \geqslant \frac{4}{k_1^2} \arctan\left(\frac{\lambda}{\sqrt{\lambda^2+k_1^2}}\tan\left(\frac{\sqrt{\lambda^2+k_1^2}}{4}L(\gamma)\right)\right)-\frac{\lambda}{k_1^2} L(\gamma).
\end{equation}
Moreover, the equality case holds only for $\lambda$-convex lunes.
\end{theorem}
\begin{remark}
Observe that when $k_1 \to 0$ the right side of inequality~(\ref{maintheq1}) tends to the right side of inequality~(\ref{ec2dth1eq1}).
\end{remark}
It is important to note that the inequality in Theorem~\ref{mainth} express the \textit{isoperimetric property} of $\lambda$-convex lunes. To make the statement above precise, we need to reformulate Theorem~\ref{mainth} in the following equivalent way.
\begin{theorem}
\label{mainthalt}
Let $\gamma$ be a closed embedded $\lambda$-convex curve lying on a two-dimensional sphere $\mathbb S^2 (k_1^2)$. If $\gamma_0 \subset \mathbb S^2(k_1^2)$ is a $\lambda$-convex lune such that $$L(\gamma) = L(\gamma_0),$$
then $$A(\gamma) \geqslant A(\gamma_0)$$ and the equality case holds if and only if $\gamma$ is a $\lambda$-convex lune.
\end{theorem}
We will prove the main result in the form of Theorem~\ref{mainthalt} and after that show its equivalence to Theorem~\ref{mainth}.
\begin{remark}
The isoperimetric property of $\lambda$-convex curves, similar to the one proved for the Euclidean plane and the one that will be proved for a sphere, also holds for a two-dimensional Lobachevsky space $\mathbb H^2 (-k_1^2)$ for any $\lambda > 0$.
\end{remark}
\section{Proof of the main result}
\label{mainres}
The principal tool for proving the main result will be Pontryagin's Maximum Principle. Let us recall all necessary definitions and a statement for a special case of this principle adapted for our needs (for a general case see~\cite[\S 1.4]{MDmO}, or equivalent but ideologically different approach in~\cite[\S 5]{ArMIT}).
\begin{definition}
\label{contrpros}
A pair of functions $(x(t), u(t))$, defined as $(x(t), u(t)) \colon [t_0,t_1] \to \mathbb R^2 \times \mathbb R$ on some fixed segment $[t_0,t_1] \subset \mathbb R$ later denoted as $\Delta$, is a \textit{controlled process} if \textit{phase variable} $x(t)$, which is a pair of coordinate functions $(x_1(t),x_2(t))$, is absolutely continuous function on $\Delta$, a \textit{control} $u(t)$ is bounded measurable function on $\Delta$, and the pair $(x(t), u(t))$ satisfies a \textit{controlled system}
\begin{equation}
\label{pmpeq1}
\begin{aligned}
\dot x (t) &= f (x(t),u(t)) \\
u(t) & \in \Delta_u
\end{aligned}
\,\, \text{ a.e. on }\Delta,
\end{equation}
where $\Delta_u \subset \mathbb R$ is a fixed segment, the function $f \colon \mathbb R^2 \times \mathbb R \to \mathbb R^2$ and its derivatives $f_x$ are continuous with respect to all variables; by the dot sign, as usual, we denote a derivative with respect to the variable $t$.
\end{definition}
For controlled processes we pose a minimization problem for the following functional
\begin{equation}
\label{pmpeq2}
\mathcal J(x(\cdot),u(\cdot)) = \int\limits_{t_0}^{t_1} F_0 (x(t),u(t)) dt \to min,
\end{equation}
with an integral constraint
\begin{equation}
\label{pmpeq3}
\int \limits_{t_0}^{t_1} F_1 (x(t),u(t)) dt = A
\end{equation}
and periodic boundary conditions
\begin{equation}
\label{pmpeq4}
x(t_0) = x(t_1) = a,
\end{equation}
where the function $F_0 \colon \mathbb R^2 \times \mathbb R \to \mathbb R$ is continuously differentiable on its domain, the function $F_1 \colon \mathbb R^2 \times \mathbb R \to \mathbb R$ is continuous together with its derivatives $(F_1)_x$, and $A \in \mathbb R$, $a \in \mathbb R^2$ are some fixed constants.
A \textit{trajectory} ${\{(x(t), u(t)) \colon t \in \Delta\}}$ that corresponds the controlled process~(\ref{pmpeq1}) is \textit{admissible} for problem~(\ref{pmpeq2}) if for this trajectory constraints~(\ref{pmpeq3}),~(\ref{pmpeq4}) are satisfied. An admissible trajectory is \textit{optimal} if it minimizes the value of the functional $\mathcal J$ among all admissible trajectories.
Let us introduce a \textit{Pontryagin's function}
$$
\mathcal H(x,u,p,\lambda_0,\lambda_1) = p \cdot f(x,u) + \lambda_1 F_1 (x,u) - \lambda_0 F_0 (x,u).
$$
We will say that an admissible trajectory ${\{(x(t), u(t)) \colon t \in \Delta\}}$ for problem~(\ref{pmpeq2}) satisfies the~\textit{Pontryagin's maximum principle} if there are real numbers $\lambda_0, \lambda_1$ and absolutely continuous on $\Delta$ function $p(t) \colon \Delta \to \mathbb R^2$ such that for them the following conditions are met:
\begin{enumerate}
\item[(i)]
$\lambda_0 \geqslant 0$ (nonnegativity condition);
\item[(ii)]
$\lambda_0 + |\lambda_1| + |p(t)| \neq 0$ for all $t \in \Delta$ (non-triviality condition);
\item[(iii)]
$\dot p(t) = - \mathcal H'_x(x(t),u(t),p(t),\lambda_0,\lambda_1)$ a.e. on $\Delta$ (adjoint system);
\item[(iv)]
$\max\limits_{u \in \Delta_u} \mathcal H(x(t),u,p(t),\lambda_0,\lambda_1) = \mathcal H(x(t),u(t),p(t),\lambda_0,\lambda_1)$ for almost all $t \in \Delta$ (maximality condition).
\end{enumerate}
Finally, we have the following theorem.
\begin{pmp}
\label{pmp}
Let ${\{(\hat x(t), \hat u(t)) \colon t \in \Delta\}}$ be an optimal trajectory for problem~(\ref{pmpeq2}). Then it satisfies the Pontryagin's maximum principle.
\end{pmp}
Now we have the necessary tool to proceed with proving the main result of the paper.
In order to use Pontryagin's Maximum Principle we need to construct a controlled system~(\ref{pmpeq1}). For this purpose let us introduce a so-called support function. Let $O \in \mathbb S^2(k_1^2)$ be a point inside a convex domain bounded by the curve $\gamma$. Let us consider on $\mathbb S^2(k_1^2)$ the polar coordinate system with the origin at the point $O$ and with the angular parameter $\theta$ with $\theta \in [0, 2\pi)$. For each geodesic ray $OL$, emanating from $O$ and forming an angle $\theta$ with some fixed direction, let us consider a geodesic perpendicular to $OL$ and that is supporting for the curve $\gamma$ at some point $P$. By strict convexity of $\gamma$ such geodesic always exists, and the point $P$ is unique. Denote $h(\theta)$ to be the distance from the point $O$ to the supporting geodesic above, measured along the ray $OL$. The function $h(\theta) \colon [0, 2\pi) \to [0, \pi/k_1)$ is a \textit{support function} for the curve $\gamma$.
We should note here that a convex curve is uniquely determined by its support function.
The next proposition shows a relationship between a support function and the \textit{radius of curvature} of a curve $\gamma \subset \mathbb S^2(k_1^2)$ later denoted by $R$. We remind that, by definition, $R(\theta) = 1/k(\theta)$, where $k$ is the geodesic curvature of $\gamma$. For our curve we will call a \textit{contact radius of curvature} the following quantity
\begin{equation*}
g(\theta)=\frac{1}{k_1}\tan\left(k_1 h(\theta)\right).
\end{equation*}
\begin{remark}
\label{remcoordsys}
Observe that for $\lambda$-convex curves we can always choose the origin of the polar coordinate system in such a way, that $$h(\theta) \leqslant \frac{1}{k_1}\arccot \frac{\lambda}{k_1} < \frac{\pi}{2 k_1}$$ or, equivalently, $g (\theta) \leqslant {1}/{\lambda}$ for all $\theta \in [0, 2\pi)$. We assume that everywhere below a coordinate system is chosen in the described way.
\end{remark}
Using the notion of a contact radius of curvature for $\lambda$-convex curves on a sphere we can prove the following proposition.
\begin{proposition}
\label{propsupfunc}
Let $\gamma \subset \mathbb S^2(k_1^2)$ be a closed embedded $\lambda$-convex curve on the sphere. If $R(\theta)$ and $g(\theta)$ are respectively the radius and the contact radius of curvature for $\gamma$, then
\begin{equation}
\label{propsupfunceq1}
R = \frac{g'' + g}{\left(1+\frac{ k_1^2 {g'}^2}{1 + k_1^2 g^2}\right)^\frac{3}{2}} \text{ for almost all $\theta \in [0,2\pi]$}
\end{equation}
(here by the prime sign we denote a derivative with respect the variable $\theta$).
\end{proposition}
\begin{remark}
If in~(\ref{propsupfunceq1}) we take a limit with $k_1 \to 0$, we will get the classical formula $R = h'' + h$, which relates a support function and the radius of curvature of a plane curve.
\end{remark}
\begin{proof}
Without loss of generality we may assume $k_1 = 1$, and let us denote $\mathbb S^2(1)$ as $\mathbb S^2$.
We start with introducing the family $l_\gamma \subset \mathbb S^2$ of all supporting geodesic lines for $\gamma$. For each geodesic $l \in l_\gamma$ let $\xi(l)$ be an outward unit normal for $l$ with respect to the curve $\gamma$. Note that if at the point $l \cap \gamma$ the geodesic $l$ is tangent to $\gamma$, then $\xi(l)$ will also be an outward normal for $\gamma$.
Let us consider a map generated by unit normals $\xi(l)$ as follows:
\begin{equation}
\label{polarmap}
\begin{aligned}
\xi \colon &l_\gamma \to \mathbb S^{2}, \\
&l \mapsto \xi(l).
\end{aligned}
\end{equation}
This map is called the \textit{polar map} of the curve $\gamma$. The curve $\gamma^\ast \subset \mathbb S^{2}$ defined as $\gamma^\ast = \xi (l_\gamma)$ is the \textit{polar image} of $\gamma$ (or the \textit{dual curve}).
Since $\gamma$ is convex, its polar map is a bijection between the set $l_\gamma$ and the set of points on $\gamma^\ast$. Moreover, $(\gamma^\ast)^\ast = \gamma$.
It appears that defining a curve using a support function is equivalent to defining its dual curve. To explain this statement, let $(t, \theta)$ be the coordinates in our polar coordinate system. From the definition of the polar map~(\ref{polarmap}) it easily follows that, assuming both $\gamma$ and $\gamma^\ast$ lies on the same sphere, the polar image $\gamma^\ast$ will be given by the equation
\begin{equation}
\label{tfordual}
t = \frac{\pi}{2} + h(\theta).
\end{equation}
From~(\ref{tfordual}) we obtain that the polar image of a circle with the radius $r$ will be a circle with the radius $\frac{\pi}{2} - r$. Note that the initial circle and its polar image have mutually reciprocal geodesic curvatures.
Using the observation above, we can deduce one important property of a polar map~(\ref{polarmap}). Namely, if $\gamma$ is a closed embedded $\lambda$-convex curve, then $\gamma^\ast$ will be a closed embedded curve through each point $P^\ast$ of which passes a circle of the curvature ${1}/{\lambda}$ such that in some neighborhood of $P^\ast$ this circle lies inside the closed domain bounded by $\gamma^\ast$. This property instantly implies that the curve $\gamma^\ast$ is in fact $C^{1,1}$-smooth, which in turn, by using~(\ref{tfordual}), implies $C^{1,1}$-smoothness of the support function $h(\theta)$.
Hence $\gamma^\ast$ is almost everywhere $C^2$-smooth, and thus the geodesic curvature $\kappa_g^\ast$ of $\gamma^\ast$ is almost everywhere well-defined. Let us show that for all values $\theta$ from $[0, 2\pi)$, where $\kappa^\ast_g$ is defined (we assume $\kappa_g^\ast \geqslant 0$), it is equal to the radius of curvature $R(\theta)$ of the curve $\gamma$, that is
\begin{equation}
\label{polarcurvprop}
\kappa_g^\ast(\theta) = R(\theta) \text{ a.e. on } [0, 2\pi).
\end{equation}
For regular curves this fact is well-known, but by the lack of a direct reference we will prove it here for completeness of exposition.
We may assume that the curve $\gamma^\ast \subset \mathbb S^2$ with $\mathbb S^2 \subset \left(\mathbb E^3,\left<\cdot,\cdot\right> \right)$ is given in an arc-length parametrization $s^\ast$ by the unit radius-vector $r(s^\ast)$. Set $\xi^\ast$ to be an inward unit normal to $\gamma^\ast$.
Since all vectors under consideration are of the unit length, and since $\gamma^\ast$ lies on a sphere, we have
\begin{equation}
\label{vectid1}
\left<r',\xi^\ast\right> = \left<r,\xi^\ast\right> = 0,
\end{equation}
\begin{equation}
\label{vectid2}
\left<r', r\right> = \left<{\xi^\ast}',\xi^\ast\right>= 0,
\end{equation}
where the prime sign denotes a derivative with respect to $s^\ast$.
From~(\ref{vectid1}) we obtain
\begin{equation}
\label{vectnorm}
\left<r,{\xi^\ast}'\right> = 0.
\end{equation}
Thus the vector $r$ is in fact a unit normal to the polar curve $\gamma$, treated as a curve in $\mathbb S^2$, and besides that $\xi^\ast$ is the radius-vector for $\gamma$.
Let us pick a point on $\gamma^\ast$ at which the geodesic curvature $\kappa_g^\ast$ is defined. For the rest of the proof of~(\ref{polarcurvprop}) we compute at this point. By definition
\begin{equation}
\label{geocurv}
\kappa_g^\ast = \left<r'',\xi^\ast \right>,
\end{equation}
Hence, using~(\ref{vectid1}),
\begin{equation}
\label{firstterm1}
\left<r',{\xi^\ast}'\right> = -\kappa^\ast_g.
\end{equation}
From~(\ref{vectid2}) and~(\ref{vectnorm}) it follows that the vector ${\xi^\ast}'$ is collinear to the vector $r'$. Together with~(\ref{firstterm1}) this collinearity implies
\begin{equation}
\label{rodrig}
{\xi^\ast}' = -\kappa^\ast_g r'.
\end{equation}
If $s$ is the arc-length parameter on $\gamma$, then from~(\ref{rodrig}) we have
\begin{equation}
\label{secondterm}
\frac{ds}{ds^\ast} = \kappa_g^\ast, \,\,\, \frac{d\xi^\ast}{ds} = - r'.
\end{equation}
From the last relation it follows that the second derivative of $\xi^\ast$ is well-defined. Thus, using~(\ref{vectnorm}) and~(\ref{firstterm1}),
\begin{equation}
\label{firstterm2}
\left<{\xi^\ast}'',r\right>=\kappa^\ast_g.
\end{equation}
Finally, using~(\ref{firstterm1}) and~(\ref{secondterm}), with respect to the normal $r$ we compute
\begin{equation}
\label{geocurvpolar}
R = \left<\frac{d^2\xi^\ast}{{ds}^2}, r \right>^{-1} = \left<{\xi^\ast}'', r\right>^{-1}\left(\frac{ds}{ds^\ast}\right)^2 = \kappa_g^\ast,
\end{equation}
which proves~(\ref{polarcurvprop}).
Therefore, by property~(\ref{polarcurvprop}) in order to get formula~(\ref{propsupfunceq1}) it suffices to calculate the geodesic curvature of $\gamma^\ast$ given in a polar coordinate system with the metric $ds^2 = dt^2 + \sin^2 t \, d\theta^2$ by equation~(\ref{tfordual}). We should pay attention that, since the origin of the coordinate system and the polar curve lie in the distinct hemispheres, the geodesic curvature of $\gamma^\ast$ given by~(\ref{tfordual}) will be negative. Hence, assuming $\kappa_g^\ast \geqslant 0$, we should reverse the sign. After standard computations we will obtain
\begin{equation}
\label{propsupfunceq8}
R= \kappa^\ast_g = \frac{h'' \cos h + 2 h'^2 \sin h + \cos^2 (h) \sin h}{\left(h'^2 + \cos^2(h)\right)^{\frac{3}{2}}}.
\end{equation}
Substituting $h(\theta) = \arctan g(\theta)$ in~(\ref{propsupfunceq8}) we come to
$$
R = \frac{g'' + g}{\left(1 + \frac{g'^2}{1 + g^2}\right)^{\frac{3}{2}}},
$$
as required. Proposition~\ref{propsupfunc} is proved.
\qed
\end{proof}
In order to prove Theorem~\ref{mainthalt} we should fix the lengths of our curves and look for a minimum of the area of convex domains bounded by these curves. To formalize this problem, besides Proposition~\ref{propsupfunc}, we will need the expression for the length $L(\gamma)$ of a curve $\gamma$ and the expression for the area $A(\gamma)$ of the convex domain bounded by $\gamma$ in terms of its contact radius of curvature $g(\theta)$ and its radius of curvature $R(\theta)$. Without loss of generality we again may assume $k_1 = \lambda = 1$.
Using~(\ref{secondterm}),~(\ref{geocurvpolar}), and polar equation~(\ref{tfordual}) of the dual curve, we obtain
\begin{equation}
\label{mainproofeq0}
\frac{ds}{d\theta} = \frac{ds}{ds^\ast} \frac{ds^\ast}{d\theta} = R \sqrt{h'^2 + \cos^2(h)} = R \frac{\sqrt{1 + g^2 + {g'}^2}}{1 + g^2}.
\end{equation}
Therefore, the length of $\gamma$ can be written in terms of $g$ and $R$ as follows:
\begin{equation}
\label{mainproofeq1}
L(\gamma) = \int \limits_0^{2\pi} ds(\theta) = \int \limits_0^{2\pi} \frac{ds}{d\theta} d\theta = \int \limits_0^{2\pi} R \frac{\sqrt{1 + g^2 + {g'}^2}}{1 + g^2} d\theta.
\end{equation}
Next, let us express the enclosed area $A(\gamma)$ in terms of $g(\theta)$. If $\{P_i \colon i \in I\} \subset \gamma$ is a no more then countable set of vertexes of the curve $\gamma$, then by the Gauss~-- Bonnet formula
\begin{equation}
\label{mainproofeq21}
A(\gamma) = 2\pi - \int \limits_\gamma \frac{1}{R(s)} ds - \sum \limits_{i \in I} \varphi_i,
\end{equation}
where $\varphi_i$ are the jump angles of the tangent at the vertex $P_i$, $i\in I$. Using~(\ref{mainproofeq0}), we can rewrite~(\ref{mainproofeq21}) in the following way:
\begin{equation}
\label{mainproofeq22}
A(\gamma) = 2\pi - \int \limits_{[0,2\pi] \backslash \bigcup \limits_{i \in I} [\alpha_i, \beta_i]} \frac{\sqrt{1 + g^2 + {g'}^2}}{1 + g^2} d\theta - \sum \limits_{i \in I} \varphi_i,
\end{equation}
where the intervals $[\alpha_i, \beta_i]$ of values for $\theta$ correspond to the vertex $P_i$. Let us show that, in fact, for all $i \in I$
\begin{equation}
\label{mainproofeq23}
\int \limits_{\alpha_i}^{\beta_i} \frac{\sqrt{1 + g^2 + {g'}^2}}{1 + g^2} d\theta = \varphi_i.
\end{equation}
Indeed, from~(\ref{propsupfunceq1}) it follows that the contact radius of curvature of the vertex $P_i$ is equal to
\begin{equation}
\label{mainproofeq24}
g(\theta) = \tan(u_i) \cos(\theta - \theta_i),
\end{equation}
where $(u_i, \theta_i)$ are polar coordiantes of $P_i$, and $u_i$ is the distance from the origin $O$ to $P_i$. Substituting~(\ref{mainproofeq24}) into the left side of~(\ref{mainproofeq23}), we get
\begin{equation}
\label{mainproofeq25}
\begin{aligned}
\int \limits_{\alpha_i}^{\beta_i} \frac{\sqrt{1 + g^2 + {g'}^2}}{1 + g^2} d\theta &= \cos u_i \int \limits_{\alpha_i}^{\beta_i} \frac{1}{\cos^2 u_i + \sin^2 u_i \cos^2(\theta - \theta_i)} d\theta\\
&=\left. \arctan \left(\cos u_i \cdot \tan(\theta - \theta_i)\right) \right|_{\alpha_i}^{\beta_i}.
\end{aligned}
\end{equation}
Suppose $\varphi_{\alpha_i}$ and $\varphi_{\beta_i}$ are the angles between the corresponding to $\alpha_i$ and $\beta_i$ semi-tangents to $\gamma$ and the coordinate line $u = u_i$ at $P_i = (u_i,\theta_i)$; then
\begin{equation}
\label{mainproofeq26}
\varphi_i = \varphi_{\beta_i} - \varphi_{\alpha_i}.
\end{equation}
Moreover, straightforward computations show that $$\tan \varphi_{\alpha_i} = \cos u_i \cdot \tan (\alpha_i - \theta_i), \,\, \tan \varphi_{\beta_i} = \cos u_i \cdot \tan (\beta_i - \theta_i).$$ The last equalities together with~(\ref{mainproofeq25}) and~(\ref{mainproofeq26}) proves assertion~(\ref{mainproofeq23}).
Therefore, combining~(\ref{mainproofeq22}) and~(\ref{mainproofeq23}), we finally obtain
\begin{equation}
\label{mainproofeq27}
A(\gamma) = \int \limits_0^{2\pi} \left(1 - \frac{\sqrt{1 + g^2 + {g'}^2}}{1 + g^2} \right)d\theta.
\end{equation}
Thus we need to minimize $A(\gamma)$ taking into account~(\ref{propsupfunceq1}), and setting $L(\gamma) = L_0 = const$.
Let us interpret this problem as an optimal control problem with $t = \theta$, $u(t) = R(t)$, $x_1(t)=g(t)$, and $x_2(t) = \dot x_1(t) = g' (\theta)$.
Since $\gamma$ is a $\lambda$-convex curve with $\lambda=1$, we have the restriction
\begin{equation}
\label{mainproofeq3}
0 \leqslant u(t) \leqslant 1 \text{ a.e. on }[0, 2 \pi].
\end{equation}
Taking into consideration~(\ref{mainproofeq3}), and rewriting~(\ref{propsupfunceq1}),~(\ref{mainproofeq1}), and~(\ref{mainproofeq27}) using the notations introduced above, we come to the following formal problem:
\begin{equation}
\label{optcontrsys}
\begin{aligned}
&\int \limits_0^{2\pi} \left(1 - \frac{\sqrt{1 + x_1^2 + x_2^2}}{1 + x_1^2}\right) dt \rightarrow \min \\
&\int \limits_0^{2\pi} u \frac{\sqrt{1 + x_1^2 + x_2^2}}{1 + x_1^2} dt = L_0 \\
&\left\{
\begin{aligned}
&\dot x_1 = x_2 \\
&\dot x_2 = u \left(\frac{1 + x_1^2 + x_2^2}{1 + x_1^2}\right)^\frac{3}{2} - x_1
\end{aligned} \right.\text{ a.e. on }[0, 2\pi]\\
&0 \leqslant u(t) \leqslant 1 \text{\, a.e. on }[0, 2\pi]\\
&x_1(0) = x_1 (2 \pi) \\ &x_2(0) = x_2 (2 \pi).
\end{aligned}
\end{equation}
Moreover, in problem~(\ref{optcontrsys}) the control $u(t)$ is bounded measurable function on $[0,2 \pi]$, and the phase variable $x(t)$ defined as $x(t) = (x_1(t), x_2(t))$ is absolutely continuous function on $[0,2\pi]$ since $h(\theta) \in C^{1,1}[0,2\pi]$ and $h(\theta) < \pi/2$ (see Remark~\ref{remcoordsys}). In addition, all the functions used in the functional, the integral constraint and the controlled system are continuous with respect to all variables. The same smoothness condition holds for derivatives with respect to $x$ of the mentioned functions.
Therefore, the pair $(x,u)$ that satisfies the controlled system from~(\ref{optcontrsys}) is a controlled process in the sense of Definition~\ref{contrpros}, and if $(x,u)$ also satisfies the integral constraint and the boundary conditions of problem~(\ref{optcontrsys}), then the corresponding trajectory $\{(x(t),u(t)) \colon t\in[0,2\pi]\}$ is an admissible trajectory.
By Blaschke's selection theorem (see~\cite{Bla}) the posed problem of minimizing the area bounded by convex curves while keeping their lengths fixed has a solution. Hence the formalized version~(\ref{optcontrsys}) of the problem also has a solution. Thus in our case Pontryagin's Maximum Principle is a criterion for optimality of admissible trajectories.
If $\{\left(\hat x(t), \hat u(t)\right) \colon t \in [0, 2\pi]\}$ is an optimal trajectory for problem~(\ref{optcontrsys}), then by Pontryagin's Maximum Principle it must satisfy the conditions (i)~-- (iv). Let us rewrite them specifically for our problem. The maximum principle implies that there exist numerical multipliers $\lambda_0$, $\lambda_1 \in \mathbb R$ (with $\lambda_0 \geqslant 0$) and absolutely continuous on $[0,2\pi]$ functions $p_1(t)$ and $ p_2(t)$ such that they don't equal zero simultaneously, and the functions $p_i$ are the solutions of the adjacent system almost everywhere on $[0,2\pi]$, which in our case reads
\begin{equation}
\label{adjp1}
\dot p_1 = p_2 \left(1 + \frac{3 u x_1 x_2^2 \sqrt{1 + x_1^2 + x_2^2}}{\left(1 + x_1^2\right)^{\frac{5}{2}}}\right) + \frac{x_1 (\lambda_0 + \lambda_1 u) \left(1 + x_1^2 + 2 x_2^2\right)}{\left(1 + x_1^2\right)^2 \sqrt{1 + x_1^2 + x_2^2}},
\end{equation}
\begin{equation}
\label{adjp2}
\dot p_2 = -p_1 - p_2 \frac{3 u x_2 \sqrt{1 + x_1^2 + x_2^2}}{\left(1 + x_1^2\right)^{\frac{3}{2}}} - \frac{x_2 (\lambda_0 + \lambda_1 u)}{\left(1 + x_1^2\right) \sqrt{1 + x_1^2 + x_2^2}},
\end{equation}
and Pontryagin's function, that for~(\ref{optcontrsys}) has the form
\begin{equation}
\label{ham}
\begin{aligned}
\mathcal H (x, u, p, \lambda_0, \lambda_1) &= p_1 x_2 + p_2 \left( u \left(\frac{1 + x_1^2 + x_2^2}{1 + x_1^2}\right)^\frac{3}{2} - x_1\right) \\
&+ \lambda_1 \left(u \frac{\sqrt{1 + x_1^2 + x_2^2}}{1 + x_1^2}\right) - \lambda_0 \left(1 - \frac{\sqrt{1 + x_1^2 + x_2^2}}{1 + x_1^2}\right),
\end{aligned}
\end{equation}
satisfies the maximality condition
\begin{equation}
\label{mincond}
\mathcal H(\hat x, \hat u, p, \lambda_0, \lambda_1) = \max \limits_{u \in [0, 1]} \mathcal H(\hat x, u, p, \lambda_0, \lambda_1) \text{ a.e. on }[0, 2\pi].
\end{equation}
Let us analyze~(\ref{mincond}) in more details. In our case Pontryagin's function~(\ref{ham}) is linear with respect to $u$ and be written as $$\mathcal H = u \mathcal H_1 + \mathcal H_2,$$
where
\begin{equation}
\label{hamh1}
\mathcal H_1 = \lambda_1 \frac{\sqrt{1 + x_1^2 + x_2^2}}{1 + x_1^2} + p_2 \left( \frac{1 + x_1^2 + x_2^2}{1 + x_1^2}\right)^\frac{3}{2}.
\end{equation}
From~(\ref{mincond}) it follows that the optimal control for problem~(\ref{optcontrsys}) must be of the form (here and below, for brevity, we will denote the the optimal solution $(\hat x, \hat u)$ as $(x, u)$)
\begin{equation}
\label{contr1}
u(t) = \left\{
\begin{aligned}
&1, \text{ for } \mathcal H_1 > 0 \\
&0, \text{ for } \mathcal H_1 < 0 \\
&\text{undefined}, \text{ for } \mathcal H_1 = 0
\end{aligned}
\right. \text{ a.e. on }[0, 2\pi].
\end{equation}
In order to determine the control completely, we need to consider a so-called \textit{singular trajectory}, that is an admissible for~(\ref{optcontrsys}) trajectory on which $\mathcal H_1$ is identically zero for some interval $(t_1,t_2) \subset [0,2\pi]$. For such trajectories the maximum principle doesn't give any information about control's behavior.
From~(\ref{hamh1}) the condition $\mathcal H_1 = 0$ is equivalent to that on the singular trajectory
\begin{equation}
\label{singularextrp2}
p_2 = - \lambda_1 \frac{\sqrt{1 + x_1^2} }{1 + x_1^2 + x_2^2}.
\end{equation}
Substituting~(\ref{singularextrp2}) into differential equation~(\ref{adjp2}) and solving for $p_1(t)$ after all necessary cancellations we will get that on the singular trajectory
\begin{equation}
\label{singularextrp1}
p_1 = x_2 \frac{\lambda_1 x_1 \sqrt{1 + x_1^2} - \lambda_0 \sqrt{1 + x_1^2 + x_2^2} }{(1 + x_1^2)(1 + x_1^2 + x_2^2)}.
\end{equation}
At the same time, the functions $p_1(t)$ and $p_2(t)$ have to satisfy remaining equation~(\ref{adjp1}) from the adjacent system. If we substitute $p_1$ and $p_2$ into~(\ref{adjp1}) and simplify it using the expressions for $\dot x_1$ and $\dot x_2$ from~(\ref{optcontrsys}), then we will come to the equality
\begin{equation*}
\label{adjsextrcons}
\frac{\lambda_1 - \lambda_0 u}{\left(1 + x_1^2\right)^{\frac{3}{2}}}=0.
\end{equation*}
Hence on the whole interval $(t_1,t_2)$ the equality
\begin{equation}
\label{conscond}
\lambda_1 - \lambda_0 u(t) = 0
\end{equation}
must hold.
If $\lambda_0 = 0$, then $\lambda_1=0$, and from~(\ref{singularextrp2}) and~(\ref{singularextrp1}) it follows that $p_1 \equiv p_2 \equiv 0$ on $(t_1,t_2)$. This contradicts the non-triviality condition (ii). Thus we may assume $\lambda_0 = 1$ and further consider only so-called \textit{normal trajectories} for which such a condition is met. With this in mind, equality~(\ref{conscond}) is possible only if $u(t) \equiv \lambda_1$ on the whole interval $(t_1,t_2)$. Since $\lambda_1$ is a constant real number and thus doesn't depend on the interval $(t_1,t_2)$, we conclude that the optimal control is equal $\lambda_1$ on any piece of the singular trajectory.
Therefore, (\ref{contr1}) can be rewritten as
\begin{equation}
\label{contr2}
u(t) = \left\{
\begin{aligned}
1,\,\, &\text{ for } \mathcal H_1 > 0 \\
0,\,\, &\text{ for } \mathcal H_1 < 0 \\
\lambda_1, &\text{ for } \mathcal H_1 = 0
\end{aligned}
\right. \text{ on }[0, 2\pi].
\end{equation}
Since the control can be redefined on a set of measure zero if necessary, we indeed may assume that~(\ref{contr2}) holds on the whole segment $[0,2\pi]$. Also, because $u(t) \in [0, 1]$, then $\lambda_1$ have to satisfy the inequality $1 \geqslant \lambda_1 \geqslant 0$.
Note that from the geometric point of view the singular trajectory for~(\ref{optcontrsys}) corresponds to a circle of curvature equal to $1/\lambda_1$ if $\lambda_1 \neq 0$, and this circle is a set where the bang-bang control $u$ switches its value.
Let us show now that, in fact, in problem~(\ref{optcontrsys}) no arc of the singular extremal can be optimal. For this purpose let us consider the necessary condition for an arc of a singular trajectory to be a part of an optimal trajectory.
Recall that a natural number $q$ is an \textit{order} of the singular trajectory $(x^*,u^*)$ if it is the minimal number such that
$$\frac{\partial}{\partial u}\frac{d^{2q}}{dt^{2q}}\mathcal H_1 \left|_{(x^*,u^*)}\right. \neq 0, $$
where the time derivatives are taken with respect to the corresponding controlled system (see~\cite{KKM}).
An arc of a singular trajectory of order $q$ satisfies the so-called \textit{generalized Legendre~-- Clebsch condition} (see~\cite{KKM}) if along this arc
\begin{equation}
\label{LCC}
(-1)^q \frac{\partial}{\partial u} \left[\frac{d^{2q}}{dt^{2q}}\mathcal H_1\right] \leqslant 0.
\end{equation}
The necessary condition mentioned above states the following (see~\cite{KKM}, and also~\cite{B}): if an arc of a singular trajectory is a part of an optimal trajectory, then this arc must satisfy generalized Legendre~-- Clebsch condition.
In our case we have the singular trajectory of order 1 ($q=1$). It can be easily verified by a straightforward computation. Using~(\ref{singularextrp2}),~(\ref{singularextrp1}), and the assumption $\lambda_0=1$, along the singular trajectory we have
$$
-\frac{\partial}{\partial u}\frac{d^{2}\mathcal H_1}{dt^{2}} = \frac{\left(1 + x_1^2 + x_2^2\right)^\frac{3}{2}}{\left(1 + x_1^2\right)^3} > 0.
$$
The last inequality contradicts necessary condition~(\ref{LCC}). Hence inclusion of the singular trajectory in the solution of problem~(\ref{optcontrsys}) is not optimal. Therefore, on the segment $[0, 2\pi]$ we have a bang-bang control with only values 0 or 1.
From the geometric point of view we obtained that an optimal curve must consist of circular arcs of curvature equal to $1$. Thus the solution of~(\ref{optcontrsys}) belongs to the class of $1$-convex polygons with possibly infinite number of vertexes. It appears that for such a class we have the following geometric proposition, which is the discrete version of Theorem~\ref{mainthalt}.
\begin{proposition}
\label{mainthspecialcase}
Let $\gamma$ and $\tilde\gamma$ be respectively a $\lambda$-convex polygon and a $\lambda$-convex lune on the sphere $\mathbb S^2(k_1^2)$. If $$L(\gamma) = L(\tilde\gamma),$$
then $$A(\gamma) \geqslant A(\tilde\gamma).$$
Moreover, the equality holds if and only if $\gamma$ and $\tilde\gamma$ are congruent.
\end{proposition}
\begin{proof}
We will divide the proof into two steps: (I) when $\gamma$ is a centrally symmetric curve; (II) the general case.
I. Suppose that $\gamma$ is a centrally symmetric $\lambda$-convex polygon. If at any point on $\gamma$ there exist a unique tangent geodesic, then $\gamma$ is a circle of constant curvature equal to $\lambda$. This is possible only when $L(\gamma) = {2\pi}/{\sqrt{k_1^2 + \lambda^2}}$. If $L(\gamma) < {2\pi}/{\sqrt{k_1^2 + \lambda^2}}$, then the curve has at least one pair of centrally symmetric vertexes, that is points at which left and right semi-tangents do not coincide. If we have precisely one pair, then $\gamma$ is a lune.
Suppose that $\gamma$ has two distinct pairs of centrally symmetric vertexes $A$ and $\bar A$, $B$ and $\bar B$ (see Fig.~\ref{pic1} (a)).
\begin{figure}[h]
\begin{center}
\begin{minipage}[H]{0.49\linewidth}
\begin{center}
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm,scale=0.25]
\clip(-5.77,-12.87) rectangle (13.24,2.78);
\fill[fill=black,fill opacity=0.1] (-2.38,0) -- (8.17,1.76) -- (11.62,0) -- cycle;
\fill[fill=black,fill opacity=0.1] (9.85,-9.84) -- (-0.71,-11.6) -- (-4.15,-9.84) -- cycle;
\fill[fill=black,fill opacity=0.1] (-4.82,-6.51) -- (-2.38,0) -- (-4.15,-9.84) -- cycle;
\fill[fill=black,fill opacity=0.1] (12.29,-3.33) -- (9.85,-9.84) -- (11.62,0) -- cycle;
\draw [shift={(-2.38,0)}] (0,0) -- (-100.22:0.83) arc (-100.22:0:0.83) -- cycle;
\draw [shift={(9.85,-9.84)}] (0,0) -- (79.78:0.83) arc (79.78:180:0.83) -- cycle;
\draw [shift={(11.62,0)}] (0,0) -- (180:0.83) arc (180:259.78:0.83) -- cycle;
\draw [shift={(-4.15,-9.84)}] (0,0) -- (0:0.83) arc (0:79.78:0.83) -- cycle;
\draw [dash pattern=on 6pt off 6pt] (-2.38,0)-- (11.62,0);
\draw [dash pattern=on 6pt off 6pt] (-2.38,0)-- (-4.15,-9.84);
\draw [dash pattern=on 6pt off 6pt] (-4.15,-9.84)-- (9.85,-9.84);
\draw [dash pattern=on 6pt off 6pt] (11.62,0)-- (9.85,-9.84);
\draw [shift={(6.57,-21.19)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.5:1.97,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(-0.51,-19.54)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.02:1.18,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(0.9,11.35)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.5:1.97,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw [shift={(7.97,9.7)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.02:1.18,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw(20.73,4.11) -- (24.89,4.11);
\draw(21.68,1.65) -- (25.84,1.65);
\draw [shift={(18,-3.68)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=3.26:3.41,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(17.69,-11.24)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=2.63:2.93,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(-10.22,1.4)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=2.63:2.93,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw [shift={(-10.54,-6.16)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=3.26:3.41,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw (-2.38,0)-- (3.3,2.4);
\draw (-2.38,0)-- (-5.39,-5.38);
\draw [shift={(11.62,0)}] (180:0.83) arc (180:259.78:0.83);
\draw [shift={(11.62,0)}] (180:0.62) arc (180:259.78:0.62);
\draw [shift={(-4.15,-9.84)}] (0:0.83) arc (0:79.78:0.83);
\draw [shift={(-4.15,-9.84)}] (0:0.62) arc (0:79.78:0.62);
\draw (-3.9,1.06) node[anchor=north west] {$A$};
\draw (11.5,1.06) node[anchor=north west] {$B$};
\draw (-5.7,-9.45) node[anchor=north west] {${\bar{B}}$};
\draw (9.55,-9.45) node[anchor=north west] {$\bar{A}$};
\draw (4.1,-0.05) node[anchor=north west] {${a}$};
\draw (2.4,-8.5) node[anchor=north west] {${a}$};
\draw (-3.1,-4) node[anchor=north west] {${b}$};
\draw (9.14,-4) node[anchor=north west] {${b}$};
\draw (-2.2,-0.3) node[anchor=north west] {$\alpha$};
\draw (-2.95,-8.32) node[anchor=north west] {$\beta$};
\draw (7.04,-11.12) node[anchor=north west] {$\gamma$};
\fill [color=black] (-2.38,0) circle (3.0pt);
\fill [color=black] (-4.15,-9.84) circle (3.0pt);
\fill [color=black] (9.85,-9.84) circle (3.0pt);
\fill [color=black] (11.62,0) circle (3.0pt);
\draw [dash pattern=on 6pt off 6pt] (-4.15,-9.84) -- (11.62,0);
\end{tikzpicture}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[h]{0.49\linewidth}
\begin{center}
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm,scale=0.25]
\clip(-11.17,-9.75) rectangle (12.87,2.86);
\fill[fill=black,fill opacity=0.1] (-2.38,0) -- (8.17,1.76) -- (11.62,0) -- cycle;
\fill[fill=black,fill opacity=0.1] (4.21,-6.71) -- (-6.35,-8.47) -- (-9.79,-6.71) -- cycle;
\fill[fill=black,fill opacity=0.1] (-8.29,-3.66) -- (-2.38,0) -- (-9.79,-6.71) -- cycle;
\fill[fill=black,fill opacity=0.1] (10.11,-3.05) -- (4.21,-6.71) -- (11.62,0) -- cycle;
\draw [shift={(-2.38,0)}] (0,0) -- (-137.85:0.83) arc (-137.85:0:0.83) -- cycle;
\draw [shift={(4.21,-6.71)}] (0,0) -- (42.15:0.83) arc (42.15:180:0.83) -- cycle;
\draw [shift={(11.62,0)}] (0,0) -- (180:0.83) arc (180:222.15:0.83) -- cycle;
\draw [shift={(-9.79,-6.71)}] (0,0) -- (0:0.83) arc (0:42.15:0.83) -- cycle;
\draw [dash pattern=on 5pt off 5pt] (-2.38,0)-- (11.62,0);
\draw [dash pattern=on 5pt off 5pt] (-2.38,0)-- (-9.79,-6.71);
\draw [dash pattern=on 5pt off 5pt] (-9.79,-6.71)-- (4.21,-6.71);
\draw [dash pattern=on 5pt off 5pt] (11.62,0)-- (4.21,-6.71);
\draw [shift={(6.57,-21.19)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.5:1.97,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(-0.51,-19.54)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.02:1.18,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(-4.74,14.48)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.5:1.97,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw [shift={(2.33,12.83)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.02:1.18,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw(20.73,4.11) -- (24.89,4.11);
\draw(21.68,1.65) -- (25.84,1.65);
\draw [shift={(11.52,-15.36)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=2.61:2.76,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(6.65,-21.15)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.97:2.28,variable=\t]({1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+1*23*sin(\t r)});
\draw [shift={(-4.83,14.44)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=1.97:2.28,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw [shift={(-9.69,8.65)},line width=1.2pt,fill=black,fill opacity=0.1] plot[domain=2.61:2.76,variable=\t]({-1*23*cos(\t r)+0*23*sin(\t r)},{0*23*cos(\t r)+-1*23*sin(\t r)});
\draw (-2.38,0)-- (3.3,2.4);
\draw (-2.38,0)-- (-8.05,-2.42);
\draw [shift={(11.62,0)}] (180:0.83) arc (180:222.15:0.83);
\draw [shift={(11.62,0)}] (180:0.62) arc (180:222.15:0.62);
\draw [shift={(-9.79,-6.71)}] (0:0.83) arc (0:42.15:0.83);
\draw [shift={(-9.79,-6.71)}] (0:0.62) arc (0:42.15:0.62);
\draw (-3.8,1.25) node[anchor=north west] {$A$};
\draw (11.3,1.25) node[anchor=north west] {${B}$};
\draw (-11.45,-6.5) node[anchor=north west] {${\bar{B}}$};
\draw (3.8,-6.5) node[anchor=north west] {$\bar{A}$};
\draw (3.05,0.15) node[anchor=north west] {${a}$};
\draw (-2.9,-5.5) node[anchor=north west] {${a}$};
\draw (-5.68,-2.5) node[anchor=north west] {${b}$};
\draw (6.5,-2.5) node[anchor=north west] {${b}$};
\draw (-2.4,-0.6) node[anchor=north west] {$\alpha_1$};
\draw (-8.75,-5.1) node[anchor=north west] {$\beta_1$};
\draw (1.43,-8) node[anchor=north west] {$\gamma$};
\fill [color=black] (-2.38,0) circle (3.0pt);
\fill [color=black] (-9.79,-6.71) circle (3.0pt);
\fill [color=black] (4.21,-6.71) circle (3.0pt);
\fill [color=black] (11.62,0) circle (3.0pt);
\end{tikzpicture}
\end{center}
\end{minipage}
\begin{minipage}[h]{1\linewidth}
\begin{tabular}{p{0.49\linewidth}p{0.49\linewidth}}
\vskip 0.05mm \centering (a) & \vskip 0.05mm \centering (b)
\end{tabular}
\end{minipage}
\caption{Centrally-symmetric $\lambda$-convex polygon $\gamma$ before (a) and after (b) the four-bar linkage isoperimetric deformation around the vertex $A$.}
\label{pic1}
\end{center}
\end{figure}
Let us consider the geodesic quadrilateral $AB\bar A\bar B$. Since $\gamma$ is centrally symmetric, the opposite sides of this quadrilateral are equal. Let us introduce the notations: $a = \left|AB\right| = \left|\bar A\bar B\right|$, $b = \left|A\bar B\right|=\left|\bar A B\right|$, $\alpha = \angle \bar B A B = \angle \bar B \bar A B$, $\beta = \angle A B \bar A = \angle A \bar B \bar A$. In the spherical geometry we have a fact, which is a complete analog of the elementary fact from the Euclidean geometry.
\begin{lemma}
\label{lemarea}
If $f(\alpha)$ is the area of a spherical triangle $BA\bar B$ on $\mathbb S^2(k_1^2)$ with $|BA| = a$, $|A\bar B| = b$ and with $\angle BA \bar B = \alpha$, then the function $f(\alpha)$ on the interval $[0,\pi]$ has a unique maximum point $\alpha_0$, and $f$ is strictly monotonous on $[0,\alpha_0)$ and $(\alpha_0, \pi]$ (see Fig.~\ref{pic1} (a)).
\end{lemma}
\begin{proof}
Without loss of generality we may assume that $k_1=1$.
In the spherical case we can get the formula for the area of a triangle completely similar to the one in the Lobachevsky geometry (see~\cite[formula 2.9]{Wal}, and also~\cite{Bil}). In particular, if $g(\alpha) = \tan \frac{f(\alpha)}{2}$, then
\begin{equation*}
\label{lemareaeq1}
g(\alpha) = \frac{\sin a \sin b \sin \alpha}{(1 + \cos a)(1 + \cos b) + \sin a \sin b \cos \alpha}.
\end{equation*}
We compute:
$$
g'(\alpha) = \frac{\sin a \sin b \left( \cos \alpha (1 + \cos a)(1 + \cos b) + \sin a \sin b \right)}{\left((1 + \cos a)(1 + \cos b) + \sin a \sin c \cos \alpha \right)^2}.
$$
It is easy to see that on the interval $[0,\pi]$ the derivative $g'$ is equal zero if and only if $\cos(\alpha) = - \tan \frac{a}{2} \tan \frac{b}{2}$. Because the cosine is monotonous on $[0, \pi]$, the last equation has only one solution $\alpha_0$. Since $g(0)=g(\pi) = 0$, and $g \geqslant 0$, we obtain that $\alpha_0$ is a maximum point of the function $g(\alpha)$, and thus of the function $f(\alpha)$. The trivial observation that $f$ is strictly monotonous on $[0,\alpha_0)$ and $(\alpha_0, \pi]$ finishes the proof of the lemma.
\qed
\end{proof}
From Lemma~\ref{lemarea} it follows that either $\alpha = \beta = \alpha_0$, or $\left(\alpha - \alpha_0\right)\left(\beta - \alpha_0\right) < 0$. In both cases there is an angle ($\alpha$ or $\beta$) whose increase leads to decrease of the area of the quadrilateral $AB\bar A \bar B$. Without loss of generality we can assume that this angle is $\alpha$. Note that increase of $\alpha$ implies decrease of $\beta$.
Now we introduce an isoperimetric deformation of the initial curve that preserves its $\lambda$-convexity, doesn't change the length, decreases the number of pairs of vertexes, and decreases the bounded area. For this we will use the idea of the four-bar linkage method due to Steiner (see~\cite{Bla}, and also~\cite{PR}).
Let us fix the arcs of the curve $\gamma$ between the points $A$, $B$, $\bar A$, and $\bar B$ and assume that at these points we have hinges. Since the four-bar linkage $AB\bar A \bar B$ is not rigid, the whole construction gains mechanical flexibility. Now, let us increase the angle $ B A \bar B$ by rotating the links $AB$ and $A\bar B$ around the hinge $A$ until the angle between semi-tangents to $\gamma$ at $A$ will become equal to $\pi$ (see Fig.~\ref{pic1} (a, b)). Recalling that with such a move the area of the quadrilateral $AB\bar A \bar B$ decreases, and since we have fixed the arcs of $\gamma$, we obtain that the whole area bounded by $\gamma$ will decrease. At the same time the length of $\gamma$ will remain unchanged. Finally, since $\gamma$ is centrally symmetric, and, besides, the angle at the vertexes $B$ and $\bar B$ of $\gamma$ can only decrease, the deformed curve will be still $\lambda$-convex and centrally symmetric.
Therefore, we have found an isoperimetric deformation of our $\lambda$-convex polygon into another $\lambda$-convex polygon such that this deformation decreases the bounded area and reduces the number of vertexes by $2$. Because by Blaschke's selection theorem there exists a centrally symmetric $\lambda$-convex polygon that bounds the least area among all such polygons with the same length, we get that it can be only a $\lambda$-convex lune. The case I of Proposition~\ref{mainthspecialcase} is proved.
II. We are ready to prove the general case. Let $PQ$ be a diameter of $\gamma$, that is a longest geodesic segment joining a pair of points on the curve. And let $\gamma_+$ and $\gamma_-$ be the arcs into which $P$ and $Q$ divide $\gamma$. We will denote $L_{+} = L(\gamma_{+})$ and $L_- = L(\gamma_{-})$. With such notations we have
\begin{equation}
\label{lemspecialcaseeq2}
L(\gamma) = L_+ + L_-.
\end{equation}
Using $\gamma_+$ and $\gamma_-$ let us construct $\lambda$-convex polygons $\gamma_1$ and $\gamma_2$ by reflecting, respectively, $\gamma_+$ and $\gamma_-$ symmetrically with respect to the midpoint of the diameter $PQ$. From the first variation formula it follows that at $P$ and $Q$ there exist supporting for $\gamma$ geodesics such that the geodesic $PQ$ is orthogonal to them. This fact guaranties that the curves $\gamma_i$ are indeed $\lambda$-convex polygons. Observe that
\begin{equation}
\label{lemspecialcaseeq3}
L(\gamma_1) = 2 L_+,\, L(\gamma_2) = 2 L_-.
\end{equation}
Let $\tilde \gamma_i$ be $\lambda$-convex lunes such that $L(\gamma_i) = L(\tilde \gamma_i)$ for $i \in \{1,2\}$. Hence for $\tilde \gamma_i$ and $\gamma_i$ we can apply the case I to obtain
\begin{equation}
\label{lemspecialcaseeq1}
A(\gamma_i) \geqslant A(\tilde\gamma_i), \, i \in \{1,2\}.
\end{equation}
Let $\tilde A$ be an area bounded by a lune $\tilde \gamma$ considered as a function of its length $L(\tilde \gamma)$, that is $\tilde A \left(L (\tilde \gamma)\right) = A(\tilde \gamma)$. To proceed we have to know explicitly the function $\tilde A$ and its convexity properties. These facts are summarized in the following lemma, whose proof consists of a straightforward computation and thus omitted.
\begin{lemma}
\label{lemluneest}
If $\tilde \gamma \subset \mathbb S^2(k_1^2)$ is a $\lambda$-convex lune of length $L$, then
\begin{equation}
\label{lemluneesteq1}
\tilde A (L) = \frac{4}{k_1^2} \arctan\left(\frac{\lambda}{\sqrt{\lambda^2+k_1^2}}\tan\left(\frac{\sqrt{\lambda^2+k_1^2}}{4} L\right)\right)-\frac{\lambda}{k_1^2} L.
\end{equation}
Moreover, $\tilde A$ is a strictly convex function on the whole its domain of definition.
\end{lemma}
Using strict convexity of $\tilde A$ from Lemma~\ref{lemluneest} and taking into consideration equalities~(\ref{lemspecialcaseeq2}),~(\ref{lemspecialcaseeq3}) and inequalities~(\ref{lemspecialcaseeq1}), we estimate:
\begin{equation*}
\begin{aligned}
A(\gamma) &= \frac{1}{2}\left(A(\gamma_1) + A(\gamma_2)\right) \geqslant \frac{1}{2}\left(A(\tilde \gamma_1) + A(\tilde \gamma_2)\right) \\
&= \frac{1}{2}\left(\tilde A(2 L_+) + \tilde A(2 L_-)\right) \geqslant \tilde A\left(\frac{2 L_+ +2 L_-}{2}\right) = \tilde A (L(\gamma)) = A(\tilde \gamma),
\end{aligned}
\end{equation*}
where $\tilde \gamma$ is the $\lambda$-convex lune of the same length as $\gamma$.
To finish the proof of Proposition~\ref{mainthspecialcase} we need to consider the equality case. Note that in the chain of inequalities shown above the equality case, in a view of strict convexity of $\tilde A$ and since the case I, is possible only if $\gamma_1$ and $\gamma_2$ are congruent lunes. But this implies that the polygon $\gamma$ is a $\lambda$-convex lune too, which is equivalent to congruence of $\gamma$ and $\tilde \gamma$. Proposition~\ref{mainthspecialcase} is proved.
\qed
\end{proof}
With the help of Proposition~\ref{mainthspecialcase} we are now ready to accomplish the proof of Theorem~\ref{mainthalt}. Indeed, by Proposition~\ref{mainthspecialcase} the only solution of our isoperimetric minimization problem for the bounded area in the class of $1$-convex polygons will be a $1$-convex lune. Hence, by Pontryagin's Maximum Principle it will also be the only solution of general problem~(\ref{optcontrsys}). Theorem~\ref{mainthalt} is proved. Theorem~\ref{mainth} directly follows from Theorem~\ref{mainthalt} and Lemma~\ref{lemluneest}.
\section{Dual results}
In this section with the help of polar maps we will obtain the dual results to the main results of our paper.
Let $\gamma \subset \mathbb S^2(k_1^2)$ be a $C^{1,1}$-smooth curve on a two-dimensional sphere. Consider some fixed vector field $X$ parallel along $\gamma$. Since a parallel field is a solution of a system of first order ODEs, smoothness of our curve will suffice to construct the field $X$. Let $\varphi(s)$ be the angle between $X$ and the tangent to $\gamma$ taken at the point corresponding to a arc-length parameter $s$ of $\gamma$. We will call \textit{upper geodesic curvature} $\overline{\kappa_g}$ and \textit{lower geodesic curvature} $\underline{\kappa_g}$ of $\gamma$ at a point corresponding to the arc length parameter $s_0$ the quantities
$$
\overline{\kappa_g} (s_0) = \limsup \limits_{s \rightarrow s_0} \frac{\varphi(s) - \varphi(s_0)}{s-s_0}, \,\,
\underline{\kappa_g} (s_0) = \liminf \limits_{s \rightarrow s_0} \frac{\varphi(s) - \varphi(s_0)}{s-s_0}.
$$
Since $\gamma$ is a $C^{1,1}$-smooth curve, both its upper and lower geodesic curvatures are well defined, particularly, they are finite and don't depend on the choice of the vector field $X$ (see~\cite[p.~252]{doC}). If $\gamma$ is $C^m$-smooth with $m \geqslant 2$ at some point, then at this point $\overline{\kappa_g} = \underline{\kappa_g}=\kappa_g$.
Observe that if lower curvature of a curve satisfies the inequality $\underline{\kappa_g} \geqslant 0$, then this curve is locally convex. One more crucial observation we should make here is that if $\lambda \geqslant \overline{\kappa_g} \geqslant \underline{\kappa_g}\geqslant 0$ at each point on $\gamma$, then through any point $P$ on the curve passes a tangent to $\gamma$ circle of geodesic curvature equal to $\lambda$ such that in some neighborhood of $P$ this circle lies from the convex side of $\gamma$. The converse is also true, namely, an existence of a circle with curvature equal to $\lambda$ that touches a convex curve from its convex side (locally) at any point implies a boundedness of the curve's upper curvature. Furthermore, from the proof of Proposition~\ref{propsupfunc} it follows that convex curves with $\lambda \geqslant \overline{\kappa_g} \geqslant \underline{\kappa_g}\geqslant 0$ on a sphere $\mathbb S^2(k_1^2)$ are just polar images of $k$-convex curves with $k=k_1^2/\lambda$. Note that similarly to the above it is possible to define upper and lower curvatures for general convex curves. In such terminology $k$-convex curves on a sphere are precisely those with $\underline {\kappa_g} \geqslant k$.
In order to state the main result of this section, we need the following definition. A \textit{racetrack curve corresponding $\lambda$} with $\lambda > 0$ is a boundary of the convex hull for two equal circles of curvature equal to $\lambda$. Such curves appear as solutions of the maximization problem for the circumscribed radius of curve, provided that its length is fixed and geodesic curvature satisfies the inequality $1 \geqslant \underline{\kappa_g} \geqslant \overline{\kappa_g} \geqslant -1$ (see~\cite{H,AB} for the two-dimensional case, and~\cite{M} for curves in $\mathbb E^n$).
Note that in the spherical case a racetrack curve corresponding $\lambda$ is just the polar image of a $1/\lambda$-convex lune.
Now we are ready to formulate the result dual to Theorem~\ref{mainth}.
\begin{theorem}
\label{dualth}
Let $\gamma \subset \mathbb S^2(k_1^2)$ be a closed embedded curve whose upper and lower geodesic curvatures satisfy the inequality $0 \leqslant \underline{\kappa_g} \leqslant \overline{\kappa_g} \leqslant \lambda$ with some constant $\lambda \geqslant 0$. If $\tilde\gamma \subset \mathbb S^2(k_1^2)$ is a racetrack curve corresponding $\lambda$ such that
$$
A(\gamma) = A(\tilde\gamma),
$$
then
$$
L(\gamma) \leqslant L(\tilde\gamma).
$$
Moreover, equality holds if and only if $\gamma$ and $\tilde \gamma$ are congruent.
\end{theorem}
The stated theorem is equivalent to the following theorem.
\begin{theorem}
\label{dualthalt}
If $\gamma \subset \mathbb S^2 (k_1^2)$ is a closed embedded curve whose upper and lower geodesic curvatures satisfy the inequality $0 \leqslant \underline{\kappa_g} \leqslant \overline{\kappa_g} \leqslant \lambda$, then
\begin{equation}
\label{dualthsphcase}
\begin{aligned}
L(\gamma) &\leqslant \frac{2\pi}{k_1} + \frac{1}{\lambda} \left[2 \pi - k_1^2 A(\gamma)\right] - \\
&- \frac{4}{k_1} \arctan \left(\frac{k_1}{\sqrt{\lambda^2 + k_1^2}}\tan\left(\left[2\pi - k_1^2 A(\gamma)\right]\frac{\sqrt{\lambda^2 + k_1^2}}{4\lambda}\right)\right).
\end{aligned}
\end{equation}
Moreover, equality holds if and only if $\gamma$ is a racetrack curve corresponding $\lambda$.
\end{theorem}
\begin{proof}[Proof of Theorem~\ref{dualth}]
Theorem~\ref{dualth} follows from Theorem~\ref{mainth} and Theorem~\ref{mainthalt}. Indeed, if $\gamma$ and $\gamma^\ast$ are two polar curves on a sphere, then from the Gauss~-- Bonnet formula using the computations from the proof of Proposition~\ref{propsupfunc} it easily follows that
\begin{equation}
\label{arealengthdual}
k_1 L(\gamma^\ast) = 2 \pi - k_1^2 A(\gamma).
\end{equation}
And since $\left(\gamma^{\ast}\right)^\ast = \gamma$, from~(\ref{arealengthdual}) we also have
\begin{equation}
\label{arealengthdual2}
k_1^2 A(\gamma^\ast) = 2 \pi - k_1 L(\gamma).
\end{equation}
Let $\gamma^\ast$ and $\tilde\gamma^\ast$ be the polar images of the curves $\gamma$ and $\tilde \gamma$. As we showed above, the curve $\gamma^\ast$ will be a closed embedded ${k_1^2}/{\lambda}$-convex curve, and $\tilde\gamma^\ast$ will be a ${k_1^2}/{\lambda}$-convex lune. Using (\ref{arealengthdual}), and the assertion of the theorem, we obtain
$$
k_1 L(\gamma^\ast) = 2 \pi - k_1^2 A(\gamma) = 2 \pi - k_1^2 A(\tilde \gamma) = k_1 ^2 L(\tilde\gamma^\ast).
$$
Thus by Theorem~\ref{mainthalt},
$$
A(\gamma^\ast) \geqslant A(\tilde\gamma^\ast),
$$
from which using~(\ref{arealengthdual2}), the inequality of Theorem~\ref{dualth} follows directly. The equality case is proved automatically. Theorem~\ref{dualth} is proved.
\qed
\end{proof}
Equivalence of Theorems~\ref{dualth} and~\ref{dualthalt} is obtained by a straightforward computation of the length of a racetrack curve corresponding $\lambda$ assuming it bounds a fixed area. Even more, inequality~(\ref{dualthsphcase}) easily follows from inequality~(\ref{maintheq1}) with the help of formula~(\ref{arealengthdual}), and the discussed above polar correspondence between $C^{1,1}$-smooth curves with $0 \leqslant \underline{\kappa_g} \leqslant \overline{\kappa_g} \leqslant \lambda$ on $\mathbb S^2(k_1^2)$ and $k_1^2/\lambda$-convex curves on $\mathbb S^2(k_1^2)$.
\begin{remark}
We can also get the dual version of Theorem~\ref{ec2dth1}. Indeed, on the Euclidean plane we can define convex curves of bounded upper ($\overline{\kappa}$) and lower ($\underline{\kappa}$) curvature, and race-track curves in the same way as it was done in the spherical case. But since on the Euclidean plane a polar map depends on the choice of a polar center, and hence there is no absolute polarity as we have on a sphere, the dual result on the plane is not an immediate corollary of Theorem~\ref{ec2dth1}. However, by applying Pontryagin's Maximum Principle once more similarly to~\cite{BorDr1} with curvature of a curve as a control parameter, it is easy to get the desired result. More precisely, if $\gamma \subset \mathbb E^2$ is a closed embedded curve with $0 \leqslant \underline{\kappa} \leqslant \overline{\kappa} \leqslant \lambda$ ($\lambda > 0$), then
$$
L(\gamma) \leqslant \lambda A(\gamma) + \frac{\pi}{\lambda},
$$
and the equality holds only for a race-track curve corresponding $\lambda$.
\end{remark}
\begin{acknowledgements}
The authors would like to thank an anonymous referee for useful comments and suggestions that helped to improve the exposition.
\end{acknowledgements}
|
\section{Introduction}
The biological function of proteins has long been associated with their
ordered, often globular, structures. It is now clear, however,
that many critical cellular functions---in signaling and cell-cycle regulation
in particular---are performed by intrinsically disordered proteins (IDPs)
that lack a unique fold. IDPs are depleted in hydrophobic but enriched
in charged and polar amino
acids~\cite{Uversky00,tompa12,FormanKay13,vanderLee14,Chen15}.
Many IDPs are polyampholytes with both positively
and negatively charged monomers~\cite{Higgs91, Dobrynin04}. Recently,
some polyampholytic IDPs were found to function not only as individual
molecules but also collectively by undergoing phase
separation on a mesoscopic length scale to form
condensed liquid-phase non-amyloidogenic IDP-rich droplets that may
encompass RNA and other
biomolecules~\cite{Brangwynne2009, Rosen12, McKnight12, Lee13, Nott15}.
This phase behavior is the basis of functional membraneless
organelles in the cell. Without a membrane, these organelles can
respond rapidly to environmental stimuli. They are critical for cellular
integrity, homeostasis~\cite{Nott15, Dundr10},
and the spatial-temporal control of gene regulation and cell
growth~\cite{Lee13, wright14, julicher14, Nott15, ElbaumGarfinkle15, Brangwynne15}.
In view of the biological/biomedical importance of this newly discovered
phenomenon, insights into its physics would be valuable.
In line with charge effects
on the size of individual IDP molecules~\cite{Mao10, Marsh10, Muller10, Das13},
electrostatics figured prominently in
computational~\cite{Ruff15} and experimental~\cite{Nott15, Quiroz15} analysis
of IDP phase separation. Polymer theory emphasized the sensitivity
of polyampholyte phase behavior to charge pattern~\cite{Higgs91, Dobrynin04},
but only a few simple patterns were
considered~\cite{Gonzalez94, Cheong05, Jiang06}.
No systematic approach has been put forth to apply those ideas to
understand phase behaviors of genetically coded proteins.
Inasmuch as IDP phase separation is concerned, the theoretical discussion
to date remains at the mean-field
level~\cite{Lee13, julicher14, Nott15, Brangwynne15},
which precludes sequence dependence to be addressed.
In this Letter, we take a step forward by developing an analytical theory
for sequence-specific electrostatics in polyampholyte phase separation,
aiming to synergize theory and experiment and lay the groundwork for
further theoretical advances.
While our theory is applicable to any charge sequence,
an impetus for our effort was the recent experiments
on the RNA helicase Ddx4, the N-terminal of which is an IDP,
that can undergo {\it in vitro} and in cell phase
separation under ambient conditions~\cite{Nott15}. Ddx4
is essential for the assembly and maintenance of the related nuage in mammals,
P-granules in worms, as well as pole plasm and polar granules in
flies~\cite{Nott15, Liang94}. The wildtype sequence of the
residue 1--236 N-terminal disordered region of Ddx4 (termed Ddx4$^{\rm N1}$)
may be seen by sliding-window averaging as a series of alternating charge
blocks abounding with negatively charged aspartic (D) and glutamic (E) acids
and positively charged arginines (R) and lysines (K) \cite{Nott15}.
In the absence of this block-charge pattern,
a charge-scrambled mutant Ddx4$^{\rm N1}$CS
does not phase separate~\cite{Nott15}. Thus the sequence-specific charge
pattern---not total positive and total negative charges per se---is
crucial for Ddx4 phase behavior.
Since it is challenging to synthesize
nonbiological polyampholytes with specific charge patterns~\cite{Dobrynin04},
theories have focused on the quenched ensemble
average of random sequences~\cite{Higgs91, Wittmer93}
or limited to diblock~\cite{Gonzalez94}
or at most four-block charge patterns~\cite{Cheong05, Jiang06}.
In contrast, a high diversity of protein sequences are readily synthesized by
the cellular machinery. In view of this expanded experimental horizon,
the availability of a large repertoire of IDP sequences
for phase separation studies is foreseeable, with the new physics that
is likely to ensue.
With this in mind, we present below a random phase approximation (RPA)
theory~\cite{Borue88, Wittmer93, Gonzalez94,Mahdi00, Ermoshkin03a, Ermoshkin04}
for any charge pattern and, as an example, apply our theory to elucidate
Ddx4 phase behavior.
As in prior applications of RPA, the particle density in our theory
is assumed to be rather homogeneous to permit an approximate account
based solely on two-body correlations of density fluctuation~\cite{Borue88}.
Electrostatic
free energy is similarly approximated by a Gaussian integral over charge
fluctuations. Other limitations of RPA, such as in its treatment of
short-range electrostatic interactions, are well documented
(e.g.~\cite{Borue88,Mahdi00}).
Approximations notwithstanding, RPA represents
a significant improvement, especially amidst the current interest
in IDP phase separation \cite{julicher14, Brangwynne15}, over
Debye-H\"{u}ckel theory for
biological coacervation~\cite{Overbeek57} in that RPA
embodies chain connectivity~\cite{Borue88,Mahdi00}
and hence allows for an explicit account of the charge
pattern along the chain sequence~\cite{Wittmer93,Gonzalez94}.
Based on previous approaches~\cite{Wittmer93, Gonzalez94},
our theory is for a system of aqueous polyampholytes, small
counterions, and salt. In contrast to self-repelling, strongly charged
polyelectrolytes that necessitate a modified
RPA~\cite{Mahdi00, Ermoshkin03b},
we focus on polyampholytes that are nearly neutral, i.e., with
approximately equal numbers of acidic and basic residues
(e.g. Ddx4$^\text{N1}$ and Ddx4$^\text{N1}$CS). Each
polyampholyte consists of $N$ monomers
(amino acid residues) with
charges $\{\sigma_i\} = \{ \sigma_1,\sigma_2,\dots,\sigma_N \}$, where
$\sigma_i$ is in units of the electronic charge $e$.
The counterions and salt are monovalent in this formulation.
The densities of monomers, counterions, and salt are denoted, respectively,
by $\rho_m$, $\rho_c$, and $\rho_s$. The number of monovalent counterions
is taken to be equal to the net charge of polyampholytes, viz.,
$\rho_c = \rho_m\left|\sum_i \sigma_i\right|/N$.
The free energy $F$ of our system per volume $V$ in units of $k_{\rm B}T$
is
\begin{equation}
f \equiv \frac{F}{V k_{\rm B}T} = -s + f_{\rm el} \; ,
\end{equation}
where $k_{\rm B}$ is Boltzmann's constant and $T$ is absolute temperature,
$s$ is the entropic contribution, and $f_{\rm el}$ accounts for electrostatic
interactions. As IDPs have few hydrophobic residues, as a first step in
our development, we assume that electrostatics is the dominant enthalpic
contribution while neglecting presumably weaker short-range attractive
forces. For simplicity, all monomers of the polyampholytes as well as the
counterions and salt ions are taken to be of equal size with length scale $a$.
The entropic term that accounts for excluded volume follows directly
from Flory-Huggins (FH) theory~\cite{Flory53, deGennes79, chandill94}:
\begin{equation}
\begin{aligned}
-s a^3 = & \frac{\phi_m}{N} \ln \phi_m + \phi_c \ln\left(\phi_c \right) +
2\phi_s \ln(\phi_s) \\
&+ (1\!-\!\phi_m\!-\!\phi_c\!-\!2\phi_s)\ln(1\!-\!\phi_m\!-
\!\phi_c\!-\!2\phi_s),
\label{eq:entropy}
\end{aligned}
\end{equation}
where $\phi_m$, $\phi_c$, $\phi_s$ are, respectively, the volume ratios
$\rho_m a^3$, $\rho_c a^3$, and $\rho_s a^3$.
A uniform, phase-independent $\phi_s$ is assumed here for simplicity
as previous work suggested that change in salt concentration is
insignificant upon biological coacervation~\cite{Overbeek57}.
The $f_{\rm el}$ term for electrostatics is computed by
RPA~\cite{Wittmer93, Mahdi00, Ermoshkin03a, Ermoshkin04}:
\begin{equation}
f_{\rm el} = \frac{1}{2}\int \frac{d^3 k}{(2\pi)^3}
\left\{\ln[\det(1 + \hat{G}_k\hat{U}_k)] - \mathrm{Tr}
(\hat{\rho}\;\hat{U}_k) \right\},
\label{eq:f_el}
\end{equation}
where $\hat{G}_k$, $\hat{U}_k$, and $\hat{\rho}$ are $(N+2)\times(N+2)$
matrices. The logarithmic term containing $\hat{G}_k$ for basic, ``bare''
density correlation without electrostatic effects and $\hat{U}_k$ for
Coulombic interactions, both defined for the reciprocal $k$-space, is
a result of standard Gaussian integrations in RPA theory.
The second trace term subtracts the self electrostatic energy
for all charged densities to eliminate the unphysical divergence of the
first term for $k\!\to\!\infty$~\cite{Wittmer93, Ermoshkin03a, Ermoshkin04}.
The density matrix
\begin{equation}
\hat{\rho} = \left(
\begin{array}{cc}
(\rho_m/N) \hat{I}_N & 0 \\
0 & \hat{\rho}_I
\end{array}\right)
\end{equation}
is a diagonal matrix, with $\hat{I}_N$ being the $N$-dimensional identity
and $\hat{\rho}_I$ a $2\times 2$ diagonal matrix for the
positive and negative monovalent ions, namely
$((\hat{\rho}_I)_{11}, (\hat{\rho}_I)_{22})=(\rho_s\!+\!\rho_c, \rho_s)$ or
$(\rho_s, \rho_s\!+\!\rho_c)$ when the net polyampholyte charge is,
respectively, negative or positive.
The bare correlation matrix $\hat{G}_k$ combines the monomer-monomer
correlation for a Gaussian chain and the density matrix for the
small monovalent ions:
\begin{equation}
\hat{G}_k = \left(
\begin{array}{cc}
(\rho_m/N) \hat{G}_\mathrm{M}(k) & 0 \\
0 & \hat{\rho}_I
\end{array}\right),
\end{equation}
where $\hat{G}_\mathrm{M}(k)$ is the $N\times N$ matrix for Gaussian chains, with
elements $\hat{G}_\mathrm{M}(k)_{ij} = \exp(-(ka)^2|i-j|/6)$~\cite{DoiEdwardsbook}.
The interaction matrix $\hat{U}_k$ is the Fourier transform of
the Coulomb potential with a short-range physical cutoff on the scale of
monomer size, $U(r) = l_B(1-e^{-r/a})/r$,
which in $k$-space becomes~\cite{Ermoshkin03a, Ermoshkin04}
\begin{equation}
\hat{U}_k = \frac{4\pi l_B}{k^2[1+(ka)^2]}|
q\rangle \langle q | \equiv \lambda(k)| q\rangle \langle q | .
\end{equation}
Here $l_B = e^2/(4\pi\epsilon_0\epsilon k_{\rm B} T)$ is the Bjerrum length,
$\epsilon_0$ is vacuum permittivity,
$\epsilon$ is the dielectric constant, $| q \rangle$ is the column vector
for the charges of the monomers and monovalent ions, and
$\langle q | \equiv | q \rangle^{\rm T}$ is the transposed row vector,
with components
$q_i = \sigma_i$ for $1 \leq i \leq N$, $q_{N+1} =1$, and $q_{N+2} = -1$.
The determinant in Eq.~(\ref{eq:f_el}) can now be simplified as
\begin{equation}
\det(1 + \hat{G}_k\hat{U}_k) = 1+ \lambda(k) \langle q | \hat{G}_k | q \rangle
\label{eq:Sylvester}
\end{equation}
by Sylvester's identity~\cite{Borue88}.
For the analysis below, we define a reduced temperature
$T^* \equiv {a}/{l_B} = {4\pi\epsilon_0\epsilon k_{\rm B} Ta}/{e^2}$.
\begin{figure}
\includegraphics[width=\columnwidth]{block_polys.eps}
\caption{ Phase diagrams of $N=240$
salt-free polyeampholytes with 4, 6, 8, and 12 alternating
charge blocks of $\sigma_\alpha^\text{block} = (-1)^{\alpha-1}\sigma$;
$\phi_m$ and $T^*$ are, respectively, dimensionless polymer volume ratio
and temperature (see text).
A dilute and a condensed phase exist, respectively, above
and below each phase-boundary curve.
(a) $\sigma=1$ and (b) $\sigma=0.5$.
The black dots are the critical points $(\phi_\mathrm{cr},T^*_\mathrm{cr})$.
Polyampholytes with fewer blocks and stronger $\sigma$
phase separate at higher $T^*$.}
\label{fig:ps_block_diff}
\end{figure}
We first consider a simple salt-free
($\rho_s = 0$) solution of polyampholytes consisting of $n$ alternating
charge blocks (labeled by $\alpha$, $\beta=1$,2, $\dots, n$), each with
one charge per $1/\sigma$ monomer, i.e.
$ \sigma_\alpha^\text{block}= (-1)^{\alpha-1}\sigma$
and length $L \!=\! N/n$~\cite{Wittmer93}.
The correlation matrix $\hat{G}_\mathrm{M}(k)$ in this case
is an $n\times n$ matrix for the blocks, with
\begin{equation}
\!\!\hat{G}_\mathrm{M}^\text{block}(k)_{\alpha\beta} = \left\{
\begin{aligned}
& \frac{1+\zeta}{1-\zeta}L\sigma - \frac{2\zeta(1-\zeta^{L\sigma})}{(1-\zeta)^2} \, , \quad \alpha \!=\! \beta , \\
& \frac{\zeta^{(|\alpha-\beta|-1)L\sigma+1}}{(1-\zeta)^2}(1-\zeta^{L\sigma})^2 \, , \; \alpha \!\neq\! \beta ,
\end{aligned}
\right.
\label{eq:Gmk_elements}
\end{equation}
where $\zeta = \exp[-(ka)^2/(6\sigma)]$.
Results for block polyampholytes that are neutral
($n$ even, $\rho_c =0$; Fig.~\ref{fig:ps_block_diff}) indicate that,
when $N$ is fixed,
the tendency to phase separate decreases (i.e., requires a lower $T^*$)
with increasing $n$ (Fig.~\ref{fig:ps_block_diff}(a)).
This predicted behavior offers insights into Ddx4 behavior (see below)
and is consistent with theoretical findings
from a charged hard-sphere chain model~\cite{Jiang06} and grand canonical
Monte Carlo simulations~\cite{Cheong05}.
While a stronger $\sigma$ leads generally to higher
phase separation $T^*$ (and thus higher critical temperature $T^*_\mathrm{cr}$),
the critical concentration $\phi_\mathrm{cr}$ is determined mainly
by the block number $n$, and barely by $\sigma$
[cf. curves for the same $n$ in Fig.~\ref{fig:ps_block_diff}(a) and (b)].
When an increasing $n$ arrives at a strictly
alternating polyampholyte with $\sigma = 1/L$, the
contribution of $\hat{G}_\mathrm{M}^\text{block}(k)$ in
$\langle q | \hat{G}_k | q \rangle$,
\begin{equation}
\frac{1}{N}\langle \sigma | \hat{G}_\mathrm{M}^{\rm block}(k) | \sigma \rangle
= \frac{1-\zeta}{1+\zeta}\sigma +
\frac{1}{N}\frac{2\zeta[1-(-1)^n \zeta^{N\sigma}]}{(1+\zeta)^2},
\label{eq:sGs_alternating}
\end{equation}
contains the second $O(1/N)$ term that diminishes as $N\to\infty$.
In that limit, the first term leads to $f_{\rm el} \propto \phi_m^2$ after
integration, defining an effective FH parameter
$\chi \propto \sigma^{3/2}/T^{*2}$ for electrostatics~\cite{Wittmer93}.
When $N$ is finite, the second term enhances phase separation,
resulting in a $\phi_\mathrm{cr}$ much smaller than the $\phi_\mathrm{cr} =1/\sqrt{N}$
predicted by FH theory. For example, for $N=240$, $\sigma = 1$ and 0.5,
$\phi_\mathrm{cr}$ $=$ 0.0213 and 0.0192, respectively. Both $\phi_\mathrm{cr}$ values
are much smaller than the FH value of $1/\sqrt{240}=0.06$.
\begin{figure}
\includegraphics[width=\columnwidth]{Ddx4_seqs_R.eps}
\caption{ (a) The Ddx4$^\text{N1}$ (top) and Ddx4$^\text{N1}$CS (bottom)
amino acid sequences. Positively (+), negatively ($-$) charged
(basic and acidic), and aromatic residues are highlighted, respectively,
in red, turquoise, and yellow.
(b) Charge patterns of the sequences ($+$: red; $-$: blue) indicate
that the charges exhibit more block-like properties (repeated red
or repeated blue)
and are less evenly dispersed in Ddx4$^\text{N1}$
than in Ddx4$^\text{N1}$CS (cf. Fig.6A of \cite{Nott15}).
(c) Theoretical phase diagrams computed by Eq.~(\ref{eq:f_el}) under
salt-free conditions.
(d) Ddx4$^\text{N1}$ phase diagrams [Eq.~(\ref{eq:f_el})]
for different monovalent salt (NaCl) concentrations. }
\label{fig:Ddx4_N1-N1CS_compare}
\end{figure}
We now apply the full theory to Ddx4 by considering the exact
Ddx4$^\text{N1}$ and charged scrambled Ddx4$^\text{N1}$CS sequences
(Fig.~\ref{fig:Ddx4_N1-N1CS_compare}(a)). There are 32
positively and 36 negatively charged residues in either
sequence (Fig.~\ref{fig:Ddx4_N1-N1CS_compare}(b)),
the two polyampholytes are thus nearly neutral~\cite{Nott15}.
We assign $\sigma_i = 1$ to each of the 28 R and 4 K residues,
$\sigma_i = -1$ to each of the 18 D and 18 E residues,
$\sigma_i = 0$ to all other residues. We then substitute these
$\{ \sigma_i \}$ values into Eq.~(\ref{eq:f_el}) for $| q \rangle$
(now a $(236+2)$-component vector) to
compute the two sequences' $f_{\rm el}$ and the corresponding
phase diagrams.
The $\sigma_i$'s are taken to be constant because we are
interested primarily in physiological and/or experimental pH in
the range of 7.0--8.0~\cite{Nott15}, although amino acid charges can
change significantly if pH variation is larger.
Comparing the two phase boundaries
in the absence of salt ($\rho_s=0$ but $\rho_c\ne 0$)
indicates that charge scrambling leads to an approximately three fold
decrease in critical temperature (Fig.~\ref{fig:Ddx4_N1-N1CS_compare}(c)),
underscoring once again the importance of charge
pattern in polyampholyte phase behavior and is in qualitative
agreement with the experimental observation that Ddx4$^\text{N1}$
phase separates whereas Ddx4$^\text{N1}$CS does not~\cite{Nott15}.
Interestingly, the predicted $\phi_\mathrm{cr}$ of Ddx4$^\text{N1}$CS is smaller than
that of Ddx4$^\text{N1}$, which, according to Fig.~\ref{fig:ps_block_diff},
suggests that Ddx4$^\text{N1}$CS is akin to a sequence with a weaker $\sigma$
but has longer charge blocks than Ddx4$^\text{N1}$.
Apparently, the sequential proximity of opposite charges
in Ddx4$^{\rm N1}$CS results in a much weaker effective $\sigma$.
We next explore salt effects on Ddx4$^\text{N1}$ phase behavior by
equating $a^3$ with the volume of a
single water molecule, in which case $\phi_m$ $=236$[Ddx4$^{\rm N1}$]/55.5 M
because the molarity of water is 55.5. We consider $\phi_s=$
0.0018, 0.0027, 0.0036, and 0.0054, which are calculated in the same
manner, respectively, for [NaCl]=100, 150, 200, and 300 mM~\cite{Nott15}.
To compare with Fig.~4 of \cite{Nott15},
we focus first on the range of [Ddx4$^{\rm N1}$]=0--400$\mu$M.
Fig.~\ref{fig:Ddx4_N1-N1CS_compare}(d)
shows a clear decreasing propensity
(i.e., requiring a lower $T^*$) to phase separate with increasing salt.
This theoretical salt dependence is expected of Ddx4$^\text{N1}$
as a nearly neutral
polyampholyte (whereas phase separation of highly charged polyelectrolytes
can be promoted by salt~\cite{Dobrynin05}), and is qualitatively
consistent with experiment~\cite{Nott15}.
While Fig.~\ref{fig:Ddx4_N1-N1CS_compare}(d) provides a
conceptual rationalization, it does not match quantitatively
with experimental salt dependence: For [Ddx4$^{\rm N1}$] $\approx 150\mu$M,
the ratio of absolute
temperatures at the experimental phase boundaries for [NaCl] $=100$mM
and $300$mM is about $1.2$ (between $\approx 0^\circ$
and $60^\circ$C)~\cite{Nott15}, but the corresponding theoretical
ratio is much higher at $\approx 1.7$.
This mismatch means that certain ``background'' Ddx4 interactions that are
less dependent on salt have not been
taken into account by our theory. Indeed, aromatic interactions
are expected to contribute significantly to Ddx4 phase properties,
as to other IDP behaviors~\cite{Chen15, Song13}, because a Ddx4$^\text{N1}$
mutant with wildtype charge pattern but with 9 of its 14 phenylalanines (F)
replaced by alanines does not phase separate~\cite{Nott15}. Moreover,
repetitive phenylalanine-glycine (FG) patterns in IDPs are known to
drive phase separation~\cite{Schmidt15, Brangwynne15}. These observations
suggest strongly that cation-$\pi$ and $\pi$-$\pi$ stacking interactions
can play central roles in IDP phase
separation~\cite{Nott15, Schmidt15, Brangwynne15, Vernon16}.
\begin{figure}
\includegraphics[width=\columnwidth]{RPAonly_vs_RPA_FH_R.eps}
\caption{(a) Theoretical phase diagrams of Ddx4$^\text{N1}$ under different
[NaCl]
values, computed by RPA theory augmented by FH short-range interactions
(curves), with fitted $\epsilon=29.5$, $\varepsilon_h = 0.15$, and
$\varepsilon_s = -0.3$ (see text). Experimental data~\cite{Nott15}
are included for comparison (symbols; same color code). (b) The salt-free
phase diagrams of Ddx4$^{\rm N1}$ and Ddx4$^{\rm N1}$CS with and without the
FH interaction in Eq.(\ref{FH-aug}).}
\label{fig:RPAonly_vs_RPA_FH}
\end{figure}
With this consideration, we augment our RPA theory for
Ddx4$^\text{N1}$ with a mean-field account of $\pi$-interactions,
which are known to be spatially short-ranged~\cite{Ma97, Meyer03} and
may therefore be formulated, as a first approximation, by a
salt-independent FH term,
\begin{equation}
f_{\rm FH} = \chi\phi_m(1-\phi_m) = \left(\frac{\varepsilon_h}{T^*}+
\varepsilon_s\right)\phi_m(1-\phi_m),
\label{FH-aug}
\end{equation}
where $\varepsilon_h$ and $\varepsilon_s$ are the enthalpic and entropic
contributions, respectively, to the FH interaction parameter $\chi$.
To compare theory with experiment quantitatively, values for the monomer
length scale $a$ and the dielectric constant $\epsilon$ are needed
to convert $T^*$ to actual temperature.
We take $a$ to be the C$\alpha$-C$\alpha$ distance 3.8\AA, and
let $\epsilon$ be a fitting parameter, as $\epsilon$
of an aqueous protein solution can vary widely between
$\approx$ 2 and 80~\cite{Pitera01, Warshel06}. By treating
$\varepsilon_h$ and $\varepsilon_s$ also as global fitting parameters,
a reasonably good fit is achieved with experimental results
for all four available [NaCl] values~\cite{Nott15}
(Fig.~\ref{fig:RPAonly_vs_RPA_FH}(a))
by using a value of $\epsilon=29.5$ that is physically plausible
for a mixture of protein ($\epsilon\approx 2$--$4$) and water
($\epsilon\approx 80$).
As a self-consistency check,
we compare the salt-free phase diagrams of
the wildtype and charge-scrambled Ddx4 sequences computed with the augmented
FH interaction. Ddx4$^{\rm N1}$CS is then predicted to never
phase separate at temperature appreciably above 0$^\circ$C
(Fig.~\ref{fig:RPAonly_vs_RPA_FH}(b)), consistent with no observation
of Ddx4$^{\rm N1}$CS phase separation in experiment under
physiological conditions~\cite{Nott15}.
Our model's ability to rationalize the diverse phase behaviors
of Ddx4$^{\rm N1}$ and Ddx4$^{\rm N1}$CS simultaneously suggests that it can
be applied to predict/rationalize the phase behaviors of other permutations
of Ddx4$^{\rm N1}$ sequences in the future.
Interestingly, the $\phi_\mathrm{cr}$ of Ddx4$^{\rm N1}$CS
is dramatically shifted from 0.0080 by the FH term to 0.046, which
is in the order of the FH critical point $1/\sqrt{236} = 0.065$,
indicating that the electrostatic interaction
in Ddx4$^{\rm N1}$CS is weak and its phase behavior is determined
mainly by the augmented FH interactions. In contrast, the critical point
of the wildtype Ddx4$^{\rm N1}$ is barely shifted by
the augmented FH term, implying that its phase behavior is dominated
by electrostatics.
The fitted $\varepsilon_h = 0.15$ and $\varepsilon_s = -0.3$
in Fig.~\ref{fig:RPAonly_vs_RPA_FH} correspond
to a favorable enthalpy of $\Delta H = -0.44$ kcal/mol and an unfavorable
entropy of $\Delta S = -0.60$ cal/mol K$^{-1}$.
Notwithstanding the limitations of our model such that part of these
augmented FH parameters may be needed to correct for the inaccuracies of RPA
itself, the fitted quantities are in line with our
assumption that they should account approximately for favorable
$\pi$-interactions. The
cation-$\pi$ attractive enthalpy between the lysine NH$_4^+$ group
and an aromatic ring is about 20 kcal/mol in gas-phase ab initio
simulation~\cite{Ma97}, and, because of the
resonant $\pi$-electron on its guanidinium group~\cite{Marsili08},
the attraction between an arginine and an aromatic ring can be even
stronger, though cation-$\pi$ interaction strength can be weakened
in an aqueous environment~\cite{Gallivan99}.
Considering there are 32 cations and 22 aromatic rings (14 phenylalanines,
3 tryptophans, and 5 tyrosines) on Ddx4$^\text{N1}$ and Ddx4$^\text{N1}$CS,
the effective $\Delta H$ for
cation-$\pi$ interactions in the
FH parameter may be estimated to be at least of order
$-20(22\times 32/236^2) =-0.25$kcal/mol (note that $\pi$-$\pi$ energies
and beyond-pairwise multi-body interactions
are not included in this estimation), which matches reasonably
well with the fitted value. Moreover, formation of cation-$\pi$
and $\pi$-$\pi$ pairs invariably entail sidechain entropy losses, which
are consistent with a fitted negative value for $\Delta S$.
We notice that curvature of the
phase boundary increases with increasing $\varepsilon_h$. In view of
the experimental trend of decreasing phase boundary
curvature with increasing salt, this observation suggests that
the theory-experiment match may be improved
by allowing $\varepsilon_h$ to decrease slightly with [NaCl].
Such a model may be justifiable because, with increasing salt
concentration, aromatic rings are more likely to bind to surrounding
salt cations rather than engaging with the positively charged lysines
and arginines. A detailed analysis of this possibility, however,
is beyond the scope of the present work.
In summary, we have developed a general analytical theory for
sequence-specific electrostatic effects on polyampholyte phase separation.
Going beyond mean-field theories that do not consider sequence
information~\cite{Nott15}, our theory provides physical
underpinnings for experimental Ddx4 behaviors in terms of the charge pattern
along its sequence and probable $\pi$-interactions. Theory and
experiment both suggest that an alternating charge pattern that maintains
reasonably high average positive or negative charge densities over a length
scale encompassing at least several amino acid residues is required for Ddx4
phase separation~\cite{Nott15}. It would be instructive to investigate how
this principle is manifested in other IDPs. In any event,
it would be useful to apply our theory to other IDP
polyampholytes and to generalize the present formulation to incorporate
sequence-specific non-electrostatic interactions.
\begin{acknowledgments}
We thank Patrick Farber, Lewis Kay, Jianhui Song, and Robert Vernon for
helpful discussions. This work was supported by a Canadian Cancer Society
Research Institute grant to J.D.F.-K. and H.S.C. as well as a Canadian
Institutes of Health Research grant to H.S.C. We are grateful to SciNet
of Compute Canada for providing computational resources.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\subsubsection*{Abstract}
{\small In the present note we review some recent results for a class of singular perturbation problems for a Navier-Stokes-Korteweg system with Coriolis force.
More precisely, we study the asymptotic behaviour of solutions when taking incompressible and fast rotation limits simultaneously, in a constant capillarity regime.
Our main purpose here is to explain in detail the description of the phenomena we want to capture, and the mathematical derivation of the system of equations.
Hence, a huge part of this work is devoted to physical considerations and mathematical modeling.}
\paragraph*{2010 Mathematics Subject Classification:}{\small 35B25
(primary);
35Q35,
35B40,
35Q86,
76D45,
76M45,
76U05,
82C24
(secondary).}
\paragraph*{Keywords:}{\small Capillarity; Earth rotation; Navier-Stokes-Korteweg system; singular perturbation problem; low Mach, Rossby and Weber numbers;
variable rotation axis; Zygmund regularity.}
\section{Introduction} \label{s:intro}
In the present note, we consider a mathematical model for compressible viscous fluids whose dynamics is mainly influenced by two
effects: strong surface tension and fast rotation of the ambient reference system (the Earth for us, but it is also the case of e.g. other planets or stars).
We intend to review some recent results for a class of singular perturbation problems for our system. More precisely, we study the asymptotic
behaviour of solutions when taking incompressible and fast rotation limits simultaneously, in a constant capillarity regime.
The analysis we present has mostly been performed in \cite{F_2015}-\cite{F_2016}: we refer to these works for more details and further references.
The main purpose of this work is rather to explain in detail the physical description of the phenomena we want to capture, and their translation at the mathematical level.
Hence, in a first moment (see Section \ref{s:model}) we derive the mathematical model we want to study, namely a Navier-Stokes-Korteweg system with degenerate viscosity coefficient
and Coriolis term.
It is a compressible Navier-Stokes type system, with an additional term due to capillarity and which depends on higher order space derivatives of the density.
The form of the so-called Korteweg stress tensor we consider here was firsty introduced by Dunn and Serrin in \cite{D-S} (see also Subsection \ref{ss:i_cap} below
for further details); it is nothing but a choice among all the possible ones: different forms lead to different models, which are relevant in e.g. quantum hydrodynamics or shallow water theory.
The reason of considering a density-dependent viscosity, which vanishes in vacuum regions, is twofold.
The first motivation comes from modeling purposes: indeed, in general the viscosity of a fluid depends both on density and temperature, even in the Newtonian case (see e.g. \cite{Chap-Cow}).
The second one is purely mathematical: although this choice complicates the analysis, since no information can be deduced
on the velocity field $u$ when the density $\rho$ vanishes, it enables us to exploit a fundamental underlying structure of the equations, the so-called \emph{BD entropy structure},
which in turn allows to gain additional regularity for $\rho$ from the capillarity term. We refer to Subsection \ref{ss:cap-effects} below for more details.
Finally, the additional Coriolis term comes from the fact that we are interested in dynamics of geophysical fluids, for which effects due to Earth rotation cannot be neglected.
At the mathematical level, its presence in the equations of motion does not involve any problem as far as one is interested in classical energy estimates, which is usually the case. On the contrary,
here it will cause some troubles in the analysis, due to its (apparent) incompatibility with the BD structure. We will explain better this point in Section \ref{s:math_prop}, where
we turn our attention to the mathematical properties of our system.
In the last part of the paper (see Sections \ref{s:asympt} and \ref{s:proof}), we specialize on the physically relevant case of variable rotation axis,
namely depending both on latitude and longitude.
As said at the beginning, we show a result on asymptotic behaviour of weak solutions, performing together the fast rotation and incompressible limits in the regime of constant capillarity.
Besides, we refine the result of \cite{F_2016} concerning minimal regularity assumptions on the fluctuation of the axis, introducing Zygmund-type conditions on the gradient of
the variation function. We refer to the above mentioned works \cite{F_2015}-\cite{F_2016} for further results, for an overview of related studies and additional references.
Let us complete this brief introduction by a detailed description of the physical effects we want to capture here, namely Earth rotation and surface tension.
\subsection{Earth rotation and the Coriolis force} \label{ss:i_rot}
Coriolis force is an inertial force, which shows up because of the choice of a rotating reference frame for describing a motion.
It causes a travelling object to curve its trajectory: actually this deflection is just apparent, and it is ``seen'' only by an observer in the rotating frame.
Coriolis force depends on the rotation speed of the reference frame and on the speed of the object through their cross product;
in particular, for motions on the Earth it depends on the projection of the rotation axis on the local vertical (i.e. the latitude).
Obviously, Earth is a rotating frame, and we have to keep into account this kind of effects when describing an event: indeed Coriolis force can be experienced in several ways in everyday life.
Its consequences are especially evident in the motion of geophysical fluids, due to the large scales
involved (see also the discussion below). In fact, rotation is one of the two main ingredients which distinguish geophysical flows, the other
one being stratification.
Quite surprisingly, centrifugal forces can be completely neglected at this level, since it is by far exceeded by gravity, which essentially rubs out its effects (up to small corrections
in the geometry of the planet). We refer to \cite{CR} for further details.
We can justify the relevance of Coriolis force in geophysical fluid dynamics by arguments based on orders of magnitude, and in fact one can speak about \emph{fast rotation}, since the
time of the physical process is much larger than the one taken by the Earth to make a revolution. This will be our approach throughout this work;
we will give more details in Subsection \ref{ss:singular}. For the time being, let us make a brief overview of the main physical phenomena determined by the Earth rotation.
The first remarkable effect is that fast rotation induces a sort of rigidity on the fluid motion, as discovered first by the physicist Taylor. More precisely,
it creates an asymmetry between vertical (i.e. parallel to the rotation axis) and horizontal
(i.e. in the plane orthogonal to the rotation axis) components of the velocity field, and the fluid particles tend to move on vertical columns.
Of course, such a vertical rigidity (known as \emph{Taylor-Proudman theorem}) is not perfectly realized in oceanic and athmospheric circulation, mainly because the rotation is not fast
enough and the density not uniform enough to hide other ongoing physical processes.
These discrepancies, as well as other deviations from idealized configurations, generate and propagate along \emph{waves}, which can be of various nature: one speaks about Kelvin, Poincar\'e,
Rossby and topographic waves. We refer to \cite{CR} and \cite{Ped} for a more indeep discussion. As far as we are concerned, we will deal with Poincar\'e and Rossby waves:
the former are due to non-vanishing vertical components of the velocity field of the fluid, while the latter are due to ``variations'' of the rotation axis together with latitude
(we will be more precise later on, see in particular Subsection \ref{ss:rot-effects}).
Actually, this terminology does not completely fit the cases we encounter here: indeed, we will consider slightly compressible fluids, which produce then acoustic waves
(as it is well-known, see e.g. \cite{Light}) due to the compressible part of the velocity field. Hence, our analysis will rather reveal the interactions between these
different classes of waves (acoustic and geophysical). Let us remark that, although having variable density, our fluids are not exactly stratified in the common sense
of geophysics, since we neglect here the gravitational force for the sake of simplicity: so, in particular we will not see vertical motions.
The vertical rigidity sancioned by the Taylor-Proudman theorem is somehow incompatible also with the physical condition which prescribes the fluid to be at rest at
the boundary (i.e. Dirichlet boundary conditions). Such an apparent incompatibility is actually explained by the formation of boundary layers, called \emph{Ekman layers}, which
are due to vertical friction. Vertical friction effects are very small if compared with rotation (this assertion can be justified again by scaling arguments),
and they can be omitted in a first approximation, but they cause the motion, roughly speaking, to present two distinct behaviours: in the interior friction is usually negligible,
while near the boundary and across a small distance (which is in fact the boundary layer) it acts to slow down the interior velocity up to zero.
Physical considerations show that this process is not purely horizontal, but it creates a vertical velocity, the so-called \emph{Ekman pumping},
which occurs throughout the depth of the fluid. This vertical component of the velocity produces a global circulation effect from the interior to the Ekman
layer and conversely, which is reponsible for the dissipation of a huge amount of kinetic energy.
As a final remark, we point out that actually two factors (among others) account for substantial corrections to this quite idealized description: turbulence
(geophysical flows have very large Reynolds number) and stratification. We do not detail this point here, for which we rather refer to Chapter 4 of \cite{Ped} and
Chapter 5 of \cite{CR}.
\medbreak
Let us now consider fluids which are influenced by the presence of a large surface tension.
\subsection{Capillarity: some physical insights} \label{ss:i_cap}
Capillarity is a physical effect which is intimately linked with surface tension of a fluid, and indeed these two words (i.e. capillarity and surface tension) are often used as synonims,
at least by mathematicians. This will be also our point of view throughout all this paper.
Capillary phenomena are connected with large variations of the density function in small regions: for instance whenever different fluids touch each other, or when matter presents different states.
In everyday life, capillarity can be commonly experienced observing fluids in thin tubes: this sentence includes a lot of simple events, like drinks filling a straw or
being absorbed by a paper (or other material) towel. Likewise, surface tension is responsible for droplets formation and breakup. Capillarity is suspected to play a fundamental role
even in phenomena linked with superfluids (like e.g. liquid helium or ultracold atomic gases); superfluids appear in astrophysics, high-energy physics and quantum gravity theories.
In a more rigorous way, from the physical viewpoint surface tension plays a fundamental role in describing phenomena which occur at the interface between
two (miscible or immiscible) fluids, see e.g. \cite{A-McF-W}.
In classical fluid mechanics approach, the interface is assumed to be of zero thickness and it is modeled then as an evolving free boundary. This description, although being successfull
in many contexts, turns out to fail in various situations: in particular, it breaks down whenever the physical process takes place at lenght scales which are comparable to
the thickness of the interface. In this case, \emph{diffuse-interface} models provide an alternative description, which is able to overtake problems of that kind. The leading idea of
these models is that quantities which were localized in the interfacial surface before, are now distributed throughout a region of possibly very small but non-zero thickness.
It is exactly in this interfacial region that density experiences steep (but still smooth) changes of value.
Diffuse-interface models are used also in the modern theory of capillarity. Previously (at around the beginning of the XIX century), capillary phenomena were explained as
the result of the interaction between short-ranged attraction forces and ``repulsive forces'' between molecules; these latter were in fact interpreted by Laplace as
a byproduct of an internal intermolecular pressure. Laplace's description evolved through Maxwell and Lord Rayleigh up to van der Waals: although being based on a static
view of matter, this theory is in fact able to explain many physical phenomena in a quite accurate way. We refer to Chapter 1 of \cite{Row-Wid} for many additional details.
From both points of view (the one of short-ranged interactions and the other one of diffuse-interfaces), the relevant feature is that density undergoes large variations in regions of small thickness.
This was already remarked by Korteweg, who proposed a new constitutive law for the stress tensor of the fluid under consideration:
an ``elastic'' component had to be added to the classical form of Cauchy and Poisson, which had to depend on the density and its derivatives up to second order
(which corresponds, in modern terminology, to fluids of grade $2$). Adopting a rational mechanics approach, in \cite{D-S} Dunn and Serrin showed the incompatibility of
the form proposed by Korteweg with the classical principle of thermodynamics, and they proposed a new version of the capillarity stress tensor, which can actually be generalized to
model fluids of any grade $N$.
Finally, it is the Dunn--Serrin form of the Korteweg stress tensor which is used in mathematical studies: see e.g. \cite{B-D-L}, \cite{B-D_2003} and \cite{BG-D-Desc_2007} (this last reference
concerns the case of inviscid fluids); we also refer to \cite{BG-D-Desc-J} for the study of traveling waves and stability issues linked with inviscid capillary flows, and
to \cite{B-C-N-V} for an interesting reformulation of the Korteweg tensor and related properties.
In Subsection \ref{ss:korteweg} we will derive the Dunn--Serrin form of the Korteweg tensor by use of variational arguments, in the special case of fluids of grade $2$.
We refer to the introductions of \cite{F_2015}-\cite{F_2016} for further references and mathematical results concerning related capillary models.
\subsection{Relevance of mixing rotation and capillarity}
We conclude the introduction by justifying our study, where we look at both rotation and surface tension effects in the motion.
First of all, we remark that the diffuse-interface approach is relevant also for single component fluids with variable density. Therefore, it is pertinent
to consider a stress tensor of Korteweg type in the context of geophysical flows, assuming that internal forces generate a significant part of the internal energy
of the fluid.
Notice also that interface mechanisms play a relevant role in propagation of internal waves; keeping them into account is hence fundamental in describing athmosphere and ocean dynamics
(see Chapter 10 of \cite{CR} about this point). Adding a capillarity term in the equations of motion can be seen as a (maybe rough, but still
interesting, from our point of view) attempt of capturing these phenomena.
Finally, as we will see in detail in Subsection \ref{ss:singular}, our equations present a strong similarity with a $2$-D shallow water system, used to describe the so-called
\emph{betaplane model} for geophysical dynamics close to the equatorial zone (see e.g. \cite{B-D-Met}, \cite{G} and \cite{G-SR_Mem}). Our study can be put in relation with those works:
here we consider a $3$-D domain and a degenerate viscosity coefficient.
\subsubsection*{Acknowledgements}
The study presented in this note has been widely enriched by interesting discussions with \textit{Sylvie Benzoni-Gavage}, \textit{Didier Bresch} and \textit{Isabelle
Gallagher}. The author wants to express all his gratitude to them.
This work was partially supported by the LABEX MILYON (ANR-10-LABX-0070) of Universit\'e de Lyon, within the program ``Investissement d'Avenir''
(ANR-11-IDEX-0007) operated by the French National Research Agency (ANR).
\section{A model for rotating capillary fluids} \label{s:model}
In this section we introduce our model, namely a Navier-Stokes-Korteweg system with Coriolis force.
It presents a term which keeps into account capillarity effects, the so-called Korteweg stress tensor, and another one which is due to the rotation of the Earth.
We start by reviewing some physical background and explaining how these effects can be translated at the mathematical level.
This will lead us to write down the equations of motion in Subsection \ref{ss:motion}. In the final Subsection \ref{ss:singular},
we will adopt the singular perturbation analysis point of view, presenting the rescaled system.
\subsection{On the Korteweg stress tensor} \label{ss:korteweg}
Let us consider first surface tension effects on fluid motion.
As pointed out in the introduction, historically capillarity phenomena were described as a consequence of short-ranged (attractive and repulsive) forces which are produced between
molecules. On the other hand, more recent theories are based on a diffuse-interface approach, where however the interfacial region can have a very small thickness.
In both contexts, it is then quite natural to assume that the classical stress tensor of the fluid under consideration has to be modified, in order to include
terms which depend on the gradient and higher order derivatives of the density. Here, however, we want to take a different point of view: namely, we are going to derive its form (one of the
possible choices) by variational principles, postulating just the specific expression of the free energy. We refer to \cite{A-McF-W} and \cite{BG_Levico} for more details.
For simplicity, let us consider a non-uniform single-component fluid at equilibrium, and let us denote by $\rho$ its density and by $\theta$ its temperature.
We suppose that the \emph{Helmholtz free energy} functional takes the form
\begin{equation} \label{eq:Helmholtz}
\mathcal{E}[\rho,\theta]\,:=\,\int_\Omega\left(\rho\,e_0(\rho,\theta)\,+\,\frac{1}{2}\,k(\rho,\theta)\,|\nabla\rho|^2\right)\,dx
\end{equation}
in some domain $\Omega\subset\mathbb{R}^3$, where $dx$ is the volume measure.
Of course, for a fluid not necessarily at rest, the contribution coming from the kinetic energy
$$
\mathcal{E}_k[\rho,u]\,:=\,\frac{1}{2}\,\int_\Omega\rho\,|u|^2\,dx
$$
has to be added to $\mathcal{E}$, where $u\in\mathbb{R}^3$ represents the velocity field of the fluid.
In formula \eqref{eq:Helmholtz}, the function $e_0$ is the bulk free energy density per unit mass, and the first term in the integral, i.e.
$$
\mathcal{E}_0[\rho,\theta]\,:=\,\int_\Omega\mathcal L_0[\rho,\theta]\,dx\,,\qquad\qquad\mbox{ with }\qquad \mathcal L_0[\rho,\theta]\,:=\,\rho\,e_0(\rho,\theta)\,,
$$
represents the ``classical'' energy of a fluid at rest. The second term in the integral takes into account the presence of capillarity effects in the fluid,
in which large density variations are no more negligible and contribute to free energy excess of the interfacial region. The positive function $k(\rho,\theta)$ is the internal
energy coefficient, also called capillarity coefficient; it is possible to give it a more precise meaning in the context of statistical mechanics,
but this point goes beyond the scopes of this discussion, and we will not say more about it here.
From now on, we restrict our attention to the case of \emph{isothermal fluids}. So let us set (with no loss of generality) $\theta\equiv1$ and forget about it in the notations.
The equilibrium conditions are obtained by minimizing the function $\mathcal{E}$ under the constraint of constant mass $\mathcal M\,:=\,\int_\Omega\rho\,dx$.
Therefore, defined the Lagrangian function
\begin{equation} \label{eq:E-L}
\mathcal{L}[\rho,\nabla\rho]\,:=\,\rho\,e_0(\rho)\,+\,\frac{1}{2}\,k(\rho)\,|\nabla\rho|^2\,-\,\lambda\,\rho
\end{equation}
(where $\lambda$ is a Lagrange multiplier), by Hamilton's principle one finds the Euler-Lagrange equation
$$
\frac{1}{2}\,k'(\rho)\,|\nabla\rho|^2\,+\,k(\rho)\,\Delta\rho\,-\,\partial_\rho\bigl(\rho\,e_0(\rho)\bigr)\,+\,\lambda\,=\,0\,.
$$
On the other hand, both $\mathcal L_1[\rho,\nabla\rho]\,:=\,\mathcal{L}_0[\rho]\,+\,k(\rho)\,|\nabla\rho|^2/2$ and the mass constraint are independent of spatial coordinates,
and therefore they are invariant by the action of the vector fields $\partial_j$, for all $1\leq j\leq 3$. Then, by application of Noether's theorem we get the relation
${\rm div}\, J\,=\,0$, where we have defined the $2$-tensor
$$
J\,:=\,\mathcal{L}\,{\rm Id}\,\,-\,\nabla\rho\,\otimes\,\frac{\partial\mathcal L}{\partial(\nabla\rho)}\,.
$$
Replacing $\mathcal L$ by its definition and using \eqref{eq:E-L} to find $\lambda$, we arrive at the expression
\begin{equation} \label{eq:static-stress}
J\,=\,-\,k(\rho)\,\nabla\rho\otimes\nabla\rho\,+\,\left(-\,\rho^2\,e_0'(\rho)\,+\,\frac{1}{2}\bigl(k(\rho)\,+\,\rho\,k'(\rho)\bigr)\,|\nabla\rho|^2\,+\,\rho\,k(\rho)\,\Delta\rho\right){\rm Id}\,\,.
\end{equation}
We remark that $\Pi\,:=\,\rho^2\,e_0'(\rho)$ is the standard thermodynamic pressure. It is classically given by $\Pi\,=\,\rho\,\partial_\rho\mathcal L_0\,-\,\mathcal L_0$; this formula allows us to
define a \emph{generalized pressure} for capillary fluids as
$$
\Pi_g\,:=\,\rho\,\partial_\rho\mathcal L\,-\,\mathcal L\,=\,\Pi\,+\,\frac{1}{2}\,\bigl(\rho\,k'(\rho)\,-\,k(\rho)\bigr)\,|\nabla\rho|^2\,.
$$
Then, with this definition, $J$ assumes the form
$$
J[\rho,\nabla\rho]\,=\,-\,k(\rho)\,\nabla\rho\otimes\nabla\rho\,+\,\Bigl(-\,\Pi_g\,+\,k(\rho)\,{\rm div}\,\bigl(\rho\,\nabla\rho\bigr)\Bigr)\,{\rm Id}\,\,.
$$
Let us mention that $J$, given by formula \eqref{eq:static-stress} and derived here in the static case, actually represents
the reversible part of the stress tensor for a capillary fluid which is not at equilibrium (of course, viscosity is missing in the present discussion).
We refer to \cite{A-McF-W} for additional details about this point, which will be used in Subsection \ref{ss:motion} to write the equations of motion.
Finally, for later use, we define the \emph{Korteweg stress tensor} as
\begin{equation} \label{eq:K-stress}
K[\rho,\nabla\rho]\,:=\,-\,k(\rho)\,\nabla\rho\otimes\nabla\rho\,+\,\frac{1}{2}\Bigl(\bigl(k(\rho)\,+\,\rho\,k'(\rho)\bigr)\,|\nabla\rho|^2\,+\,\rho\,k(\rho)\,\Delta\rho\Bigr)\,{\rm Id}\,\,.
\end{equation}
\begin{rem} \label{r:k}
Of course, different forms of $K[\rho,\nabla\rho]$ are possible, depending on the form of the interfacial energy $\mathcal E\,-\,\mathcal E_0$. This leads to several models, which
are relevant in various contexts, like e.g. in quantum hydrodynamics or in shallow water theory.
We refer to \cite{BG_Levico} and to the introductions of \cite{F_2015}-\cite{F_2016} for more details about this point.
\end{rem}
\begin{rem} \label{r:phase-b}
The previous form of the total energy \eqref{eq:Helmholtz} has special importance when $\mathcal L_0$ is \emph{not} a convex function of $\rho$, like
e.g. for van der Waals equations of state. In these cases, $\mathcal E$ may admit minimizers which are not necessarily constant density profiles and for which, in particular,
two different phases are permitted to coexist. This allows one to capture phase boundaries and their dynamical evolution (see also Section 1 of \cite{BG_Levico} for
a more indeep discussion).
\end{rem}
\subsection{Adding the rotation of the Earth} \label{ss:rotation}
We now turn our attention to Earth rotation and its mathematical description.
Given two frames $\mathcal F$ and $\mathcal F'$ in non-intertial relative motion, let us call $x(t)$ the position of a moving body with respect to the first origin
(say, the ``absolute'' position) and $x'(t)$ the position with respect to the second one (say, the position in the ``non-inertial'' frame). Denote by $o(t)$ the vector position of the origin
of $\mathcal F'$ with respect to the origin of $\mathcal F$,
and suppose also that $\mathcal F'$ is rotating with respect to $\mathcal F$, so that
$$
x(t)\,=\,o(t)\,+\,R(t)\,x'(t)\,,
$$
for some rotation matrix $R(t)$. We set $v(t)$ and $v'(t)$ to be the velocities in $\mathcal F$ and $\mathcal F'$ respectively, and $a(t)$ and
$a'(t)$ to be the accelerations: it is a standard matter (see e.g. \cite{Arnold} or \cite{Tartar}) to derive, from principles of Newtonian mechanics, the formula
$$
a(t)\,=\,a_o(t)\,+\,R(t)\Bigl(a'(t)\,+\,\xi(t)\times v'(t)\,+\,\dot{\xi}(t)\times x'(t)\Bigr)\,+\,R(t)\,\xi(t)\times\Bigl(v'(t)\,+\,\xi(t)\times x'(t)\Bigr)\,,
$$
where we have denoted by $a_o$ the relative acceleration of $\mathcal F'$ with respect to $\mathcal F$, by $\xi(t)$ the vector associated to the
skew-symmetric matrix $\,^tR(t)\,\dot{R}(t)$ (since we are in $\mathbb{R}^3$), and by a dot the time derivative.
The term $\xi\times(\xi\times v')$is related, together with the relative acceleration $a_0$, to the centrifugal force. It can be easily seen that it derives from a potential,
which slightly changes the gravitational potential creating the so-called geopotential. As we have said in Subsection \ref{ss:i_rot}, this term can be safely neglected
in a first approximation. The term $\dot{\xi}\times x'$ can be ingnored as well at this stage, since we suppose no variations in time of the Earth rotation.
Finally, the term $2\,\xi\times v'$ is the \emph{Coriolis acceleration}; notice that, being orthogonal to the velocity, it makes no work and so
it is not seen in the energy balance (also, energy being an invariant, its form does not depend on the choice of the reference frame).
To sum up, after the previous approximations, the absolute acceleration reduces to the formula
\begin{equation} \label{eq:accel}
a(t)\,\simeq\,R(t)\Bigl(a'(t)\,+\,2\,\xi(t)\times v'(t)\Bigr)\,.
\end{equation}
Let us now place on the surface of the Earth, which we approximate to a perfect sphere (we neglect again effects due to centrifugal force). We also assume that it rotates
around its north pole-south pole axis: at any given latitude $\varphi$, this direction departs from the local vertical of the angle $\pi/2-\varphi$, and the Coriolis
force consequently varies. More precisely, let us fix a local chart on the Earth surface, such that the $x^1$-axis describes the longitude, oriented eastward, the $x^2$-axis describes the latitude,
oriented northward, and the $x^3$-axis describes the altitude, oriented upward; let us denote by $(e^1,e^2,e^3)$ the corresponding orthonormal basis.
Then, the Earth's rotation vector can be written in the following way:
$$
\xi\,=\,|\xi|\,\bigl(\cos\varphi\,e^2\,+\,\sin\varphi\,e^3\bigr)\,.
$$
We are now going to make several assumptions.
First of all, for the sake of simplicity we neglect the curvature of the Earth, so that spherical coordinates are treated as cartesian coordinates. This approximation
is justified if we restrict our attention on regions of our planet which are not too extended with respect to the radius of the Earth (see also \cite{SR}).
In addition, we suppose that the rotation axis is parallel to the $x^3$-axis (i.e. we take $\varphi=\pi/2$), which is quite reasonable if we place far enough from the equatorial zone.
Nonetheless, in order to still keep track of the variations due to the latitude, we postulate the presence of a suitable smooth function $\mathfrak{c}$ of the ``horizontal'' variables
$x^h\,=\,(x^1,x^2)$.
Therefore, under the previous assumptions we can write
$$
\xi\,=\,\mathfrak{c}(x^h)\,e^3\,.
$$
In the light of the previous discussion, the equations of motion will be perturbed by adding one term (see the next subsection), i.e. the Coriolis operator
\begin{equation} \label{eq:coriolis}
\mathfrak{C}(\rho,u)\,:=\,\mathfrak{c}(x^h)\,e^3\,\times\,\rho\,u\,,
\end{equation}
where, as in the previous subsection, $\rho$ is the density of the fluid
and $u$ its velocity field. We remark that, since Coriolis force makes no work, this term does not contribute to the energy balance, neither to the pontential part $\mathcal E$ nor to
the kinetic part $\mathcal E_k$ (both defined in the previous subsection).
\begin{rem} \label{r:coriolis}
Notice that, in view of our assumptions, we have arrived at a quite simple form of the Coriolis operator. This form is nonetheless widely accepted in the mathematical community,
and it already enables one to capture different physical effects linked with rotation (see e.g. book \cite{C-D-G-G} and the references therein, and works \cite{G-SR_2006} and \cite{G-SR_Mem},
among many others).
Treating the most general form of the Coriolis operator goes beyond the scopes of our presentation.
\end{rem}
Precise hypotheses on the variation function $\mathfrak c$ will be described in Subsections \ref{ss:hypotheses} and \ref{ss:var-result} below.
There, we will introduce also further simplifications to the model, in order to tackle the mathematical problem.
For the time being, let us introduce the relevant system of equations.
\subsection{Writing down the equations of motion} \label{ss:motion}
In the present subsection we finally present the system of PDEs which describes the physical phenomena we detailed above.
Some preliminary simplifications are in order, to deal with an accessible mathematical problem.
\subsubsection{Some assumptions} \label{sss:mot_assump}
As already said, we omit here effects due to centrifugal and gravitational forces. We will also introduce (see Subsection \ref{ss:var-result}) some restrictions on the variations of
the rotation axis, namely on the function $\mathfrak c$ of formula \eqref{eq:coriolis}.
Moreover, we do not consider other relevant physical quantities: e.g. temperature variations, salinity (in the case of oceans), wind force (on the surface of the ocean).
Therefore, our fluid will be described by its density $\rho\geq0$ and its velocity field $u\in\mathbb{R}^3$. Moreover, we will restrict our attention to the case of
\emph{Newtonian fluids}, for which the viscous stress tensor is linearly dependent on the gradient of the velocity field: we suppose it to be given by the relation
$$
S[\rho,u]\,=\,\nu_1(\rho)\,Du\,+\,\nu_2(\rho)\,{\rm div}\, u\,{\rm Id}\,\,,\qquad\qquad\mbox{ with }\qquad Du\,=\,\nabla u\,+\,^t\nabla u\,,
$$
for suitable functions $\nu_1$ and $\nu_2$ of the density only. As already pointed out, considering density--dependent viscous coefficients is very important
at the level of mathematical modeling (but again, we are missing effects linked with temperature variations).
Here and until the end of the paper, we make the simple choice $\nu_1(\rho)\,=\,\nu\,\rho$ and $\nu_2(\rho)\equiv0$, for some constant $\nu>0$.
Finally, let us immediately fix the spacial domain: neglecting the curvature of the Earth surface for simplicity (as remarked in the previous subsection),
we consider evolutions on the infinite slab
$$
\Omega\,:=\,\mathbb{R}^2\,\times\;]0,1[
$$
(but the main result, i.e. Theorem \ref{t:sing_var} below, works also on $\mathbb{T}^2\,\times]0,1[\;$). Hence, we are going to consider a very simple geometry of the domain, which in particular
introduce some rigidity at the boundary: we ignore variations at the surface of the Earth for atmospheric circulation, and for ocean in its topografy and at the free surface
(what is called the \emph{rigid lid approximation}).
Let us reveal in advance that, for the sake of simplicity, we want to avoid boundary layers formation in the present study, and rather focus on the other physical effects
linked with fast rotation (recall the discussion in Subsection \ref{ss:i_rot}). Therefore, in dealing with the singular perturbation problem, we will take \emph{complete slip}
boundary conditions for $\Omega$: we refer to Subsection \ref{ss:hypotheses} for the precise assumptions.
\subsubsection{The system of equations} \label{sss:eq}
Thanks to these hypotheses, we can now write down the equations describing the dynamics of our geophysical flow. We do not present here the details of the precise derivation, which can
be found in many textbooks: see e.g. \cite{Feir}, \cite{F-N} and \cite{Gill}.
The first relation is deduced from the principle of mass conservation, namely $\mathcal M\,\equiv\,cst$, where $\mathcal M$ has been introduced in Subsection \ref{ss:korteweg}.
Differentiating it with respect to time and using the divergence theorem, one gets the equation
\begin{equation} \label{eq:mass}
\partial_t\rho\,+\,{\rm div}\,\bigl(\rho\,u\bigr)\,=\,0\,.
\end{equation}
Gathering conservation of momentum is just a matter of applying Newton's second law $F\,=\,m\,a$, with the acceleration $a$ which is given
by \eqref{eq:accel} and with the mass $m$ replaced by the density (mass per unit of volume) $\rho$. Passing in Eulerian coordinates, so that $d/dt\,=\,\partial_t\,+\,u\cdot\nabla$,
and denoting by $\Psi$ the stress tensor\footnote{We recall that the \emph{stress tensor} yields the force per unit of surface which the part of the fluid in contact with an ideal surface element
imposes on the other part of the fluid on the other side of the same surface element.}, we find
$$
\rho\,\bigl(\partial_tu\,+\,u\cdot\nabla u\bigr)\,+\,\mathfrak{C}(\rho,u)\,=\,{\rm div}\,\Psi\,,
$$
where we have also made use of \eqref{eq:coriolis}. By previous assumptions, in our case $\Psi$ is given by a more general form of the Stoke law, namely
$$
\Psi\,=\,S[\rho,u]\,+\,J[\rho,\nabla\rho]\,=\,S[\rho,u]\,-\,\Pi(\rho)\,{\rm Id}\,\,+\,K[\rho,\nabla\rho]\,,
$$
where $J$ and $K$ have been defined in \eqref{eq:static-stress} and \eqref{eq:K-stress} respectively. Recall that $\Pi$ is the thermodynamical pressure of the fluid:
we will specify its precise form in Subsection \ref{ss:hypotheses} below.
At this point, we make the choice $k(\rho)\equiv k$ constant in \eqref{eq:K-stress}, which we can take equal to $1$ with no loss of generality.
It is not just a simplification: at the mathematical level, this precise form of the Korteweg stress tensor is exactly the one which combines well with the previous fixed values of the
viscosity coefficients $\nu_1$ and $\nu_2$, in order to exploit the BD structure of our system (recall the discussion in the Introduction, and see Paragraph \ref{sss:BD} below for more details).
Finally, making use of \eqref{eq:mass} and of the specific form of $K[\rho,\nabla\rho]$, it is easy to check that we arrive at the balance law
\begin{equation} \label{eq:momentum}
\partial_t\left(\rho u\right)\,+\,{\rm div}\,\bigl(\rho u\otimes u\bigr)\,+\,\,\nabla\Pi(\rho)\,-\,\nu\,{\rm div}\,\bigl(\rho Du\bigr)\,-\,\rho\nabla\Delta\rho\,+\,\mathfrak{C}(\rho,u)\,=\,0\,,
\end{equation}
which expresses the conservation of momentum.
\begin{rem} \label{r:phase-b_2}
Notice that we can read system \eqref{eq:mass}-\eqref{eq:momentum} as a dynamical description
of propagation of interfaces and phase boundaries. Recall also the discussions in Subsections \ref{ss:i_cap} and \ref{ss:korteweg}, and in particular Remark \ref{r:phase-b}.
\end{rem}
\subsection{The singular perturbation viewpoint} \label{ss:singular}
We have presented above the equations of motion we are interested in, namely \eqref{eq:mass}-\eqref{eq:momentum}.
Here we want to introduce relevant physical adimensional parameters in the equations, namely the Rossby, Mach and Weber numbers. Sending these parameters to $0$ will allow us to perform an
asymptotic analysis of the fast rotation and incompressible limits, in the regime of constant capillarity.
Let us first explain the main motivations for a singular perturbation study.
\subsubsection{Motivations} \label{sss:sing_motiv}
To assess the importance of some process in a particular situation, it is common practice in Physics to introduce and compare dimensional quantities
expressing the orders of magnitude of the variables under consideration, and to ignore the ``small'' ones.
More precisely, performing a rescaling in the equations allows to select dimensionless parameters: sending them to $0$ or to $+\infty$ leads to some simplifications
in the equations and possibly to a better understanding of the main phenomena which take place. The limit process
means that we are overlooking some negligible features of the experience, or focusing on some special ones.
Scale analysis is an efficient tool exploited both theoretically and in numerical experiments: it allows to reduce the complexity of the physical system under consideration, selecting just
relevant quantities, or the computational cost of simulations. In addition, it reveals to be very fruitful in real world applications.
This approach is expecially relevant for large-scale processes, like geophysical flows we want to consider here.
Actually most, if not all mathematical models used in fluid mechanics (like e.g. incompressible Navier-Stokes equations) rely on an asymptotic analysis of
more complicated systems.
Nonetheless, a huge part of the available literature on scale analysis is based on formal asymptotic limits of (supposed to exist) solutions with respect to one or more singular parameters.
The main goal of the mathematical approach to singular perturbation problems is to provide a rigorous justification of the limit model employed in the physical approximation, and possibly to
capture corrections (the equations being non-linear in general, amplification of small quantities might occur) or further underlying effects (like e.g. the way the
quantities which are negligible in a certain regime are filtered off in the asymptotic limit).
\subsubsection{The Rossby, Mach and Weber numbers} \label{sss:sing_numb}
In view of the previous discussion, in the present paragraph our aim is to identify the physical parameters which are relevant with respect to the phenomena
we want to capture.
Let us consider rotation first. To establish its importance, a \textsl{na\"if} approach may be to compare time scales, and hence the time $\tau$ of one revolution
to the life $T$ of the entire physical process. But it is easy to realize that in geophysical flows also lenght scale can play an important role. Therefore, it is customary to take
rather the ratio between the time of one revolution and the time required to cover a distance $L$ at speed $U$.
More precisely, denoted by $L$ the typical lenght scale, by $U$ the typical velocity scale and by $\Theta$ the rotation rate (namely, $\Theta\,=\,2\pi/\tau$),
we define the \emph{Rossby number} as the ratio
$$
Ro\,:=\,\frac{U}{\Theta\,L}\,.
$$
It compares advection to Corilolis force: rotation plays an important role whenever $Ro\,\lesssim\,1$ (see also Sections 1.5 and 3.6 of \cite{CR}).
A common simplification in geophysical fluid dynamics (see e.g. \cite{CR} and \cite{Ped}) is the Boussinesq approximation, which states the incompressibility of the velocity
field. Here, we generlize this assumption (as it is quite usual, see e.g. \cite{F-G-N}-\cite{F-G-GV-N}) by considering slightly compressible fluids.
The compressibility is measured by the so-called \emph{Mach number}, defined as
$$
Ma\,:=\,\frac{U}{c}\,,
$$
where $c$ denotes the speed of sound in the medium: more $Ma$ is small, more the fluid behaves as incompressible. This number will appear in front of the pressure term.
We refer also to \cite{F-N} for further details.
Finally, to evaluate the importance of capillarity, we compare typical inertial forces to stabilizing molecular cohesive forces, or (which is the same) kinetic energy to surface
tension energy. Hence, we introduce the \emph{Weber number} as
$$
We\,:=\,\frac{\rho_0\,U^2\,L}{\sigma}\,,
$$
where we set $\rho_0$ to be the reference density of the fluid and $\sigma$ the typical surface tension. Having $We$ small means that capillarity forces are relevant
in the ongoing physical process.
\subsubsection{Introducing the scaling in the equations} \label{sss:sing_eq}
We now introduce the previous parameters in the mathematical model, getting a rescaled system.
The mass equation \eqref{eq:mass} is not modified by this scale analysis: the Rossby, Mach and Weber numbers come into play in the momentum balance only.
More precisely, equation \eqref{eq:momentum} becomes
$$
\partial_t\left(\rho u\right)\,+\,{\rm div}\,\bigl(\rho u\otimes u\bigr)\,+\,
\dfrac{1}{Ma\,^2}\,\nabla\Pi(\rho)\,-\,\nu\,{\rm div}\,\bigl(\rho Du\bigr)\,-\,
\dfrac{1}{We}\,\rho\nabla\Delta\rho\,+\,\dfrac{1}{Ro}\,\mathfrak{C}(\rho,u)\,=\,0\,.
$$
As a final step, we take the point of view of the singular perturbation analysis. Namely, we fix a parameter
$\varepsilon\,\in\,]0,1]$, and we set
$$
Ma\,=\,Ro\,=\,\varepsilon\qquad\qquad \mbox{ and }\qquad\qquad We\,=\,\varepsilon^2\,.
$$
Accordingly, the previous momentum equation becomes
\begin{equation} \label{eq:mom-sing}
\partial_t\left(\rho u\right)\,+\,{\rm div}\,\bigl(\rho u\otimes u\bigr)\,+\,\dfrac{1}{\varepsilon^2}\,\nabla\Pi(\rho)\,-\,\nu\,{\rm div}\,\bigl(\rho Du\bigr)\,-\,
\dfrac{1}{\varepsilon^2}\,\rho\nabla\Delta\rho\,+\,\dfrac{1}{\varepsilon}\,\mathfrak{C}(\rho,u)\,=\,0\,.
\end{equation}
For all $\varepsilon$ fixed, thanks to previous results on Navier-Stokes-Korteweg system (Coriolis force can be easily handled at this level),
we know the existence of a weak solution $(\rho_\varepsilon,u_\varepsilon)$ to equations \eqref{eq:mass}-\eqref{eq:mom-sing}.
Our mathematical problem is then to study the asymptotic behaviour of the family $\bigl(\rho_\varepsilon,u_\varepsilon\bigr)_\varepsilon$
in the limit $\varepsilon\,\rightarrow\,0$. We recall that this corresponds to making the incompressible and fast rotation limit together, in the regime of constant capillarity.
Indeed, we remark that the same system could be obtained from \eqref{eq:mass}-\eqref{eq:momentum} using also a different approach (see \cite{J-L-W} and \cite{F_2015}),
in order to investigate the long-time behaviour of solutions.
Namely, one can perform the scaling $t\mapsto\varepsilon t$, $u\mapsto\varepsilon u$ and
$\nu\mapsto\varepsilon\nu$, set the capillarity coefficient $k=\varepsilon^{2\alpha}$, for some $0\leq\alpha\leq1$, and divide everything by $\varepsilon^2$ (i.e. the highest power of $\varepsilon$).
One finally bumps exactly into \eqref{eq:mom-sing} when making the choice $\alpha=0$, which means $k=1$: this is why we call this scaling the ``constant capillarity regime''.
Taking different values of $\alpha$ was investigated in \cite{F_2015}: we refer to this paper for results in both the regimes of constant and vanishing capillarity,
for the whole range of vanishing rates (namely, for any fixed $\alpha\in[0,1]$).
We conclude this part by pointing out that different scalings were allowed, in principle, for the Mach and Rossby numbers.
However, on the one hand our choice $Ma=Ro=\varepsilon$ is in accordance with previous related works, see e.g. \cite{F-G-N}, \cite{F-G-GV-N}. This is the right scaling
in order to recover the \emph{geostrophic balance}, i.e. the equilibrium between strong Coriolis force and pressure (see e.g. \cite{CR}, \cite{Ped}, \cite{SR}).
On the other hand, we have to remark the strong analogy between our equations \eqref{eq:mass}-\eqref{eq:momentum} and the viscous shallow water system, studied in e.g. \cite{B-D_2003}-\cite{B-D-Met}
(which however makes sense only in $2$-D, because it derives when averaging vertical motions). We also refer to \cite{G}-\cite{G-SR_Mem} for a related system,
relevant in the investigation of the betaplane model for equatorial dynamics, where however the viscosity coefficients are taken to be constant.
In shallow water equations, the term corresponding to the pressure (density is actually replaced by the depth variation function) is rescaled according to
the so-called \emph{Froude number}: if $N$ denotes the stratification frequency (we have the relation $N^2\,\sim\,\partial_3\rho$) and $H$ the typical depth of the domain, we define
$$
Fr\,:=\,\frac{U}{N\,H}\,.
$$
This number measures the importance of vertical stratification with respect to the (main) horizontal flow.
Pysical considerations reveal that for large-scale motions, in general, the relation $Fr\,^2\,\lesssim\,Ro$ must hold; moreover, a special regime
occurs when rotation and stratification effects are comparable, namely when $Fr\,\sim\,Ro$, a case which corresponds to take the so-called \emph{Burger number} proportional to $1$.
We refer to Chapter 9 of \cite{CR} for more details about this point.
By the parallel between our system and the shallow water system, there is a strict correspondence between $Fr$ and $Ma$: taking then the scaling $Ma=Ro=\varepsilon$ is a way to
capture such a special regime when $Fr\,\sim\,Ro$.
\section{Mathematical properties of the model} \label{s:math_prop}
In the present section we want to give some insights on the main mathematical features our model enjoys.
These properties are quite general, since they are determined by capillarity on the one side and by Earth rotation on the other side. Nonetheless,
for clarity of exposition and for sake of conciseness, we prefer to focus on our special working setting: let us fix our assumptions first.
\subsection{Fixing the working hypotheses} \label{ss:hypotheses}
Let us recall here our working setting, and introduce further important assumptions in order to tackle the mathematical problem.
We fix the domain $\Omega\,:=\,\mathbb{R}^2\,\times\,\,]0,1[\,$. Recall that, accordingly, we split $x\in\Omega$ as $x=(x^h,x^3)$.
In $\mathbb{R}_+\times\Omega$, and for a small parameter $0<\varepsilon\leq1$, let us consider the rescaled Navier-Stokes-Korteweg system
\begin{equation} \label{eq:NSK-sing}
\begin{cases}
\partial_t\rho\,+\,{\rm div}\,\left(\rho u\right)\,=\,0 \\[1ex]
\partial_t\left(\rho u\right)\,+\,{\rm div}\,\bigl(\rho u\otimes u\bigr)\,+\,
\dfrac{1}{\varepsilon^2}\,\nabla\Pi(\rho)\,-\,\nu\,{\rm div}\,\bigl(\rho Du\bigr)\,-\,
\dfrac{1}{\varepsilon^2}\,\rho\nabla\Delta\rho\,+\,\dfrac{1}{\varepsilon}\,\mathfrak{C}(\rho,u)\,=\,0\,.
\end{cases}
\end{equation}
Here, we suppose that the Coriolis operator $\mathfrak C$ is given by formula \eqref{eq:coriolis},
where $\mathfrak{c}$ is a scalar function of the horizontal variables only. For the time being, let us assume
$\mathfrak c\,\in\,W^{1,\infty}(\mathbb{R}^2)$: this hypothesis will allow us to establish uniform bounds in Subsection \ref{ss:cap-effects}.
See also Theorem \ref{t:weak} and the subsequent remark in Subsection \ref{ss:weak}
In addition, we assume that the pressure $\Pi(\rho)$ can be decomposed into $\Pi\,=\,P\,+\,P_c$, where $P$ is the classical
pressure law, given by the Boyle relation
\begin{equation} \label{eq:def_P}
P(\rho)\,:=\,\frac{1}{2\gamma}\,\rho^\gamma\,,\qquad\qquad\mbox{ for some }\quad1\,<\,\gamma\,\leq\,2\,,
\end{equation}
and the second term is the so-called \emph{cold pressure} component, for which we take the expression
\begin{equation} \label{eq:def_cold}
P_c(\rho)\,:=\,-\,\frac{1}{2\gamma_c}\,\rho^{-\gamma_c}\,,\qquad\qquad\mbox{ with }\qquad 1\,\leq\,\gamma_c\,\leq\,2\,.
\end{equation}
For us, $\gamma_c=2$ always. The presence of the $1/2$ is just a normalization in order to have $\Pi'(1)=1$: this fact simplifies some computations in the sequel.
Having a cold component in the pressure law is important in the theory of existence of weak solutions (this is our motivation in assuming it), but it plays no special role in the singular
perturbation analysis.
As said in Subsection \ref{ss:motion}, we supplement system \eqref{eq:NSK-sing} by complete slip boundary conditions,
in order to avoid the appearing of boundary layers effects.
Namely, if $n$ denotes the unitary outward normal to the boundary $\partial\Omega$ of the domain (simply,
$\partial\Omega=\{x^3=0\}\cup\{x^3=1\}$), we impose
\begin{equation} \label{eq:bc}
\left(u\cdot n\right)_{|\partial\Omega}\,=\,0\,,\qquad
\left(\nabla\rho\cdot n\right)_{|\partial\Omega}\,=\,0\,,\qquad
\bigl((Du)n\times n\bigr)_{|\partial\Omega}\,=\,0\,.
\end{equation}
\begin{rem} \label{r:period-bc}
Equations \eqref{eq:NSK-sing}, supplemented by boundary conditions \eqref{eq:bc},
can be recasted as a periodic problem with respect to the vertical variable, in the new domain
$$
\widetilde \Omega\,=\,\mathbb{R}^2\,\times\,\mathbb{T}^1\,,\qquad\qquad\mbox{ with }\qquad\mathbb{T}^1\,:=\,[-1,1]/\sim\,,
$$
where $\sim$ denotes the equivalence relation which identifies $-1$ and $1$.
Indeed, the equations are invariant if we extend $\rho$ and $u^h$ as even functions with respect to $x^3$, and $u^3$ as an odd function.
In what follows, we will always assume that such modifications have been performed on the initial data, and
that the respective solutions keep the same symmetry properties.
\end{rem}
Here we will always consider initial data $(\rho_0,u_0)$ such that $\rho_0\geq0$ and
\begin{equation} \label{eq:initial}
\begin{cases}
\dfrac{1}{\varepsilon}\left(\rho_0\,-\,1\right)\;\in\;L^\gamma(\Omega)\qquad\mbox{ and }\qquad
\dfrac{1}{\varepsilon}\left(\dfrac{1}{\rho_0}\,-\,1\right)\;\in\;L^{\gamma_c}(\Omega) \\[1ex]
\sqrt{\rho_0}\,u_0\;,\;\nabla\sqrt{\rho_0}\;,\;\dfrac{1}{\varepsilon}\nabla\rho_0\quad\in\;L^2(\Omega)\,.
\end{cases}
\end{equation}
When taking a family $\bigl(\rho_{0,\varepsilon},u_{0,\varepsilon}\bigr)_\varepsilon$ of initial data, we will require that conditions \eqref{eq:initial} hold true for any
$\varepsilon\in\,]0,1]$, and \emph{uniformly} in $\varepsilon$ (see also the hypotheses at the beginning of Subsection \ref{ss:var-result}).
At this point, we introduce the internal energy functions $h(\rho)$ and $h_c(\rho)$ in the following way: we require that
$$
\begin{cases}
h''(\rho)\,=\,\dfrac{P'(\rho)}{\rho}\,=\,\rho^{\gamma-2}\qquad\qquad\mbox{ and }\qquad\qquad
h(1)\,=\,h'(1)\,=\,0\,, \\[2ex]
h_c''(\rho)\,=\,\dfrac{P_c'(\rho)}{\rho}\,=\,\rho^{-\gamma_c-2}\qquad\qquad\mbox{ and }\qquad\qquad
h_c(1)\,=\,h_c'(1)\,=\,0\,.
\end{cases}
$$
Let us define the classical energy
\begin{equation} \label{eq:def_E_2}
E_\varepsilon[\rho,u](t) \;:=\;\int_\Omega\left(\dfrac{1}{\varepsilon^2}\,h(\rho)\,+\,\dfrac{1}{\varepsilon^2}\,h_c(\rho)\,+\,
\dfrac{1}{2}\,\rho\,|u|^2\,+\,
\dfrac{1}{2\,{\varepsilon^2}}\,|\nabla\rho|^2\right)dx\,,
\end{equation}
and the BD entropy function
\begin{equation} \label{eq:def_F}
F_\varepsilon[\rho](t)\;:=\;\frac{\nu^2}{2}\int_\Omega\rho\,|\nabla\log\rho|^2\,dx\;=\;
2\,\nu^2\int_\Omega\left|\nabla\sqrt{\rho}\right|^2\,dx\,.
\end{equation}
Finally, let us denote by $E_\varepsilon[\rho_0,u_0]\,\equiv\,E_\varepsilon[\rho,u](0)$ and by
$F_\varepsilon[\rho_0]\,\equiv\,F_\varepsilon[\rho](0)$ the same energies, when computed on the initial data $\bigl(\rho_0,u_0\bigr)$.
\begin{rem} \label{r:en_initial}
Notice that, under our assumptions, we deduce the existence of a ``universal constant'' $C_0>0$ such that
$$
E_\varepsilon[\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}]\,+\,F_\varepsilon[\rho_{0,\varepsilon}]\,\leq\,C_0\,.
$$
\end{rem}
\subsection{Features due to capillarity} \label{ss:cap-effects}
We now enter more in detail in the description of the properties of our model. Here we look at capillarity effects.
The main point is to establish energy estimates (in Paragraph \ref{sss:BD}), which will be of two kinds: classical energy estimates and BD entropy estimates.
\subsubsection{On the BD entropy structure} \label{sss:BD}
As we have already remarked in the Introduction, the combination of capillarity and density-dependent viscosity coefficient gives a special mathematical
structure to our system, the so-called \emph{BD structure}, which consists of a second energy conservation and which in turn allows to exploit the presence of the Korteweg tensor
in the equations, proving higher order regularity estimates for $\rho$.
This additional regularity is fundamental both in the theory of weak solutions and in the study of the singular perturbation problem.
The BD entropy structure was first discovered by Bresch and Desjardins for Korteweg models (see also paper \cite{B-D-L}), but it was then generalized by the same authors to
several other systems having density-dependent viscosity coefficients.
More precisely, the BD structure is a fundamental tool in weak solutions theory for compressible fluids with degenerate (i.e. vanishing with vacuum) viscosity
coefficients. For further details, the reader is referred to \cite{B-D-Zat} and the references therein (see also the introductions of \cite{F_2015}-\cite{F_2016}).
\medbreak
We now present energy estimates and additional uniform bounds for the family of solutions $\bigl(\rho_\varepsilon,u_\varepsilon\bigr)_\varepsilon$ to system \eqref{eq:NSK-sing}.
Since the Coriolis term makes no work, classical energy estimates are easy to find. This is not the case for the BD entropy: the main point is then to control
the rotation term uniformly in $\varepsilon$. We refer to papers \cite{F_2015}-\cite{F_2016} for the proofs of the estimates.
Of course, the precise computations are rigorous for smooth enough solutions: see Subsection \ref{ss:weak} for more details.
We start by proving classical energy estimates.
\begin{prop} \label{p:E}
Let $(\rho,u)$ be a smooth solution to system \eqref{eq:NSK-sing} in $\mathbb{R}_+\times\Omega$, related to the initial datum $\bigl(\rho_0,u_0\bigr)$.
Then, for all $\varepsilon>0$ and all $t\in\mathbb{R}_+$, one has
$$
\frac{d}{dt}E_\varepsilon[\rho,u]\,+\,\nu\int_\Omega\rho\,|Du|^2\,dx\,=\,0\,.
$$
\end{prop}
From the previous statement, we deduce the first class of uniform bounds.
\begin{coroll} \label{c:E}
Let $\gamma_c=2$, and let $\bigl(\rho_\varepsilon,u_\varepsilon\bigr)_\varepsilon$ be a family of smooth solutions to system \eqref{eq:NSK-sing} in $\mathbb{R}_+\times\Omega$, related to
initial data $\bigl(\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}\bigr)_\varepsilon$, and assume that the initial energy $E_\varepsilon[\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}]<+\infty$.
Then one has the following properties, uniformly in $\varepsilon\in\,]0,1]$:
$$
\sqrt{\rho_\varepsilon}\,u_\varepsilon\;,\;\frac{1}{\varepsilon}\,\nabla\rho_\varepsilon\quad\in\;L^\infty\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)\quad\mbox{ and }\quad
\sqrt{\rho_\varepsilon}\,Du_\varepsilon\;\in\;L^2\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)\,.
$$
Moreover, we also get
$$
\frac{1}{\varepsilon}\left(\rho_\varepsilon\,-\,1\right)\,\in\,L^\infty\bigl(\mathbb{R}_+;L^\gamma(\Omega)\bigr)\qquad\mbox{ and }\qquad
\frac{1}{\varepsilon}\left(\frac{1}{\rho_\varepsilon}\,-\,1\right)\,\in\,L^\infty\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)\,.
$$
\end{coroll}
\begin{rem} \label{r:rho_L^2}
In particular, under our assumptions we have
$$
\left\|\rho_\varepsilon\,-\,1\right\|_{L^\infty(\mathbb{R}_+;L^2(\Omega))}\,\leq\,C\,\varepsilon\,.
$$
\end{rem}
BD entropy estimates are harder to establish. Let us present the final estimate, without giving the details on how controlling the Coriolis force uniformly in $\varepsilon$.
We remark that at this point the hypothesis $\nabla_h\mathfrak c\in L^{\infty}$ comes into play in the proof (see also Remark \ref{r:existence} below).
\begin{prop} \label{p:F}
Let $(\rho,u)$ be a smooth solution to system \eqref{eq:NSK-sing} in $\mathbb{R}_+\times\Omega$, related to the initial
datum $\bigl(\rho_0,u_0\bigr)$.
Then, there exists a positive constant $C$, such that, for all $T\in\mathbb{R}_+$ fixed, one has
$$
\sup_{t\in[0,T]}F[\rho](t)\,+\,\frac{\nu}{\varepsilon^2}\int^T_0\!\!\int_\Omega\left|\nabla^2\rho\right|^2\,dx\,dt\,+\,
\frac{\nu}{\varepsilon^2}\int^T_0\!\!\int_\Omega\Pi'(\rho)\,\left|\nabla\sqrt{\rho}\right|^2\,dx\,dt\,\leq\,C\,(1+T)\,.
$$
The constant $C$ depends just on the viscosity coefficient $\nu$ and on the energies of the initial data $E_\varepsilon[\rho_0,u_0]$ and $F_\varepsilon[\rho_0]$.
\end{prop}
Let us point out here that the pressure term can be also written as
$$
\frac{\nu}{\varepsilon^2}\int_\Omega\Pi'(\rho)\,\left|\nabla\sqrt{\rho}\right|^2\,=\,
\frac{C_\gamma\,\nu}{4\,\varepsilon^2}\int_\Omega\left|\nabla\left(\rho^{\gamma/2}\right)\right|^2\,+\,\frac{C_{\gamma_c}\,\nu}{4\,\varepsilon^2}\int_\Omega\left|\nabla\left(\rho^{-\gamma_c/2}\right)\right|^2\,,
$$
for some positive constants $C_\gamma$ and $C_{\gamma_c}$. In particular, when $\gamma_c=2$ (which will be our case), $C_{\gamma_c}=1$.
From Proposition \ref{p:F} we infer the following consequences.
\begin{coroll} \label{c:F}
Let $\gamma_c=2$, and let $\bigl(\rho_\varepsilon,u_\varepsilon\bigr)_\varepsilon$ be a family of smooth solutions to system \eqref{eq:NSK-sing} in $\mathbb{R}_+\times\Omega$, related to
initial data $\bigl(\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}\bigr)_\varepsilon$, and assume that the initial energy $E_\varepsilon[\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}]<+\infty$.
Then one has the following bounds, uniformly for $\varepsilon>0$:
$$
\begin{cases}
\nabla\sqrt{\rho_\varepsilon}\;\in\;L^\infty_{loc}\bigl(\mathbb{R}_+;L^2(\Omega)\bigr) \\[1ex]
\dfrac{1}{\varepsilon}\,\nabla^2\rho_\varepsilon\;,\quad\dfrac{1}{\varepsilon}\,\nabla\!\left(\rho_\varepsilon^{\gamma/2}\right)\;,\quad
\dfrac{1}{\varepsilon}\,\nabla\!\left(\dfrac{1}{\rho_\varepsilon}\right)\qquad \in\;L^2_{loc}\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)\,.
\end{cases}
$$
In particular, the family $\bigl(\varepsilon^{-1}\,(\rho_\varepsilon-1)\bigr)_\varepsilon$ is bounded in $L^p_{loc}\bigl(\mathbb{R}_+;L^\infty(\Omega)\bigr)$
for any $2\leq p<4$.
\end{coroll}
\subsubsection{Additional bounds} \label{sss:bounds}
Let us continue and deduce further uniform bounds on a family $\bigl(\rho_\varepsilon,u_\varepsilon\bigr)_\varepsilon$ of solutions to system \eqref{eq:NSK-sing} in $\mathbb{R}_+\times\Omega$, related to
initial data $\bigl(\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}\bigr)_\varepsilon$ such that $E_\varepsilon[\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}]<+\infty$.
First of all, working a little bit one can establish, uniformly in $\varepsilon$:
\begin{eqnarray*}
\bigl(u_\varepsilon\bigr)_\varepsilon & \subset & L^\infty_T\bigl(L^2\bigr)+L^2_T\bigl(L^{3/2}\bigr)\,\hookrightarrow\,
L^2_T\bigl(L^{3/2}_{loc}\bigr) \\
\bigl(Du_\varepsilon\bigr)_\varepsilon & \subset & \Bigl(L^2_T\bigl(L^2\bigr)
+L^1_T\bigl(L^{3/2}\bigr)\Bigr)\,\cap\,\Bigl(L^2_T\bigl(L^2+L^1\bigr)\Bigr)\,.
\end{eqnarray*}
In particular, $\bigl(Du_\varepsilon\bigr)_\varepsilon$ is uniformly bounded in $L^1_T\bigl(L^{3/2}_{loc}\bigr)$; therefore,
by Sobolev embeddings we gather also the additional continuous inclusion $\bigl(u_\varepsilon\bigr)_\varepsilon\,\subset\,L^1_T\bigl(L^3_{loc}\bigr)$.
Furthermore, we also infer the uniform bounds
\begin{eqnarray}
\bigl(\rho_\varepsilon\,u_\varepsilon\bigr)_\varepsilon & \subset & L^\infty_T\bigl(L^2+L^{3/2}\bigr)\,\cap\,
\Bigl(L^\infty_T\bigl(L^2\bigr)+L^2_T\bigl(L^2\bigr)\Bigr) \nonumber \\%\label{sing-b:rho-u} \\
\bigl(D(\rho_\varepsilon\,u_\varepsilon)\bigr)_\varepsilon & \subset & L^2_T\bigl(L^2+L^{3/2}\bigr)\,+\,L^\infty_T\bigl(L^1\bigr)\;\hookrightarrow\;
L^2_T\bigl(L^1_{loc}\bigr)\,. \label{sing-b:D_rho-u}
\end{eqnarray}
In particular, we deduce that $\bigl(\rho_\varepsilon\,u_\varepsilon\bigr)_\varepsilon$ is a bounded family in
$L^\infty_T\bigl(L^{3/2}_{loc}\bigr)\,\cap\,L^2_T\bigl(L^2\bigr)$.
For the sake of completeness let us also establish uniform bounds on quantities related to $\rho^{3/2}_\varepsilon\,u_\varepsilon$.
First of all, we get
$$
\bigl(\rho^{3/2}_\varepsilon\,u_\varepsilon\bigr)_\varepsilon\;\subset\;L^\infty_T\bigl(L^2+L^{3/2}\bigr)\,\cap\,
\Bigl(L^\infty_T\bigl(L^2\bigr)+L^2_T\bigl(L^2\bigr)\Bigr)\,;
$$
on the other hand, we have also the uniform embedding
\begin{equation} \label{sing-b:D(rho-u)}
\Bigl(D\bigl(\rho^{3/2}_\varepsilon\,u_\varepsilon\bigr)\Bigr)_\varepsilon\;\subset\;L^2_T\bigl(L^2(\Omega)+L^{3/2}(\Omega)\bigr)\,.
\end{equation}
Therefore (see Theorem 2.40 of \cite{B-C-D}), we infer that
$\bigl(\rho^{3/2}_\varepsilon\,u_\varepsilon\bigr)_\varepsilon$ is uniformly bounded in $L^2_T\bigl(L^3(\Omega)\bigr)$.
Finally, from this fact combined with the usual decomposition $\sqrt{\rho_\varepsilon}\,=\,1+(\sqrt{\rho_\varepsilon}-1)$
and Sobolev embeddings, it follows also that
$$
\bigl(\rho_\varepsilon^2\,u_\varepsilon\bigr)_\varepsilon\;\subset\;L^2_T\bigl(L^2(\Omega)\bigr)\,.
$$
\subsection{Existence of weak solutions} \label{ss:weak}
At this point, let us spend a few words about the global in time existence of weak solutions to our system.
First of all, let us recall the definition. The integrability properties are justified by previous energy estimates.
\begin{defin} \label{d:weak}
Fix initial data $(\rho_0,u_0)$ satisfying the conditions in \eqref{eq:initial}, with $\rho_0\geq0$.
We say that $\bigl(\rho,u\bigr)$ is a \emph{weak solution} to system \eqref{eq:NSK-sing}-\eqref{eq:bc}
in $[0,T[\,\times\Omega$ (for some $T>0$) with initial datum $(\rho_0,u_0)$ if the following conditions are verified:
\begin{itemize}
\item[(i)] $\rho\geq0$ almost everywhere, and one has that
$\varepsilon^{-1}\bigl(\rho-1\bigr)\,\in\,L^\infty\bigl([0,T[\,;L^\gamma(\Omega)\bigr)$,
$\varepsilon^{-1}\bigl(1/\rho-1\bigr)\,\in\,L^\infty\bigl([0,T[\,;L^{\gamma_c}(\Omega)\bigr)$, $\varepsilon^{-1}\nabla\rho$
and $\nabla\sqrt{\rho}\;\in L^\infty\bigl([0,T[\,;L^2(\Omega)\bigr)$ and
$\varepsilon^{-1}\nabla^2\rho\in L^2\bigl([0,T[\,;L^2(\Omega)\bigr)$;
\item[(ii)] $\sqrt{\rho}\,u\,\in L^\infty\bigl([0,T[\,;L^2(\Omega)\bigr)$ and $\sqrt{\rho}\,Du\,\in L^2\bigl([0,T[\,;L^2(\Omega)\bigr)$;
\item[(iii)] the mass and momentum equations are satisfied in the weak sense: for any scalar function
$\phi\in\mathcal{D}\bigl([0,T[\,\times\Omega\bigr)$ one has the equality
$
-\int^T_0\int_\Omega\biggl(\rho\,\partial_t\phi\,+\,\rho\,u\,\cdot\,\nabla\phi\biggr)\,dx\,dt\,=\,\int_\Omega\rho_0\,\phi(0)\,dx\,,
$
and for any vector-field $\psi\in\mathcal{D}\bigl([0,T[\times\Omega;\mathbb{R}^3\bigr)$ one has
\begin{eqnarray}
& & \hspace{-0.3cm}
\int_\Omega\rho_0u_0\cdot\psi(0)dx\,=\,\int^T_0\!\!\int_\Omega\biggl(-\rho u\cdot\partial_t\psi\,-\,\rho u\otimes u:\nabla\psi\,-\,
\frac{1}{\varepsilon^2}\Pi(\rho){\rm div}\,\psi\,+ \label{eq:weak-momentum} \\
& & \qquad +\,\nu \rho Du:\nabla\psi\,+\,\frac{1}{\varepsilon^2}\rho\Delta\rho{\rm div}\,\psi\,+\,
\frac{1}{\varepsilon^2}\Delta\rho\nabla\rho\cdot\psi\,+\,\frac{\mathfrak{c}(x^h)}{\varepsilon}e^3\times\rho u\cdot\psi\biggr)\,dx\,dt\,; \nonumber
\end{eqnarray}
\item[(iv)] for almost every $t\in\,]0,T[\,$, the following energy inequalities hold true:
\begin{eqnarray}
& & \hspace{-0.5cm} E_\varepsilon[\rho,u](t)\,+\,\nu\int^t_0\int_\Omega\rho\,\left|Du\right|^2\,dx\,d\tau\;\leq\;E_\varepsilon[\rho_0,u_0] \label{en_est:E} \\
& & \hspace{-0.5cm} F_\varepsilon[\rho](t)\,+\,\dfrac{\nu}{\varepsilon^2}\int^t_0\int_\Omega \Pi'(\rho)\,|\nabla\sqrt{\rho}|^2\,dx\,d\tau\,+\,
\dfrac{\nu}{\varepsilon^{2}}\int^t_0\int_\Omega\left|\nabla^2\rho\right|^2\,dx\,d\tau\;\leq\;C\,(1+T)\,, \label{en_est:F}
\end{eqnarray}
for some constant $C$ depending just on $\bigl(E_\varepsilon[\rho_{0},u_0],F_\varepsilon[\rho_{0}],\nu\bigr)$.
\end{itemize}
\end{defin}
The last condition in the definition is required just in the study of the singular perturbation problem, because we have to assume \textsl{a priori} that
the family of weak solutions satisfies relevant uniform bonds.
The existence of weak solutions, in the sense of the previous definition, is guaranteed by the next result.
\begin{thm} \label{t:weak}
Let $\gamma_c=2$ in \eqref{eq:def_cold} and $\mathfrak c\,\in\,W^{1,\infty}(\mathbb{R}^2)$ in \eqref{eq:coriolis}.
For any fixed $\varepsilon>0$, consider a couple $(\rho_0,u_0)$ satisfying conditions \eqref{eq:initial}, with $\rho_0\geq0$.
Then, there exits a global in time weak solution $(\rho,u)$ to system \eqref{eq:NSK-sing},
related to the initial datum $(\rho_0,u_0)$.
\end{thm}
\begin{rem} \label{r:existence}
\begin{itemize}
\item The hypothesis $\gamma_c=2$ is assumed just for simplicity here, but it is not really necessary for existence (see also comments below).
\item The condition $\mathfrak{c}\,\in\,W^{1,\infty}$ is important in order to take advantage of the BD entropy structure of our system.
However, it can be deeply relaxed at this level, in presence of further assumptions (e.g. when friction terms are considered, as in paper \cite{B-D_2003}).
\end{itemize}
\end{rem}
The previous result can be established arguing exactly as in \cite{B-D-Zat}. Actually, the statement of \cite{B-D-Zat} holds true under more general assumptions
than ours (as for the cold component of the pressure and the viscosity coefficient, for instance). We also refer to \cite{Mu-Po-Zat}-\cite{Mu-Po-Zat_2015}.
So, we omit the proof. It is based on a quite standard approximation procedure and passage to the limit: we construct smooth solutions to a perturbed system,
which however preserves the mathematical structure of the original one. We can then deduce existence of a weak solution to our system \eqref{eq:NSK-sing} as a limit of the previous family of
smooth approximate solutions.
In particular, this construction allows to justify the integrability properties we require in Definition \ref{d:weak}, see points (i), (ii) and (iv):
as a matter of fact, uniform bounds which follow from energy estimates, established in Subsection \ref{ss:cap-effects}, will be inherited also by the weak solutions.
\subsection{Effects due to fast rotation} \label{ss:rot-effects}
In the previous parts, we have proved the existence of weak solutions for our system, and the relevant bounds they enjoy.
In this subsection we identify properties related to the effects of a strong Coriolis force.
Namely, we will establish the analogue of the \emph{Taylor-Proudman theorem} in our context. On the other hand, as already pointed out, we ignore boundary layer effects.
Propagation of waves will be treated instead in Section \ref{s:proof}.
For doing this, let us fix a family $\bigl(\rho_\varepsilon,u_\varepsilon\bigr)_\varepsilon$ of weak solutions to system \eqref{eq:NSK-sing}-\eqref{eq:bc},
related to the initial data $\bigl(\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}\bigr)_\varepsilon$ such that $E_\varepsilon[\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}]<+\infty$.
First of all, by uniform bounds we immediately deduce that $\rho_\varepsilon\,\rightarrow\,1$ (strong convergence)
in $L^\infty\bigl(\mathbb{R}_+;H^1(\Omega)\bigr)\,\cap\,L^2_{loc}\bigl(\mathbb{R}_+;H^2(\Omega)\bigr)$, with convergence rate $O(\varepsilon)$.
So, we can write $\rho_\varepsilon\,=\,1\,+\,\varepsilon\,r_\varepsilon$, with the family $\bigl(r_\varepsilon\bigr)_\varepsilon$ bounded in
the previous spaces. Then we infer that
\begin{equation} \label{eq:conv-r}
\qquad
r_\varepsilon\,\rightharpoonup\,r\qquad\qquad\qquad\mbox{ in }\quad
L^\infty\bigl(\mathbb{R}_+;H^1(\Omega)\bigr)\,\cap\,L^2_{loc}\bigl(\mathbb{R}_+;H^2(\Omega)\bigr)\,.
\end{equation}
In the same way, if we define $a_\varepsilon\,:=\,\bigl(1/\rho_\varepsilon-1\bigr)/\varepsilon$, we gather that $\bigl(a_\varepsilon\bigr)_\varepsilon$ is uniformly
bounded in $L^\infty\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)\,\cap\,L^2_{loc}\bigl(\mathbb{R}_+;H^1(\Omega)\bigr)$. So it weakly converges
to some $a$ in this space: more precisely,
\begin{equation} \label{eq:conv-a}
\qquad
a_\varepsilon\,\rightharpoonup\,-\,r\qquad\qquad\qquad\mbox{ in }\quad
L^\infty\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)\,\cap\,L^2_{loc}\bigl(\mathbb{R}_+;H^1(\Omega)\bigr)\,.
\end{equation}
Again by uniform bounds, we also deduce
\begin{equation}
\qquad\qquad\qquad
u_\varepsilon\,\rightharpoonup\,u\,\qquad\qquad\qquad\mbox{ in }\quad L^2_{loc}\bigl(\mathbb{R}_+;L^{3/2}_{loc}(\Omega)\bigr) \label{eq:conv-u}
\end{equation}
and $Du_\varepsilon\,\rightharpoonup\,Du$ in $L^2_{loc}\bigl(\mathbb{R}_+;L^1_{loc}(\Omega)\bigr)$, where we have identified $(L^1)^*$ with
$L^\infty$.
Notice also that, by uniqueness of the limit, we have the additional properties
\begin{eqnarray*}
\sqrt{\rho_\varepsilon}\,u_\varepsilon\,\stackrel{*}{\rightharpoonup}\,u & \qquad\qquad\mbox{ in }\quad & L^\infty\bigl(\mathbb{R}_+;L^2(\Omega)\bigr) \\[1ex]
\rho_\varepsilon\,u_\varepsilon\,\rightharpoonup\,u & \qquad\qquad\mbox{ in }\quad & L^2_{loc}\bigl(\mathbb{R}_+;L^2(\Omega)\bigr) \\[1ex]
\sqrt{\rho_\varepsilon}\,Du_\varepsilon\,\rightharpoonup\,Du & \qquad\qquad\mbox{ in }\quad & L^2\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)\,,
\end{eqnarray*}
where $\stackrel{*}{\rightharpoonup}$ denotes the weak-$*$ convergence in $L^\infty\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)$.
Finally, from the analysis of Paragraphs \ref{sss:BD} and \ref{sss:bounds}, we deduce some constraints the limit points $(r,u)$ have to satisfy.
This is exactly the analogue of the Taylor-Proudman theorem in our context.
\begin{prop} \label{p:TP}
Let $\bigl(\rho_\varepsilon,u_\varepsilon\bigr)_\varepsilon$ be a family of weak solutions to system \eqref{eq:NSK-sing}-\eqref{eq:bc},
related to the initial data $\bigl(\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}\bigr)_\varepsilon$ such that $E_\varepsilon[\rho_{0,\varepsilon}\,,\,u_{0,\varepsilon}]<+\infty$.
Let us define $r_\varepsilon:=\varepsilon^{-1}\left(\rho_\varepsilon-1\right)$, and let $(r,u)$ be a limit point of the sequence
$\bigl(r_\varepsilon,u_\varepsilon\bigr)_\varepsilon$.
Then $r=r(x^h)$ and $u=\bigl(u^h(x^h),0\bigr)$, with ${\rm div}\,_{\!h}u^h=0$. Moreover, we have
$$
\mathfrak{c}\,u^h\;=\,\nabla_h^\perp\bigl({\rm Id}\,\,-\,\Delta_h\bigr)r\qquad\qquad\mbox{ and }\qquad\qquad
u^h\,\cdot\,\nabla_h\mathfrak{c}\,\equiv\,0\,.
$$
\end{prop}
\begin{rem} \label{r:limit-dens}
Notice that, by the previous proposition and the fact that $r$ and $u$ belongs to $L^\infty\bigl(\mathbb{R}_+;L^2(\Omega)\bigr)$, we actually get
$r\in L^\infty\bigl(\mathbb{R}_+;H^3(\Omega)\bigr)$.
\end{rem}
The proof of the previous proposition relies in testing the mass and momentum equations respectively on smooth functions $\varphi$ and $\varepsilon\psi$.
From the equations for the velocity fields we then obtain the (quasi-)geostrophic balance relation.
Departures from geostrophy, which are determined by components of the solutions which do not respect the conditions found in the previous proposition, arise as superposition of waves.
By Proposition \ref{p:TP}, the wave propagator, i.e. the singular perturbation operator, can be defined as
\begin{equation} \label{eq:sing-op}
\begin{array}{lccc}
\mathcal{A}\,: & L^2(\Omega)\;\times\;L^2(\Omega) & \longrightarrow & H^{-1}(\Omega)\;\times\;H^{-3}(\Omega) \\
& \bigl(\,r\;,\;V\,\bigr) & \mapsto & \Bigl({\rm div}\, V\;,\;\mathfrak{c}(x^h)\,e^3\times V\,+\nabla\bigl({\rm Id}\,-\Delta\bigr)r\Bigr)\,.
\end{array}
\end{equation}
Remark that $\mathcal A$
has variable coefficients whenever $\mathfrak c$ is a non-constant function.
\section{Asymptotic limits for capillary fluids in fast rotation} \label{s:asympt}
In the present section we
perform the incompressible and fast rotation asymptotics simultaneously, while we keep the capillarity coefficient constant in
order to capture surface tension effects.
After spending a few words on the case of constant axis, we then specialize on the case of variable rotation axis.
We show the convergence of our system to a linear parabolic-type equation with variable coefficients.
Besides, here we look for minimal regularity assumptions on the variations of the axis, and we consider conditions of Zygmund-type for the function $\mathfrak{c}$.
A more complete analysis and further results can be found in \cite{F_2015}-\cite{F_2016}.
\subsection{An overview of the constant rotation axis case} \label{ss:constant}
Let us consider the case $\mathfrak c\equiv1$ first. This situation has been studied in \cite{F_2015}.
The analysis developed in Section \ref{s:math_prop} is almost enough to pass to the limit in the weak formulation of the equations,
when testing them on functions $(\varphi,\psi)$ in the kernel of the singular perturbation operator $\mathcal A$, defined in \eqref{eq:sing-op}.
The difficulty relies in passing to the limit in the convective and capillarity terms
$$
\rho_\varepsilon\,u_\varepsilon\otimes u_\varepsilon\qquad\qquad\mbox{ and }\qquad\qquad \frac{1}{\varepsilon^2}\,\Delta\rho\,\nabla\rho\,.
$$
As usual in singular perturbation problems, writing the equations (roughly speaking) in the form of a wave system
$$
\varepsilon\,\partial_tU_\varepsilon\,+\,\mathcal{A}\,U_\varepsilon\,=\,\varepsilon\,F_\varepsilon\,,
$$
one expects that components of the solutions in the kernel of $\mathcal A$, say $\mathbb{P}U_\varepsilon$, strongly converge (in suitable spaces) to a solution of the target system.
On the other hand, according to the previous equation, the projections onto the orthogonal complement of ${\rm Ker}\,\mathcal A$, say $\mathbb{Q}U_\varepsilon$, produce fast oscillations,
which therefore should weakly converge to $0$. In order to treat non-linearities, then, the first approach is to look for strong convergence properties, and especially in this case for
dispersion of the components in $\bigl({\rm Ker}\,\mathcal A\bigr)^\perp$.
The fact that $\mathcal A$ has constant coefficients for $\mathfrak c\equiv1$ simplifies the study of dispersive properties, since spectral analysis tools are available.
Direct computations show that the point spectrum of $\mathcal A$ reduces to the only eigenvalue $0$: therefore, one wants to use the celebrated RAGE theorem to prove
dispersion of $\mathbb{Q}U_\varepsilon$ in suitable norms.
Nevertheless, RAGE theorem is not directly applicable, since $\mathcal A$ is not skew-adjoint with respect to the classical $L^2$ scalar product. Then, after having localized in frequencies,
the idea is to resort to microlocal symmetrization arguments, in order to define a scalar product with respect to which $\mathcal A$ is skew-adjoint, and then RAGE theorem can be applied.
We remark that the symmetrizer involves a loss of derivatives for the density component, but this loss is safely handled because of the frequency localization
(notice however that one disposes of additional regularity for the density, provided by BD entropy estimates).
Finally, without entering into the details (for instance, precise assumptions on the initial data will be made in Subsection \ref{ss:var-result} below),
one can prove a result of the following type.
\begin{thm} \label{th:sing}
Let $\bigl(\rho_\varepsilon\,,\,u_\varepsilon\bigr)_\varepsilon$ be a family of weak solutions (in the sense of Definition \ref{d:weak} above) to system
\eqref{eq:NSK-sing}-\eqref{eq:bc} in $[0,T]\times\Omega$, related to suitable initial data
$\bigl(\rho_{0,\varepsilon},u_{0,\varepsilon}\bigr)_\varepsilon$. Suppose that the symmetriy properties of Remark \ref{r:period-bc} are verified.
Define $r_\varepsilon\,:=\,\varepsilon^{-1}\left(\rho_\varepsilon-1\right)$.
Then, up to the extraction of a subsequence, one has the convergence properties
\begin{itemize}
\item[(a)] ~$r_\varepsilon\,\rightharpoonup\,r$ in $L^\infty\bigl([0,T];H^1(\Omega)\bigr)\,\cap\,L^2\bigl([0,T];H^2(\Omega)\bigr)$;
\item[(b)] ~$\sqrt{\rho_\varepsilon}\,u_\varepsilon\,\rightharpoonup\,u$ in $L^\infty\bigl([0,T];L^2(\Omega)\bigr)$ and
$\sqrt{\rho_\varepsilon}\,Du_\varepsilon\,\rightharpoonup\,Du$ in $L^2\bigl([0,T];L^2(\Omega)\bigr)$;
\item[(c)] ~$r_\varepsilon\,\rightarrow\,r$ and $\rho_\varepsilon^{3/2}\,u_\varepsilon\,\rightarrow\,u$ (strong convergence) in $L^2\bigl([0,T];L^2_{loc}(\Omega)\bigr)$,
\end{itemize}
where $r$ and $u$ are linked by the relation found in Proposition \ref{p:TP} above. Moreover,
$r$ solves (in the weak sense) the modified Quasi-Geostrophic equation
\begin{equation} \label{eq:q-geo_0}
\partial_t\Bigl(\bigl({\rm Id}\,-\Delta_h+\Delta_h^2\bigr)r\Bigr)\,+\,\nabla^\perp_h\bigl({\rm Id}\,-\Delta_h\bigr)r\,\cdot\,\nabla_h\Delta_h^2r\,+\,
\frac{\nu}{2}\,\Delta_h^2\bigl({\rm Id}\,-\Delta_h\bigr)r\,=\,0
\end{equation}
supplemented with the initial condition $r_{|t=0}\,=\,\widetilde{r}_0$, where $\widetilde{r}_0\,\in\,H^3(\mathbb{R}^2)$ is the unique solution of
$$
\bigl({\rm Id}\,-\Delta_h+\Delta_h^2\bigr)\,\widetilde{r}_0\,=\,\int_0^1\bigl(\omega^3_0\,+\,r_0\bigr)\,dx^3\,,
$$
with $r_0$ and $u_0$ defined as the weak limits (up to extraction) of the initial data and $\omega_0\,=\,\nabla\times u_0$ the vorticity of $u_0$.
\end{thm}
\subsection{The result for variable rotation axis} \label{ss:var-result}
We get now interested in studying the incompressible and high rotation limit simultaneously, in the regime of constant capillarity and for effectively variable
rotation axis, i.e. when the function $\mathfrak{c}$ is non-constant.
Here we will consider the general instance of \emph{ill-prepared} initial data
$\bigl(\rho,u\bigr)_{|t=0}=\bigl(\rho_{0,\varepsilon},u_{0,\varepsilon}\bigr)$. Namely, we will suppose
the following assumptions:
\begin{itemize}
\item[(i)] ~$\rho_{0,\varepsilon}\,=\,1\,+\,\varepsilon\,r_{0,\varepsilon}$, with
$\bigl(r_{0,\varepsilon}\bigr)_\varepsilon\,\subset\,H^1(\Omega)\cap L^\infty(\Omega)$ bounded;
\item[(ii)] ~$1/\rho_{0,\varepsilon}\,=\,1\,+\,\varepsilon\,a_{0,\varepsilon}$, with
$\bigl(a_{0,\varepsilon}\bigr)_\varepsilon\,\subset\,L^2(\Omega)$ bounded;
\item[(iii)] ~~$\bigl(u_{0,\varepsilon}\bigr)_\varepsilon\,\subset\,L^2(\Omega)$ bounded.
\end{itemize}
Up to extraction of a subsequence, we can suppose to have the weak convergence properties
\begin{equation} \label{eq:conv-initial}
r_{0,\varepsilon}\,\rightharpoonup\,r_0\quad\mbox{ in }\;H^1(\Omega)\;,\qquad
a_{0,\varepsilon}\,\rightharpoonup\,a_0\,=\,-\,r_0\;\mbox{ and }\;
u_{0,\varepsilon}\,\rightharpoonup\,u_0\quad\mbox{ in }\;L^2(\Omega)\,.
\end{equation}
Remark that the previous assumptions respect what we required in Remark \ref{r:en_initial}, see Subsection \ref{ss:hypotheses}.
We also recall that it is at this point that we need condition (iv) of Definition \ref{d:weak}.
Let us turn our attention to the function $\mathfrak c$.
For technical reason, analogously to what done in \cite{G-SR_2006}, we need to assume that it has non-degenerate
critical points: namely, we will suppose
\begin{equation} \label{eq:non-crit}
\lim_{\delta\ra0}\;\mathcal{L}\left(\left\{x^h\,\in\,\mathbb{R}^2\;\Bigl|\;\bigl|\nabla_h\mathfrak{c}(x^h)\bigr|\,\leq\,\delta\right\}\right)\,=\,0\,,
\end{equation}
where we denoted by $\mathcal L(\mathcal O)$ the $2$-dimensional Lebesgue measure of a set $\mathcal O\,\subset\mathbb{R}^2$.
Also, fixed an admissible modulus of continuity $\mu$ (we will recall the precise definition in Paragraph \ref{sss:continuity}), we require that $\nabla_h\mathfrak{c}$ belongs to the space
\begin{equation} \label{def:Z}
\mathcal{Z}_\mu(\mathbb{R}^d)\,:=\,\biggl\{a\,\in\,L^\infty(\mathbb{R}^d;\mathbb{R})\;\biggl|\;\sup_{x\in\mathbb{R}^d}\left|a(x+y)+a(x-y)-2a(x)\right|\,\leq\,C\,\mu(|y|)\quad\forall\;|y|\leq1\biggr\}\,.
\end{equation}
Let us set $|a|_{\mathcal{Z}_\mu}$ as the smallest constant $C$ such that the previous inequality holds true, and
$\|a\|_{\mathcal{Z}_\mu}\,:=\,\|a\|_{L^\infty}+|a|_{\mathcal{Z}_\mu}$.
Finally,for notation convenience, we introduce the operator
$$
\mathfrak{D}_{\mathfrak{c}}(f)\,:=\,D_h\bigl(\mathfrak{c}^{-1}\,\nabla_h^\perp f\bigr)\,=\,\frac{1}{2}\,\left(\nabla_h\,+\,^t\nabla_h\right)\bigl(\mathfrak{c}^{-1}\,\nabla_h^\perp f\bigr)
$$
for any scalar function $f\,=\,f(x^h)$.
\begin{thm} \label{t:sing_var}
Let $1<\gamma\leq2$ in \eqref{eq:def_P} and $\mathfrak{C}(\rho,u)=\mathfrak{c}(x^h)\,e^3\times\rho\,u$,
where $\mathfrak{c}\in W^{1,\infty}(\mathbb{R}^2)$ is $\neq0$ almost everywhere and it verifies the non-degeneracy condition \eqref{eq:non-crit}.
Let us also assume that $\nabla_h\mathfrak{c}\,\in\,\mathcal{Z}_\mu$, for some admissible modulus of continuity $\mu$ which verifies the property
\begin{equation} \label{cond:mu}
\widetilde{\mu}(s)\,:=\,\mu(s)\,\log\!\left(1+\frac{1}{s}\right)\;\longrightarrow\;0\qquad\qquad\mbox{ for }\quad s\ra0\,.
\end{equation}
Let $\bigl(\rho_{0,\varepsilon},u_{0,\varepsilon}\bigr)_\varepsilon$ be initial data satisfying the hypotheses ${\rm (i)-(ii)-(iii)}$ and
\eqref{eq:conv-initial}, and let $\bigl(\rho_\varepsilon\,,\,u_\varepsilon\bigr)_\varepsilon$ be a family of corresponding weak solutions to system
\eqref{eq:NSK-sing}-\eqref{eq:bc} in $[0,T]\times\Omega$, in the sense of Definition \ref{d:weak}. Suppose that the symmetriy properties of Remark \ref{r:period-bc} are verified.
Define $r_\varepsilon\,:=\,\varepsilon^{-1}\left(\rho_\varepsilon-1\right)$.
Then, up to the extraction of a subsequence, one has the following convergence properties:
\begin{itemize}
\item[(a)] ~$r_\varepsilon\,\rightharpoonup\,r$ in $L^\infty\bigl([0,T];H^1(\Omega)\bigr)\,\cap\,L^2\bigl([0,T];H^2(\Omega)\bigr)$,
\item[(b)] ~$\sqrt{\rho_\varepsilon}\,u_\varepsilon\,\rightharpoonup\,u$ in $L^\infty\bigl([0,T];L^2(\Omega)\bigr)$ and
$\sqrt{\rho_\varepsilon}\,Du_\varepsilon\,\rightharpoonup\,Du$ in $L^2\bigl([0,T];L^2(\Omega)\bigr)$,
\end{itemize}
where $r$ and $u$ verify the relations established in Proposition \ref{p:TP}. Moreover,
$r$ solves (in the weak sense) the equation
\begin{equation} \label{eq:lim_var}
\partial_t\left(r\,-\,{\rm div}\,_{\!h}\!\left(\frac{1}{\mathfrak{c}^2}\,\nabla_h\bigl({\rm Id}\,-\Delta_h\bigr)r\right)\right)\,+\,
\nu\,\;^t\mathfrak{D}_{\mathfrak{c}}\,\circ\,\mathfrak{D}_{\mathfrak{c}}\bigl(({\rm Id}\,-\Delta_h)r\bigr)\,=\,0
\end{equation}
supplemented with the initial condition $r_{|t=0}\,=\,\widetilde{r}_1$, where $\widetilde{r}_1$ is defined by
$$
\widetilde{r}_1\,-\,{\rm div}\,_{\!h}\!\left(\frac{1}{\mathfrak{c}^2}\,\nabla_h\bigl({\rm Id}\,-\Delta_h\bigr)\widetilde{r}_1\right)\,=\,
\int_0^1\Bigl({\rm curl}_h\bigl(\mathfrak{c}^{-1}\,u^h_0\bigr)\,+\,r_0\Bigr)\,dx^3\,.
$$
\end{thm}
Notice that we have the identity
$$
\,^t\mathfrak{D}_{\mathfrak{c}}\,\circ\,\mathfrak{D}_{\mathfrak{c}}(f)\,=\,
\nabla_h^\perp\cdot\left(\frac{1}{\mathfrak{c}}\,\nabla_h\cdot\mathfrak{D}_{\mathfrak{c}}(f)\right)\,,
$$
where we used the notations ${\rm div}\, f$ and $\nabla\cdot f$ in an equivalent way.
We also remark here that, for $\mathfrak{c}\equiv1$, this operator reduces to $(1/2)\Delta_h^2f$, according to Theorem \ref{th:sing}.
\begin{rem} \label{r:var_axis}
The fact that the limit equation is linear is remarkable.
As already pointed out in \cite{G-SR_2006} and \cite{F-G-GV-N}, this corresponds to a sort of turbulent behaviour of the fluid, where all the scales are mixed and
one can identify just an average horizontal motion.
From the technical viewpoint, the motivation is that, in the case of variable rotation axis, the limit motion is much more constrained than for constant axis; correspondingly,
the kernel of the singular perturbation operator is smaller.
\end{rem}
\section{Proof of the main result} \label{s:proof}
In the present section we prove our main result, namely Theorem \ref{t:sing_var}.
It relies on compensated compactness arguments, firstly introduced by Lions and Masmoudi \cite{L-M}-\cite{L-M_CRAS} in the context of incompressible limit,
and later adapted by Gallagher and Saint-Raymond \cite{G-SR_2006} to the case of rotating fluids (see also \cite{F-G-GV-N}).
First of all, we give some insights on Zygmund conditions and the regularity hypothesis on the rotation function $\mathfrak{c}$. Then, we present how passing to the limit in the weak formulation
of the equations, and we finally derive the limit system.
\subsection{Moduli of continuity and Zygmund conditions} \label{ss:moduli}
The present subsection is devoted to present in detail the regularity class which the rotation coefficient belongs to.
First of all, we recall some basic notions and properties related to admissible moduli of continuity. Then, we switch to the analysis of Zygmund type conditions;
we conclude with a fundamental lemma, which allows us to prove Theorem \ref{t:sing_var}.
We are going to make a broad use of tools from Fourier Analysis, and especially Littlewood-Paley theory.
For the sake of conciseness, we do not present the details here, and we refer e.g. to Chapter 2 of \cite{B-C-D}.
Furthermore, for simplicity of exposition we will deal with the $\mathbb{R}^d$ case; however, everything can be adapted to the $d$-dimensional
torus $\mathbb{T}^d$, and then also to the case of $\mathbb{R}^{d_1}\times\mathbb{T}^{d_2}$.
\subsubsection{Admissible moduli of continuity} \label{sss:continuity}
In this paragraph we recall some fundamental definitions and properties about general moduli of continuity.
First of all, let us introduce the \emph{Littlewood-Paley decomposition}, based on a non-homogeneous dyadic partition of unity in the
Phase Space.
We fix a smooth radial function $\chi$ supported in the ball $B(0,2)$, equal to $1$ in a neighborhood of $B(0,1)$
and such that $r\mapsto\chi(r\,e)$ is nonincreasing over $\mathbb{R}_+$ for all unitary vectors $e\in\mathbb{R}^d$. Set
$\varphi\left(\xi\right)=\chi\left(\xi\right)-\chi\left(2\xi\right)$ and $\varphi_j(\xi):=\varphi(2^{-j}\xi)$ for all $j\geq0$.
The dyadic blocks $(\Delta_j)_{j\in\mathbb{Z}}$ are defined by\footnote{Throughout we agree that $f(D)$ stands for
the pseudo-differential operator $u\mapsto\mathcal{F}^{-1}(f\,\mathcal{F}u)$.}
$$
\Delta_j:=0\ \hbox{ if }\ j\leq-2,\quad\Delta_{-1}:=\chi(D)\qquad\mbox{ and }\qquad
\Delta_j:=\varphi(2^{-j}D)\; \mbox{ if }\; j\geq0\,.
$$
We also introduce the low frequency cut-off operators: for any $j\geq0$,
\begin{equation} \label{eq:low-freq}
S_ju\;:=\;\chi\bigl(2^{-j}D\bigr)u\;=\;\sum_{k\leq j-1}\Delta_{k}u\,.
\end{equation}
Let us now present a basic definition.
\begin{defin} \label{d:mod-cont}
A \emph{modulus of continuity} is a continuous non-decreasing function $\mu:[0,1]\,\longrightarrow\,\mathbb{R}_+$ such that $\mu(0)=0$.
It is said to be \emph{admissible} if the function $\Gamma_\mu$, defined by the relation
$$
\Gamma_\mu(s)\,:=\,s\,\mu(1/s)\,,
$$
is non-decreasing on $[1,+\infty[\,$ and it verifies, for some constant $C>0$ and any $s\geq1$,
$$
\int_s^{+\infty}\sigma^{-2}\,\Gamma_\mu(\sigma)\,d\sigma\,\leq\,C\,s^{-1}\,\Gamma_\mu(s)\,.
$$
\end{defin}
Given a modulus of continuity $\mu$, we can define the space $\mathcal{C}_\mu(\mathbb{R}^d)$ as the set of real-valued functions $a\in L^\infty(\mathbb{R}^d)$
such that
$$
|a|_{\mathcal{C}_\mu}\,:=\,\sup_{|y|\in\,]0,1]}\frac{\left|a(x+y)\,-\,a(x)\right|}{\mu(|y|)}\,<\,+\infty\,.
$$
We also define $\|a\|_{\mathcal{C}_\mu}\,:=\,\|a\|_{L^\infty}\,+\,|a|_{\mathcal{C}_\mu}$.
On the other hand, for an increasing $\Gamma$ on $[1,+\infty[\,$, we define the space $B_\Gamma(\mathbb{R}^d)$ as the set
of real-valued functions $a\in L^\infty(\mathbb{R}^d)$ such that
$$
|a|_{B_\Gamma}\,:=\,\sup_{j\geq0}\frac{\left\|\nabla S_ja\right\|_{L^\infty}}{\Gamma(2^j)}\,<\,+\infty\,,
$$
where $S_j$ is the low-frequency cut-off operator of a Littlewood-Paley decomposition, as introduced above.
We also set $\|a\|_{B_\Gamma}\,:=\,\|a\|_{L^\infty}\,+\,|a|_{B_\Gamma}$.
One has the following result (see Proposition 2.111 of \cite{B-C-D}).
\begin{prop} \label{p:cont-equiv}
Let $\mu$ be an admissible modulus of continuity. Then $\mathcal{C}_\mu(\mathbb{R}^d)\,=\,B_{\Gamma_\mu}(\mathbb{R}^d)$,
and the respective norms are equivalent. Moreover, for any $a\in\mathcal{C}_\mu(\mathbb{R}^d)$ one has
$$
\left\|\Delta_ja\right\|_{L^\infty}\,\leq\,C\,\mu(2^{-j})
$$
for all $j\geq-1$, where the constant $C$ just depend on $\|a\|_{\mathcal{C}_\mu}$.
\end{prop}
Now we want to present a commutator lemma, which is fundamental in the proof of our main result.
Let us recall first the classical result (see Lemma 2.97 of \cite{B-C-D}).
\begin{lemma} \label{l:commutator}
Let $\theta\in\mathcal{C}^1(\mathbb{R}^d)$ such that $\bigl(1+|\,\cdot\,|\bigr)\widehat{\theta}\,\in\,L^1$. There exists a constant $C$ such that,
for any Lipschitz function $\ell\in W^{1,\infty}(\mathbb{R}^d)$ and any $f\in L^p(\mathbb{R}^d)$ and for all $\lambda>0$, one has
$$
\left\|\bigl[\theta(\lambda^{-1}D),\ell\bigr]f\right\|_{L^p}\,\leq\,C\,\lambda^{-1}\,\left\|\nabla\ell\right\|_{L^\infty}\,\|f\|_{L^p}\,.
$$
\end{lemma}
Going along the lines of the proof, it is easy to see that the constant $C$ depends just on the $L^1$ norm
of the function $|x|\,k(x)$, where $k\,=\,\mathcal{F}_\xi^{-1}\theta$ denotes the inverse Fourier transform of $\theta$.
Easy modifications of the proof of Lemma \ref{l:commutator} give a variation of the previous lemma. For simplicity, we restrict our attention to the case of $\theta$ in the Schwartz
class $\mathcal{S}(\mathbb{R}^d)$.
\begin{lemma} \label{l:comm_L^p}
Let $\theta\in\mathcal{S}(\mathbb{R}^d)$ and $(p_1,p_2,q)\in[1,+\infty]^3$ such that $1/q\,=\,1+1/p_2-1/p_1$.
Then there exists a constant $C$ such that,
for any $f\in L^{p_1}(\mathbb{R}^d)$, any $\ell\in W^{1,\infty}(\mathbb{R}^d)$ and all $\lambda>0$,
$$
\left\|\bigl[\theta(\lambda^{-1}D),\ell\bigr]f\right\|_{L^{p_2}}\,\leq\,C\,
\lambda^{-1}\,\left\|\nabla\ell\right\|_{L^\infty}\,\|f\|_{L^{p_1}}\,.
$$
The constant $C$ just depends on the $L^q$ norm of the function $|x|\,k(x)$, where $k\,=\,\mathcal{F}_\xi^{-1}\theta$ as above.
\end{lemma}
Let us consider now less regular functions $\ell$. The next result is proved in \cite{F_2016}.
\begin{lemma} \label{l:comm-less}
Let $\theta\in\mathcal{C}^1(\mathbb{R}^d)$ be as in Lemma \ref{l:commutator}, and let $\mu$ be an admissible modulus of continuity. Then, there exists
a constant $C$ such that, for any function $\ell\in\mathcal{C}_{\mu}(\mathbb{R}^d)$ and any $f\in L^p(\mathbb{R}^d)$ and for all $\lambda>1$, one has
$$
\left\|\bigl[\theta(\lambda^{-1}D),\ell\bigr]f\right\|_{L^p}\,\leq\,C\,\mu(\lambda^{-1})\,\left|\ell\right|_{\mathcal{C}_\mu}\,\|f\|_{L^p}\,.
$$
The constant $C$ only depends on the $L^1$ norms of the functions $k(x)$ and $|x|\,k(x)$.
\end{lemma}
Obviously, an extension of the previous result, in the same spirit of Lemma \ref{l:comm_L^p}, holds true.
\subsubsection{Zygmund-type regularity conditions} \label{sss:Z_mu}
Let us now focus on second order conditions, namely Zygmund conditions. Given $\mu$ an admissible modulus of continuity,
we have defined the space $\mathcal{Z}_\mu$ in \eqref{def:Z}.
It is clear that $\mathcal{C}_\mu\,\subset\,\mathcal{Z}_\mu$.
Notice that, if $\mu(s)=s$ then one recovers the classical Zygmund space, while if $\mu(s)=s\,|\log s|$ then $\mathcal{Z}_\mu$
coincides with the space of log-Zygmund functions.
Zygmund and log-Zygmund conditions were introduced by Tarama \cite{Tar} in studying well-posedness of hyperbolic Cauchy problems with low regularity coefficients.
We refer to \cite{C-DS-F-M_Z-syst} and the references therein for further details and results in the same direction, and to \cite{F-Z}
for applications to control problems.
By use of Littlewood-Paley decomposition, we study here some properties of the space $\mathcal{Z}_\mu$.
The following analysis extends well-known facts about Zygmund and log-Zygmund classes (see e.g. Chapter 2 of \cite{B-C-D}, \cite{F-Z}).
First of all, we want to characterize $\mathcal{Z}_\mu$ as a special Besov-type class. Recalling Proposition \ref{p:cont-equiv},
this will imply (again) $\mathcal{C}_\mu\subset\mathcal{Z}_\mu$ in terms of dyadic blocks.
\begin{prop} \label{p:zyg}
The space $\mathcal{Z}_\mu$ coincides with the Besov-type class
$$
\mathcal{B}_\mu(\mathbb{R}^d)\,:=\,\left\{a\,\in\,L^\infty(\mathbb{R}^d;\mathbb{R})\;\Bigl|\;\left\|\Delta_ja\right\|_{L^\infty}\,\leq\,C\,\mu(2^{-j})
\quad\forall\;j\geq-1\right\}\,.
$$
Moreover, the $\mathcal{Z}_\mu$ and $\mathcal{B}_\mu$ norms are equivalent, where we have defined
$$
\left\|a\right\|_{\mathcal{B}_\mu}\,:=\,\sup_{j\geq-1}\!\left(\frac{1}{\mu(2^{-j})}\,\|\Delta_ja\|_{L^\infty}\right)\,.
$$
\end{prop}
\begin{proof}
First of all, let us take an $a\in\mathcal{Z}_\mu(\mathbb{R}^d)$. By Bernstein inequality, we immediately have
$\|\Delta_{-1}a\|_{L^\infty}\leq\|a\|_{L^\infty}$. Now let us denote by $h$ the inverse Fourier transform of $\varphi$:
since $\varphi$ is even and $\int h\,=\,\varphi(0)\,=\,0$, for any $j\geq0$ we can write
\begin{eqnarray*}
\Delta_ja(x) & = & 2^{jd}\int_{\mathbb{R}^d}h(2^jy)\,a(x-y)\,dy\;=\;2^{jd-1}\int_{\mathbb{R}^d}h(2^jy)\,\bigl(a(x+y)+a(x-y)\bigr)\,dy \\
& = & 2^{jd-1}\int_{\mathbb{R}^d}h(2^jy)\,\bigl(a(x+y)+a(x-y)-2a(x)\bigr)\,dy\,.
\end{eqnarray*}
From this we deduce that
$$
\left\|\Delta_ja\right\|_{L^\infty}\,\leq\,C\,2^{jd}\,\int_{\mathbb{R}^d}|h|(2^j\,y)\,\mu(|y|)\,dy\,.
$$
Let us split the previous integral according to the space decomposition $\mathbb{R}^d\,=\,\left\{|y|\leq2^{-j}\right\}\,\cup\,
\left\{|y|\geq2^{-j}\right\}$. For the former term, since $\mu$ is increasing we have
$$
2^{jd}\,\int_{|y|\leq 2^{-j}}|h|(2^j\,y)\,\mu(|y|)\,dy\,\leq\,\mu(2^{-j})\,\|h\|_{L^1}\,.
$$
For the latter term, instead, we make the non-decreasing function $\Gamma_\mu$ appear, and we estimate
\begin{eqnarray*}
2^{jd}\,\int_{|y|\geq2^{-j}}|h|(2^{j}\,y)\,\mu(|y|)\,dz & = &
2^{jd}\,\int_{|y|\geq2^{-j}}|h|(2^j\,y)\,\Gamma_\mu(|y|^{-1})\,|y|\,dy \\
& \leq & C\,\Gamma_\mu(2^j)\,2^{-j}\,\left\|\;|\cdot|\;h(\,\cdot\,)\right\|_{L^1}\;\leq\;C\,\mu(2^{-j})\,.
\end{eqnarray*}
We have thus proved that $\mathcal{Z}_\mu\,\subset\,\mathcal{B}_\mu$.
Let now fix $a\in\mathcal{B}_\mu(\mathbb{R}^d)$. Take $x\in\mathbb{R}^d$ and $|y|\leq1$: for any $n\in\mathbb{N}$ we have the decomposition
\begin{eqnarray*}
a(x+y)+a(x-y)-2a(x) & = & \sum_{m<n}\bigl(\Delta_m a(x+y)+\Delta_m a(x-y)-2\Delta_m a(x)\bigr)\,+ \\
& & \qquad+\,\sum_{m\geq n}\bigl(\Delta_m a(x+y)+\Delta_m a(x-y)-2\Delta_m a(x)\bigr)\,.
\end{eqnarray*}
By Taylor formula up to second order, the former sum can be estimated by the quantity
\begin{eqnarray*}
C\,|y|^2\sum_{m<n}\left\|\nabla^2\Delta_m a\right\|_{L^\infty} & \leq & C\,|y|^2\sum_{m<n}2^{2m}\,\mu(2^{-m}) \\
& \leq & C\,|y|^2\,\sum_{m<n}2^{m}\,\Gamma_\mu(2^{m})\;\leq\;C\,|y|^2\,\Gamma_\mu(2^n)\,2^n\,.
\end{eqnarray*}
For the latter, instead, we use directly the property of the dyadic blocks, finding
$$
\sum_{m\geq n}\bigl|\Delta_m a(x+y)+\Delta_m a(x-y)-2\Delta_m a(x)\bigr|\,\leq\,4\sum_{m\geq n}\left\|\Delta_ja\right\|_{L^\infty}
\,\leq\,C\sum_{m\geq n}\Gamma_\mu(2^m)\,2^{-m}\,.
$$
Since $\mu$ is admissible, we have the bound
$$
\sum_{m\geq n}\Gamma_\mu(2^m)\,2^{-m}\,\leq\,C\int_{2^n}^{+\infty}\tau^{-2}\,\Gamma_\mu(\tau)\,d\tau\,\leq\,C\,\Gamma_\mu(2^n)\,2^{-n}\,,
$$
and this finally implies
$$
\bigl|a(x+y)+a(x-y)-2a(x)\bigr|\,\leq\,C\,\mu(2^{-n})\left(|y|^2\,2^{2n}\,+\,1\right)\,.
$$
Now, the choice $|y|\,2^n\,\sim\,1$ completes the proof of the inclusion $\mathcal{B}_\mu\,\subset\,\mathcal{Z}_\mu$,
and then of the whole proposition.
\end{proof}
On the other hand, Zygmund conditions imply a control on the first variation of the function, for which one loses a logarithmic factor.
\begin{prop} \label{p:zyg-first}
For any $a\,\in\,\mathcal{Z}_\mu(\mathbb{R}^d)$, there exists $C_a>0$ such that, for all $|y|\leq1$,
$$
\sup_{x\in\mathbb{R}^d}\bigl|a(x+y)-a(x)\bigr|\,\leq\,C_a\,\mu(|y|)\,\log\!\left(1\,+\,\frac{1}{|y|}\right)\,.
$$
The constant $C_a$ just depends on $\|a\|_{\mathcal{Z}_\mu}$.
\end{prop}
\begin{proof}
Analogously to the previous proof, let us write
$$
a(x+y)-a(x)\,=\,\sum_{m<n}\bigl(\Delta_m a(x+y)-\Delta_m a(x)\bigr)\,+\,
\sum_{m\geq n}\bigl(\Delta_m a(x+y)-\Delta_m a(x)\bigr)\,.
$$
As done above, we can estimate then
\begin{eqnarray*}
\bigl|a(x+y)-a(x)\bigr| & \leq & |y|\sum_{m<n}\left\|\nabla\Delta_m a\right\|_{L^\infty}\,+\,
2\sum_{m\geq n}\left\|\Delta_m a\right\|_{L^\infty} \\
& \leq & C\,|y|\sum_{m<n}\Gamma_\mu(2^m)\,+\,C\sum_{m\geq n}\Gamma_\mu(2^m)\,2^{-m} \\
& \leq & C\,\Gamma_\mu(2^n)\left(|y|\,n\,+\,2^{-n}\right)\;=\;C\,\mu(2^{-n})\left(2^n\,n\,|y|\,+\,1\right)\,.
\end{eqnarray*}
Again, the choice $n\,\sim\,\log_2\bigl(1/|y|\bigr)$ completes the proof of the proposition.
\end{proof}
Finally, let us present the analogue of Lemma \ref{l:comm-less} for second order regularity hypotheses. This result will be fundamental in proving Theorem \ref{t:sing_var}.
\begin{lemma} \label{l:comm-zyg}
Let $\theta\in\mathcal{C}^1(\mathbb{R}^d)$ such that $\bigl(1+|\,\cdot\,|\bigr)\widehat{\theta}\,\in\,L^1$, and let $\mu$ be an admissible modulus of
continuity for which \eqref{cond:mu} holds true. Then, there exist a constant $C$ and a $\lambda_0>0$ such that, for any function
$a\in\mathcal{Z}_{\mu}(\mathbb{R}^d)$ and any $f\in L^p(\mathbb{R}^d)$ and for all $\lambda\geq\lambda_0$, one has
$$
\left\|\bigl[\theta(\lambda^{-1}D),a\bigr]f\right\|_{L^p}\,\leq\,
C\,\mu(\lambda^{-1})\,\log(1+\lambda)\,\left\|a\right\|_{\mathcal{Z}_\mu}\,\|f\|_{L^p}\,.
$$
The constant $C$ only depends on the $L^1$ norms of the functions $k:=\mathcal{F}_\xi^{-1}\theta$ and $|\cdot|\,k$, while $\lambda_0$
just depends on $\mu$.
\end{lemma}
\begin{proof}
We start by writing the identity
$$
\bigl[\theta(\lambda^{-1}D),a\bigr]f\,=\,\lambda^d\int_\mathbb{R}^dk\bigl(\lambda(x-y)\bigr)\,f(y)\,\bigl(a(x)\,-\,a(y)\bigr)\,dy\,.
$$
Therefore, by use of Proposition \ref{p:zyg-first} above and Young inequality, we are reconducted to bound
$$
\lambda^d\,\left\|k(\lambda\,\cdot\,)\,\widetilde{\mu}(\,|\cdot|\,)\right\|_{L^1}\,=\,
\lambda^d\,\int_{\mathbb{R}^d}|k|(\lambda\,z)\,\widetilde{\mu}(|z|)\,dz\,,
$$
where $\widetilde{\mu}$ has been defined in \eqref{cond:mu}. We split the integral in the regions
$\left\{|z|\leq\lambda^{-1}\right\}$ and $\left\{|z|\geq\lambda^{-1}\right\}$.
For the former term, if $\lambda$ is big enough, by condition \eqref{cond:mu} we have
$$
\lambda^d\,\int_{|z|\leq \lambda^{-1}}|k|(\lambda\,z)\,\widetilde{\mu}(|z|)\,dz\,\leq\,\widetilde{\mu}(\lambda^{-1})\,\|k\|_{L^1}\,\leq\,
C\,\mu(\lambda^{-1})\,\log(1+\lambda)\,.
$$
For the latter term, instead, we have the estimate
\begin{eqnarray*}
& & \lambda^d\,\int_{|z|\geq \lambda^{-1}}|k|(\lambda\,z)\,\widetilde{\mu}(|z|)\,dz\;=\;
\lambda^d\,\int_{|z|\geq \lambda^{-1}}|k|(\lambda\,z)\,\Gamma_\mu(|z|^{-1})\,|z|\,\log\left(1+\frac{1}{|z|}\right)\,dz \\
& & \qquad\qquad\qquad\qquad \leq C\,\Gamma_\mu(\lambda)\,\lambda^{-1}\,\log(1+\lambda)\,\left\|\;|\cdot|\,k\right\|_{L^1}\;\leq\;
C\,\mu(\lambda^{-1})\,\log(1+\lambda)\,.
\end{eqnarray*}
The lemma is hence proved.
\end{proof}
\subsection{Convergence by compensated compactness} \label{ss:proof}
In this section we start proving Theorem \ref{t:sing_var}. The argument is analogous to the one given in \cite{F_2016} for moduli of continuity. There is only one point where the Zygmund condition
comes into play, i.e. in Proposition \ref{p:regular} below, for which we have to use Lemma \ref{l:comm-zyg}.
Nonetheless, for reader's convenience, we will give here most of the details. Indeed, this allows us to describe propagation of waves.
As pointed out in Subsection \ref{ss:i_rot}, here we will study interactions of acoustic, Poincar\'e and Rossby waves: the first ones are due to the the compressible part
of $u$, the second ones to the vertical component of $u$ and the third ones to the variations of the axis.
\medbreak
By Proposition \ref{p:TP}, we have identified in \eqref{eq:sing-op} the singular perturbation operator
$\mathcal{A}$, which has variable coefficients. So, spectral analysis tools (employed in \cite{F_2015} for constant rotation axis) are out of use here. Hence,
in order to prove convergence in the weak formulation of our equations, we have to resort then to a compensated compactness argument.
Let us consider tests functions $\phi\,\in\,\mathcal{D}\bigl([0,T[\,\times\Omega\bigr)$ and
$\psi\,\in\,\mathcal{D}\bigl([0,T[\,\times\Omega;\mathbb{R}^3\bigr)$ such that the couple $(\phi,\psi)$ belongs to
${\rm Ker}\,{\mathcal A}$. Recall that, by Proposition \ref{p:TP}, they satisfy
$$
{\rm div}\,\psi\,=\,0\qquad\mbox{ and }\qquad \mathfrak{c}(x^h)\,e^3\times\psi\,+\,\nabla\bigl({\rm Id}\,-\Delta\bigr)\phi\,=\,0\,.
$$
In particular, $\psi=\bigl(\psi^h,0\bigr)$ and $\phi$ just depend on the horizontal variable $x^h\in\mathbb{R}^2$ and they are linked
by the relation $\mathfrak{c}\,\psi^h\,=\,\nabla^\perp_h\bigl({\rm Id}\,-\Delta_h\bigr)\phi$. Finally,
we infer also that $\nabla^\perp_h\bigl({\rm Id}\,-\Delta_h\bigr)\phi\cdot\nabla_h\mathfrak{c}\,=\,0$.
First of all, we evaluate the momentum equation on such a $\psi$: taking into account the previous properties, we end up with
\begin{eqnarray}
& & \hspace{-0.7cm} \int_\Omega\rho_{0,\varepsilon}\,u_{0,\varepsilon}\cdot\psi(0)\,dx\;=\;
\int^T_0\!\!\int_\Omega\biggl(-\rho_\varepsilon\,u_\varepsilon\cdot\partial_t\psi\,-\,\rho_\varepsilon\,u_\varepsilon\otimes u_\varepsilon:\nabla\psi\,+ \label{eq:sing_weak} \\
& & \qquad
+\,\nu\,\rho_\varepsilon\,Du_\varepsilon:\nabla\psi\,+\,\frac{1}{\varepsilon^2}\,\Delta\rho_\varepsilon\,\nabla\rho_\varepsilon\cdot\psi\,+\,
\frac{\mathfrak{c}(x^h)}{\varepsilon}\,e^3\times\rho_\varepsilon\,u_\varepsilon\cdot\psi\biggr)dx\,dt. \nonumber
\end{eqnarray}
The $\partial_t$ and viscosity terms do not present any difficulty in passing to the limit. On the other hand, the rotation
term can be handled by use of the weak form of the mass equation, tested on
$\widetilde{\phi}\,=\,\bigl({\rm Id}\,-\Delta_h\bigr)\phi$: we get
\begin{eqnarray*}
\frac{1}{\varepsilon}\int^T_0\!\!\int_\Omega\mathfrak{c}(x^h)\,e^3\times\rho_\varepsilon\,u_\varepsilon\cdot\psi & = &
-\,\frac{1}{\varepsilon}\int^T_0\!\!\int_\Omega\mathfrak{c}(x^h)\,\rho_\varepsilon\,u^h_\varepsilon\cdot\left(\psi^h\right)^\perp \\
& = &
-\int_\Omega r_{0,\varepsilon}\,\widetilde{\phi}(0)\,-\,\int^T_0\!\!\int_\Omega r_\varepsilon\,\partial_t\widetilde{\phi}\,,
\end{eqnarray*}
which obviously converges in the limit $\varepsilon\ra0$.
In order to deal with the transport and the capillarity terms, we want to use the structure of the system. Therefore, first of all we need
to introduce a regularization of our solutions.
\subsubsection{Regularization and description of the oscillations}
Let us set $V_\varepsilon\,:=\,\rho_\varepsilon\,u_\varepsilon$. We can write system \eqref{eq:NSK-sing} in the form
\begin{equation} \label{eq:ac-w_c}
\begin{cases}
\varepsilon\,\partial_tr_\varepsilon\,+\,{\rm div}\,\,V_\varepsilon\,=\,0 \\[1ex]
\varepsilon\,\partial_tV_\varepsilon\,+\,\Bigl(\mathfrak{c}(x^h)\,e^3\times V_\varepsilon\,+\,\nabla\bigl({\rm Id}\,\,-\,\Delta\bigr)r_\varepsilon\Bigr)\,=\,\varepsilon\,f_\varepsilon\,,
\end{cases}
\end{equation}
where we have defined $f_\varepsilon$ by the formula
\begin{eqnarray}
f_\varepsilon & := & -\,{\rm div}\,\left(\rho_\varepsilon u_\varepsilon\otimes u_\varepsilon\right)\,+\,\nu\,{\rm div}\,\left(\rho_\varepsilon Du_\varepsilon\right)\,-
\label{eq:f_veps} \\
& & \qquad -\,\frac{1}{\varepsilon^2}\nabla\Bigl(\Pi(\rho_\varepsilon)-\Pi(1)-\Pi'(1)\left(\rho_\varepsilon-1\right)\Bigr)\,+\,
\frac{1}{\varepsilon^2}\bigl(\rho_\varepsilon-1\bigr)\,\nabla\Delta\rho_\varepsilon\,. \nonumber
\end{eqnarray}
Equations \eqref{eq:ac-w_c} have to be read in the weak sense, of course. In particular, from writing
\begin{eqnarray*}
\langle f_\varepsilon,\psi\rangle & := & \int_\Omega\biggl(\rho_\varepsilon u_\varepsilon\otimes u_\varepsilon:\nabla\psi\,-\,
\nu\,\rho_\varepsilon Du_\varepsilon:\nabla\psi\,-\,\frac{1}{\varepsilon^2}\,\Delta\rho_\varepsilon\,\nabla\rho_\varepsilon\cdot\psi\,- \\
& & \hspace{-0.5cm} -\,\frac{1}{\varepsilon^2}\,(\rho_\varepsilon-1)\,\Delta\rho_\varepsilon\,{\rm div}\,\psi\,+\,
\frac{1}{\varepsilon^2}\Bigl(\Pi(\rho_\varepsilon)-\Pi(1)-\Pi'(1)\left(\rho_\varepsilon-1\right)\Bigr){\rm div}\,\psi\biggr)dx \\
& = & \int_\Omega\Bigl(f^1_\varepsilon:\nabla\psi\,+\,f^2_\varepsilon:\nabla\psi\,+\,f^3_\varepsilon\cdot\psi\,+\,f^4_\varepsilon\,{\rm div}\,\psi\,+\,
f^5_\varepsilon\,{\rm div}\,\psi\Bigr)\,dx
\end{eqnarray*}
and by a systematic use of uniform bounds,
we can easily see that $\bigl(f^1_\varepsilon\bigr)_\varepsilon$ and $\bigl(f^5_\varepsilon\bigr)_\varepsilon$ are uniformly bounded in $L^\infty_T\bigl(L^1\bigr)$,
and so is $\bigl(f^2_\varepsilon\bigr)_\varepsilon$ in $L^2_T\bigl(L^2\bigr)$; finally, $\bigl(f^3_\varepsilon\bigr)_\varepsilon$ and $\bigl(f^4_\varepsilon\bigr)_\varepsilon$
are bounded in $L^2_T\bigl(L^1\bigr)$.
Therefore, we deduce that the family $\bigl(f_\varepsilon\bigr)_\varepsilon$ is uniformly bounded in the space
$L^2_T\bigl(H^{-1}(\Omega)\,+\,W^{-1,1}(\Omega)\bigr)$, and then in particular in $L^2_T\bigl(H^{-s}(\Omega)\bigr)$ for any
$s>5/2$.
\medbreak
Now, for any $M>0$, let us consider the low-frequency cut-off operator $S_M$ of a Littlewood-Paley decomposition,
as introduced in \eqref{eq:low-freq} above, and let us define
$$
r_{\varepsilon,M}\,:=\,S_Mr_\varepsilon\qquad\qquad\mbox{ and }\qquad\qquad
V_{\varepsilon,M}\,:=\,S_MV_\varepsilon\,.
$$
The following result hods true.
\begin{prop} \label{p:regular}
For any fixed time $T>0$ and compact set $K\subset\Omega$, the following convergence properties hold, in the limit for $M\longrightarrow+\infty$:
\begin{equation} \label{reg:convergence}
\begin{cases}
\; \sup_{\varepsilon>0}\left\|r_\varepsilon\,-\,r_{\varepsilon,M}\right\|_{L^\infty_T(H^s(K))\,\cap\,L^2_T(H^{1+s}(K))}\,\longrightarrow\,0
\qquad\qquad\forall\; s<1 \\[1ex]
\; \sup_{\varepsilon>0}\left\|V_\varepsilon\,-\,V_{\varepsilon,M}\right\|_{L^2_T(H^{-s}(K))}\,\longrightarrow\,0
\qquad\qquad\forall\; s>0\,.
\end{cases}
\end{equation}
Moreover, for any $M>0$, the couple $\bigl(r_{\varepsilon,M}\,,\,V_{\varepsilon,M}\bigr)$ satisfies the approximate wave equations
\begin{equation} \label{reg:approx-w}
\begin{cases}
\varepsilon\,\partial_tr_{\varepsilon,M}\,+\,{\rm div}\,\,V_{\varepsilon,M}\,=\,0 \\[1ex]
\varepsilon\,\partial_tV_{\varepsilon,M}\,+\,\Bigl(\mathfrak{c}(x^h)\,e^3\times V_{\varepsilon,M}\,+\,\nabla\bigl({\rm Id}\,-\Delta\bigr)r_{\varepsilon,M}\Bigr)\,=\,
\varepsilon\,f_{\varepsilon,M}\,+\,g_{\varepsilon,M}\,,
\end{cases}
\end{equation}
where $\bigl(f_{\varepsilon,M}\bigr)_{\varepsilon}$ and $\bigl(g_{\varepsilon,M}\bigr)_{\varepsilon}$ are families of smooth functions satisfying
\begin{equation} \label{reg:source}
\begin{cases}
\; \sup_{\varepsilon>0}\left\|f_{\varepsilon,M}\right\|_{L^2_T(H^{s}(K))}\,\leq\,C(s,M)
\qquad\qquad\forall\; s\geq0 \\[1ex]
\; \sup_{\varepsilon>0}\left\|g_{\varepsilon,M}\right\|_{L^2_T(H^1(K))}\,\longrightarrow\,0
\qquad\qquad\mbox{ for }\quad M\rightarrow+\infty\,,
\end{cases}
\end{equation}
where the constant $C(s,M)$ depends on the fixed values of $s\geq0$, $M>0$.
\end{prop}
\begin{proof}
Keeping in mind the characterization of $H^s$ spaces in terms of Littlewood-Paley decomposition (see Chapter 2 of \cite{B-C-D}),
properties \eqref{reg:convergence} are straightforward consequences of the uniform bounds established in Paragraphs \ref{sss:BD} and \ref{sss:bounds}.
Next, applying operator $S_M$ to \eqref{eq:ac-w_c} immediately gives us system \eqref{reg:approx-w},
where, denoting by $[\mathcal{P},\mathcal{Q}]$ the commutator between two operators $\mathcal P$ and $\mathcal Q$, we have set
$$
f_{\varepsilon,M}\,:=\,S_Mf_\varepsilon\qquad\qquad\mbox{ and }\qquad\qquad
g_{\varepsilon,M}\,:=\,\bigl[\mathfrak{c}(x^h),S_M\bigr]\bigl(e^3\times V_\varepsilon\bigr)\,.
$$
By these definitions and the uniform bounds on $\bigl(f_\varepsilon\bigr)_\varepsilon$, it is easy to verify the first property in \eqref{reg:source}.
As for the second one, we need to proceed carefully.
First of all, by uniform bounds and Lemma \ref{l:commutator} we get
$$
\sup_{\varepsilon>0}\left\|g_{\varepsilon,M}\right\|_{L^2_T(L^2)}\,\leq\,C\,2^{-M}\,.
$$
As for the gradient, for any $1\leq j\leq 3$ we can write
$$
\partial_jg_{\varepsilon,M}\,=\,\left[\mathfrak{c}\,,\,S_M\right]\partial_j\left(e^3\times V_\varepsilon\right)\,+\,
\left[\partial_j\mathfrak{c}\,,\,S_M\right]\left(e^3\times V_\varepsilon\right)\,.
$$
In order to control the former term, we use Lemma \ref{l:comm_L^p} with $p_2=q=2$ and $p_1=1$. Recalling that, by \eqref{sing-b:D_rho-u},
$\left(DV_\varepsilon\right)_\varepsilon\subset L^2_T(L^1_{loc})$, for any compact $K\subset\Omega$ we get
$$
\sup_{\varepsilon>0}\left\|\left[\mathfrak{c}\,,\,S_M\right]\partial_j\left(e^3\times V_\varepsilon\right)\right\|_{L^2_T(L^2(K))}\,\leq\,C\,2^{-M}\,.
$$
For the latter term, instead, Lemma \ref{l:comm-zyg} gives us
$$
\sup_{\varepsilon>0}\left\|\left[\partial_j\mathfrak{c}\,,\,S_M\right]\left(e^3\times V_\varepsilon\right)\right\|_{L^2_T(L^2)}\,\leq\,C\,\mu(2^{-M})\,\log\bigl(1+2^{M}\bigr)\,.
$$
In the end, choosing $\eta(M)=\max\left\{2^{-M},\mu(2^{-M})\,\log\bigl(1+2^{M}\bigr)\right\}$ (which goes to $0$ when $M\rightarrow+\infty$), we get
$$
\sup_{\varepsilon>0}\left\|g_{\varepsilon,M}\right\|_{L^2_T(H^1_{loc})}\,\leq\,C\,\eta(M)
$$
for a suitable constant $C>0$, and this completes the proof of the proposition.
\end{proof}
We also have an important decomposition for the approximated velocity fields. We refer to \cite{F_2016} for the proof.
\begin{prop} \label{p:decomp}
The following decompositions hold true:
$$
V_{\varepsilon,M}\,=\,\mathbb{V}_{\varepsilon,M}\,+\,\varepsilon\,\mathcal{V}_{\varepsilon,M}\qquad\qquad\mbox{ and }\qquad\qquad
DV_{\varepsilon,M}\,=\,\mathbb{D}_{\varepsilon,M}\,+\,\varepsilon\,\mathcal{D}_{\varepsilon,M}\,,
$$
where, for any compact set $K\subset\Omega$ and any $s\geq0$ one has
$$
\begin{cases}
\left\|\mathbb{V}_{\varepsilon,M}\right\|_{L^2_T\bigl(L^2(K)\cap L^3(K)\bigr)}\,+\,
\left\|\mathbb{D}_{\varepsilon,M}\right\|_{L^2_T\bigl(L^2(K)\bigr)}\,\leq\,C(K) \\[1ex]
\left\|\mathcal{V}_{\varepsilon,M}\right\|_{L^2_T\bigl(H^s(K)\bigr)}\,+\,
\left\|\mathcal{D}_{\varepsilon,M}\right\|_{L^2_T\bigl(H^s(K)\bigr)}\,\leq\,C(K,s,M)\,,
\end{cases}
$$
for suitable positive constants $C(K)$, $C(K,s,M)$ depending just on the quantities in the brackets.
\end{prop}
Before proceeding, let us introduce some useful notations. More precisely, we recall the following decomposition: for a vector-field $X$, we write
\begin{equation} \label{dec:vert-av}
X(x)\,=\,\langle X\rangle(x^h)\,+\,\widetilde{X}(x)\,,\quad\qquad\mbox{ where }\quad
\langle X\rangle(x^h)\,:=\,\int_{\mathbb{T}}X(x^h,x^3)\,dx^3\,.
\end{equation}
Notice that $\widetilde{X}$ has zero vertical average, and therefore we can write $\widetilde{X}(x)\,=\,\partial_3\widetilde{Z}(x)$,
with $\widetilde{Z}$ having zero vertical average as well.
We also set $\widetilde{Z}\,=\,\mathcal{I}(\widetilde{X})\,=\,\partial_3^{-1}\widetilde{X}$.
\subsubsection{The capillarity term}
First of all, let us deal with the surface tension term in \eqref{eq:sing_weak}. Notice that it can be rewritten as
$\;\int^T_0\int_\Omega\Delta r_\varepsilon\,\nabla r_\varepsilon\cdot\psi\;$, for any smooth test function $\psi$.
Thanks to the next lemma, we reconduct ourselves to study the convergence in the case of regular density functions.
\begin{lemma} \label{l:capill_approx}
For any $\psi\,\in\,\mathcal{D}\bigl([0,T[\,\times\Omega;\mathbb{R}^3\bigr)$, we have
$$
\lim_{M\rightarrow+\infty}\,\limsup_{\varepsilon\ra0}\,\left|\int^T_0\!\!\int_\Omega\Delta r_\varepsilon\;\nabla r_\varepsilon\,\cdot\,\psi\,dx\,dt\,-\,
\int^T_0\!\!\int_\Omega \Delta r_{\varepsilon,M}\; \nabla r_{\varepsilon,M}\,\cdot\,\psi\,dx\,dt\right|\,=\,0\,.
$$
\end{lemma}
Then, for any $\psi\,\in\,\mathcal{D}\bigl([0,T[\,\times\Omega;\mathbb{R}^3\bigr)\,\cap\,{\rm Ker}\,{\mathcal{A}}$
we have to consider the convergence of the term (pay attention to the signs)
\begin{eqnarray*}
\hspace{-0.3cm} \int^T_0\!\!\int_\Omega\Delta r_{\varepsilon,M}\;\nabla r_{\varepsilon,M}\cdot\psi\,dx\,dt & = &
-\,\int^T_0\!\!\int_\Omega\bigl({\rm Id}\,-\Delta\bigr)r_{\varepsilon,M}\;\nabla r_{\varepsilon,M}\cdot\psi\,dx\,dt\,+ \\
& & \qquad\qquad\qquad +\,\int^T_0\!\!\int_\Omega r_{\varepsilon,M}\;\nabla r_{\varepsilon,M}\cdot\psi\,dx\,dt\,.
\end{eqnarray*}
Notice that $r_{\varepsilon,M}\,\nabla r_{\varepsilon,M}\,=\,\nabla\left(r_{\varepsilon,M}\right)^2/2$: therefore, since ${\rm div}\,\psi=0$,
by integration by parts we get that the latter item on the right-hand side is identically $0$.
Hence, in the end we have to deal only with the remainder
\begin{eqnarray}
-\,\int^T_0\!\!\int_\Omega\bigl({\rm Id}\,-\Delta\bigr)r_{\varepsilon,M}\;\nabla r_{\varepsilon,M}\cdot\psi & = &
-\,\int^T_0\!\!\int_\Omega\bigl({\rm Id}\,-\Delta_h\bigr)\langle r_{\varepsilon,M}\rangle\;\nabla_h\langle r_{\varepsilon,M}\rangle\cdot\psi\,- \label{comp-cpt:dens} \\
& & \qquad\qquad
-\,\int^T_0\!\!\int_\Omega\langle\bigl({\rm Id}\,-\Delta\bigr)\widetilde{r}_{\varepsilon,M}\;\nabla\widetilde{r}_{\varepsilon,M}\rangle\cdot\psi\,, \nonumber
\end{eqnarray}
where, using the notations of \eqref{dec:vert-av}, $\widetilde{r}_{\varepsilon,M}$ denotes the mean-free part of $r_{\varepsilon,M}$.
\subsubsection{The convective term}
We now deal with the convective term. Once again, the first step is to reduce
the study to the case of smooth vector fields $V_{\varepsilon,M}$.
\begin{lemma} \label{l:conv_approx}
For any $\psi\,\in\,\mathcal{D}\bigl([0,T[\,\times\Omega;\mathbb{R}^3\bigr)$, we have
$$
\lim_{M\rightarrow+\infty}\,\limsup_{\varepsilon\ra0}\,\left|\int^T_0\!\!\int_\Omega\rho_\varepsilon u_\varepsilon\otimes u_\varepsilon:\nabla\psi\,dx\,dt\,-\,
\int^T_0\!\!\int_\Omega V_{\varepsilon,M}\otimes V_{\varepsilon,M}:\nabla\psi\,dx\,dt\right|\,=\,0\,.
$$
\end{lemma}
Next, recall equation \eqref{eq:sing_weak}: paying attention once again to the right signs, by the previous lemma we have just to pass to
the limit in the term
\begin{eqnarray*}
& & \hspace{-0.5cm} -\,\int^T_0\!\!\int_\Omega V_{\varepsilon,M}\otimes V_{\varepsilon,M}:\nabla\psi\,=\,
\int^T_0\!\!\int_\Omega {\rm div}\,\left(V_{\varepsilon,M}\otimes V_{\varepsilon,M}\right)\,\cdot\,\psi \\
& & \qquad\qquad\quad =\,\int^T_0\!\!\int_\Omega {\rm div}\,_h\left(\langle V^h_{\varepsilon,M}\rangle\otimes\langle V^h_{\varepsilon,M}\rangle\right)\,\cdot\,\psi\,+\,
\int^T_0\!\!\int_\Omega {\rm div}\,_h\left(\langle \widetilde{V}^h_{\varepsilon,M}\otimes\widetilde{V}^h_{\varepsilon,M}\rangle\right)\,\cdot\,\psi \\
& & \qquad\qquad\quad =\,\int^T_0\!\!\int_\Omega\left(\mathcal{T}^1_{\varepsilon,M}\,+\,\mathcal{T}^2_{\varepsilon,M}\right)\,\cdot\,\psi\,.
\end{eqnarray*}
For notational convenience, from now on we will generically denote by $\mathcal{R}_{\varepsilon,M}$ any remainder, i.e. any term satisfying the property
\begin{equation} \label{eq:remainder}
\lim_{M\rightarrow+\infty}\,\limsup_{\varepsilon\ra0}\,\left|\int^T_0\int_\Omega \mathcal{R}_{\varepsilon,M}\,\cdot\,\psi\,dx\,dt\right|\,=\,0
\end{equation}
for all test functions $\psi\,\in\,\mathcal{D}\bigl([0,T[\,\times\Omega;\mathbb{R}^3\bigr)\,\cap\,{\rm Ker}\,{\mathcal{A}}$.
We give a sketch of how dealing with the terms $\mathcal{T}^1_{\varepsilon,M}$ and $\mathcal{T}^2_{\varepsilon,M}$, referring to Subsection 4.3.3 of \cite{F_2016}
for the details.
\paragraph{\bf Handling $\mathcal{T}^1_{\varepsilon,M}$}
Since we are dealing with smooth functions, we can integrate by parts: we get
\begin{eqnarray*}
\mathcal{T}^1_{\varepsilon,M} & = & {\rm div}\,_{\!h}\!\left(\langle V^h_{\varepsilon,M}\rangle\otimes\langle V^h_{\varepsilon,M}\rangle\right)\;=\;
{\rm div}\,_{\!h}\!\bigl(\langle V^h_{\varepsilon,M}\rangle\bigr)\;\langle V^h_{\varepsilon,M}\rangle\,+\,
\langle V^h_{\varepsilon,M}\rangle\cdot\nabla_h\left(\langle V^h_{\varepsilon,M}\rangle\right) \\
& = & {\rm div}\,_{\!h}\bigl(\langle V^h_{\varepsilon,M}\rangle\bigr)\;\langle V^h_{\varepsilon,M}\rangle\,+\,
\dfrac{1}{2}\,\nabla_h\left(\left|\langle V^h_{\varepsilon,M}\rangle\right|^2\right)\,+\,
{\rm curl}_h\langle V^h_{\varepsilon,M}\rangle\;\langle V^h_{\varepsilon,M}\rangle^\perp\,.
\end{eqnarray*}
Notice that the second term is a perfect gradient, and then it vanishes when tested against a function in the kernel of the singular perturbation operator.
For the first term, we take advantage of system \eqref{reg:approx-w}: averaging the first equation with respect to $x^3$
and multiplying it by $\langle V^h_{\varepsilon,M}\rangle$, we arrive at
$$
{\rm div}\,_h\bigl(\langle V^h_{\varepsilon,M}\rangle\bigr)\;\langle V^h_{\varepsilon,M}\rangle\;=\;-\,\varepsilon\,\partial_t\langle r_{\varepsilon,M}\rangle\,\langle V^h_{\varepsilon,M}\rangle\;=\;
\mathcal{R}_{\varepsilon,M}\,+\,\varepsilon\,\langle r_{\varepsilon,M}\rangle\,\partial_t\langle V^h_{\varepsilon,M}\rangle\,,
$$
since $\varepsilon\,\partial_t\bigl(\langle r_{\varepsilon,M}\rangle \,\langle V^h_{\varepsilon,M}\rangle\bigr)$ is a remainder in the sense specified by relation
\eqref{eq:remainder}. We use now the horizontal part of \eqref{reg:approx-w} (again, after taking the vertical average),
multiplied by $\langle r_{\varepsilon,M}\rangle$: paying attention to the signs, we get
$
\varepsilon\,\langle r_{\varepsilon,M}\rangle\,\partial_t\langle V^h_{\varepsilon,M}\rangle
\,=\,-\,\mathfrak{c}(x^h)\,\langle r_{\varepsilon,M}\rangle\,\langle V^h_{\varepsilon,M}\rangle^\perp\,+\,
\bigl({\rm Id}\,-\Delta_h\bigr)\langle r_{\varepsilon_M}\rangle\,\nabla_h\langle r_{\varepsilon_M}\rangle\,+\,\mathcal{R}_{\varepsilon,M}\,,
$
where we used also the properties proved in Proposition \ref{p:regular} and we included in the remainder term also the perfect gradient.
Inserting this relation into the expression for $\mathcal{T}^1_{\varepsilon,M}$, we find that this term equals
\begin{equation} \label{eq:T^1_a}
\mathcal{X}_{\varepsilon,M}\,\langle V^h_{\varepsilon,M}\rangle^\perp
\,+\,\bigl({\rm Id}\,-\Delta_h\bigr)\langle r_{\varepsilon_M}\rangle\,\nabla_h\langle r_{\varepsilon_M}\rangle\,+\,\mathcal{R}_{\varepsilon,M}\,,
\end{equation}
where, for notational convenience, we set $\mathcal{X}_{\varepsilon,M}\,:=\,{\rm curl}_h\langle V^h_{\varepsilon,M}\rangle\,-\,\mathfrak{c}(x^h)\,\langle r_{\varepsilon,M}\rangle$.
In order to deal with the first term in the right-hand side, the idea is to decompose $V^h_{\varepsilon,M}$ in the orthonormal basis
(up to normalization) $\left\{\nabla_h\mathfrak{c}\,,\,\nabla^\perp_h\mathfrak{c}\right\}$. Of course, this can be done in the region when
$\nabla_h\mathfrak{c}$ is far from $0$: therefore, we proceed carefully.
First of all, we notice that, after some manipulations, we can write
\begin{equation} \label{eq:curl-cr}
\varepsilon\,\partial_t\mathcal{X}_{\varepsilon,M}\,=\varepsilon\,{\rm curl}_h\langle f^h_{\varepsilon,M}\rangle\,+\,{\rm curl}_h\langle g^h_{\varepsilon,M}\rangle\,+\,\langle V^h_{\varepsilon,M}\rangle\,\cdot\,\nabla_h\mathfrak{c}(x^h)\,.
\end{equation}
Notice that, thanks to Proposition \ref{p:regular}, there exists a function $\eta\geq0$, with $\eta(M)\longrightarrow0$ for $M\rightarrow+\infty$,
such that, for any compact $K\,\subset\,\Omega$,
\begin{equation} \label{est:curl_rem}
\sup_{\varepsilon>0}\left\|{\rm curl}_h\langle g^h_{\varepsilon,M}\rangle\right\|_{L^2_T\bigl(L^2(K)\bigr)}
\,\leq\,\eta(M)\,.
\end{equation}
Then, fixed a $b\in\mathcal{C}^{\infty}_0(\mathbb{R}^2)$, with $0\leq b(x^h)\leq1$, such that $b\equiv1$ on $\left\{|x^h|\leq1\right\}$ and
$b\equiv0$ on $\left\{|x^h|\geq2\right\}$, we define
$$
b_M(x^h)\,:=\,b\left(\bigl(\eta(M)\bigr)^{\!-1/2}\,\nabla_h\mathfrak{c}(x^h)\right)\,.
$$
Now we are ready to deal with the first term in the right-hand side of \eqref{eq:T^1_a}.
On the one hand, using the decomposition
and the bounds established in Proposition \ref{p:decomp}, we deduce that, for any compact $K\subset\Omega$,
\begin{eqnarray*}
\left\|b_M\,\mathcal{X}_{\varepsilon,M}\,\langle V^h_{\varepsilon,M}\rangle^\perp\right\|_{L^1([0,T]\times K)} & \leq &
\varepsilon\,C(M)\,+\,C\,\left\|b_M\right\|_{L^6(K)} \\
& \leq & \varepsilon\,C(M)\,+\,C\,\left(\mathcal{L}\left\{x^h\in\mathbb{R}^2\;\bigl|\;\left|\nabla_h\mathfrak{c}(x^h)\right|\,\leq\,2\,\sqrt{\eta(M)}\right\}\right)^{\!1/6}\,.
\end{eqnarray*}
Therefore, thanks to hypothesis \eqref{eq:non-crit}, we infer that this term is a remainder, in the sense specified by
relation \eqref{eq:remainder}.
On the other hand, for $\nabla_h\mathfrak{c}$ far from $0$, we can write
$
\left(1-b_M\right)\mathcal{X}_{\varepsilon,M}\langle V^h_{\varepsilon,M}\rangle^\perp
=\left(1-b_M\right)\mathcal{X}_{\varepsilon,M}
\left(\frac{\langle V^h_{\varepsilon,M}\rangle\cdot\nabla_h\mathfrak{c}}{\left|\nabla_h\mathfrak{c}\right|^2}\nabla_h^\perp\mathfrak{c}+
\frac{\langle V^h_{\varepsilon,M}\rangle^\perp\cdot\nabla_h\mathfrak{c}}{\left|\nabla_h\mathfrak{c}\right|^2}\nabla_h\mathfrak{c}\right).
$
We observe that the latter term in the right-hand side is identically $0$ when tested against a $\psi\in{\rm Ker}\,\mathcal{A}$.
For the former term, instead, we use the expression found in \eqref{eq:curl-cr}: after some manipulations we get
\begin{eqnarray*}
\left(1-b_M\right)\,\mathcal{X}_{\varepsilon,M}\frac{\langle V^h_{\varepsilon,M}\rangle\cdot\nabla_h\mathfrak{c}}{\left|\nabla_h\mathfrak{c}\right|^2}\,
\nabla_h^\perp\mathfrak{c} & = & \frac{\varepsilon\,\left(1-b_M\right)\,\partial_t\left|\mathcal{X}_{\varepsilon,M}\right|^2}{2\,\left|\nabla_h\mathfrak{c}\right|^2}\,\nabla_h^\perp\mathfrak{c}\,- \\
& & -\,\frac{\left(1-b_M\right)\,\mathcal{X}_{\varepsilon,M}}{\left|\nabla_h\mathfrak{c}\right|^2}
\left(\varepsilon\,{\rm curl}_h\langle f^h_{\varepsilon,M}\rangle\,+\,{\rm curl}_h\langle g^h_{\varepsilon,M}\rangle\right)\,\nabla_h^\perp\mathfrak{c}\,,
\end{eqnarray*}
which is again a remainder $\mathcal{R}_{\varepsilon,M}$, thanks to Proposition \ref{p:decomp} and property \eqref{est:curl_rem}.
In the end, putting all these facts together, we have proved that (paying attention again to the right signs)
\begin{equation} \label{eq:T^1_final}
\mathcal{T}^1_{\varepsilon,M}\,=\,\bigl({\rm Id}\,-\Delta_h\bigr)\langle r_{\varepsilon_M}\rangle\,\nabla_h\langle r_{\varepsilon_M}\rangle\,+\,\mathcal{R}_{\varepsilon,M}\,.
\end{equation}
\paragraph{\bf Dealing with $\mathcal{T}^2_{\varepsilon,M}$}
Let us now consider the term $\mathcal{T}^2_{\varepsilon,M}$: exactly as done above, we can write
$$
\mathcal{T}^2_{\varepsilon,M}\,=\,\langle {\rm div}\,_{\!h}\bigl(\widetilde{V}^h_{\varepsilon,M}\bigr)\;\widetilde{V}^h_{\varepsilon,M}\rangle\,+\,
\dfrac{1}{2}\,\langle\nabla_h\left|\widetilde{V}^h_{\varepsilon,M}\right|^2\rangle\,+\,
\langle {\rm curl}_h\widetilde{V}^h_{\varepsilon,M}\;\left(\widetilde{V}^h_{\varepsilon,M}\right)^\perp\rangle\,.
$$
Let us focus on the last term for a while: with the notations introduced in \eqref{dec:vert-av}, we have
$$
\left({\rm curl}\widetilde{V}_{\varepsilon,M}\right)^h\,=\,\partial_3\widetilde{W}^h_{\varepsilon,M}\qquad\mbox{ and }\qquad
\left({\rm curl}\widetilde{V}_{\varepsilon,M}\right)^3\,=\,{\rm curl}_h\widetilde{V}^h_{\varepsilon,M}\,=\,\widetilde{\omega}^3_{\varepsilon,M}\,,
$$
where we have defined $\widetilde{W}^h_{\varepsilon,M}\,:=\,\left(\widetilde{V}^h_{\varepsilon,M}\right)^\perp\,-\,
\partial_3^{-1}\nabla^\perp_h\widetilde{V}^3_{\varepsilon,M}$.
For these quantities, from \eqref{reg:approx-w}, taking the mean-free part and the ${\rm curl}$ we deduce
\begin{equation} \label{eq:mean-free}
\begin{cases}
\varepsilon\,\partial_t\widetilde{W}^h_{\varepsilon,M}\,-\,\mathfrak{c}\,\widetilde{V}^h_{\varepsilon,M}\,=\,\left(\partial_3^{-1}\,
{\rm curl}\left(\varepsilon\,\widetilde{f}_{\varepsilon,M}\,+\,\widetilde{g}_{\varepsilon,M}\right)\right)^h \\[1ex]
\varepsilon\,\partial_t\widetilde{\omega}^3_{\varepsilon,M}\,+\,{\rm div}\,_{\!h}\bigl(\mathfrak{c}\,\widetilde{V}^h_{\varepsilon,M}\bigr)\,=\,{\rm curl}_h\left(\varepsilon\,\widetilde{f}^h_{\varepsilon,M}\,+\,
\widetilde{g}^h_{\varepsilon,M}\right)
\end{cases}
\end{equation}
Making use of the relations above and of Propositions \ref{p:regular} and \ref{p:decomp}, we get
\begin{eqnarray*}
{\rm curl}_h\widetilde{V}^h_{\varepsilon,M}\left(\widetilde{V}^h_{\varepsilon,M}\right)^\perp & = &
\frac{\varepsilon}{\mathfrak c}\partial_t\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\widetilde{\omega}^3_{\varepsilon,M}-\frac{\widetilde{\omega}^3_{\varepsilon,M}}{\mathfrak c}
\left(\partial_3^{-1}{\rm curl}\left(\varepsilon\widetilde{f}_{\varepsilon,M}+\widetilde{g}_{\varepsilon,M}\right)\right)^{h,\perp} \\
& = & \frac{1}{\mathfrak c}\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\,{\rm div}\,_{\!h}\bigl(\mathfrak{c}\,\widetilde{V}^h_{\varepsilon,M}\bigr)\,+\,\mathcal{R}_{\varepsilon,M}\,.
\end{eqnarray*}
Hence, including also the gradient term into the remainders and making some esy manipulations, we arrive at the equality
\begin{eqnarray*}
\mathcal{T}^2_{\varepsilon,M} & = & \langle {\rm div}\,\widetilde{V}_{\varepsilon,M}\,\left(\widetilde{V}^h_{\varepsilon,M}\,+\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\right)\rangle\,- \\
& & \qquad -\,\langle \partial_3\widetilde{V}^3_{\varepsilon,M}\,\left(\widetilde{V}^h_{\varepsilon,M}\,+\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\right)\rangle\,+\,
\langle \frac{1}{\mathfrak c}\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\,\widetilde{V}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c} \rangle\,+\,\mathcal{R}_{\varepsilon,M}\,.
\end{eqnarray*}
The second term on the right-hand side is actually another remainder.
As for the first term, instead, we use the equation for the density in \eqref{reg:approx-w} to obtain
$$
{\rm div}\,\widetilde{V}_{\varepsilon,M}\left(\widetilde{V}^h_{\varepsilon,M}+\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\right)\,=\,
\mathcal{R}_{\varepsilon,M}\,+\,\varepsilon\,\widetilde{r}_{\varepsilon,M}\;\partial_t\!\left(\widetilde{V}^h_{\varepsilon,M}+\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\right)\,.
$$
Now, by equations \eqref{eq:mean-free} and \eqref{reg:approx-w} again, it is easy to see that
$$
\varepsilon\,\widetilde{r}_{\varepsilon,M}\;\partial_t\!\left(\widetilde{V}^h_{\varepsilon,M}+\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\right)\,=\,
\mathcal{R}_{\varepsilon,M}\,+\,\nabla\widetilde{r}_{\varepsilon,M}\,\bigl({\rm Id}\,-\Delta\bigr)\widetilde{r}_{\varepsilon,M}\,,
$$
and therefore we find (with attention to the right sign)
$$
\mathcal{T}^2_{\varepsilon,M}\,=\,\langle \nabla\widetilde{r}_{\varepsilon,M}\,\bigl({\rm Id}\,-\Delta\bigr)\widetilde{r}_{\varepsilon,M}\rangle\,+\,
\langle \frac{1}{\mathfrak c}\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\,\widetilde{V}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c} \rangle\,+\,\mathcal{R}_{\varepsilon,M}\,.
$$
Now we have to deal with the second term in this last identity. Once again, we take advantage of the decomposition along the basis $\left\{\nabla_h\mathfrak{c}\,,\,\nabla^\perp_h\mathfrak{c}\right\}$.
For simplicity of exposition, we omit the cut-off away from the region $\{\nabla_h\mathfrak c=0\}$ and we just give a sketch of the argument, since it
is analogous to what done above for $\mathcal{T}^1_{\varepsilon,M}$.
First of all, we write
$$
\frac{\widetilde{V}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c}}{\mathfrak c}\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\,=\,\frac{\widetilde{V}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c}}{\mathfrak c}
\left(\widetilde{W}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c}\,\frac{\nabla_h^\perp\mathfrak{c}}{\left|\nabla_h\mathfrak{c}\right|^2}\,+\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\cdot\nabla_h\mathfrak{c}\,
\frac{\nabla_h\mathfrak{c}}{\left|\nabla_h\mathfrak{c}\right|^2}\right).
$$
As before, the last term in the right-hand side vanishes when tested against a smooth $\psi\in{\rm Ker}\,{\mathcal A}$.
Next, from the first equation in \eqref{eq:mean-free} we get
$$
\widetilde{V}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c}\,=\,\frac{1}{\mathfrak c}\left(\varepsilon\,\partial_t\widetilde{W}^h_{\varepsilon,M}\,-\,\left(\partial_3^{-1}\,
{\rm curl}\left(\varepsilon\,\widetilde{f}_{\varepsilon,M}\,+\,\widetilde{g}_{\varepsilon,M}\right)\right)^h\right)\cdot\nabla_h\mathfrak{c}\,.
$$
Therefore, we obtain that
\begin{eqnarray*}
\frac{1}{\mathfrak c}\,\left(\widetilde{W}^h_{\varepsilon,M}\right)^\perp\,\widetilde{V}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c} & = & \frac{\varepsilon}{2\,\mathfrak c^2}\,
\partial_t\left|\widetilde{W}^h_{\varepsilon,M}\cdot\nabla_h\mathfrak{c}\right|^2\,\frac{\nabla_h^\perp\mathfrak{c}}{\left|\nabla_h\mathfrak{c}\right|^2}\,- \\
& & \quad-\,\frac{1}{\mathfrak c^2}\,\left(\partial_3^{-1}\,{\rm curl}\left(\varepsilon\,\widetilde{f}_{\varepsilon,M}\,+\,\widetilde{g}_{\varepsilon,M}\right)\right)^h\cdot\nabla_h\mathfrak{c}\;
\frac{\nabla_h^\perp\mathfrak{c}}{\left|\nabla_h\mathfrak{c}\right|^2}\,,
\end{eqnarray*}
which is obviously a remainder in the sense of relation \eqref{eq:remainder}.
In the end, we have discovered that (paying attention to the right sign)
\begin{equation} \label{eq:T^2_final}
\mathcal{T}^2_{\varepsilon,M}\,=\,\langle \nabla\widetilde{r}_{\varepsilon,M}\,\bigl({\rm Id}\,-\Delta\bigr)\widetilde{r}_{\varepsilon,M}\rangle\,+\,\mathcal{R}_{\varepsilon,M}\,.
\end{equation}
\subsection{The limit equation}
Let us sum up what we have just proved. In order to pass to the limit in equation \eqref{eq:sing_weak}, we needed to treat the non-linearities coming
from the capillarity term and the convection term.
Putting relations \eqref{comp-cpt:dens}, \eqref{eq:T^1_final} and \eqref{eq:T^2_final} all together, we finally discover that
$$
\int^T_0\!\!\int_\Omega\biggl(-\,\bigl({\rm Id}\,-\Delta\bigr)r_{\varepsilon,M}\;\nabla r_{\varepsilon,M}\cdot\psi\,-\,V_{\varepsilon,M}\otimes V_{\varepsilon,M}:\nabla\psi\biggr)\,=\,
\int^T_0\!\!\int_\Omega\mathcal{R}_{\varepsilon,M}\cdot\psi\,,
$$
which immediately implies, together with Lemmas \ref{l:capill_approx} and \ref{l:conv_approx}, that
$$
\lim_{M\rightarrow+\infty}\,\lim_{\varepsilon\ra0}
\int^T_0\!\!\int_\Omega\biggl(\frac{1}{\varepsilon^2}\,\Delta\rho_\varepsilon\,\nabla\rho_\varepsilon\cdot\psi\,-\,\rho_\varepsilon\,u_\varepsilon\otimes u_\varepsilon:\nabla\psi\biggr)dx\,dt\,=\,0\,.
$$
Then, thanks to the previous computations, we can pass to the limit in the weak formulation of our system: we obtain
$$
\int^T_0\!\!\int_\Omega\biggl(-u\cdot\partial_t\psi-r\partial_t\bigl({\rm Id}\,-\Delta_h\bigr)\phi+\nu Du:\nabla\psi\biggr)dxdt\,=\,
\int_\Omega\bigl(u_0\cdot\psi(0)+r_0\bigl({\rm Id}\,-\Delta_h\bigr)\phi(0)\bigr)dx
$$
for any $(\phi,\psi)$ test functions belonging to the kernel of the singular perturbation operator ${\mathcal{A}}$.
Recall that, in particular, this implies the relation $\mathfrak{c}\,\psi^h\,=\,\nabla_h^\perp\bigl({\rm Id}\,-\Delta_h\bigr)\phi$.
Furthermore, also $(r,u)\,\in\,{\rm Ker}\,{\mathcal{A}}$: then we have the properties ${\rm div}\, u\equiv0$, $u=\bigl(u^h,0\bigr)$ and
$\mathfrak{c}\,u^h\,=\,\nabla_h^\perp\bigl({\rm Id}\,-\Delta_h\bigr)r$.
Setting $X(r)\,=\,\bigl({\rm Id}\,-\Delta_h\bigr)r$ and $\widetilde{\phi}\,=\,\bigl({\rm Id}\,-\Delta_h\bigr)\phi$, and
using that all the functions depend just on the horizontal variables, straightforward computations yield to
\begin{eqnarray*}
-\int^T_0\int_\Omega u\cdot\partial_t\psi\,dx\,dt & = &
\int^T_0\int_{\mathbb{R}^2}{\rm div}\,_{\!h}\!\left(\frac{1}{\mathfrak{c}^2}\,\nabla_hX(r)\right)\,\partial_t\widetilde{\phi}\,dx^h\,dt \\
\nu\int^T_0\int_\Omega Du:\nabla\psi\,dx\,dt & = &
\nu\int^T_0\int_{\mathbb{R}^2}\mathfrak{D}_{\mathfrak{c}}\bigl(X(r)\bigr)\,:\,\nabla_h\bigl(\mathfrak{c}^{-1}\,\nabla^\perp_h\widetilde{\phi}\,\bigr)\,dx^h\,dt \\
& = & \nu\int^T_0\int_{\mathbb{R}^2}\,^t\mathfrak{D}_{\mathfrak{c}}\,\circ\,\mathfrak{D}_{\mathfrak{c}}\bigl(X(r)\bigr)\,\widetilde{\phi}\,dx^h\,dt\,.
\end{eqnarray*}
Inserting these equalities into the previous relation finally completes the proof of Theorem \ref{t:sing_var}.
{\small
|
\section{Some examples and best-practices}
\section{Introduction}
After the amazing discovery of the relation between perturbative gauge theory and twistor string theory by Witten \cite{Witten:2003nn}, there have been several developments on computing scattering matrices in various theories from a moduli space on a punctured sphere \cite{Roiban:2004yf,Cachazo:2012da,Cachazo:2012kg,Huang:2012vt,Cachazo:2013iaa}. Cachazo, He and Yuan (CHY) proposed the equations governing the map from the space of kinematic invariants to the moduli space to be the same in each case and independent of the particular spacetime dimension. This led them to search for a more general formulation of scattering matrices in arbitrary dimension. Deriving some inspiration from a formula for MHV gravity amplitudes due to Hodges \cite{Bern:1998sv,Hodges:2012ym,Nguyen:2009jk}, CHY went on to discover their new formulation for amplitudes in a range of theories in \cite{Cachazo:2013gna,Cachazo:2013hca,Cachazo:2013iea}, and later \cite{Cachazo:2014nsa,Cachazo:2014xea}. This so called CHY formulation produces tree level $n$-point scattering amplitudes for massless particles in arbitrary dimension by means of $(n-3)$ moduli integrations localizing so called scattering equations. The scattering equations first appeared in the work of Fairlie and Roberts \cite{FairlieRoberts}, and later Gross and Mende \cite{GrossMende}, as well as more recently Witten \cite{Witten:2004cp}, and from the string theory classical worldsheet perspective in\footnote{The author thanks P. Caputa for pointing out this last point.} \cite{Caputa:2011zk,Caputa:2012pi}. Soon after the CHY equations made their appearance, the scalar and gluon cases were proven directly \cite{Dolan:2013isa} by means of BCFW recursion relations \cite{Britto:2004ap,Britto:2005fq}. Subsequently generalizations appeared, extending the formulation in terms of scattering equations to involve i.e. massive particles \cite{Dolan:2013isa,Naculich:2014naa,Naculich:2015zha,Naculich:2015coa}, fermions \cite{Weinzierl:2014ava}, supersymmetric theory \cite{Adamo:2015gia,Adamo:2015hoa}, one-loop amplitudes \cite{Adamo:2013tsa,Geyer:2015bja,Geyer:2015jch}, QCD related amplitudes \cite{delaCruz:2015raa}, off-shell amplitudes \cite{Lam:2015mgu}, or comparison to a string theory setting \cite{Mason:2013sva,Bjerrum-Bohr:2014qwa,Casali:2015vta}.\\
The most direct approach to evaluate amplitudes in CHY formulation was to try and find solutions to the scattering equations in general \cite{Weinzierl:2014vwa,Lam:2014tga}, or solve at special kinematics \cite{Kalousios:2013eca,Naculich:2014rta}. The scattering equations could also be reformulated in a polynomial form \cite{Dolan:2014ega,He:2014wua}. However, it became clear that solving scattering equations is very non-trivial and is not the most convenient way of evaluating amplitudes. Subsequently, techniques that avoid explicit solving of scattering equations started to emerge \cite{Kalousios:2015fya}. Contour deformations in the moduli integrals led to diagrammatic prescriptions that can be used to evaluate separate amplitude building blocks \cite{Cachazo:2015nwa,Baadsgaard:2015voa,Baadsgaard:2015ifa,Baadsgaard:2015hia,Lam:2015sqb}. An algebraic approach to evaluating scattering amplitudes in CHY formulation involving so-called companion matrices was suggested in \cite{Huang:2015yka}. For a comparison of this method with an elimination theory based technique see \cite{Cardona:2015eba}. One further algebraic technique involving polynomial inversion of moduli differences on the support of the ideal spanned by scattering equations, as well as the Bezoutian matrix to evaluate amplitudes was presented in \cite{Sogaard:2015dba}. Elimination theory was applied to scattering equations in polynomial form to obtain single variable polynomials \cite{Cardona:2015ouc,Dolan:2015iln}. Loop level integrands have been shown to follow from higher dimensional massless tree-level amplitudes \cite{Cachazo:2015aol,Feng:2016nrf}. Some further progress on evaluating CHY amplitudes was made in \cite{Lam:2016tlk}, diagrammatic techniques were generalized to compute higher order poles \cite{Huang:2016zzb}, and a double cover deformation of the moduli space led to evaluation of more general amplitude types as well \cite{Gomez:2016bmv}. Finally, monodromy relations were applied to Yang-Mills amplitudes in CHY representation to facilitate evaluation \cite{Bjerrum-Bohr:2016juj}.\\
In this work we start by developing a polynomial degree reduction procedure for multivariate polynomials in $\sigma$-moduli on the support of gauge fixed scattering equations for any $n$. As a consequence we realize that the most general multivariate polynomial in $\sigma$-moduli can be reduced to contain what we call \textit{ladder type} monomials only, with multivariate degree of at most $\frac{(n-3)(n-4)}{2}$ and coefficients rational in kinematic data. We say such a fully reduced polynomial is of \textit{standard form}. Application of Hilbert's strong Nullstellensatz as well as our degree reduction procedure conceptually allows us to find a standard form polynomial expression for rational functions in the $\sigma$-moduli. Making use of the above findings, a CHY amplitude integrand of any theory at any $n$ can be converted to a corresponding standard form polynomial. This general structural constraint is one of the main findings of the current work. After the polynomial reduction is carried out, we use the global residue theorem to derive a prescription to evaluate CHY amplitudes by collecting simple residues at infinity only. We note that only highest degree ladder type monomials contribute to any such amplitude integral, and since we find only simple poles the evaluation step is trivial. The difficulty is shifted towards finding standard form polynomial integrands for CHY amplitudes. We demonstrate the prescription on explicit examples of amplitude integrands at tree and one-loop level.\\
This paper is organized as follows. In section \ref{sec:CHY} we review the CHY formulation of tree-level scattering amplitudes for massless $\phi^3$ scalar theory as an example. As a warm up, section \ref{sec:warmup} shows a five point amplitude calculation to motivate our further investigation in section \ref{sec:reduction}. Section \ref{app:degred} describes the degree reduction of multivariate polynomials to the standard form, and section \ref{sec:polinv} extends the reduction procedure to rational functions, on the support of gauge fixed scattering equations. Subsequently, section \ref{sec:contours} describes the global residue theorem based proof for our amplitude evaluation prescription after polynomial reduction is applied to the integrands. In section \ref{sec:examples} we give explicit examples on how amplitudes are evaluated making use of our new method. We go on to consider 1-loop amplitudes in section \ref{sec:loop}, where we determine gauge fixed polynomial scattering equations that are free of singular solutions in the forward limit. Section \ref{sec:loopexpl} contains a few amplitude evaluation examples at 1-loop. We conclude in section \ref{sec:conclusion}. Appendix \ref{app:ratmom} suggests a simple method to generate real rational on-shell momenta based on Euclid's Pythagorean triple parametrization.
\paragraph{Note added:}{$~$}\\
When this work was being prepared for submission, J. Bosma, M. S{\o}gaard and Y. Zhang released a paper with similar results in \cite{Bosma:2016ttj}.
\section{CHY formulation of tree level scattering amplitudes}
\label{sec:CHY}
The Cachazo-He-Yuan (CHY) formulation of tree-level scattering amplitudes for massless particles in arbitrary dimension was introduced in \cite{Cachazo:2013gna,Cachazo:2013hca}. In CHY representation, the map of kinematic data to the moduli space is governed by the rational scattering equations
\begin{align}
\label{ratf}
f_a=\sum_{b=1,b\neq a}^n\frac{k_a\cdot k_b}{\sigma_a-\sigma_b}~~~~~~\forall a\in\{1,2,...,n\}.
\end{align}
Dolan and Goddard transformed the original amplitude expression to involve polynomial scattering equations \cite{Dolan:2014ega}. In what follows, it will be more convenient for us to work with polynomial scattering equations, therefore we will use the latter form for i.e. an $n$-point scalar $\phi^3$ amplitude in the examples to follow:
\begin{align}
\label{CHYDG}
A_n=\int \left(\prod_{{c=1}\atop{c\neq q,p,w}}^n d\sigma_c\right)(\sigma_{qp}\sigma_{pw}\sigma_{wq})\left(\prod_{1\leq i<j\leq n}\sigma_{ij}\right)\left(\prod_{a=2}^{n-2}\delta\left(\tilde{h}_a\right)\right)\frac{1}{(\sigma_{12}\sigma_{23}...\sigma_{n1})^{2}}.
\end{align}
Here the indices $1\leq q <p<w\leq n$ are fixed and can be chosen arbitrarily without changing the result. Minkowski momenta of scattering external particles are denoted $k_i$, and the difference of moduli is abbreviated as $\sigma_{ij}=\sigma_i-\sigma_j$. There are $n-3$ moduli integrations and the same amount of delta functions, such that the integral reduces to a sum over the solutions to the system of the scattering equations in the delta function arguments
\begin{align}
\label{scatOrig}
\tilde{h}_i\equiv \sum_{\{q_1,...,q_i\}\subset\{1,2,...,n\}}\mathfrak{s}_{q_1,...,q_i}\prod_{j=1}^i \sigma_{q_j}=0.
\end{align}
In this formula the summation is over all possible unordered subsets of $i$ different numbers $\{q_1,...,q_i\}$ out of the integer sequence from $1$ to $n$. Due to momentum conservation and massless on-shell conditions, the kinematic variables
\begin{align}
\label{kinS}
\mathfrak{s}_{q_1,...,q_i}=\frac{1}{2}\left(\sum_{j=1}^ik_{q_j}\right)^2
\end{align}
are only non-zero when at least $2$ or at most $n-2$ indices are provided. Therefore, exactly $n-3$ scattering equations (\ref{scatOrig}) from $\tilde{h}_2$ through $\tilde{h}_{n-2}$ are nontrivial.\\
In the following we will be working with the particular gauge choice $\sigma_1=\infty$, $\sigma_2=0$ and $\sigma_3=1$ for convenience. For this purpose we define the gauge fixed polynomial scattering equations:
\begin{align}
\label{scat}
h_i\equiv\left(\lim_{\sigma_1\rightarrow \infty}\frac{1}{\sigma_1}\tilde{h}_{i+1}\right)|_{{\sigma_2=0}\atop{\sigma_3=1}}=0~~~,~~~\forall i\in\{1,2,...,n-3\}.
\end{align}
Correspondingly, we will fix the free indices in (\ref{CHYDG}) as $q=1,p=2,w=3$.
\section{Warm up: five point tree level scalar amplitude}
\label{sec:warmup}
At five points we have two scattering equations:
\begin{align}
h_1=&\sigma _4 \mathfrak{s}_{1,4}+\sigma _5 \mathfrak{s}_{1,5}+\mathfrak{s}_{1,3}=0,\notag\\
h_2=&\sigma _4 \sigma _5 \mathfrak{s}_{2,3}+\sigma _5 \mathfrak{s}_{2,4}+\sigma _4 \mathfrak{s}_{2,5}=0.\notag
\end{align}
The gauge fixed scattering amplitude for scalars becomes
\begin{align}
\label{eq:ampl5sc}
A^{\phi^3}_5=\oint \frac{d\sigma_4d\sigma_5}{ h_1 h_2}\frac{\sigma _4 \sigma _5\left(1-\sigma _5\right)}{\left(1-\sigma _4\right) \left(\sigma _4-\sigma
_5\right)},
\end{align}
where the delta functions have been mapped to simple poles as usual, and the integration contour is such that both poles are localized. We would like to transform the integrand such that an evaluation via contour deformation becomes simpler. For that end, consider the following equality
\begin{align}
\label{5ptansatz}
\sigma _4 \sigma _5\left(1-\sigma _5\right)\hateq \left(1-\sigma _4\right) \left(\sigma _4-\sigma_5\right)N^{\phi^3}_5
\end{align}
where $\hateq$ shall denote equivalence on the support of scattering equations. Here $N^{\phi^3}_5$ clearly corresponds to the explicit integrand part of (\ref{eq:ampl5sc}). We claim that (\ref{5ptansatz}) can be realized i.e. by the following Ansatz
\begin{align}
\label{5ptNansatz}
N^{\phi^3}_5=c_1 \sigma_4+c_2\sigma_5.
\end{align}
To show that this is indeed the case, we can first solve $ h_1=0$ for either $\sigma_4$ or $\sigma_5$, and solve $ h_2=0$ for $\sigma_4\sigma_5$:
\begin{align}
\label{5ptsc1s}
\sigma _4 =-\frac{ \mathfrak{s}_{1,5}}{\mathfrak{s}_{1,4}}&\sigma _5-\frac{\mathfrak{s}_{1,3}}{\mathfrak{s}_{1,4}},~~~~~~\sigma _5=-\frac{\mathfrak{s}_{1,4}}{\mathfrak{s}_{1,5}}\sigma _4 -\frac{\mathfrak{s}_{1,3}}{\mathfrak{s}_{1,5}},\\
\label{5ptsc2s}
&\sigma _4 \sigma _5= - \frac{\mathfrak{s}_{2,4}}{\mathfrak{s}_{2,3}}\sigma _5- \frac{\mathfrak{s}_{2,5}}{\mathfrak{s}_{2,3}}\sigma _4.
\end{align}
Then we start with (\ref{5ptansatz}) making use of (\ref{5ptNansatz}), expand both sides of the equation, and iterate the following substitution rules:
\begin{enumerate}
\item Whenever we encounter a monomial featuring both $\sigma_4$ and $\sigma_5$, we isolate the highest power of $\sigma_4\sigma_5$, substitute in the right hand side of (\ref{5ptsc2s}) and expand - this leads to an overall multivariate degree reduction in monomials.
\item Whenever we encounter a monomial featuring $\sigma_4$ xor $\sigma_5$ to a power higher than one, we isolate a single power of $\sigma_4$ xor $\sigma_5$ respectively, substitute it by the right hand side of the respective equation in (\ref{5ptsc1s}) and expand - this leads either to an overall degree reduction in monomials, or to creation of new $\sigma_4\sigma_5$ terms.
\end{enumerate}
Iterating the above two steps a few times reduces both sides of (\ref{5ptansatz}) to only the two monomials $\sigma_4$ and $\sigma_5$ with some constant coefficients.\footnote{The exact coefficients are not necessarily unique and might depend on the order of substitutions during the reduction.} Collecting all terms on one side of the equation and demanding that the overall coefficients of monomials $\sigma_4$ and $\sigma_5$ vanish identically, we obtain a set of two linear equations in two unknowns $c_1$ and $c_2$. Solving these equations yields one possible solution for the Ansatz $N^{\phi^3}_5$, i.e.
\begin{align}
c_1=&\frac{\mathfrak{s}_{1,4}\mathfrak{s}_{2,5} \left(\left(\mathfrak{s}_{1,3}+\mathfrak{s}_{1,5}\right)
\mathfrak{s}_{2,4}+\mathfrak{s}_{1,4} \left(\mathfrak{s}_{2,3}+2
\mathfrak{s}_{2,4}+\mathfrak{s}_{2,5}\right)\right)}{
\left(\left(\mathfrak{s}_{1,3}+\mathfrak{s}_{1,4}\right)
\left(\mathfrak{s}_{2,3}+\mathfrak{s}_{2,4}\right)-\mathfrak{s}_{1,5}
\mathfrak{s}_{2,5}\right) \left(\mathfrak{s}_{1,3}
\mathfrak{s}_{2,3}-\left(\mathfrak{s}_{1,4}+\mathfrak{s}_{1,5}\right)
\left(\mathfrak{s}_{2,4}+\mathfrak{s}_{2,5}\right)\right)},\notag\\
c_2=&\frac{\mathfrak{s}_{2,4}(\mathfrak{s}_{1,3}+\mathfrak{s}_{1,4})}{\mathfrak{s}_{1,5}
\mathfrak{s}_{2,5}-\left(\mathfrak{s}_{1,3}+\mathfrak{s}_{1,4}
\right) \left(\mathfrak{s}_{2,3}+\mathfrak{s}_{2,4}\right)}+\frac{\mathfrak{s}_{1,5}\mathfrak{s}_{2,4}}{\left(\mathfrak{s}_{1,4}+\mathfrak{s}_{1,5}\right)
\left(\mathfrak{s}_{2,4}+\mathfrak{s}_{2,5}\right)-\mathfrak{s}_{1,3}
\mathfrak{s}_{2,3}}.\notag
\end{align}
Deforming the integration contours to infinity consecutively, we find only simple poles and get \footnote{In what follows, we give more details on this, from the point of view of global residue theorem.}
\begin{align}
\label{eq:ampl5scRes}
A^{\phi^3}_5=&\oint d\sigma_4 d\sigma_5\frac{N^{\phi^3}_5}{ h_1 h_2}=\frac{c_1}{\mathfrak{s}_{1,4} \mathfrak{s}_{2,3}}-\frac{c_2}{\mathfrak{s}_{1,5}
\mathfrak{s}_{2,3}}.
\end{align}
Using momentum conservation and the fact that all external particles are massless, we can re-express the above in the following familiar form
\begin{align}
A^{\phi^3}_5=\frac{1}{\mathfrak{s}_{1,2}
\mathfrak{s}_{3,4}}+\frac{1}{\mathfrak{s}_{5,1}\mathfrak{s}_{2,3} }+\frac{1}{\mathfrak{s}_{4,5}\mathfrak{s}_{1,2}}+\frac{1}{\mathfrak{s}_{3,4}\mathfrak{s}_{5,1} }+\frac{1}{\mathfrak{s}_{2,3}
\mathfrak{s}_{4,5}},
\end{align}
confirming that the result we found is indeed the correct five point massless scalar amplitude in $\phi^3$ theory. In the following section we will generalize the above technique to all $n$.
\section{Amplitude structure and evaluation prescription}
\label{sec:reduction}
Our plan is to show that any multivariate polynomial on the support of scattering equations can be written in a specific monomial structure we call the \textit{standard form}. Subsequently, we show that any rational function that is finite and non-vanishing on the support of scattering equations can be written as a standard form polynomial. Lastly, we apply these findings to amplitude integrands, convert them to standard form polynomials and evaluate the amplitude by means of the global residue theorem while collecting simple pole residues at infinity only.
\subsection{Degree reduction of polynomials to a standard form}
\label{app:degred}
In this section we start with an arbitrary multivariate polynomial $N$ in the $n-3$ different $\sigma$-moduli that are not gauge fixed (substitute $n\rightarrow n+2$ everywhere for 1-loop), and show that any such polynomial can be degree reduced to a very specific form.
\paragraph{Conventions:} Consider a generic monomial $M$ within polynomial $N$ separately:
\begin{align}
\label{genmonom}
M=C \sigma_{q_{1}}^{p_{1}}\sigma_{q_{2}}^{p_{2}}...\sigma_{q_{m_{max}}}^{p_{m_{max}}}.
\end{align}
$C$ is an overall constant, $q_{1}\neq q_{2}\neq...\neq q_{m_{max}}$ label the different $\sigma$-moduli appearing in the monomial $M$, while $p_{1},p_{2},...,p_{m_{max}}$ are the corresponding powers of each $\sigma$-modulus. We choose to always order all $\sigma$-moduli within each monomial such that $p_{1}\leq p_{2}\leq...\leq p_{m_{max}}$. For convenience we define $p_{0}\equiv 0$ for all $M$. Since there are at most $n-3$ different non-gauge fixed $\sigma$-moduli, we have $0\leq m_{max}\leq n-3$ in general.\footnote{The case $m_{max}=0$ corresponds to only $C$ being present in (\ref{genmonom}).}
\paragraph{Definition 1:} We define a monomial $M$ as introduced in the conventions above to be of \textit{ladder type} if its moduli powers satisfy $0\leq p_{j}-p_{j-1}\leq 1$ for all $j\in\{1,2,...,m_{max}\}$ when $m_{max}>0$, and iff additionally the property $0\leq m_{max}\leq n-4$ is satisfied.
\paragraph{Definition 2:} We define a multivariate polynomial in the non-gauge fixed $\sigma$-moduli to be of \textit{standard form} if it consists of \textit{ladder type} monomials only, with coefficients rational in kinematic data. See Table \ref{tab:ladmon} for some examples of ladder type monomials.
\begin{table}[h]
\centering
\begin{tabular}{ | c | c | c |}
\hline
$n=4$ & $n=5$ & $n=6$ \\ \hline
$1$ & $1,~\sigma_4,~\sigma_5$ & $1,~\sigma_4,~\sigma_5,~\sigma_6,~\sigma_4\sigma_5,~\sigma_4\sigma_6,~\sigma_5\sigma_6,~\sigma_5\sigma_4^2,~\sigma_6\sigma_4^2,~\sigma_6\sigma_5^2,~\sigma_4\sigma_5^2,~\sigma_4\sigma_6^2,~\sigma_5\sigma_6^2$ \\ \hline
\end{tabular}
\caption{Examples of all ladder type monomials for the first few $n$. ($\sigma_1,\sigma_2,\sigma_3$ gauge fixed.)}
\label{tab:ladmon}
\end{table}
\paragraph{Theorem 1:}\textit{On the support of the ideal spanned by scattering equations, an arbitrary regular multivariate polynomial $N$ in the $n-3$ non-gauge fixed moduli, with coefficients rational in kinematic data, is equivalent to at least one }standard form \textit{polynomial $N'$ that consists of ladder type monomials only, with coefficients rational in kinematic data.}
\paragraph{Proof:} To prove this we use flow arguments induced by scattering equation based transformations in the space of moduli powers within monomials. The arguments consist of the following two steps.
\paragraph{Step 1: Reduction of monomials to $0\leq p_{j}-p_{j-1}\leq 1$ for all $j\in\{1,2,...,m_{max}\}$}{$~$}\\
Consider a generic monomial of an arbitrary polynomial
\begin{align}
\label{arbimonom}
C \underbrace{\sigma_{q_{1}}^{p_{1}}\sigma_{q_{2}}^{p_{2}}...\sigma_{q_{j-1}}^{p_{j-1}}}_{w_1\text{ terms}}\underbrace{\sigma_{q_{j}}^{p_{j}}...\sigma_{q_{m_{max}}}^{p_{m_{max}}}}_{w_2\text{ terms}},
\end{align}
for some fixed $1\leq j\leq m_{max}$, ordered as $p_{1}\leq p_{2}\leq...\leq p_{m_{max}}$ and such that $p_{j}-p_{j-1}>1$, so that the monomial is non-ladder type. Also note that $0\leq (w_1+w_2=m_{max})\leq n-3$. If we want to transform this monomial into a sum over ladder type monomials, we first have to reduce the discrepancy $p_{j}-p_{j-1}>1$ to $0\leq p_{j}-p_{j-1}\leq 1$. We employ the scattering equations to do that as follows.\\
The general structure of gauge fixed polynomial scattering equations $ h_a=0$ for $a=1,...,n-3$ is such that $ h_a$ features all possible multilinear monomials of degree $a$ and $a-1$ respectively. Therefore, we can solve the scattering equation $ h_{w_2}=0$ for the monomial $\underbrace{\sigma_{q_{j}}...\sigma_{q_{m_{max}}}}_{w_2\text{ terms}}$:
\begin{align}
\label{Pfunct}
\sigma_{q_{j}}...\sigma_{q_{m_{max}}}=\sigma_{q_{j}}...\sigma_{q_{m_{max}}}-\frac{ h_{w_2}}{(\partial_{\sigma_{q_{j}}}...\partial_{\sigma_{q_{m_{max}}}} h_{w_2})}.
\end{align}
The derivatives in the denominator isolate the coefficient of monomial $\sigma_{q_{j}}...\sigma_{q_{m_{max}}}$ within $ h_{w_2}$. This coefficient is canceled for the corresponding summand in the numerator and the pure monomial is subtracted. Therefore, the right hand side of (\ref{Pfunct}) features all possible multilinear monomials of degree $w_2-1$ and all multilinear monomials of degree $w_2$ except for $\sigma_{q_{j}}...\sigma_{q_{m_{max}}}$.\\
We can now isolate $(\sigma_{q_{j}}...\sigma_{q_{m_{max}}})^{\floor{\frac{p_{j}-p_{j-1}}{2}}}$ moduli from the $w_2$ terms in (\ref{arbimonom}), substitute them by the right hand side of (\ref{Pfunct}) to the power $\floor{\frac{p_{j}-p_{j-1}}{2}}$ and expand.\footnote{The notation $\floor{x}$ means the floor function, returning the biggest integer $\leq x$.} Since each multilinear monomial of a certain degree is unique up to a constant factor, this has the effect that in each of the resulting terms
\begin{itemize}
\item the power of at least one modulus in the $w_2$ terms is reduced by at least one,
\item the power of at least one modulus in the $w_1$ terms is increased by at least one\footnote{Note that the power of this modulus could have been zero initially.}, or the overall degree is reduced.
\end{itemize}
Since the above guarantees a non-zero flow in the distribution of $\sigma$-moduli powers away from $w_2$ terms either into the $w_1$ terms or into overall degree reduction, iteration of the substitution rule for all $j$ and each monomial in the resulting terms is bound to reach a fixed point. This fixed point is straightforwardly given by the state where all monomials obey $0\leq p_{j}-p_{j-1}\leq 1$ for all $j\in\{1,2,...\}$ within each respective monomial, since then $\floor{\frac{p_{j}-p_{j-1}}{2}}=0$ for all $j$ and no substitutions can be carried out any more.
\paragraph{Step 2: Reduction of monomials to $m_{max}\leq n-4$}{$~$}\\
After step 1 is applied to all monomials in a polynomial $N$, it can still contain monomials with the maximal number of different moduli $m_{max}=n-3$:
\begin{align}
\label{almostladder}
C \sigma_{q_{1}}^{p_{1}}\sigma_{q_{2}}^{p_{2}}...\sigma_{q_{n-3}}^{p_{n-3}},
\end{align}
with $p_{1}=1$ and $0\leq p_{j}-p_{j-1}\leq 1$ for all $j\in\{2,3,...,n-3\}$. Similar to (\ref{Pfunct}), we can solve the gauge fixed polynomial scattering equation $ h_{n-3}=0$ for the single highest degree multilinear monomial $\sigma_{q_{1}}\sigma_{q_{2}}...\sigma_{q_{n-3}}$. Since that yields only multilinear terms of degree $n-4$, this necessarily leads to a degree reduction. We isolate the highest power of $\sigma_{q_{1}}\sigma_{q_{2}}...\sigma_{q_{n-3}}$ from monomials such as (\ref{almostladder}), make the substitution obtained from $ h_{n-3}=0$ and expand. Due to the guaranteed degree reduction in this step, we are again bound to iteratively reach a fixed point. This fixed point is trivially given by the condition $m_{max}\leq n-4$ for all resulting monomials $M$, since then no highest degree multilinear monomial can be isolated within the monomials, and therefore no substitutions can be carried out any more.
\paragraph{Conclusion}{$~$}\\
Step 1 and 2 above can be applied consecutively and iteratively to an arbitrary multivariate polynomial $N$. Due to the guaranteed degree reduction in step 2, both fixed points are bound to be reached simultaneously eventually. Therefore, we have shown that any polynomial $N$ on the support of scattering equations can be cast into a \textit{standard form} $N'$ containing only ladder type monomials.\footnote{The coefficients stay rational in kinematic data since we only used a finite number of additions and multiplications, and the coefficients in the scattering equations are rational as well.} Note that the degrees of the ladder type monomials $M_{lt}$ are $0\leq \text{deg}(M_{lt}) \leq\frac{(n-3)(n-4)}{2}$ at $n$ points. The full set of pure ladder type monomials at any $n$ is symmetric in all moduli. This homogeneity follows from the homogeneity of scattering equations that are used to achieve this form.
\subsection{Polynomial reduction of rational expressions}
\label{sec:polinv}
\paragraph{Theorem 2:}\textit{On the support of the ideal spanned by the scattering equations, any regular\footnote{By regular we mean non-infinite and non-zero on all solutions to the scattering equations.} multivariate rational function $\frac{P}{Q}$ in the $n-3$ non-gauge fixed moduli, where $P$ and $Q$ are polynomials with rational coefficients in kinematic data, is equivalent to at least one }standard form \textit{polynomial $N'$ that consists of ladder type monomials only, with rational coefficients in kinematic data.}
\paragraph{Proof:} Similar to some ideas of \cite{Sogaard:2015dba}, we will make use of Hilbert's Nullstellensatz. Consider the following equation involving the set of gauge fixed polynomial scattering equations $ h_m$ and multivariate polynomials in the $\sigma$-moduli $P,Q,a$ and $a_m$ for $m\in\{1,2,...,n-3\}$
\begin{align}
\label{Shilbert}
a\,Q+\sum_{m=1}^{n-3}{a_m h_m}=P.
\end{align}
The strong version of the Nullstellensatz guarantees that we can always find polynomials $a$ and $a_m$ for given polynomials $P$ and $Q$ such that (\ref{Shilbert}) is satisfied, as long as the $a,a_m,P$ and $Q$ do not share common roots among themselves and with the set of scattering equation polynomials $ h_m$. Considering the situation at the locus of solutions to the scattering equations, this simplifies to
\begin{align}
\label{Shilbert2}
a\,Q\hateq P,
\end{align}
where we use the symbol $\hateq$ to denote equivalence modulo the ideal spanned by the scattering equations. Thus, $a\hateq\frac{P}{Q}$ is a polynomial expression for a rational function.\footnote{Dividing by $Q$ is allowed since it is per assumption non-zero at the locus of solutions to the scattering equations.} Due to Theorem 1, a standard form polynomial $N'\hateq a$ must exist, which concludes the proof.
\paragraph{Construction:} In the above proof we used the fact that a standard form polynomial $N'\hateq a$ must exist, however the proof was not constructive. To construct an explicit $N'$ corresponding to a given rational function $\frac{P}{Q}$ we have to work harder. In principle, this step could be realized by various techniques. Here we will make use of an ad hoc procedure as follows.\\
Since the ladder type monomials span a complete polynomial basis with rational coefficients on the support of scattering equations, we can make an ansatz $\tilde{N}'$ containing all ladder type monomials with unfixed coefficients to parametrize our ignorance of what $N'$ actually is: $\tilde{N}'Q-P\hateq 0$. Making use of an implementation of the degree reduction procedure for Theorem 1, we can find a standard form polynomial $H'$ such that $H'\hateq \tilde{N}'Q-P\hateq0$. Demanding that the overall coefficient of each monomial in $H'$ vanishes separately, sets up a number of linear equations in (at least) the same number of unknown coefficients of $\tilde{N}'$.\footnote{Since there is a finite number of ladder type monomials at any $n$, the number of unfixed coefficients in $\tilde{N}'$ is at least equal to the number of monomials in a most general resulting standard form $H'$. If $H'$ has fewer than the maximum number of monomials, then the amount of unfixed coefficients is greater than the number of linear equations.} Solving this set of equations fixes the coefficients and yields an $N'$.
\\
In practice, in many cases of interest the ansatz for $\tilde N'$ does not require all ladder type monomials to be present to find a valid standard form polynomial $N'$. This reduces the dimension and complexity of the linear set of equations one has to solve. Additionally, we will see in the next section that only the coefficients of the highest degree ladder type monomials have a non-vanishing contribution to an amplitude integral.
\subsection{Collecting residues}
\label{sec:contours}
In this section we concentrate on tree level amplitudes for concreteness. However, at every step in the following it should be clear that essentially the same logic applies to the loop level integrands. Therefore, the result we find is valid in general.
\paragraph{Theorem 3:}
\textit{Any amplitude integral of the general shape}\footnote{Here, again, we consider the formulation where the delta functions have been mapped to simple poles with appropriate integration contours. Factors of $2\pi i$ are suppressed.}
\begin{align}
\label{contamp}
A_n=\oint\frac{N(\sigma_4,\sigma_5,...,\sigma_n)}{\prod_{j=1}^{n-3} h_j}\prod_{i=4}^nd\sigma_i ,
\end{align}
\textit{where $h_j$ for $j=1,...,n-3$ are gauge fixed scattering equation polynomials, $N(\sigma_4,\sigma_5,...,\sigma_n)$ is a} standard form \textit{polynomial in the $n-3$ non-gauge fixed moduli and the integration contour is initially localized at the locus of scattering equation solutions, can be evaluated by the following anti-symmetrized sum over the $(n-3)!$ different orders of consecutive infinity residues\footnote{The square brackets in $\sigma_{[n}=\infty,~...,~\sigma_{4]}=\infty$ denote anti-symmetrization with respect to the moduli indices, so that i.e. $\text{Res}_{\sigma_{[5}=\infty,~\sigma_{4]}=\infty}=\frac{1}{2!}(\text{Res}_{\sigma_{5}=\infty}\text{Res}_{\sigma_{4}=\infty}-\text{Res}_{\sigma_{4}=\infty}\text{Res}_{\sigma_{5}=\infty})$. The right most residue operation always acts first.}}
\begin{align}
\label{finalA}
A_n=(-1)^{n-3} (n-3)!~\text{Res}_{\sigma_{[n}=\infty,~...,~\sigma_{5}=\infty,~\sigma_{4]}=\infty}\left[\frac{N}{\prod_{j=1}^{n-3} h_j}\right].
\end{align}
\paragraph{Note:} Instead of calculating the $(n-3)!$ residues to evaluate the amplitude integral, it is possible to employ an integrand deformation in which the $h_i$'s are replaced by their leading homogeneous parts lt$(h_i)$. With this, the sum over residues equals one single residue at the origin by the transformation law of multivariate residues. This is an efficient alternative approach \cite{Cattani:1994}.\footnote{The author thanks the JHEP referee for pointing this out.}
\paragraph{Proof of Theorem 3:} Starting with (\ref{contamp}), it is straightforward to realize that any contour deformation away from the locus defined by the solutions to the scattering equations can possibly yield other residues only at infinity.\\
Decompose the numerator polynomial of the integrand into monomials $N=\sum_i N_i$. By additivity of integrals, consider the contour integral in pieces involving just one monomial $N_q=M\propto \prod_{r=4}^n\sigma_r^{a_{r}}$ at a time, where the integer powers $a_{r}\geq 0$ are such that $M$ is a ladder type monomial. Planning to investigate residues at infinity, we perform the substitution $\sigma_i\rightarrow1/\sigma_i$ and $d\sigma_i\rightarrow-d\sigma_i/\sigma_i^2$ for $i\in\{4,5,...,n\}$, to focus on residues at zero instead, so that
\begin{align}
\label{int1}
\oint\frac{\prod_{r=4}^n\sigma_r^{a_{r}}}{\prod_{j=1}^{n-3} h_j}\prod_{i=4}^nd\sigma_i\rightarrow \oint\frac{(-1)^{n-3}}{\left(\prod_{j=1}^{n-3}\hat{ h}_j\right)\left(\prod_{r=4}^n\sigma_r^{a'_{r}}\right)}\prod_{i=4}^nd\sigma_i,
\end{align}
where $a'_{r}=(a_{r}-n+5)$ is an abbreviation for the new integer exponents, and $\hat{ h}_j$ can be conveniently obtained from the gauge fixed scattering equations in the slightly different gauge $\sigma_1=0,~\sigma_2=\infty,~\sigma_3=1$.\footnote{The set of scattering equations is invariant under simultaneous inversion $\sigma\rightarrow1/\sigma$ of all $\sigma$-moduli (up to overall $\sigma$-moduli factors that here are accounted for by the powers $a'_{r}$), as long as we also invert the values of the gauge fixed moduli.}\\
Next we apply the global residue theorem (GRT), as for instance described in detail in \cite{ArkaniHamed:2009dn}. Consider a contour integral in $n-3$ variables over an integrand $1/f_1f_2...f_{n-3}$, such that the contours localize all possible poles in the integrand $f_i=0,\forall i$. Since all possible residues are collected in this way, it follows from the GRT that the result must be zero:
\begin{align}
\label{GRT}
\text{Res}_{\{f_1,f_2,...,f_{n-4},f_{n-3}\}}=0.
\end{align}
Using the above in our integrand of interest in (\ref{int1}), assign $f_i=\hat{ h}_i$ for $i=\{1,2,...,n-4\}$ and $f_{n-3}=\hat{ h}_{n-3}\prod_{r=4}^n{\sigma_r}^{a'_{r}}$. This clearly takes all possible poles into consideration, so that eq. (\ref{GRT}) is satisfied. Expand the global residue as a sum over the poles in $f_{n-3}$:
\begin{align}
\label{GRTdecomp}
\text{Res}_{\{f_1,f_2,...,f_{n-4},f_{n-3}\}}=\text{Res}_{\{f_1,f_2,...,f_{n-4},\hat{ h}_{n-3}\}}+\sum_{t=4}^{n}\text{Res}_{\{f_1,f_2,...,f_{n-4},{\sigma_t}^{a'_{t}}\}}=0.
\end{align}
The first summand corresponds to (\ref{int1}), so that we can re-express it in terms of the other $n-3$ residues $\text{Res}_{\{f_1,f_2,...,f_{n-4},\hat{ h}_{n-3}\}}=-\sum_{t=4}^{n}\text{Res}_{\{f_1,f_2,...,f_{n-4},{\sigma_t}^{a'_{t}}\}}$. Whenever partial poles in a multivariate residue calculation depend on one variable only, single variable complex analysis can be used to integrate out the corresponding residue separately. In our case each $\text{Res}_{\{f_1,f_2,...,f_{n-4},{\sigma_t}^{a'_{t}}\}}$, among other poles, involves a pole $1/{\sigma_t}^{a'_{t}}$ dependent on a single variable $\sigma_t$, which we will now integrate out separately.\\
Considering that $a'_{t}=(a_{t}-n+5)$ for each $t$, only highest degree ladder type monomials have a non-vanishing contribution to the integral, since exactly one of their $a_{t}$ satisfies $a_{t}=n-4$ which produces a simple pole as $1/{\sigma_t}^{a'_{t}}$. For all other ladder type monomials we have $0\leq a_{t}<n-4$ such that $a'_{t}\leq 0$ and $1/{\sigma_t}^{a'_{t}}$ ceases to be a pole and thus no residue is present.\\
To keep track of the correct contour orientation in the remaining variables, we anti-commute $d\sigma_t$ to one side in the integration measure $d\sigma_4\wedge d\sigma_5\wedge ...\wedge d\sigma_n=(\pm)_t d\sigma_t \prod^n_{i=4,i\neq t}(\wedge d\sigma_i)$. This produces an overall plus or minus $(\pm)_t$ dependent on the initial position $t$. Thus, we have
\begin{align}
\label{step1}
\oint\frac{\prod_{i=4}^nd\sigma_i}{\left(\prod_{j=1}^{n-3}\hat{ h}_j\right)\left(\prod_{r=4}^n\sigma_r^{a'_{r}}\right)}=\left\{ \begin{matrix}
- \oint\sum_{t=4}^n(\pm)_t\left(\frac{\prod_{{i=4}\atop{i\neq t}}^nd\sigma_i}{\left(\prod_{j=1}^{n-3}\hat{ h}_j\right)\left(\prod_{{r=4}\atop{r\neq t}}^n\sigma_r^{a'_{r}}\right)}\right)_{\sigma_t=0} & \text{ for }a'_{t}=1 \\
&\\
0 & \text{ for }a'_{t}<1 \end{matrix}, \right.
\end{align}
with the saturation $a'_{t}=1$ occurring for exactly one of the moduli in each highest degree ladder type monomial (nevertheless, we sum over all $\sum_{t=4}^n$ since it is not known a priori which label $t$ is going to yield the contribution).\footnote{
In terms of the expression in original variables on the left hand side of (\ref{int1}), this structurally means
\begin{align}
\label{int2}
\oint\frac{\prod_{r=4}^n\sigma_r^{a_{r}}}{\prod_{j=1}^{n-3} h_j}\prod_{i=4}^nd\sigma_i=&- \oint\sum_{u=4}^n(\pm)_u \text{Res}_{\sigma_u=\infty}\left[\frac{\prod_{r=4}^n\sigma_r^{a_{r}}}{\prod_{j=1}^{n-3} h_j}\right]\prod_{{i=4}\atop{i\neq u}}^nd\sigma_i,
\end{align}
where we imply that there are at most first order poles at infinity.}\\
As $\sigma_t=0$ in (\ref{step1}) is set, we find that $\hat{ h}_{n-3}$ reduces to a single monomial $\hat{ h}_{n-3}|_{\sigma_t=0}\propto \prod_{{j=4}\atop{j\neq t}}^n\sigma_j$ by general structure of scattering equations. Therefore, the non-vanishing contribution schematically becomes
\begin{align}
\label{int3}
\oint\sum_{t=4}^n(\pm)_t&\left(\frac{\prod_{{i=4}\atop{i\neq t}}^nd\sigma_i}{\left(\prod_{j=1}^{n-3}\hat{ h}_j\right)\left(\prod_{{r=4}\atop{r\neq t}}^n\sigma_r^{a'_{r}}\right)}\right)_{\sigma_t=0}= \oint\sum_{t=4}^n(\pm)_t\frac{C_{t}\prod_{{i=4}\atop{i\neq t}}^nd\sigma_i}{\left(\prod_{j=1}^{n-4}(\hat{ h}_j|_{\sigma_t=0})\right)\left(\prod_{{r=4}\atop{r\neq t}}^n\sigma_r^{a''_{r,t}}\right)}.
\end{align}
Here $C_{t}$ is one over the constant coefficient of the single monomial that survives as we take $\hat{ h}_{n-3}|_{\sigma_t=0}\propto \prod^n_{j=4,j\neq t}\sigma_j$, while the moduli of this monomial are accounted for by the new powers ${a''}_{r,t}$. The remaining $n-4$ scattering equation denominators $\hat{ h}_{j}|_{\sigma_t=0}$ now have the same monomial structure as scattering equation polynomials at $n-1$ points. Therefore, we can treat each summand in the sum over $t$ in (\ref{int3}) the same way as the initial expression (\ref{int1}), except now there is one fewer contour to integrate in each case. Thus, we can iterate. Noticing that by general structure of polynomial scattering equations we always get single monomials as more and more $\sigma_i$ are set to zero:
\begin{align}
\hat{ h}_{n-3}|_{\sigma_t=0}\propto \prod_{{j=4}\atop{j\neq t}}^n\sigma_j,~~~~~\hat{ h}_{n-4}|_{\sigma_t=0,~\sigma_l=0}\propto \prod_{{j=4}\atop{j\neq t,l}}^n\sigma_j,~~~~~\hat{ h}_{n-5}|_{\sigma_t=0,~\sigma_l=0,~\sigma_c=0}\propto \prod_{{j=4}\atop{j\neq t,l,c}}^n\sigma_j,~~~~~\text{etc.}\notag
\end{align}
ensures that each time a residue in a $\sigma$-modulus is collected, the remaining set of non-trivial scattering equation polynomials in the denominators is effectively reduced by one, as one of the scattering equation polynomials reduces to a single monomial and produces simple poles for the next iteration. With this, the above steps may be iterated from (\ref{int1}) to (\ref{int3}) $n-3$ times, while always expanding the resulting terms and summing over the process applied to one term at a time. Formally, each iteration adds one more level of signed infinity residue operations to (\ref{int2}). At the end of the day, when all contours have been treated, we are left with an anti-symmetrized sum over consecutive residue operations
\begin{align}
&\oint\frac{\prod_{r=4}^n\sigma_r^{a_{r}}}{\prod_{j=1}^{n-3} h_j}\prod_{i=4}^nd\sigma_i = (-1)^{n-3} (n-3)!~\text{Res}_{\sigma_{[n}=\infty,~...,~\sigma_{5}=\infty,~\sigma_{4]}=\infty}\left[\frac{\prod_{r=4}^n\sigma_r^{a_{r}}}{\prod_{j=1}^{n-3} h_j}\right].
\end{align}
This straightforwardly yields the full amplitude as we sum over all numerator monomials in the integrand, so that our final result for the amplitude is (\ref{finalA}). This concludes the proof.\\
\indent Due to the structure of standard form polynomials on the support of scattering equations we could rely on the fact that all residues we collect come from simple poles only. However, a straightforward generalization of the above steps yields the same result (\ref{finalA}) even for cases where $N$ is not a standard form polynomial and higher order residues are present.\\
It is interesting to note that the above procedure replaces a summation over $(n-3)!$ scattering equation solutions by a summation over the $(n-3)!$ different $(n-3)$-fold consecutive infinity residues in the $\sigma$-moduli. When $N$ is a standard form polynomial, all residues come from simple poles, such that the map from the integrand to the final result is trivial. With this the difficulty of the problem is shifted towards finding a standard form polynomial numerator $N$. Applying the degree reduction procedure described in the previous section this corresponds to solving a linear set on the order of $(n-3)!$ equations.
\subsection{Tree level amplitude examples}
\label{sec:examples}
In the following we demonstrate the evaluation prescription (\ref{finalA}) on $\phi^3$ scalar amplitudes at tree level. We also consider specific examples that otherwise require the more advanced evaluation techniques in order to be solved.
\subsubsection{Six point tree level scalar example}
At six points the three scattering equations are given by:
\begin{align}
h_1&=\sigma _4 \mathfrak{s}_{1,4}+\sigma _5 \mathfrak{s}_{1,5}+\sigma _6 \mathfrak{s}_{1,6}+\mathfrak{s}_{1,3}=0,\notag\\
h_2&=\sigma _4 \mathfrak{s}_{1,3,4}+\sigma _5 \mathfrak{s}_{1,3,5}+\sigma _6 \mathfrak{s}_{1,3,6}+\sigma _4
\sigma _5 \mathfrak{s}_{1,4,5}+\sigma _4 \sigma _6 \mathfrak{s}_{1,4,6}+\sigma _5 \sigma _6
\mathfrak{s}_{1,5,6}=0,\notag\\
h_3&=\sigma _4 \sigma _5 \sigma _6 \mathfrak{s}_{2,3}+\sigma _5 \sigma _6 \mathfrak{s}_{2,4}+\sigma _4 \sigma _6
\mathfrak{s}_{2,5}+\sigma _4 \sigma _5 \mathfrak{s}_{2,6}=0.\notag
\end{align}
The gauge fixed scattering amplitude for scalars is given by
\begin{align}
\label{eq:ampl6sc}
A^{\phi^3}_6=\oint \frac{d\sigma_4d\sigma_5d\sigma_6}{ h_1 h_2 h_3}\frac{\sigma _4 \left(1-\sigma _5\right) \sigma _5 \left(1-\sigma _6\right) \left(\sigma _4-\sigma _6\right) \sigma _6}{\left(1-\sigma _4\right) \left(\sigma _4-\sigma _5\right) \left(\sigma _5-\sigma _6\right)}.
\end{align}
Applying partial fraction decomposition as well as transformations by rational scattering equations (\ref{ratf}), we can rewrite the integrand of (\ref{eq:ampl6sc}) as
\begin{align}
\label{partfrac6}
\frac{\sigma _4 \left(1-\sigma _5\right) \sigma _5 \left(1-\sigma _6\right) \left(\sigma _4-\sigma _6\right) \sigma _6}{\left(1-\sigma _4\right) \left(\sigma _4-\sigma _5\right) \left(\sigma _5-\sigma _6\right)}\hateq \frac{P_1}{\sigma_4-\sigma_5}+P_2
\end{align}
where $P_1$ and $P_2$ are polynomials. To reduce the rational part to a polynomial, we take
\begin{align}
\label{6ptansatz}
P_1\hateq \left(\sigma _4-\sigma _5\right) P_3
\end{align}
with the following standard form Ansatz:\footnote{Ladder type monomials with base length $m_{max}=n-4$ appear to be a sufficient monomial basis.}
\begin{align}
P_3=c_1 \sigma _5 \sigma _4^2+c_2 \sigma _4 \sigma _5^2+c_3 \sigma _6 \sigma _4^2+c_4 \sigma _4\sigma _6^2+c_5 \sigma _6 \sigma _5^2+c_6 \sigma _5 \sigma _6^2 +c_7 \sigma _5 \sigma _4+c_8 \sigma _6 \sigma _4+c_9 \sigma _5 \sigma_6.\notag
\end{align}
There are nine constants $c_i$ with $i=1,2,...,9$ we have to fix. We apply the reduction procedure of section \ref{app:degred} to both sides of (\ref{6ptansatz}), collect all terms on one side of the equation and demand that the overall coefficient in front of each monomial vanishes. This produces a set of nine linear equations in nine unknowns. Solving the set of linear equations fixes the nine unknown coefficients and thus yields a polynomial $P_3$. With this, also reducing $P_2$ to contain ladder type monomials only, a standard form numerator polynomial $N^{\phi^3}_6\hateq P_2+P_3$ is obtained. It takes a direct implementation of the polynomial reduction algorithm in \textit{Mathematica} and a linear solver just a few seconds to find a valid analytic $N^{\phi^3}_6$ result, without much effort spent on optimization.\footnote{If we start with the left hand side of eq. (\ref{partfrac6}) instead, as in $\sigma _4 \left(1-\sigma _5\right) \sigma _5 \left(1-\sigma _6\right) \left(\sigma _4-\sigma _6\right) \sigma _6\hateq \left(1-\sigma _4\right) \left(\sigma _4-\sigma _5\right) \left(\sigma _5-\sigma _6\right)N^{\phi^3}_6$, it takes the polynomial reduction algorithm and linear solver, with a few tweaks, about a minute to obtain a different more complicated analytic version of $N^{\phi^3}_6$.} We can evaluate the amplitude making use of prescription (\ref{finalA}):
\begin{align}
\label{eq:ampl6scRes}
A^{\phi^3}_6&=(-1)^3 3!~\text{Res}_{\sigma_{[6}=\infty,~\sigma_{5}=\infty,~\sigma_{4]}=\infty}\left[\frac{N^{\phi^3}_6}{\prod_{j=1}^{n-3} h_j}\right].
\end{align}
The result is completely analytic and about one page long. It can be simplified making use of momentum conservation and on-shell conditions by hand, which is somewhat tedious. Instead we set up a basis of physical poles and fix the coefficients by multiple evaluation on different kinematic points as follows.\\
As in \cite{Huang:2015yka}, the physical poles are given by $\mathfrak{s}_{1,2},\mathfrak{s}_{2,3},\mathfrak{s}_{3,4},\mathfrak{s}_{4,5},\mathfrak{s}_{5,6},\mathfrak{s}_{6,1},\mathfrak{s}_{1,2,3},\mathfrak{s}_{2,3,4}$ and $\mathfrak{s}_{3,4,5}$. By dimensional analysis we see that each term in the amplitude should have three different poles. This means the complete basis is given by $\binom{9}{3}=84$ different triple pole combinations with unknown coefficients. Making use of the procedure described in appendix \ref{app:ratmom}, we can generate $84$ different rational kinematic points and evaluate the amplitude and the basis $84$ times. This sets up a linear set of $84$ equations in the same number of unknowns. Solving this set of equations fixes the coefficients (which turn out to be exactly $1$ or $0$) and yields the simplified $6$-point scalar tree level amplitude in terms of physical poles
\begin{align}
\label{eq:ampl6scRes}
A^{\phi^3}_6=-&\left( \frac{1}{\mathfrak{s}_{1,2} \mathfrak{s}_{3,4} \mathfrak{s}_{5,6}}+\frac{1}{\mathfrak{s}_{1,2} \mathfrak{s}_{5,6} \mathfrak{s}_{1,2,3}}+\frac{1}{\mathfrak{s}_{2,3} \mathfrak{s}_{5,6}
\mathfrak{s}_{1,2,3}}+\frac{1}{\mathfrak{s}_{1,6} \mathfrak{s}_{2,3} \mathfrak{s}_{2,3,4}}+\frac{1}{\mathfrak{s}_{1,6} \mathfrak{s}_{3,4} \mathfrak{s}_{2,3,4}}\right.\\
&+\frac{1}{\mathfrak{s}_{2,3}
\mathfrak{s}_{5,6} \mathfrak{s}_{2,3,4}}+\frac{1}{\mathfrak{s}_{3,4} \mathfrak{s}_{5,6} \mathfrak{s}_{2,3,4}}+\frac{1}{\mathfrak{s}_{1,2} \mathfrak{s}_{3,4}
\mathfrak{s}_{3,4,5}}+\frac{1}{\mathfrak{s}_{1,6} \mathfrak{s}_{3,4} \mathfrak{s}_{3,4,5}}+\frac{1}{\mathfrak{s}_{1,6} \mathfrak{s}_{2,3} \mathfrak{s}_{4,5}}\notag\\
&\left.+\frac{1}{\mathfrak{s}_{1,2}
\mathfrak{s}_{1,2,3} \mathfrak{s}_{4,5}}+\frac{1}{\mathfrak{s}_{2,3} \mathfrak{s}_{1,2,3} \mathfrak{s}_{4,5}}+\frac{1}{\mathfrak{s}_{1,2} \mathfrak{s}_{3,4,5}
\mathfrak{s}_{4,5}}+\frac{1}{\mathfrak{s}_{1,6} \mathfrak{s}_{3,4,5} \mathfrak{s}_{4,5}}\right),\notag
\end{align}
which is equivalent to summing Feynman diagrams in $\phi^3$ theory and agrees with the result found in \cite{Huang:2015yka}.
\subsubsection{Six point tree level - first special example}
Here we will give an example that is very hard to do with less advanced versions of diagrammatic integration rule techniques.\footnote{The author thanks J. Bourjaily for pointing this out and suggesting this test integrand.} It involves integrating the following terms over the CHY measure
\begin{align}
\frac{1}{\sigma _{2,3}^4 \sigma _{4,5}^4 \sigma _{6,1}^4}.
\end{align}
Multiplying with the CHY measure and applying our gauge we get
\begin{align}
U_1=\oint \frac{d\sigma_4d\sigma_5d\sigma_6}{ h_1 h_2 h_3}\frac{\left(1-\sigma _4\right) \sigma _4 \left(1-\sigma _5\right) \sigma _5 \left(1-\sigma _6\right) \left(\sigma _4-\sigma _6\right) \left(\sigma _5-\sigma _6\right) \sigma _6}{\left(\sigma _4-\sigma _5\right)^3}.
\end{align}
In order to polynomially reduce the effective rational integrand, we write
\begin{align}
\label{special1}
\left(1-\sigma _4\right) \sigma _4 \left(1-\sigma _5\right) \sigma _5 \left(1-\sigma _6\right) \left(\sigma _4-\sigma _6\right) \left(\sigma _5-\sigma _6\right) \sigma _6\hateq\left(\sigma _4-\sigma _5\right)^3N
\end{align}
where we use the following standard form polynomial Ansatz
\begin{align}
N=c_1 \sigma _5 \sigma _4^2+c_2 \sigma _4 \sigma _5^2+c_3 \sigma _6 \sigma _4^2+c_4 \sigma _4\sigma _6^2+c_5 \sigma _6 \sigma _5^2+c_6 \sigma _5 \sigma _6^2 +c_7 \sigma _5 \sigma _4+c_8 \sigma _6 \sigma _4+c_9 \sigma _5 \sigma_6.\notag
\end{align}
We have to find nine constants $c_1,c_2,...,c_9$. A completely analytic result is directly accessible applying our procedure, yet not very readable.\footnote{Here an analytic $N$ can be obtained from the polynomial reduction algorithm and a linear solver within $1$ to $2$ minutes. This timing probably could be substantially improved by optimization.} We do not expect the result to be given by pure physical poles either. Therefore, we will instead demonstrate an explicit exact evaluation of the integral on the following kinematic point, which was generated making use of the procedure described in appendix \ref{app:ratmom}:
\begin{align}
\label{kin6}
k_1^\mu&=(20,~\,\, 20,~\,\,\,~ 0, \,\,\,\,\,\,\,\, 0),&k_4^\mu&=(\,\,\,\,60, -48, \,\,\,\,\,~0, -36),\notag\\
k_2^\mu&=(25, -20, \,\,\,\,15, \,\,\,\,\,\,\,\, 0),&k_5^\mu&=(-80, \,\,\,\,48,\,\,\,\, 64, \,\,\,\,\,~0),\\
k_3^\mu&=(39,~\,\,\,~ 0, -15, \,\,\,\,36),&k_6^\mu&=(-64, \,\,\,\,\,~0, -64, \,\,\,\,\,~0).\notag
\end{align}
First we apply the degree reduction procedure of section \ref{app:degred} to both sides of equation (\ref{special1}) and collect all monomials on one side. The vanishing of the overall coefficient of each monomial separately produces a set of linear equations. Solving this set of equations yields
\begin{align*}
&{\scriptstyle c_5= \frac{7059649218217401322274}{3974168469797996315755} ~~~~\,~~,~~~ c_6= -\frac{5529649875686983344959}{15896673879191985263020} ~~~,~~~ c_2= \frac{12838684423}{1662217245} ~~,~~~ c_4= \frac{354818034905}{57180273228} }\\
& {\scriptstyle c_7=-\frac{5774994253402805042003591}{2146050973690918010507700} ~~,~~~ c_8= -\frac{466431129022169341083793}{343368155790546881681232} ~~,~~~ c_9= -\frac{70384223902707859416469}{158966738791919852630200}~~,~~~c_1=c_3=0.}
\end{align*}
Using this in the Ansatz for $N$ above, we obtain a standard form numerator polynomial and can apply (\ref{finalA}) to evaluate the integral:
\begin{align}
\label{u1result}
U_1&=(-1)^3 3!~\text{Res}_{\sigma_{[6}=\infty,~\sigma_{5}=\infty,~\sigma_{4]}=\infty}\left[\frac{N}{\prod_{j=1}^{n-3} h_j}\right]\notag\\
&=-\frac{c_2}{\mathfrak{s}_{1,5} \mathfrak{s}_{2,3} \mathfrak{s}_{1,4,5}}+\frac{c_4}{\mathfrak{s}_{1,6} \mathfrak{s}_{2,3} \mathfrak{s}_{1,4,6}}+\frac{c_5}{\mathfrak{s}_{1,5} \mathfrak{s}_{2,3}
\mathfrak{s}_{1,5,6}}-\frac{c_6}{\mathfrak{s}_{1,6} \mathfrak{s}_{2,3} \mathfrak{s}_{1,5,6}}\notag\\
&=\frac{14174374134763}{40854136935339786240000}.
\end{align}
Note that indeed properly only the coefficients of highest degree ladder type monomials appear in the final result.\\
Alternatively, we can solve the scattering equations numerically and obtain a numerical approximation for $U_1$, which agrees with (\ref{u1result}).
\subsubsection{Six point tree level - second special example}
Another example that is impossible to do with less advanced diagrammatic integration rule techniques involves integrating the following terms over the CHY measure\footnote{Again, the author thanks J. Bourjaily for pointing this out and suggesting this test integrand.}
\begin{align}
\frac{1}{\sigma _{2,3}^2 \sigma _{3,4}^2 \sigma _{4,2}^2 \sigma _{1,5}^2 \sigma _{5,6}^2 \sigma _{6,1}^2}.
\end{align}
Combining this with the CHY measure and applying our usual gauge we have
\begin{align}
U_2=\oint \frac{d\sigma_4d\sigma_5d\sigma_6}{ h_1 h_2 h_3}\frac{\left(1-\sigma _5\right) \left(\sigma _4-\sigma _5\right) \sigma _5 \left(1-\sigma _6\right) \left(\sigma _4-\sigma _6\right) \sigma _6}{\left(1-\sigma _4\right) \sigma _4 \left(\sigma _5-\sigma _6\right)}.
\end{align}
In order to polynomially reduce the effective rational integrand, we write the equation
\begin{align}
\label{special2}
\left(1-\sigma _5\right) \left(\sigma _4-\sigma _5\right) \sigma _5 \left(1-\sigma _6\right) \left(\sigma _4-\sigma _6\right) \sigma _6\hateq\left(1-\sigma _4\right) \sigma _4 \left(\sigma _5-\sigma _6\right)N
\end{align}
where we use the following standard form polynomial Ansatz
\begin{align}
N=c_1 \sigma _5 \sigma _4^2+c_2 \sigma _4\sigma _5^2 +c_3 \sigma _6 \sigma _4^2+c_4 \sigma _4\sigma _6^2 +c_5 \sigma _6\sigma _5^2 +c_6 \sigma _5 \sigma _6^2+c_7 \sigma _5 \sigma _4+c_8 \sigma _6 \sigma _4+c_9 \sigma _5 \sigma_6.\notag
\end{align}
So that again there are nine constants $c_1,c_2,...,c_9$ to be fixed. Just as before, we can proceed completely analytically, yet the result would be too large to report.\footnote{Here, again, an analytic $N$ can be obtained from the polynomial reduction algorithm and a linear solver within $1$ to $2$ minutes. This timing probably could be substantially improved by optimization.} Therefore, we will illustrate the procedure by evaluating the integral on the kinematic point (\ref{kin6}) instead.\\
First we apply the degree reduction procedure of section \ref{app:degred} to both sides of equation (\ref{special2}) and collect all monomials on one side of the equation. Demanding that the overall coefficient of each monomial vanishes separately provides us with a set of linear equations. Solving the set of equations we obtain
\begin{align*}
&c_5= \frac{162215379551}{1221259549104} ~~~~~~,~~~ c_6= \frac{5662761717335}{17097633687456} ~~\,~~,~~~ c_4= -\frac{92500}{133623} ~~~,~~~ c_2= \frac{39458}{133623}~~~, \\
& c_7= -\frac{23433636506339}{34195267374912} ~~~,~~\,\, c_8=
\frac{329688097714075}{273562138999296} ~~~,~~~ c_9= -\frac{3664568494697}{3256692130944},~~~c_1=c_3=0.
\end{align*}
Plugging this into the Ansatz for $N$ above, we therefore have obtained a standard form numerator polynomial and can use (\ref{finalA}) to evaluate the integral:
\begin{align}
\label{u2result}
U_2&=(-1)^3 3!~\text{Res}_{\sigma_{[6}=\infty,~\sigma_{5}=\infty,~\sigma_{4]}=\infty}\left[\frac{N}{\prod_{j=1}^{n-3} h_j}\right],\notag\\
&=-\frac{c_2}{\mathfrak{s}_{1,5} \mathfrak{s}_{2,3} \mathfrak{s}_{1,4,5}}+\frac{c_4}{\mathfrak{s}_{1,6} \mathfrak{s}_{2,3} \mathfrak{s}_{1,4,6}}+\frac{c_5}{\mathfrak{s}_{1,5} \mathfrak{s}_{2,3}
\mathfrak{s}_{1,5,6}}-\frac{c_6}{\mathfrak{s}_{1,6} \mathfrak{s}_{2,3} \mathfrak{s}_{1,5,6}}\notag\\
&=-\frac{2407}{15692753534976}.
\end{align}
Note that again properly only the coefficients of the highest degree ladder type monomials enter the final result. Additionally, it is clear that the calculation for this example structurally follows exactly the same steps and has the same level of complexity as the previous two examples, which would have been different from the point of view of applying diagrammatic integration rules to evaluate the integral.\\
Alternatively, we can solve the scattering equations numerically and obtain a numerical approximation for $U_2$, which agrees with (\ref{u2result}).
\subsubsection{Eight point tree level scalar amplitude}
At eight points there are five scattering equations. The gauge fixed scattering amplitude for scalars reads\footnote{Where $\sigma_3=1$ is implied.}
\begin{align}
\label{eq:ampl8sc}
A^{\phi^3}_8=\oint \frac{\prod_{i=4}^8 d\sigma_i}{\prod_{j=1}^{5} h_j}\frac{\sigma _4 \sigma _5 \sigma _6 \sigma _7 \sigma _8 \sigma _{3,5} \sigma _{3,6} \sigma _{3,7} \sigma _{3,8} \sigma _{4,6} \sigma _{4,7} \sigma _{4,8} \sigma _{5,7} \sigma _{5,8} \sigma _{6,8}}{\sigma _{3,4} \sigma_{4,5} \sigma _{5,6} \sigma _{6,7} \sigma _{7,8}}.
\end{align}
We will demonstrate an explicit evaluation of the amplitude. Making use of the procedure described in appendix \ref{app:ratmom}, we generate some on-shell kinematic data
\begin{align}
\label{kin8}
k_1^\mu&=(\,\,\,\,-54, \,\,\,\,-54, ~\,\,\,\,\,\,\,\,0,\,\,\,\,\,\,\,\, 0),&k_5^\mu&=(-85,\,\,\,\, 0,\,\,\,\, 75,\,\,\,\, 40),\notag\\
k_2^\mu&=(-246,\,\,\,\,\,\,\,\, 54, -240, \,\,\,\,\,\,\,\,0),&k_6^\mu&=(\,\,\,\,50,\,\,\,\, 0, -30, -40),\\
k_3^\mu&=(\,\,\,\,260, \,\,\,\,100, \,\,\,\,240, \,\,\,\,\,\,\,\,0),&k_7^\mu&=(-34,\,\,\,\, 0,\,\,\,\, 30, -16),\notag\\
k_4^\mu&=(\,\,\,\,125, -100,\,\,\,\, -75, \,\,\,\,\,\,\,\,0),&k_8^\mu&=(-16,\,\,\,\, 0,\,\,\,\,\,\,\,\, 0,\,\,\,\, 16).\notag
\end{align}
We want to find an effective integral expression
\begin{align}
A_8^{\phi^3}=\oint \frac{\prod_{i=4}^8 d\sigma_i}{\prod_{j=1}^{5} h_j}N^{\phi^3}_8,
\end{align}
where $N^{\phi^3}_8$ is a standard form polynomial satisfying
\begin{align}
\label{8ptRedAnsatz}
\sigma _4 \sigma _5 \sigma _6 \sigma _7 \sigma _8 \sigma _{3,5} \sigma _{3,6} \sigma _{3,7} \sigma _{3,8} \sigma _{4,6} \sigma _{4,7} \sigma _{4,8} \sigma _{5,7} \sigma _{5,8} \sigma _{6,8}\hateq \sigma _{3,4} \sigma_{4,5} \sigma _{5,6} \sigma _{6,7} \sigma _{7,8}N^{\phi^3}_8
\end{align}
on the support of the ideal spanned by the scattering equations. As an Ansatz for $N^{\phi^3}_8$ we take the $375$ different ladder type monomials with $m_{max}=n-4=4$. At eight points, polynomially reducing the complete right hand side of (\ref{8ptRedAnsatz}) proves to be time consuming. Therefore, we instead perform a much simpler polynomial reduction of the expression $\sigma_iN_{\sigma_i}\rightarrow N'_{\sigma_i}$ for $i=4,...,8$ with the same Ansatz for $N_{\sigma_i}$.\footnote{The resulting polynomial $N'_{\sigma_i}$ features the same monomials as $N_{\sigma_i}$, but with the coefficients mixed by the reduction procedure.} These results can now be straightforwardly linearly combined as in $(\sigma_i-\sigma_j)N\rightarrow N'_{\sigma_i}-N'_{\sigma_j}\equiv N'_{\sigma_{ij}}$. Additionally, we can nest them by computing the reduction in steps of one degree at a time $(\sigma_i-\sigma_j)(\sigma_a-\sigma_b)N\rightarrow (\sigma_i-\sigma_j)N'_{\sigma_{ab}}\rightarrow N''_{\sigma_{ij}\sigma_{ab}}$, where in the second step we treat the complete monomial coefficients of $N'_{\sigma_{ab}}$ as simple unknowns and substitute their structure back in once the reduction has been performed. Clearly, we can apply the nesting as many times as required. Therefore, the polynomial reduction of $\sigma_iN_{\sigma_i}$ is the only building block we need to construct the complete effective numerator polynomial $N^{\phi^3}_8$.\\
Furthermore, it is more convenient to fractionally decompose the integrand in $(\ref{eq:ampl8sc})$. The numerators and denominators of each of the resulting fractions have smaller polynomial degree, so that the complexity of finding a polynomial reduction for each of these fractions separately is reduced compared to the original expression.\\
Once the polynomial reduction is complete, we collect all terms in (\ref{8ptRedAnsatz}) on one side of the equation and demand the vanishing of all overall monomial coefficients separately. This gives us $375$ linear equations in the same number of unknowns. Solving these equations, we fix the unknown coefficients and obtain the effective standard form numerator polynomial $N^{\phi^3}_8$. With this, prescription (\ref{finalA}) is easily evaluated:
\begin{align}
A_8^{\phi^3}=&(-1)^55!~\text{Res}_{\sigma_{[8}=\infty,~\sigma_{7}=\infty,~\sigma_{6}=\infty,~\sigma_{5}=\infty,~\sigma_{4]}=\infty}\left[\frac{N^{\phi^3}_8}{\prod_{j=1}^{n-3} h_j}\right]=\frac{1360947997721}{22293435818142720000000000000}.\notag
\end{align}
This is an exact result since we did not invoke any floating point calculations at any step. Alternatively, we can approximately solve the scattering equations numerically and evaluate $A_8^{\phi^3}$ on the solutions, which yields agreement.
\section{CHY formulation of 1-loop level scattering amplitudes}
\label{sec:loop}
At one loop, $n$-point scattering equations have been shown to follow from $(n+2)$-point tree level scattering equations with two massive particles by taking the forward limit of the two massive momenta \cite{He:2015yua}. The tree level scattering equations with two massive particles are given by \cite{Naculich:2014naa,Naculich:2015zha}:
\begin{align}
\label{massSC}
&~~~E_a=\sum_{{b=1}\atop{b\neq a}}^{n+2}\frac{\mathfrak{p}_{a,b}}{\sigma_{ab}} ~~~\text{for}~~~a\in\{1,2,...,n\},\\
E_{n+1}=\sum_{b=1}^{n}\frac{\mathfrak{p}_{n+1,b}}{\sigma_{n+1,b}} &+\frac{\mathfrak{p}_{n+1,n+2}+m^2}{\sigma_{n+1,n+2}} ~~~,~~~E_{n+2}=\sum_{b=1}^{n}\frac{\mathfrak{p}_{n+2,b}}{\sigma_{n+2,b}} -\frac{\mathfrak{p}_{n+1,n+2}+m^2}{\sigma_{n+1,n+2}}, \notag
\end{align}
where two particles are massive with the same mass $k_{n+1}^2=k_{n+2}^2=m^2$. Here we have introduced a shorthand notation\footnote{When all momenta are massless and on-shell, we have $\mathfrak{p}_{\alpha(1),\alpha(2),...,\alpha(q)}=\mathfrak{s}_{\alpha(1),\alpha(2),...,\alpha(q)}$ from (\ref{kinS}).}
\begin{align}
\label{kinP}
\mathfrak{p}_{\alpha(1),\alpha(2),...,\alpha(q)}\equiv\sum_{\{\beta(1),\beta(2)\}\subset\{\alpha(1),\alpha(2),...,\alpha(q)\}}k_{\beta(1)}\cdot k_{\beta(2)}~~~~~~\text{for integer }q>1.
\end{align}
The sum is over all unordered subsets of two numbers out of a set of $q$ numbers. In the context of 1-loop CHY amplitudes, equations (\ref{massSC}) and (\ref{kinP}) also naturally arise from the formalism described in \cite{Cardona:2016bpi}, without the need to impose them.\footnote{The author thanks C. Cardona and H. Gomez for pointing this out.}\\
In the following we will require the scattering equations in polynomial form. To obtain them, we can for instance apply an appropriate transformation to (\ref{massSC}). However, we should proceed carefully, since in the forward limit
\begin{align}
k_{n+1}^\mu\to -l^\mu~~~,~~~k_{n+2}^\mu\to l^\mu
\end{align}
the set of equations (\ref{massSC}) admits singular solutions with $\sigma_{ij}\to 0$ for some $i\neq j$, if $E_{n+1}$ and $E_{n+2}$ are taken into consideration. Such singular solutions have no physical contribution to the amplitudes of relevant theories \cite{He:2015yua,Cachazo:2015aol}. Therefore, we will use $(n-1)$ independent equations $E_a$ with $a\leq n$ in order to exclude the singular solutions. It is straightforward to check that the transformation we are looking for is given by
\begin{align}
\label{massSCtrafo}
\tilde{h}_a^{p,q,v} =\sum_{{i=1}\atop{i\neq p,q,v}}^{n+2}\sigma_{ip}\sigma_{iq}\sigma_{iv}Y^{a-2}_{p,q,v,i}E_i~~~\text{for}~~~a\in\{2,3,...,n\},
\end{align}
where
\begin{align}
Y^x_{p,q,v,i}=\left\{\begin{matrix}
\sum_{\{\alpha(1),...,\alpha(x)\}\subset\{1,...,n+2\}/\{p,q,v,i\}}\prod_{j=1}^x\sigma_{\alpha(j)} & \text{ for }0<x\leq n-2,\\
1 & \text{for }x=0,\\
0 & \text{for $x<0$ and $x>n-2$}.
\end{matrix}\right.
\end{align}
The range in the index $a$ is set to correspond to (\ref{scatOrig}). Indices $p,q,v$ label the three different massive scattering equations (\ref{massSC}) that are dropped. As we expect, $\tilde{h}_a^{p,n+1,n+2}$ yields the same results regardless of the choice of $p$, so in the following we can consider $\tilde{h}_a^{1,n+1,n+2}$ for convenience. We can compactly write this result as
\begin{align}
\label{massSCpoly}
\tilde{h}_a^{1,n+1,n+2} = \sum_{\{\alpha(1),...,\alpha(a)\}\subset\{1,2,...,n+2\}}(\mathfrak{p}_{\alpha(1),...,\alpha(a)}+m^2\delta_{\alpha,\{n+1,n+2\}})\prod_{j=1}^a \sigma_{\alpha(j)}=0,
\end{align}
for integer $2\leq a \leq n$. Here we used a generalized Kronecker delta
\begin{align}
\delta_{\alpha,\{n+1,n+2\}}=\left\{{{1~~~\text{ if }~~~\{n+1,n+2\}\subset\{\alpha(1),...,\alpha(a)\}}\atop {0~~~\text{ if }~~~\{n+1,n+2\}\nsubset\{\alpha(1),...,\alpha(a)\}}}\right..
\end{align}
As long as we consider $\tilde{h}_a^{1,n+1,n+2}$ in the massive case before taking the forward limit, the scattering equations have the full set of $(n-1)!$ solutions. Knowing that the forward limit is singular in nature, we should check whether any singular solutions resurge in (\ref{massSCpoly}) due to the transformation (\ref{massSCtrafo}) having been applied. Indeed, if we choose to gauge fix $\sigma_1,\sigma_{n+1}$ and $\sigma_{n+2}$, it is straightforward to see that the trivial solution $\sigma_i=\sigma_1$ for $i=2,3,...,n$ is now present in the forward limit,\footnote{Setting $\sigma_i=\sigma_1$ for $i=2,3,...,n$ causes all scattering equations to be proportional to $\mathfrak{p}_{1,2,...,n}$, which vanishes in the forward limit.} additionally to the $(n-1)!-2(n-2)!$ expected regular solutions. Luckily, we can remove this trivial solution by fixing the gauge $\sigma_1=\infty$.\footnote{The fact that the trivial solution can be projected out by a gauge choice indicates that its contribution is not physical.} For convenience we will also fix $\sigma_{n+1}=0,~\sigma_{n+2}=1$. Thus, we will work with the following representation of gauge fixed polynomial scattering equations with two massive particles
\begin{align}
\label{scatL}
h_i\equiv\left(\lim_{\sigma_1\rightarrow \infty}\frac{1}{\sigma_1}\tilde{h}_{i+1}^{1,n+1,n+2}\right)|_{{\sigma_{n+1}=0}\atop{\sigma_{n+2}=1}}=0~~~,~~~\forall i\in\{1,2,...,n-1\},
\end{align}
which has a smooth forward limit containing only regular solutions of interest.\footnote{We use the same symbol $ h$ as for tree level scattering equations here, since it is always clear from context which scattering equations are in use.} It will be convenient to treat the forward limit as a regulator whenever the kinematics in the limit becomes singular.\\
For $\tilde{h}_a^{1,n+1,n+2}$ the transformation Jacobian is $(-1)^{n+1}[\prod_{i=2}^n\sigma_{1i}][\prod_{1<j<q\leq n+2}^{j\leq n}\sigma_{jq}]$. Therefore, possibly up to a minus sign we have the usual CHY measure for polynomial scattering equations
\begin{align}
d\mu=\left(\prod_{{c=1}\atop{c\neq q,p,w}}^{n+2} d\sigma_c\right)(\sigma_{qp}\sigma_{pw}\sigma_{wq})\left(\prod_{1\leq i<j\leq n+2}\sigma_{ij}\right)\left(\prod_{a=2}^{n}\delta\left(\tilde{h}_a^{1,n+1,n+2}\right)\right).
\end{align}
Recall that we gauge fixed the moduli $q=1,~p=n+1,~w=n+2$. To test our evaluation procedure at one-loop level, we will consider the bi-adjoint scalar $\phi^3$ theory as proposed in \cite{He:2015yua}, which can be written as
\begin{align}
A_n^{1-loop,\phi^3}=\int\frac{d^Dl}{(2\pi)^D}\frac{1}{l^2}\lim_{{k_{n+1}\to -l}\atop{k_{n+2}\to ~l}}\int d\mu\left(\sum_{\gamma\in\text{cyclic}\{1,2,...,n\}}PT(n+2,\gamma,n+1)\right)^2,
\end{align}
where
\begin{align}
PT(n+2,\gamma,n+1)=\frac{1}{\sigma_{n+2,\gamma(1)}\sigma_{\gamma(1),\gamma(2)}...\sigma_{\gamma(n),n+1}\sigma_{n+1,n+2}}.
\end{align}
However, our evaluation method applies more generally to any integrand that is rational in $\sigma$-moduli and is being integrated over the measure $d\mu$.
\subsection{One-loop amplitude examples}
\label{sec:loopexpl}
\subsubsection{Two point 1-loop scalar amplitude}
At two points and 1-loop there is one scattering equation, given by\footnote{Since the forward limit makes the kinematics singular, we use it as a parametrization.}
\begin{align}
h_1=\sigma _2 \mathfrak{p}_{1,2}+\mathfrak{p}_{2,3}=0.
\end{align}
The gauge fixed amplitude amounts to
\begin{align}
A_2^{1-loop,\phi^3}=\int\frac{d^Dl}{(2\pi)^D}\frac{1}{l^2}\lim_{{k_{3}\to -l}\atop{k_{4}\to ~l}}\oint \frac{d\sigma_2}{ h_1}\frac{1}{\left(1-\sigma _2\right) \sigma _2}.
\end{align}
We require a standard form numerator polynomial $N_{1-\text{loop}}^{2,\phi^3}$ such that $1\hateq \left(1-\sigma _2\right) \sigma _2 N_{1-\text{loop}}^{2,\phi^3}$ with the standard form Ansatz $N_{1-\text{loop}}^{2,\phi^3}=c_1$. Making use of the scattering equation, we polynomially reduce the right hand side, collect all terms on one side of the equation and in doing so obtain one linear equation in one unknown. Solving this equation and applying momentum conservation yields:
\begin{align}
N_{1-\text{loop}}^{2,\phi^3}=\frac{\mathfrak{p}_{1,2}^2}{\mathfrak{p}_{2,3}\mathfrak{p}_{2,4}}.
\end{align}
Prescription (\ref{finalA}) suggests the calculation
\begin{align}
(-1)^1(1!)\text{Res}_{\sigma_2=\infty}\left[\frac{1}{ h_1}\frac{\mathfrak{p}_{1,2}^2}{\mathfrak{p}_{2,3} \mathfrak{p}_{2,4}}\right]=\frac{\mathfrak{p}_{1,2}}{\mathfrak{p}_{2,3} \mathfrak{p}_{2,4}}.
\end{align}
If we solve the scattering equation instead $\sigma_2=-\frac{\mathfrak{p}_{2,3}}{\mathfrak{p}_{1,2}}$, we get exactly the same result
\begin{align}
\sum_{{ h=0}\atop{\text{solutions}}}\frac{1}{\det\left([\partial_i h_j]\right)}\frac{1}{\left(1-\sigma _2\right) \sigma _2}=\frac{\mathfrak{p}_{1,2}}{\mathfrak{p}_{2,3} \mathfrak{p}_{2,4}}.
\end{align}
In the forward limit we have $\mathfrak{p}_{1,2}\to 0$ while $\mathfrak{p}_{2,3}$ and $\mathfrak{p}_{2,4}$ stay finite. Therefore, the 1-loop integrand vanishes.
\subsubsection{Three point 1-loop scalar amplitude}
At three points and 1-loop there are two scattering equations, given by
\begin{align*}
h_1&=\sigma _2 \mathfrak{p}_{1,2}+\sigma _3 \mathfrak{p}_{1,3}+\mathfrak{p}_{1,5}=0,\\
h_2&=\sigma _3 \mathfrak{p}_{2,4}+\sigma _2 \mathfrak{p}_{3,4}+\sigma _2 \sigma _3 \mathfrak{p}_{4,5}=0.
\end{align*}
The gauge fixed amplitude can be written as
\begin{align*}
A_3^{1-loop,\phi^3}=\int\frac{d^Dl}{(2\pi)^D}\frac{1}{l^2}\lim_{{k_{4}\to -l}\atop{k_{5}\to ~l}}\oint \frac{d\sigma_2d\sigma_3}{ h_1 h_2}\frac{-\left(\sigma _2^2+\sigma _3^2-\left(\sigma _2+1\right) \sigma _3\right){}^2}{\left(1-\sigma _2\right) \sigma _2 \left(1-\sigma _3\right)
\left(\sigma _2-\sigma _3\right) \sigma _3}.
\end{align*}
Therefore, we consider the following equality in order to find a standard form effective numerator polynomial $N_{1-\text{loop}}^{3,\phi^3}$
\begin{align*}
-\left(\sigma _2^2+\sigma _3^2-\left(\sigma _2+1\right) \sigma _3\right){}^2\hateq \left(1-\sigma _2\right) \sigma _2 \left(1-\sigma _3\right)
\left(\sigma _2-\sigma _3\right) \sigma _3 N_{1-\text{loop}}^{3,\phi^3},
\end{align*}
with the standard form Ansatz $N_{1-\text{loop}}^{3,\phi^3}=c_1 \sigma_2+c_2\sigma_3$. We apply the reduction procedure of section \ref{app:degred} to both sides of this equation, collect all terms on one side and demand that the overall coefficient in front of each monomial vanishes separately. This sets up two linear equations in two unknowns $c_1,~c_2$. Solving for the unknowns yields a numerator polynomial $N_{1-\text{loop}}^{3,\phi^3}$. Using prescription (\ref{finalA}) and simplifying via five-point momentum conservation and on-shell conditions with two massive particles we get the result
\begin{align}
\label{loopPhi33}
&(-1)^22!\,\text{Res}_{\sigma_{[3}=\infty,~\sigma_{2]}=\infty}\left[\frac{N_{1-\text{loop}}^{3,\phi^3}}{ h_1 h_2}\right]=\\
&=-\frac{1}{\mathfrak{p}_{1,2}}\left(\frac{1}{\mathfrak{p}_{3,5}}+\frac{1}{\mathfrak{p
}_{3,4}}\right)-\frac{1}{\mathfrak{p}_{2,3}}\left(\frac{1}{\mathfrak{p}_{1,5}}+\frac{1}{\mathfrak{p}_{1,4}}\right)-\frac{1}{\mathfrak{p}_{1,3}}\left(\frac{1}{\mathfrak{p}_{2,5}}+\frac{1}{\mathfrak{p}_{2,4}}\right)-\frac{1}{\mathfrak{p}_{1,5} \mathfrak{p}_{2,4}}-\frac{1}{\mathfrak{p}_{2,5} \mathfrak{p}_{3,4}}-\frac{1}{\mathfrak{p}_{1,4} \mathfrak{p}_{3,5}}.\notag
\end{align}
Alternatively, we can solve the scattering equations and obtain the two solutions $(\sigma_{2,+},\sigma_{3,+})$ and $(\sigma_{2,-},\sigma_{3,-})$ with
\begin{align}
\sigma_{2,\pm}&=\frac{\mathfrak{p}_{1,3} \mathfrak{p}_{3,4}-\mathfrak{p}_{1,5} \mathfrak{p}_{4,5}}{2
\mathfrak{p}_{1,2} \mathfrak{p}_{4,5}}-\frac{\mathfrak{p}_{2,4}}{2
\mathfrak{p}_{4,5}}\pm \frac{\sqrt{\left(\mathfrak{p}_{1,2}
\mathfrak{p}_{2,4}-\mathfrak{p}_{1,3} \mathfrak{p}_{3,4}+\mathfrak{p}_{1,5}
\mathfrak{p}_{4,5}\right){}^2-4 \mathfrak{p}_{1,2} \mathfrak{p}_{1,5} \mathfrak{p}_{2,4}
\mathfrak{p}_{4,5}}}{2 \mathfrak{p}_{1,2} \mathfrak{p}_{4,5}},\notag\\
\sigma_{3,\pm}&=\frac{\mathfrak{p}_{1,2} \mathfrak{p}_{2,4}-\mathfrak{p}_{1,5} \mathfrak{p}_{4,5}}{2
\mathfrak{p}_{1,3} \mathfrak{p}_{4,5}}-\frac{\mathfrak{p}_{3,4}}{2
\mathfrak{p}_{4,5}}\mp \frac{\sqrt{\left(\mathfrak{p}_{1,2}
\mathfrak{p}_{2,4}-\mathfrak{p}_{1,3} \mathfrak{p}_{3,4}+\mathfrak{p}_{1,5}
\mathfrak{p}_{4,5}\right){}^2-4 \mathfrak{p}_{1,2} \mathfrak{p}_{1,5} \mathfrak{p}_{2,4}
\mathfrak{p}_{4,5}}}{2 \mathfrak{p}_{1,3} \mathfrak{p}_{4,5}}.\notag
\end{align}
Evaluating the integral on these solutions, summing the contributions and simplifying by means of momentum conservation and on-shell conditions directly leads to exactly the same result (\ref{loopPhi33}).\\
In the forward limit, terms $\mathfrak{p}_{i,4},\mathfrak{p}_{i,5}$ with $i\in{1,2,3}$ stay finite while $\mathfrak{p}_{i,j}$ with $i,j\in\{1,2,3\}$ tend to zero. Therefore, we first rewrite each of the three different terms in parenthesis in (\ref{loopPhi33}) analogously to the following
\begin{align}
-\frac{1}{\mathfrak{p}_{1,2}} \left(\frac{1}{\mathfrak{p}_{3,5}}+\frac{1}{\mathfrak{p}_{3,4}}\right)=-\frac{1}{\mathfrak{p}_{3,4} \mathfrak{p}_{3,5}}\left(\frac{\mathfrak{p}_{3,4}+\mathfrak{p}_{3,5}}{\mathfrak{p}_{3,4}+\mathfrak{p}_{3,5}+\frac{1}{2}(k_4+k_5)^2}\right).
\end{align}
We may parametrize the forward limit as $k_4^\mu=-(l^\mu+\tau q^\mu_4)$ and $k_5^\mu=(l^\mu+\tau q^\mu_5)$ with $\tau\to0$ and finite $q^\mu_4\neq q^\mu_5$. With this, at leading order we find
\begin{align}
\frac{1}{\mathfrak{p}_{3,l+\tau q_4} \mathfrak{p}_{3,l+\tau q_5}}\left(\frac{\tau\mathfrak{p}_{3,q_5}-\tau\mathfrak{p}_{3,q_4}}{\tau\mathfrak{p}_{3,q_5}-\tau\mathfrak{p}_{3,q_4}+\tau^2\frac{1}{2}(q_5-q_4)^2}\right)=\frac{1}{(\mathfrak{p}_{3,l})^2}+O(\tau).
\end{align}
Therefore, the one-loop integrand at three points in bi-adjoint scalar $\phi^3$ theory is given by
\begin{align}
A_3^{1-loop,\phi^3}&=\int\frac{d^Dl}{(2\pi)^D}\frac{1}{l^2}\left( \frac{1}{\mathfrak{p}
_{1,l} \mathfrak{p}_{2,l}}+\frac{1}{\mathfrak{p}_{1,l}
\mathfrak{p}_{3,l}}+\frac{1}{\mathfrak{p}_{2,l}
\mathfrak{p}_{3,l}}+\frac{1}{\mathfrak{p}_{1,l}^2}+\frac{1}{\mathfrak{p}_{2,l}^2}+\frac{1}{\mathfrak{p}_{3,l}^2}\right)\\
&=\int\frac{d^Dl}{(2\pi)^D}\frac{1}{l^2}\left(\frac{1}{\mathfrak{p}_{1,l}^2}+\frac{1}{\mathfrak{p}_{2,l}^2}+\frac{1}{\mathfrak{p}_{3,l}^2}\right),
\end{align}
since the first three terms vanish by three-point momentum conservation. Since we might be interested in the 1-loop 3-point amplitude as a vertex correction, it would make sense to consider the momenta $k_1,k_2,k_3$ to be off-shell $-$ then the above result is non-trivial. In case when $k_1,k_2,k_3$ are on-shell, all appearing integrals are scaleless.
\section{Conclusion and outlook}
\label{sec:conclusion}
In this work we started with the CHY formulation of scattering amplitudes in arbitrary dimension. We then developed the degree reduction procedure of section \ref{app:degred} and applied it alongside the strong Nullstellensatz to show that any rational function can be written as a standard form polynomial on the support of scattering equations. Making use of this conversion for CHY amplitude integrands, we derived an evaluation prescription that allows to find an amplitude purely from collecting consecutive simple residues at infinity only.\\
Summing over all possible ladder type shapes and taking into account the multiplicity due to available subsets of non-gauge fixed moduli that are used to compose the shapes, we realize that the total number of different ladder type monomials at any $n$ is given by $N^{\text{ladd}}_n=s(n-3)$, where the function $s(x)$ is
\begin{align*}
s(0)=1,~~~~~s(x)=\sum_{i=0}^{x-1}\binom{x}{i}s(i).
\end{align*}
Upon inspection, the $s(x)$ turn out to be equivalent to so called ordered Bell numbers, or Fubini numbers. For large $x$ these numbers asymptote to $x s(x-1)\approx \ln(2)s(x)$, so that the number of ladder type monomials grows quicker than factorially with $n$.\\
In all explicit amplitude examples we studied above, it was sufficient to consider the subset of ladder type monomials with highest base length $m_{max}=n-4$ to find standard form polynomials corresponding to relevant rational functions. By the counting above, at any $n$ there are $N^{\text{ladd}}_{m_{max}=n-4}=(n-3)s(n-4)$ such ladder type monomials.\\
It is well known that gauge fixed scattering equations have $(n-3)!$ different solutions at tree level \cite{Cachazo:2013gna,Cachazo:2013hca}. In \cite{Cardona:2015ouc,Dolan:2015iln} it was shown that gauge fixed polynomial scattering equations can be transformed to a different form such that $\sigma_i-P_i(\sigma_n)=0$ for $i\in\{4,5,...,n-1\}$ and $P_n(\sigma_n)=0$, where the $P_i(\sigma_n)$ are univariate polynomials in $\sigma_n$. The polynomial $P_n(\sigma_n)$ is of highest degree $(n-3)!$ and accomodates the $(n-3)!$ different solutions. Reducing multivariate polynomials over this transformed system of equations trivially leaves $(n-3)!$ univariate monomials (i.e. $1,\sigma_n,\sigma_n^2,...,\sigma_n^{(n-3)!-1}$) as a minimal basis for the quotient ring of multivariate polynomials over the ideal spanned by scattering equations $Q=R/\langle h_1,h_2,...,h_{n-3}\rangle$. Therefore, the dimension of the quotient ring is $\dim_R(Q)=(n-3)!$ and thus, in the present case, we can similarly expect only $(n-3)!$ of ladder type monomials to be linearly independent on the support of the ideal spanned by scattering equations $\langle h_1,h_2,...,h_{n-3}\rangle$. Here, a natural candidate for such a minimal basis would be the $(n-3)!$ highest degree ladder type monomials. At first glance it might seem that restricting to this minimal basis could increase computational efficiency, since this sets up a minimal linear system of equations in the polynomial construction of rational terms and makes the resulting coefficients unique. However, on a second thought it becomes apparent that modifying the polynomial reduction algorithm such as to eliminate the tail of lower degree ladder type monomials is highly non-trivial and would introduce a large computational overhead before the linear system of equations is set up. Therefore, employing more than the minimal amount of ladder type monomials to keep polynomial reduction simple at the expense of working with larger linear systems of equations appears to be more convenient.
\\
One nice feature of the above procedure is that it works in exactly the same fashion at any $n$ and for amplitudes of any theory in CHY formulation due to the inherent structure of CHY integrands: While the complexity of the kinematic part of a CHY amplitude integrand in a theory like i.e. pure Yang-Mills or gravity is greater compared to massless scalars, the integrand still always is a rational function in the $\sigma$-moduli, such that the conceptual steps towards finding the amplitude described in previous sections still remain exactly the same, making the procedure universal. Furthermore, since all relevant residues for any amplitude or partial term in consideration are always collected from simple poles at infinity only, each generic evaluation step is of low complexity and the difficulty is shifted towards finding standard form polynomial expressions for the originally rational amplitude integrands. The polynomial reduction procedure that addresses this problem can be implemented algorithmically in general, so that the amplitude evaluation becomes automated for general input, which is one further strength of the current approach.\\
One problem that is bound to appear as we choose higher values for $n$, is the question of efficiency. The number of linear equations and corresponding number of unknowns increases as $(n-3)s(n-4)$ if we apply the construction step of section \ref{sec:polinv}. Even though other techniques to find the reduced form might exist, this kind of limitation is bound to appear whenever a solution is formulated algorithmically involving a sequence of structural steps leading from a certain input to an output of a different structure. Therefore, as a possible direction for further investigation it might be interesting to search for general $n$-point integrands of standard polynomial form in various theories of interest directly, eliminating the necessity for the polynomial reduction procedure. Additionally, knowing that only the highest degree ladder type monomials contribute to any integral, finding just the coefficients for the minimal basis of highest degree ladder type monomials based on some general physical arguments would be equivalent to obtaining a direct closed form expression for the amplitude, since the remaining contour integration is trivial.
|
\section{Introduction}
Consider the following problem: suppose $\rho \in L_1(\Omega)$ is an unknown function, given observations of the form
$\int \varphi_i^a \rho \, d x = m_i^a$, where $\{ \varphi_i \}_{i=1}^N$ is a set of independent functions with the property of partition of unity(POU),
i.e., $\sum_{i=1}^N\varphi_i^a=1$, we want to estimate the value
$m_i^b=\int \varphi_j^b \rho \, dx$ with respect to a different set of independent functions $\{ \varphi_j^b\}_i^M$.
If we require $f$ to have a better regularity, then the classical pointwise interpolation problem is just a special case of such a formulation by setting $\varphi_i^a=\delta(x-x_i)$ and $\varphi_j^b=\delta(x-y_j)$ for two different point sets $\{ x_i\}_{i=1}^N$ and $\{ y_j\}_{j=1}^{M}$, where $\delta(\cdot)$ is the Dirac delta function.
In this paper, we focus on the case that $\varphi_i^a=\chi(T_i^a)$ is the characteristic function for $T_i^a$ and $T_i^a$ is an element of a tessellation/mesh $\T^a=\{T_1^a,\ldots,T_N^a \}$ of $\Omega \subset R^d$. Similarly, $\varphi_i^b=\chi(T_i^b)$ is the characteristic function on $T_i^b$, and $T_i^b $ is an element of another tessellation $ \T^b= \{ T_1^b, \ldots, T_M^b\}$ of $\Omega$. Such a generalised interpolation problem arises when it is required to transfer information from the old tessellation to a new one. For example, in the Arbitrary Lagrangian-Eulerian(ALE) method \cite{H74,L05}, or a moving mesh method, when the old tessellation becomes poor or severely distorted, one has to rezoning the tessellation and to map physical data on an old tessellation to a new one. The process of transfer the physical variables based on the observations $m_i^a=\int \varphi_i^a \rho \,dx$, $i=1,\ldots,N$, on an old mesh $\T^a$ to $m_i^b=\int \varphi_i^b \rho dx$, $i=1,\ldots,M$, on a new mesh $\T^b$, is referred to as \textit{remapping}.
As one of the three main steps in the ALE framework, remapping has kept researchers busy in the past decades. Currently there are two main streams of remapping schemes: one is the \textit{Cell-Intersection-Based Donor Cell(CIB/DC)}. Such an approach is conceptually simple; the information $\int \varphi_i^b \rho dx $ is aggregated from intersections between old cells and the new cells:
\begin{linenomath}
\begin{equation}
\int \varphi_i^b \rho dx = \sum_{j}\int_{\sup \varphi_i^b \cap \sup \varphi_j^a} \varphi_j^a \rho dx. \label{eq:ai}
\end{equation}
\end{linenomath}
Once this information is aggregated exactly, then the conservative property of the generalised interpolation follows due to the property of partition of unity:
\begin{linenomath}
$$
\sum_{i}^N \int_\Omega \varphi_i^a \rho dx = \int \sum_{i=1}^N \varphi_i^a \rho dx = \int_{\Omega} \rho dx =\int_{\Omega } \sum_{j=1}^M \varphi_j^b \rho dx = \sum_{j=1}^M \int_\Omega \varphi_j^b \rho dx.
$$
\end{linenomath}
Such an approach based on set or support intersection operation is also referred to as \textit{aggregated intersection Based Donor Cell(AIB/DC)} method.
The remapping error depends only on the approximation/reconstruction of the density function on the old cells whenever the integral over the overlays between the old and new cells is computed exactly. Precisely, piecewise constant reconstruct of the density function on the old cell grantees first order accuracy and convergence and piecewise linear construction grantees second order accuracy according to the standard approximation theory. Dukowicz in \cite{D84} considered applying the Gauss divergence theory to calculate the overlap(s) between an old cell and a new one. His procedure requires to transform a physical general quadrilateral meshes to the standard regular grid. Ramshaw in \cite{Ramshaw1985,Ramshaw1986} improves Dukowicz' result, the improved method works on the physical mesh directly without any coordinate transform. A first order three dimensional remapping scheme based on intersecting arbitrary polyhedra was consider by Grandy \cite{grandy1999} and Azarenok \cite{Azarenok2009}. These approaches are first order methods. Dukowicz pointed out that the first order CIB/DC based remmapping scheme is equivalent to upwind difference of the advective term of a continuous equation \cite[p.412]{D84} \footnote{It originally reads "It can be easily shown that the continuous rezone with constant cell density corresponds to donor cell(or upwind) differencing of the advective terms" }. In \cite{dukowicz2000}, Dukowicz and Baumgardner articulated this equivalence. Second order CIB/DC methods were constructed by using piecewise linear reconstruction of the density function on the old cells \cite{dukowicz1987accurate}. When the mass in a new cell are calculated by mass transfer from an old cell through the overlapped regions, the CIB/DC approach is also referred to as the \textit{flux-intersection-based} approach (generalised flux in fact). This approach surfers from programming difficulties as we shall demonstrate, but it works for all kinds of tessellations. In particular, it is still used when \textit{mixed cells}(cells contains several materials) appear \cite{B11}.
The other method is the \textit{Faced-Based Donor-Cell(FB/DC) method} or the \textit{swept-region based methods}. In such an approach, the mass transfer between \textit{a home/departing cell} and a \textit{target/arriving cell} is calculated by surface integral(line integral in $R^2$) according to the Green formula or the divergence theory. The transfer is based on the \textit{fluxing area or swept region} between the arriving/target cell and neighbouring cells of the home cell. Dukowicz and Baumgardner shows how the mass exchanges between a home cell and a target cell \cite{dukowicz2000}. In such an approach, there is flexibility to approximate the the transferred mass, the mass in the swept region is not necessarily calculated exactly, simplification and neat approximation to the flux bring efficient and high order methods, see \cite{L11}\cite{margolin03} for details. And it is easier to generalize into three dimensional space compared with the CIB/DC approach. See \cite{Garimella2007} for example. Such an approach is also referred to the \textit{flux-based/face-based} method. It is valid when the Lagrangian and rezoned mesh have the same connectivity and are close to each other.
The FB/DC approach is simpler for the CIB/DC methods due to the \textit{degeneracy of the intersections}. Increasing the robustness of a CIB/DC methods requires a special procedure to handle such degeneracies. Some authors believe the CIB/DC approach is very complicated. This argument is basically true but is mathematical imprecise, because as we shall see similar degeneracy also arise for the FB/DC method. For both method, neglecting such degeneracy will result flawed procedure. The ability to handle such degeneracy decides the robustness of a remapping procedure.
While the techniques vary between the two class methods, the essence of these mapping schemes is to compute or approximate the transfer of mass between a \textit{home cell or departing cell} in the old tessellation and a \textit{target cell or arriving cell} in the new tessellation. In the CIB/DC approach, this mass transfer is equivalent to computing the integration/quadrature on the intersections. Calculating the cell intersections and a proper quadrature rule (in terms of volumetric or surface integral) severs as the mathematical foundation of this approach. In the FB/DC method, calculating the \textit{fluxing/swept area} between the old faces and a good reconstruction on the faces and vertices is critical. Since the density function $\rho$ is usually approximated by piecewise functions on the old tessellation. Lower regularity appears on the faces of the old tessellation. Therefore, accurate computing the remapped mass or other physical variables requires summarize the underlying quantity piecewisely on the overlapped regions in the old tessellation to avoid singularity. That is why an exact remapping scheme requires to calculate the intersections or the swept/fluxing area between the old and new mesh. We shall mention that calculating or approximation the overlapped regions between the old and new tessellations and a reconstruction scheme (for the density function or flux) are two aspects of an remapping scheme. There are intensive discussion on reconstruction scheme, see \cite{Cheng2008,ms03,Z11} and reference therein, while there are relative sparse discussion on how to calculate the overlapped regions except a couple of noticeable examples \cite{Azarenok2009}, \cite{grandy1999} and \cite{dukowicz2000}.
The aim of this paper is to articulate the mathematical formulation of these problems, present detailed classification of intersection types between a quadrilateral and another, and provide simple method to calculate the intersections between two quadrilateral mesh of the same connectivity based on the classification. The classification is simper and clearer than that in \cite[p.327, Table I]{dukowicz2000}. According the classification, one can quickly identify a degeneracy when it may arise. This helps to develop a robust software. When there is non degeneracy of the overlapped region, we demonstrate that the present approach only requires 34 programming cases(17 when considering symmetry), while the CIB/DC approach used as in \cite{Ramshaw1985,Ramshaw1986} requires 98 programming cases(Neither Dukowicz nor Ramshaw gave this quantitative result, the number is derived from Fig.~\ref{fig:cswap}, Fig.~\ref{fig:branch} in this paper and Fig.1 in \cite{Ramshaw1986}). When consider degeneracy, more benefit can be obtained. We shall demonstrates how the degeneracy of the intersection depends the so called \textit{singular intersection points}, and how the computational amount depends on the singular intersection points. As far as we known this is the first result on how the computational complexity depends on an underlying problem.
The paper is enlightened by the fact that calculating the intersection of a line and a polygon only takes $O(\log(p))$ complexity \cite[p.2, Fig.1]{CD87}, where $p$ is the number of vertices of a polygon. When considering how a continues piecewise linear curve intersects with a mesh, the computations can be further reduced. In either the CIB/DC approach or the FB/DC approach, we will reduce cell intersections or the swept region to a case that an edge/face intersects with a \textit{local frame}(see Definition \ref{def:lf}). By focusing on how an edge intersects with old cells and exploring the structure of the mesh, the intersection area between the new and old tessellation, and the union of the fluxing/swept area can be traveled in $\mathcal{O}(n)$ time (Theorem \ref{thm:1}),where $n$ is the number of elements in the underlying mesh or tessellation. Our starting point is a direct cell-intersection formula based on purely set operation. It is conceptually simple and serves as the basis for the proposed algorithms. In parallel, we shall show the basic formula for the FB/DC method based on the \textit{fluxing/swept region}. It is derived from the Green formula. Here are some preliminary for presenting our results.
\
\section{Preliminary}
A \textit{tessellation} $\T=\{T_1,T_2,\ldots,T_N\}$ of a domain $\Omega$ is a partition of $\Omega$ such that $\bar{\Omega} =\cup_{i=1}^N T_i$ and $\dot{T}_i \cap \dot{T}_j =\emptyset $, where $\dot{T}_i$ is the interior of the \textit{element} or \text{cell} $T_j$ and $\bar{\Omega}$ is the closure of $\Omega$.
A \textit{admissible} quadrilateral tessellation has no hanging node on any edge of the tessellation. Precisely, if $T_i\cap T_j \neq \emptyset$, then $T_i $ and $T_j$ can only share a common vertex or a common edge. We shall also refer to such a tessellation as a quadrilateral mesh.
A admissible quadrilateral tessellation can be indexed only by its vertices. We use the conversional notation as follows. The first three terms are identical to that used in \cite{margolin03}
\begin{itemize}
\item $P_{i,j}$, for $i=1:M, j=1:N$ are the vertices;
\item $F_{i+ \frac{1}{2},j }$ for $i=1:M-1,j=1:N$ and $F_{i, j+\frac{1}{2}}$ for $i=1:M,j=1:N-1$ are the edges with vertices $P_{i,j}$ and $P_{i+1,j}$;
\item $C_{i+\frac{1}{2},j+\frac{1}{2}}$ stands for the quadrilateral cell $P_{i,j} P_{i,j+1} P_{i+1,j+1}P_{i,j+1}$ for $1\le i \le M-1$ and $1\le j \le N+1$. $C_{i+\frac{1}{2},j+\frac{1}{2}}$ is an element of $\T$, we shall also referred to it as $T_{i,j}$ with integer subindex for convenience;
\item $x_i(t)$ is the piecewise curve which consists all the face of $F_{i,j+\frac{1}{2}}$ for $j=1:N-1$, $y_j(t)$ is the piece wise curve which consists all the faces of $F_{i+\frac{1}{2},j}$ for $i=1:M-1$.
\end{itemize}
Let $\T^a$ and $\T^b$ be two admissible quadrilateral mesh. The vertices of $\T^a$ and $\T^b$ are denoted as $P_{i,j}$ and $Q_{i,j}$ respectively, if there is a one-to-one map between $P_{i,j}$ and $Q_{ij}$, we shall say that the two tessellation share the same \textit{logical structure} or \textit{connectivity}.
Here we shall also assume the two admissible of the same logical structure of the same domain have the following property:
\begin{ass}[A1]
For the two admissible quadrilateral mesh $\T^a$ and $\T^b$, each vertex $Q_{ij}$ of $\T^b$ can only lie in the interior of the union of the cells $C^a_{i\pm \frac{1}{2},j\pm \frac{1}{2} }$ of $\T^a$.
\label{ass:1}
\end{ass}
\begin{ass}[A2]
For the two admissible quadrilateral mesh $\T^a$ and $\T^b$, each curve $x_i^b(t)$ has at most one intersection point with $y_j^a(t)$, so does for $y_j^b(t)$ and $x^a_i(t)$.
\label{ass:2}
\end{ass}
\begin{ass}[A3]
The intersection point of an new face and an old face lies in the middle of the two faces.
\end{ass}
For convenience, we also introduce the following notation and definition. They will be frequently refereed to in the remaining of this paper.
\begin{definition}
A local patch $\LP_{i,j}$ of an admissible quadrilateral mesh consists an element $ T_{i,j}$ and its neighbours in $\T$. Take an interior element of $T_{i,j}$ as example,
\begin{linenomath}
\begin{equation}
\LP_{i,j}:=\{ T_{i,j}, T_{i\pm1,j}, T_{i,j\pm1}, T_{i\pm1,j\pm1} \}.
\end{equation}
\end{linenomath}
The index set of $\LP_{i,j}$ are denoted as
\begin{linenomath}
\begin{equation}
\mathcal{J}_{i,j}=\{ (k,s): T_{k,s} \in \LP_{i,j} \}.
\end{equation}
\end{linenomath}
\end{definition}
The local patch is similar to the notation $ \mathscr{C_i}= \cup_k C_k$ such that $\tilde{C}_i \in \mathscr{C}(C_i)$ in previous publication like \cite[eq.2.1]{ms03} and \cite[eq.1]{B11}, which is the smallest patch in the old tessellation which contains $C_i$, and it depends on the how a new cell intersect with the old cells. Here, the local patch is a fixed patch and is independent of how the cell intersects. Under the Assumption~\ref{ass:1}, $\mathscr{C}(T_{i,j}) \subset \LP_{i,j}$. The introduce of a local patch for quadrilateral mesh is for programming simplicity.
Let $\T^a$ and $\T^b$ be two admissible quadrilateral mesh of the same domain with the Assumption~\ref{ass:1}.
\begin{definition}
An \textit{invading set} of the element $T_{i,j}^b$ with respect to the local patch $\LP^a_{i,j}$ is defined as
\begin{linenomath}
\begin{equation}
\mathcal{I}_{i,j}^b=(T^b_{i,j} \cap \LP^a_{i,j}) \backslash (T^b_{i,j}\cap T^a_{i,j}). \label{eq:I}
\end{equation}
\end{linenomath}
\end{definition}
\begin{definition}
An \textit{occupied set} of the element $T^a_{i,j}$ with respect to the local patch $\LP^b_{i,j}$ is defined as
\begin{linenomath}
\begin{equation}
\Oc_{i,j}^a=(T^a_{i,j} \cap \LP^b_{i,j}) \backslash (T^a_{i,j}\cap T^b_{i,j}). \label{eq:O}
\end{equation}
\end{linenomath}
\end{definition}
\begin{definition}
A \textit{swept area or fluxing area} is the area which is enclosed by a quadrilateral polygon with an edge in $\T^a$ and its counterpart edge in $\T^b$. For example, the swept area enclosed by the quadrilateral polygon with edges $F^a_{i+\frac{1}{2},j}$ and $F^b_{i+\frac{1}{2},j}$ will be denoted as $\partial F_{i+\frac{1}{2},j}$. $\partial^{b+} F_{k,s}$ stands for boundary of the fluxing/swept area $\partial F_{k,s}$ is ordered such that direction of edges of the cell $T^b_{i,j}$ are counterclockwise in the cell $T^b_{i,j}$. Precisely
\begin{linenomath}
\begin{align*}
\partial^{b+}F_{i+\frac{1}{2},j} &=Q_{i,j}Q_{i,j+1}P_{i,j+1}P_{i,j}, \\
\partial^{b+} F_{i+1, j+\frac{1}{2}}&=Q_{i+1,j}Q_{i+1,j+1}P_{i+1,j+1}P_{i+1,j}, \\
\partial^{b+}F_{i+\frac{1}{2},j+1} &=Q_{i+1,j+1}Q_{i,j+1}P_{i+1,j+1}P_{i,j+1}, \\
\partial^{b+} F_{i, j+\frac{1}{2}} &=Q_{i,j+1}Q_{i,j}P_{i,j}P_{i,j+1}.
\end{align*}
\end{linenomath}
where ${Q_{i,j}Q_{i+1,j}Q_{i+1,j+1}Q_{i,j+1}}$ are list in the counterclockwise order.
\end{definition}
The invading set $\I^b_{i,j}$ has no interior intersection with the occupied set $\Oc^a_{i,j}$, while the fluxing areas associated to two connect edges of can be overlapped.
The invading and occupied sets consist of the whole intersection between an old cell and a new cell, while corners of a fluxing area may only be part of an intersection between an old cell and a new cell.
Fig.~\ref{fig:IO}, Fig.~\ref{fig:swepta} and Fig.~\ref{fig:swept2} illustrate such differences. In a local patch, the union of the occupied and invading set is a subset of the union of the swept/fluxing area of a home cell $T^a_{i,j}$. However, the difference (extra corner area), if any, will be self-canceled when summing all the \textit{signed} fluxing area, see the north-west and south-east color region in Fig.~\ref{fig:swepta} and Fig.~\ref{fig:swept2}.
\begin{figure}[!t]
\centering
\subfloat[\label{fig:IO}]{\includegraphics[width=0.22\textwidth]{def_areas_01}}
\subfloat[\label{fig:swepta}]{\includegraphics[width=0.22\textwidth]{def_areas_02}}
\subfloat[\label{fig:swept2}]{\includegraphics[width=0.22\textwidth]{def_areas_03}}
\subfloat[\label{fig:swap}]{\includegraphics[width=0.22\textwidth]{def_areas_04}}
\caption{(a) the invading set $\I^b_{i,j}$ (red) and occupied set $\Oc^a_{i,j}$ (b) the swept/fluxing area $\partial F_{i\pm \frac{1}{2},j}$ (c) The swept/fluxing area $\partial F_{i,j\pm \frac{1}{2}}$ (d) the swept area (left) v.s. the local swap set(right) in a local frame.}
\label{fig:IOswept}
\end{figure}
\begin{definition}
A \textit{local frame} consist an edge and its neighbouring edges. Taking the edge $F_{i,j+\frac{1}{2}}$ as an example,
\begin{linenomath}
\begin{equation}
\LF_{i, j+\frac{1}{2}}=\{ F_{i,j\pm \frac{1}{2}}, F_{i,j+\frac{3}{2}}, F_{i\pm\frac{1}{2},j}, F_{i\pm \frac{1}{2},j+1} \}.
\end{equation}
\end{linenomath}
\label{def:lf}
\end{definition}
Figure \ref{fig:lf} illustrates the local frame $\LF_{i,j+\frac{1}{2}}$.
\begin{figure}
\centering
\subfloat[\label{fig:swapping} swap regions]{\includegraphics[width=0.3\textwidth]{swapping_03.eps}} \quad
\subfloat[$\LF_{i,j+\frac{1}{2}}$\label{fig:lf}]{\includegraphics[width=0.3\textwidth]{caseseps_26.eps}}
\subfloat[\label{fig:abcd}]{\includegraphics[width=0.3\textwidth]{caseseps_25.eps}}
\caption{Illustration of the swept/swap region and the local frame}
\end{figure}
\begin{definition}
The region between the two curve $x^a_i(t)$ and $x^b_{i}(t)$ in $\Omega $ is referred to a vertical \textit{swap region}. The region between $y_j^a(t)$ and $y_j^b(t)$ in the domain $\Omega$ is referred to as a horizonal swap region. The region enclosed by $x^a_{i}(t), x_i^b(t) , y^b_j(t)$ and $y_{j+1}^b(t) $ is referred to as a \textit{local swap region} in the local frame $\LF_{i,j+\frac{1}{2}}$.
\end{definition}
We introduce this two definition to tell the difference between the CIB/DC methods and the FB/DC method. In a CIB/DC method, a vertical swap region is divided into local swap region, while in the FB/DC method, a vertical swap region is divided by local swept regions. Such a difference results in different programming difficult.
\begin{definition}
The intersection points of $x^a_i(t)$ and $x^b_i(t)$ or $y^a_j(t)$ and $y^b_j(t)$ will be referred to as \textit{singular intersection points}. The total number of singular intersection points between $x^a_i(t)$ and $x^a_i(t)$ is denoted as $ns_{xx}$ for $ 1\leq i \leq M$, and the total singular intersection points between $y^a_j(t)$ and $y_j^b(t)$ will be denoted as $ns_{yy}$.
\end{definition}
\section{Facts and results}
Let $\T^a$ and $\T^b$ be two admissible quadrilateral mesh with the Assumption A1, then the following facts hold \begin{fact}
The element $T^b_{i,j}$ of $\T^b$ locates in the interior of a local patch $\LP^a_{i,j}$ of $\T^a$. \end{fact}
\begin{fact}
The face $F^b_{i,j+\frac{1}{2}}$ locates in the local frame $\LF^a_{i,j+\frac{1}{2}}$.
\end{fact}
\begin{fact}
An inner element of $\T^b$ has at least $4^4$ possible ways to intersect with a local patch in $\T^a$.
\end{fact}
\begin{fact}
If there is no singular point in the local swap region between $x_i^a(t)$ and $x_i^b(t)$ in $\Omega$ for some $i \in \{2, 3, \ldots,M-1 \}$, the local swap region consists of $2(N-1)-1$ polygons. Each singular intersection point in the swap region will bring one more polygon.
\end{fact}
\subsection{Basic lemma}
The following results serve as the basis of the CIB/DC and FB/DC methods.
\begin{lemma}
Let $\T^a$ and $\T^b$ be two admissible mesh of the same structure. Under the Assumption A1, we have
\begin{itemize}
\item [(a)]
\begin{linenomath}
\begin{equation}
\mu(T^b_{i,j}) = \mu(T^a_{ij}) -\mu(\mathcal{O}^a_{i,j}) + \mu (\mathcal{I}^b_{i,j}), \label{eq:c}
\end{equation}
\end{linenomath}
where $\mu( \cdot )$ is the area of the underlying set, $\mathcal{I}^b_{i,j}$ is the invading set of $T^b_{i,j}$ and $\Oc^a_{i,j}$ is the occupied set of $T^a_{i,j}$ defined in \eqref{eq:I} and \eqref{eq:O} respectively.
\item [(b)]
\begin{linenomath}
\begin{equation}
\mu(T^b_{i,j}) =\mu(T^a_{i,j}) + \overrightarrow{\mu} (\partial F^{b+}_{i +\frac{1}{2},j}) + \overrightarrow{\mu} (\partial^{b+} F_{i+1,j+\frac{1}{2}})
+ \overrightarrow{\mu} (\partial F^{b+}_{i +\frac{1}{2},j+1}) + \overrightarrow{\mu} (\partial^{b+} F_{i,j+\frac{1}{2}})
,
\end{equation}
\end{linenomath}
where $\overrightarrow{\mu}$ stands for the signed area calculated by directional line integrals.
\end{itemize}
\label{lem:b}
\end{lemma}
\begin{proof}
(a)
According to the identities
\begin{linenomath}
\begin{align*}
T^b_{i,j} &=\left (T^b_{i,j} \cap \LP^a_{i,j}) \backslash (T^b_{i,j} \cap T^a_{i,j}) \right) \bigcup (T^b_{i,j} \cap T^a_{i,j} )=\I^b_{i,j} \bigcup (T^a_{i,j} \cap T^b_{i,j}), \\
T^b_{i,j} &=\left (T^a_{i,j} \cap \LP^b_{i,j}) \backslash (T^a_{i,j} \cap T^b_{i,j}) \right) \bigcup (T^a_{i,j} \cap T^b_{i,j} )=\Oc^b_{i,j} \bigcup (T^a_{i,j} \cap T^b_{i,j}),
\end{align*}
\end{linenomath}
we have
\begin{linenomath}
\begin{align}
\mu(T^b_{i,j})= \mu (\I^b_{i,j}) +\mu(T^a_{i,j} \cap T^b_{i,j}), \label{eq:c1}\\
\mu(T^a_{i,j})=\mu(\Oc^a_{i,j}) +\mu (T^a_{i,j} \cap T^b_{i,j}). \label{eq:c2}
\end{align}
\end{linenomath}
Employ \eqref{eq:c1} and \eqref{eq:c2} and cancel $\mu (T^a_{i,j} \cap T^b_{i,j}) $, we obtain the result.
(b)The area of a quadrilateral polygon with vertices $v_1v_2v_3v_4$ counterclockwise arranged can be calculated according to the Green formula by line integrals.
\begin{linenomath}
\begin{equation}
\iint_{\{v_1v_2v_3v_4 \}} dxdy = {\ointctrclockwise}_{\overrightarrow {v_1v_2}+\overrightarrow {v_2v_3}+ \overrightarrow {v_3v_4} +\overrightarrow {v_4v_1}} xdy.
\end{equation}
\end{linenomath}
Then one can verify that
\begin{linenomath}
\begin{align}
{\ointctrclockwise}_{\partial T^b_{i,j}} xdy &= {\ointctrclockwise}_{\partial T^a_{i,j}} xdy + \sum_{k,s}\oint_{\partial^{b+} F_{k,s}} xdy, \label{eq:cf}
\end{align}
\end{linenomath}
where $(k,s) \in \{ (i+\frac{1}{2},j), (i+1, j+\frac{1}{2}), (i+\frac{1}{2},j+1) , (i,j+\frac{1}{2}) \}$.
With the help of Fig.~\ref{fig:IOswept}, one can check that those line integrals which does not go along an edge of $T^b_{i,j}$ on the right hand side of \eqref{eq:cf} are canceled.
\qed
\end{proof}
\begin{remark}
If we write
\begin{linenomath}
$$
\I_{i,j}^b =\bigcup_{(k,s) \in \mathcal{J}_{i,j}^a} ( \I_{i,j}^b\cap T_{k,js}^a), \quad \Oc_{i,j}^a =\bigcup_{(k,s)\in \mathcal{J}_{i,j}^a} (\Oc_{i,j}^a \cap T_{k,s}^a).
$$
\end{linenomath}
and define
\begin{linenomath}
\begin{equation}
FA_{i,j\rightarrow k,s}= ( \I_{i,j}^b\cap T_{k,s}^a) - (\Oc_{i,j}^a \cap T_{k,s}^a) \textit{ for } (k,s) \in \mathcal{J}_{i,j}^a.
\end{equation}
\end{linenomath}
then
\begin{linenomath}
\begin{equation}
\mu(T_{ij}^b) =\mu(T_{i,j}^a)+\sum_{(k,s) \in \mathcal{J}_{i,j}^a} \mu(FA_{i,j\rightarrow k,s}). \label{eq:gfl}
\end{equation}
\end{linenomath}
$FA_{i,j\rightarrow k,s}$ corresponds to the so called \textit{generalised (mass) flux } in \cite[eq. 19]{B11} and \cite[eq. 3.12]{margolin03}. The \textit{generalised flux} terms can have up to 9 terms in a local patch while the (physical) flux area can only have 4 terms.
\end{remark}
Suppose a density function is a piecewise function on the tessellation $\T^a$ of $\Omega$. To avoid the interior singularity, we have to calculate the mass on $T^b_{i,j}$ piecewisely to avoid interior singularity as
\begin{linenomath}
\begin{equation}
\int_{T_{i,j}^b} \rho d \Omega = \sum_{(k,s)\in \mathcal{J}^a_{i,j}} \int_{T^b_{i,j} \cap T^a_{k,s}} \rho d\Omega = \int_{T^a_{i,j}} \rho d\Omega - \int_{\I^b_{i,j}} \rho d\Omega -\int_{\Oc^a_{i,j}} \rho d\Omega.
\end{equation}
\end{linenomath}
The following result is a directly consequence of Lemma \ref{lem:b}.
\begin{corollary}
Let $\T^a$ and $\T^b$ be two admissible quadrilateral mesh of $\Omega$, and $\rho$ is a piecewise function on $\T^a$, then the mass on each element of $\T^b$ satisfies
\begin{linenomath}
\begin{equation}
m(T_{i,j}^b) =m(T^a_{i,j}) -m(\Oc^a_{i,j})+m(\I^b_{i,j}). \label{eq:CIB}
\end{equation}
\end{linenomath}
and
\begin{linenomath}
\begin{equation}
m(T_{i,j}^b)=m(T^a_{i,j}) + \sum_{k,s} m(\partial^{b+} F_{k,s}). \label{eq:FB}
\end{equation}
\end{linenomath}
where $(k,s) \in \{ (i+\frac{1}{2},j), (i+1, j+\frac{1}{2}), (i+\frac{1}{2},j+1) , (i,j+\frac{1}{2}) \}$, $ m(\partial^{b+} F_{k,s})$ is directional mass, the sign is consistent with the directional area of $ \partial^{b+} F_{k,s}$.
\end{corollary}
The formulas \eqref{eq:CIB} and \eqref{eq:FB} are the essential formula for the CIB/DC method and FB/DC method respectively. The so called \textit{flux-intersection-based} approach based on the formula \eqref{eq:gfl} is equivalent to the CIB/DC approach.
It is easy to find that
\begin{linenomath}
$$
\bigcup_{i,j} \Oc^a_{i,j} =\bigcup_{i,j} \I^b_{i,j} = \bigcup_{i,j} (\partial F_{i + \frac{1}{2},j} \cup \partial F_{i,j+\frac{1}{2}})
$$
\end{linenomath}
This is the total swap region and the fluxing area. Since the swap region is nothing but the union of the the intersections between elements in $\T^a$ and $\T^b$, we shall introduce the following definition to characterise the ways of the intersections.
\begin{theorem}
Let $\T^a$ and $\T^b$ be two admissible quadrilateral mesh of a square in $R^2$ with the Assumption A1 and A2. $ns_{xx}$ and $ns_{yy}$ be the singular intersection numbers between the vertical and horizonal edges of the two mesh. If there is no common edge in the interior of $\Omega$ between $\T^a$ and $\T^b$. Then the swapping region of the two mesh consists of
\begin{linenomath}
\begin{equation}
3(N-1)(M-1)-2((M-1)+(N-1))+1 +ns_{xx} +ns_{yy}. \label{eq:comp}
\end{equation}
\end{linenomath}
polygons.
\label{thm:1}
\end{theorem}
\begin{proof}
We first consider the case where there is no singular intersection points.
There are $2(N-1)$ horizontal curves, say, $$y_1^a(t)=y_1^b(t), y_2^a(t), y_2^b(t), \ldots, y_{N-1}^a(t),y_{N-1}^b(t), y_N^a(t)=y_N^b(t).$$
For each pair of $x_i^a(t),x_i^b(t)$, $i=2,\ldots,M-1$, there are $2(N-1)-1$ quadrilateral polygons between them. And there are $N-2$ polygons between
$\max_{x}\{ x_i^a(t),x_i^b(t) \}$, $\min_{x} \{ x_{i+1}^a(t), x_{i+1}^b(t) \}$ for $i=1, M-1$ and $y_j^a(t)$ and $y_j^b(t)$ for $j=2,\ldots,N-1$. Therefore there are in total
\begin{equation}
3(N-1)(M-1)-2((M-1)+(N-1))+1.
\end{equation}
The result is follows because each singular intersection points divides one not simple polygon into two simple polygons. \qed
\end{proof}
Notice that $(N-1)(M-1)$ is the number of the elements of the tessellation of $\T^a$ and $\T^b$. Then \eqref{eq:comp} implies for the CIB/DC method, the swap region can be computed in $O(n)$ time when every the overall singular intersection points is bounded in $\mathcal{O}(n)$, where $n$ is the number of the cells. The complexity depends on the singular intersection points. Such singular intersection points depends on the underlying problem, for example, a rotating flow can bring such singular intersections.
\subsection{Intersection between a face and a local frame}
\begin{table}
\centering
\caption{Cases of intersections of a vertical edge with a local frame}
\begin{tabular}{cccc}
\hline
& \multicolumn{3}{c}{ Intersection \# with horizonal/vertical frames (H\#/V\#)} \\ \cline{2-4}
& H0 & H1 & H2 \\\hline
0 & shrunk & shifted & stretched \\
V1& diagonally shrunk & diagonally shifted & diagonally stretched \\
V2& --& shifted & stretched \\
V3& --& -- & diagonally stretched \\
\hline
\end{tabular}
\label{tab:case}
\end{table}
For the two admissible quadrilateral mesh $\T^a$ and $\T^b$ of the same connectivity, we classify the intersections between a face $F^b_{i,j+\frac{1}{2}}$ and the local frame $\LF^a_{i,j+\frac{1}{2}}$ into six groups according to the relative position of the vertices $Q_{i,j}$ and $Q_{i,j+1}$ in the local frame $\LF^a_{i,j+\frac{1}{2}}$. The point $Q_{i,j+1}$ can locate in $A_1$,$A_2$, $A_3$ and $A_4$ in the local frame of $\LF_{i,j+\frac{1}{2}}$ in Figure~\ref{fig:abcd}, and the point $Q_{i,j}$ can locate in $B_1$,$B_2$, $B_3$ and $B_4$ region. Compared with the face $F^a_{i,j+\frac{1}{2}}$, the face $F^b_{i,j+\frac{1}{2}}$ can be
\begin{itemize}
\item shrunk: $A_3B_2$ and $A_4B_1$;
\item shifted: $A_1B_1$, $A_2B_2$, $A_3B_3$ and $A_4,B_4$;
\item stretched: $A_1B_2$ and $A_2B_3$;
\item diagonally shrunk: $A_3B_1$ and $A_4B_3$;
\item diagonally shifted: $A_1B_2$, $A_2B_1$, $A_3B_2$, and $A_4B_3$;
\item diagonally stretched: $A_1B_3$ and $A_2B_4$.
\end{itemize}
And then for each group, we choose one representor/generator to check the intersection numbers between the face $F^a_{i,j+\frac{1}{2}}$ and the local frame $\LF^a_{i,j+\frac{1}{2}}$. Finally, according to intersection numbers of the $F^b_{i,j+\frac{1}{2}}$ with the horizonal edges and vertical edges in the local frame $\LF^a_{i,j+\frac{1}{2}}$, we classify the intersection cases into six groups in Table \ref{tab:case} and 17 symmetric cases in Fig.~\ref{fig:case}.
\begin{fact}
Let $\T^a$ and $\T^b$ be two admissible quadrilateral mesh of the same structure. Under the Assumption A1, A2 and A3, an inner edge $F^b_{i,j+\frac{1}{2}}$ of $\T^b$ has 17 symmetric ways to intersect with the local frame $\LF^a_{i,j+\frac{1}{2}}$. A swept/fluxing area has 17 possible symmetric cases with respect to the local frame.
\end{fact}
\begin{figure}[!t]
\centering
\subfloat[ H0V0\label{fig:h0v0}]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_01.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_02.eps}} \,
\subfloat[ H0V1\label{fig:h0v1}]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_03.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_04.eps}}
\center{ shrunk and diagonally shrunk cases} \\
\centering
\subfloat[H1V0\label{fig:h1v0}]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_05.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_06.eps}} \,
\subfloat[H1V0]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_07.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_08.eps}}
\subfloat[H1V2A]{\includegraphics[height=1.6cm,width=1.2cm]{cases2_01.eps}
\includegraphics[height=1.6cm,width=1.2cm]{cases2_02.eps}} \,
\subfloat[H1V2B\label{fig:h1v2}]{\includegraphics[height=1.6cm,width=1.2cm]{cases2_03.eps}
\includegraphics[height=1.6cm,width=1.2cm]{cases2_04.eps} \,
}
\center{ shifted cases } \\
\subfloat[H1V1A\label{fig:h1v1}]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_09.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_10.eps}}\,
\subfloat[H1V1A]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_11.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_12.eps}}\,
\subfloat[H1V1B]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_13.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_14.eps}}\,
\subfloat[H1V1B\label{fig:h1v1c}]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_15.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_16.eps}}
\center{ diagonally shifted cases}\\
\subfloat[H2V0\label{fig:h2v0}]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_17.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_18.eps}} \,
\subfloat[H2V2A]{\includegraphics[height=1.6cm,width=1.2cm]{cases2_05.eps}
\includegraphics[height=1.6cm,width=1.2cm]{cases2_06.eps}} \,
\subfloat[H2V2B\label{fig:h2v2b}]{\includegraphics[height=1.6cm,width=1.2cm]{cases2_07.eps}
\includegraphics[height=1.6cm,width=1.2cm]{cases2_08.eps}} \,
\center{stretched cases.}
\subfloat[H2V1A\label{fig:h2v1}]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_19.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_20.eps}} \,
\subfloat[H2V1B]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_21.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_22.eps}} \,
\subfloat[H2V1C]{\includegraphics[height=1.6cm,width=1.2cm]{caseseps_23.eps}
\includegraphics[height=1.6cm,width=1.2cm]{caseseps_24.eps}}
\subfloat[H2V3\label{fig:h2v3}]{\includegraphics[height=1.6cm,width=1.2cm]{cases2_09.eps}
\includegraphics[height=1.6cm,width=1.2cm]{cases2_10.eps}}
\center{diagonally stretched cases}
\caption{Intersections between a face(dashed line) and a local frame}
\label{fig:case}
\end{figure}
\subsection{Fluxing/swept area and local swap region in a local frame}
For the FB/DC method, once the intersection between $F^b_{i,j+\frac{1}{2}}$ between $\LF^a_{i,j+\frac{1}{2}}$ is determined, then the shape of the swept area $\partial F_{i,j+\frac{1}{2}}$ will be determined. There are total 17 symmetric cases as shown in Fig. \ref{fig:case}. On contrast, the local swap region requires additional effort to be identified.
The vertex $Q_{i,j}$ lies on the line segment $y^b_{j}(t)$, according the Assumption A2, $y^b_{j}(t)$ can only have one intersection with $x^a_i(t)$. This intersection point will be referred to as $V_1$.
The line segment $Q_{i,j}V_1$ can have 0 or 1 intersection with the local frame $\LF_{i,j+\frac{1}{2}}$ except $V_1$ itself; the intersection point, if any, will be denoted as $V_2$.
Table \ref{tab:V} describes how the line segment $Q_{i,j}V_1$ intersects with the local frame according to the relative position of $Q_{i,j}$. The north and south boundary of the local swap region therefore can have 1 or 2 intersection with the local frame. Therefore each case in Figure \ref{fig:case} can result up to four possible local swap region. We use the intersection number between the up and south boundary and the local frame to classify the four cases, see Figure \ref{fig:cswap} for an illustration. For certain cases, it is impossible for the point $Q_{i,j}V_1$ have two intersection points with the local frame, 98 possible combinations are illustrated in Figure~\ref{fig:branch}.
\begin{table}[!t]
\centering
\caption{Intersection cases of the up/down edge of a local swap region}
\begin{tabular}{cc|c|c|c|c}
\hline
& & \multicolumn{4}{c}{position of $Q_{i,j}$ in the local frame} \\\cline{3-6}
\# & points & $B_1$ & $B_2$ & $B_3$ & $B_4$ \\ \hline
1 & $V_1=$& $F^b_{i-\frac{1}{2},j} \cap F^a_{i,j+\frac{1}{2}}$ & $F^b_{i+\frac{1}{2},j} \cap F^a_{i,j+\frac{1}{2}}$ & $F^b_{i+\frac{1}{2},j} \cap F^a_{i,j-\frac{1}{2}}$ & $F^b_{i-\frac{1}{2},j} \cap F^a_{i,j-\frac{1}{2}}$ \\
\hline
2 & $V_1=$ & $F^b_{i-\frac{1}{2},j} \cap F^a_{i,j-\frac{1}{2}}$ & $F^b_{i+\frac{1}{2},j} \cap F^a_{i,j-\frac{1}{2}}$ & $F^b_{i+\frac{1}{2},j} \cap F^a_{i,j+\frac{1}{2}}$ & $F^b_{i-\frac{1}{2},j} \cap F^a_{i,j+\frac{1}{2}}$ \\
& $V_2=$ & $F^b_{i-\frac{1}{2},j} \cap F^a_{i+\frac{1}{1},j}$ & $F^b_{i+\frac{1}{2},j} \cap F^a_{i-\frac{1}{2},j}$ & $F^b_{i+\frac{1}{2},j} \cap F^a_{i-\frac{1}{2},j}$ & $F^b_{i-\frac{1}{2},j} \cap F^a_{i+\frac{1}{2},j}$ \\
\hline
\end{tabular}
\label{tab:V}
\end{table}
\begin{figure}[!t]
\centering
\subfloat[U1S1]{\includegraphics[height=1.6cm,width=1.2cm]{shadow_01.eps}
\includegraphics[height=1.6cm,width=1.2cm]{shadow_05.eps}}\,
\subfloat[U2S2]{\includegraphics[height=1.6cm,width=1.2cm]{shadow_02.eps}
\includegraphics[height=1.6cm,width=1.2cm]{shadow_06.eps}} \,
\subfloat[U2S1]{\includegraphics[height=1.6cm,width=1.2cm]{shadow_03.eps}
\includegraphics[height=1.6cm,width=1.2cm]{shadow_07.eps}} \,
\subfloat[U2S2]{\includegraphics[height=1.6cm,width=1.2cm]{shadow_04.eps}
\includegraphics[height=1.6cm,width=1.2cm]{shadow_08.eps}}
\caption{The shapes of local swap region based on the case of H0V0 in Fig.~\ref{fig:case}.}
\label{fig:cswap}
\end{figure}
\begin{figure}
\centering
\footnotesize
\resizebox{\textwidth}{0.8\textheight}{
\input{branches.tizk}
}
\caption{Classification tree for all possible cases of a local swap region. $Q_{i+1,j+1}$ is on of the vertex of $F^b_{i,j+1/2}$, $L$, $R$ $U$ and $D$ stand for the relative position of $Q_{i,j}$ in the local frame $\LF^a_{i,j+\frac{1}{2}}$. The edge $F^b_{i,j+\frac{1}{2}}$ can be shifted(st), shrunk(sk), stretched(sh), diagonally shifted(dst), diagonally shrunk(dsk), and diagonally stretched(dsh). }
\label{fig:branch}
\end{figure}
\begin{fact}
Let $\T^a$ and $\T^b$ be two admissible quadrilateral mesh of the same structure. Under the Assumption A1, A2 and A3, an local swap region has up 98 cases with respect to a local frame.
\end{fact}
Form the Fig.1 in \cite{D84} and \cite{Ramshaw1986}, we shall see that Ramshaw's approach requires at least 98 programming cases Under the Assumption A1, A2 and A3, while the swept region approach requires only 34 programming cases. This is a significant improvement. When the Assumption A3 fails, or the so called degeneracy of the overlapped region arises, more benefit can be obtained.
\subsection{Degeneracy of the intersection and signed area of a polygon}
\begin{figure}[!t]
\centering
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_01.eps}}
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_03.eps}}
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_11.eps}}
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_12.eps}}
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_16.eps}}
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_13.eps}}
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_17.eps}}
\subfloat[ ]{\includegraphics[height=1.6cm,width=1.2cm]{degenerate_18.eps}}
\caption{Some degeneracy of the intersections. The swept/fluxing area for (a)-(e) degenerates to be a triangle plus a segment, for (f),(g) and (h), the intersection degenerate to a line segment. While for the swap region, an overlap line segment of a new edge and the old edge can be viewed a degeneracy of a quadrilaterals. A overlap of the horizontal and vertical intersection point can be viewed as the degeneracy of a triangular.}
\end{figure}
The intersection of $T^b_{i,j}$ and $T^a_{k,s}$ for $(k,s) \in \mathcal{J}^a_{i,j}$ can be an polygon, an edge or even only a vertex. The degeneracy of the intersection was believed as one of the difficulty of the challenge of the CIB/DC method\cite[p.273]{ms03}. In fact, some cases can be handled by the Green formula to calculate the planar polygon with line integrals. Suppose the vertices of a polygon are arranged in counterclockwise, say, $P_1P_2,\ldots, P_s$, then it area can be calculated as
\begin{linenomath}
\begin{align}
\iint dxdy &= {\ointctrclockwise}_{\overrightarrow {P_1P_2}+\overrightarrow {P_2P_3}+ \cdots +\overrightarrow {P_sP_1}} xdy \nonumber \\
& =\sum_{k=2}^{s+1} \frac{(y_k-y_{k-1})(x_k+x_{k-1})}{2}=
\frac{1}{2} \sum_{k=1}^{s-1}(x_ky_{k+1}-y_kx_{k+1}), \label{eq:darea}
\end{align}
\end{linenomath}
where $P_{s+1}=P_1$. This is due to the fact
\begin{linenomath}
$$
\int_{\overrightarrow{P_1P_2}} x dy = \int_{x_1}^{x_2} x \frac{y_2 -y_1}{x_2-x_1} dx = \frac{(y_2-y_1)(x_1+x_2)}{2}.
$$
\end{linenomath}
The formula \eqref{eq:darea} can handle any polygon including the degenerate cases: a polygon with $s+1$ vertices degenerate to one with $s$ vertices, a triangle degenerates to a vertex or a quadrilateral polygonal degenerates to a line segment. Such degenerated cases arise when one or two of the vertices of the the face $F^b_{i,j+\frac{1}{2}}$ lie on the local frame $\LF^a_{i,j+\frac{1}{2}}$, or the horizontal intersection points overlaps with the vertical intersection points. The later cases bring no difficulty. What really brings difficulties are the cases when $Q_{ij}$ locates in the vertical lines in the local frame or the faced overlaps with partial of the vertical lines in the local frame.. To identify such degeneracies, one only need more flags to identify whether such cases happens when calculating the intersecting points between face $F^b_{i,j+\frac{1}{2}}$ and $F^a_{i,j\pm\frac{1}{2}}$ and $F^a_{i,j+\frac{3}{2}}$.
\subsection{Assign a new vertex to an old cell}
As shown in Fig.~\ref{fig:branch}, the relative position of a new vertex in an old local frame is the basis to classify all the intersection possibilities. This can also be obtained by the Green formula for the singed area of a polygon. We shall denote the signed area of $P_{i,j}P_{i,j+1}Q_{i,j}$ as $A_1$, $P_{i,j}P_{i,j+1}Q_{i,j}$ as $A_2$, $P_{i,j-1}P_{i,j}Q_{i,j}$ as $A_3$ and $P_{i,j-1}P_{i,j}Q_{i,j}$ as $A_4$. Then the vertex $Q_{i,j}$ can be assigned according to Algorithm~\ref{alg:assign}. This determine the first two level of branches of the classification tree in Fig. \ref{fig:branch}.
\begin{algorithm}
\caption{Assign current new vertex to an old cell}
\label{alg:assign}
\begin{algorithmic}
\State{ Compute $A_1$ }
\If {$A_1\ge 0$}
\State{ Compute $A_2$}
\If { $A_2\leq 0$}
\State{ \Return flag='RU'; }\Comment{$Q_{i,j} \in C_{i+\frac{1}{2},j+\frac{1}{2}}$}
\Else
\State{Compute $A_3$}
\If {$A_3\geq 0$}
\State{ \Return flag='LU';} \Comment{$Q_{i,j} \in C_{i-\frac{1}{2},j+\frac{1}{2}}$}
\Else
\State{ \Return flag='RD';} \Comment{$Q_{i,j} \in C_{i-\frac{1}{2},j-\frac{1}{2}}$}
\EndIf
\EndIf
\Else
\State{ Compute $A_4$}
\If { $A_4 \leq 0$}
\State{ \Return flag='RD'; }\Comment{$Q_{i,j} \in C_{i+\frac{1}{2},j-\frac{1}{2}}$}
\Else
\State{Compute $A_3$}
\If {$A_3<0$}
\State{ \Return flag='LD';} \Comment{$Q_{i,j} \in C_{i-\frac{1}{2},j-\frac{1}{2}}$}
\Else
\State{ \Return flag='LU';} \Comment{$Q_{i,j} \in C_{i-\frac{1}{2},j+\frac{1}{2}}$}
\EndIf
\EndIf
\EndIf
\end{algorithmic}
\end{algorithm}
\begin{figure}[!t]
\centering
\subfloat[]{\includegraphics[width=0.32\textwidth, height=0.32\textwidth]{swapping_03}}
\subfloat[\label{fig:VS}]{\includegraphics[width=0.32\textwidth, height=0.32\textwidth]{swapping_01}}
\subfloat[\label{fig:HS}]{\includegraphics[width=0.32\textwidth, height=0.32\textwidth]{swapping_02}}
\caption{Illustration for the vertical and horizontal swap and swept regions}
\end{figure}
\subsection{Alternative direction sweeping}
To calculate all intersections in the swap area and swept/fluxing area, we can apply the alternative direction idea: view the union of the swap area of two admissible quadrilateral mesh of a domain as the union of (logically) vertical strips (shadowed region in Fig.~\ref{fig:VS}) and horizonal strips (shadowed area in Fig.~\ref{fig:HS}). One can alternatively sweep the vertical and horizontal swap strips. Notice that the horizonal strips can be viewed as the vertical strip by exchanging the x-coordinates and the y-coordinates. Therefore, one can only program the vertical sweep case. Each sweep calculate the intersections in the vertical/horizontal strips chunk by chunk. For the CIB/DC method, each chunk is a local swap region, while the FB/DC method, each chunk is a fluxing/swept area.
For the CIB/DC method, one can avoid to repeat calculating the corner contribution by thinning the second sweeping. The first vertical sweep calculate all the intersection areas in the swap region. The second sweep only calculates the the swap region due to the intersection of $T^b_{i,j} \cap T^a_{i,j\pm1}$. In contrast, in the FB/DC methods, the two sweep is totaly symmetric, repeat calculating the corner contribution is necessary.
\section{Applications}
Consider the following two kinds of grids.
\subsection{Tensor product grids}
The mesh on the unit square $[0,1] \times [0,1]$ is generated by the following function
\begin{linenomath}
\begin{equation}
\begin{cases}
x(\xi, \eta, t)=(1-\alpha(t)) \xi +\alpha(t) \xi^3, \\
y(\xi, \eta, t)=(1-\alpha(t)) \eta^2 \\
\alpha(t)= \sin(4\pi t)/2, \quad \xi,\eta, t \in[0,1].
\end{cases}
\label{eq:tensor}
\end{equation}
\end{linenomath}
This produce a sequnce of tensor product grids $ {x_{i,j}^n}$ given by
\begin{linenomath}
\begin{equation}
x_{i,j}^n=x(\xi_i,\eta_i, t^n), y_{i,j}^n=y(\xi_i, \eta_j, t^n).
\end{equation}
\end{linenomath}
where $\xi_i $ and $\eta_j$ are $nx$ and $ny$ equally spaced points in $[0,1]$. For the old grid, $t_1=1/(320+nx)$, $t_2=2t_1$. We choose $nx=ny=11,21,31,41,\ldots,101$.
\begin{figure}
\centering
\subfloat[ Tensor grids ]{\includegraphics[width=0.45\textwidth]{demomesh11_01.eps}}
\subfloat[ Random grids ]{\includegraphics[width=0.45\textwidth]{demomesh11_02.eps}}
\caption{Illustration of the tensor grids and the random grids. }
\end{figure}
\subsection{Random grids}
A random grid is a perturbation of a uniform grid,
\begin{linenomath}
\begin{equation}
\begin{cases}
x_{ij}^n= \xi_i + \gamma r_i^n h, \\
y_{ij}^n= \eta_j+ \gamma r_j^n h.
\end{cases}
\end{equation}
\end{linenomath}
where $\xi_i$ and $\eta_j$ are constructed as that in the above tensor grids. $h=1/(nx-1)$. We use $\gamma=0.4$ as the old grid and $\gamma=0.1$ as the new grid, $nx=ny$.
\subsection{Testing functions}
\begin{figure}
\centering
\subfloat[Franke function \label{fig:franke} ]{\includegraphics[width=0.32\textwidth]{sasfig2_01.eps}}
\subfloat[tanh function \label{fig:tanh} ]{\includegraphics[width=0.32\textwidth]{sasfig2_06.eps}}
\subfloat[peak function \label{fig:peak} ]{\includegraphics[width=0.32\textwidth]{sasfig2_11.eps}}
\caption{Contour lines of the test functions on $101 \times 101$ equally spaced mesh in the unite square. }
\label{fig:test}
\end{figure}
We use three examples as the density function, the first one is the \texttt{franke} function in Matlab.
\begin{linenomath}
\begin{align}
\rho_1(x,y)=franke(x,y)= &.75\exp(-((9x-2)^2 + (9y-2)^2)/4) + \nonumber \\& .75\exp(-((9x+1).^2)/49 - (9y+1)/10) + \nonumber \\
&.5\exp(-((9x-7)^2 + (9y-3)^2)/4) - \nonumber \\& .2\exp(-(9x-4)^2 - (9y-7)^2).
\end{align}
\end{linenomath}
The second example is a continuous function with sharp gradient. It is a shock like function.
\begin{linenomath}
\begin{equation}
\rho_2(x,y)=\tanh(y-15x+6)+1.2.
\end{equation}
\end{linenomath}
The third example is the peak function used in \cite{margolin03}
\begin{linenomath}
\begin{equation}
\rho_3(x,y)=
\begin{cases}
0, & \sqrt{(x-0.5)^2+(y-0.5)^2} >0.25; \\
\max\{ 0.001, 4(0.25-r) \}, & \sqrt{(x-0.5)^2+(y-0.5)^2} \leq 0.25.
\end{cases}
\end{equation}
\end{linenomath}
Fig.~\ref{fig:test} illustrates the contour lines of the test functions. The initial mass of on the old cell are calculated by a fourth order quadrature, the remapped density function is calculated by the exact FB/DC method: the swept region is calculated exactly. The density are assumed to be a piecewise constant on each old cell. Since the swept/flux area are calculated exactly. The remapped error only depends on the approximation scheme to the density function on the old cell. The $L_\infty$ norm
\begin{linenomath}
\begin{equation}
\|\rho^{*} -\rho \|_{\infty} =\max_{i_j}| \rho^{h}_{i+\frac{1}{2},j+\frac{1}{2}}- \rho(x_{i+\frac{1}{2},j+\frac{1}{2}}) |
\end{equation}
\end{linenomath}
is expected in the order of $\mathcal{O}(h)$ for piecewise constant approximation to the density function on the old mesh. While the $L_1$ norm
\begin{linenomath}
$$
\|m^*-m \|_{\infty}=\max_{i,j} | (\rho^{h}_{i+\frac{1}{2},j+\frac{1}{2}}- \rho(x_{i+\frac{1}{2},j+\frac{1}{2}}))\mu(C_{i+\frac{1}{2}, j+\frac{1}{2}}) |.
$$
\end{linenomath}
is expected to be in the order of $\mathcal{O}(h^3)$. We don't use the $L_1$ normal like in other publications, because when plot the convergence curve in in the same figure, the $L_1$ norm and the $L_\infty$ norm for the density function converges also the same rate.
Fig.~\ref{fig:error} demonstrates the convergence of the remmapping error based on the piecewise constant reconstruction of the density function in the old mesh.
\begin{figure}
\centering
\subfloat[ Tensor grids ]{\includegraphics[width=0.45\textwidth]{sasfig2_16.eps}}
\subfloat[ Random grids ]{\includegraphics[width=0.45\textwidth]{sasfig_16.eps}}
\caption{Convergence of the error between the remapped density functions and the true density functions.$F$: the Franke function, P: the peak function, and T: tanh function. The error are scaled by the level on the coarse level.}
\label{fig:error}
\end{figure}
\begin{figure}
\centering
\subfloat[ franke ]{\includegraphics[width=0.32\textwidth]{sasfig2_05.eps}}
\subfloat[ tanh ]{\includegraphics[width=0.32\textwidth]{sasfig2_10.eps}}
\subfloat[ peak ]{\includegraphics[width=0.32\textwidth]{sasfig2_15.eps}}
\caption{Remapped error of the density function on $101 \times 101$ tensor grids.}
\end{figure}
\begin{figure}
\centering
\subfloat[ franke ]{\includegraphics[width=0.31\textwidth]{sasfig_05.eps}}
\subfloat[ tanh ]{\includegraphics[width=0.34\textwidth]{sasfig_10.eps}}
\subfloat[ peak ]{\includegraphics[width=0.31\textwidth]{sasfig_15.eps}}
\caption{Remapped error of the density function on $101 \times 101$ random grinds.}
\end{figure}
\begin{figure}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_02.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_03.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_04.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_07.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_08.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_09.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_12.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_13.eps}}
\subfloat[ ]{\includegraphics[width=0.32\textwidth]{sasfig_14.eps}}
\caption{Contour lines of the remapped density functions on $nx \times ny $ random grids. $nx=ny=11$(left), 21(middle) and 51(right).}
\end{figure}
\section{Discussion}
Computing the overlapped region of a Lagrangian mesh(old mesh) and a rezoned mesh(new mesh) and reconstruction (the density or flux) on the Lagrangian mesh are two aspects of a remapping scheme. According to the way how the overlapped (vertical) strips are divided, a remmapping scheme can be either an CIB/DC approach or an FB/DC approach. Both approach are used in practice. The CIB/DC methods is based on pure geometric or set operation. It is conceptually simple, however Fig.\ref{fig:branch} shows the complexity of computing all the intersections of the two quadrilateral mesh of the same connectivity, it requires 98 programming cases for the non-degenerate intersections to cover all the possible intersection cases which are more than 256. On contrast, the FB/DC approach is based on the physical law, flux exchanges. Mathematically, this is based on the Green formula and the line integral formula. Fig. \ref{fig:case} shows for the FB/DC method, it only requires 34 programming cases for the non-degenerate intersections. This approach is attractive for the case when the two mesh share the same connectivity. Here we present method to calculate fluxing/swept area or the local swap area. They are calculated exact. The classification on the intersection types can help us to identify the possible degenerate cases, this is convenient when develop a robust remapping procedure. Based on the Fact 3, we know there are at least 256 possible ways for a new cell to intersect with an old tessellation. But according to Fig.\ref{fig:case}, we can tell there are more cases than 256. What is the exactly possibilities? This problem remains open as far as we know.
\bibliographystyle{spmpsci}
|
\section*{Note added}
During the ICHEP 2016 conference, hence after the submission of this work to the journal,
both ATLAS and CMS updated their earlier results by employing
an increased data sample
based on some 12 fb$^{-1}$ of LHC Run 2 data per experiment: see the talks by
B. Lenzi (ATLAS) and C. Rovelli (CMS) at
{\tt https://indico.cern.ch/event/432527/contributio}\\
{\tt ns/1072431/}. The new LHC results around 750 GeV are now compatible with the SM at the level of less than
$2\sigma$. Therefore, data in the $\sim750$ GeV region are, at present, consistent with a statistical
fluctuation. However, it remains of importance to explore the high mass region for two reasons.
Firstly, the fact that both ATLAS and CMS initially recorded a very similar excess calls for a closer scrutiny than normal of
data as more and more luminosity will accrue.
Secondly, a coordinated effort between theorists and
experimentalists is presumably required in order to carefully address all possibilities explaining potential anomalies,
from the more evident ones (like resonances) to the more subtle ones (like those studied here).
But let us first recap the situation. Tab.~\ref{tab:ICHEP} does so in statistical terms before and after
the ICHEP 2016 conference. It should also be noted that, while CMS retains a moderate excess
at 750 GeV, ATLAS now favours a mass value slightly lower (730 GeV) or higher (770 GeV) than previously. In short,
the local significance is still slightly pronounced (at the $2\sigma$ level) while we estimate the global one to be rather
poor (at the $1\sigma$ level). However, two considerations are in order. On the one hand,
notice that, with twice as much luminosity as the present one, it would be possible
to re-obtain a $\approx3\sigma$ significance per experiment. On the other hand, it should be appreciated that
only the $\sigma\times$ BR hypothesis has been tested. Hence, the dynamics we propose, i.e., a less pronounced threshold enhancement (which is approximately modelled by an asymmetric Landau distribution) than a resonant one (which is approximately modelled by a BW), as manifest from Fig.~\ref{fig:BW-Threshold}, remains untested to this day. Unfortunately, with data currently unavailable, we are not in the position to test it ourselves. We look forward to the LHC experiments to tackle this challenge rapidly.
\begin{table}
\setlength{\tabcolsep}{5pt}
\begin{tabular}{c||c|c}
\hline
{} & {ATLAS} & {CMS} \\
\hline\multicolumn{3}{c}{}\\[-8pt]
\multicolumn{3}{c} {~~Pre-ICHEP} \\[2pt]
\hline
local $p$-value & $3.9\sigma$ & $3.4\sigma$\\
limit on $\sigma\times$~BR, NWA & $\approx12$ fb & $\approx 14$ fb\\
fitted $\sigma\times$~BR, NWA & $\approx8$ fb & $\approx 4$ fb\\
\hline\multicolumn{3}{c}{}\\[-8pt]
\multicolumn{3}{c} {~~Post-ICHEP} \\[2pt]
\hline
local $p$-value & $2\sigma$ & $2\sigma$\\
limit on $\sigma\times$~BR, NWA(wide) & $\approx2$ fb & $\approx 2(4)$ fb\\
fitted $\sigma\times$~BR, NWA(wide) & $\approx1$ fb & $\approx 1(2)$ fb\\
\hline
\end{tabular}
\caption{\label{tab:ICHEP} ATLAS and CMS statistical findings before and after ICHEP.}
\end{table}
\begin{figure}
\centering
\includegraphics[height=0.9\linewidth,angle=90]{./BW-Threshold-crop}
\caption{Normalised (to 1) differential diphoton mass distributions at the 13 TeV LHC for the processes
$q\bar q\to Z^*, {\rm Box}\to \gamma\gamma$ (solid), with $m_{\rm VLQ}=375$ GeV and $\mu=1$
(VLQ induced), as well as
$gg\to {\rm Higgs}\to \gamma\gamma$ (dotted), with $M_{\rm Higgs}=750$ GeV, $\Gamma_{\rm Higgs}=50$ GeV and
SM-like couplings (CP-even scalar mediated), after the cuts
$p^T_\gamma>20$ GeV, $|\eta_\gamma|<2.5$ and $M_{\gamma\gamma}>100$ GeV.
CTEQ(5L) with $Q=\mu=\sqrt{\hat s}$ is used \cite{Lai:1999wy}.}
\label{fig:BW-Threshold}
\end{figure}
\vspace*{0.25cm}
\noindent
{\emph{Acknowledgements}} SM and LP are supported in part through the NExT Institute and the STFC Consolidated Grant ST/J000396/1. FM is supported in part through the {\em Rita Levi Montalcini 2009 - MIUR} grant. FM wishes to thank Chiara Rovelli for the useful discussions.
\bibliographystyle{apsrev4-1}
|
\part{\Large Reeb components with complex leaves}
\setcounter{section}{-1}
\section{\large Introduction}
The aim of this article is to begin a study of
Reeb components in foliation with comlpex leaves of codimension one,
especially focused on the symmetry in the real $3$-dimensional case.
Quite often we call them leafwise complex foliations.
Recall that a $(p+1)$-dimensional Reeb component is
a compact manifold $R=D^p\times S^1$ with a (smooth) foliation
of codimension one, whose leaves are graphs of smooth functions
$f : \mathrm{int} D^p \to {\Bbb R}$
where $\lim_{z \to \partial D^p}f(z) =+\infty$,
and a compact leaf which is the boundary $S^{p-1}\times S^1$.
Here we identify $R$ with $(D^p \times {\Bbb R}) / {\Bbb Z}$.
See also the figures in Section 2.
Foliations of codimension one with complex leaves are
drawing attentions in several complex variables because
it appears as the Levi foliations of
Levi-flat real hypersurfaces in complex manifolds.
A simple construction of Hopf manifolds admits
a Levi-flat real hypersurface, whose Levi foliation
consists of a pair of Reeb components.
This construction is generalized in Nemirovskii's examples \cite{Ne}.
They have non-trivial linear holonomy along toral leaves.
In this paper, from rather topological points of view,
we study Reeb components with all kinds of holonomy.
Of course,
the case where the holonomy is {\it flat to the identity},
{\it i.e.}, it is infinitely tangent to the identity at the origin,
is included.
Such Reeb components appear in turbulization of a codimension one
foliation along a closed transversal, or in pasting constructions
like Dehn surgery of 3-manifolds.
The present paper is organized as follows.
In Part I we study Reeb components with
complex leaves and their symmetries,
as well as leafwise complex foliations in general.
Part II is devoted to the study of
Schr\"oder's functional equation on the half line,
whose results are the core part of the study of
the automorphism group of Reeb components.
In Section 1 we fix the basic notions on foliations
with complex leaves. They are also called
{\it leafwise complex foliations}.
After these preparations, Reeb components and the turbulization are
reviewed in the context of leafwise complex foliations
in Section 2.
Here we also review the notion of {\it tameness} of Reeb components,
which we always assume in studying the symmetries.
Then, we study the symmetries of a
Reeb component with complex leaves on a 3-manifold.
In section 3 we investigate the structure of
the group of automorphisms of
Reeb components with complex leaves of
complex-dimension 1.
Except for the case where the centralizer
of the holonomy diffeomorphism in
${\mathit Diff}^\infty [0, \infty)$ is
not exactly known,
we completely determine the structure of
the automorphism group.
This exception happens for some of
diffeomorphisms which is flat to the identity.
In anyway, we see that the automorphism group
is of finite dimensional if the linear holonomy
of the boundary leaf is non-trivial.
On the other hand, if the linear holonomy is
trivial, the automorphism group always
contains an infinite dimensional vector space.
(Theorem \ref{MainThm}).
Such a clear contrast results from the
analysis of Schr\"oder's equation
in Part II.
Similar results are obtained for Reeb components of
complex leaf dimension $2$.
They are explained by one of the authors in \cite{Ho}.
In Section 4 some direct corollaries to the
results in Section 3 are stated.
For example, the automorphism group of
a Reeb foliation with complex leaves on the
three sphere is understood.
The study of moduli space of tame Reeb component
is studied by Meersseman and Verjovsky \cite{MV}.
The moduli exhibits to a certain degree a similar phenomena
to those of compact complex manifolds,
especially concerning the finite dimensionality. .
As to automorphism groups
our result tells that
only the Reeb components with non-trivial
linear holonomy on the boundary shows such a similarity.
The second part of the paper is devoted to
the study of
Schr\"oder's equation on the half line.
It is in a form which is looking for
eigen solutions for a pull-back operator.
Here the pull-back diffeomorphism is nothing but
the holonomy of the Reeb component when it is
applied to Part I.
In fact the results are the main ingredients
in describing the automorphism groups
of Reeb components in Part I.
In section 5,
we describe the space of solutions
to
Schr\"oder's equation.
We also extend the values of the equation
to ${\Bbb C}^2$ or still higher dimensional case,
which is used
in \cite{Ho}.
For a diffeomorphism of the half line with
non-trivial linear part, the computation is
easy and in fact is well-known.
For expanding diffeomorphisms with trivial
linear part at the origin,
the proofs are given in the subsequent sections.
We present two different proofs.
In Section 6
a direct proof is given
for the diffeomorphisms which are flat to the identity.
This proof has a similar flavor to
one by the center manifold theory,
which is given in the final section.
In Section 7,
a proof given
for
diffeomorphisms with non-trivial finite jets,
{\it i.e.}, those with the Taylor expansion at the origin
different from the identity.
This proof relies on
Takens' normal form \cite{Ta} and the classical
Fourier series on smooth function on the circle.
Neither of the proofs in Section 7 nor 8
works in other cases.
Then in the final section,
we give a unified proof
which is applicable for both of the above cases.
The main tool is the center manifold theory
of partially hyperbolic dynamical systems.
This proof was suggested by Masayuki Asaoka.
Throughout this article, we assume manifolds and
foliations to be smooth unless otherwise stated.
The authors are deeply grateful to the members of Saturday Seminar
at TITech, especially to Takashi Inaba,
for exciting discussions and valuable comments,
as well as to Masayuki Asaoka for suggesting us
a proof in Part II and
for introducing us the background.
They also would like to express their gratitude
to Laurent Meersseman and Alberto Verjovsky
for their guidance on the basic materials in
Section 1 and 2.
\vspace{30pt}
\noindent
{\bf \Large Part I :
Reeb components with complex leaves}
\section{\large Basic definitions}
Let $M$ be a $(2n+q)$-dimensional smooth manifold
and ${\cal F}$ be a smooth foliation of codimension $q$
on $M$ and let $p=2n$ be the dimension of leaves.
In this section and in the next,
$n$ and $q$ are reserved
for the complex leaf dimension and the codimension.
We refer general basics for foliation theory to \cite{CC}.
\begin{df}[Leafwise complex foliation, cf. \cite{MV}]
\label{LeafwiseComplexStructure}
{\rm \quad A smooth foliation ${\cal F}$ on a smooth manifold
$M$ is said to be a
{\it leafwise complex foliation}
or
{\it foliation with complex leaves}
if there exists a system of local smooth foliated coordinate charts
$(U_\lambda, \varphi_\lambda)$
where $\varphi_\lambda : U_\lambda \to V_\lambda \subset
{\Bbb C}^n\times {\Bbb R}^q=\{(z_1,\, \cdots, z_n, y_1, \,\cdots, y_q)\}$
is a smooth diffeomorphism onto an open set $V_\lambda$
such that the coordinate change
$(w_1,\, \cdots, w_n, t_1, \,\cdots, t_q)=
\gamma_{\mu \lambda}(z_1, \,\cdots, z_n, y_1, \,\cdots, y_q)$
is smooth, $t_j$'s depend only on $y_k$'s ($j,k =1, \,\cdots , q$),
and when $y_k$'s are fixed $w_l$'s are holomorphic in $z_m$'s, where
$\gamma_{\mu \lambda}: \varphi_\lambda(V_\lambda \cap U_\mu)
\to \varphi_\mu(V_\lambda \cap U_\mu)$.
}
\end{df}
\begin{rem}{\rm \quad
Instead of assuming local coordinate
system as above, it is also natural to
consider the following condition
that
the smooth foliation ${\cal F}$ admits
a smooth almost complex structure $J$ acting on
the tangent bundle $T{\cal F}$ to the foliation,
which is integrable on each leaves,
namely there exists local holomorphic coordinates on each leaves.
We assume here $J$ is smooth on the ambient manifold $M$.
(Of course $J$ becomes more than smooth in each leaf. )
This might appear slightly weaker than Definition 1.1,
but
eventually they are equivalent to each other.
To prove from the weaker to the stronger
is nothing but
the parametric version of Newlander-Nirenberg's theorem.
The Newlander-Nirenberg theorem \cite{NN} in the usual sense
claims that an almost complex manifold $(L,J)$
admits a complex structure
(a holomorphic local coordinate system) if
the Nijenhuis tensor $N_J$ vanishes.
In dead in \cite{NN}
Newlander and Nirenberg mentioned
in the very last paragraph
that the parametric version holds.
For the case of $n=1$,
even the parametric version seems to be classically known,
{\it e.g.}, see \cite{Mo}.
It is also well known that if an almost complex structure
$J$ is real analytic
on $2n$-dimensional real analytic manifold $L$,
the Newlander-Nirenberg theorem has a simple geometric proof.
See, for example, Appendix A4 of \cite{Hu}.
Using this argument,
if we can take such a smooth foliated chart
$(u_1, u_2, \cdots, u_{2n}, y_1, \cdots , y_q)$
that $J$ is real analytic on $(u_1, \cdots, u_{2n})$,
we can show the existence of a local coordinate system
in Definition 1.1.
It should be also remarked that
the case of Levi-flat real hypersurface $M$ in
an $(n+1)$-dimensional complex manifold $W$,
the stronger one is easily satisfied.
(If $M$ is of class $C^r$, then we can only assure that
$TM$ is of class $C^{r-1}$,
so that the resultant local coordinate system is assured
to be of class $C^{r-1}$. )
}
\end{rem}
\begin{df}{\rm \quad
A diffeomorphism between two foliated manifolds
with complex leaves is said to be an {\em isomorphism}
between leafwise complex foliations iff it preserves the foliations
and gives rise to biholomorphisms between leaves.
An {\it automorphism} is an isomorphisms between the same one.
}
\end{df}
In this paper we are mainly concerned
with foliation of codimension one.
In particular,
our interest will be focused on Reeb components of real dimension $3$,
namely in the case of $n=1$ and $q=1$.
As we see from the examples of Nemirovskii \cite{Ne}
even a real analytic Levi-flat hypersurface in a complex manifold
can admit Reeb components in its Levi foliation.
In such a case, the holonomy along the toral boundary leaf
has a non-trivial linear part.
Apart from Levi-flat real hypersurfaces,
for example, if we perform a turbulization
we easily find various leafwise complex foliations
admitting Reeb components with holonomy
flat to the identity.
See the next section for more detail.
\section{\large Reeb components with complex leaves}
In this section we review a particular construction of
Reeb component with complex leaves and a process of turbulization
which produces a new Reeb component in a leafwise complex foliation.
In order to make pasting construction easier,
we introduce the following notions.
Let $(R, {\cal F}, J)$ (or simply $R$ for short)
be a Reeb component with leafwise complex structure
of complex leaf dimension $n$ and $(H,J_H)$ be its boundary
leaf.
\begin{df}\label{TameFlat}
{\rm
The Reeb component $R$ has a {\it tame} boundary
(or `$R$ is {\it tame} at boundary' for short,
or even shorter `$R$ is {\it tame}')
with respect to a product coordinate $H\times[0,\varepsilon)$
of a collar neighborhood of $H$
if it gives rise to a smooth foliation with
leafwise complex structure
when it is pasted with the product foliation
$(H\times(-\varepsilon, 0],
\{H\times\{x\}\vert x\in(-\varepsilon, 0]\}, J_H)$
along their boundary.
Here each leaf $H\times\{x\}$ has the same complex structure
as $H$ when identified with the natural projection.
Namely,
the Reeb component is extended to the outside as a product foliation.
}
\end{df}
The notion of tameness was introduced in \cite{MV}.
\begin{rem}{\rm \quad
If we forget the leafwise complex structure and consider
the same notion only as foliation of codimension one,
it does not depend on the choice of product coordinate
on the positive side and the tameness implies
exactly that the holonomy is tangent to the identity
to the infinite order. This is because the set of expanding
diffeomorphisms of the half line $[0,\infty)$
which are infinitely tangent to the identity is
an open convex cone and invariant under conjugation
by any diffeomorphism.
Also remark that the tameness depends only
on the smooth projection of the collar neighborhood
to the boundary, which the product coordinate defines.
If two projections have the same infinite jets on the boundary,
the tameness notion coincides for the two.
}
\end{rem}
\begin{df}\label{simple}
{\rm
The leafwise complex structure of a Reeb component $R$ is {\it simple}
around boundary
(or $R$ has a {\it simple} complex structures around boundary)
if the boundary has a collar neighborhood $U\cong H\times [0, \varepsilon)$
such that
the restriction of the projection $U=H\times [0, \varepsilon) \to H$
to each leaf in $U$ is holomorphic.
This notion should also be understood relative to the projection
from a collar neighborhood to the boundary. }
\end{df}
The notions of tameness and simpleness apply not only to Reeb components
but also to more general leafwise complex foliations
of codimension one with a compact leaf or a boundary leaf.
Clearly if a Reeb component has simple complex structures around the boundary
and the holonomy of the boundary leaf is infinitely tangent to the identity,
it is tame with respect to the appropriate projection.
The tameness condition prohibits
unexpected wild behaviour around boundary.
In particular in the case of complex leaf dimension $=1$,
it induces a strong consequence due to Meersseman and Verjovsky.
See the following subsection.
\subsection{Reeb component by Hopf construction}
Let us recall the Hopf construction
which is one of the standard ways to construct
Reeb components.
This construction gives rise to a tame Reeb component
if the holonomy $\varphi$ is infinitely tangent to
the identity at $\,x=0\,$.
\begin{constr}\label{Hopf-1}{\rm (Hopf construction)\quad
Let $\varphi\in\mathit{Diff}^\infty([0,\infty))$
a diffeomorphism of the half line ${\Bbb R}_{\geq 0}=[0, +\infty)$
satisfying
$\varphi(x)-x>0$ for $x>0$,
namely the origin is an expanding unique fixed point.
Also take a (local) biholomorphic diffeomorphism
$G\in \mathit{Diff}^\mathit{hol}({\Bbb C}^n, O)$
which is expanding. This implies that for some
small neighborhood $D$ of the origin $O$ with smooth boundary
$G(\mathrm{int}D)\supset \overline{D}$ and
$\displaystyle \cap_{k=1}^{\infty}G^{-k}(D)=\{O\}$.
Now take $\displaystyle U= \cup_{k=1}^\infty G^k(D) \subset {\Bbb C}^n$.
Then on $\tilde{R}=U\times[0,\infty)\setminus
\{(O,0)\}\subset {\Bbb C}^n \times {\Bbb R}$, take the restriction $\tilde{\cal F}$ of the
product foliation $\{{\Bbb C}^n\times\{x\}\}$
together with the natural complex structure on leaves
and a diffeomorphism
$T=G\times\varphi$ on $U\times[0,\infty)\setminus
\{(O,0)\}$). Practically we take fairy simple diffeomorphisms
such as linear maps as $G$
so that $U$ becomes the whole ${\Bbb C}^n$.
Then on the quotient $R=\tilde{R}/T^{\Bbb Z} $ a foliation ${\cal F}$ with
complex leaves is naturally induced.
From the construction, it is simple around the boundary.
If the holonomy is infinitely tangent to the identity
it is also tame with respect to the coordinate in the construction.
The boundary $U\setminus\{O\}/G^{\Bbb Z}$ is a complex manifold which is
a so called {\it Hopf manifold}.
In the case $n=1$ it is an elliptic curve and the construction is
equivalent to one with linear map as $G$.
}
\end{constr}
\begin{rem}{\quad \rm It is well known as an elementary fact
in complex dynamical systems
that assuming $G(\mathrm{int}D)\supset \overline{D}$
for a bounded connected domain $D\subset {\Bbb C}^n$ is enough to conclude that
there is a unique linearly expanding fixed point
in $D$ and $D$ is included in the attracting basin
of $G^{-1}$.
}
\end{rem}
\begin{thm}\label{MV}
{\rm
(Meersseman-Verjovsky, \cite{MV})\quad
Any tame Reeb component with complex leaves
of complex dimension $1$
is isomorphic to
one of those given by the Hopf construction.
}
\end{thm}
We present a couple of extensions
(variants) of the above
construction.
\begin{constr}\quad{\rm
Now, let us take the product not with
the half line but with the whole real line ${\Bbb R}$.
Let $M$ and $\Phi\in \mathit{Diff}^\infty_+
({\Bbb R})$ be as follows.
\vspace{5pt}
\\
\qquad
$\bullet$ $ M=(U\times{\Bbb R}\setminus\{(O,0)\})/T'^{\Bbb Z}, \quad
T'=G\times\Phi$,
\vspace{3pt}
\\
\qquad
$\bullet$ $x=0$ is an expanding unique fixed point of $\Phi$.
\vspace{5pt}
\\
$M$ consists of two Reeb components
and in exactly the same way as above
a foliation with leafwise complex structure is induced on $M$.
Note that in this construction and in the previous one,
the holonomy of the toral leaf is given by $\Phi$
and by $\varphi$ respectively.
}
\end{constr}
\begin{constr}\label{HopfSurface}\quad{\rm
Next, we further extend the previous consruction to
obtain a Reeb component as a part of a Levi-flat
hypersurface in a Hopf surface.
We take
$\Phi$ to be a
linear expansion in order to extend
it a biholomorphic (in fact linear) expansion
$\tilde\Phi$ of ${\Bbb C}$.
Note that ${\Bbb R}$ is an invariant subspace in ${\Bbb C}$.
Fix the expansion ratio $\mu >1$
and take the followings;
\begin{eqnarray*}
W=(U\times{\Bbb C}\setminus\{O\})/T''^{\Bbb Z}, \,\, \quad\quad
T''=G\times\tilde\Phi, \quad
\tilde\Phi(z)=\mu z \,\,\,(z\in{\Bbb C}),
\\
M=(U\times{\Bbb R}\setminus\{(O,0)\})/T'^{\Bbb Z}, \quad
T'=G\times\Phi, \quad \,
\Phi= \tilde\Phi\vert_{\Bbb R}\, .\,\, \qquad \qquad
\end{eqnarray*}
$W$ is an $(n+1)$-dimensional Hopf manifold,
$M$ is its Levi-flat real hypersurface
with the Levi foliation
consisting of two Reeb components, and
a unique compact leaf is the Hopf manifold $(U\setminus\{O\})/G^{\Bbb Z}$
of $\dim_{\Bbb C} =n$.
}
\end{constr}
\begin{prob}\quad{\rm
Theorem \ref{MV}
due to Meersseman and Verjovsky
poses the following questions.
We assume the complex leaf dimension to be one.
Provided that two Reeb components
with leafwise complex structures
have the same boundary holonomy and their boundary leaves
are biholomorphic to each other, are they isomorphic
as leafwise complex foliations?
Does there exist a Reeb component with complex leaves
which is not isomorphic to a tame one but with
holonomy infinitely tangent to the identity?
Or does there exist one which is not
isomorphic to any of those given
by the Hopf construction?
One more similar but subtle question is to ask whether if
a tame Reeb component is always isomorphic to
one given by the Hopf construction.
The second form of question seems less difficult and
negative.
Anyway, those questions are asking what should be the
complete invariants to determine Reeb components
without assuming the tameness. }
\end{prob}
\begin{constr}\label{preparation}\quad
{\rm We introduce one more
construction,
which is a preparation for turbulization.
Take $\tilde{M}=
({\Bbb C}^n\times {\Bbb R}) \setminus\{O\}\times(-\infty, 0]$ and restrict
the product action $\hat{T}=G \times\Psi$ to $\tilde{M}$,
where $\Psi$ is an orientation preserving diffeomorphism
of ${\Bbb R}$ which fixes $0$,
expanding on $[0,\infty)$,
and {\it contracting} on the negative side
$(-\infty, 0]$,
{\it i.e.},
$\Psi(x) > x $ for $x<0$.
On $\tilde M$ we take (the restriction of)
the horizontal foliation $\tilde {\cal F}$.
Then take the quotient
$(M,{\cal F}, J_{\cal F}) =
(\tilde M,\tilde {\cal F}, J_\mathrm{std})/\hat T ^{\Bbb Z}$.
The non-negative part is nothing but the Reeb component
constructed in \ref{Hopf-1}
regarding $\varphi=\Psi\vert_{[0,\infty)}$.
The non-positive side $(N,{\cal G})=(M,{\cal F})\vert_{x\leq 0}$
remains non-compact and
is in fact a foliated
$(-\infty, 0]$-bundle with holonomy
$\psi= \Psi\vert_{(-\infty, 0]}$.
If we remove the boundary compact leaf $\{x=0\}$
from the non-positive side
$(N, {\cal G})$,
it is isomorphic to $({\Bbb C}^n\setminus \{O\})\times S^1$.
For a better description of turbulization process,
let us be more precise about this identification.
This is done by embedding $\hat T$ in a 1-parameter family.
Take a smooth curve $G_t$ in $\mathit{GL}(2;{\Bbb C})$
and also a smooth curve $\psi_t$ in
$\mathit{Diff}^\infty((-\infty, 0])$
satisfying the following conditions.
$$
\psi_k=\psi^k \,\,\, (k\in {\Bbb Z}), \quad
\psi_{t+1}=\psi\circ\psi_t \,\,\, (t\in {\Bbb R}), \quad
\frac{\partial\psi_t(x)}{\partial t}>0\,\, (\forall x, t),
$$
$$
G_k=G^k \,\,\, (k\in {\Bbb Z}), \quad
G_{t+1}=G\circ G_t \,\,\, (t\in {\Bbb R}).
$$
Then, fixing (any) $x_0<0$,
$x=\psi_t\,(x_0)$ gives a diffeomorphism between
$(-\infty, 0)\,(\ni x)$ and ${\Bbb R}(\ni t)$.
Then the identification of
$(z,x)\in({\Bbb C}^n\setminus\{O\})\times (-\infty, 0)$ with
$(w,t)\in({\Bbb C}^n\setminus\{O\})\times {\Bbb R}$
by
$(z=G_t(w), x=\psi_t(x_0))$
conjugates $\hat T\vert_{x<0}$ into $(w,t)\mapsto(w, t+1)$.
Of course an easy way to choose such a 1-parameter family
$\,\{\psi_t\}\,$ is to take a 1-parameter subgroup.
Take a smooth vector field $\,\rho(x)\frac{d}{dx}\,$ on
$\,(-\infty, 0]\,$ with $\,\rho(x)>0\,$ for $\,x<0\,$ and
$\,\rho(0)=0\,$.
Then putting $\,\psi_t=\exp(tX)\,$ we obtain such
$\,\psi_t\,$ with $\,\psi=\psi_1\,$.
If we choose $\,\rho(x)\,$ to be flat at $\,x=0\,$,
$\,\psi_t\,$ is infinitely tangent to the identity at $x=0$.
Even if $\,\psi\,$ does not ly
It is worth remarking that this identification gives rise to
a partial compactification of horizontally foliated manifold
$(({\Bbb C}^n\setminus\{O\})\times S^1,
\{({\Bbb C}^n\setminus\{O\})\times\{t\}\}$
by a Hopf manifold $N$ as a boundary leaf so as to
obtain $(N, {\cal G})$.
If we take diffeormorphisms $\psi_t$
infinitely tangent to the identity at the origin,
we obtain a tame structure.
Also on the non-negative side,
by taking smilar family $\varphi_t$
for $\varphi=\varphi_1$, we also obtain a tame structure on
the non-negative side.
Once we obtain tame ones on both side with the same
complex structure on the boundaries, we can paste them
to obtain a smooth structure.
}
\end{constr}
\subsection{Turbulization
in $L\times S^1$}\label{standard turbulization}
Here we review the turbulization,
which is classically well-known modification of
a foliation of codimension one
to yield a new Reeb component.
We start from a standard situation.
\begin{constr}\label{standard turbulization}{\rm \quad
Let $(M,{\cal F})$ be a leafwise complex foliation of codimension one
and assume that there is an embedded
solid torus $U={\mathrm{int }}D^{2n} \times S^1$
on which the the induced foliation
is $\{{\mathrm{int }}D^{2n} \times \{*\}\}$ and
the induced complex structure is also the canonical ones
on each
${\mathrm{int }}D^{2n} \times \{*\}
\cong {\mathrm{int }}D^{2n}$
$\subset {\Bbb C}^n$. Let $(w,t)$
denote the natural coordinate of
$U={\mathrm{int }}D^{2n} \times S^1$
where $S^1$ is regarded as ${\Bbb R}/{\Bbb Z}$.
Then we remove
$\{O\}\times S^1$ from $U$
and let
$U^*$
denote the result.
Using the coordinate
$(w,t)$ $U^*$
is identified with an open subset
of the negative side of Construction \ref{preparation},
together with leafwise complex foliations.
Therefore we can compactify this end
with the Hopf manifold as in Construction \ref{preparation}
and also
if we add positive side of Construction \ref{preparation}
we obtain a leafwise complex foliated manifold
without boundary
but with a new Reeb component.
For this construction
we can choose any of $G_t$, $\psi_t$, and $\Phi$ as in
Construction \ref{preparation}.
The above process including adding the positive side is
the leafwise complex version of the {\it turbulization}.
See also Figure 1 below.
}
\end{constr}
\subsection{General case}
It is easy to find a closed transversal to a foliation of codimension one,
namely, an embedded circle which is transverse to the foliation,
unless the manifold is open and the foliation is too simple.
Like in the case of
a smooth foliation without leafwise complex structure,
it is always possible to perform the turbulization
in a tubular neighborhood of any closed transversal
regarding leafwise complex structure.
This fact also belongs a kind of folklore, while below
it is reviewed.
\begin{thm}{\rm
\quad
Let $(M^{2n+1}, {\cal F}, J)$ be a smooth leafwise complex foliation
of codimension one and $K\subset M$ is a closed transvesal,
namely there exists a smooth embedding $f:S^1 \to M$
which is transverse to the foliation ${\cal F}$ with
its image $f(S^1)=K$.
Then, there exists a tubular neighborhood
$U\cong K \times {\mathrm{int}\,}D^{2n}$ such that
the restricted foliation $(U, {\cal F}_U, J\vert_{{\cal F}_U})$
is isomorphic to the standard one
$(S^1 \times {\mathrm{int}\,}D^{2n},
{\cal F}_0=\{t\}\times {\mathrm{int}\,}D^{2n}, J_0)$
and through this isomorphism
$K$ is identified with $S^1\times \{O\}$.
In particular, we can perform the standard turbulization
\ref{standard turbulization}
in $U$.
}
\end{thm}
This theorem is a direct corollary to the following lemma.
\begin{lem}
{\rm
\quad
The group
$\mathit{Diff}^{\mathit{hol}}({\Bbb C}^n, O)$
of germs of holomorphic diffeomorphisms
of $({\Bbb C}^n, O)$ which fix the origin is pathwise connected.
}
\end{lem}
The lemma immediately follows from the two facts
that $GL(n;{\Bbb C})$ is pathwise connected
and that such a germ with identical linear part can
be joined by a straight segment to the identity.
\subsection{Dehn surgery in $\dim=3$ vs. higher dimensional turbulization}
In order to close the section,
this subsection provides with some remarks concerning the possibility
of pasting the Reeb component in a different way in a turbulization.
In the rest of this section,
we assume the holonomy $\Psi$, and eventually
$\varphi$ and $\psi$,
to be infinitely tangent to the identity at the origin,
so that it is easier to past two peices along their boundarie.
\begin{rem}{\rm \quad
If we forget the leafwise complex structure and treat foliations
only as smooth objects, basically there are two ways to perform
the turbulization. The one has been already described above and
is indicated in Figure 1.
For the other one we can reverse the top and bottom
of the Reeb component (Figure 2). This is because the cyclic
(universal for $n\geq 2$) covering of the boundary leaf is
${\Bbb R}^{2n}\setminus\{O\}\cong S^{2n}\setminus\{N, S\}$ and
two ends are exchangeable by a diffeomorphism.
However, as a complex manifold, ${\Bbb C}^n\setminus\{O\}$ has
one convex end and the another concave.
For the case of complex leaf dimension $n$ greater than one,
these two ends are not exchangeable.
In particular, for $n\geq 2$,
the turbulization for leafwise complex foliations does not
change the homotopy class of the tangent bundle.
}
\end{rem}
\noindent
\begin{figure}[H]
\begin{center}
\includegraphics[width=6cm]{turb1.eps}
\qquad
\includegraphics[width=6cm]{turb2.eps}
\\
\hspace*{\fill}
Figure 1
\hspace*{\fill}
\hspace*{\fill}
Figure 2
\hspace*{\fill}
\\
\hspace*{\fill}
\begin{minipage}{119mm}
{\small \mbox{}
\\
The green lines indicate the bounndary
leaves of Reeb components.
The axes of the rotational
symmetries of the Reeb components,
which are not drawn in the figure,
correspond to $\{O\}\times {\Bbb R}_+$.
}
\end{minipage}\vspace{-10pt}
\hspace*{\fill}
\end{center}
\end{figure}
\begin{rem}{\rm \quad In the case of complex leaf dimension one,
the `upside-down' construction always works.
Namely, in Construction \ref{Hopf-1}, $z \leftrightarrow z^{-1}$
always induces an biholomorphism on the boundary elliptic curve.
Therefore we can regard it as a turbulization inthecated in Figure 2.
If the boundary elliptic curve admits a complex multiplication,
namely finite but discrete symmetries of order 2, 3 or 4,
removing the Reeb component and pasting it back with one of
those symmetries is a special kind of Dehn surgeries.
More generally, in the process of turbulization,
after removing a tubular neighborhood of
the closed transversal and compactify the new boundary
(namely after the process of Construction \ref{preparation}),
instead of filling up with a Reeb component as explained above,
we fill up the boundary as follows.
We prepare another
Reeb component with a different complex structure.
If their boundaries match up through some diffeomorphism,
we can fill up the boundary with that Reeb component.
In this way, a Dehn surgery corresponding to
any element of the mapping class group
${\mathcal M}_1(\cong \mathit{}SL(2;{\Bbb Z}))$
of a $2$-dimensional torus $T^2$ is realized
for a closed transversal in a leafwise complex
codimension one foliation of $n=1$.
}
\end{rem}
\section{\large Symmetries of $3$-dimensional Reeb components }
In this section we compute the group of automorphisms of
a Reeb component of dimension $3$ with complex leaves,
which is given by the Hopf construction.
In order to fix notations,
we present our objects again.
Let $\tilde R$ be ${\Bbb C}\times[0,\infty) \setminus \{(0,0)\}$,
take $\lambda\in {\Bbb C}$ with $\vert\lambda\vert>1$ and
a diffeomorphism $\varphi\in\mathit{Diff}^\infty([0,\infty))$
which is expanding, namely, satisfying
$\varphi(x)-x>0$ for $x>0$.
Let $G$ denote the linear expansion of ${\Bbb C}$ defined
as the multiplication by $\lambda$
and $T:\tilde{R} \to \tilde{R}$ be $T=G \times \varphi$.
Then we obtain a Reeb component
$(R, {\cal F}, J)=
(\tilde{R}, \tilde{{\cal F}}, J_\mathrm{std}) /T^{\Bbb Z}$
as the quotient, as well as the boundary elliptic curve
$H={\Bbb C}\setminus\{0\}/G^{\Bbb Z}$.
Here, on the upstairs the leaves of the foliation $\tilde{{\cal F}}=
\{{\Bbb C}\times\{x\} \vert x>0\}\sqcup \{{\Bbb C}^
\times\{0\}\, \}$
are equipped with the natural complex structure
$J_\mathrm{std}$ which is inherited by those of ${\cal F}$.
\vspace{8pt}
We classify the diffeomorphisms $\varphi$
into the following three cases
according to the nature of its jet at $x=0$.
Here, $\varphi^{(i)}$ denotes the $i$-th derivative of
$\varphi(x)$
and $j^i\varphi(0)$ denotes the $i$-th jet at $x=0$.
\vspace{5pt}
\\
\noindent
\quad Case (1) : $\varphi'(0)=\mu >1$ ,
\vspace{2pt}
\\
\noindent
\quad Case (2) : for some $n \geq 2$ ,
$j^{n-1}\varphi(0)=j^{n-1}\mathit{id}(0)\,$
and
$\,\,\varphi^{(n)}(0)>0$ ,
\vspace{2pt}
\\
\noindent
\quad Case (3) :
$j^\infty\varphi(0)=j^\infty\mathit{id}(0)\,$ .
\vspace{8pt}
The discussions in Subsection \ref{LiftRest}
does not depend on the above classification.
As reviewed in Section \ref{centralizer},
the structure of the centralizer of $\varphi$
is very subtle for a certain class in Case (3).
For the rest of Case (3) and for Case (1) and (2),
the centralizer is fairy simple.
The main result of this paper is the
computation of the automorphisms
which fixes the boundary and
the transverse space.
This is in deed the main results of Part II
of this paper.
Concerning this part, for Case (2) and for Case (3)
the results are the same.
For Case (1), such automorpsims are very few.
The internal structure
of the automophism group
for Case (2) and (3) rather depends only on
the nature of the centralizer
(see Subsection \ref{structure}).
The centralizer itself is included in the
automorphism group.
Except for this part,
the extendability to the outside is basically the same
for any $\,\varphi\,$ in Case (2) and (3).
\subsection{Lift to $\tilde R$ and restriction to $H$}
\label{LiftRest}
Let us consider the group $\mathit{Aut} (R, {\cal F}, J)$, which
is also denoted by $\mathit{Aut} R$ for short, of
all foliation preserving diffeomorphisms of $R$ whose
restriction to each leaf is holomorphic.
Also we consider the group of holomorphic diffeomorphisms
$\mathit{Aut} H$ of the boundary elliptic curve $H$
as well as its identity component $\mathit{Aut}_0 H$ which
is isomorphic to $T^2$ and can be identified with $H$ itself.
\begin{prop}{\rm
\quad The image of the restriction map
$r_H:\mathit{Aut} R \to \mathit{Aut} H$ is
exactly $\mathit{Aut}_0 H$.
}
\end{prop}
{\it Proof.} \quad
If we regard
$\mathit{Aut} H / \mathit{Aut}_0 H$
as a subgroup of $\mathit{SL}(2;{\Bbb Z})$,
in most cases it is just $\{\pm E\}$ where $E$ denotes
the identity matrix.
In a few cases where the elliptic curve $H$ admits
complex multiplications, they are of order 3,4, or of 6
and a kind of `rotations' on the universal covering,
{\it i.e.}, elliptic matrices in $\mathit{SL}(2;{\Bbb Z})$.
In any of those cases, no element in
$\mathit{Aut} H \setminus \mathit{Aut}_0 H$
preserves the direction of holonomy and thus none
extends to $R$ as a foliation preserving diffeomorphism.
On the other hand, any element in $\mathit{Aut}_0 H$
is obtained as the quotient of the scalar multiplication
$m_a : {\Bbb C}^* \to {\Bbb C}^*$ by some nonzero complex number $a$.
The automorphism $m_a\times \mathrm{id}_{[0,\infty)}$ of
$\tilde R$ clearly descends to $R$ and defines an element
in $\mathit{Aut} R$.
\hspace*{\fill}$\square$
\\
\\
By this proposition, the study of the structure of
$\mathit{Aut} R$ breaks into two parts,
that of the kernel $\mathit{Aut}(R, H)$
and the study of the restriction map $r_H$.
\\
Now it is easier to look at the lifts of automorphisms
on $\tilde R$.
Any element $f \in \mathit{Aut} R$ has a
lift $\tilde f\in\mathit{Aut}(\tilde R, \tilde {\cal F}, J_\mathrm{std})$
($=\mathit{Aut}\tilde R$)
which takes the form
$$
\tilde f(z,x)=(\xi(z,x), \eta(x))
$$
in ${\Bbb C}\times [0,\infty)$-coordinate.
A lift $\tilde f$ should commutes with the covering
transformation $T$, because, $T\circ \tilde f = \tilde f \circ T^k$
for some $k\in {\Bbb Z}$ but it is easy to see that $k=1$
when it is restricted to the boundary.
Therefore an element in
$\mathit{Aut}\tilde R$
is a lift of some element in $\mathit{Aut} R$
if and only if it commutes with $T$.
Let
$\mathit{Aut}(\tilde R ; T)$
denote the centralizer of $T$ in
$\mathit{Aut}\tilde R$,
namely, the group of all such lifts.
It contains an abelian subgroup
$\{m_a\times \mathrm{id}_{[0,\infty)}\vert a\in{\Bbb C}^*\}\cong{\Bbb C}^*$.
This subgroup injectively descends to a subgroup of
$\mathit{Aut}R$
which restricts exactly to
$\mathit{Aut}_0 H \cong {\Bbb C}^*/\lambda^{\Bbb Z}$.
It is important to remark that
whether $\mathit{Aut}_0 H$ admits a homomorphic section
is not a trivial question.
Postponing this question until the end of this section,
we go on an easier way.
Let us introduce one more subgroup
$\mathit{Aut}(\tilde R, \tilde H; T)$ of
$\mathit{Aut}(\tilde R ; T)$
which consists of all elements
which act trivially on the boundary $\tilde H$.
Any element
$f \in \mathit{Aut}(R, H)$ has a unique lift to
an element $\tilde f\in\mathit{Aut}(\tilde R, \tilde H; T)$
Namely,
\begin{cor}{\rm \quad
$\mathit{Aut}(R, H)$
is isomorphic to
$\mathit{Aut}(\tilde R, \tilde H; T)$.
}
\end{cor}
Again, let
$\,g\in \mathit{Aut}\tilde R\,$
be presented
in the form
$\,
g(z,x)=(\xi(z,x), \eta(x))
\,$.
\begin{lem}\label{principal lemma}
{\rm \quad
The element $\,g\,$
in $\,\mathit{Aut}\tilde R\,$
belongs to
$\mathit{Aut}(\tilde R ; T)$
if and only if
it satisfies the following conditions.
\begin{itemize}
\item[(a)] $\xi(z,x)= a z + b(x)$, $b(0)=0$
for some
$b\in C^\infty([0,\infty), {\Bbb C})$ and $a\in{\Bbb C}^*$.
\item[(b)] $b(\varphi(x))=\lambda b(x)$.
\item[(c)]
$\varphi\circ \eta = \eta \circ \varphi$,
namely, $\eta \in Z_\varphi=$ the centralizer of
$\varphi$ in $\mathit{Diff}^\infty([0,\infty))$.
\end{itemize}
Further more,
$g$ belongs to $\mathit{Aut}(\tilde R, \tilde H ; T)$
if and only if the above conditions are satisfied
with $a=1$.
}
\end{lem}
{\it Proof.}\quad
Let us first show the {\it only if} direction,
then the {\it if} direction will become almost trivial.
Assume
$g\in\mathit{Aut}(\tilde R ; T)$.
$\xi(z,x)$ is smooth and holomorphic in $z$. If $x$ is fixed,
$\xi(\,\cdot\,, x): \
\to\
$ is
a holomorphic automorphism
even in the case where
$x=0$
because the origin is a removable singularity,
it is a linear map with nontrivial linear term.
Therefore it is written in the following form;
$\xi(z, x)=a(x)z + b(x)$ where $a(x), b(x)
\in C^\infty([0,\infty), {\Bbb C})$
with $a(x)\ne0$ and $b(0)=0$.
These also apply to elements in
$\mathit{Aut}\tilde R$.
Now look at the commutation relation
$\,g\circ T=T\circ g\,$.
This implies
$$
(a(\varphi(x))\lambda z + b(\varphi(x)), \eta(\varphi(x))
=(\lambda a(x)z + \lambda b(x), \varphi(\eta(x)))\, .
$$
Thus we obtain (b) and (c).
This also tells us that $a(\varphi(x))=a(x)$,
so that for any $x\geq0\,$ we have
$\displaystyle a(x)=\lim_{n\to\infty}a(\varphi^{-n}(x))=a(0)$
and (a) is concluded.
For $g\in\mathit{Aut}(\tilde R, \tilde H ; T)$
we just need to confirm that $a=1$.
\hspace*{\fill} $\square$
\begin{rem}{\rm\quad
The condition (b)
appears as Equation (I) in Part II.
Solving this
Schr\"oder type functional equation
on the half line $\,[0, \infty)\,$ for
given $\,\lambda\,$ and $\,\varphi\,$
is the main theme in Part II.
}
\end{rem}
\begin{cor}\label{identification}
{\rm \quad
$\mathit{Aut} R$ is naturally isomorphic to
$\mathit{Aut}(\tilde R; T)/T^{\Bbb Z}$.
}
\end{cor}
\subsection{Centralizer of
$\varphi$ in $\mathit{Diff}^\infty([0,\infty))$}
\label{centralizer}
For an expanding diffeomorphism
$\,\varphi\in \mathit{Diff}^\infty([0,\infty))\,$,
concerning its centralizer in
$\,\mathit{Diff}^\infty([0,\infty))\,$
and the embeddability
in a (smooth) 1-parameter subgroup,
the followings are known.
\begin{thm}{\rm\quad
1) \quad For Case (1) thanks to Sternberg's
linearization \cite{St}
and
for Case (2) thank to Takens' normal form \cite{Ta},
there exists a smooth vector field
$\displaystyle\,X=\rho(x)\frac{d}{dx}\,$ on $\,[0,\infty)\,$
such that $\varphi=\exp{X}$ and the centralilzer
$Z_\varphi$ exactly coincides
with the 1-parameter subgroup $\,\exp(tX)\,; \,
t\in {\Bbb R}\}\,$ generated by $X$.
\\
2)\quad In Case (3), if there exists a smooth
vector field
$\displaystyle\,X=\rho(x)\frac{d}{dx}\,$ on $\,[0,\infty)\,$
with $\varphi=\exp{X}$,
then its centralizer
$Z_\varphi$ in coincides
with the 1-parameter subgroup
$\,\exp(tX)\,; \, t\in {\Bbb R}\}\,$.
}
\end{thm}
In general, the centralizer $Z_\varphi$ of
$\varphi$ which is infinitely tangent to the identity
at $\,x=0\,$ is known
to be fairy wild (see \cite{Ey}).
For an expanding $\,\varphi\,$,
it is known that
there exists a unique $C^1$-vector field
$\displaystyle X_\varphi=\rho(x)\frac{d}{dx}$
on the half line $[0,\infty)$,
which is of class $C^\infty$ on $(0, +\infty)$,
in such a way that
the exponential map
$\exp X_\varphi$
coincides with $\varphi$ (see \cite{Sz} and \cite{Na}).
This vector field is often called
the {\it Szekerez vector field} of $\varphi$.
If the Szekerez vector field $X_\varphi$ is of class
$C^\infty$ on $[0,\infty)$,
namely $\rho(x)$ is smooth and flat at $x=0$,
then the centralizer
$Z_\varphi$ coincides with the 1-parameter family
$\{\exp (tX_\varphi)=\varphi^t \,;\, t\in {\Bbb R}\}$
generated by $X_\varphi$.
In general case, using the order of real numbers,
the centralizer $Z_\varphi$ turns out to be a totally
ordered abelian group which contains $\{\varphi^{\Bbb Z}\}\cong{\Bbb Z}$.
Therefore it is uniquely identified with a certain
subgroup of the additive group ${\Bbb R}$ under the identification
$\{\varphi^{\Bbb Z}\}\cong{\Bbb Z}$.
Depending on $\varphi$,
$Z_\varphi$ can be
far beyond the normal expectation,
{\it e.g.}, it can be
${\Bbb Z}$, ${\Bbb Q}$, or
${\Bbb Z}\oplus{\Bbb Z}\alpha$ where
$\alpha$ is a Liouville number \cite{Ey},
or far more complicated.
The topology on $Z_\varphi$ through this identification
with natural topology of ${\Bbb R}$ coincides with
the one induced from
the $C^0$-topology on
$\mathit{Diff}^\infty([0,\infty))$.
We should also remark that for
$\,\varphi\,$ which is infinitely tangent to the identity
at $\,x=0\,$,
so is any element of $\,Z_\varphi,$.
At present it is not known whether $Z_\varphi\cong{\Bbb R}$
implies the smoothness of $X_\varphi$ at $x=0$.
This is a subtle point in Hilbert's problem No. 5 when
it is stated in the context of a continuous
homomorphism from
a Lie group to a group of diffeomorphisms.
The difficulty occurs when the orbit is not compact.
\subsection{Structure of $\mathit{Aut} R$}
\label{structure}
Upon all the previous preparations
we are able to describe the structure of
$\mathit{Aut} R$ as follows.
\begin{prop}{\rm \quad Let $R$ be a Reeb component
of real dimension $3$ which is given
by the Hopf construction.
\\
1)\quad
The group $\mathit{Aut} R$
of automorphisms of the Reeb component $R$
admits a following
sequence of extensions by abelian groups
$\,\mathit{Aut}_0 H\,$, $\,Z_\varphi\,$,
and $\,{\mathcal K}_{\lambda, \varphi}\,$,
$$
0 \to \mathit{Aut}(R, H) \to \mathit{Aut} R
\to \mathit{Aut}_0 H \to 0
$$
$$0 \to {\mathcal K}_{\lambda, \varphi} \to \mathit{Aut}(R, H)
\to Z_\varphi \to 0
\\
$$
where $\mathit{Aut}_0 H \cong {\Bbb C}^*/\lambda^{\Bbb Z}$ is
the multiplication by the constant linear part $a$
mod $\,\lambda^{\Bbb Z}\,$ as
described in Lemma \ref{principal lemma},
$Z_\varphi$ is the centralizer of $\varphi$
which is explained in the previous subsection,
and
${\mathcal K}_{\lambda, \varphi}$
is identified with the space of solutions to
the equation (b) in Lemma \ref{principal lemma}.
As explained in Part II,
$\,{\mathcal K}_{\lambda, \varphi}\,$ is isomorphic to an infinite dimensional
vector space ${\mathcal Z}_{\lambda, \varphi}$ in Case (2) and (3),
while in Case (1) it is a complex vector space
of dimension 1 or 0 according to
the resonance condition $\,\lambda=(\varphi'(0))^n$
for some $\,n\in{\Bbb N}\,$ or not.
}
\end{prop}
The following is an important consequence to
Theomre \ref{Main}
in Part II on $\,{\mathcal K}_{\lambda, \varphi}\cong{\mathcal Z}_{\lambda, \varphi} \,$ in Case (2) and (3).
\begin{cor}\label{extension}{\rm
\quad
If we paste $H\times (-\varepsilon, 0]$ to $R$
along the boundary $H$,
in Case (3) any element of
$\mathit{Aut}(R, H)$
extends to the other side, being the identity on
$H\times (-\varepsilon, 0]$,
as a diffeomorphisms of class $C^\infty$.
In Case (2) the same applies to $\,{\mathcal K}_{\lambda, \varphi}\,$.
}
\end{cor}
\noindent
{\it Proof} of Proposition. \quad
The first step of the extensions is obtained
by looking at the action on the boundary,
and once we assume that
the action on the boundary is trivial,
the second extension is obtained by looking at the
action on the vertical line $\{0\}\times[0,\infty)$.
We can interpret it as an action on the leaf space.
\qed
In the above,
the first extension does not yield a non-abelian group.
Using the identification
$\mathit{Aut} R\cong \mathit{Aut}(\tilde R; T)/T^{\Bbb Z}$
in Corollary \ref{identification},
we obtain a better description
not only from the above point of view
but also from that of the question
whether the restriction map
$r_H:\mathit{Aut} R \to \mathit{Aut}_0 H$
admits a homomorphic section.
Note that $Z_\varphi$ admits a section to
$\mathit{Aut}(R, H) \subset \mathit{Aut} R$.
An element $f \in \mathit{Aut}(\tilde R; T)/T^{\Bbb Z}$
admits a presentation $f(z,x)=(az+b(x), \eta(x))$
up to $T^{\Bbb Z}$ where $T(z,x)=(\lambda z, \varphi(x))$.
Therefore ignoring $b(x)$ from this presentation
and assigning $f \mapsto (a,\eta)$ (mod $(\lambda, \varphi)^{\Bbb Z}$),
we obtain a surjective homomorphism
$\mathit{Aut}(\tilde R; T)/T^{\Bbb Z} \twoheadrightarrow
({\Bbb C}^*\times Z_\varphi)/(\lambda, \varphi)^{\Bbb Z}$ to an abelian group.
Also, by setting $b(x)=0$, we see this abelian group
can be realized as a subgroup of $\mathit{Aut}(\tilde R; T)/T^{\Bbb Z}$.
This enables us to describe the structure
of $\mathit{Aut} R$ as follows.
\begin{thm}\label{MainThm}{\rm
\quad
The automorphism group
$\mathit{Aut} R\cong \mathit{Aut}(\tilde R; T)/T^{\Bbb Z}$
is isomorphic to the semi-direct product
$$
{\mathcal K}_{\lambda, \varphi}\,
\rtimes \,\left\{
({\Bbb C}^*\times Z_\varphi)/(\lambda, \varphi)^{\Bbb Z}
\right\}
$$
where $a\in{\Bbb C}^*$ acts on $b(x)\in {\mathcal K}_{\lambda, \varphi}$
by multiplication
$b(x)\mapsto a^{-1}b(x)$, {\it i.e.}, the conjugation in
the affine transformations of each leaf,
and
$\eta\in Z_\varphi$ acts
by $b(x)\mapsto b(\eta(x))$.
}
\end{thm}
{\it proof.}\quad
Let us only verify the action of $a$.
The conjugation by the multiplication by $a$
is $[z\mapsto z+b(x)] \mapsto
[z \mapsto a^{-1}(az + b(x))=z+a^{-1}b(x)]$.
\hspace*{\fill} $\square$
\\
To close this section, consider the liftability of
$\mathit{Aut}_0 H$
to $\mathit{Aut} R$.
This is nothing but the liftability of the surjective
homomorphism
$$
({\Bbb C}^*\times Z_\varphi)/(\lambda, \varphi)^{\Bbb Z}
\twoheadrightarrow
{\Bbb C}^*/\lambda^{\Bbb Z} \, .
$$
Here we assume the continuity of splitting,
otherwise the question should include thinking about
non-continuous homomorphism
${\Bbb R} \to {\Bbb R}$ with $1\mapsto 1$.
If the centralizer $Z_\varphi$ is the total of ${\Bbb R}$,
it implies $Z_\varphi$
is a
$C^0$-family
of 1-parameter subgroup
$\{\eta_t\, ;\, t\in{\Bbb R}\}$
in $\mathit{Diff}^\infty([0,\infty))$
with $\varphi = \eta_1$.
Then we obtain easily a lift
defined as
$$
a (\mathrm{mod}\, \lambda^{\Bbb Z} )
\mapsto
(a, \eta_{t(a)})\,
(\mathrm{mod}\, (\lambda, \varphi)^{\Bbb Z} )\, ,
\quad
t(a)=\frac{\log \vert a\vert}
{\log\vert\lambda\vert}
\, .
$$
The converse is almost the same.
If we have a continuous lift
to $\mathit{Aut}(\tilde R; T)/T^{\Bbb Z}$,
choose a value of $\log \lambda$ and
look at the lift of a circle subgroup
$e^{t\log \lambda}$ ($0\leq t\leq 1$) to
a continuous path in
$\mathit{Aut}(\tilde R; T)$
starting from the identity.
Then its projection to $Z_\varphi$
gives rise to a 1-parameter family
in $Z_\varphi$ starting from the identity
which ends at
$\varphi$.
If this curve is smooth, it implies that
the Szekeres vector field $X_\varphi$ of
$\varphi$ is smooth.
Thus we obtain the following.
\begin{thm}\label{splitting}{\rm
\quad The restriction map
$r_H:\mathit{Aut} R \to \mathit{Aut}_0 H$ admits
a continuous [{\it resp.} smooth] homomorphic section
if and only if the centralizer $Z_\varphi$
is isomorphic to ${\Bbb R}$ as an ordered abelian group
[{\it resp.} the Szekeres vector field $X_\varphi$ is smooth].
In such cases,
$\mathit{Aut} R$ admits not only
a structure of twice semi-direct products
$$
\mathit{Aut} R \cong
({\mathcal K}_{\lambda, \varphi} \rtimes Z_\varphi)
\rtimes \mathit{Aut}_0 H
\cong
({\mathcal K}_{\lambda, \varphi} \rtimes {\Bbb R})
\rtimes {\Bbb R}^2/{\Bbb Z}^2 \, ,
$$
but also a structure of simple semi-direct product
$$
\mathit{Aut} R \cong
{\mathcal K}_{\lambda, \varphi} \rtimes
(\mathit{Aut}_0 H \times Z_\varphi)
\cong
{\mathcal K}_{\lambda, \varphi} \rtimes
({\Bbb R}^2/{\Bbb Z}^2 \times {\Bbb R}).
$$
of two abelian groups,
where the action of the right group on the left
is continuous with respect to the smooth topology
on ${\mathcal K}_{\lambda, \varphi}\cong{\mathcal Z}_{\lambda, \varphi} \subset C^\infty([0, \infty);{\Bbb C})$
in Case (2) and (3)
and ${\mathcal K}_{\lambda, \varphi}\cong{\Bbb C}$ or $\{0\}$ in Case (1)
[{\it resp}. smooth in a usual sense].
}
\end{thm}
\begin{rem}{\rm \quad
We saw that in Case (2) and (3)
the autmorphism groups are of infinite
dimension, while in Case (1) it is of finite
dimension and shows a similarity to
compact complex manifold.
In fact, in Case (3), not only any Reeb component
is realized in a Hopf surface
in Construction \ref{HopfSurface},
but also any of its automorphims
extends to the ambient Hopf surface.
Undr the setting $n=1$, the Hopf surface $W$
is obtained as $W=({\Bbb C}^2\setminus \{O\})/T''$
where $T''(z, w)=(\lambda\cdot z,\, \mu\cdot w)$.
Then a Levi-flat hypersurface $M= ({\Bbb C}\times{\Bbb R} \setminus\{O\})/T''$
contains a Reeb component $R$ of Case (3)
for $\lambda$ and $\mu\,$. It is easy to see that
the automorpshim
$(z, w)\mapsto (a\cdot z + b w^p\, , \, c\cdot w)\,$
of the Hopf surface $W$
restricts to $R$
where $a$, $c \in {\Bbb C}^*$ and $b\in{\Bbb C}$ are
arbitrary constants in the resonant case $\lambda=\mu^p$ and
$b=0$ in the non-resonant case.
Therefore any
of $\mathit{Aut} R$ extends to $W$.
On the other hand,
a recent work of Koike and Ogawa \cite{KO}
seems suggesting that Reeb componets of Case (3)
never appers in a Levi-flat hypersurfaces
in a complex surface.
Our result also mildly suggests that
the same might apply to Case (2).
Even in Case (1) it should be still confirmed
whether if the automorphism extends to the ambient surface
in the case where the Reeb component appears
as a part of a Levi-flat hypersurface which bounds
a Stein surface.
}
\end{rem}
\section{Reeb foliations}
The automorphism group
of a leafwise complex foliation
on a closed 3-manifold which consists of
two Reeb components is now easy to compute.
In this section we assume
that the relevant holonomy is infinitely
tangent to the identity at the origin,
because it must be so
except for two special cases where
two Reeb components are pasted in the same direction
of the holonomies or exactly in the inverse direction,
in both of which cases the pasting yields
$\,S^2 \times S^1\,$.
Let $R_{\varphi, \lambda}$ be the Reeb component
which we dealt with in the previous section.
For another pair of a diffeomorphism
$\psi \in \mathit{Diff}^\infty([0,\infty))$
which is also expanding and infinitely tangent to the identity
at the origin
and a constant
$\mu\in{\Bbb C}$ with $\vert\mu\vert>1$, take the Reeb component
${R}_{\psi, \mu}$ and let $\overline{R}_{\psi, \mu}$
denote the mirror of ${R}_{\psi, \mu}$, namely
the one which we obtain by reversing the
the transverse orientation.
It is done by replacing $x$ and $\varphi(x)$
with $-x$ and $-\varphi(-x)$ in the Hopf construction.
For example if $\lambda=\mu$ we can paste
$R_{\varphi, \lambda}$ and $\overline{R}_{\psi, \mu}$
along the common boundary $H={\Bbb C}^*/\lambda^{\Bbb Z}$
by the identity of $H$
to obtain a leafwise complex foliation on $S^2\times S^1$.
In general according to the pasting element
$\in \mathit{SL}(2;{\Bbb Z})$ we can choose appropriately
$\lambda$ and $\mu$ and paste them.
The foliation on $S^3$ obtained in such a way is called
the Reeb foliation.
Corollary \ref{extension} yields the following results.
\begin{thm}{\rm
\quad
Let $(M, {\cal F}, J)$ be a leafwise complex foliation
which is obtained by pasting
$R_{\varphi, \lambda}$ and $\overline{R}_{\psi, \mu}$.
Then its group of automorphism is
naturally isomorphic to the fibre product
of $\mathit{Aut} R_{\varphi, \lambda}$
and
$\mathit{Aut} \overline{R}_{\psi, \mu}$
with respect to $\mathit{Aut}_0 H$.
If the centralizer $Z_\psi$ is isomorphic to
${\Bbb R}$ as an ordered abelian group, then
$\mathit{Aut} R_{\varphi, \lambda}$ is continuously
realized as a subgroup in the resulting group of automorphisms.
}
\end{thm}
\begin{thm}{\rm
\quad
If $(M^3, {\cal F}_1, J_1)$ is obtained from
$(M^3, {\cal F}_0, J_0)$ by turbulization along
a closed transversal and the resulting Reeb component
is isomorphic to $\mathit{Aut} R_{\varphi, \lambda}$,
the group $\mathit{Aut}(M^3, {\cal F}_1, J_1)$ naturally contains
a subgroup which is isomorphic to
$\mathit{Aut} (R_{\varphi, \lambda}, H)$.
}
\end{thm}
\begin{rem}{\rm\quad
In both of above theorems, the automorphism group
contains an infinite dimensional vector space $\mathcal Z$
or one more copy.
Thus even in the case of closed manifolds,
the automorphism group of leafwise complex foliation
can be fairy large.
This presents a clear contrast between
a leafwise complex foliation on a compact manifold
and a complex structure on a compact manifold.
Due to Meersseman and Verjovsky \cite{MV}
in the study of moduli spaces,
they present similar features
as far as we deal with
tame leafwise complex foliation.
}
\end{rem}
\vspace{30pt}
\noindent
{\bf \Large Part II : Schr\"oder's equation on the half line}
\\
\noindent
We study the functional equations on the half line $[0, \infty)$
which appeared in Section 3.
The simpler one takes the form
$$
\beta\circ \varphi (x) = \lambda \beta(x)
$$
for a fixed diffeomorphism
$\varphi$ and a constant $\lambda$.
Ernst Schr\"oder
started to study
a similar (in fact, formally the same)
functional equation on the unit disk $D$
in the complex plane ${\Bbb C}$
under the complex analytic setting
in \cite{Sch} in the late 19th century.
Not only because
it is just the natural eigenvalue problem
for the pull-back operator
to look for $\beta$ and $\lambda$
for a given $\varphi$,
also Schr\"oder initiated complex dynamical studies
and was interested in the iteration of compositions
maybe in the context of Newton's method.
According to the development of the complex dynamics
the problems that were treated in these epoch
has become fairy well-understood.
Recently the studies in this direction seem to be
aiming at higher dimensional cases.
For the history of an early stage of the complex dynamics
and Schr\"oder's functional equation,
we have two nice references [Al1, 2].
Our aim is to solve the equations
\begin{eqnarray*}
{\mathrm{Equation \,\, (I)\,}} &:& \beta\circ \varphi (x)
= \lambda \beta(x)
\\
{\mathrm{Equation \, (II)}} &:&
\beta_1\circ \varphi (x) = \lambda \beta_1(x),
\quad
\beta_2\circ \varphi (x)
= \lambda \beta_2(x) + \beta_1(x)
\end{eqnarray*}
on the half line $[0, \infty)$
for an expanding diffeomorphism
$\varphi \in {\mathit{Diff}}^\infty([0, \infty))$
and a complex constant $\lambda$ with $|\lambda | >1$.
Equation (II) is generalized to still higher dimensional case
(II') which is expressed by using vector notations as
$$
\qquad \!\!\!
{\mathrm{Equation\,(II')}}\,\,\, :\, \, \,
\boldsymbol\beta\circ \varphi(x)
=A \boldsymbol\beta(x) \qquad \qquad \qquad \qquad \qquad \qquad
\qquad \quad
$$
where $\varphi$ is as above,
$A=(a_{ij})$ is an $M\times M$ matrix
any of its eigenvalues
has the absolute value greater than $1$,
and
$\boldsymbol\beta(x)
= {}^t(\beta_1(x), \cdots , \beta_M(x))
\in C^\infty([0, \infty); {\Bbb C}^N)$
is the unknown function.
Of course, the problem is easily reduced to the case
where $A$ is a single non-trivial Jordan block.
So we can assume that $A$ has a unique eigenvalue
$\lambda$ as above and eventually
$a_{ij}=\lambda$ for $i=j$, $a_{ij}=1$ for $i=j+1$,
and $a_{ij}=0$ otherwise.
It is easily seen that in the case of $\varphi'(0)=1$,
which is of our main concern,
there is no analytic solution but
$\beta(x)$ or $\boldsymbol\beta(x)\equiv 0$.
On the other hand, if $\varphi'(0)>1$,
for Equation (I) the space of solution is trivial or
of dimension 1, depending on the resonance of
$\lambda$ and $\varphi'(0)$.
These are in fact exactly the same even when
working on $D\in {\Bbb C}$.
Therefore Schr\"oder's equation
exhibits very distinctive feature
when it is considered on the half line
with $\varphi'(0)=1$.
The space of solutions to Equation (I)
turns out to be an infinite dimensional vector space
which is in a sense
isomorphic to $C^\infty(S^1; {\Bbb C})$
whenever $\varphi'(0)=1$.
In the subsequent sections,
first we describe the space of solutions much clearer,
and then we give two different proofs.
It is to be remarked that as $\varphi$ is expanding,
our problem is essentially that on the germs around $x=0$.
For a given germ of $\varphi$,
we can extend $\varphi$ to the whole of $[0, \infty)$
as a realization of the germ as far as $\,\,\varphi(x)>x\,\,$
is satisfied for $\,x>0\,$.
Then the same applies to $\beta(x)$
because once it is given as a germ around $x=0$,
it is automatically and uniquely extended to the whole half line
by the equation itself.
The results on Equation (I)
are used in Part I of this paper.
Those for (II) serve in \cite{Ho} to determine
the automophism groups of Reeb components
with complex leaves of complex dimension 2.
When we extend our results on the automorphism groups
to higher dimensional cases,
the results on (II') are necessary.
\section{\large The space of solutions}
In this section we give precise statements of our results
on the Schr\"oder type functional equations (I), (II), and (II')
and describe the spaces of their smooth solutions.
\\
For an expanding diffeomorphism
$\varphi\in\mathit{Diff}^\infty([0, \infty))$
and a complex number $\lambda$ with $\vert\lambda\vert > 1$,
we consider Equation (I), (II), and (II').
First consider these equations
(not on the whole $[0, \infty)$ but) on $(0,\infty)$.
Then, Equation (I) has a lot of solutions
and if we fix any solution
$\beta^*(x)\in C^\infty((0,\infty); {\Bbb C})$
which never vanishes, {\it i.e.}, $\beta^*(x)\ne 0$
for $x>0$,
then each solution corresponds to a
smooth function on $S^1=(0,\infty)/\varphi^{\Bbb Z}$
by assigning $\beta \mapsto \beta/\beta^*$.
This gives a bijective correspondence between
the space ${\mathcal Z}={\mathcal Z}_{\lambda, \varphi}$ of solutions to (I)
on $(0, \infty)$
and $C^\infty(S^1; {\Bbb C})$ as vector space.
This correspondence will be more precise
in Section \ref{Fourier}
in fixing the coordinate on the circle.
In Section \ref{Fourier},
we will make this correspondence
more precise,
fixing
the coordinate on $\,S^1\,$ and the choice of
$\,\beta^*\,$.
Also take the space ${\mathcal S}={\mathcal S}_{\lambda, \varphi}$ of solutions to Equation (II)
on $(0, \infty)$. If we assign $\beta_1$ to a solution
$(\beta_1, \beta_2) \in \mathcal S$, we obtain the projection
$P_1: \mathcal S \to \mathcal Z$.
Here the kernel
of $P_2$ is nothing but $\mathcal Z$.
We also see that the projection $P_1$ is surjective because
for any $\beta_1\in \mathcal Z$
$$
\beta_2(x)= \frac{1}{\lambda\log\lambda} \beta_1(x)\log\beta^*(x)
$$
gives a solution $(\beta_1, \beta_2)\in \mathcal S$,
where for $\log\lambda$ and $\log\beta^*(x)$
any (smooth) branch can be taken.
Therefore, as a vector space,
$\mathcal S$ has a structure such that
$$
0 \to \mathcal Z \to \mathcal S \to \mathcal Z \to 0
$$
is a short exact sequence.
For Equation (II') we simply repeat this extension.
Let
${\mathcal S}_M$
denote the set of solutions on $(0, \infty)$.
For
$1\leq m<M$,
${\mathcal S}_m$
is naturally identified with
a quotient
$\{{}^t(\beta_1,\, \cdots , \, \beta_m) \,\vert\,
{}^t(\beta_1, \, \cdots ,\, \beta_M) \in {\mathcal S}_M
\}$
of ${\mathcal S}_n$.
Each projection
$
{\mathcal S}_m \to {\mathcal S}_{m-1}
$
is surjective because
the multiplication
$
\times \frac{1}{\lambda\log\lambda}\beta^*
$
is a linear right inverse
and its kernel coincides with $\mathcal Z$.
${\mathcal S}_1$ is nothing but $\mathcal Z$ and
${\mathcal S}_2$ the above ${\mathcal S}$ as well.
Let $\,\varphi^*\,$ denote the pull-back by $\,\varphi\,$.
Then Equation (I) is expressed as
$\,\,(\varphi^* - \lambda)\beta=0\,\,$.
Now, Equation (II) is nothing but
$\,\,(\varphi^* - \lambda)^2\beta_2=0\,$,
where we put $\,\,\beta_1\,\,$
by setting
$\,\,(\varphi^* - \lambda)\beta_2=\beta_1\,$.
Inductively we see Equation (II')
is nothing but
$\,\,(\varphi^* - \lambda)^M\beta_M=0\,\,$ while
$\,\,(\varphi^* - \lambda)\beta_m=\beta_{m-1}\,\,$
for $\,m=2,3,4, \cdots, M\,$ might also be regarded as
auxiliary equations. \vspace{8pt}
As in Section 3 of Part I,
we devide the situation into the following three cases
according to the nature of the jet of $\varphi$ at $x=0$.
For the second and the third cases,
the statements of our results are the same.
Here, $f^{(i)}$ denotes the $i$-th derivative of $f(x)$
and $j^if(0)$ denotes the $i$-th jet at $x=0$.
\vspace{5pt}
\\
\noindent
\quad Case (1) : $\varphi'(0)=\mu >1$ ,
\vspace{2pt}
\\
\noindent
\quad Case (2) : for some $n \geq 2$ ,
$j^{n-1}\varphi(0)=j^{n-1}\mathit{id}(0)\,$
and
$\,\,\varphi^{(n)}(0)>0$ ,
\vspace{2pt}
\\
\noindent
\quad Case (3) :
$j^\infty\varphi(0)=j^\infty\mathit{id}(0)\,$ .
\vspace{8pt}
Let us state the results. We start with the easiest case.
\begin{thm}\label{linear expansion}{\rm
\quad Consider Case (1).
\\
1) (Resonant case)\quad If $\,\,\lambda=\mu^n\,\,$
is satisfied for some $\,n\in{\Bbb N}$ ,
then the space of solutions ${\cal K}$
to Equation (I)
is a complex vector space of dimension 1.
For Equation (II'), $\,\boldsymbol\beta\,\,$ satisfies
(II') if and only if
$\,\,\beta_1= \cdots =\beta_{n-1}\equiv 0\,\,$ and
$\,\,\beta_n\in{\cal K}\,\,$ hold.
Therefore in total the space of solution is also
1-dimensional.
\vspace{3pt}
\\
2)(Non-resonant case)\quad If no positive integer
$n\in{\Bbb N}$ satisfies $\lambda=\mu^n$,
then there exists no solution to (I) but
$\beta(x)\equiv 0$, and we have ${\cal K}=\{0\}$.
For Equation (II') the same applies.
Namely the only smooth solutions on $\,[0,\infty)\,$
is $\,\,\boldsymbol\beta(x)\equiv 0$ .
}\end{thm}
Remark here that if $\varphi(x)=\mu x$,
the solution to (I) in resonant case
is nothing but $\beta(x)=\mathrm{const}\cdot x^n$ .
Also remark that accordingly $\lambda$ is a
positive real number.
This result is so easy that the proof is given here.
\vspace{8pt}
\\
\noindent
{\it Proof.} \quad
From Sternberg's linearization theorem \cite{St},
there exists a diffeomorphism
$\,h\in{\mathit{Diff}}^\infty([0, \infty))\, $
which conjugates $\varphi$ into the linear one
$\, h^{-1}\circ\varphi\circ h(x)=\mu x\,$.
Therefore in solving the equations,
from the first we can assume
$\,\varphi(x)=\mu x\,$ .
Therefore the equation (I) takes the following form.
$$
\beta(\mu x)=\lambda\beta(x)\qquad
\mathrm{for}\,\,\, x\in [0,\infty)\, .
$$
In both cases,
by differentiating Equation (I) for arbitrary many times
at $x=0$, we see that the Taylor expansion at $x=0$ can be
non-trivial only at the degree $n=\log\lambda/\log\mu$.
Therefore in the resonant case,
$\beta(x)$ must be in the form
$\beta(x)=c\cdot x^n +f(x)$ where $f(x)$ is a flat function,
and in the non-resonant case, the same form with $c=0\,$.
Then, in the resonant case,
as $c\cdot x^n$ is a solution to (I), so is $f(x)\,$.
Therefore in both cases, it is enough to check that
any non-trivial flat function can no be a solution.
If we had a non-trivial flat
solution $f(x)\,$,
it would contradict as follows.
Take $x_0 \in (0, \infty)$ with $f(x_0)\ne 0$ and
look at $f(\mu^{-k}x_0)=\lambda^{-k}f(x_0)$ for $k\in{\Bbb N}$.
On the other hand, as $f$ is flat we have
$\lim_{x\to 0} f(x)/x^l =0$ for any $l\in {\Bbb N}$.
So large enough $l$ ($\geq |\log\lambda/\log\mu|$)
gives rise to a contradiction.
This is well-known also as a fact (even for
higher dimensional case)
that a weighted-homogeneous function
is smooth at the origin only when
it is a polynomial.
For Equation (II), from the above result,
we assume $\beta_1(x)=cx^n$ in the resonant case.
Then a similar computation for $\beta_2$
implies $c=0$. Therefore we have $\beta_1=0$ and thus
$\beta_2=c'x^n$ for some $c'\in {\Bbb C}$.
In the non-resonant case,
the argument for (I) suffices.
For (II'), the argument for (II) works as an inductive step.
\qed
The results for Case (2) and (3)
can be stated together.
\begin{thm}\label{Main}\quad{\rm
For both of Case (2) and Case (3),
and for any $\,\,\lambda\in {\Bbb C}\,\,$ with $\,\,|\lambda|>1\,$,
the followings hold. \vspace{3pt}
\\
1) Any solution $\,\,\beta\in\mathcal Z_{\lambda, \varphi}\,$
to Equation (I)
on $\,\,(0, \infty)\,\,$
extends to $\,\,[0,\infty)\,\,$
so as to be a smooth function which is flat at $\,x=0\,$,
{\it i.e.}, the $k$-th jet satisfies
$\,\,j^k\beta(0)=0\,\,$ for any $\,k=0,1,2, \,\cdots$.
In other words, the space $\,\,{\cal K}\,\,$ of all solutions to (I)
considered on $\,\,[0,\infty)\,\,$
coincides with $\,\,\mathcal Z_{\lambda, \varphi}\,\,$.
\vspace{3pt}
\\
2) The same applies to Equation (II) and (II').
Namely by putting $\,{\boldsymbol\beta}(0)=0\,$,
any $\,\,{\boldsymbol\beta}\in {\mathcal S}_M\,\,$
is smooth and flat at $\,x=0\,$.
}
\end{thm}
In the next section we give a proof for Case (3),
which is more direct and simpler than
one given in the final section.
In this proof,
we directly estimate the derivatives of any order
of $\beta\circ\varphi$ for $\beta \in \mathcal Z$.
The higher derivatives of a composite function
is complicated and described in the formula
of Fa\`a di Bruno. We do not need its full length.
Unfortunately this method does not work for Case (2).
A proof by a different approach for Case (2) is
given in the following section.
It is based on two theories.
One is Takens' normal forms \cite{Ta}
for germs around the origin
of $\,{\mathit{Diff}}^\infty([0,\infty))\,$
and of vector fields on $[0, \infty)$
with non-trivial finite order jets.
The other one is a classical theory of
Fourier expansion of $C^\infty(S^1; {\Bbb C})$.
Unfortunately again, this method
seems difficult to apply to Case (3).
Then in the final section we give a proof
relying on the center manifold theory,
which covers both of Case (2) and (3).
This proof is suggested by Masayuki Asaoka.
It might be worth remarking that
when a proposition is proved in the framework of
hyperbolic dynamical systems,
quite often it is also proved in Fourier analysis,
and vice versa.
Here we might observe a similar phenomenon.
\section{\large Direct proof for Case (3)}
We prove Theorem \ref{Main} for Case (3),
namely,
in the case where $\varphi$ is flat to the identity,
by a direct estimate of the derivatives of $\beta(x)$
of an arbitrary order when $x \to 0$.
In order to clarify the strategy
it might be suggested to the readers to check
$\displaystyle\lim_{x\to +0}\beta(x)=0$
and
$\displaystyle\lim_{x\to +0}\beta'(x)=0\,$,
{\it i.e.}, for $k=0, 1$,
which are easy and reviewed in Proposition \ref{1jet},
and then the the second jet $k=2$).
Looking at up to the case $k=3$ might make
the roll of the following lemma clearer.
\begin{lem}\label{Faa di Bruno}\quad{\rm
The $n$-th derivative $\{\beta(\varphi(x))\}^{(n)}$
is written in the following form for $n\in{\Bbb N}$.
$$
\{\beta(\varphi(x))\}^{(n)}
=(\varphi'(x))^n \cdot\beta^{(n)}(\varphi(x))
+
\sum_{k=1}^{n-1}\Phi_{n,k}\cdot\beta^{(k)}(\varphi(x))\, .
$$
Here, $\Phi_{n,k}$ is an integral polynomial in
$\varphi'(x)$, $\varphi''(x)$,$\cdots$, $\varphi^{(n)}(x)$,
without constant term and no term is of monomial
only in $\varphi'(x)$.
}
\end{lem}
This lemma is easily seen by the induction,
but in fact it is a corollary to
the well-known formula of Fa\`a di Bruno
({\it e.g.}, see \cite{Ri}, \cite{Ro}, or textbooks on calculus).
It is independent of our assumption on $\varphi$ and
is valid for any composite functions.
On the other hand the flatness of $\varphi$ at the origin
implies
$(\varphi'(x))^n \to 1$ and $\Phi_{n,k}\to 0$
when $x \to 0+0$.
\\
Now let us prove 1) of Theorem \ref{Main}. Let $\beta$ be
a solution to (I) on $(0,\infty)$.
From the equation it is easy to see that $\beta(x)\to 0$
when $x \to 0+0$.
Now fix any integer $N$.
$\beta'(x)\to 0$ is also easy to see, but for higher
derivatives, in a natural estimate the lower derivatives
are involved. Thus the basic strategy is not to estimate
the higher derivatives by induction on the order,
but to estimate them all together up to the
fixed order $N$.
From Equation (I) and the above lemma
we have the following computation.
\begin{eqnarray*}
\sum_{n=1}^{N}\vert\beta^{(n)}(x)\vert
&=&
\frac{1}{\vert\lambda\vert}
\sum_{n=1}^{N}
\left\vert
\{\beta(\varphi(x))\}^{(n)}
\right\vert
\\
&\leq&
\frac{1}{\vert\lambda\vert}
\sum_{n=1}^{N}
\left\{
(\varphi'(x))^n \cdot\vert\beta^{(n)}(\varphi(x))\vert
+
\sum_{k=1}^{n-1}\vert\Phi_{n,k}\vert
\cdot
\vert\beta^{(k)}(\varphi(x))\vert
\right\}
\\
&\leq&
\frac{1}{\vert\lambda\vert}
\sum_{k=1}^{N}
\left((\varphi'(x))^k +
\sum_{n=k+1}^{N}\vert\Phi_{n,k}
\vert\right)
\cdot
\vert\beta^{(k)}(\varphi(x))\vert
\end{eqnarray*}
As is remarked above, we know $(\varphi'(x))^k \to 1$
and
$\sum_{n=k+1}^{N}\vert\Phi_{n,k}
\vert \to 0$ when $x\to 0$. Therefore there exists
$b_N>0$ such that for $x\in (0,b_N]$
we have
$$(\varphi'(x))^k +
\sum_{n=k+1}^{N}\vert\Phi_{n,k}
\vert \leq \sqrt{\vert\lambda\vert}
\quad
\mathrm{for}\,\,\,\, k=1, 2, \cdots , N .
$$
This implies for any $x \in (0, b_N]$
$$
\sum_{n=1}^{N} \vert \beta^{(n)}(x) \vert
\leq
\frac{1}{\sqrt{\vert\lambda\vert}}
\sum_{n=1}^{N} \vert \beta^{(n)}(\varphi(x)) \vert .
$$
Put
M=\max
\{
\sum_{n=1}^{N} \vert \beta^{(n)}(x) \vert
\,; \,
x\in[b_N, \varphi(b_N)]
\}
$
and define $m(x)\in{\Bbb N}$ for $x\in (0, b_N)$
so that
$\varphi^{m(x)}\in[b_N, \varphi(b_N))$.
Then, the above inequality implies
$$
\sum_{n=1}^{N} \vert \beta^{(n)}(x) \vert
\leq
M\cdot \sqrt{\vert\lambda\vert\,}^{-m(x)}
$$
for $x\in (0, b_N)$.
Because `$x\to 0+0$' is equivalent to `$m(x)\to\infty$',
we obtained the convergence
$$
\beta^{(n)}(x) \to 0 \quad (x\to 0+0)
\qquad \mathrm{for} \quad n=1,\cdots, N .
$$
This completes the proof of 1).
\\
Let us outline the proof of 2) for $M=2$.
We extend the basic strategy of the proof of 1)
in the following sense. When we estimate the derivatives
of $\beta_2$, naturally those of $\beta_1$ are involved.
Therefore we will estimate the derivatives of $\beta_2$
and $\beta_1$ all together up to a fixed order $N$,
even though the flatness of $\beta_1$ is already proved in 1).
First we fix
$\varepsilon>0$ so small that
$\displaystyle
\varepsilon \leq
\vert\lambda\vert^{\frac{5}{4}}-
\vert\lambda\vert
$
is satisfied.
Now take any solution $(\beta_1, \beta_2) \in \mathcal S$
and put $\tilde\beta_1=\varepsilon^{-1}\beta_1$.
Then instead of Equation (II), $\tilde\beta_1$ and $\beta_2$ satisfy
\vspace{5pt}
\\
$
\quad \mathrm{Equation\,\, (\tilde{II})} :\quad
\tilde\beta_1(\varphi(x))=\lambda \tilde\beta_1(x)
, \;\;\;
\beta_2(\varphi(x))
=\lambda \beta_2(x)+ \varepsilon\tilde\beta_1(x)
\, .
$
\vspace{5pt}
\\
Then, from ($\tilde{\mathrm{II}}$) we have
\begin{eqnarray*}
\frac{e^{i\theta}\lambda - \varepsilon}
{\lambda}\tilde\beta_1(\varphi(x))
+
\beta_2(\varphi(x))
&=&
e^{i\theta}\lambda\tilde\beta_1(x)
+
\lambda\beta_2(x)
\end{eqnarray*}
and consider the $n$-th derivatives of both sides.
For any $\theta\in{\Bbb R}$ and $n=1, \cdots , N$, we have
$$
\vert e^{i\theta}\tilde\beta_1^{(n)}(x) + \beta_2^{(n)}(x)\vert
\leq
\frac{1}{\vert\lambda\vert}
\left(
\frac{\vert\lambda\vert + \varepsilon}{\vert\lambda\vert}
\vert \{\tilde\beta_1(\varphi(x))\}^{(n)}\vert
+
\vert \{\beta_2(\varphi(x))\}^{(n)}\vert\right)
$$
Because the right hand side is independent of $\theta$,
using the inequality
$$
\left\vert
\frac{e^{i\theta}\lambda - \varepsilon}{\lambda}
\right\vert
\leq
\frac{\vert\lambda\vert + \varepsilon}{\vert\lambda\vert}
\leq
\vert\lambda\vert^\frac{1}{4}
\quad
\mathrm{for\,\,\, any}\,\,\, \theta\in{\Bbb R}
$$
we obtain
$$
\vert \tilde\beta_1^{(n)}(x)\vert + \vert\beta_2^{(n)}(x)\vert
\leq
\frac{1}{\vert\lambda\vert^\frac{3}{4}}
\left(\vert \{\tilde\beta_1(\varphi(x))\}^{(n)}\vert
+
\vert \{\beta_2(\varphi(x))\}^{(n)}\vert\right).
$$
Applying Lemma \ref{Faa di Bruno} to $\tilde\beta_1(\varphi(x))$ and
to $\beta_2(\varphi(x))$
for $n=1, \cdots , N$,
from the same argument as in 1) we obtain
$$
\sum_{n=1}^{N}
\left(\vert \tilde\beta_1^{(n)}(x)\vert + \vert \beta_2^{(n)}(x)\vert\right)
\leq
\frac{1}{\vert\lambda\vert^\frac{1}{4}}
\sum_{n=1}^{N}
\left(\vert \tilde\beta_1^{(n)}(\varphi(x))\vert
+ \vert \beta_2^{(n)}(\varphi(x))\vert\right)
$$
for $x\in (0, b_N]$, where $b_N$ is exactly the same as in the proof of 1).
\hspace*{\fill}$\square$
\\
Now it is almost straight forward to further generalize
this proof for Equation (II').
\section{\large Proof by
Fourier series for Case (2)}\label{Fourier}
A proof of Theorem \ref{Main} for Case (2) is given here.
It relies on two big tools.
The first one is
Takens' normal form which plays a similar roll
as Sternberg's linearization in the proof of
Theorem \ref{linear expansion}.
The second one is a classical Fourier expansion/series
of smooth functions on the circle.
Takens' normal form enables us to
consider (countably many) simple linear homogeneous
ordinary differential equations instead of
considering Equation (I).
Equation (II) and (II') correspond to inhomogeneous
or vector valued case.
As we will see below,
we have an ODE for each choice of the value of
$\,\log \lambda\,$ ,
Our functional equation (I) and
countably many ODE's are related
by Fourier expansion and series.
\subsection{
Takens' normal form and Fourier basis
}
\begin{thm}\label{Takens}
{\rm(Takens, \cite{Ta})\quad
Let $\,\varphi \in {\mathit{Diff}}^\infty([0, \infty))\,$
be in Case (2). \vspace{3pt}
\\
1)\quad
There exists a diffeomorphism
$h \in \mathit{Diff}^\infty([0,\infty))$
which conjugates $\varphi$ into a diffeomorphism
$\psi\in \mathit{Diff}^\infty([0,\infty))$
of the following polynomial type
on $[0, x_1]$ ($\exists x_1>0$)
$$
\psi(x)=x + x^n + \alpha x^{2n-1}
\quad \mathrm{and} \quad
\psi = h^{-1}\circ\varphi\circ h \, .
$$
The coefficient $\alpha\in{\Bbb R}$
is determined by the $(2n-1)$-jet of $\varphi$ at $x=0$.
\vspace{3pt}
\\
\noindent
2) \quad
There also exists a diffeomorphism
$k \in \mathit{Diff}^\infty([0,\infty))$ which conjugates
$\varphi$
in a neighborhood of the origin
into
the exponential map
$$
\exp
X
= k^{-1}\circ\varphi\circ k \, .
$$
of the vector field
$$
X=\rho(x) \frac{d}{dx}\, ,
\quad
\rho(x)= x^n + a x^{2n-1} \, .
$$
The the coefficients $\alpha$ in 1) and
$a$ here are related by $a=\alpha - n/2\,$.
}
\end{thm}
\begin{rem}
{\rm \quad 1) \quad
The second statement follows from the first one,
because a direct computation shows
$$
\exp(x^n + a x^{2n-1})
\frac{d}{dx} = x + x^n + \left(a+\frac{n}{2}\right)x^{2n-1}
+ [\mathrm{\,higher\,\, order\,\, terms}]\, .
$$
2)\quad Takens also gave a normal form for vector fields.
It takes almost the same form but we do not need it here.
}
\end{rem}
Thanks to Takens' theorem,
we can conjugate our equations by a smooth diffeomorphism
and are allowed to assume that the holonomy
$\varphi$ is of the form
$$
\varphi=\exp
\, ,
\quad
=\rh
(x)\frac{d}{dx}\, ,
\quad
\rh
(x)=x^n + a x^{2n-1}\,\,
\mathrm{on} \,\,
[0, x_0]
$$
for some $n\geq 2$, $a \in {\Bbb R}$, and $x_0>0\,$.
We also assume that
$\,\rh
(x)>0\,$ on $\,(0, \infty)\,$ and
$\,\rh
(x)\equiv 1\,$ on $\,(x_1, \infty)\,$
for some $x_1>x_0\,$.
\vspace{8pt}
We consider the following ordinary differential equation on
$(0, \infty)$
\vspace{8pt}
\\
\qquad Equation (I-$\Lambda$) : \quad \qquad
$\beta'(x)=\frac{\Lambda }{\rho(x)}\beta(x)\, .
$
\vspace{8pt}
\\
\noindent
This is of course equivalent
to the following ODE in the variable $t$.
$$
\left.\frac{d}{dt}\right\vert_{t=0}
\beta(\exp(tX)(x))=\Lambda\cdot\beta(x)
$$
Therefore any solution $\beta$ is presented as
$\beta(\exp(tX)(x_0))=e^{\Lambda t}\cdot\beta(x_0)
=C\cdot e^{\Lambda t}$
for a constant $C\in{\Bbb C}$.
It is also clear that $\beta$ satisfies
Equation (I) on
$(0, \infty)$.
In the variable $x$,
$\beta(x)$ is presented as
$\beta(x)=C \cdot
\exp\left(
\Lambda \int_{x_0}^x \frac{1}{\rho(u)}
du\right)
$.
In particular on $(0, x_0)$, we have
$$\beta(x)=C \cdot
\exp\left(
(R + i\,{\mathrm{Im} \Lambda})
\int_{x_0}^x \frac{1}{u^n(1+au^{n-1})}
du
\right)
$$
where the real part
$R=\log\vert\lambda\vert$
is positive.
Now we choose $\Lambda_{\langle 0 \rangle}$
as a value of $\log \lambda$
and fix it.
Then other general values of $\log\lambda$
are given as $\Lambda_{\langle l \rangle}= R +i\theta_l$,
$\theta_l =\theta_0 + 2l\pi$ for $l\in {\Bbb Z}$.
\vspace{8pt}
Here we make the correspondence between
the space ${\mathcal Z}_{\lambda, \varphi}$ of solutions of (I) considered on
$(0, \infty)$ and $C^\infty(S^1;{\Bbb C})$ more precise.
As a coordinate on the circle $S^1$ we take
$\theta = t$ (mod $2\pi$) where
$x(t)=\beta(\exp(tX)(x_0))$ is assumed.
Now for each $l\in {\Bbb Z}$,
let $\beta_{\langle l \rangle}$ denote the solution to
the ODE (I-$\Lambda_{\langle l \rangle}$)
which satisfies $\beta_{\langle l \rangle}(x_0)=1$.
Therefore we easily know that
$\beta_{\langle l \rangle}(x(t))=
e^{2\pi l t \cdot i}\cdot
\beta_{\langle 0 \rangle}(x(t))\,$.
Take $\beta_{\langle 0 \rangle}$ as $\beta^*$
in defining the correspondence.
Then $\beta_{\langle 0 \rangle}$ corresponds
to the constant function $1$ on $S^1$ and in general
$\beta_{\langle l \rangle}$
corresponds to
$\check{\beta}_{\langle l \rangle} \in C^\infty(S^1;{\Bbb C})\,$,
namely,
$$
\check{\beta}_{\langle l \rangle}(\theta)=e^{2\pi l \theta \cdot i}
\qquad \mathrm{for}\,\,\, l \in {\Bbb Z}\, ,
$$
so that
$\check{\beta}_{\langle l \rangle}$'s ($l \in {\Bbb Z}$)
form the standard Fourier basis
for $C^\infty(S^1; {\Bbb C})$.
The following is well-known and well fits into our situation.
\begin{thm}{\rm (see {\it e.g.}, \cite{Ka})\quad
The infinite sum with coefficients $c_l\in{\Bbb C}$
$$
\sum_{l=-\infty}^{\infty}c_l\cdot e^{i\theta}
$$
defines a smooth function on
$\theta\in S^1={\Bbb R}/2\pi{\Bbb Z}\,\,\,$
if and only if\,\, the sequence of coefficients
$\{c_k\}_{k\in{\Bbb Z}}$
is rapidly decreasing, namely it satisfies
$$
\sum_{l=-\infty}^{\infty}|l|^d|c_l|<\infty
\quad \mathrm{for\,\, any}\,\,\, d\in{\Bbb N}
\, .
$$
}\end{thm}
Therefore any $\beta\in{\mathcal Z}_{\lambda, \varphi}$ is given as
an infinite sum
$$
\beta=\sum_{l=-\infty}^{\infty}c_l\cdot
\beta_{\langle l \rangle}
$$
with a rapidly decreasing sequence of coefficients
$\{c_k\}_{k\in{\Bbb Z}}$.
\subsection{Proof of Main Theorem}
What we have to prove in this section
is stated as follows.
\begin{thm}\label{restatement}
{\rm\quad
For any rapidly decreasing
$\{c_k\}_{k\in{\Bbb Z}}\,$,
$\beta=\sum_{l=-\infty}^{\infty}c_l\cdot
\beta_{\langle l \rangle}$
is extended to
$[0,\infty)$ as $\beta(0)=0$
and is smooth and flat at $x=0$.
}
\end{thm}
Let us verify this for each base.
\begin{prop}{\rm\quad
The solution $\beta(x)$ to (I-$\Lambda$) is extended to
$[0,\infty)$ as $\beta(0)=0$,
and then $\beta(x)$ is smooth and
flat at $x=0$.
}
\end{prop}
\noindent
{\it Proof.}\quad
It is easy to compute the integration but
we only need to remark
that for some $\delta>0$ and any $x\in (0, \delta)$ we have
$$
\left|\int_{x_0}^x \frac{1}{u^n(1+au^{n-1})}du\right|
\geq
\left|\frac{1}{2x}\right|\, .
$$
The derivative of $\beta$ of order $k\in {\Bbb N}$
is a multiplication of $\beta$ and some rational function in
the variable $x$.
Therefore for any $k \in {\Bbb N}$ we have
$$
\vert\beta^{(k)}(x)\vert
\leq
\vert \mathrm{a\,\, rational\,\, function} \vert
\times \exp(-\frac{1}{2x})
\, \to 0 \,\, (x\to 0)
$$
which suffices to show
the smoothness and flatness of $\,\beta\,$ at $\,x=0\,$.
\vspace{8pt}
\qed
In order to proceed further,
we need to take
a slightly closer look at those rational functions.
Recall that $n$, $a$, and $\lambda$ are already fixed.
\begin{lem}{\rm\quad
For $k\in {\Bbb N}$ and $j=1, \cdots , k$, there exists a
fixed polynomial $Q_{k,j}(x)$ which satisfies
on $(0, x_0)$
$$
\beta_{\langle l \rangle}^{(k)}(x)=
\left\{
\frac{1}{P(x)^k}\sum_{j=1}^k Q_{k,j}(x)(R +i\theta_l)^j
\right\}
\beta_{\langle l \rangle}(x),
\quad
P(x)=x^n + a x^{2n-1}
$$
and $Q_{k,j}(x)$ is a linear combination of
multiplications of $(k-j)$-many of
$P(x)$, $P'(x)$, $\cdots$, $P^{(k-j)}(x)$,
with total degree of differentiation $(k-j)$.
}\end{lem}
For example,
$Q_{k,k}(x)=1$, $Q_{k,k-1}(x)=k(1-k)P'(x)/2$, and so on.
The lemma is easily proved by the induction on $k$.
\vspace{5pt}
Let us develop $(R +i\theta_l)^j$ into
a polynomial of $l$ as follows.
$$
(R +i\theta_l)^j
=
(R +i(\theta_0+2\pi l))^j
=\sum_{d=0}^jR_{j,d}l^d
$$
Here the constants $R_{j,d}$
($j\in{\Bbb N}$, $d=0$, $\cdots$, $j$)
are determined by
$R$ and $\theta_0$.
\vspace{8pt}
\\
\noindent
{\it Proof} of Theorem \ref{restatement}. \quad
For a rapidly decreasing sequence
$$
\{c_k\}_{k\in{\Bbb Z}}
\quad
\mathrm{with}
\quad
\sum_{l=-\infty}^{\infty}|l|^d|c_l| = M_d < \infty
\,\,\,
\mathrm{for}\,\, \forall d \in {\Bbb N}\cup\{0\}
$$
take
$
\beta(x)=\sum_{l=-\infty}^{\infty}c_l\cdot
\beta_{\langle l \rangle}(x)
$.
Then we have the following estimate;
\begin{eqnarray*}
|\beta^{(k)}(x)|
&=&
\left|
\sum_{l=-\infty}^{\infty}c_l\cdot
\beta_{\langle l \rangle}^{(k)}(x)
\right|
\\
&=&
\left|
\sum_{l=-\infty}^{\infty}c_l\cdot
\left\{
\frac{1}{P(x)^k}\sum_{j=1}^k Q_{k,j}(x)(R +i\theta_l)^j
\right\}
\beta_{\langle l \rangle}(x)
\right|
\\
&\leq&
\left|
\frac{1}{P(x)^k}
\right|
\left|
\sum_{j=1}^k Q_{k,j}(x)
\sum_{l=-\infty}^{\infty}c_
\left(
\sum_{d=0}^jR_{j,d}l^d
\right)
\beta_{\langle l \rangle}(x)
\right|
\\
&=&
\left|
\frac{1}{P(x)^k}
\right|
\left\{
\sum_{j=1}^k
\sum_{d=0}^j
Q_{k,j}(x)
R_{j,d}
\left(
\sum_{l=-\infty}^{\infty}c_l\cdot l^d
\right)
\beta_{\langle l \rangle}(x)
\right|
\\
&\leq&
\left\{
\left|
\frac{1}{P(x)^k}
\right|
\sum_{j=1}^k
\sum_{d=0}^j
|Q_{k,j}(x)|
|R_{j,d}|M_d
\right\}
|\beta_{\langle 0 \rangle}(x)| \to 0 \quad (x \to 0+0)
\end{eqnarray*}
because the last $\{\cdots\}$ is a rational function
when $x$ is close to $0$.
Also this computation shows the validity of the first equality.
\hspace{\fill}$\square$
\begin{rem}{\rm \quad
For the equation (II),
the smoothness and flatness of
$\beta\log \beta_{\langle 0 \rangle}$
for a solution $\beta$ to (I)
follow from more or less the same
arguments, because
$\log \beta_{\langle 0 \rangle}
= (R + i\theta_0) \int_{x_0}^x\frac{1}{P(u)}du$.
}\end{rem}
\section{\large Unified proof for Case (2) and Case (3)}
The proof of Theorem \ref{Main} given in this section
relies on the theory of center manifolds and
the idea of graph transformation.
For this theory, refer to a nice book by Shub
\cite{Sh},
in particular, Appendix III to Chapter 5.
First we review the center manifold theorem in
a form which is suitable in and focused to our context.
\begin{thm}\label{center}{\rm(Theorem III. 2,
\cite{Sh},
modified) \quad
Let $T : E \to E$ be a continuous linear endomorphism
on a Banach space $E$ with a $T$-invariant decomposition
$E=E_1\oplus E_2$ into closed subspaces.
For the restrictions $T_i : E_i \to E_i$ ($i=1,\, 2$)
of $T$ we assume that
$T_1$ is an isomorphism and
there exist positive constants $0<\mu^* < \lambda^*$
satisfying the following conditions.
\begin{eqnarray*}
\Vert T_1({\bf v})\Vert > \mu^*\Vert {\bf v}\Vert
\quad
\mathrm{for\,\,all}\,\, {\bf v}\ne 0 \in E_1 \, ,
\\
\Vert T_2({\bf v})\Vert < \lambda^*\Vert {\bf v}\Vert
\quad
\mathrm{for\,\,all}\,\, {\bf v}\ne 0 \in E_2 \, .
\end{eqnarray*}
Then there exists a real number $\varepsilon^*>0$
such that if a $C^r$-map $\Phi : E \to E$
($r\geq 1$) satisfies $\Phi(0)=0$ and
$\mathit{Lip}(\Phi - T)<\varepsilon^*$,
then we have the followings.
\vspace{5pt}
\\
1)\quad The set
$\,\,\displaystyle W_1= \cap_{n\geq 0} \Phi^n(S_1)\,\,$
where
$\,S_1=\{({\bf v}_1, {\bf v}_2)\in E_1\times E_2 \, ; \,
\Vert {\bf v}_1\Vert \geq \Vert {\bf v}_2\Vert \}\,$
is the graph of a $C^1$-map $\,g:E_1 \to E_2\,$
with
$\mathit{Lip(g)}\leq 1$ and is invariant by $\Phi$,
namely,
$\Phi(W_1)=W_1\,$.
\vspace{3pt}
\\
2)\quad If $\,\lambda^*<(\mu^*)^r\,$ holds, then the map $\,g\,$
is of $\,C^r$.
}
\end{thm}
In this theorem $\mathit{Lip}$ denotes the Lipschitz constant
of a Lipschitz map, {\it i.e.},
$\displaystyle \mathit{Lip}(f)
= \sup\{\Vert f({\bf v_2})-f({\bf v_1})\Vert /
\Vert{\bf v_2}-{\bf v_1} \Vert
\, ; \, {\bf v_2}\ne {\bf v_1}, \,\,
{\bf v_1}, {\bf v_2} \in E
\}\,
$.
In order prove Theorem \ref{Main} concerning Equation (I),
we take $E_1={\Bbb R}$, $E_2={\Bbb C}$, $T_1=\mathit{id}_{\Bbb R}$, and
$T_2$ is a scalar multiplication by $\lambda^{-1}\,$.
Thereofre the real number $\varepsilon^*$
in the theorem is determined by $\lambda$.
The very virtue of this theorem is that
higher order regularities are assured
only by estimates
on 1-jets.
\vspace{3pt}
Let us explain a rough idea before getting into
the details.
As $\Phi$, the map
$(x, z)\mapsto (\tilde\varphi^{-1}(x), \lambda^{-1}z)$
or its modification will be taken. Here
$\tilde\varphi(x)=\varphi(x)$ for $x\geq 0$.
If we apply this theorem by taking
$\Phi(x, z)=(\tilde\varphi^{-1}(x), \lambda^{-1}z)$,
we just obtain $g(x)\equiv 0$ and nothing more.
The basic strategy is, not exactly but roughly;
for any $\beta \in \mathcal Z$
and for any $\varepsilon>0\,$,
we look for an appropriate $\Phi$ with
$\mathit{Lip}(\Phi - T)<\varepsilon$,
so that the resultant $g$ coincides with $\tilde\beta$
for $x<\delta$ for some $\delta>0$.
Here $\tilde\beta$ is an extension of $\beta$
to ${\Bbb R}$ by taking $\tilde\beta|_{(-\infty, 0]}\equiv 0$.
Before these arguments,
we need to take appropriate modifications of $\varphi$,
and for $\beta$ we choose $\Phi$ in a suitable way.
\vspace{5pt}
Let us start the proof of the theorem for Equation (I).
We fix a smooth function $h\in C^\infty({\Bbb R};[0,1])$
satisfying
$$
h(x)\equiv 0\,\,\, \mathrm{on} \,\,(-\infty,
1/3]
\quad
\mathrm{and}
\quad
h(x)\equiv 1\,\,\, \mathrm{on} \,\,
2/3,\infty)\, .
$$
Now take $\varphi\in{\mathit{Diff}}^\infty([0,\infty))$
in either of Case (2) or (3) and
$\lambda\in {\Bbb C}\,$ as well.
Take any extension of $\varphi$ in ${\mathit{Diff}}^\infty({\Bbb R})$.
For abuse of notation,
it is denoted by $\varphi$ again.
In Case (3), of course we can take $\varphi$
so as to be the identity on the negative side
$(-\infty, 0]$.
First, by the following lemma,
we
modify it
away from the origin so as to be
suitable for the center manifold theory
while its germ is not changed.
\begin{lem}\label{modification}{\rm \quad
For $\delta>0$ define $\tilde\varphi_\delta$ as follows.
\begin{eqnarray*}
\tilde\varphi_\delta(x)&=&h\left(\frac{-x}{\delta}\right)\cdot x
+ \left(1-h\left(\frac{-x}{\delta}\right)\right) \cdot \varphi(x)
\quad \mathrm{for}\,\, x\leq 0 \, ,
\\
\tilde\varphi_\delta(x)&=&h\left(\frac{x}{\delta}\right)\cdot
(x + \delta^2)
+ \left(1-h\left(\frac{x}{\delta}\right)\right) \cdot \varphi(x)
\quad \mathrm{for}\,\, x\geq 0 \, .
\end{eqnarray*}
Then we have
$$
\lim_{\delta\to 0+0}
\mathit{Lip}(\tilde\varphi_\delta - \mathit{id}_{\Bbb R}) = 0\, .
$$
In particular, we have the followings.
\vspace{5pt}
\\
1)\quad For small enough $\,\delta$,
$\tilde\varphi_\delta\,$ is in $\,{\mathit{Diff}}^\infty({\Bbb R})\,$
and expanding on $[0,\infty)$.
\vspace{3pt}
\\
2)\quad The germ of $\tilde\varphi_\delta$ around $x=0$
is the same as that of $\varphi\,$.
\vspace{3pt}
\\
3) \quad
$\tilde\varphi_\delta|_{(-\infty, -\delta]}
=\mathit{id}_{(-\infty, -\delta]}\,\,$
and
$\,\,\tilde\varphi_\delta|_{[\delta, \infty)}
=\mathit{id}_{[\delta, \infty)} + \delta^2\,$.
\vspace{3pt}
\\
4) \quad $\displaystyle
\lim_{\delta\to 0+0}
\mathit{Lip}(\tilde\varphi_\delta^{-1} - \mathit{id}_{\Bbb R}) = 0\, .
$
}
\end{lem}
\noindent
{\it Proof. }\quad From a direct computation using
$\,\varphi'(0)=1$,
the uniform convergence
$\,\tilde\varphi_\delta'(x) \to 0\,$
when $\,\delta\to 0+0\,$ is
easily obtained.
Then, because the support of $\,\,\tilde\varphi_\delta'- 1\,\,$
is contained in $\,[-\delta, \delta]\,$
and $\,\tilde\varphi_\delta\,$ converges to
$\,\mathit{id}_{\Bbb R}\,$
uniformly, we obtain the above estimate for the Lipschitz
constant. The statements 1) - 4) follow naturally.
\qed
For each
$\,\tilde\varphi_\delta|_{[0, \infty)}
\in {\mathit{Diff}}^\infty([0, \infty))\,$,
take any solution
$\beta \in {\mathcal Z}_{\lambda, {\tilde\varphi_\delta}}\,$
and the extension $\,\tilde\beta\,$ to $\,{\Bbb R}\,$
as explained above.
Our objective is to prove that
$\tilde\beta$ is smooth on $\,{\Bbb R}\,$.
At least for $\tilde\beta'(0)$,
not only we see it eaily
but we need it for our proof.
This fact is true even for the case
$\varphi'(0)> 1$ as far as
$\varphi'(0)<|\lambda|\,$.
\begin{prop}\label{1jet}{\rm \quad
$\tilde\beta$ is of $C^1\!$, namely,
$\displaystyle \lim_{x\to 0+0}\beta(x)=\lim_{x\to 0+0}\beta'(x)=0\,$
holds.
}
\end{prop}
\noindent{Proof.}\quad
Only $\displaystyle \lim_{x\to 0+0}\beta'(x)=0\,$
is verified.
From Equation (I), we have
$$
\beta'(\varphi(x))=(\varphi'(x))^{-1}\lambda\beta'(x)\, .
$$
From the condition there exist
$\,x_1\!>\!0\,$ and $\,\nu\!>\!1\,$
such that
$\,\varphi^{-1}|\lambda|\!<\!\nu\,$ holds
on $\,[0,x_1]\,$.
Take $M=\max|\beta'(x)|$ on the fundamental domain
$[\varphi^{-1}(x_1), x_1]$
of the action of $\varphi$ on $(0, \infty)\,$,
we see that when $\,x\,$ approaches to $0$ in $(0, x_1)$,
each time it passes through a smaller fundamental domain,
$|\beta'(x)|$ shrinks by $\nu^{-1}\,$. \qed
Next step is to look for a suitable
$\Phi : {\Bbb R}\times {\Bbb C} \to {\Bbb R}\times {\Bbb C} \,$.
First put
$\Phi_0(x,z)=({\tilde\varphi_\delta}^{-1}(x), \lambda^{-1} z)$.
The graph of $\,\tilde\beta\,$
is invariant under $\,\Phi_0\,$,
while the only invariant one contained in $W_1$
in the center manifold theorem is the real axis
${\Bbb R} \times \{0\}\, $, because any non-trivial
solution $\beta$ grows exponentially.
To avoid this inconvenience,
consider a diffeomorphism
$$H_c(x,\, z) = \left(x, \,
z + h\left(\frac{x}{c}\right)\tilde\beta(x)\right)$$
of $\,{\Bbb R}\times {\Bbb C}\,$ depending on
the parameter $\,c>0\,$.
$H_c$ is the identity on $\{x\leq c/3\}\,$.
Then by $H_c$ we take the conjugate
$$\Phi_c = H_c^{-1}\circ \Phi_0 \circ H_c\, .$$
\begin{lem}\label{conjugation}
{\rm \quad For any $\,\varphi$,
small enough $\,\delta$,
and $\beta$, the followings hold.
\vspace{5pt}
\\
1)\quad $\displaystyle \lim_{c\to 0+0}
\mathit{Lip}(\Phi_c - \Phi_0)=0\,$.
\vspace{3pt}
\\
2)\quad The graph of
$\,\,(1-h(\frac{x}{c}))\tilde\beta(x)\,\,$ is
invariant under $\,\Phi_c\,$.
}
\end{lem}
\noindent
{\it Proof.}\quad
Even though $\,H\,$ is exponentially
away from the identity according to $\beta$ growing
when $\,x\to \infty\,$,
thanks to the fact that
$\,H\,$ and $\,\Phi_0\,$ commute
to each other on $\,\{x\geq \tilde\varphi_\delta(c)\}\,$,
the result of conjugation does not go away.
Precisely the lemma is proved by direct computations as follows.
From the definition and Equation (I) we have
\begin{eqnarray*}
H_c^{-1}\circ\,\Phi_c\!\!\!\!&\circ&\!\!\!\! H_c(x, z)
\,- \, \Phi_0(x, z)
\\
&=&
\left(0\,,\,\,
\lambda^{-1}h\left(\frac{x}{c} \right)\tilde\beta(x)
- h\left(\frac{\tilde\varphi_\delta^{-1}(x)}{c} \right)
\tilde\beta(\tilde\varphi_\delta^{-1}(x))
\right)
\\
&=&
\left(0\,,\,\,
\lambda^{-1}\tilde\beta(x)
\left(
h\left(\frac{x}{c} \right)
- h\left(\frac{\tilde\varphi_\delta^{-1}(x)}{c}
\right)
\right)\right)\, .
\end{eqnarray*}
Therefore in order to conclude
$\displaystyle \lim_{c\to 0+0}
\mathit{Lip}(\Phi_c - \Phi_0)=0$,
it is enough to show the uniform convergence
of the derivative of the second component with respect to $x$
to $0$ when c tends to $0$.
The estimates concerning $c\to 0$
which appear below are uniform in $x$.
Let us make this point clearer.
The second component of the above has the support
contained in $[0, \tilde\varphi_\delta(c) ]$
as a function on $x$.
As we assumed that
$\displaystyle \lim_{x\to 0+0}\tilde\varphi_\delta(x) \to 1$,
taking $c>0$ small enough,
we can also assume that
$\tilde\varphi_\delta(c) \leq 2c\,$ and
it is enough to verify the estimates on $[0, 2c]\,$.
Now for example,
as we remarked in the above proposition we know
$|\tilde\beta(x)|=o(|x|)$ and hence
we have
$\max\{|\beta(x)|\, ; \, 0\leq x \leq 2c
\}
= o(c)\,$.
Now we show the derivative of the second component
with respect to $x$ uniformly converges to 0
(namely $o(1)$) when $c\to 0$.
We can forget about $\lambda^{-1}$
because it is just a constant.
\begin{eqnarray*}
\frac{d}{dx}
\left\{
\tilde\beta(x)
\left(
h\left(\frac{x}{c} \right)
- h\left(\frac{\tilde\varphi_\delta^{-1}(x)}{c}\right)
\right)
\right\}
\qquad
\qquad
\qquad
\qquad
\qquad
\quad
{}
\\
=\,
\tilde\beta'(x)
\left(
h\left(\frac{x}{c} \right)
- h\left(\frac{\tilde\varphi_\delta^{-1}(x)}{c}\right)
\right)
\qquad
\qquad
\qquad
{}
\\
\qquad
\qquad
+
\,\,
\frac{\tilde\beta(x)}{c}
\left(
h'\left(\frac{x}{c} \right)
-
(\tilde\varphi_\delta^{-1})'(x)
h'\left(\frac{\tilde\varphi_\delta^{-1}(x)}{c}\right)
\right)
\, .
\end{eqnarray*}
The first term is of $o(1)$ because $\tilde\beta'(x)=o(1)$
and
$|h(\frac{x}{c})
- h(\frac{\tilde\varphi_\delta^{-1}(x)}{c})|\leq 1\,$.
As to the second term,
$h'=O(1)$, $\tilde\beta'=O(1)$,
and $\tilde\beta(x)=o(c)$ as remarked above.
Therefore 1) is proved.
\vspace{5pt}
A direct computation verifies 2).
It is also understood from the arrangement
of $\Phi_c$,
because $H_c$ sends the graph of
$\,\,(1-h(\frac{x}{c}))\tilde\beta(x)\,\,$
to that of $\tilde\beta(x)$
which is invariant under $\Phi_0$.
\qed
Now we are ready to apply the center manifold theorem
to proof of Theorem \ref{Main} 1).
Recall that our
$\,T\,$ was fixed as $\,T(x, z)=(x, \lambda z)\,$.
From out settings,
we can take such $\mu^*$ and $\lambda^*$ that
\vspace{3pt}
\\
\quad $\bullet\,$
$\,\,\mu^*$ is as close to 1 as we want
as far as $\mu^*<1$ is satisfied
\\
and
\\
\quad $\bullet\,$
$\,\,\lambda^*$ is as close to $\lambda^{-1}$ as we want
as far as $\mu^* > \lambda^{-1}$ is satisfied.
\vspace{3pt}
\\
Therefore for an arbitrary fixed $\,r\in {\Bbb N}\,$
we take
$\mu^*$ and $\lambda^*$
so that
$\lambda^*<(\mu^*)^r$
is also satisfied.
From these choices
$\,\varepsilon^*\,$
in the center manifold theorem is also fixed.
Then by Lemma \ref{modification}
we can find $\delta>0$ so that
$\displaystyle\,\mathit{Lip}(\tilde\varphi_\delta - \mathit{id}_{\Bbb R})
< \varepsilon^*/2 \,$.
This means $\,\mathit{Lip}(\Phi_0 - T)
< \varepsilon^*/2\,$ because
$\displaystyle\,\mathit{Lip}(\Phi_0 - T) =
\mathit{Lip}(\tilde\varphi_\delta - \mathit{id}_{\Bbb R})\,
$.
For this $\,\delta$, by Lemma \ref{conjugation}
we can find $\,c>0\,$ so that
$\displaystyle\,\mathit{Lip}(\Phi_c - \Phi_0) < \varepsilon^*/2\,$.
Therefore the center manifold theorem is applicable
to $T$ and to our $\Phi_c$ as $\Phi$ in the theorem.
The second statement of Lemma \ref{conjugation} tells
that the graph of
$\,\,(1-h(\frac{x}{c}))\tilde\beta(x)\,\,$
is invariant by $\,\Phi_c\,$.
From the arrangement it is also clear that
the graph is contained in the sector $S_1$
in the center manifold theorem.
Therefore the graph is nothing but $W_1$
in the center manifold theorem and $g$
in the theorem turns out to be
$\,\,(1-h(\frac{x}{c}))\tilde\beta(x)\,\,$
in our case.
Therefore it is conluded that this function
is of $C^r\,$ and so is $\tilde\beta$.
We are free to improve the choice of
$\mu^*$ and $\lambda^*$ to obtain
another arbitrary $r\in {\Bbb N}\,$.
As a conclusion, $\tilde\beta$
is of $C^\infty$.
This implies nothing but
the fact that
$\beta$ is smooth and flat at $x=0\,$
and completes the proof.
\vspace{5pt}
It is easy to arrange the proof
for Equation (II').
The space $\,E_2\,$ is now taken to be $\,{\Bbb C}^M\,$
and the operator
$\,T\,$ to be $\,A^{-1}\,$.
We should remark here that
by change of basis, $\,A\,$ can be conjugate to one
which is arbitrarily close to
$\,\lambda\cdot E\,$ where $E$ denotes
the identity matrix.
This enables us to choose $\mu^*$ and $\lambda^*$
in the same say as in the above proof.
\begin{rem}{\rm
\quad Instead of using the center manifold theorem,
we can also arrange the proof so as to rely on the
$C^r$ section theorem due to Hirsch-Pugh-Shub
(cf. \cite{Sh}),
which even proves the center manifold theorem.
We also need the conjugation by $H_c$ to obtain $\Phi_c$
and then we may look at the space of functions
which are supported on a large enough interval $[-R, R]$.
}
\end{rem}
|
\section{MDP description}\label{appendix::model}
In this section we describe briefly the computation method of the attack policy that maximizes the fraction of attacked blocks, against a defender policy of the form $\sigma_{\alpha}\equiv k$, for some constant $k$ (that may depend on $\alpha$). A block $b$ of the honest network is successfully attacked if the published chain above it is of length $k$ or more, including $b$ in the count, and the attacker then overrides it by publishing a longer chain (or matching it). A given state of the form $(\sa,\sh)$ represents the lengths of the attacker's and the network's chain, respectively, counted above the latest fork (i.e., the latest block adopted by both parties). Thus, if $\sa>\sh\ge k$, then the attacker can attack the $\sh+1-k$ blocks at the bottom of the honest chain (above the fork)---these blocks have $k$ confirmations, hence were accepted by the policy $\sigma$. The remaining $k-1$ blocks are indeed overridden but not attacked, since the policy didn't accept them yet, hence the transactions in them were not considered safe yet by their recipients. This is the main difference from selfish mining, where the attacker is rewarded for these blocks as well.
When the attacker abandons the attack and adopts, all blocks in the chain it adopted will be accepted and never attacked, since future attack blocks contain them in their history.
Accordingly, we grant the honest network a reward of $\sh$ whenever the attacker adopts. When the attacker adopts, we reward the attacker $\sh-(k-1)$ (this is the number of blocks it successfully attacked), and reward the honest network $k$ blocks (this complements the attacker's reward to the chain's new length $\sh+1$, and is needed for appropriate normalization). Further complexity arises due to the possibility of the attacker matching the honest chain's length (rather than succeeding it by 1). Whether this \ma action is feasible, for a given state $(\sa,\sh)$, is encoded in a third field called $fork$ with possible values: $irrelevant,relevant,active$.
For further details, and for a description of an algorithm that uses this MDP to maximize the fractional non-linear objective -- refer to~\cite{OPT_SELFISH_MINING}.
\begin{table}[h]
\caption{A description of the transition and reward matrices of the MDP. The third column contains the probability of transiting from the state specified in the left-most column, under the action specified therein, to the state on the second one. The corresponding two-dimensional reward (that of the attacker and that of the honest nodes) is specified on the right-most column.\label{table::PandR}}
\begin{tabular}{|K{37mm}|K{34mm}|K{23mm}|K{25mm}|}
\hline
\textbf{State $\times$ Action} & \textbf{State} & \textbf{Probability} & \textbf{Reward}\\
\hline
\multirow{2}{*}{$(\sa,\sh,\cdot),adopt$} & $(1,0,irrelevant)$ & $\alpha$ & \multirow{2}{*}{$(0,\sh)$}\\
\ & $(0,1,irrelevant)$ & $1-\alpha$ & \\
\hline
\multirow{2}{*}{$(\sa,\sh,\cdot),override^\dagger$} & $(\sa-\sh,0,irrelevant)$ & $\alpha$ & \multirow{2}{*}{$(\sh-k+1,k)^\mathsection$}\\
\ & $(\sa-\sh-1,1,relevant)$ & $1-\alpha$ & \\
\hline
\multirow{2}{*}{\makecell{$(\sa,\sh,irrelevant),wait$\ \\ $(\sa,\sh,relevant),wait$}} & $(\sa+1,\sh,irrelevant)$ & $\alpha$ & (0,0)\\
\ & $(\sa,\sh+1,relevant)$ & $1-\alpha$ & (0,0)\\
\hline
\multirow{3}{*}{ \makecell{$(\sa,\sh,active),wait$\ \\ $(\sa,\sh,relevant),match^\ddagger$}} & $(\sa+1,\sh,active)$ & $\alpha$ & (0,0)\\
\ & $(\sa-\sh,1,relevant)$ & $\gamma\cdot(1-\alpha)$ & $(\sh-k+1,k-1)^\mathparagraph$\\
\ & $(\sa,\sh+1,relevant)$ & $(1-\gamma)\cdot(1-\alpha)$ & (0,0)\\
\hline
\end{tabular}
\noindent\begin{scriptsize}$^\dagger$feasible only when $\sa>\sh$\end{scriptsize} \\ \noindent\begin{scriptsize}$^\ddagger$feasible only when $\sa\ge\sh$\end{scriptsize} \\
\noindent\begin{scriptsize}$^\mathsection$if $\sh<k-1$, then the reward is $(0,\sh+1)$, as no block was actually attacked.\end{scriptsize}\\ \noindent\begin{scriptsize}$^\mathparagraph$if $\sh<k-1$, then the reward is $(0,\sh)$, , as no block was actually attacked.\end{scriptsize}
\end{table}
\section{Conclusions}
We presented a variety of different interpretations of the security of a single transaction in the Bitcoin system, and matching advice regarding the number of confirmations merchants should await in order to properly secure their transactions. Our suggested prescription can be summarized, in short, as follows:
\begin{itemize}
\item Transactions whose timings can be assumed to be non-adversarial can be protected via $\epsilon$-arbitrary-robust policies, such as the $\sigma^{arb}$ family presented in Section~\ref{sec::premining}.
\item Merchants engaging in medium-valued transactions at a regular rate ought defend against a reversal of non-negligible fractions of the transactions they have authorized, in the long term (the $\epsilon$-fractional-robustness security model, coupled with the $\sigma^{frac}$ policy).
\item Recipients of large transactions are advised to commit to policies such as $\sigma^{total}$ (Section~\ref{sec::logarithmic}) which waits a time logarithmic in the chain's length, thereby guaranteeing themselves $\epsilon$-total-robustness, i.e., security from even a single reversal of any of their payments.
\item Light clients are advised to use a policy specifically protecting against the generalized Vector76 attack ($\sigma^{spv}$, Section~\ref{sec::vec76}).
\end{itemize}
Indeed, regarding the latter point, we demonstrated a clear case in which light nodes are less secure than full nodes solely due to the fact that they do not relay blocks further. This observation suggests several mitigation techniques, including sending requests for recent blocks and relaying them to the network, which would in fact imply that hybrids between light nodes and full nodes can be more secure.
We have further shown that it is difficult to argue that attackers lose revenue from mining if they are trying to attack. Instead, Bitcoin can be considered to give guarantees lower bounding the losses of merchants.
Many research directions remain. The models here should be further adapted to settings with more significant delay. Several hybrid guarantees can also be explored. One such example is a fractional guarantee for a merchant that receives transactions occasionally (but not in every block), or attackers that are more limited in selecting the timing of their attacks (e.g., if a store is only open during the daytime). Finally, it would be interesting to evaluate the guarantees of variants of the protocol such as Bitcoin-NG~\cite{BitcoinNG} and Ethereum~\cite{ETHEREUM} against similar attacks, as they employ slightly different rules to manage the blockchain.
\section{The profit of an attacker}\label{sec::profit_of_attacker}
Arguably, one might hope that carrying out double spending attacks would be costly to the attacker, due to the loss in potential profit the attacker could gain from participating honestly in the mining effort. This, presumably, will disincentivize attackers from committing to long attack-strategies which waste their resources.
Alas, the work of Sapirshtein et. al.~\cite{OPT_SELFISH_MINING} observes that this is not the case in general. An attacker with sufficient hashrate or a significant $\gamma$ can actually combine profitable selfish mining with double spending attacks. Their idea is simple: Every profitable selfish mining scheme arrives with positive probability at states of the form $(\sa,\sh)$, $\sa>\sh+k$, from which the attacker can plan a definitely successful double spending attack (against the policy $\sigma\equiv k$). In this section we aim at quantifying the potential profit for the attacker under optimal combinations of these attacks.
To this end, we adapt the MDP used above to a setup in which the attacker maximizes its returns from both the block reward and fees, and from possible double spending attacks it manages to perform. We make the simplifying assumption that the reward is fixed between blocks, and thus we reward the attacker with one unit for every block of his which is part of the longest chain. We additionally reward it with $\mathcal{R}$ units of reward for every successful double spend.
Following~\cite{ES}, we normalize the rewards by the length of the main chain.
Figure~\ref{fig::cost3} depicts the profit of an attacker that carries out such an attack. The dashed line corresponds to honest mining without double spending attacks (where a miner with $\alpha$ of the hash rate gains an $\alpha$-fraction of the rewards), and the other curves correspond to the profit computed by the MDPs for different values of $\gamma$. Consistent with the results of~\cite{ES,OPT_SELFISH_MINING}, at $\gamma=1$ the attacker is always profitable (at any $\alpha$), and occasionally attacks with double spending attacks. Here, the gains from double spending are assumed to be $\mathcal{R}=2$ block rewards.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.8]{figures/fig_cost_2.eps}
\caption{The profit of an attacker from carrying optimal combinations of selfish mining and double spend attacks on the $\sigma\equiv 6$ confirmations policy, assuming a successful double spent transaction is worth three times the value of an ordinary block.}
\label{fig::cost3}
\end{figure}
Figure~\ref{fig::cost05} depicts the gains for different values of double spend. The different plots correspond to different reward values from successful double spending. It is interesting to see the expected result: given that rewards from double spending increase, smaller miners can choose to deviate from honest behavior and gain (honest mining is again represented by the dashed line).
\begin{figure}[ht]
\centering
\includegraphics[scale=0.8]{figures/fig_cost_0.5.eps}
\caption{The profit of an attacker from carrying optimal combinations of selfish mining and double spend attacks on the $\sigma\equiv6$ confirmations policy, assuming that it is able to match a fraction of $\gamma=0.5$ of the honest nodes.}
\label{fig::cost05}
\end{figure}
\section{Defending the long-term fraction of double spent transactions}
In this section we focus on finding the level of fractional-robustness of policies $\sigma^{frac}_{\alpha,\gamma}$ that wait for some given constant number of confirmations. We investigate what fraction of all blocks that the policy accepts are later overridden.
The robustness of $\sigma^{frac}_{\alpha,\gamma}$ is not derived analytically, but is rather computed by an algorithm that finds the optimal attack policy. To compute the optimal attack, we follow the technique introduced in~\cite{OPT_SELFISH_MINING}, that encodes the decision problem of an attacker as a sequence Markov Decision Problems (MDPs). These then encode the action that the attacker takes at each state: for every length of attacker chain $\sa$ and length of honest chain $\sh$ (measured from the block they fork at), the attacker needs to decide whether it continues to build atop his chain, abandons his efforts and starts a new fork, or publishes blocks to succeed by one (or, with less success-certainty, match) the length of the network's chain thereby overriding it (provided he has enough blocks to do this). The transition and reward matrices are summarized in Appendix~\ref{appendix::model}.
The main difference from the algorithm presented in~\cite{OPT_SELFISH_MINING} is that the latter used this technique to compute optimal selfish mining attacks, and to maximize the number of attacker blocks in the chain. In contrast, we reward the attacker differently (as its objective here is different): The attacker is rewarded 1 unit for every successful block that the network accepted (i.e., had enough confirmations) and that the attacker managed to later remove from the chain. This reward is normalized by the number of all accepted blocks. Due to this normalization, the output of this computation equals the expected number of attacker blocks over the expected total number of accepted blocks, which in turn equals the left-hand side of~(\ref{eq_def_fractional}).
Recall that the parameter $\gamma$ encodes the probability that a chain is overridden when it is matched in length. Since honest nodes adopt the first chains that ehy receive, in case of ties, the ability of the attacker to push his block first to a significant fraction of the nodes dictates the chances that the next block will be built on top of its chain.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.85]{figures/fig_fractional_6.eps}
\caption{\label{fig::frac}The fraction of accepted blocks that an optimal attacker can double spend against a defender that uses 6 confirmations to accept as a function of the attacker's hashrate $\alpha$. The different curves correspond to different values of $\gamma$. Rosenfeld's result is also plotted for comparison.}
\end{figure}
Figure~\ref{fig::frac} depicts the results obtained for a policy with 6 confirmations, as computed on an MDP that was truncated to consider chains of length up to 60 blocks (the MDP analyzed in~\cite{OPT_SELFISH_MINING} is infinite and needs to be truncated for a numeric solution). The figure depicts the fraction of blocks an optimal attacker may double spend, for different values of $\gamma$. This essentially measures the $\epsilon$-robustness of the policy $\sigma\equiv6$.
The results of Rosenfeld for the probability of attack on a block~\cite{MENI} are included for comparison. It is interesting to note that the fraction of blocks that can be attacked is in fact lower than predicted by Rosenfeld (for any $\gamma$). This is because his analysis (and Satoshi's as well) consider an attack on a single block that goes on infinitely, that is, the attacker is assumed to never give up and to try to catch up with the chain no matter how far behind he is. In contrast, an attacker that aims to maximize the fraction of blocks it successfully attacks must occasionally give up and restart the attack if he is far behind. This effect is demonstrated in these results (note that, on the other hand, our model allows the attacker to double spend several blocks at once. These results demonstrate that the effective $\epsilon$ lowers nonetheless).
We similarly present the percentage of double spent blocks for different numbers of confirmations in Table~\ref{table::fractional}. Each cell was computed separately with its own optimal policy.
\begin{table*}[ht]
\caption{The fraction of the network's blocks that an attacker with a given hashrate ($\alpha$) successfully attacks, when using an optimal attack policy, given the number of confirmations the acceptance policy waits for ($conf$).}
\label{table::fractional}
{\ \ }\\
\centering
{\footnotesize
\begin{tabular}{|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|}
\hline
$\alpha\backslash conf$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\ \hline
2\% & 0.08\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% \\ \hline
6\% & 0.69\% & 0.12\% & 0.03\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% \\ \hline
10\% & 1.89\% & 0.52\% & 0.16\% & 0.05\% & 0.02\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% \\ \hline
14\% & 3.70\% & 1.34\% & 0.53\% & 0.23\% & 0.10\% & 0.05\% & 0.02\% & $\approx$ 0\% & $\approx$ 0\% & $\approx$ 0\% \\ \hline
18\% & 6.16\% & 2.75\% & 1.34\% & 0.69\% & 0.36\% & 0.20\% & 0.11\% & 0.06\% & 0.04\% & 0.02\% \\ \hline
22\% & 9.37\% & 4.92\% & 2.80\% & 1.66\% & 1.02\% & 0.64\% & 0.41\% & 0.27\% & 0.18\% & 0.12\% \\ \hline
26\% & 13.47\% & 8.12\% & 5.34\% & 3.63\% & 2.52\% & 1.78\% & 1.28\% & 0.92\% & 0.67\% & 0.49\% \\ \hline
30\% & 18.71\% & 13.20\% & 9.63\% & 7.19\% & 5.48\% & 4.23\% & 3.30\% & 2.60\% & 2.07\% & 1.66\% \\ \hline
34\% & 26.46\% & 20.57\% & 16.36\% & 13.29\% & 10.99\% & 9.17\% & 7.69\% & 6.49\% & 5.51\% & 4.71\% \\ \hline
38\% & 36.54\% & 31.04\% & 26.95\% & 23.60\% & 20.77\% & 18.37\% & 16.39\% & 14.66\% & 13.16\% & 11.84\% \\ \hline
42\% & 50.32\% & 46.42\% & 42.99\% & 39.91\% & 37.17\% & 34.73\% & 32.49\% & 30.43\% & 28.56\% & 26.84\% \\ \hline
46\% & 69.53\% & 67.65\% & 65.84\% & 64.06\% & 62.33\% & 60.64\% & 59\% & 57.38\% & 55.79\% & 54.23\% \\ \hline
48\% & 81.48\% & 80.59\% & 79.66\% & 78.72\% & 77.75\% & 76.77\% & 75.76\% & 74.73\% & 73.67\% & 72.59\% \\ \hline
50\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% \\ \hline
\end{tabular}
}
\end{table*}
\subsection{Optimal policies}\label{subsec::optimal_policies} We now present the optimal policies returned by our algorithm, in two particular setups. Table~\ref{tbl::policy_26_0} describes the policy for an attacker with $\alpha=0.26, \gamma=0$ (here \ma is of no consequence). The row numbers correspond to the length of the attackers branch $\sa$ and the columns to the length of the honest network's branch $\sh$.
Notice that here the attacker does not override the network's chain and receive rewards until its branch is of length three at least, as a successful attack requires the merchant sees three confirmations above its chain before the attack is released. Note, additionally, that the attacker does not give up on his attack when he is just slightly behind. If his chain is relatively long, he will not abandon it unless he is at least 3 blocks behind.
Table~\ref{tbl::policy_26_0.5} similarly corresponds to $\alpha=0.26, \gamma=0.5$. Each entry in it contains a string of three characters, corresponding to the possible status of the honest network: if it is working only on its own branch, if the attacker can possibly match the length of its branch and split its resources, and if it is already split between two branches of equal length ($fork$: $irrelevant,relevant,active$).\footnote{E.g., the string ``wm${*}$'' in entry $(\sa,\sh)=(3,3)$ reads: ``in case a fork is $irrelevant$ (that is, the previous state was $(2,3)$), $wait$; in case it is $relevant$ (the previous state was $(3,2)$), $match$; the case where a fork is already $active$ is not reachable''.} Actions are abbreviated to their initials: $\textbf{\emph{a}}dopt, \textbf{\emph{o}}verride, \textbf{\emph{m}}atch, \textbf{\emph{w}}ait$, while `$*$' represents an unreachable state.
\begin{table}[ht]
\caption{Optimal actions for an attacker with $\alpha=0.26,\gamma=0$, against a policy that waits for two confirmations, when the merchant accepts transactions after 2 confirmations. The row and column indices correspond to $\sa$ and $\sh$, respectively.
Actions are: $\textbf{\emph{a}}dopt, \textbf{\emph{o}}verride, \textbf{\emph{m}}atch, \textbf{\emph{w}}ait$, or `$*$' unreachable.}
\label{tbl::policy_26_0}
{\ \ }\\
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$\sa\backslash\sh$ & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\ \hline
0& w & a & $*$ & $*$ & $*$ & $*$ & $*$ & $*$ & $*$ & $*$ & $*$\\ \hline
1& w & w & w & a & $*$ & $*$ & $*$ & $*$ & $*$ & $*$ & $*$\\ \hline
2& w & w & w & w & a & w & a & $*$ & $*$ & $*$ & $*$\\ \hline
3& w & w & o & w & w & w & w & a & $*$ & $*$ & $*$\\ \hline
4& w & w & w & o & w & w & w & w & a & $*$ & $*$\\ \hline
5& w & w & w & w & o & w & w & w & w & a & $*$\\ \hline
6& w & w & w & w & w & o & w & w & w & w & a\\ \hline
7& w & w & w & w & w & w & o & w & w & w & w\\ \hline
8& w & w & w & w & w & w & w & o & w & w & w\\ \hline
9& w & w & w & w & w & w & w & w & o & w & w\\ \hline
10 & w & w & w & w & w & w & w & w & w & o & w \\ \hline
\end{tabular}
\end{table}
\begin{table}[ht]
\centering
\caption{Optimal actions for an attacker with $\alpha=0.26, \gamma=0.5$, for states $(\sa,\sh,\cdot)$ with $\sa,\sh\le 6$, when the merchant accepts transactions after 2 confirmations.
The row and column indices correspond to $\sa$ and $\sh$, respectively.
Actions are: $\textbf{\emph{a}}dopt, \textbf{\emph{o}}verride, \textbf{\emph{m}}atch, \textbf{\emph{w}}ait$, or `$*$' unreachable. The three entries at each cell are: $fork$: $irrelevant,relevant,active$}
\label{tbl::policy_26_0.5}
{\ \ }\\
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline $\sa\backslash\sh$ & 0 & 1 & 2 & 3 & 4 & 5 & 6 \\ \hline
0 & w$**\!$& aa$*\!$& $*\!*\!*\!$& $*\!*\!*\!$&$*\!*\!*\!$&$*\!*\!*\!$&$*\!*\!*\!$\\ \hline
1 & w$**\!$&$*$w$*\!$& w$**\!$& a$**\!$&$*\!*\!*\!$&$*\!*\!*\!$&$*\!*\!*\!$\\ \hline
2 & w$**\!$& ww$*\!$& wm$*\!$& w$**\!$& a$**\!$&$*\!*\!*\!$&$*\!*\!*\!$\\ \hline
3 & w$**\!$& ww$*\!$& www & wm$*\!$& w$**\!$& a$**\!$&$*\!*\!*\!$ \\ \hline
4 & w$**\!$& ww$*\!$& www & wmw & wm$*\!$& w$**\!$& a$**\!$ \\ \hline
5 & w$**\!$& ww$*\!$& www & www & omw & wm$*\!$& w$**\!$ \\ \hline
6 & w$**\!$& ww$*\!$& www & www &$*$mw & omw & wm$*\!$ \\ \hline
\end{tabular}
\end{table}
\section{Introduction}
Users of the Bitcoin system~\cite{SATOSHI} rely on the irreversibility of monetary transfers when using the currency. In particular, merchants that accept bitcoins, must be assured that once a payment has been accepted, it will not be reversed or routed to a different destination, and that they can safely dispense products and services in exchange for the funds.
Payments may be rerouted or canceled if, for example, an attacker tries to send two conflicting transaction requests to the system in an attempt to send the same funds to two different destinations. The system cannot allow money to be used twice and thus one of the two conflicting payments must be rejected eventually. It is important that the recipient of the canceled payment is not fooled into thinking he has received the payment in the interim. Such an attack is called a \emph{double spending} attack. Indeed, Bitcoin's most important innovation is its solution to this very problem.
Bitcoin's core data structure -- The Blockchain -- contains a record of all transactions that have been accepted by the system. Each block is a batch of accepted transactions that contains additionally the cryptographic hash of its predecessor in the chain, as well as a cryptographic proof-of-work. Blocks are created by nodes that solve this proof-of-work and in return collect fees from transactions embedded in their block and from newly minted money as well. These nodes are often called \emph{miners}.
In case several chains form, due to the concurrent action of miners, Bitcoin nodes accept the longest chain as the record of transactions that have occurred,\footnote{In fact the chain representing the highest cumulative amount of computational power is chosen. This is usually the longest chain.} and ignore transactions not contained in this chain. This re-selection of the set of accepted transactions may cause some payments to be canceled, which may be abused by an attacker.
To be secure against such double spending, merchants are advised to wait until their transaction is included in a block, and that several blocks are built on top of it.
The more blocks built atop a given block, the less likely it is that a conflicting longer branch will form (even under deliberate attempts).
For a transaction embedded in a block, the block containing it, and each block that follows on the main chain, is counted as an additional \emph{confirmation}.
Satoshi in his original work~\cite{SATOSHI}, as well as additional works that follow~\cite{MENI,GHOST,garay2015bitcoin}, offer a guarantee of the security of transactions in the currency. Specifically, each provides a similar theorem of the following ``flavour'':
\begin{theorem}[informal]
As long as the attacker holds less than 50\% of the computational power, and all honest nodes can communicate quickly (compared to the expected time for block creation), the probability of a transaction being reversed decreases exponentially with the number of confirmations it has received.
\end{theorem}
This work is motivated by the following argument against any such guarantee: If the attacker is allowed to choose the time it prefers to transmit the transaction, no probabilistic guarantee can be given that its attack will fail. Indeed, the attacker
may try to create blocks prior to the the transmission of transactions, in a preparatory stage that we term \emph{pre-mining}. The pre-mining stage may take a long while to succeed, but once it does, an attack can be carried out with success-probability 1.\footnote{The attacker may, alternatively, settle for fewer blocks during the pre-mining stage, and then carry out an attack with lower probability.} Figure~\ref{fig::premining_attack} illustrates such an attack.
It is important to note that the pre-mining stage need not be costly to an attacker. In fact, attackers that employ selfish mining strategies~\cite{ES,OPT_SELFISH_MINING,STUBORN_MINING} repeatedly create secret chains that are longer than those of the network and which can be additionally used to launch double spending attacks, and in fact gain as they do so (provided that the attacker is sufficiently well connected to the network~\cite{ES}, or if delays exist~\cite{OPT_SELFISH_MINING}).
\myitem{Our contributions.}
We discuss two pre-mining attacks that can be used when the attacker can choose the timing of the transaction: One attack is a generalization of the Finney attack~\cite{Finney}, and the other is a generalization of the (somewhat lesser known) Vector76~\cite{Vector76} attack. The first can be used effectively against any node, whereas the second, against nodes that do not broadcast blocks such as lightweight clients that do not maintain a full copy of the blockchain.
We propose four different versions of security guarantees (for regular nodes), given that pre-mining may in general take place:
\begin{enumerate}
\item Defending an independently generated transaction (one whose timing does not rely on the attacker) \label{g1}
\item Defending the long-term fraction of lost transactions to the merchant
\item Defending all transactions from ever being double spent\label{g3}
\item Upper bounding the average profit of the attacker during a continuing attack
\end{enumerate}
We formalize these notions in Section~\ref{sec::model}. We then introduce three families of acceptance policies, $\sigma^{arb}$, $\sigma^{frac}$, and $\sigma^{total}$ that provide a defense of according to guarantees~\ref{g1}-\ref{g3} above.
With respect to the bound on the profit of attackers, we show that, indeed, attackers with enough mining power (still under 50\%) or superior networking capabilities can profitably launch double spending attacks, even when selfish mining schemes alone are not profitable to them.
The first guarantee above most closely matches the flavour of the theorem given by Satoshi~\cite{SATOSHI} and following works. We provide a corrected analysis that better accounts for pre-mining in this case, and maintains the general exponential decay. We highlight that this result is of slightly lesser use, as in most cases, attackers can easily control the timing of an attack: a buyer, for example, can choose the exact moment at which it enters a store to buy items, since merchants usually provide continuous service.
As it is impossible to bound well below 1 the probability that a transaction timed by the attacker will succeed, or to bound the cost for an attacker, we suggest the second security guarantee as an effective upper-bound on the losses experienced by a merchant who regularly transacts with the currency.
This guarantee corresponds to a ``safety-level'' strategy for a merchant who wishes to ensure that only a small fraction of his accepted payments are double spent. Here, we compute the optimal attack for every given policy of the merchant and thus compute its exact safety level.
The third model best applies to large valued transactions whose introduction to the blockchain may have been selected at the convenience of the attacker. We show that waiting for a fixed number of confirmations does not provide adequate security in this case. To circumvent this, we provide a policy that requires a number of confirmations logarithmic in the length of the chain, and prove that, with high probability, no transaction will \emph{ever} be attacked when sticking to this policy. The downside of this policy is of-course the fact that the number of confirmations that it requires grows (albeit incredibly slowly), as time goes on.
We now elaborate more on pre-mining attacks and why they pose a risk for a merchant that receives a transaction that was broadcast at a timing selected by the attacker.
\subsection{Pre-mining and selective attack timings}
\begin{figure*}
{\centering
\includegraphics[scale=0.5]{figures/premining_attack.eps}
}
\emph{As the attack begins the attacker starts working on a secret chain with $tx_2$ inside its first block (1). If the attacker's chain is shorter than the honest nodes', the attacker gives up and restarts the attack (2). The attacker manages to gain a lead of 2 blocks (3). He then transmits the transaction he wishes to double spend which is included in a block (4). The transaction now has enough confirmations (1-conf) and the attacker collects his rewards. He then publishes his secret chain and successfully double spends (5). Notice that once the pre-mining stage is concluded, the attack succeeds with probability 1, so miners that see $tx_1$ that is only broadcast then will always lose the funds.}
\caption{\label{fig::premining_attack} The progression of a pre-mining attack on a 1-confirmation defender}
\end{figure*}
Consider the following attack scheme against a defender that waits for $k$ confirmations:
In the pre-mining phase, the attacker begins to work on a secret branch that splits off from the most recent version of the chain. He embeds transaction $tx_2$ in this chain that conflicts the transaction $tx_1$ he wishes to double spend. If the attacker manages to create $k+1$ blocks more than the network, then he proceeds to carry out the attack. If at any point in time the network's chain is longer than the attacker's, he resets and starts a new branch, spiting off at a higher block in the public chain. Notice that this phase is in fact performed silently, and repeats until he is successful.
Once the attacker holds $k+1$ more blocks than the network's chain, he broadcasts his transaction to the network (when a large enough fee to ensure he is included in the next block), and waits for he to gain $k$ confirmations. It then releases his chain which is adopted immediately by the network, and invalidates the transaction $tx_1$.
Observe that since the attack is only visible if the attacker is going to win, the recipient of funds can never be safe---\emph{conditioned on seeing the transaction}, the attack succeeds with probability 1. More sophisticated schemes are possible, and will be discussed throughout this paper.
The restricted version of this attack for a 0-confirmation defender is simply known as \emph{The Finney Attack} (named after its discoverer, Hal Finney, one of Bitcoin's first adopters).
The key point in these pre-mining attacks is that they are not carried out at an ``arbitrary'' moment in time, but rather at a moment selected by the attacker. This leads us to the natural question: In what sense then is Bitcoin secure against such attacks?
\subsection{Guarantees}
We consider three main scenarios that a recipient of funds may face:\\
\myitem{1. Protecting an independently placed transaction.} Here we assume a transaction has been placed in a block \emph{independently of the actions of the attacker}. One example of such a transaction is a minting transaction (also known as a \emph{coinbase transaction}) that the recipient may wish to accept. In this case, the attacker could not have chosen to launch the attack once he is successful in a pre-mining stage, but may still have pre-mined blocks.\footnote{This is the scenario that most closely matches previous results in~\cite{GHOST,SATOSHI,MENI,garay2015bitcoin}, although each work has analyzed it slightly differently. In this paper we augment the analysis with a proper quantification of the attacker's pre-mining.} \\
\myitem{2. Protecting a large fraction of blocks}
Here we consider a merchant that regularly receives payments in the blockchain and wishes to upper bound the loss he may suffer due to an attacker. We wish to find the attack policy that maximizes the fraction of blocks that are accepted by the network and then removed from the chain by the attacker. We note that a single double spending attack in which a long attack chain is released may remove many blocks simultaneously, thus this case differs from the previous one. An additional difference is that, in this case, attackers must actively decide when to give up on the attack on a specific block if the odds are not in his favour so that he may attack other more recent blocks instead. Such restarts are not considered when protecting a single transaction.
We show here, as well, that waiting for a fixed number of confirmations which is a function of $\epsilon$ can provide any level of security.
\myitem{3. Protecting all accepted blocks} In this case we consider a merchant that wishes to receive funds for a transaction at a moment in time that is possibly selected by the buyer. Given that the attacker can choose to place his transaction inside a block once he already knows he is certain to succeed, the only way to be fully secure in such cases is to find a policy for accepting transactions that never accepts a block that will be double spent. As we have already discussed above, it is impossible to be secure against such a scenario by waiting for a fixed number of confirmations. While it is trivial to solve this problem by holding off acceptance of transactions indefinitely, we present a policy that guarantees that \emph{no block can be double spent} which fails only with arbitrarily low probability $\epsilon$, and requires a logarithmic number of confirmations in $1/\epsilon$ and the length of the chain. While waiting times that depend on the length of the chain and may grow are somehow unsatisfactory, we note that growth is extremely slow. Still, we believe that this is the main model that needs to be considered when dealing with extremely large transactions (e.g., when sums that are equivalent to tens of millions of USD are sent -- as was the case with several large bitcoin transactions like the FBI's seizure of the SilkRoad funds back in 2013).\\
\subsection{Related Work}
The first analysis of the resilience of Bitcoin is due to Nakamoto~\cite{SATOSHI}. His analysis considers a double spending attack without pre-mining, and is only approximate. Rosenfeld later goes on to correct the analysis~\cite{MENI}, and includes the pre-mining of a single block before it is launched (as such, it is not an attack against an arbitrarily chosen block). Rosenfeld further argues that the cost of an attack grows exponentially (which is correct as long as no block withholding is performed). Lewenberg et. al.~\cite{INCLUSIVE} also consider the exponential cost of the simple hidden-chain attack.
In a previous work~\cite{GHOST} we have extended the analysis of security for settings with delay, demonstrating that the security of Bitcoin declines as delays increase and bounding its resulting throughput. We
additionally present a time-dependent acceptance policy, applicable for the longest chain rule, that is more secure than purely structural policies (and as a result is faster to accept for a given level of security). In this work we restrict ourselves to structural policies, as it is not always guaranteed that the recipient remains online to time the creation of blocks.
Garay et. al.~\cite{garay2015bitcoin} provide a formal model for the core of the Bitcoin protocol, using a discrete time setup where blocks can be found simultaneously at each step. They define desired properties of a blockchain protocol, and prove that they are satisfied by Bitcoin, when the attacker is adequately bounded.
They derive asymptotic bounds for security (and not an explicit formula). Their analysis assumes the transaction to defend is available to all honest nodes, and hence too roughly corresponds to an attack on an independently chosen block.
Karame et.al.~\cite{karame2012two} and Bambert et. al.~\cite{bamert2013have} have both considered double spending 0-confirmation payments. The simplest pre-mining attack which applies to such payments is known as the Finney attack~\cite{Finney}.
Eyal and Sirer~\cite{ES} suggested and analyzed a particular attack that knocks out blocks of the network in order to gain more from mining. Sapirshtein et. al.~\cite{OPT_SELFISH_MINING} improved the attack to optimal policies. This work uses these techniques to analyze optimal double spending strategies.
\section{Logarithmic waiting time}\label{sec::logarithmic}
For every given acceptance policy one could ask what is the probability that at least one attack, in the course of the entire history, will be successful. This notion is formalized by $\epsilon$-total-robustness, in Definition~\ref{def_total}.
As discussed above, for any acceptance policy of the form $\sigma\equiv k$, for some constant $k$, the probability that a single attack on $\C^t$ will be successful goes to 1 as $t$ goes to infinity. Observe that achieving an arbitrary low $\epsilon$-total-robustness is trivially achievable by never accepting any transaction. Fortunately, below we show that there exists an $\epsilon$-totally-robust policy, for any $\epsilon>0$ which accepts every transaction in the blockchain after a time logarithmic in the chain's current length, as long as the block containing it still belongs to the longest chain. This result motivated the modeling of acceptance policies as taking the height of the block as an argument: It shows that considering policies not constant in the block height open up the option of achieving a strong security property unachievable otherwise.
\begin{theorem}
For any $\epsilon>0$, the policy $$\sigma^{total}_{\alpha,\gamma}\left(h\right):=C_{\alpha,\epsilon}+\lfloor\log_{b_{\alpha}}\left(h\right)\rfloor$$ is $\epsilon$-totally-robust, where $C_{\alpha,\epsilon}:=\ceil*{\frac{1}{c}\cdot \ln\left(\frac1\epsilon\cdot b_{\alpha}\cdot\left(1-e^{-c}\right)^{-1}\cdot\left(1-b_{\alpha}/(e^c)\right)^{-1}\right)}$, $b_{\alpha}:=\frac{e^c+1}{2}$, with $c:=\frac18\cdot\frac{(1-2\cdot\alpha)^2}{1-\alpha}$.
\end{theorem}
\begin{proof}
\emph{Part I:}
Let us write $(\sa,\sh,h)$ whenever the attacker's chain is $\sa$ blocks long, the honest network's chain is \sh blocks long, and the earliest block in the chain of the network (i.e., the $\sh$-th block from the tip of the honest chain) is of height $h$.
By definition, the attacker can perform a successful attack iff $\sa\ge \sh$ and $\sh\ge \sigma^{total}_{\alpha,\gamma}\left(h\right)$. Assume now that if the attacker performs \ad at some state $(\sa,\sh,h)$ then the process transits to state $(0,0,h+1)$ (instead of $(0,0,h+\sh)$).\footnote{Technically, transiting to $(0,0,h+1)$ is realized by transiting to $(1,0,h+1)$ with probability $\alpha$ or to $(0,1,h+1)$ with probability $(1-\alpha)$.\label{footnote::technical}}
That this assumption works in favour of the attacker is clear: When $h'<h$, state $(\sh,\sa,h')$ is always preferable by the attacker over state $(\sh,\sa,h)$, for any $\sa,\sh$. Indeed, these two states differ only in that a successful attack in the latter state implies a successful attack in the former as well (but not necessarily vice versa), since the condition $\sh\ge \sigma^{total}_{\alpha,\gamma}\left(h\right)$ is stronger than $\sh\ge \sigma^{total}_{\alpha,\gamma}\left(h'\right)$. In particular, $(0,0,h+1)$ is preferable to the attacker over $(0,0,h+\sh)$.
Define $h_i=b^i\;(i=0,1,2....)$. Below we abbreviate $\sigma=\sigma^{total}_{\alpha,\gamma}$.
We define the $i$th epoch by the set of states with $h_i\le h<h_{i+1}$. The sequence $(h_i)$ satisfies the property that $\sigma\left(h_{i+1}\right)-1=\sigma\left(h_{i+1}-1\right)=\dots=\sigma\left(h_{i}\right)=C_{\alpha,\epsilon}+i$.
Let $p_i$ denote the probability that the attacker manages to perform a successful attack on a block belonging to the $i$th epoch. Suffice it to show that $\sum_{i=1}^\infty p_i<\epsilon$.
By definition, every attack in the epoch between $h_i$ and $h_{i+1}$ begins at a state of the form $(0,0,h)$ with $h_i\le h<h_{i+1}$ (an attack begins after the attacker abandoned its previous attempt, by an \ad, which leads to this state). A successful attack on block $u=C^t_h$ in the $i$th epoch can be reached only after at least $\sigma\left(h_i\right)$ blocks were built by the honest network above $b$ (as it requires $\sh\ge\sigma\left(h_i\right)$), including $u$, and in particular, after at least $\sigma\left(h_i\right)$ blocks were built since the creation of $u$'s predecessor (counting the blocks of both parties, and excluding $u$'s predecessor). For any number of steps $k\ge\sigma\left(h_i\right)$, the probability that after precisely $k$ steps $\sa\ge\sh$ is at most $e^{-c\cdot k}$: By putting $Z_i:=(X_i-1)/-2$, we arrive at a sequence of i.i.d random variables that take values in $\left\{0,1\right\}$. The event $\sum_{i=1}^{k}X_i\ge 0$ is then equivalent to $\sum_{i=1}^{k}Z_i\le k/2$, with $\mathbb{E}\left[\sum_{i=1}^{k}Z_i\right]=(1-\alpha)\cdot k$. We then apply to the latter sum Chernoff's bound: $\Pr\left(Z\le(1-\delta)\cdot \E\left[Z\right]\right)$, with $Z=\sum_{i=1}^{k}Z_i$ and $\delta:=\frac12\cdot\frac{1-2\alpha}{1-\alpha}$.
However, since the number of steps $k$ is not known in general, we upper bound the success probability of an attack on $u$ by $\sum\limits_{k=\sigma\left(h_i\right)}^{\infty}e^{-c\cdot k}=e^{-c\cdot \sigma\left(h_i\right)}\cdot\left(1-e^{-c}\right)^{-1}$.
By the union bound, the probability $p_i$ of a successful attack during the $i$th epoch can be upper bounded by $(h_{i+1}-h_i)\cdot e^{-c\cdot \sigma\left(h_i\right)}\cdot\left(1-e^{-c}\right)^{-1} < h_{i+1}\cdot e^{-c\cdot \sigma\left(h_i\right)}\cdot \left(1-e^{-c}\right)^{-1}=b^{i+1}\cdot e^{-c\cdot \sigma\left(h_i\right)}\cdot \left(1-e^{-c}\right)^{-1}$.
\emph{Part II:}
In order to upper bound the probability that there exists an epoch with a successful attack we apply the union bound on the entire sequence of epochs:
\begin{align}
& \sum\limits_{i=0}^{\infty}p_i\le \left(1-e^{-c}\right)^{-1}\cdot\sum\limits_{i=1}^{\infty} b_{\alpha}^{i+1}\cdot e^{-c\cdot \sigma\left(l_i\right)} = \\&
b_{\alpha}\cdot\left(1-e^{-c}\right)^{-1}\cdot \sum\limits_{i=1}^{\infty} b_{\alpha}^i\cdot e^{-c\cdot (C_{\alpha,\epsilon}+i)}=
\\& b_{\alpha}\cdot\left(1-e^{-c}\right)^{-1}\cdot e^{-c\cdot C_{\alpha,\epsilon}}\cdot \sum\limits_{i=1}^{\infty} \left(b_{\alpha}/\left(e^c\right)\right)^i= \\ &
\label{eq::geometricsum}b_{\alpha}\cdot\left(1-e^{-c}\right)^{-1}\cdot e^{-c\cdot C_{\alpha,\epsilon}}\cdot \sum\limits_{i=1}^{\infty} (b_{\alpha}/e^c)^i < \\ & \left(b_{\alpha}\cdot\left(1-e^{-c}\right)^{-1}\cdot\left(1-b_{\alpha}/(e^c)\right)^{-1}\right)
\cdot e^{-c\cdot C_{\alpha,\epsilon}}<\epsilon.
\end{align}
The last inequality holds by the choice of $C_{\alpha,\epsilon}$, whereas the geometric series in~(\ref{eq::geometricsum}) converges due to $b_{\alpha}<e^c$, by the choice of $b_{\alpha}$.
\end{proof}
Observe that our analysis allowed the attacker to \ov whenever the random walk visits the zero (as we have bounded the probability that $\sum_{i}X_i\ge 0$). Consequently, this bound applies to an attacker of any $\gamma$-value, as it already assumes that the attacker is always able to \ma successfully.
\section{The model}\label{sec::model}
We adopt the original setup analyzed by Satoshi Nakamoto, and later by Rosenfeld, that has become a standard model of Bitcoin's operation at the bound where block creation rates are much higher than the propagation time of blocks.
Miners in the Bitcoin network create blocks with exponential inter-arrival times, with parameter $\lambda$ (in Bitcoin, lambda is 1/600 blocks/second).\footnote{$\lambda$ is in fact controlled via the difficulty of the proof-of-work that is embedded in each valid block, and the exponential inter-arrival time is a good approximation given that the proof-of-work is based on guessing inputs to a cryptographic hash function that will cause its output to land within some narrow range. This process is nearly memoryless.} Unless otherwise stated, we assume that honest nodes remain connected and can always communicate, and that the mining rate $\lambda$ remains constant over time.
Each block contains a reference to a single predecessor block (a cryptographic hash). The entire history of blocks created up to time $t$ forms thus a tree, which we denote by $\T^t$. Nevertheless, the Bitcoin protocol dictates that the valid history of transactions consists of (transactions in) the longest chain of blocks alone. Accordingly, we assume honest participants only keep track of the longest chain they have been presented with, and do not maintain the entire tree structure. We further assume that blocks propagate in the network very fast relative to $1/\lambda$, and under this assumption the honest network's chain at every point $t$ in time is uniquely determined; we denote it by $\C^t=\left(C^t_0,C^t_1,C^t_2,C^t_3,...,C^t_{height(t)}\right)$ ($height(t)$ is thus the length of the honest chain at time $t$).
The block $C^t_0$ is a unique predetermined block that any chain must have at its root, and it is also called \emph{the genesis block}.
The height of a block $b$ is its distance from the genesis (with $height\left(C^t_0\right)=0$). We denote by $\past{b}$ ($\future{b}$) the set of blocks that precede (succeed) $b$ in the chain; note that $\future{b}$ keeps developing in time, as long as $b\in \C^t$.
The attacker is assumed to own an $\alpha$ fraction of the computational power, and the rest $(1-\alpha)$ is owned by honest nodes. Thus, the attacker creates blocks at a rate of $\alpha \cdot\lambda$, and the honest participants at a rate of $(1-\alpha)\cdot\lambda$. Following~\cite{OPT_SELFISH_MINING,ES}, we assume that the attacker has some communication capabilities that may allow it to transmit blocks that it has prepared in advance to nodes, just as the honest participants are starting to propagate a block that they have created. We denote the fraction of nodes that receives the attacker's block in this case by $\gamma\in [0,1]$. If $\gamma=0$, the attacker always loses block transmission races, and if $\gamma=1$ he wins them and is in fact able to get his block first to all honest miners.
Bitcoin nodes that participate in block creation efforts are called \emph{miners}. Upon mining a block, the miner embeds in it a set of Bitcoin transactions, created by users of the system. The transactions in $b$ must not double spend transactions in $\past{b}$.
\subsection{The acceptance policy}
A \emph{merchant} is any recipient, or beneficiary, of a Bitcoin transaction.
Upon receiving a bitcoin transaction, the merchant considers it as either accepted (in case which it releases the good or service paid for), or not accepted---the latter, if it is not sufficiently convinced that it will remain forever in the longest chain. To decide this, the merchant, or ``defender'', uses an acceptance policy. We restrict our attention to acceptance policies that use only structural information that is available to a merchant that was offline, namely, the current chain $\C^t$.
We further assume that the defender is currently connected to the honest nodes, and is not isolated by an attacker.
We define ``the acceptance policy'' of the defender as a function $\sigma_{\alpha,\gamma}:\mathbb{N} \to \mathbb{N}$. The function takes as input the height of the block containing the transaction, and returns the number of blocks that must be on top of it (including itself, i.e., $\size{\future{b}}+1$) before accepting the transaction. These blocks are often referred to as \emph{confirmations}. Thus, if the merchant sees a block $b$, and $\future{b}+1=\sigma_{\alpha}(h(b))$, then the transaction is considered accepted.
Perhaps the most commonly used policy for accepting transactions is the constant policy $\sigma_{\alpha,\gamma}(h) \equiv k$ that requires a certain number of confirmations, independently of the location of the transaction in the chain. The number of confirmations, $k$, is often expressed as a function of $\alpha$ (and in our case $\gamma$ as well) to ensure the security of the chain. Our modeling, which allows for dependency on the block's height, will be justified in Section~\ref{sec::logarithmic}.
\subsection{The attack policy}
We follow~\cite{OPT_SELFISH_MINING} and define the attacker's policy as a function that determines the action of the attacker at every possible state. The attacker is assumed to be building a secret branch of the chain which he will use to later override the honest network's current chain. The attacker may take one of several actions:
\begin{itemize}
\item \ad -- it abandons its attack, and its future chains will contain the tip of the honest network's current chain.
\item \ov -- it overrides the honest network's current chain by publishing a strictly longer chain.
\item \ma -- it publishes a chain of the same length as the honest network's current one.
\item \wa -- the null action which waits for future events (i.e., block creations)
\end{itemize}
\subsection{Security properties}
We define three robustness notions that correspond to the different security guarantees suggetsed above.
For a block $b$, and acceptance policy $\sigma$, the event where $b$ is accepted by $\sigma$ is given by $\mathcal{E}_{accepted}(b) := \left\{\exists t: b=C^t_h \wedge height(t)\ge h+\sigma(h)\right\}$. Denote by $time_{accepted}(b)$ the time at which $\mathcal{E}_{accepted}(b)$ first occurred, with $time_{accepted}(b)=\infty$ if it never has occurred. The event where it is later removed from the longest chain is $\mathcal{E}_{attacked}(b):=\left\{\exists s: time_{accepted}(b)<s<\infty, b\notin \C^s\right\}$.
\begin{defn}\label{def_arbitrary}
An acceptance policy $\sigma_{\alpha,\gamma}$ is $\epsilon$-arbitrary-robust, or robust against an attack on an arbitrary transaction, iff for any fixed height $h$ and $t$ such that $height(t)\ge h$, for any attacker with parameters $\alpha$ and $\gamma$, and under any attack policy $\pi_A$:\begin{equation}\label{eq_def_arbitrary}
\Pr\left(\mathcal{E}_{attacked}(C^t_h) \mid \mathcal{E}_{accepted}(C^t_h) \right)<\epsilon.\end{equation}
\end{defn}
\begin{defn}\label{def_fractional}
An acceptance policy $\sigma_{\alpha,\gamma}$ is $\epsilon$-fractional-robust, or robust against a removal of non-negligible portions of blocks, iff for any attacker with parameters $\alpha$ and $\gamma$, and under any attack policy $\pi_A$:\begin{equation}\label{eq_def_fractional}
\lim\limits_{t\rightarrow\infty}\frac{\sum_{b\in \T^{t}}\Pr\left(\mathcal{E}_{attacked}(b)\right)}{height(t)}<\epsilon.\end{equation}
\end{defn}
\begin{defn}\label{def_total}
An acceptance policy $\sigma_{\alpha,\gamma}$ is considered $\epsilon$-totally-robust, or resilient to a double spend anywhere in the chain, iff for any attacker with parameters $\alpha$ and $\gamma$, under any attack policy $\pi_A$:
\begin{equation}\label{eq_def_total}
\Pr\left(\exists b\in \T^{\infty}:\mathcal{E}_{attacked}(b)\right)<\epsilon.\end{equation}
\end{defn}
In this paper, we introduce three families of acceptance policies, $\sigma^{arb}$, $\sigma^{frac}$, and $\sigma^{total}$ that are $\epsilon$-arbitrary-robust, $\epsilon$-fractional-robust, and $\epsilon$-totally-robust, respectively, for any $\epsilon>0$.
\section{Defending independently generated transactions} \label{sec::premining}
In this section we find policies that can be used by a defender to guarantee that transactions cannot be double spent, assuming that their timing cannot be controlled by the attacker. We show a strategy that waits for a constant number of confirmations (depending still on $\alpha$ and $\epsilon$) which is $\epsilon$-arbitrary-robust. In our analysis we fix some flaws in analysis done in previous works, specifically, we more precisely account for blocks mined before the attack.
Under the assumption that the attacker was \emph{not} involved in selecting the time at which the transaction appeared in the blockchain $\C^t$, we are able to provide a distribution over the number of blocks that it has prepared in advance and on any lead that it may have relative to the network (as it could not have conditioned the payment on some rare event). It is then possible to analyze when this transaction could be considered effectively irreversible.
This guarantee is applicable, for example, when the transaction is a minting transactions, whose timing is determined by the time at which the miner created a block.
For the moment, we focus our analysis on the case $\gamma=0$. We define an acceptance policy $\sigma^{arb}_{\alpha}=\sigma^{arb}_{\alpha,0}$ as follows:
\begin{defn}
Let $\sigma^{arb}_{\alpha}:=\min\left\{n\in\mathbb{N} : f(n,\alpha)<\epsilon\right\}$, \\where
\begin{align}\label{eq::number_of_conf}
f(n,\alpha):=&\sum\limits_{l=0}^{\infty}\frac{1-2\cdot\alpha}{1-\alpha}\cdot\left(\frac{\alpha}{1-\alpha}\right)^l\cdot \\ \nonumber &\left(\sum\limits_{m=0}^{n-l}\binom{m+n-1}{m}\cdot\alpha^m\cdot(1-\alpha)^n\cdot\left(\frac{\alpha}{1-\alpha}\right)^{n+1-m-l}+\right.\\&
\left.\sum\limits_{m=n-l+1}^{\infty}\binom{m+n-1}{m}\cdot\alpha^m\cdot(1-\alpha)^n\right) \nonumber
\end{align}
\end{defn}
\begin{theorem}
For all $\epsilon>0$, the policy $\sigma^{arb}_{\alpha}$ is $\epsilon$-arbitrary-robust.
\end{theorem}
\begin{proof}
Let $b=C^t_h$ such that $len(t)\ge h$.
Let $C^{pub}_{t}$ be the longest chain in the published tree up to time $t$, and let $C^{oracle}_{t}$ be the longest chain in the entire tree, including secret attack-blocks.
For any $z\in C^{pub}_{t}$, put $R^z_{t}:=\size{\future{z}\cap C^{oracle}_{t}\setminus C^{pub}_{t}}$ and $Q^z_{t}:=\size{\future{z}\cap C^{pub}_{t}}$. We call $\max_{z\in C^{pub}_{t'}}\left\{R^z_{t'}-Q^z_{t'}\right\}$ the \emph{pre-mined gap} of the attacker at time $t'$. In Lemma~\ref{lem::phase1} we show that a random variable $Y$ with distribution vector $\Pr(Y=n)=\frac{1-2\cdot\alpha}{1-\alpha}\cdot\left(\frac{\alpha}{1-\alpha}\right)^n$ stochastically dominates (first-order) the distribution of the random variable $\max_{z\in C^{pub}_{t'}}\left\{R^z_{t'}-Q^z_{t'}\right\}$.
Now, a necessary condition for a successful attack is that, above some $z\in\past{b}$, the attacker has managed to create a chain which is longer than the published chain above $z$ by the time when the attack is released (i.e., when the secret blocks are published). Therefore, the maximal pre-mined gap of the attacker over a block $z\in \past{b}$, by $time(b)$, can be upper bounded by a random variable with the distribution $(p_n)$. Since the creation of $b$'s predecessor, the attacker's gap over any $z\in\past{b}$ follows an ordinary random walk with drift towards negative infinity (with different $z$'s in $\past{b}$ corresponding to possibly different starting points, all bounded together by $(p_n)$). The probability that the attacker advanced $m$ blocks during the period at which the honest network created $n$ confirmation-blocks is given by $\binom{m+n-1}{m}\cdot(1-\alpha)^n\cdot\alpha^m$. The event in which the attack will succeed is then equivalent to the event that the walk will ever arrive at $X=-1$ (here we used the restriction to $\gamma=0$). For a given pre-mined gap of size $l$, this happens with probability 1, if $m>n-l$, and with probability $\left(\frac{\alpha}{1-\alpha}\right)^{m-n-l+1}$, if $m\le n-l$ (this can be derived, e.g., using a martingale method. See also in~\cite{MENI}). Altogether, we have thus shown that $f(n,\alpha)$, as defined in Equation~\ref{eq::number_of_conf}, upper bounds the probability that $b$ will ever be reversed.
\end{proof}
\begin{lem}\label{lem::phase1}
For a fixed time $t'$,
$$\Pr\left(\max_{z\in C^{pub}_{t'}}\left\{R^z_{t'}-Q^z_{t'}\right\} \ge n \right)\le \left(\frac{\alpha}{1-\alpha}\right)^n.$$
\end{lem}
\begin{proof}
If an attacker aims to maximize the value of $\max_{z\in C^{pub}_{t'}}\left\{R^z_{t'}-Q^z_{t'}\right\}$', then its optimal strategy is as follows: It begin mining at time $t=0$, right after the creation of the genesis block, and whenever $\sh<\sa$ it performs \ad, resetting the attack above the tip of $\sh$. To see that this is optimal, simply observe that if the honest network has a positive lead over the attacker at time $t$, then by adopting $z_{tip}$ the attacker will have at least as high as a gap (i.e., $R^z_{t'}-Q^z_{t'}$) over $z=z_{tip}$ than it would have had above any other $z\in\past{z_{tip}}$ had it not adopted $z_{tip}$.
Consider now a random walk on the non-negative integers with a reflecting barrier at the origin: At position $k$, the probability to move one step to the right is $\alpha$, and to the left is $(1-\alpha)$. At the origin, the probability to move to the right is $\alpha$ and to stay in place is $(1-\alpha)$.
the transition probability matrix $P$ is accordingly. Denote by $Y_n:=\sum_{i=0}^{n}X_i$ the location of the walk after $n$ steps were made. The stationary distribution of the process $(Y_n)$ is $p_n:=\left(\frac{1-2\cdot\alpha}{1-\alpha}\right)\cdot \left(\frac{\alpha}{1-\alpha}\right)^{n}$, because if $Y$ is the distributed according to the limiting distribution then, for $n>0$: \begin{align*}
&\Pr\left(Y=n\right)=(1-\alpha)\cdot \Pr\left(Y=n+1\right)+\alpha\cdot\Pr\left(Y=n-1\right) = \\& (1-\alpha)\cdot p_{n+1}+\alpha\cdot p_{n-1} = (1-\alpha)\cdot \left(\frac{1-2\cdot\alpha}{1-\alpha}\right)\cdot \left(\frac{\alpha}{1-\alpha}\right)^{n+1} \\& +\alpha\cdot \left(\frac{1-2\cdot\alpha}{1-\alpha}\right)\cdot \left(\frac{\alpha}{1-\alpha}\right)^{n-1} = \ratio^n\cdot\frac{1-2\cdot\alpha}{1-\alpha}=p_n,
\end{align*}
and for $n=0$: $\Pr\left(Y=0\right)=(1-\alpha)\cdot \Pr\left(Y=0\right)+(1-\alpha)\cdot\Pr\left(Y=1\right)$, implying $\Pr(Y=0)=\frac{1-2\cdot \alpha}{1-\alpha}=p_0$.
Denote by $t_i$ the creation time of the $i$th block in $C^{oracle}_{t_{acc}}$.
We claim that $\max_{z\in C^{pub}_{t_i}}\left\{R^z_{t_i}-Q^z_{t_i}\right\}$ has the same probability distribution as $Y_i$. We prove it by an induction on $i$. For $i=0$, $t_0=0$. At time $0$, following the creation of the $genesis$ block, the value of $\max_{z\in C^{pub}_{t_i}}\left\{R^z_{t_i}-Q^z_{t_i}\right\}=\left(R^{genesis}_0-Q^{genesis}_0\right)$ is 0, as $\size{\future{genesis}\cap C^{oracle}_0}=0$; and likewise $Y_0=0$. Assume we have proved this for $i$, and we now prove it for $i+1$. With probability $\alpha$, the attacker creates the block at time $t_{i+1}$. In that case, it increases by 1 $R^{z}_{t_i}-Q^{z}_{t_i}$ for every $z\in C^{pub}_{t_i}$, and in particular $\max_{z\in C^{pub}_{t_i}}\left\{R^z_{t_i}-Q^z_{t_i}\right\}$ increases by 1; likewise, $Y_{i}$ increases by 1 with probability $\alpha$, since this is the probability of the $(i+1)$th step being towards positive infinity.
With probability $(1-\alpha)$, the honest network created the $(i+1)$th block. In that case, $R^z_{t_i}-Q^z_{t_i}$ decreases by 1 for every block $z\in C^{pub}_{t_i}$, whereas for the new block, $R^{z_{t_{i+1}}}_{t_i}-Q^{z_{t_{i+1}}}_{t_i}=0-0=0$. Thus, the value of $\max_{z\in C^{pub}_{t_{i+1}}}\left\{R^z_{t_i}-Q^z_{t_i}\right\}$ is the maximum between $\max_{z\in C^{pub}_{t_{i}}}$ $\left\{R^z_{t_i}-Q^z_{t_i}\right\} -1$ and 0. Similarly, if $Y_i>0$ then with probability $(1-\alpha)$ the $(i+1)$th step is towards negative infinity decreasing its value by 1, whereas if $Y_i=0$ then with probability $(1-\alpha)$ the value of $Y_{i}$ remains 0 at te $(i+1)$th step.
As argued above, the attacker does not lose anything from beginning its attack above the genesis block (recall its aim is to maximize the success-probability and not to minimize costs). Moreover, the process $(Y_n)$ forms an ergodic process: It is aperiodic because at the origin there's a positive probability to stay in place, and it is positive recurrent since the walk has a positive biased towards negative infinity.
Moreover, it can be shown that the stationary distribution of this process stochastically dominates (first-order) the distribution of $Y=(Y_n)$: $\Pr\left(Y=n\right) =\left(\frac{1-2\cdot\alpha}{1-\alpha}\right)\cdot\left(\frac{\alpha}{1-\alpha}\right)^n$. In particular, $\Pr\left(\max_{z\in C^{pub}_{t'}}\left\{R^z_{t'}-Q^z_{t'}\right\} \ge n \right)\le \left(\frac{1-2\cdot\alpha}{1-\alpha}\right)\sum_{k=n}^{\infty}\left(\frac{\alpha}{1-\alpha}\right)^k=\left(\frac{\alpha}{1-\alpha}\right)^n$.
\end{proof}
Note that the analysis above assumed $\gamma=0$, as the success probbaility of an attack in case of a tie was given by $\frac{\alpha}{1-\alpha}$.
The bound can be generalized to be valid for all $\gamma>0$ by waiting for an additional confirmation. This completes the description of the family $\sigma^{arb}_{\alpha,\gamma}$.
The formulas provided in the definition of $\sigma^{arb}_{\alpha}$ do not give a good feel for the security for the results. Following~\cite{MENI,SATOSHI}, we present the results in a table for some representative values.
Table~\ref{table::premining} illustrates the number of confirmations needed for an attacker of various sizes. This is to be contrasted with the table appearing in~\cite{MENI}, where the author assumed a constant pre-mining of 1 block before the transaction is transmitted (thus the analysis there is not of an arbitrarily timed transaction).
\begin{table*}[ht]
\caption{The probability of a successful attack on an arbitrary block ($\epsilon-arbitrary-robustness$), given the attacker's hashrate ($\alpha$) and the number of confirmations the acceptance policy waits for ($conf$). The calculation includes consideration for pre-mining.}
\label{table::premining}
\centering
{\ \ }\\
{\footnotesize
\begin{tabular}{|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|K{10mm}|}
\hline $\alpha\backslash conf$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\ \hline
2\% & 0.24\% & 0.02\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% \\ \hline
6\% & 2.16\% & 0.42\% & 0.09\% & 0.02\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% \\ \hline
10\% & 5.98\% & 1.85\% & 0.60\% & 0.20\% & 0.07\% & 0.03\% & $\approx$0\% & $\approx$0\% & $\approx$0\% & $\approx$0\% \\ \hline
14\% & 11.66\% & 4.88\% & 2.11\% & 0.93\% & 0.42\% & 0.19\% & 0.09\% & 0.04\% & 0.02\% & $\approx$0\% \\ \hline
18\% & 19.13\% & 9.94\% & 5.32\% & 2.90\% & 1.60\% & 0.89\% & 0.50\% & 0.28\% & 0.16\% & 0.09\% \\ \hline
22\% & 28.27\% & 17.33\% & 10.89\% & 6.95\% & 4.48\% & 2.91\% & 1.91\% & 1.25\% & 0.83\% & 0.55\% \\ \hline
26\% & 38.90\% & 27.17\% & 19.36\% & 13.97\% & 10.17\% & 7.45\% & 5.49\% & 4.06\% & 3.01\% & 2.23\% \\ \hline
30\% & 50.70\% & 39.33\% & 30.98\% & 24.64\% & 19.73\% & 15.88\% & 12.84\% & 10.41\% & 8.46\% & 6.89\% \\ \hline
34\% & 63.23\% & 53.37\% & 45.55\% & 39.14\% & 33.81\% & 29.31\% & 25.49\% & 22.21\% & 19.39\% & 16.95\% \\ \hline
38\% & 75.80\% & 68.45\% & 62.25\% & 56.85\% & 52.09\% & 47.85\% & 44.03\% & 40.58\% & 37.45\% & 34.56\% \\ \hline
42\% & 87.35\% & 83.09\% & 79.31\% & 75.86\% & 72.68\% & 69.72\% & 66.95\% & 64.33\% & 61.83\% & 59.44\% \\ \hline
46\% & 96.26\% & 94.88\% & 93.61\% & 92.41\% & 91.27\% & 90.17\% & 89.10\% & 88.05\% & 86.99\% & 85.82\% \\ \hline
48\% & 98.98\% & 98.59\% & 98.23\% & 97.88\% & 97.54\% & 97.21\% & 96.88\% & 96.54\% & 96.15\% & 95.60\% \\ \hline
50\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% & 100\% \\ \hline
\end{tabular}
}
\end{table*}
\section{The Generalized Vector76 pre-mining attack}\label{sec::vec76}
In this section we present the generalized Vector76 attack (the original attack was suggested by a user named Vector76 in the bitcoinTalk forums to possibly explain a successful double spending attack against the MyBitcoin e-wallet~\cite{Vector76}).
The attack is a form of pre-mining attack that the attacker can work on in secret until he is guaranteed to be successful. In this case as well, hybrid methods that trade off a shorter preparation time in exchange for lower success probabilities exist. It further assumes that the victim is unable to relay blocks to the main network, e.g., if he uses a light weight client that receives cryptographic proof of the attack, but not the blocks themselves. The attack is then \emph{easier} to execute compared to a regular pre-mining attack, since it requires the attacker to generate less of a lead on the honest network, and in fact, in a reverse twist, relies upon the network to confirm the double spending payment.
The important aspect of this attack is that it draws a clear distinction between full nodes and light node implementations that do not relay blocks, and hence demonstrates that light nodes need in fact to wait for additional confirmations to be equally secure to full nodes.
The attack proceeds as follows:
\begin{enumerate}
\item The attacker starts working on a secret branch of the chain. It embeds the transaction $tx_1$ (that it later wishes to reverse) in this block.
\item If the defender requires $\sigma\equiv k$ confirmations, the attacker needs to build an additional $k-1$ blocks on top of the one containing $tx_1$ (for a total of $k$ confirmations). He attempts to do so, in secret.
\item If his branch of the chain is longer than that of the honest chain, at some point \emph{after} he has $k$ confirmations for $tx_1$, then it shows the $k$ confirmations to the lightweight client which then that accepts it as the legitimate chain, since it is the longest one.
\item The attacker then transmits a conflicting transaction $tx_2$ to the honest network. As the honest network is not aware of the attacker's chain, the former's chain will grow long enough for $tx_2$ to be accepted by all nodes (and eventually even the attacked one).
\end{enumerate}
Figure~\ref{fig::vector76_attack} depicts the attack. Again, notice that a crucial stage in the success of the attack is that the honest network does not adopt the block containing $tx_1$.
\begin{figure*}
{\centering
\includegraphics[scale=0.5]{figures/vector76_attack.eps}
}
\emph{As the attack begins the attacker starts working on a secret chain with $tx_1$ inside its first block (1). If the attacker's chain is too far behind it may restart the attack (2). The attacker manages to gain a lead of 1 blocks, but has the two confirmations on his $tx_1$ needed to convince the victim (3). He then reveals the secret chain to the victim (that does not relay it), and collects an item in exchange. He then transmits the double spending transactions $tx_2$ to the network which is then included in a block (4). The network continues to mine atop $tx_2$ and it eventually prevails (5).}
\caption{\label{fig::vector76_attack} The progression of a generalized Vector76 attack on a 2-confirmation defender}
\end{figure*}
\myitem{Difficulty of the attack} The requirement for success in the pre-mining phase (after which the attack succeeds with probability 1) is that after constructing $k$ blocks or more, the attacker has a lead of 1 block over the network's chain. In a successful regular attack (whether it contains pre-mining or not), against a similar $k$-confirmation defender, the network has constructed $k$ blocks on top of the transaction (including the block that includes it), and at some point the attacker succeeds to lead over the network, and he thus has at least $k$ such blocks of his own in his branch. It is therefore easy to see that events in which successful double spending attacks occur are strictly contained in events in which the generalized Vector76 attack is successful (and succeeds with probability 1). The above argument thus shows that light nodes are strictly less secure than regular nodes, and need to wait for more confirmations (we include below an analysis that quantifies the effect of waiting longer).
In contrast, a regular (generalized Finney) pre-mining phase that
leads to a successful attack with probability 1 has much stricter requirements: the attacker needs to lead by $k+1$ blocks over the network (again, here he can launch a Vector76 attack as well, since he has built at least $k$ blocks and leads over the network).
\myitem{Resets and attacker strategy}
The generalized Vector76 attack described above can in fact be improved if attackers pick a better policy regarding restarts of the attack. This policy can again be found by solving MDPs that reward successful attacks. We leave this for future work.
\myitem{The Vector76 attack against full nodes} The Vector 76 attack can also be applied to full nodes if the attacker can somehow manage to send his chain with $k$ confirmations to the victim while a similar length chain is being propagated through the network. In this case, even once the defender relays the block, the network will not adopt it, since it is of equal height to the one created by the attacker. The attacker naturally needs to time the transmission right, and if he misses, his original block may be adopted by some fraction of the honest nodes (a model with a parameter similar to $\gamma$ that is used in selfish mining may capture this).
\myitem{Analytical guarantees of security}
Below we provide an analysis of the security of lightweight nodes against the generalize Vector76 attack. As the transactions that the node accepts are conditioned
on the attack's current state, the arbitrary-block security guarantee does not apply. Fortunately, we can upper bound the safety level against any attack, as follows.
The policy defined below is of the form $\sigma^{spv}_{\alpha}=\sigma^{spv}_{\alpha,\gamma}$, and applies for all $\gamma$.
\begin{defn}\label{def::vec76defense}
Let $$\sigma^{spv}_{\alpha}:=\min\left\{k\in\mathbb{N} : g(k,\alpha)<\epsilon\cdot(1-\alpha)\right\},$$ where
\begin{align}\label{eq::vec76defense}
g(k,\alpha):=&\sum\limits_{l=0}^{\infty}\frac{1-2\cdot\alpha}{1-\alpha}\cdot\left(\frac{\alpha}{1-\alpha}\right)^l\cdot \\ \nonumber &\left(\sum\limits_{n=0}^{k+l}\binom{n+k-1}{n}\cdot\alpha^k\cdot(1-\alpha)^n+\right. \\ &\left.\sum\limits_{n=k+l+1}^{\infty}\binom{n+k-1}{n}\cdot\alpha^{n-l}\cdot(1-\alpha)^{k+l}\right) \label{eq::defened76}.
\end{align}
\end{defn}
\begin{theorem}
For any $\epsilon>0$, the policy $\sigma^{spv}_{\alpha}$ is $\epsilon$-fractional-robust.
\end{theorem}
While the technique used in the following proof upper bounds the success-probability of an attack on an arbitrary block, we stress that its result cannot be interpreted as a security-guarantee for an arbitrary block. Transactions in this scheme are explicitly conditioned on the attack's state before the transmission of the transaction to the merchant. Nonetheless, this theorem shows that the merchant can guarantee any $\epsilon$-fractional robustness, by waiting long enough.
\begin{proof}
Let $b$ be an arbitrary block of the attacker. Let $T\gg0$ and denote by $N_T$ the total number of blocks created in the system up to time $T$. Finally, denote by $I_b$ the indicator random variable of the event where block $b$ participates in a successful attack. Under the generalized Vector76 attack, the attacker never publishes its secret chain. Therefore, eventually, all of the transactions in the honest network's chain will be accepted, as it grows indefinitely. Additionally, whenever an attacker block participates in a successful attack, by definition, the policy must have accepted a transaction in that block. Therefore, assuming a roughly constant number of transactions per block, and denoting the entire set of attacker blocks by $att$, we obtain:
\begin{align*}
& \lim\limits_{t\rightarrow\infty}\frac{\sum_{u\in \T^{t}}\Pr\left(\mathcal{E}_{attacked}(u)\right)}{height(t)}= \\ & \lim\limits_{t\rightarrow\infty}\frac{\frac1t\cdot\sum_{u\in \T^{t}}\Pr\left(\mathcal{E}_{attacked}(u)\right)}{\frac1t\cdot height(t)}= \\ & \lim\limits_{t\rightarrow\infty}\frac{\Pr_b\left(\mathcal{E}_{attacked}(b)\right)}{\frac1t\cdot height(t)},
\end{align*}
where the probability here is also over the choice of $b$ in $T^{\infty}$. This identity holds due to the Law of Large Numbers. The last expression is upper bounded by $\frac{\Pr_b\left(\mathcal{E}_{attacked}(b)\right)}{1-\alpha}$, as the longest chain grows at least at the rate of the honest network's chain.
We now provide an upper bound on $\Pr_b\left(\mathcal{E}_{attacked}(b)\right)$.
Fix the attacker's policy $\pi_A$.
The attacker can utilize $b$ to carry out a generalized Vector76 attack if and only if, at some point in time, the following conditions are met: $\sa>\sh$ and the number of blocks in the attacker's chain above $b$ is at least $k$. Indeed, if the first condition is not met then the merchant will count zero confirmations for its transaction in the honest network's chain, and will not accept. Likewise, if the second one is not met, the merchant will not see $k$ confirmations in the attacker's chain, and will not accept. Assume that block $b$ has $k$ confirmations (including itself). The probability that, in a period of time in which the attacker created $k$ blocks, the honest network created $n$ blocks is given by $\binom{n+k-1}{n}\cdot\alpha^k\cdot(1-\alpha)^n$.
As proven in Section~\ref{sec::premining}, the lead that the attacker gained over the network's chain, prior to mining $b$, can be upper bounded by the distribution vector $(p_l)_{l=0}^{\infty}$, with $p_l=\frac{1-2\cdot\alpha}{1-\alpha}\cdot\ratio^l$. In consequence, given $n$, the probability that the attacker will ever succeed in bypassing the honest network's chain is 1, if $l+k\ge n$, and $\ratio^{n-(l+k)}$, if $n>l+k$ (here we used the worst-case assumption that $\gamma=1$). Therefore, the probability that $b$ participates in a successful attack is upper bounded by $g(k,\alpha)$, which is the sum of $p_l\cdot \binom{n+k-1}{n}\cdot\alpha^k\cdot(1-\alpha)^n\cdot\ratio^{(n-(l+k))^+}$ over all $l$ and $n$.
Therefore, the expression~(\ref{eq::vec76defense}) upper bounds the probability that an arbitrary block of the attacker will have $k$ confirmations and at the same time will be part of a chain longer than the honest network's.
All in all, we obtain that the fraction of attacker blocks (out of the total number of accepted blocks) can be upper bounded by $\frac{g(k,\alpha)}{1-\alpha}$. In particular, $\sigma^{spv}_{\alpha}:=g(n,\alpha,\epsilon)$ is $\epsilon$-fractional-robust.
\end{proof}
|
\section{Introduction}
IRC\,+10420 is a yellow hypergiant (YHG) star. These objects are evolved massive stars, which
present extreme initial mases and very high luminosities ($\mathrm{log}(L/\mbox{$L_{\odot}$})\sim$\,5.6, $M_{init}\,\raisebox{-.4ex}{$\stackrel{>}{\scriptstyle \sim}$}\, 20\mbox{$M_{\odot}$}$). In fact,
this particular YHG has a luminosity of
$L\sim 5 \times 10^5 \mbox{$L_{\odot}$}$ and has an estimated initial mass of $M_{init} \sim 50 \mbox{$M_{\odot}$}$ \citep[][]{tiffany2010,nieuwen00}.
Yellow hypergiants are post red-supergiant stars (RSGs) evolving toward higher temperatures
in the HR diagram. { In particular, the spectral type of \irc\ has changed from F8Ia to A5Ia in just 20 yr \citep{Klochkova97}.}
These objects are thought to lose as much as one half of their initial masses during the RSG phase \citep[e.g.,][]{MM88}.
In addition, during their post-RSG evolution the YHGs encounter an instability region, called
the yellow void \citep{dejager98}, which results in a new episode of effective mass ejection.
These outburst were recently detected for $\rho$ Cas \citep{lobel2003}.
As a result of these outbursts, the ionized wind of these stars becomes optically thick and the resultant stellar
wind spectrum mimics that of a lower $T_\mathrm{eff}$ star
\citep{CrossingYV}.
As the ejected material dilutes into the interstellar medium (ISM), the apparent $T_\mathrm{eff}$ increases again.
These apparent $T_\mathrm{eff}$ oscillations are frequently called bouncing against the yellow void.
Since the real $T_\mathrm{eff}$ continues increasing, at a certain moment, as the ejected material dilutes out, the YHG stars would
eventually appear at the { just beyond the high-temperature edge} of the yellow void.
However, \citet{Smith04} found that the { high-temperature edge} of the yellow void is coincident with the S Doradus instability strip.
These sources then form a pseudophotosphere that keeps the source in the { low-temperature edge} of the yellow void in the HR diagram.
Therefore, these authors suggested that the evolution of the YHGs remain hidden until they become slash stars and finally
enter in the Wolf-Rayet phase.
While these mass ejections are predicted to be very important for these objects,
only three YHGs, IRC+\,10420 and AFGL\,2343 \citep{cc07}, and
recently IRAS\,17163--3907 \citep{muller2015}, have shown molecular emission.
The kinematics and structure of the molecular envelopes around IRC+\,10420 and AFGL\,2343 were studied in detail by \citet{cc07}
thanks to high angular resolution interferometric maps of \doce.
They found that, while these objects showed slight departures from the spherical symmetry, the data could
be reasonably modeled by adopting an isotropic mass loss with large variations with time.
In particular, for the YHG IRC+\,10420 they found a detached circumstellar envelope (CSE) with an extent
of $5 \times 10{^{17}}$cm expanding at velocities of $\sim$37 km/s.
Two strong mass ejection episodes, which occurred within a lapse of
1200 years and reached a mass loss rate of 3 $\times 10^{-4}$\mbox{$M_{\odot}$~yr$^{-1}$}, are responsible for the formation of this CSE.
The total
mass derived for the CSE around this object is $\sim 1 \mbox{$M_{\odot}$}$.
\citet{gqtesis} showed that the ejection of this material could be explained in a similar way as the ejections
presented by the low mass AGB stars (i.e., the mass ejection is driven by radiation pressure on the dust grains). Also,
\citet{gqtesis} argued that all the molecular material observed by \citet{cc07} was only ejected
during the YHG phase. Any gas ejected during the previous RGS phase should have been rapidly diluted in the ISM
and photodissociated by ISM ultraviolet (UV) radiation field.
The chemistry of IRC\,+10420 is particularly rich \citep{chemyhg}.
The following species were detected in the CSE around this object: CO, $^{13}$CO, HCN, CN, H$^{13}$CN, SiO, $^{29}$SiO, SO, SiS, HCO$^{+}$,
CN, HNC, HN$^{13}$C, and CS. Surprisingly, some species, such as HCN and HNC, showed particularly high abundances
compared with the O-rich AGB stars, which are the low mass counterpart of YHGs \citep[see, e.g.,][]{bujarrabal94}.
This is explained by an enrichment in nitrogen due to the hot bottom burning process \citep{boothroyd93},
which is present for stars with
masses above $\sim$3 \mbox{$M_{\odot}$}. This process transforms $^{12}$C into nitrogen, which changes the composition of the material that is
transported to the photosphere and later on expelled. The nitrogen enrichment is expected
to lead to a N-rich chemistry for the most massive
and evolved stars. This was recently confirmed by \citet{NO10420}. These authors found an abnormally
high abundance of NO in IRC\,+10420, directly related to an enhancement of the elemental abundance of nitrogen.
On the other hand, SiO emission was found to come from regions
located at 10$^{17}$\,cm from the star, far from the radius where dust formation takes place
and where the SiO is expected to be largely depleted in the gas \citep{ccsio}. These authors suggested
that this SiO emission is related to a spherical shock front that heats up the grains,
releasing some amount of Si back to the gas phase.
Recently, \citet{teyssier2012} observed selected transitions of NH$_3$, OH, H$_2$O, CO, and $^{13}$CO toward this object with HIFI.
They also showed that the model derived by \citet{cc07} for low-$J$ CO transitions
was applicable, with minor changes, to high excitation lines.
In this paper we present a full line survey of IRC\,+10420
obtained with the IRAM 30 m telescope in the atmospheric windows
centered at wavelengths of 1\,mm and 3\,mm. The survey
confirms that the chemistry of this object is particularly rich.
We detected 22 molecular species for which we determined the column
density and rotational temperature.
\section{Observations}
We used the IRAM 30 m radiotelescope to obtain a complete line
survey at 3\,mm and 1\,mm for the YHG IRC\,+10420.
We observed IRC\,+10420
at the position coordinates (J2000) 19\h26\m48$\mbox{\rlap{.}$^{\rm s}$}$10,
+11$\rm^o$21${'}$17\secp0 with an $\rm{v_{LSR}}$ = 76\,km/s.
The observations were obtained during February 2012. We used the EMIR receiver, simultaneously
using the receivers E090 and E230 with a bandwidth of 4\,GHz at two
polarizations.
We used 16 different setups to cover the atmospheric windows at
3\,mm and 1\,mm. Each setup was observed for one hour.
We used the wobbler switching mode to minimize the ripples in the baselines.
The system temperatures
during the observations were in the range 100--250\,K for the E090 receiver
and between 200 and 425\,K
for the E230 receiver. The weather conditions during the observations
were good with an amount of precipitable water vapor ranging between
2 and 6 mm.
The backends used were WILMA (spectral resolution 2\,MHz) and the 4\,MHz
filter bank. Owing to the large width of the line profiles
observed in IRC\,+10420 ($\sim$60 km/s), this resolution is enough to
resolve the line profiles.
The pointing correction was checked frequently and, therefore, we expect
pointing errors of $\sim$3''. The telescope beam size at 3\,mm is 21--29$''$ and 9--13$''$ at 1\,mm.
The Atmospheric Transmission Model (ATM) is adopted
at the IRAM 30 m \citep{Cerni85,Pardo01}.
The data presented is calibrated in antenna
temperature ($T^*_A$). The calibration error is expected to be 10\% at 3\,mm and
30\% at 1\,mm.
The data were processed using the GILDAS package\footnote{See URL
http://www.iram.fr/IRAMFR/GILDAS/}. The baselines were subtracted using
only first grade polynomials.
\section{Line identification}
The full spectra obtained at 3 and 1\,mm are presented in Figs.\ref{Fig3mm}\&\ref{Fig1mm}.
The sideband rejection of the EMIR receiver is higher than 10\,db
and therefore only the strongest lines present their counterparts
in the image band. The spectral features whose origin is
the image band are labeled in red characters in Figs. \ref{Fig3mm}\&\ref{Fig1mm}.
The line identification was performed using the online catalogs of CDMS\footnote{http://www.astro.uni-koeln.de/cdms/catalog} \citep{cdms},
JPL\footnote{http://spec.jpl.nasa.gov/} \citep{jpl}, Splatalogue\footnote{http://www.splatalogue.net/},
and the catalog incorporated into the radiative
transfer code MADEX \citep{MADEX}.
In Tables A.1\&A.2, we list all the lines detected in the line survey in the 3\,mm and
1\,mm bands, respectively. In these tables we also present
the velocity-integrated intensity of the lines, the peak intensity and the {rms of the noise level} for a spectral
resolution of 2\,MHz. As a result of the large width of the lines of \irc,
the hyperfine structure of the transitions of CN,
NO, and NS cannot be resolved. For these transitions we present
the global integrated intensity including all hyperfine components.
Thanks to the large frequency coverage of the spectral survey, we were able to confirm
the presence of the species detected by inspecting their relative strengths expected for local thermodynamic equilibrium
(LTE) conditions.
We detected a total of 106 spectral lines, which we identified as arising from 22 species
including different isotopologues.
The list of the species detected can be found in Table\,1.
\begin{table}
\caption{Molecular species detected}
\label{table:1}
\centering
\begin{tabular}{l c | c c }
\hline\hline
Molecule & Isotopologues & Molecule & isotopologues \\
\hline
CO & $^{13}$CO &HCN &H$^{13}$CN\\
SiO & $^{29}$SiO, $^{30}$SiO,Si$^{18}$O&HNC&HN$^{13}$C\\
SiS&&CH$_3$OH&\\
SO&$^{34}$SO&CN&\\
SO$_2$&&PN&\\
CS&&NO&\\
HCO$^{+}$&&NS&\\
N$_2$H$^{+}$&&&\\
\hline
\end{tabular}
\end{table}
\section{LTE modeling of the line emission.}
Once we successfully identified the emission lines observed, we aimed to estimate the
physical conditions of the regions where these species are located.
In order to estimate the excitation temperatures and column densities
of the different species observed we used a LTE approach as that described by \citet{rtd}.
These authors showed that the rotational diagram of a certain molecule follows the
next relationship.
\begin{equation}
\mathrm{ln} \left(\frac{N_\mathrm{u}}{g_\mathrm{u}}\right) =
{\mathrm{ln}} \left(\frac{N}{Z(T_\mathrm{rot})}\right) - \mathrm{ln}\ C_{\tau} - \frac{E_\mathrm{u}}{kT_\mathrm{rot}}
,\end{equation}
\noindent where $N_\mathrm{u}$ and $g_\mathrm{u}$ are the total column density and statistical weight
of the upper level, respectively, $N$ is the total column density, $T_\mathrm{rot}$ is the rotational temperature,
$Z(T_\mathrm{rot})$ is the partition function, $E_\mathrm{u}$ is the energy of the upper level,
and $C_{\tau}$ is the term that takes the opacity effects into account,
which corresponds to $C_{\tau} = \tau / (1-e^{-\tau})$. The value of $\mathrm{ln}({N_\mathrm{u}}/{g_\mathrm{u}})$ is
directly proportional to the integrated intensity of the molecular line observed \citep{rtd}.
Here we use the opacity-corrected column density $N'_\mathrm{u} = N_\mathrm{u}\ C_{\tau}$. Therefore, relation
(1) can be rewritten as
\begin{equation}
\mathrm{ln} \left(\frac{N'_\mathrm{u}}{g_\mathrm{u}}\right) =
{\mathrm{ln}} \left(\frac{N}{Z(T_\mathrm{rot})}\right) - \frac{E_\mathrm{u}}{kT_\mathrm{rot}}
.\end{equation}
This relationship allows us to estimate $T_\mathrm{rot}$ and $N$ of a molecule by fitting a straight
line to the rotational diagram.
As an approximation for the opacity, we use that presented by \citet{chemyhg},
\begin{equation}
\tau = \mathrm{ln}\ \left( \frac{1}{1-\frac{T_\mathrm{mb}}{S_\nu}\frac{\Omega_\mathrm{A}}{\Omega_\mathrm{S}}} \right)
,\end{equation}
\noindent where $T_\mathrm{mb}$ is the peak temperature in main beam temperature scale, $S_\nu$ is the source function, and
$\Omega_\mathrm{S}/\Omega_\mathrm{A}$ is the geometrical dilution factor between source
size and telescope main beam.
{ We adopt the source size from \citet{chemyhg}, i.e., an angular radius of
11\secp0 for} \doce\ and \trece, and 3\secp3 { for the rest of the molecular species.
{ We also took this beam dilution correction into account in the calculation of ${N_\mathrm{u}}/{g_\mathrm{u}}$,
in particular by correcting the integrated intensity of each molecular transition.}
The source function is calculated as follows:}
\begin{equation}
S_\nu = \frac{h\,\nu}{k}\left( \frac{1}{e^{\frac{h\,\nu}{k\,T_\mathrm{ex}}}-1} - \frac{1}{e^{\frac{h\,\nu}{k\,T_\mathrm{bg}}}-1} \right)
,\end{equation}
\noindent where $T_\mathrm{ex}$ is the excitation temperature and $T_\mathrm{bg}$ is the temperature of the cosmic background.
The term $C_\tau$ is different for each rotational transition and
depends on $T_\mathrm{rot}$ via the source function. As a result of this, $T_\mathrm{rot}$ cannot
be directly obtained by fitting a straight line to equation (1). In order to solve
this problem, we used an iterative procedure. We introduce an initial rotational temperature in the opacity term (1000\,K) and
fit to the expression in Eq.\,(1). The value of $T_\mathrm{rot}$ obtained is then introduced as the new temperature to compute
the opacity corrections. This process was repeated until the difference between the
temperature introduced in the opacity term and that derived was below a certain tolerance (0.1\,K).
At $T_\mathrm{rot} = 1000$\,K, the opacity term $\mathrm{ln}(C_\tau)$ is in general negligible at the
frequencies covered by the surveys presented and
$N_\mathrm{u} = N'_\mathrm{u}$. In that sense, assuming an initial rotational temperature of 1000\,K for the calculation of the
opacity is equal to assume $\tau << 1$.
An example of the effect of the opacity correction on the rotational diagram is shown in Fig.\,\ref{tau}.
The opacity correction applied can only account for moderate values of the optical
depth. For high opacities the line emission we detect
comes only from the outer layers of the circumstellar envelope. Because of this, we impose a higher limit to the opacity of
$\tau = 2$.
Furthermore, in case the opacity is remarkably high, and if there is a large opacity variation for the different
transitions, the method fails to obtain estimates of $T_\mathrm{rot}$. As shown by \citet{rtd},
when the opacity is high, the rotational diagram given by eq.\,(1) deviates from the straight line expected for low opacities
The optical depth of the transition does not vary proportionally to the energy of the transition.
These authors showed that
for linear molecules in LTE conditions the $J_\mathrm{up, \tau max}$ of the transition with the maximum opacity is $J_\mathrm{up, \tau max} = 4.6
\sqrt{T_\mathrm{rot}(K)/B_\mathrm{o}(\mathrm{GHz})}$ at a frequency of $\nu(\mathrm{GHz})_{\tau max} = 9.13 \sqrt{T_\mathrm{rot}(K) B_\mathrm{o}(\mathrm{GHz})}$.
Therefore, depending on the $T_\mathrm{rot}$ of the gas and of the transitions observed, the opacity effects
could lead to both higher or lower slopes for the fitting than that expected for low opacities.
Only in those cases in which we can estimate the $T_\mathrm{k}$ of the
gas where a particular emission
arises by any other method, as deriving it from fitting the rotational diagrams of low-abundance isotopologues, can
we obtain a limit to the column density of a given molecule.
This can be accomplished for instance by ignoring those lines available
with frequencies closer to $\nu(\mathrm{GHz})_{\tau max}$.
In case we only observe two transitions, we can obtain a realistic derivation of
the value of $T_\mathrm{rot}$ if the opacity from both lines
is relatively low. If the optical depth significantly
affects the intensity of the lines, however, we could only obtain a first
approximation to the value of $T_\mathrm{rot}$ assuming optically thin emission.
We use the molecular distribution of the two shells described by \citet{cc07}.
{ As mentioned}, in a similar approach as \citet{chemyhg}, we assume that the molecular emission from all molecules but
those of CO and $^{13}$CO come from the inner shell. In particular, the diameter of the emitting region would be
11$''$ for CO and $^{13}$CO and 3\secp3 for the rest of the molecules.
\subsection{Column density and rotational temperature determination}
As expected, the method described above is not well suited for certain molecules because of the high opacity of some
of their transitions. These molecules are HCN, H$^{13}$CN, SiO, $^{29}$SiO, and $^{30}$SiO.
In the particular cases of SiO and their isotopologues,
the origin of the emission of these molecules was claimed to be related to the presence of a shock front \citep{ccsio}.
However, recently \citet{teyssier2012} were not able to detect high-$J$ lines with HIFI. In addition, no vibrationally excited lines
were detected in the current survey. This shows that the excitation temperature of the SiO emission is not be particularly
high. Since we cannot estimate the kinetic temperature of the gas where this emission arises, we cannot rely on the method proposed
to derive the abundance and excitation temperature of this molecule, but just to constrain these values.
We assumed that the emission is optically thin to obtain a
lower limit to the abundance.
Also, since only two lines were detected for the rest of the molecules, which presented opacity effects that avoided the convergence of the
procedure described above (HCN, H$^{13}$CN),
we also assumed that their emission is optically thin.
\begin{figure}[h!]
\centering
\includegraphics[angle=0,width=9cm]{SO_diftau.eps}
\caption{Rotational diagram for SO. Empty squares represent the values for the SO transitions
assuming a negligible opacity. Filled squares show the values for $N_\mathrm{u} / g_\mathrm{u}$
corrected for opacity effects.}
\label{tau}%
\end{figure}
For those molecules for which only a single line was observed, we had to impose a certain value of
$T_\mathrm{rot}$ to estimate the column density.
In case the molecule has an isotopologue for which we could determine the $T_\mathrm{rot}$,
we assumed the same $T_\mathrm{rot}$ value for both isotopologues. In the rest of the cases,
the value of $T_\mathrm{rot}$ was imposed to be the mean $T_\mathrm{rot}$ value obtained for all the
molecules located in the same region of the envelope, assuming the molecule distribution proposed by \citet{chemyhg}. As mentioned above, we
assume that the emission from all molecules but $^{12}$CO and $^{13}$CO
arises from the inner shell observed by \citet{cc07}. We found
that $\langle T_\mathrm{rot} \rangle = 11.7 \pm 5.7\,K$ for this inner inner
shell.
The column density for these molecules was obtained by modeling the observed spectra assuming LTE
conditions with the radiative transfer code MADEX.
In the case lines with hyperfine structures (CN, NO, and NS), we used the LTE approach instead of the rotational diagram analysis in MADEX to reproduce the
observed profiles. Since MADEX computes the line opacity, we were able to estimate whether the column densities obtained
are a lower limit to the value of $N$.
We also confirmed that the results obtained for the column density and $T_\mathrm{rot}$ using the rotational diagrams were
compatible with the results derived adopting the LTE approximation within the MADEX code.
In the particular case of HCO$^{+}$, although only one line was
detected, we could estimate a lower limit to the column density
and an upper limit to the rotational temperature thanks to an upper
limit to the intensity of the $J= 3-2$ transition, lying at 267.557\,GHz.
The rotational diagrams are presented in Fig.\,\ref{rtd} for those species for which we estimated the opacity and in
Fig.\,\ref{rtd2} for those molecules for which we assumed an optically thin regime.
The $N$ and $T_\mathrm{rot}$ results obtained are presented in Table 2.
{ In this table, the error of the column density
is only presented
for those species with more that two transitions detected and for which the opacity has not been imposed by a certain value.
For the species with only two transitions detected, the error derived is unrealistically low, as the only source of error in the fitting
of the two points by a line is the error in the intensity of the line, which in general was very low. The species, as SiS or SO, for which
a larger number of species were observed, present more realistic errors and therefore they are prompted in the table. }
{ We used MADEX to generate
synthetic spectra with the parameters presented in table 2 to produce a global estimate of the error in the abundance determination
for the rest of the species. The uncertainty in the determination of the abundance was
found to be on average $\sim$15\%.}
\begin{table*}
\caption{Column densities and rotational temperatures. $^*$: tau fixed to a value of 10$^{-3}$. }
\centering
\begin{tabular}{l c c c c c}
\hline\hline
Molecule & $N$ (cm$^{-2}$) & $T_{rot}$ (K) & $\langle X \rangle$&N. Obs. Line & Comment\\
\hline
CO & $2.6 \times 10^{17}$ &$24.9\pm 4.3$ & $5.3 \times 10^{-4}$ &2 & \\
$^{13}$CO & $3.7 \times 10^{16}$ &$56.4\pm8.9$ & $7.5 \times 10^{-5}$ &2 & \\
SiO & $>4.8 \times 10^{15}$ &$10.0\pm2.2$ & $>1.3 \times 10^{-6}$ &3 & $^*$ \\
$^{29}$SiO & $>8.3 \times 10^{14}$ &$9.7\pm2.9$ & $>2.8 \times 10^{-7}$ &3 & $^*$ \\
$^{30}$SiO & $>8.1 \times 10^{15}$ &$9.1\pm0.3$ & $>2.7 \times 10^{-7}$ &3 & $^*$ \\
Si$^{18}$O & $1.7 \times 10^{14}$ &$9.1\pm0.3$ & $5.7 \times 10^{-8}$ &1 & \\
HCN & $>3.4 \times 10^{15}$ &$>2.4\pm0.1$ & $>1.1 \times 10^{-6}$&2 & $^*$ \\
H$^{13}$CN & $>4.2 \times 10^{14}$ &$>6.0\pm0.3$ & $>1.4 \times 10^{-7}$ &2 & $^*$ \\
HNC & $2.9 \times 10^{14}$ &$11.3\pm0.9$ & $9.7 \times 10^{-8}$ &2 & \\
HN$^{13}$C & $8.0 \times 10^{13}$ &$11.3\pm0.9$ & $2.7 \times 10^{-8}$&1 & \\
CN & $4.0 \times 10^{15}$ &$5.0\pm0.5$ & $1.3 \times 10^{-6}$ &2 & \\
CS & $3.7 \times 10^{14}$ &$7.9\pm0.4$ & $1.2 \times 10^{-7}$ &2 & \\
PN & $1.1 \times 10^{14}$ &$7.9\pm0.4$ & $3.7 \times 10^{-8}$ &2 & \\
NO & $3.0 \times 10^{16}$ &$9.0\pm0.1$ & $6.7 \times 10^{-6}$ &1 & \\
NS & $5.0 \pm 0.5 \times 10^{15}$ &$8.0\pm0.5$ & $1.7 \pm 0.2 \times 10^{-6}$ &4 & \\
SiS & $2.1 \pm 0.1 \times 10^{14}$ &$20.5\pm2.0$ & $7.0 \pm 0.5 \times 10^{-8}$ &5 & \\
SO & $3.3 \pm 1.2 \times 10^{16}$ &$11.4\pm1.2$ & $1.1 \pm 0.4 \times 10^{-6}$ &9 & \\
$^{34}$SO & $9.0 \times 10^{14}$ &$11.4\pm1.2$ & $2.4 \times 10^{-7}$&1 & \\
SO$_2$ & $1.0 \pm 0.4 \times 10^{15}$ &$22.1\pm5.2$ & $3.4 \pm 1.0 \times 10^{-7}$ &14 & \\
HCO$^{+}$ & $>2 \times 10^{13}$ &$<9.2\pm0.5$ & $>6.7 \times 10^{-9}$&2 & \\
N$_2$H$^{+}$ & $4.4 \times 10^{13}$ &$11.7\pm5.7$ & $1.5 \times 10^{-8}$&1 & \\
CH$_3$OH & $2.1 \times 10^{14}$ &$8.6\pm0.5$ & $7.0 \times 10^{-8}$&2 & \\
\hline
\end{tabular}
\end{table*}
\section{Results}
The results obtained for the column density
in the previous section allow us to understand the characteristics
of the chemical processes occurring in the CSE around the massive evolved star IRC\,+10420 in greater detail.
To obtain the fractional abundance of each molecule relative to H$_2$, $X$, we
adopted the density profile presented by \citet{cc07}. This density profile allows us to calculate
the column density of H$_2$ at each region of the shell and, therefore, to determine the fractional
abundance of each molecule as $X_\mathrm{mol} = N_\mathrm{mol}/N_\mathrm{H_2}$.
{ These fractional abundances are also shown in Table\,2.}
We discuss the results for the species detected separately.
\subsection{$^{12}$CO and $^{13}$CO}
In their work, \citet{cc07}
assumed that to derive the H$_2$ density profile of the envelope around \irc, the fractional abundance of $^{12}$CO in IRC\,+10420 was
the standard abundance found for { AGB stars}, i.e., $3 \times 10^{-4}$.
From this assumption, these authors derived the amount of mass ejected by the star.
Since we used the density profiles obtained by these authors,
we expected to find similar values for the fractional abundance of \doce\ to those assumed
by \citet{cc07}.
In fact, the ratio between both $^{12}$CO abundances is $\sim 1.8$, which is higher than that
derived in this work.
The LVG modeling approach carried out by \citet{cc07} estimates the excitation more accurately than
the LTE approach. This factor provides an estimate of the error expected by the method used in this work,
compared with more accurate radiative transfer models.
{ The $^{12}$CO abundance derived here is similar to that obtained toward the RSG VY\,CMa by \citet{ziurys09}
for the red flow, while those abundances obtained by these authors for the spherical flow and blue flow are lower.}
Even though the \doce\ abundance is somehow fixed, and since the opacity effects do not critically affect the density fitting of both $^{12}$CO and $^{13}$CO, we can
estimate the $^{12}$C/$^{13}$C ratio.
It is found to be $\sim$7, which is slightly below the range found by \citet{cisotopic} for O-rich evolved stars ($^{12}$C/$^{13}$C $= 10-35$).
This was expected in any case, as the hot bottom burning process transforms $^{12}$C in $^{14}$N in massive stars as \irc.
\subsection{CN, HCN, and HNC}
The fractional abundance obtained for CN in this object is $1.3 \times 10^{-6}$.
This abundance is high compared with
O-rich evolved stars ($6.6 \times 10^{-8}$), as already noticed by \citet{cnbachiller}.
This value is 3.4 times higher than that derived in \citet{chemyhg}. In that work we used an approximation to
simplify the hyperfine structure of CN to a simple structure {of two Hund rotational levels
without hyperfine structure}. This approximation is only reasonable
for optically thin emissions, however, the approach presented in the current work
reveals that the hyperfine transitions present moderate opacities.
Therefore, in this particular case, the most accurate approach to estimate the abundance and rotational temperature of CN is that presented
in this work, {i.e., generating a synthetic spectra to fit the different hyperfine lines.}
Since the formation of CN is mainly produced by the photodissociation of HCN and HNC
due to UV radiation, a high CN value is expected for those objects presenting a hot central
star surrounded by very diluted material, as in the proto-PN phase. While this could be the case for the
yellow hypergiant IRC\,+10420, the ratio $X_\mathrm{CN}/X_\mathrm{HCN}$, which traces the
UV field, is remarkably low ($\sim$ 0.12). In addition, \citet{Alcolea2013} found a weak emission
of HCN $J$= 13--12, which reveals that HCN is not highly excited in this CSE.
The photodissociation of HCN in IRC\,+10420 might not be very effective because of the high density of the ejected material
\citep[$n \sim 2 \times 10^4$ cm$^{-3}$,][]{cc07}.
The reason for the high abundance of CN, as well as for that of HCN, is most probable an enhanced nitrogen abundance
in the photosphere of the star (see below).
In contrast, the fractional abundance obtained for HNC is $9.7 \times 10^{-8}$.
The opacities obtained for the HNC lines are relatively low (0.2 for HNC $J=1-0$ and 0.5 for HNC $J=3-2$)
and, therefore, the abundance obtained is probably a good approximation to reality.
This value is slightly higher than
the mean values obtained for the O-rich AGB stars \citep[][$\langle X_\mathrm{HNC} \rangle = 8.2 \times 10^{-8}$]{bujarrabal94}.
The $^{12}$C/$^{13}$C ratio derived from the abundances obtained for HNC and its $^{13}$C isotopologue
is $\sim$ 3.6. As mentioned in Sect.\,5.1., this ratio is remarkably low compared with the
standard value found for the AGB stars, which is consistent with the low value deduced from CO (Sect.\,5.1).
{ It is also particularly relevant to compare these abundances
with those of the RSG VY\,CMa, as IRC\,+10420 is expected to be a post-RSG object.
The lower limit for the HCN abundance found for \irc\ is similar to the lowest value found
for VY\,CMa \citep{ziurys09}. Also, the HNC abundances found are similar for both objects.
On the contrary, the abundance found for CN by these authors is $\sim$72 times lower in VY\,CMa than in IRC\,+10420.
This is most probably a sign of the nitrogen enrichment of the photosphere of the star along its
RSG -- post-RSG evolution.}
\subsection{Si-bearing species}
The presence of
{ SiO in most of the evolved stars (AGBs, RSGs, and YHGs) is believed to be restricted to the inner layers}
of the circumstellar envelopes
where the dust is
being formed \citep[see, e.g.,][]{lucas92}. The refractory materials, such as silicon, are rapidly attached to the grains that
are formed and removed from the gas.
\citet{cc07} showed that the amount of molecular gas in the inner regions of the circumstellar envelope
around \irc, where the SiO emission would be expected to arise, is negligible. Because of this, in principle,
the anticipated intensity of the Si-rich lines would be low. Despite this, theses species are particularly abundant in
\irc .
In \irc the SiO emission was found to come from a shell located
at $\sim 10^{17}$cm \citep{ccsio}. Because of the large distance from the star where SiO emission is located,
these authors suggested that SiO emission is actually tracing a shocked region, where the dust grains have
been heated, evaporating part of the silicon attached to them.
{ In their study of the molecular emission of IK\,Tau \citet{Gobrecht16} found that while both SiS and SiO are
destroyed by the shock, SiS molecules are reformed in the post-shocked regions with the same abundance as before it is destroyed.
This might indicate that both Si-bearing species can be located in different regions of the ejecta of the evolved stars.}
The lines observed of SiO and its isotopologues are optically thick, as shown by \citet{chemyhg}. Therefore, the values obtained for the
column densities are lower limits.
The LTE approximation is not an accurate option for SiO to determine the
physical conditions of the region where this molecule is located. Solving the radiative transfer equations
using an LVG approximation would be a better option (see Sect.\,6.1).
On the contrary, the SiS emission is only affected by small opacity effects.
While, as already mentioned, the origin of both SiO and SiS emission is different in the AGB stars and in \irc a comparison of the abundance ratio between both Si-bearing molecules from an O-rich evolved star and \irc\ gives a
similar result of { $X_\mathrm{SiO}/X_\mathrm{SiS} \raisebox{-.4ex}{$\stackrel{>}{\scriptstyle \sim}$} 10$}.
{ At the distances from the star at which \citet{ccsio} found the SiO shell, the reaccretion of the SiO in the grains
would be extremely slow because of the dust dilution lasting hundreds of years. Therefore, once these molecules evaporated from the grains, the
SiO/SiS abundance ratio would be expected to be similar to that found in the innermost layers of the star.
\subsection{Cations: N$_2$H$^{+}$ and HCO$^{+}$.}
The formation of these two cations is strongly coupled to the cosmic-ray
ionization rate of H$_2$ ($\zeta$) via the cation H$_3^{+}$. Chemical modeling
calculations similar to those carried out by \citet{NO10420} and \citet{sanchezcontreras15} indicate that at
the low edge of the range of values of $\zeta$ in the Galaxy (10$^{-17}$--10$^{-15}$ s$^{-1}$;
\citealt{Ionizationrate}), the formation of HCO$^{+}$ in O-rich envelopes is dominated by the reactions
\begin{equation}
\rm CO^{+} + H_2 \rightarrow HCO^{+} + H, \label{reac:co+_h2}
\end{equation}
\begin{equation}
\rm C^{+} + H_2O \rightarrow HCO^{+} + H, \label{reac:c+_h2o}
\end{equation}
where CO$^{+}$ is formed by the reaction between C$^{+}$ and OH. At high values
of $\zeta$, HCO$^{+}$ is mainly formed by the reaction
\begin{equation}
\rm H_3^{+} + CO \rightarrow HCO^{+} + H_2,
\end{equation}
where H$_3^{+}$ is formed in the reaction between H$_2$ and H$_2^{+}$, where the latter
is produced by the cosmic-ray ionization of H$_2$. In contrast, the main
formation route to N$_2$H$^{+}$, independent of the value of $\zeta$, is the reaction
\begin{equation}
\rm H_3^{+} + N_2 \rightarrow N_2H^{+} + H_2.
\end{equation}
These chemical calculations indicate that in IRC~+10420, where there is an
important nitrogen enhancement \citep{NO10420}, HCO$^{+}$ should be much more
abundant than N$_2$H$^{+}$ if $\zeta$ is on the order of 10$^{-17}$ s$^{-1}$,
while if it is approximately 10$^{-15}$ s$^{-1}$ both cations could reach comparable
abundances. A high ionization rate could be driven by cosmic rays but also
by soft X-rays emitted by the central stars. In fact, \citet{zhang7027} found
that the abundance of N$_2$H$^{+}$ in the planetary nebulae NGC\,7027 was abnormally
high compared with the rest of evolved stars, and this abundance was related to its particularly
strong X-ray field.
In addition to this, HCO$^{+}$ can also be formed in the presence of an ionizing
shock front by reactions~(\ref{reac:co+_h2}) and (\ref{reac:c+_h2o}), as found
by \citet{Rawlings2004}.
In the case of \irc\,we find a lower limit for the abundance of HCO$^{+}$ of
$6.7 \times 10^{-9}$. This lower limit for the abundance corresponds to the upper limit obtained for
the HCO$^{+}$ $J= 3-2$ transition. In contrast, the abundance derived for N$_2$H$^{+}$ is
$1.5 \times 10^{-8}$. The abundance ratio $X_\mathrm{N_2H^{+}}/X_\mathrm{HCO^{+}}$
is $\raisebox{-.4ex}{$\stackrel{<}{\scriptstyle \sim}$}\, 2.2$. This value is very high compared with NGC~7027, where
$X_\mathrm{N_2H^{+}}/X_\mathrm{HCO^{+}} = 0.07$ \citep{zhang7027}. In fact, the
N$_2$H$^{+}$ abundance derived in \irc\ is the second highest value found in an evolved
star. The highest fractional abundance has been found for the extreme massive
star Eta Carinae \citep[$X_\mathrm{N_2H^{+}} = 2\times 10^{-7}$;][]{Loinard12}.
The high N$_2$H$^{+}$/HCO$^{+}$ abundance ratio found in \irc\ points to a high
ionization rate driven by cosmic rays or X-rays, although shocks could also
be at the origin of these two molecules. In the first hypothesis, the formation
of both cations would be chemically coupled, as H$_3^{+}$ would participate in both
cases, and we could expect a similar emission distribution for HCO$^{+}$ and N$_2$H$^{+}$.
Whether the main ionization mechanism is cosmic rays or X-rays is difficult to say.
Recently, \citet{xrays} did not detect X-rays in IRC\,+10420 using the XMM-Newton
Space telescope. This does not necessarily imply that X-rays are not an important
ionizing source in this object as they could be obscured to a large extent by the
massive envelope. In the second hypothesis, the emission of these two cations would
tend to follow that of other molecules such as SiO, whose emission in \irc\
probably related to a shock front \citep{ccsio}. Additional single-dish observations
of N$_2$H$^{+}$, in particular $J = 2-1$ and
$J=3-2$, would allow us to determine the temperature of gas traced by N$_2$H$^{+}$
and, therefore,,accurately disentangle its origin. If the N$_2$H$^{+}$ emission is a
consequence of the interaction of X-rays on the innermost regions of the ejected
material the kinetic temperature of the gas traced by this molecule should be high,
while the temperature should be low if its origin were the interaction of the cosmic rays with gas at the outer
regions of the envelope .
In any case, the particularly high abundance of N$_2$H$^{+}$ found for \irc\ corresponds directly to a high abundance of N$_2$.
\subsection{CH$_3$OH}
We have identified two lines of CH$_3$OH. The number of molecular transitions of this molecule
that fall within the
observed ranges is relatively large (35 lines with excitation temperatures above 500\,K in the 3\,mm
and 47 in the 1\,mm band).
In order to confirm that the two lines actually correspond to CH$_3$OH, we
used the column density and rotational temperature derived from the rotational diagrams
to produce
a synthetic spectra
with the MADEX code, assuming LTE conditions, to reproduce the intensity of the rest of the nondetected CH$_3$OH lines
from the column density and temperature obtained from the rotational diagram.
{ We found that all the remaining lines either present an intensity that prevented their detection or
are blended with other more intense lines}. Furthermore, the model fits a feature at 241.8\,GHz, which was not identified
and was found to be related to the presence of several CH$_3$OH lines in these region of the spectra (see Fig.\,\ref{ch3oh241}).
The identification of these lines confirm the presence of CH$_3$OH in this object and the abundance and temperature derived for this molecule.
{ It has been suggested that the formation of CH$_3$OH might take place in warm
environments \citep{Hartquist95}, such as the environment found in the CSE around IRC\,+10420.
}
\begin{figure}[h!]
\centering
\includegraphics[angle=0,width=9cm]{ch3oh_241m.eps}
\caption[]{Spectral feature that corresponds to CH$_3$OH emission lines in the range 241.8\,GHz.}
\label{ch3oh241}%
\end{figure}
\subsection{N-rich chemistry}
The most remarkable result of this survey is the detection of a wide number of
N-bearing molecules. These molecules are HCN, HNC, PN, NS, NO, CN, and N$_2$H$^{+}$ plus some isotopologues.
In addition to these molecules, \citet{Menten1995} detected NH$_3$.
Our abundance estimates revealed high abundances found for CN, HCN, HNC, and N$_2$H$^{+}$ when compared with
O-rich evolved stars \citep[see, e.g.,][]{bujarrabal94}. In particular, \irc\ presents{ one of the highest} N$_2$H$^{+}$
abundance found for an evolved star.
These facts could be explained by an enrichment in nitrogen due to the hot bottom burning process \citep[HBB;][]{boothroyd93}
which, as mentioned, is present for stars with
masses above $\sim$3 \mbox{$M_{\odot}$}.
For the most massive and evolved stars, the nitrogen enrichment is expected
to lead to a N-rich chemistry. In addition, the $^{12}$C/$^{13}$C isotopic ratio is therefore expected to be
low, as has been found for \irc\ (see Sect.\,5.1\&5.2).
Recently, \citet{NO10420} has shown that to reproduce the NO
profiles, the initial abundance of nitrogen in the photosphere of the star has to be
significantly enhanced with respect to the standard values expected for the O-rich evolved stars,
i.e., the enhancement expected as a result of the HBB is confirmed.
{ We obtained a lower limit to the column density of P$^{15}$N to compare the $^{14}$N/$^{15}$N isotopic ratio with that found for
N-rich Nova CK\,Vul by \citet{kaminsky15}. The lower limit of the integrated intensity was calculated
following \citet{chemyhg}. We choose this molecule as it does not present hyperfine transitions and since the PN emissions
detected are optically thin. We found an upper limit for the integrated intensity $W < 4 \times 10^{-3}$ K\,\mbox{km~s$^{-1}$}, which results
in a column density of $N < 4 \times 10^{13}$ cm$^{-2}$. The upper limit of the isotopic ratio so derived is $^{14}$N/$^{15}$N $>$ 3.
This low value is similar to the range found by \citet{kaminsky15} for CK\,Vul ($4-26$). Deeper integrations toward selected
$^{15}$N-bearing molecules would allow us to confirm this ratio.}
A paper devoted to a detailed study of the N-rich chemistry in \irc\ is under preparation.
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,width=14cm]{10420_10yr.eps}
\caption[]{Comparison between the spectra observed by \citet{chemyhg} (black lines) and
that are presented in this work (red lines). The temperatures are in mean beam temperature scale.
}
\label{Comp}%
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,width=14cm]{comp_2000.eps}
\caption[]{SiO line fitting for the observations presented in \citet{chemyhg}.
The temperature derived for the dust shell is 750\,K.
The line fitting is depicted by a red line. In the upper left corner
we present the calibration factor applied to the observed spectra.}
\label{Comp2000}%
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,width=14cm]{comp.eps}
\caption[]{SiO line fitting for the observations presented in this work.
The temperature derived for the dust shell is 550\,K.
The line fitting is depicted by a red line. In the upper left corner
we present the calibration factor applied to the observed spectra.}
\label{Compnow}%
\end{figure*}
\section{Time variations of the SiO lines}
The abundances we found for SiO and $^{29}$SiO are $\sim$9 and $\sim$4 times lower that those found by \citet{chemyhg},
respectively.
A careful comparison between the line profiles obtained by \citet{chemyhg} and those presented here reveal
that while the intensities of most of the molecular lines do not present significant changes, the $J$=5--4 lines of SiO and $^{29}$SiO are
much weaker (see Fig.\ref{Comp}).
The yellow hypergiants are known to present large variations in their effective temperatures \citep{nieuwen00}.
These changes in their effective temperatures would result in a remarkable variations in the emission of certain
lines due to the IR pumping. SiO lines are particularly affected by this effect. A decrease in the $T_{\mathrm{eff}}$
of IRC\,+10420 would result in a decrease of the SiO emission, especially for the high-$J$ lines.
A second possible explanation for the variation in the SiO line emission is that the silicon evaporated from the
dust grains during the shock is reattached to the grains. However, { as mentioned above,} at the large distances where \citet{ccsio} located
the SiO emission, the re-absorption of SiO would last hundreds of years.
\citet{nieuwen00} found that the $T_{\mathrm{eff}}$ of IRC\,+10420 showed an increase of $\sim$ 3000\,K in the
last 20 years, reaching a value of 8500 in 1997.
These authors predicted the effective temperature of \irc\ to grow continuously in its evolution toward blue regions
of the HR diagram. The decrease in the intensity of the high-$J$ lines can confirm this scenario, as the dust ejected
during the last encounter with the yellow void dilutes into the ISM.
In order to obtain an accurate view of the properties of the SiO emission, we used an LVG approximation to
model the lines observed. We adopted the extent of the SiO shell from \citet{ccsio}, the density
profile from \citet{cc07}, and the SED model from \citet{Dinh09}.
We found that we are unable to reproduce the SiO emission from IRC\,+10420 assuming the
mass distribution derived by \citet{cc07}.
\citet{WongThesis} found a similar difficulty with the observed SiO $J$ = 1 -- 0 VLA maps. This author suggested that
the emission of SiO is located in clumps, which were ejected from localized regions of the photosphere of the star.
\citet{WongThesis} argue therefore that the properties of these clumps are not tied with the global mass ejection traced by CO.
This author suggested a two-shell model to reproduce the SiO $J=1-0 $ and $J=2-1$ emission maps. These maps consisted of a radial
density law, $n_\mathrm{H_2} = 2.3 \times 10^{4}$ cm$^{-3} \times (r/10^{17}$cm)$^{-0.7}$, and an abundance profile, which present two different
regions with a constant SiO abundance: $X_\mathrm{SiO} = 1.25 \times 10^{-6}$ between $5 \times 10^{16}$ cm and $2.35 \times 10^{17}$ cm, and
$X_\mathrm{SiO} = 6 \times 10^{-8}$ between $2.35 \times 10^{17}$ cm and $4.2 \times 10^{17}$ cm.
As already mentioned, the IR emission from the star has an important impact on the intensities of the SiO lines via the
infrared pumping.
The emission from the star is obscured by the optically thick dust layers. To reproduce the infrared emission from the star,
\citet{WongThesis} assumed that
the continuum source consisted in an optically thick shell of dust at 400\,K and at a radius of 8 10$^{15}$cm. This assumption was based on
the flux observed at 8$\mu{m}$ by \citet{Jones1993} in March 1992.
However, the IR flux from this object is known to vary and the low-$J$ SiO lines are less sensitive to
this variation. Therefore, while \citet{WongThesis} fitted the observations with a dust temperature of 400\,K, a higher temperature could
have also fitted the observations.
While the flux observed by \citet{Jones1993} at 8 $\mu{m}$ is $\sim$1000\,Jy, the flux observed with ISO in 2001 at this same wavelength
is only $\sim 350$\,Jy\footnote{Data obtained from the ISO data archive http://iso.esac.esa.int/ida/}. In fact, \citet{egan2003} showed that
in 1997 the flux at 8.28 $\mu{m}$ was 148.7\,Jy. This confirms the large flux variations that this object has undergone.
Therefore, \citet{WongThesis} probably overestimated the 8 $\mu{m}$ flux in their calculations of the abundance and temperature.
Taking into account these variations, we used the ISO observations from
year 2001 to fit the SiO lines presented by \citet{chemyhg}
observed in year 2000. The flux of 350\,Jy can be simulated, following
\citet{WongThesis}, by a blackbody with a radius of 8 10$^{15}$cm,
and a temperature of 331\,K. We also assumed the gas distribution of
\citet{WongThesis}.
Simultaneously we fitted the observations presented in the current work.
We left as free parameter the IR flux from the star for the 2012 observations, and
allowed minor modifications to the
SiO fractional abundance of both regions suggested by \citet{WongThesis}.
We found that in order to fit the two sets of data, we needed to decrease the SiO fractional
abundance in the inner region to 10$^{-5}$. The temperature of the dust shell
obtained for the new set of data is 200\,K, which corresponds to a 8 $\mu{m}$ flux
of $\sim$8\,Jy.
This reveals that an accurate knowledge on the
variations of the IR flux of these objects is crucial to obtain
realistic abundance estimates of species, such as
SiO, whose emission is strongly correlated with the IR emission.
The YHGs are known to present two main pulsation periods \citep{dejager98}. These consist of a
short period of several hundred of days corresponding to
a quiescent pulsation \citep{PercyKim14,lecoroller03}, and a larger period of
tens of years in which these objects present a quasi-explosive mass ejection \citep{dejager98}
A detailed study on the flux variation of this object, as well as high angular resolution observations of the SiO
distribution, are essential to understand the SiO emission observed in IRC\,+10420.
Also, future SiO observations are fundamental to confirm whether the source of the IR variation
is related to the expansion of the dust shell, which would result in a constant
decrease of this flux, or to the quiescent pulsation or a new explosive mass ejection
period, in case future observations present an increase of the SiO lines.
The first scenario would support the idea that \irc\ is finally crossing the yellow void,
while the latter would suggest that this object is undergoing a new
encounter with the yellow void.
Future observations of the SiO emission lines{, in addition to bolometric observations at 8 $\mu{m}$}
would allow us track this decrease in the IR flux from the dust{ and to better constrain the properties of the SiO emitting region.}
A new increase in the intensity of high-$J$ SiO emission lines could also be related to new mass ejections.
\section{Conclusions}
We have performed a full line survey in the atmospheric windows at 1 and 3\,mm for the yellow
hypergiant IRC\,+10420. We detected a total of 22 species. We have obtained LTE abundances for these species.
These abundances showed a particularly high value of all N-bearing species compared with standard O-rich evolved
stars. In particular, the abundance of N$_2$H$^{+}$ is found to be the second highest found for an evolved star.
These results seem to confirm the nitrogen enrichment found by \citet{NO10420}.
In addition to this, the isotopic $^{12}$C/$^{13}$C ratio is particularly low ($\sim$7--3.6, see Sects.\,5.1\&5.2).
{ This difference cannot be explained by the errors in the fitting of the abundance (see Sect.4.1), therefore this
variation in the isotopic ration seems to be genuine.}
Both effects are expected
for massive stars in which hot bottom burning process takes place, transforming $^{12}$C into $^{14}$N.
Also, the nitrogen enrichment and the $^{13}$C/$^{12}$C ratio increase with time as the hot bottom burning
is active \citep{Marigo2007}. The outer regions of the envelope, where only CO was detected, trace gas ejected between 1200 and 3000
years before the ejection
of the innermost shell \citep{cc07}. Therefore we might expect a highest $^{12}$C/$^{13}$C. The differences in the ratios derived from CO and HNC might
be caused by opacity effects. However, for high opacities we might expect higher abundances of $^{12}$CO, and therefore a lower
$^{12}$C/$^{13}$C ratio. Most probably, the difference in the ratios found for both regions are due to the continuous effect of the HBB along
the last 3800 years. High angular resolution maps of optically thin emission lines of $^{12}$C and $^{13}$C species could confirm this result.
High-$J$ SiO line intensities observed in this paper are remarkably low compared with the same transitions observed by
\citet{chemyhg}, while the variation of the low-$J$ SiO lines observed is low. This variation was also observed in HCN and CS, while
HNC, CN, SiS, and HCO$^{+}$ line intensity have remained constant.
The effective temperature of the star has a direct effect on the SiO emission via infrared pumping
\citep[see, e.g.,][]{cerni2014c}.
The SiO line intensity variation probably corresponds to a recent decrease in the IR flux emitted from the
dust, which has a direct effect on the intensity of the SiO lines via infrared pumping.
\irc\ has been claimed to be evolving blueward in the HR diagram crossing the yellow void, i.e.,
to present a constant increase in its $T_\mathrm{eff}$.
The changes in the intensity of the SiO lines revealed a decrease in IR flux from the dust around the star.
Tracking the evolution of the IR flux of this object is important to determine its evolutionary status, i.e., if the source
is actually crossing the yellow void or if it is undergoing a new mass ejection.
{ Future SiO observations in combination with IR 8 $\mu{m}$ observations would allow us to trace the temperature of the dust
around evolved stars in particular around
massive evolved stars, which will allow us to disentangle their evolutionary status and to constrain the
properties of the SiO emitting region. }
\begin{acknowledgements}
The research leading to these results has received funding from the European Research Council
under the European Union's Seventh Framework Programme (FP/2007-2013) / ERC Grant
Agreement n. 610256 (NANOCOSMOS).
We would also like to thank the Spanish MINECO for funding support from grants CSD2009-00038,
AYA2009-07304 \& AYA2012-32032.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
The main result of this paper is a characterization of the smooth Schubert varieties of a rational homogeneous variety $G/P$ that are integrals of a canonically defined subbundle of $T(G/P)$. Here $G$ is a complex, semisimple Lie group, $P$ is a parabolic subgroup of $G$, and $T(G/P)$ is the holomorphic tangent bundle. Every rational homogeneous variety $G/P$ admits a unique minimal, bracket--generating, $G$--homogeneous subbundle $T^1 \subset T(G/P)$ of the (holomorphic) tangent bundle. The variety $G/P$ is cominuscule if and only if $T^1 = T(G/P)$; equivalently, $G/P$ admits the structure of a Hermitian symmetric space. These are the simplest rational homogeneous varieties; examples include the Grassmannian $\mathrm{Gr}(k,\mathbb C} \def\cC{\mathcal C^n)$ of $k$--planes in $\mathbb C} \def\cC{\mathcal C^n$. From the work of Lakshmibai--Weyman, Brion--Polo, and J.~Hong we have
\begin{theorem}[{\cite{MR1703350, MR2276624, MR1080976}}] \label{T:comin}
In the case that $G/P$ is cominuscule, the smooth Schubert varieties $X \subset G/P$ are homogeneously embedded cominuscule $G'/P'$, and are indexed by subdiagrams of the Dynkin diagram $\sD$ of $G$.
\end{theorem}
\noindent Hong--Mok generalized Theorem \ref{T:comin}, using the deformation theory to prove
\begin{theorem}[{\cite{HongMok2013}}] \label{T:HM}
Suppose that $G$ is simple and $P$ is a maximal parabolic with the property that the associated simple root is not short. Then the smooth Schubert varieties $X \subset G/P$ are homogeneously embedded $G'/P'$, and are indexed by subdiagrams of the Dynkin diagram $\sD$ of $G$.
\end{theorem}
The main result (Theorem \ref{T:smooth}) of the paper is another generalization of Theorem \ref{T:comin}. We say that a Schubert variety $X \subset G/P$ is ``horizontal'' if it is tangent to $T^1$ at smooth points (\S\ref{S:HSV}).
\begin{theorem} \label{T:smooth}
Let $X \subset G/P$ be a horizontal Schubert variety. If $X$ is smooth, then $X$ is a product of homogeneously embedded, rational homogeneous subvarieties $X(\sD') \subset G/P$ corresponding to subdiagrams $\sD' \subset \sD$. Moreover, each $X(\sD')$ is cominuscule.
\end{theorem}
\noindent The proof of Theorem \ref{T:smooth} is based on the characterization of horizontal Schubert varieties (HSV) in \cite{MR3217458}, and Theorem \ref{T:HM}. Note that, in general, a smooth Schubert variety $X \subset {G/P}$ need not be homogeneous (Example \ref{eg:SG}). As will be seen in the course of the proof, the subdiagrams $\sD'$ may be explicitly characterized.
As an application of the main result we characterize the maximal parabolics $P \subset G$ with the property that the Schubert variety swept out by the horizontal lines passing through a point is itself horizontal. This, and the material developed in \S\ref{S:lines}, is motivated by anticipated applications to a current project of the authors \cite{KR1}.
\section{Horizontal Schubert varieties}
\subsection{Flag manifolds} \label{S:rhv}
Let $G$ be a connected, complex semisimple Lie group, and let $P \subset G$ be a parabolic subgroup. The homogeneous manifold $G/P$ admits the structure of a rational homogeneous variety as follows. Fix a choice of \emph{Cartan} and \emph{Borel subgroups} $H \subset B \subset P$. Let $\mathfrak{h}} \def\olfh{\overline{\mathfrak h} \subset \mathfrak{b}} \def\tfb{{\tilde\fb} \subset\mathfrak{p}} \def\olfp{\overline{\mathfrak p}\subset{\mathfrak{g}}} \def\tfg{\tilde\fg$ be the associated Lie algebras. The choice of Cartan determines a set of \emph{roots} $\Delta = \Delta({\mathfrak{g}}} \def\tfg{\tilde\fg,\mathfrak{h}} \def\olfh{\overline{\mathfrak h}) \subset \mathfrak{h}} \def\olfh{\overline{\mathfrak h}^*$. Given a root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in \Delta$, let ${\mathfrak{g}}} \def\tfg{\tilde\fg^\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \subset {\mathfrak{g}}} \def\tfg{\tilde\fg$ denote the \emph{root space}. Given a subspace $\mathfrak{s} \subset {\mathfrak{g}}} \def\tfg{\tilde\fg$, let
$$
\Delta(\mathfrak{s}) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in\Delta \ | \ {\mathfrak{g}}} \def\tfg{\tilde\fg^\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \subset \mathfrak{s} \} \,.
$$
The choice of Borel determines \emph{positive roots} $\Delta^+ = \Delta(\mathfrak{b}} \def\tfb{{\tilde\fb}) = \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in\Delta \ | \ {\mathfrak{g}}} \def\tfg{\tilde\fg^\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \subset \mathfrak{b}} \def\tfb{{\tilde\fb} \}$. Let $\mathscr{S} = \{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_1,\ldots,\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_r\}$ denote the \emph{simple roots}, and set
\begin{equation} \label{E:I}
I \ = \ I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ i \ | \ {\mathfrak{g}}} \def\tfg{\tilde\fg^{-\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_i} \not\subset \mathfrak{p}} \def\olfp{\overline{\mathfrak p} \}\,.
\end{equation}
Note that
$$
I(\mathfrak{b}} \def\tfb{{\tilde\fb}) \ = \ \{1,\ldots,r\} \,,
$$
and $I = \{\mathtt{i}} \def\sfi{\mathsf{i}\}$ consists of single element if and only if $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ is a \emph{maximal parabolic}.
Let $\{ \omega} \def\olw{\overline \omega_1,\ldots,\omega} \def\olw{\overline \omega_r\}$ denote the \emph{fundamental weights} of ${\mathfrak{g}}} \def\tfg{\tilde\fg$ and let $V$ be the irreducible ${\mathfrak{g}}} \def\tfg{\tilde\fg$--representation of highest weight
\begin{equation} \label{E:mu}
\mu \ = \ \mu_I \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \sum_{i \in I} \omega} \def\olw{\overline \omega_i \,.
\end{equation}
Assume that the representation ${\mathfrak{g}}} \def\tfg{\tilde\fg \to \mathrm{End}} \def\tEuc{\mathrm{Euc}(V)$ `integrates' to a representation $G \to \mathrm{Aut}(V)$ of $G$. (This is always the case if $G$ is simply connected.) Let $o \in \mathbb P} \def\cP{\mathcal P V$ be the \emph{highest weight line} in $V$. The $G$--orbit $G \cdot o \subset \mathbb P} \def\cP{\mathcal P V$ is the \emph{minimal homogeneous embedding} of $G/P$ as a rational homogeneous variety.
\begin{remark}[Non-minimal embeddings] \label{R:nonminemb}
More generally, suppose that $V$ is the irreducible $G$--representation of highest weight $\tilde \mu = \sum_{i \in I} a^i\,\omega} \def\olw{\overline \omega_i$ with $0 < a^i \in \bZ$. Again, the $G$--orbit of the highest weight line $o \in \mathbb P} \def\cP{\mathcal P V$ is a homogeneous embedding of $G/P$. We write $G/P \hookrightarrow \mathbb P} \def\cP{\mathcal P V$. The embedding is minimal if and only if $a^i=1$ for all $i \in I$. For example, the Veronese re-embedding $v_d(\mathbb P} \def\cP{\mathcal P^n) \subset \mathbb P} \def\cP{\mathcal P\,\tSym^d\mathbb C} \def\cC{\mathcal C^{n+1}$ of $\mathbb P} \def\cP{\mathcal P^n$ is if minimal if and only if $d = 1$. (Here $V = \tSym^d\mathbb C} \def\cC{\mathcal C^{n+1}$ has highest weight $d\omega} \def\olw{\overline \omega_1$.)
\end{remark}
The variety $G/P$ is sometimes indicated by circling the nodes of the Dynkin diagram (Appendix \ref{S:dynkin}) corresponding to the index set $I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p})$.
\subsection{Schubert varieties} \label{S:schub}
In the case that $G$ is classical (one of $\mathrm{SL}} \def\tSO{\mathrm{SO}_n\mathbb C} \def\cC{\mathcal C$, $\tSO_n\mathbb C} \def\cC{\mathcal C$ or $\mathrm{Sp}} \def\tSpin{\mathit{Spin}_{2n}\mathbb C} \def\cC{\mathcal C$), the Schubert varieties of a flag variety $G/P$ may be described geometrically as degeneracy loci, cf.~ Example \ref{eg:schub1}. However, these descriptions, aside from not generalizing easily to the exceptional groups, are type dependent. We will utilize a representation theoretic description of the Schubert varieties that allows a uniform treatment across all flag varieties. This section does little more than establish notation for our discussion of Schubert varieties. The reader interested in greater detail is encouraged to consult \cite{MR3217458} and the references therein.
Given a simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_i\in\mathscr{S}$, let $(i) \in \mathrm{Aut}(\mathfrak{h}} \def\olfh{\overline{\mathfrak h}^*)$ denote the corresponding \emph{simple reflection}. The \emph{Weyl group} $\sW \subset \mathrm{Aut}(\mathfrak{h}} \def\olfh{\overline{\mathfrak h}^*)$ of ${\mathfrak{g}}} \def\tfg{\tilde\fg$ is the group generated by the simple reflections $\{ (i) \ | \ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_i\in\mathscr{S} \}$. A composition of simple reflections $(i_1) \circ (i_2) \circ\cdots\circ(i_t)$, which are understood to act on the left, is written $(i_1 i_2 \cdots i_t) \in \sW$. The \emph{length} of a Weyl group element $w$ is the minimal number
$$
|w| \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \tmin\{ \ell \ | \ w = (i_1 i_2 \cdots i_\ell) \}
$$
of simple reflections necessary to represent $w$.
Let $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} \subset \sW$ be the subgroup generated by the simple reflections $\{(i) \ | \ i \not \in I\}$. Then $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ is naturally identified with the Weyl group of ${\mathfrak{g}}} \def\tfg{\tilde\fg^0_\mathrm{ss}$. The rational homogeneous variety $G/P$ decomposes into a finite number of $B$--orbits
$$
G/P \ = \ \bigcup_{\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w \in \sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}\backslash \sW} B w^{-1} o
$$
which are indexed by the right cosets $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}\backslash\sW$. The \emph{$B$--Schubert varieties} of $G/P$ are the Zariski closures
$$
X_w \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \overline{B w^{-1} o }\,.
$$
\begin{remark}\label{R:schubStab}
Observe that the stabilizer $P_w$ of $X_w$ in $G$ contains $B$, and is therefore a parabolic subgroup of $G$.
\end{remark}
\noindent
More generally, any $G$--translate $g X_w$ is a Schubert variety (of type $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w$). Define a partial order on $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}\backslash \sW$ by defining $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w \le \sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} v$ if $X_w \subset X_v$; then $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}\backslash\sW$ is the \emph{Hasse poset}.
Each right coset $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}\backslash \sW$ admits a unique representative of minimal length; let
$$
\sW^\mathfrak{p}} \def\olfp{\overline{\mathfrak p}_\tmin \ \simeq \ \sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} \backslash W
$$
be the set of minimal length representatives. Given $w \in \sW^\mathfrak{p}} \def\olfp{\overline{\mathfrak p}_\tmin$, the Schubert variety $w X_w$ is the Zariski closure of $N_w \cdot o$, where $N_w \subset G$ is a unipotent subgroup with nilpotent Lie algebra
\begin{equation} \label{E:nw}
\mathfrak{n}} \def\sfn{\mathsf{n}_w \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \bigoplus_{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in \Delta(w)} {\mathfrak{g}}} \def\tfg{\tilde\fg^{-\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}} \ \subset \ {\mathfrak{g}}} \def\tfg{\tilde\fg^{-}
\end{equation}
given by
\begin{equation} \label{E:Dw}
\Delta(w) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \Delta^+ \,\cap \, w(\Delta^-) \,.
\end{equation}
Moreover, $N_w \cdot o$ is an affine cell isomorphic to $\mathfrak{n}} \def\sfn{\mathsf{n}_w$, and $\tdim\,X_w = \tdim\,\mathfrak{n}} \def\sfn{\mathsf{n}_w = |\Delta(w)|$. Indeed
$$
T_o (w X_w) \ = \ \mathfrak{n}} \def\sfn{\mathsf{n}_w \,.
$$
For any $w \in \sW^\mathfrak{p}} \def\olfp{\overline{\mathfrak p}_\tmin$ we have
\begin{equation}\label{E:len=dim}
|w| \ = \ |\Delta(w)| \ = \ \tdim\,X_w \,.
\end{equation}
(The first equality holds for all $w \in \sW$, cf.~\cite[Proposition 3.2.14(3)]{MR2532439}.)
\begin{example}[Schubert varieties in $\tOG(2,\mathbb C} \def\cC{\mathcal C^m)$] \label{eg:schub1}
Let $\nu$ be a nondegenerate, symmetric bilinear form on $\mathbb C} \def\cC{\mathcal C^m$, and let
\[
\tOG(2,\mathbb C} \def\cC{\mathcal C^m) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ E \in \mathrm{Gr}(2,\mathbb C} \def\cC{\mathcal C^m) \ | \ \left.\nu\right|_{E} = 0\}
\]
be the orthogonal grassmannian of $\nu$--isotropic $2$--planes in $\mathbb C} \def\cC{\mathcal C^m$. Fix a basis $\{ e_1 , \ldots , e_m\}$ of $\mathbb C} \def\cC{\mathcal C^m$ with the property that
\[
\nu( e_a , e_b ) \ = \delta} \def\od{{\overline \d}} \def\ud{\underline \d^{m+1}_{a+b} \,.
\]
Then
\[
\mathscr{F}^p \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \tspan_\mathbb C} \def\cC{\mathcal C\{ e_1 , \ldots , e_p \} \ \subset \ \mathbb C} \def\cC{\mathcal C^m
\]
defines a flag $\mathscr{F}^{\hbox{\tiny{$\bullet$}}}$ with the property that
\begin{equation} \label{E:nu-iso-F}
\nu(\mathscr{F}^p , \mathscr{F}^{m-p}) \ = \ 0 \,.
\end{equation}
Any flag $\mathscr{F}^{\hbox{\tiny{$\bullet$}}}$ satisfying \eqref{E:nu-iso-F} is called \emph{$\nu$--isotropic}.
Given $1 \le a < b \le m$ with $a+b \not=m+1$ and $a,b \not=r+1$, define
\begin{subequations} \label{SE:schubOG}
\begin{equation} \label{E:sOG1}
X_{a,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}}) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ E \in \tOG(2,\mathbb C} \def\cC{\mathcal C^m) \ | \
\tdim\,E\cap\mathscr{F}^a \ge 1 \,,\ E \subset \mathscr{F}^b \} \,\,.
\end{equation}
If $m = 2r+1$, then every Schubert variety of $\tOG(2,\mathbb C} \def\cC{\mathcal C^{2r+1})$ is $G$--congruent to one of \eqref{E:sOG1}. If $m = 2r$ is even, define
\[
\tilde \mathscr{F}^r \ = \ \tspan_\mathbb C} \def\cC{\mathcal C\{e_1,\ldots,e_{r-1}\,,\, e_{r+1} \} \,.
\]
Note that $\mathscr{F}^1 \subset \cdots\subset\mathscr{F}^{r-1} \subset \tilde\mathscr{F}^r \subset \mathscr{F}^{r+1}\subset \cdots$ is also a $\nu$--isotropic flag. Set
\begin{eqnarray}
\label{E:sOG2}
\tilde X_{a,r}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}}) & \stackrel{\hbox{\tiny{dfn}}}{=} & \{ E \in \tOG(2,\mathbb C} \def\cC{\mathcal C^m) \ | \
\tdim\,E\cap\mathscr{F}^a \ge 1 \,,\ E \subset \tilde\mathscr{F}^r \}
\,, \quad a \le r-1,\\
\label{E:sOG3}
\tilde X_{r,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}}) & \stackrel{\hbox{\tiny{dfn}}}{=} & \{ E \in \tOG(2,\mathbb C} \def\cC{\mathcal C^m) \ | \
\tdim\,E\cap\tilde\mathscr{F}^r \ge 1 \,,\ E \subset \mathscr{F}^b \}
\,,\quad b \ge r+2 \,.
\end{eqnarray}
\end{subequations}
Every Schubert variety of $\tOG(2,\mathbb C} \def\cC{\mathcal C^{2r})$ is $G$--congruent to one of \eqref{SE:schubOG}.
\end{example}
\subsection{The horizontal subbundle}
This section is a very brief review of the horizontal sub-bundle $T^1 \subset T(G/P)$, which may be characterized as the unique $G$--homogeneous, bracket--generating subbundle of the holomorphic tangent bundle. The reader interested in greater detail is encouraged to consult \cite{MR3217458} and the references therein.
We will need an representation theoretic description of the horizontal sub-bundle, and for this it will be convenient to recall the notion of a ``grading element.''
\subsubsection{Grading elements} \label{S:grelem}
Let $\{ \ttS^1 , \ldots , \ttS^r\}$ be the basis of $\mathfrak{h}} \def\olfh{\overline{\mathfrak h}$ dual to the simple roots. A \emph{grading element} is any member of $\tspan_\bZ\{ \ttS^1,\ldots,\ttS^r\}$; these are precisely the elements $\ttT\in \mathfrak{h}} \def\olfh{\overline{\mathfrak h}$ of the Cartan subalgebra with the property that $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttT) \in \bZ$ for all roots $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in \Delta$. Since $\ttT$ is semisimple (as an element of $\mathfrak{h}} \def\olfh{\overline{\mathfrak h}$), the Lie algebra ${\mathfrak{g}}} \def\tfg{\tilde\fg$ admits a direct sum decomposition
\begin{subequations} \label{SE:grading}
\begin{equation}
{\mathfrak{g}}} \def\tfg{\tilde\fg \ = \ \bigoplus_{\ell\in\bZ} {\mathfrak{g}}} \def\tfg{\tilde\fg^\ell
\end{equation}
into $\ttT$--eigenspaces
\begin{equation} \label{E:grT}
{\mathfrak{g}}} \def\tfg{\tilde\fg^\ell \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ \xi \in {\mathfrak{g}}} \def\tfg{\tilde\fg \ | \ [\ttT,\xi] = \ell \xi \} \,.
\end{equation}
In terms of root spaces, we have
\begin{equation} \label{E:gr1} \renewcommand{\arraystretch}{1.3}
\begin{array}{rcl}
\displaystyle {\mathfrak{g}}} \def\tfg{\tilde\fg^\ell & = &
\displaystyle \bigoplus_{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttT)=\ell} {\mathfrak{g}}} \def\tfg{\tilde\fg^\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \,,\quad \ell \not=0 \,,\\
\displaystyle {\mathfrak{g}}} \def\tfg{\tilde\fg^0 & = &
\displaystyle \mathfrak{h}} \def\olfh{\overline{\mathfrak h} \ \oplus \ \bigoplus_{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttT)=0} {\mathfrak{g}}} \def\tfg{\tilde\fg^\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \,.
\end{array}
\end{equation}
\end{subequations}
The $\ttT$--eigenspace decomposition is a \emph{graded Lie algebra decomposition} in the sense that
\begin{equation}\label{E:gr}
\left[ {\mathfrak{g}}} \def\tfg{\tilde\fg^\ell , {\mathfrak{g}}} \def\tfg{\tilde\fg^m \right] \ \subset \ {\mathfrak{g}}} \def\tfg{\tilde\fg^{\ell+m} \,,
\end{equation}
a straightforward consequence of the Jacobi identity. It follows that ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$ is a Lie subalgebra of ${\mathfrak{g}}} \def\tfg{\tilde\fg$ (in fact, reductive), and each ${\mathfrak{g}}} \def\tfg{\tilde\fg^\ell$ is a ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$--module. The Lie algebra
\begin{equation} \label{E:p}
\mathfrak{p}} \def\olfp{\overline{\mathfrak p} \,=\, \mathfrak{p}} \def\olfp{\overline{\mathfrak p}_\ttT \ = \ {\mathfrak{g}}} \def\tfg{\tilde\fg^0 \,\oplus \,{\mathfrak{g}}} \def\tfg{\tilde\fg^+
\end{equation}
is the \emph{parabolic subalgebra determined by the grading element $\ttT$}. See \cite[Section 2.2]{MR3217458} for details.
\subsubsection{Minimal grading elements} \label{S:TvE}
Two distinct grading elements may determine the same parabolic $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$. As a trivial example of this, given a grading element $\ttT$ and $0 < d \in \bZ$, both $\ttT$ and $d \ttT$ determine the same parabolic. Among those grading elements $\ttT$ determining the same parabolic \eqref{E:p}, only one will have the property that ${\mathfrak{g}}} \def\tfg{\tilde\fg^{\pm1}$ generates ${\mathfrak{g}}} \def\tfg{\tilde\fg^\pm$ as an algebra. That canonical grading element is defined as follows. Given a parabolic $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ subalgebra (and choice of Cartan and Borel $\mathfrak{h}} \def\olfh{\overline{\mathfrak h} \subset \mathfrak{b}} \def\tfb{{\tilde\fb} \subset \mathfrak{p}} \def\olfp{\overline{\mathfrak p}$), the \emph{grading element associated to $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$} is
\begin{equation} \label{E:ttE}
\ttE \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \sum_{i \in I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p})} \ttS^i \,.
\end{equation}
The reductive subalgebra ${\mathfrak{g}}} \def\tfg{\tilde\fg^0 = {\mathfrak{g}}} \def\tfg{\tilde\fg^0_\mathrm{ss} \oplus \mathfrak{z}} \def\olfz{\overline \fz$ has center $\mathfrak{z}} \def\olfz{\overline \fz = \tspan_\mathbb C} \def\cC{\mathcal C\{ \ttS^i \ | \ i \in I \}$ and semisimple subalgebra ${\mathfrak{g}}} \def\tfg{\tilde\fg^0_\mathrm{ss} = [{\mathfrak{g}}} \def\tfg{\tilde\fg^0,{\mathfrak{g}}} \def\tfg{\tilde\fg^0]$. A set of simple roots for ${\mathfrak{g}}} \def\tfg{\tilde\fg^0_\mathrm{ss}$ is given by $\mathscr{S}({\mathfrak{g}}} \def\tfg{\tilde\fg_0) = \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_j \ | \ j \not \in I\}$.
\subsubsection{Definition}
The (holomorphic) tangent bundle of $G/P$ is the $G$--homogeneous vector bundle
\[
T (G/P)\ = \ G \times_P ({\mathfrak{g}}} \def\tfg{\tilde\fg/\mathfrak{p}} \def\olfp{\overline{\mathfrak p}) \,.
\]
Let $\ttE$ be the minimal grading element associated with $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ (\S\ref{S:TvE}), and consider the associated grading \eqref{SE:grading}. By \eqref{E:gr} and \eqref{E:p}, the quotient ${\mathfrak{g}}} \def\tfg{\tilde\fg^{\ge-1}/\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ is a $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$--module. Therefore,
\begin{equation} \label{E:ipr}
T^1 \ \stackrel{\hbox{\tiny{dfn}}}{=} \ G \times_P ({\mathfrak{g}}} \def\tfg{\tilde\fg^{\ge-1}/\mathfrak{p}} \def\olfp{\overline{\mathfrak p})
\end{equation}
defines a homogeneous, holomorphic subbundle of $T(G/P)$. This subbundle is the \emph{horizontal subbundle}. We have $T^1 = T(G/P)$ if and only if $G/P$ is cominuscule.
A \emph{horizontal submanifold} is an integral manifold of $T^1$; that is, a horizontal submanifold is any connected complex submanifold $M \subset G/P$ with the property that $T_zM \subset T^1_z$ for all $z \in M$, or any irreducible subvariety $Y \subset {G/P}$ with the property that $T_yY \subset T^1_y$ for all smooth points $y \in Y$.
\subsection{Horizontal Schubert varieties} \label{S:HSV}
The condition that the Schubert variety $X_w$ be horizontal is equivalent to $\Delta(w) \subset \Delta({\mathfrak{g}}} \def\tfg{\tilde\fg_1)$, where $\Delta(w)$ is given by \eqref{E:Dw}, cf.~\cite[Theorem 3.8]{MR3217458}. A convenient way to test for this condition is as follows. Let
$$
\varrho \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \sum_{i=1}^r \omega} \def\olw{\overline \omega_i \ = \ \tfrac{1}{2} \sum_{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in\Delta^+} \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}
$$
be the sum of the fundamental weights (which is also half the sum of the positive roots). Define
\begin{equation} \label{E:dfn_rho}
\varrho_w \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \varrho \,-\, w(\varrho) \ = \ \sum_{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in\Delta(w)} \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \,.
\end{equation}
(See \cite[(5.10.1)]{MR0114875} for the second equality.) Then
$$
|w| \ \le \ \varrho_w(\ttE) \ \in \ \bZ \,,
$$
and equality holds if and only if $\Delta(w) \subset \Delta({\mathfrak{g}}} \def\tfg{\tilde\fg_1)$; equivalently, $X_w$ is horizontal if and only if $\varrho_w(\ttE) = |w|$. See \cite[Section 3]{MR3217458} for details. Let
$$
\sW_\mathrm{h} \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ w \in \sW^\mathfrak{p}} \def\olfp{\overline{\mathfrak p}_\tmin \ | \ \varrho_w(\ttE) = |w| \}
$$
be the set indexing the Schubert variations of Hodge structure.
Lemma \ref{L:sOG} and Corollary \ref{C:sOG} are continuations of Example \ref{eg:schub1}.
\begin{lemma}[Horizontal Schubert varieties in $\tOG(2,\mathbb C} \def\cC{\mathcal C^m)$] \label{L:sOG}
Recall the Schubert varieties \eqref{SE:schubOG} of $\tOG(2,\mathbb C} \def\cC{\mathcal C^m)$.
\begin{a_list_emph}
\item The Schubert variety $X_{a,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ is horizontal if and only if $\mathscr{F}^a \subset (\mathscr{F}^b)^\perp$.
\item The Schubert varieties $\tilde X_{a,r}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$, are all horizontal.
\item None of the Schubert varieties $\tilde X_{r,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ are horizontal.
\end{a_list_emph}
\end{lemma}
\begin{corollary}[Maximal horizontal Schubert varieties in $\tOG(2,\mathbb C} \def\cC{\mathcal C^m)$] \label{C:sOG}
The maximal (with respect to containment) HSV in $\tSO(2,\mathbb C} \def\cC{\mathcal C^m)$, where $m \in \{2r,2r+1\}$, are: the $X_{a,m-a}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$, with $1 \le a \le r-1$; and
$\tilde X_{r-1,r}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$, if $m = 2r$. Each of these maximal HSV is of dimension $m-4$.
\end{corollary}
\begin{proof}[Proof of Lemma \ref{L:sOG}]
By definition, a Schubert variety $X_w \subset \tOG(2,\mathbb C} \def\cC{\mathcal C^m)$ is horizontal (equivalently, satisfies Griffiths's transversality condition) if and only if
\begin{equation} \label{E:MR3217458}
\left. \mathrm{d}} \def\sfd{\mathsf{d} E \right|_{N_w \cdot o} \ \subset \ E^\perp \ \stackrel{\hbox{\tiny{dfn}}}{=} \
\{ v \in \mathbb C} \def\cC{\mathcal C^m \ | \ \nu(E,v) = 0 \}
\end{equation}
on the Schubert cell $N_w \cdot o$. \smallskip
(a) It is straight--forward to see that the condition $\mathscr{F}^a \subset (\mathscr{F}^b)^\perp$ implies $X_{a,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ is horizontal. Suppose that $\mathscr{F}^a \not\subset (\mathscr{F}^b)^\perp$; equivalently, $a+b > m$. Then the condition $a+b \not=m+1$ (see Example \ref{eg:schub1}) forces $a+b \ge m+2$. It follows that
\[
E(t) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \tspan_\mathbb C} \def\cC{\mathcal C\{ e_b + t e_{m+1-a} \,,\,
e_a - t e_{m+1-b} \} \ \in \ X_{a,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}}) \,.
\]
(In fact, $E(t)$ lies in the Schubert cell.) As \eqref{E:MR3217458} clearly fails for $E(t)$, we see that $X_{a,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ is not horizontal. This establishes assertion (a) of the lemma. \smallskip
(b) As noted in Example \ref{eg:schub1}, the Schubert varieties $\tilde X_{a,r}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ are defined for $a < r$. Since $\mathscr{F}^a \subset \tilde\mathscr{F}^r = (\tilde\mathscr{F}^r)^\perp$, it is immediate that the $\tilde X_{a,r}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ are horizontal. \smallskip
(c) Recall that the Schubert varieties $\tilde X_{r,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ are defined for $r+2 \le b$, cf.~ Example \ref{eg:schub1}. It follows that
\[
E(t) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \tspan_\mathbb C} \def\cC{\mathcal C\{ e_b + t e_r \,,\,
e_{r+1} - t e_{m+1-b} \} \ \in \ \tilde X_{r,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}}) \,.
\]
(As above, $E(t)$ lies in the Schubert cell.) As \eqref{E:MR3217458} clearly fails for $E(t)$, we see that $\tilde X_{r,b}(\mathscr{F}^{\hbox{\tiny{$\bullet$}}})$ is not horizontal. This establishes assertion (c) of the lemma.
\end{proof}
\section{Smooth horizontal Schubert varieties} \label{S:smooth}
A \emph{homogeneously embedded, homogeneous submanifold} of $G/P$ is a submanifold of the form
$$
g\,\{ aP \in G/P \ | \ a \in G' \} \ = \ g\, G'/P \,,
$$
where $g \in G$ and $G'$ is a closed Lie subgroup of $G$. Such a submanifold is isomorphic to $G'/P'$, where $P' = G' \cap P$.
Distinguished amongst the homogeneously embedded, homogeneous submanifolds of $G/P$ are the rational homogeneous subvarieties $X(\sD')$ corresponding to subdiagrams $\sD'$ of the Dynkin diagram $\sD$ of ${\mathfrak{g}}} \def\tfg{\tilde\fg$. The correspondence is as follows: Identifying the nodes of $\sD$ with the simple roots $\mathscr{S}$ of ${\mathfrak{g}}} \def\tfg{\tilde\fg$, to any subdiagram we may associate a semisimple Lie subalgebra ${\mathfrak{g}}} \def\tfg{\tilde\fg' \subset {\mathfrak{g}}} \def\tfg{\tilde\fg$ by defining ${\mathfrak{g}}} \def\tfg{\tilde\fg'$ to be the subalgebra generated by the root spaces $\{{\mathfrak{g}}} \def\tfg{\tilde\fg^{\pm\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}} \ | \ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \in \sD' \}$. Note that the Dynkin diagram of ${\mathfrak{g}}} \def\tfg{\tilde\fg'$ is naturally identified with the subdiagram $\sD'$. Let $G' \subset G$ be the corresponding closed, connected semisimple Lie subgroup. For such $G'$, the subgroup $P' = G' \cap P$ is a parabolic subgroup. The $G'$--orbit of $P \in G/P$ is isomorphic to $G'/P'$ and is the \emph{homogeneously embedded, rational homogeneous subvariety $X(\sD') \subset G/P$ corresponding to subdiagram $\sD' \subset \sD$}.
Define the index set $J = \{ j \ | \ {\mathfrak{g}}} \def\tfg{\tilde\fg^{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_j} \not\subset {\mathfrak{g}}} \def\tfg{\tilde\fg' \}$. Let $\ttF = \sum_{j \in J}\,\ttS^j$ be the corresponding grading element (Section \ref{S:grelem}), and let ${\mathfrak{g}}} \def\tfg{\tilde\fg = \oplus {\mathfrak{g}}} \def\tfg{\tilde\fg^\ell$ be the $\ttF$--eigenspace decomposition \eqref{SE:grading}. Then ${\mathfrak{g}}} \def\tfg{\tilde\fg' = [{\mathfrak{g}}} \def\tfg{\tilde\fg^0 , {\mathfrak{g}}} \def\tfg{\tilde\fg^0]$ is the semisimple component of the reductive subalgebra ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$. Let $Q = P_J$ be the corresponding parabolic subgroup of $G$ (with Lie algebra $\mathfrak{q}} \def\bq{\mathbf{q} = {\mathfrak{g}}} \def\tfg{\tilde\fg^0 \oplus {\mathfrak{g}}} \def\tfg{\tilde\fg^+$.) It follows from the discussion of \cite[\S2.7.1]{MR2090671}\footnote{There is a typo in \cite[\S2.7.1]{MR2090671}; see \S\ref{S:tits_pt} for the corrected statement.} that
\begin{equation} \label{E:TvD}
\hbox{$X(\sD') \subset G/P$ is the Tits transformation $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(Q/Q)$ of the point in
$Q/Q \in G/Q$.}
\end{equation}
This, with \cite[Lemma 2.4]{MR3130568} (alternatively, see Lemma \ref{L:tits}), yields the following
\begin{lemma} \label{L:subD}
The homogeneously embedded, rational homogeneous subvarieties $X(\sD') \subset G/P$ corresponding to subdiagrams $\sD' \subset \sD$ are smooth Schubert varieties of $G/P$.
\end{lemma}
\begin{remark} \label{R:Dw}
Let $\ttE$ be the grading element \eqref{E:ttE} associated with $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$. It follows from the discussion above, that a Schubert variety $X_w$ is of the form $X(\sD')$ if and only if $\Delta(w) = \Delta({\mathfrak{g}}} \def\tfg{\tilde\fg'_+)$, where ${\mathfrak{g}}} \def\tfg{\tilde\fg' = \oplus {\mathfrak{g}}} \def\tfg{\tilde\fg'_\ell$ is the $\ttE$--eigenspace decomposition of ${\mathfrak{g}}} \def\tfg{\tilde\fg'$ and $\Delta(w)$ is the set \eqref{E:Dw}. By Section \ref{S:HSV}, $X(\sD')$ is a horizontal if and only if ${\mathfrak{g}}} \def\tfg{\tilde\fg' = {\mathfrak{g}}} \def\tfg{\tilde\fg'_1 \oplus {\mathfrak{g}}} \def\tfg{\tilde\fg'_0 \oplus {\mathfrak{g}}} \def\tfg{\tilde\fg'_{-1}$; equivalently, $X(\sD')$ is Hermitian symmetric.
\end{remark}
The main result of the paper is
\begin{mt}
Let $G/P$ be a rational homogeneous variety, and let $X \subset G/P$ be a horizontal Schubert variety. If $X$ is smooth, then $X$ is a product of homogeneously embedded, rational homogeneous subvarieties $X(\sD') \subset G/P$ corresponding to subdiagrams $\sD' \subset \sD$. Moreover, each $X(\sD')$ is cominuscule.
\end{mt}
\noindent Since a homogeneously embedded Hermitian symmetric space is necessarily smooth, Theorem \ref{T:smooth} is equivalent to: \emph{A HSV is smooth if and only if it is a horizontal homogeneously embedded Hermitian symmetric space.}
Theorem \ref{T:smooth} is proved in Sections \ref{S:smprf}--\ref{S:prf4}.
\medskip
\begin{example}[Symplectic Grassmannians] \label{eg:SG}
In this example we will (a) illustrate the construction of $X(\sD')$ from $\sD'\subset \sD$, and (b) observe that not every smooth Schubert variety of a rational homogeneous variety is homogeneous; in particular, the assumption that $X$ be horizontal in Theorem \ref{T:smooth} is essential.
Let $\nu$ be a non-degenerate, skew-symmetric bilinear form on $\mathbb C} \def\cC{\mathcal C^{2r}$. Fix $\mathtt{i}} \def\sfi{\mathsf{i} \le r$. The symplectic grassmannian
\[
\mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r}) \ = \
\left\{ E \in \mathrm{Gr}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r}) \ \left| \ \left.\nu\right|_E = 0 \right.\right\}
\]
is a rational homogeneous variety $G/P$, where $G = \mathrm{Aut}(\mathbb C} \def\cC{\mathcal C^{2r},Q) = \mathrm{Sp}} \def\tSpin{\mathit{Spin}_{2r}\mathbb C} \def\cC{\mathcal C$ and $P = P_\mathtt{i}} \def\sfi{\mathsf{i}$ is the maximal parabolic subgroup associated to the $\mathtt{i}} \def\sfi{\mathsf{i}$--th simple root. The Dynkin diagram of $G$ (containing $r$ nodes) is
{
\setlength{\unitlength}{3pt}
\begin{picture}(25,3)(0,-1)
\multiput(0,0)(5,0){2}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\multiput(15,0)(5,0){3}{\circle*{1}}
\multiput(8,0)(2,0){3}{\circle*{0.4}}
\put(15,0){\line(1,0){5}}
\put(20,0.3){\line(1,0){5}}
\put(20,-0.3){\line(1,0){5}}
\put(22,-0.75){{\footnotesize{$\langle$}}}
\end{picture}
}.
Fix a $Q$--isotropic flag $F^1 \subset F^2 \subset \cdots \subset F^{2r} = \mathbb C} \def\cC{\mathcal C^{2r}$; here $F^d$ is a complex linear subspace of dimension $d$ and $\nu(F^d , F^{2r-d}) = 0$. Given $0 \le a < \mathtt{i}} \def\sfi{\mathsf{i} < b \le 2r-a$,
$$
X_{a,b} \ = \ \{ E \in \mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r}) \ | \ F^a \subset E \subset F^b \}
$$
is a Schubert variety. The variety $X_{a,b}$ is the homogeneous submanifold $X(\sD')$ associated to a Dynkin subdiagram $\sD' \subset \sD$ if and only if either $b \le r$, or $b = 2r-a$. In the first case ($b \le r$), the subdiagram $\sD'$ corresponds to the simple roots $\mathscr{S}' = \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_{a+1},\ldots,\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_{b-1}\}$; we have ${\mathfrak{g}}} \def\tfg{\tilde\fg' \simeq \fsl_{b-a}\mathbb C} \def\cC{\mathcal C$ and $X_{a,b} \simeq \mathrm{Gr}(\mathtt{i}} \def\sfi{\mathsf{i}-a,\mathbb C} \def\cC{\mathcal C^{b-a})$; these Schubert varieties are horizontal. In the second case ($b = 2r-a$), the subdiagram $\sD'$ corresponds to the simple roots $\mathscr{S}' = \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_{a-1},\ldots,\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_r\}$; we have ${\mathfrak{g}}} \def\tfg{\tilde\fg'\simeq \mathfrak{sp}} \def\fsu{\mathfrak{su}_{2(r-a)}\mathbb C} \def\cC{\mathcal C$ and $X_{a,b} = \mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i}-a,\mathbb C} \def\cC{\mathcal C^{2(r-a)})$; these $X_{a,b}$ are horizontal if and only if $\mathtt{i}} \def\sfi{\mathsf{i}=r$ (equivalently, the symplectic grassmannian is a Hermitian symmetric Lagrangian grassmannian).
We illustrate this for $G/P=\mathrm{SG}(4,\mathbb C} \def\cC{\mathcal C^{16})$; that is, $\mathtt{i}} \def\sfi{\mathsf{i}=4$ and $r=8$. The encircled subdiagram
\begin{center}
\setlength{\unitlength}{3pt}
\begin{picture}(35,3)(0,-1)
\multiput(0,0)(5,0){8}{\circle*{1}}
\put(15,0){\circle{1.9}}
\put(0,0){\line(1,0){30}}
\put(30,0.3){\line(1,0){5}}
\put(30,-0.3){\line(1,0){5}}
\put(32,-0.75){{\footnotesize{$\langle$}}}
\put(15,0){\oval(25,4)}
\end{picture}
\end{center}
corresponds to ${\mathfrak{g}}} \def\tfg{\tilde\fg' = \fsl_6\mathbb C} \def\cC{\mathcal C$ and
$$
X_{1,7} \ = \ \{ E \in \mathrm{SG}(4,\mathbb C} \def\cC{\mathcal C^{16}) \ | \ F^1 \subset E \subset F^7 \}
\ = \ \mathrm{Gr}(3,F^7/F^1) \ \simeq \ \mathrm{Gr}(3,\mathbb C} \def\cC{\mathcal C^6) \,.
$$
Likewise, the encircled subdiagram
\begin{center}
\setlength{\unitlength}{3pt}
\begin{picture}(35,3)(0,-1)
\multiput(0,0)(5,0){8}{\circle*{1}}
\put(15,0){\circle{1.9}}
\put(0,0){\line(1,0){30}}
\put(30,0.3){\line(1,0){5}}
\put(30,-0.3){\line(1,0){5}}
\put(32,-0.75){{\footnotesize{$\langle$}}}
\put(22.5,0){\oval(30,4)}
\end{picture}
\end{center}
corresponds to ${\mathfrak{g}}} \def\tfg{\tilde\fg' = \mathfrak{sp}} \def\fsu{\mathfrak{su}_{12}\mathbb C} \def\cC{\mathcal C$ and
\[
X_{2,14} \ = \ \{ E \in \mathrm{SG}(4,\mathbb C} \def\cC{\mathcal C^{16}) \ | \ F^2 \subset E \}
\ = \ \mathrm{SG}(2,(F^2)^\perp/F^2) \ \simeq \ \mathrm{SG}(2,\mathbb C} \def\cC{\mathcal C^{12}) \,.
\]
Returning to the general case, we have $\mathtt{i}} \def\sfi{\mathsf{i}=r$ if and only if ${G/P}$ is a Hermitian symmetric space. In this case, the $T^1 = T(G/P)$ so that every Schubert variety is a HSV. Moreover, a Schubert variety is smooth if and only if it is the homogenous submanifold associated to a Dynkin subdiagram \cite{MR1703350}.
In contrast to the Hermitian symmetric case, if $\mathtt{i}} \def\sfi{\mathsf{i} < r$ and $b = 2n-a-1$, then $X_{a,b}$ is smooth, but not homogeneous, cf.~\cite{MR2356323}. These Schubert varieties are not horizontal.
\end{example}
\begin{remark}[Maximal parabolic associated to non--short root] \label{R:HM}
If $P$ is a maximal parabolic corresponding to a non--short simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$, then all smooth Schubert varieties of ${G/P}$ are homogeneously embedded, rational homogeneous varieties $X(\sD')$ corresponding to connected subdiagrams $\sD'\subset\sD$ containing the $\mathtt{i}} \def\sfi{\mathsf{i}$--th node, cf.~\cite[Proposition 3.7]{HongMok2013}. For example, the smooth Schubert varieties in the Grassmannian $\mathrm{Gr}(k,\mathbb C} \def\cC{\mathcal C^n)$ are all homogeneous.
\end{remark}
\subsection{Proof of Theorem \ref{T:smooth}: outline} \label{S:smprf}
The proposition is proved in three steps. First we reduce to the case that $P$ is maximal (Section \ref{S:prf1}). Then, the result \cite[Proposition 3.7]{HongMok2013} of Hong and Mok establishes the proposition in the case that the associated simple root is not short (Section \ref{S:prf2}). Third, we address the short root case (Section \ref{S:prf3}).
\subsection{Reduce to the case that $P$ is maximal parabolic} \label{S:prf1}
Suppose that $X$ is a HSV. Let $w \in W^\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ be the associated Weyl group element (Section \ref{S:schub}), so that $X = X_w = \overline{B w^{-1}o}$. The condition that the Schubert variety be horizontal is equivalent to $\Delta(w) \subset \Delta({\mathfrak{g}}} \def\tfg{\tilde\fg_1)$ (Section \ref{S:HSV}). By definition $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \in\Delta({\mathfrak{g}}} \def\tfg{\tilde\fg_1)$ if and only if $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttE) = 1$. Equivalently, $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttS^i) = 1$ for exactly one $i \in I$, and $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttS^j) = 0$ for all other $j \in I$. Therefore, we have a disjoint union
$$
\Delta(w) \ = \ \bigsqcup_{i \in I} \Delta_i(w)\,,
$$
where
$$
\Delta_i(w) \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}\in\Delta(w) \ | \ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttS^i) = 1 \} \,.
$$
It is straightforward to confirm that both $\Delta_i(w)$ and $\Delta^+ \backslash \Delta_i(w)$ are closed. Therefore, $\Delta_i(w)$ determines a Schubert variety $X_i \subset G/P$, cf.~\cite[Remark 3.7(c)]{MR3217458}. Since $\Delta_i(w) \subset \Delta(w)$ it follows from \cite[Lemma 8.1 \& Corollary 8.3]{MR3217458} that $X_i \subset X$. Whence
\[
X \ = \ \prod_{i \in I} X_i \,.
\]
Let $\sD$ denote the Dynkin diagram of $G$. Given $i \in I$, let $\sD \backslash \{ I \backslash\{i\}\}$ denote the subdiagram of $\sD$ obtained by removing all nodes corresponding to $j \in I \backslash\{i\}$ (and their adjacent edges). Let $\sD_i$ denote the connected component of $\sD\backslash \{ I \backslash\{i\}\}$ containing the $i$-th node. Let $G_i \subset G$ denote the closed, connected semisimple Lie subgroup of $G$ corresponding to $\sD_i$. Let $\cO_i \subset G/P$ be the $G_i$--orbit of $o$. Then $\cO_i \simeq G_i/(G_i\cap P)$ is rational homogeneous variety containing $X_i$ as a Schubert subvariety. Moreover, $X_i$ is horizontal for the IPR on $\cO_i$. Finally, note that $G_i\cap P$ is a maximal parabolic subgroup of $G_i$ -- the corresponding index set is just $\{i\}$. Since $X$ is smooth if and only if each $X_i$ is smooth, to prove the proposition, it suffices to show that $X_i$ is a homogeneously embedded Hermitian symmetric space.
This reduces the proof of the proposition to the case that $P$ is a maximal parabolic, which we now assume. Let $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ denote the associated simple root.
\subsection{The case that $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ is not a short root} \label{S:prf2}
Suppose that $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ is not a short root of $G$. By \cite[Proposition 3.7]{HongMok2013}, the Schubert variety $X \subset G/P_\mathtt{i}} \def\sfi{\mathsf{i}$ is the homogeneously embedded, rational homogeneous subvariety corresponding to a subdiagram of $\sD$.
\subsection{The case that $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ is a short root}\label{S:prf3}
In this case $G/P_\mathtt{i}} \def\sfi{\mathsf{i}$ is one of the following five rational homogeneous varieties:
\begin{bcirclist}
\item
Let $\nu$ be a nondegenerate symmetric bilinear form on $\mathbb C} \def\cC{\mathcal C^{2r+1}$. Then the \emph{orthogonal grassmannian}
$$
\mathrm{OG}(r,\mathbb C} \def\cC{\mathcal C^{2r+1}) \ = \
\left\{ E \in \mathrm{Gr}(r,\mathbb C} \def\cC{\mathcal C^{2r+1}) \ \left| \ \left.\nu\right|_E = 0 \right.\right\}
$$
of maximal, $\nu$--isotropic subspaces is the rational homogeneous variety $B_r/P_r = \tSpin_{2r+1}\mathbb C} \def\cC{\mathcal C/P_r$, where $P_r$ is the maximal parabolic associated to the short simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_r$.
\item
Let $\nu$ be a nondegenerate skew-symmetric bilinear form on $\mathbb C} \def\cC{\mathcal C^{2r}$. Then the \emph{symplectic grassmannian}
$$
\mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r}) \ = \
\left\{ E \in \mathrm{Gr}(r,\mathbb C} \def\cC{\mathcal C^{2r+1}) \ \left| \ \left.\nu\right|_E = 0 \right.\right\}
$$
of $\mathtt{i}} \def\sfi{\mathsf{i}$--dimensional $\nu$--isotropic subspaces is the rational homogeneous variety $C_r/P_\mathtt{i}} \def\sfi{\mathsf{i} = \mathrm{Sp}} \def\tSpin{\mathit{Spin}_{2r}\mathbb C} \def\cC{\mathcal C/P_\mathtt{i}} \def\sfi{\mathsf{i}$, where $P_\mathtt{i}} \def\sfi{\mathsf{i}$ is the maximal parabolic associated to the simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$. The simple root is short if and only if $\mathtt{i}} \def\sfi{\mathsf{i} < r$.
\item The exceptional $F_4/P_3$, $F_4/P_4$ or $G_2/P_1$.
\end{bcirclist}
Unfortunately, the Hong--Mok argument does not appear to extend to $\mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r})$, $F_4/P_3$ or $G_2/P_1$. Instead, we will see that, for each of the five ${G/P}$ above, Theorem \ref{T:smooth} follows from either (i) the Brion--Polo classification of smooth minuscule Schubert varieties, or (ii) existing descriptions of the HSV, or a combination of both.
\subsubsection{The case that ${G/P} = \mathrm{OG}(r,\mathbb C} \def\cC{\mathcal C^{2r+1})$}
In this case, ${G/P}$ is minuscule, cf.~\cite{MR1782635}. Brion and Polo have shown that the smooth Schubert varieties of minuscule $G/P$ are homogeneous submanifolds. More precisely, let $Q \supset B$ denote the stabilizer of $X$, cf.~Remark \ref{R:schubStab}. Then $X = Q\cdot o$ by \cite[Proposition 3.3(a)]{MR1703350}.
Let $\mathfrak{q}} \def\bq{\mathbf{q} = \mathfrak{q}} \def\bq{\mathbf{q}^0 \oplus \mathfrak{q}} \def\bq{\mathbf{q}^+$ be the graded decomposition \eqref{E:p} associated to associated to the parabolic $\mathfrak{q}} \def\bq{\mathbf{q}$ (and choices $\mathfrak{q}} \def\bq{\mathbf{q} \supset \mathfrak{b}} \def\tfb{{\tilde\fb} \supset \mathfrak{h}} \def\olfh{\overline{\mathfrak h}$). If $Q^0 \subset Q$ is the closed Lie subgroup with Lie algebra $\mathfrak{q}} \def\bq{\mathbf{q}^0$, then $Q \simeq Q^0 \times \mathfrak{q}} \def\bq{\mathbf{q}^+ = Q^0 \times \mathrm{exp}(\mathfrak{q}} \def\bq{\mathbf{q}^+)$, by \cite[Theorem 3.1.3]{MR2532439}. Therefore, $X = Q \cdot o = Q^0 \cdot o = Q^0_\mathrm{ss}\cdot o$, where $Q^0_\mathrm{ss}$ is the closed semisimple Lie subgroup with Lie algebra $\mathfrak{q}} \def\bq{\mathbf{q}^0_\mathrm{ss} = [\mathfrak{q}} \def\bq{\mathbf{q}^0,\mathfrak{q}} \def\bq{\mathbf{q}^0]$. The semisimple $Q^0_\mathrm{ss}$ is the subgroup of $G$ associated to the subdiagram $\sD' = \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \in \sD \ | \ {\mathfrak{g}}} \def\tfg{\tilde\fg^\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \in \mathfrak{q}} \def\bq{\mathbf{q}^0 \}$ (Section \ref{S:grelem}). Thus, $X = X(\sD')$ is the homogeneously embedded, rational homogeneous subvariety associated to a subdiagram $\sD'\subset\sD$.
\subsubsection{The case that ${G/P} = \mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r})$}
Recall that $\mathtt{i}} \def\sfi{\mathsf{i} < r$. In this case, the marked Dynkin diagram associated to $\mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r})$ is
\begin{center}
\setlength{\unitlength}{3pt}
\begin{picture}(45,2)(0,-1)
\multiput(0,0)(5,0){2}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\multiput(8,0)(2,0){3}{\circle*{0.4}}
\multiput(15,0)(5,0){3}{\circle*{1}}
\put(15,0){\line(1,0){10}}
\multiput(28,0)(2,0){3}{\circle*{0.4}}
\multiput(35,0)(5,0){3}{\circle*{1}}
\put(35,0){\line(1,0){5}}
\put(40,0.3){\line(1,0){5}}
\put(40,-0.3){\line(1,0){5}}
\put(42,-0.675){{\scriptsize{$\langle$}}}
\put(-0.8,1.5){\footnotesize{1}}
\put(4.2,1.5){\footnotesize{2}}
\put(19.2,1.5){\footnotesize{$\mathtt{i}} \def\sfi{\mathsf{i}$}}
\put(44.2,1.5){\footnotesize{$r$}}
\put(20,0){\circle{1.9}}
\end{picture}
\end{center}
It follows from the proof of \cite[Proposition 3.11]{MR3217458}\footnote{See, in particular, Step 3 of Section 7.3 in \cite{MR3217458}; in the case under consideration, $t = 1$ and $i_t = \mathtt{i}} \def\sfi{\mathsf{i}$.} that there is a unique maximal Schubert variation of Hodge structure; it is the homogeneous submanifold $Z' \simeq \mathrm{Gr}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{r})$ associated to the circled subdiagram
\begin{center}
\setlength{\unitlength}{3pt}
\begin{picture}(45,2)(0,-1)
\multiput(0,0)(5,0){2}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\multiput(8,0)(2,0){3}{\circle*{0.4}}
\multiput(15,0)(5,0){3}{\circle*{1}}
\put(15,0){\line(1,0){10}}
\multiput(28,0)(2,0){3}{\circle*{0.4}}
\multiput(35,0)(5,0){3}{\circle*{1}}
\put(35,0){\line(1,0){5}}
\put(40,0.3){\line(1,0){5}}
\put(40,-0.3){\line(1,0){5}}
\put(42,-0.675){{\scriptsize{$\langle$}}}
\put(20,0){\circle{1.9}}
\put(20,0){\oval(43,3)}
\end{picture}
\end{center}
So, any smooth HSV in $\mathrm{SG}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^{2r})$ is a smooth Schubert variety of $\mathrm{Gr}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^r)$. These are well-known to be precisely the homogeneously embedded, rational homogeneous subvarieties corresponding to connected subdiagrams of the the marked Dynkin diagram for $\mathrm{Gr}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^r)$ that contain the $\mathtt{i}} \def\sfi{\mathsf{i}$--th node, cf.~\cite[Section 9.3]{MR1782635}. (Alternatively, this follows from the Brion--Polo result, since $\mathrm{Gr}(\mathtt{i}} \def\sfi{\mathsf{i},\mathbb C} \def\cC{\mathcal C^r)$ is minuscule, or Remark \ref{R:HM}.)
\subsubsection{The case that ${G/P} = F_4/P_3$}
By \cite[Example 5.16]{MR3217458}, the only Schubert variations of Hodge structure are (the trivial $o \in G/P$ and) the $\mathbb P} \def\cP{\mathcal P^1 \subset \mathbb P} \def\cP{\mathcal P^2$ corresponding to the two subdiagrams
\begin{center}
\setlength{\unitlength}{4pt}
\begin{picture}(15,1)(0,-1)
\multiput(0,0)(5,0){4}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\put(5,0.3){\line(1,0){5}}
\put(5,-0.3){\line(1,0){5}}
\put(10,0){\line(1,0){5}}
\put(6.5,-0.6){{\footnotesize{$\rangle$}}}
\put(10,0){\circle{1.9}}
\put(10,0){\oval(4.5,3)}
\end{picture}
%
\hsp{40pt}
%
\begin{picture}(15,1)(0,-1)
\multiput(0,0)(5,0){4}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\put(5,0.3){\line(1,0){5}}
\put(5,-0.3){\line(1,0){5}}
\put(10,0){\line(1,0){5}}
\put(6.5,-0.6){{\footnotesize{$\rangle$}}}
\put(10,0){\circle{1.9}}
\put(12.5,0){\oval(9,3)}
\end{picture}
\end{center}
Therefore, any smooth HSV in $F_4/P_3$ is a homogeneously embedded, rational homogeneous subvariety corresponding to a subdiagram of $\sD$.
\subsubsection{The case that $G/P = F_4/P_4$}
By \cite[Example 5.17]{MR3217458}, the only Schubert variations of Hodge structure are (the trivial $o \in {G/P}$ and) the $\mathbb P} \def\cP{\mathcal P^1 \subset \mathbb P} \def\cP{\mathcal P^2$ corresponding to the two subdiagrams
\begin{center}
\setlength{\unitlength}{4pt}
\begin{picture}(15,1)(0,-1)
\multiput(0,0)(5,0){4}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\put(5,0.3){\line(1,0){5}}
\put(5,-0.3){\line(1,0){5}}
\put(10,0){\line(1,0){5}}
\put(7,-0.6){{\footnotesize{$\rangle$}}}
\put(15,0){\circle{1.9}}
\put(15,0){\oval(4.5,3)}
\end{picture}
%
\hsp{40pt}
%
\begin{picture}(15,1)(0,-1)
\multiput(0,0)(5,0){4}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\put(5,0.3){\line(1,0){5}}
\put(5,-0.3){\line(1,0){5}}
\put(10,0){\line(1,0){5}}
\put(6.5,-0.6){{\footnotesize{$\rangle$}}}
\put(15,0){\circle{1.9}}
\put(12.5,0){\oval(9,3)}
\end{picture}
\end{center}
Therefore, any smooth HSV in $F_4/P_4$ is a homogeneously embedded, rational homogeneous subvariety corresponding to a subdiagram of $\sD$.
\subsubsection{The case ${G/P} = G_2/P_1$}
This variety is the quadric hypersurface $\cQ^5 \subset \mathbb P} \def\cP{\mathcal P^6$. By \cite[Example 5.30]{MR3217458}, the only HSV are (the trivial $o \in {G/P}$ and) the smooth $\mathbb P} \def\cP{\mathcal P^1 \subset {G/P}$ which is the homogeneous submanifold corresponding to the circled subdiagram
\begin{center}
\setlength{\unitlength}{4pt}
\begin{picture}(15,1)(0,-1)
\multiput(0,0)(5,0){2}{\circle*{1}}
\put(0,0){\line(1,0){5}}
\put(0,0.3){\line(1,0){5}}
\put(0,-0.3){\line(1,0){5}}
\put(2.5,-0.6){{\footnotesize{$\langle$}}}
\put(0,0){\circle{1.7}}
\put(0,0){\oval(4,3)}
\end{picture}
\end{center}
Therefore, any smooth HSV in $G_2/P_1$ is a homogeneously embedded, rational homogeneous subvariety corresponding to a subdiagram of $\sD$.
\subsection{Fini} \label{S:prf4}
We have shown, in Sections \ref{S:prf1}--\ref{S:prf3}, that given any rational homogeneous variety $G/P$ (here the parabolic is arbitrary -- $P$ need not be maximal) any smooth HSV $X\subset {G/P}$ is a product homogeneously embedded, rational homogeneous subvarieties $X_i = X(\sD_i)$ corresponding to a subdiagram of $\sD_i \subset \sD$. The condition that $X$ (and therefore each $X_i$) be horizontal forces $X_i$ to be Hermitian symmetric. This completes the proof of Theorem \ref{T:smooth}.
\begin{remark}
It is possible that \cite[Theorem 2.6]{MR1703350} can be used to prove Theorem \ref{T:smooth}, as it was used to establish the homogeneity of minuscule and cominuscule Schubert varieties in \cite{MR1703350}. The arguments of \cite{MR1703350} are representation theoretic in nature; in the proof of Theorem \ref{T:smooth}, we elected for the more geometric approach of \cite{HongMok2013}.
\end{remark}
\section{Lines on flag varieties} \label{S:lines}
Throughout this section we
\begin{quote}
\emph{assume $P$ is a maximal parabolic subgroup, and identify $G/P$ with its image under the minimal homogeneous embedding $G/P \hookrightarrow \mathbb P} \def\cP{\mathcal P V$}
\end{quote}
of Section \ref{S:rhv}. The main result of this section is Proposition \ref{P:Xo} which characterizes the flag manifolds $G/P$ with the property that the Schubert variety swept out by the horizontal lines through a point is itself horizontal. The proposition is applied in \cite{KR1} to study
Fix a highest weight vector $0\not=v \in V$, so that $[v] = o \in \mathbb P} \def\cP{\mathcal P V$ is the highest weight line. By \eqref{E:p}, the tangent space $T_o(G/P)$ is naturally identified with ${\mathfrak{g}}} \def\tfg{\tilde\fg/\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ as a $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$--module, and with ${\mathfrak{g}}} \def\tfg{\tilde\fg^-$ as a ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$--module; for the most part, we will work with the latter identification. The set of embedded, linear $\mathbb P} \def\cP{\mathcal P^1 \subset \mathbb P} \def\cP{\mathcal P V$ containing $o$ and tangent to $G/P$ at that point is in bijection with $\mathbb P} \def\cP{\mathcal P {\mathfrak{g}}} \def\tfg{\tilde\fg^- = \mathbb P} \def\cP{\mathcal P\,T_o(G/P)$. To be precise, given a tangent line $[\xi] \in \mathbb P} \def\cP{\mathcal P {\mathfrak{g}}} \def\tfg{\tilde\fg^-$, we have
$$
\mathbb P} \def\cP{\mathcal P^1 \ = \ \mathbb P} \def\cP{\mathcal P^1(o,[\xi]) \ \stackrel{\hbox{\tiny{dfn}}}{=}\ \mathbb P} \def\cP{\mathcal P \, \tspan_\mathbb C} \def\cC{\mathcal C\{v , \xi(v)\} \subset \mathbb P} \def\cP{\mathcal P V \,.
$$
Making use of this identification, let
$$
\tilde \cC_o \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ [\xi] \in \mathbb P} \def\cP{\mathcal P \,{\mathfrak{g}}} \def\tfg{\tilde\fg^- \ | \ \mathbb P} \def\cP{\mathcal P^1(o,[\xi]) \subset G/P \}
\ = \ \{ \mathbb P} \def\cP{\mathcal P^1 \subset \mathbb P} \def\cP{\mathcal P V \ | \ o \in \mathbb P} \def\cP{\mathcal P^1 \subset G/P \}
$$
be the set of lines on $G/P$ passing through $o$. (For a general embedding $G/P \hookrightarrow \mathbb P} \def\cP{\mathcal P V$, not necessarily minimal, $\tilde \cC_o$ is defined to be the variety of minimal rational tangents, cf.~\cite{MR1748609}.) The subvariety of lines tangent to ${\mathfrak{g}}} \def\tfg{\tilde\fg^{-1} \subset {\mathfrak{g}}} \def\tfg{\tilde\fg^- \simeq T_o(G/P)$ is
\begin{equation}\label{E:Co}
\cC_o \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ [\xi] \in \tilde \cC_o \ | \ \xi \in {\mathfrak{g}}} \def\tfg{\tilde\fg^{-1} \}
\ = \ \{ \mathbb P} \def\cP{\mathcal P^1 \subset (G/P) \ | \ \mathbb P} \def\cP{\mathcal P^1 \ni o
\hbox{ is horizontal}\}\,.
\end{equation}
Let
\begin{equation}\label{E:Xo}
X \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \bigcup_{\mathbb P} \def\cP{\mathcal P^1 \in \cC_o} \mathbb P} \def\cP{\mathcal P^1
\end{equation}
be the variety swept out by the lines that pass through $o$ and are horizontal.
\begin{proposition} \label{P:Xo}
Assume that $P$ is a maximal parabolic subgroup of $G$. Then
\begin{a_list_emph}
\item $X$ is a cone over $\cC_o$ with vertex $o$ and a Schubert variety.
\item $X$ is horizontal if and only if the simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ associated with the maximal parabolic $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ is not short.
\end{a_list_emph}
\end{proposition}
\noindent The proposition is proved in Section \ref{S:linespt}.
\subsection{Properties of $\cC_0$} \label{S:Pmax}
We now recall two properties of $\cC_o$ in the case that $P$ is maximal (equivalently, $I = \{\mathtt{i}} \def\sfi{\mathsf{i}\}$ and $\ttE = \ttS^\mathtt{i}} \def\sfi{\mathsf{i}$). The results that follow are due to \cite{MR1966752}, where $\cC_o$ and $\tilde\cC_o$ are discussed for arbitrary (not necessarily maximal) $P$.
\begin{a_list_bold}
\item
Since $P$ is maximal, ${\mathfrak{g}}} \def\tfg{\tilde\fg^{-1}$ is an irreducible ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$--module with highest weight line ${\mathfrak{g}}} \def\tfg{\tilde\fg^{-\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}}$, cf. \cite[Theorem 8.13.3]{MR928600}. The variety of lines $\cC_o \subset \mathbb P} \def\cP{\mathcal P {\mathfrak{g}}} \def\tfg{\tilde\fg^{-1}$ is the $G^0$--orbit of this highest weight line, cf.~\cite[Theorem 4.3]{MR1966752}. In particular, $\cC_o$ is a rational homogeneous variety; indeed,
$$
\cC_o \ \simeq \ G^0/(G^0\cap Q)\,,
$$
where $Q \supset B$ is the parabolic subgroup defined by
\begin{equation} \label{E:Iq}
I(\mathfrak{q}} \def\bq{\mathbf{q}) \ = \ \{ j \ | \ {\mathfrak{g}}} \def\tfg{\tilde\fg^{-\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_j} \not\subset \mathfrak{q}} \def\bq{\mathbf{q}\}
\ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ j \ | \ \langle \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i} , \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_j \rangle \not=0\} \,.
\end{equation}
That is, the simple roots indexed by $I(\mathfrak{q}} \def\bq{\mathbf{q})$ are those adjacent to $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ in the Dynkin diagram of ${\mathfrak{g}}} \def\tfg{\tilde\fg$; cf.~\cite[Proposition 2.5]{MR1966752}. With only a few exceptions, $\cC_o$ is a $G_0$--cominuscule variety; equivalently, $\cC_o \simeq G^0/(G^0\cap Q)$ admits the structure of a compact Hermitian symmetric space, cf.~\cite[Proposition 2.11]{MR1966752}.
\item
If the simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ associated with the maximal parabolic $P$ is not short, then $\cC_o = \tilde \cC_o$, cf.~\cite[Theorem 4.8.1]{MR1966752}. If the simple root is short, then $\tilde \cC_o$ is the union of two $P$--orbits, and open orbit and its boundary $\cC_o$.
\end{a_list_bold}
\subsection{Uniruling of $G/P$} \label{S:uniruling}
We continue with the assumption that $P$ is maximal; however, analogous statements follow for unirulings on general $G/P$. For the more general statements, it is convenient to use Tits correspondences, which are briefly reviewed in Section \ref{S:tits}.
Given $x = g o \in G/P$, with $g \in G$, let $\cC_x = g \cC_o$ denote the corresponding set of lines through $x$. (It is an exercise to show that $\cC_x$ is well-defined; that is, $\cC_x$ does not depend on our choice of $g$ yielding $x = go$.) Then
$$
\cC \ \stackrel{\hbox{\tiny{dfn}}}{=} \ \{ \mathbb P} \def\cP{\mathcal P^1 \ | \ \mathbb P} \def\cP{\mathcal P^1 \in \cC_x \,,\ x\in G/P\}
\ = \ \bigcup_{g \in G} g \,\cC_o \,.
$$
forms a uniruling of $G/P$. (As will be shown in Corollary \ref{C:pt}, this uniruling is parameterized by $G/Q$ -- that is, $\cC \simeq G/Q$.)
\begin{remark}\label{R:tC}
More generally, the set of \emph{all} lines on $G/P$ is $\tilde \cC = \cup_{g\in G} \, g \,\tilde\cC_o$.
\begin{a_list_emph}
\item It follows from definition \eqref{E:Co} and the homogeneity of the IPR, that
\begin{equation} \label{E:Cogen}
\cC \ = \ \{ \mathbb P} \def\cP{\mathcal P^1 \in \tilde\cC \ | \ \mathbb P} \def\cP{\mathcal P^1 \hbox{ is horizontal}\}
\end{equation}
is precisely the set of lines on $G/P$ that are integrals of the IPR.
\item As noted in Section \ref{S:Pmax}(b), if the simple root associated to the maximal parabolic $P$ is not short, then $\tilde \cC = \cC$ consists of a single $G$--orbit. If the simple root is short, then $\tilde\cC$ consists of two $G$--orbits, an open orbit and its boundary $\cC$, cf.~\cite[Theorem 4.3]{MR1966752}.
\end{a_list_emph}
\end{remark}
\subsection{Tits correspondences} \label{S:tits}
Tits correspondences describe homogeneous unirulings of a rational homogeneous variety $G/P$ by homogeneously embedded, rational homogeneous subvarieties $G'/P'$; these unirulings may be used to clarify the geometry of $G/P$. The material in Sections \ref{S:tits_schub}--\ref{S:linespt} is taken from \cite{MR3130568, CR2}. Given two standard parabolics $P$ and $Q=P_J$, the intersection $P \cap Q$ is also a standard parabolic. (Note that $I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}\cap\mathfrak{q}} \def\bq{\mathbf{q}) = I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}) \cup I(\mathfrak{q}} \def\bq{\mathbf{q})$.) There is a natural double fibration, called the \emph{Tits correspondence}, given by the diagram in Figure \ref{f:tits}; here the maps $\eta$ and $\tau$ are the natural projections.
\begin{figure}[h]
\caption{Tits correspondence}
\setlength{\unitlength}{4pt}
\begin{picture}(33,12)(0,0)
\put(0,0){$G/P$}
\put(10,6.8){\vector(-1,-1){4}}
\put(10.5,8.5){$G/(P\cap Q)$}
\put(23,6.8){\vector(1,-1){4}}
\put(28,0){$G/Q$}
\put(6,5.5){$\eta$}
\put(25.5,5){$\tau$}
\end{picture}
\label{f:tits}
\end{figure}
Given a subset $\Sigma\subset G/Q$, the \emph{Tits transform} is $\mathcal{T}(\Sigma) : = \eta(\tau^{-1}(\Sigma))$.
\subsection{Tits transform of a point} \label{S:tits_pt}
The Tits transform of a point $y \in G/Q$ will play a crucial r\^ole in our discussion of $G/P$ and its Schubert varieties. What follows is a brief review of \cite[\S2.7.1]{MR2090671}.\footnote{There is a typo in \cite[\S2.7.1]{MR2090671}; on page 87, $S$ and $S'$ should be swapped. This is corrected for in the discussion above.}
The Tits transform $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(Q/Q)$ of the point $Q/Q \in G/Q$ is the $G'$--orbit $G'(P/P) \subset G/P$, where $G'$ is the semisimple subgroup of $G$ whose Dynkin diagram $\sD'$ is obtained from the diagram $\sD$ of $G$ by removing the nodes $I(\mathfrak{q}} \def\bq{\mathbf{q})\backslash I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p})$. Therefore, $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(Q/Q) \simeq G'/P'$, where $P' = G' \cap P$ is the parabolic subgroup of $G'$ with index set \eqref{E:I} given by $I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}') = I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}) \backslash I(\mathfrak{q}} \def\bq{\mathbf{q})$.
Since any point $y \in G/Q$ is of the form $y = g Q/Q$ for some $g \in G$, and $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(g Q/Q) = g\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(Q/Q)$, it follows that $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(y) \simeq G'/P'$ and
\begin{subequations} \label{SE:tits}
\begin{equation} \label{E:titsa}
\begin{array}{c}
\hbox{\emph{$G/P$ is uniruled by subvarieties $\Sigma$ isomorphic to $G'/P'$,}} \\
\hbox{\emph{and the uniruling is parameterized by $G/Q$.}}
\end{array}
\end{equation}
Moreover, $G/(P\cap Q)$ is the incidence space for this uniruling. Precisely,
\begin{equation} \label{E:incidence}
G/(P\cap Q) \ = \ \{ (x,\Sigma) \in G/P \times G/Q \ | \ x \in \Sigma \}\,.
\end{equation}
\end{subequations}
\begin{remark}[$\mathbb P} \def\cP{\mathcal P^1$--unirulings of $G/P$ for maximal $P$] \label{R:P1}
In the case that $P \subset G$ is a maximal parabolic (Section \ref{S:rhv}), there is a unique $G$--homogeneous variety $G/Q$ parameterizing a uniruling of $G/P$ by lines $\mathbb P} \def\cP{\mathcal P^1$; it may be identified by inspection of the Dynkin diagram $\sD$ of ${\mathfrak{g}}} \def\tfg{\tilde\fg$ as follows. The maximality of $P$ is equivalent to $I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}) = \{\mathtt{i}} \def\sfi{\mathsf{i}\}$ for some $\mathtt{i}} \def\sfi{\mathsf{i}$. In order to obtain a uniruling by $\mathbb P} \def\cP{\mathcal P^1$s, we must choose $Q$ (equivalently, the index set $I(\mathfrak{q}} \def\bq{\mathbf{q})$) so that $G'/P' = \mathbb P} \def\cP{\mathcal P^1$. To that end, let $J = \{ j \not= \mathtt{i}} \def\sfi{\mathsf{i} \ | \ (\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i},\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_j) \not=0 \}$ index the nodes in the Dynkin diagram that are adjacent to the $\mathtt{i}} \def\sfi{\mathsf{i}$--th node. Then $G'/P' \simeq \mathbb P} \def\cP{\mathcal P^1$ (equivalently, $G/Q$ parameterizes a uniruling of $G/P$ by $\mathbb P} \def\cP{\mathcal P^1$s) if and only if $J \subset I(\mathfrak{q}} \def\bq{\mathbf{q})$. When $I(\mathfrak{q}} \def\bq{\mathbf{q}) = J$ we say that \emph{$G/Q$ is the smallest rational $G$--homogeneous variety parameterizing a uniruling of $G/P$ by lines $\mathbb P} \def\cP{\mathcal P^1$}.
\end{remark}
\subsection{Tits transform of a Schubert variety} \label{S:tits_schub}
The Tits transform preserves Schubert varieties; that is, if $Y \subset G/Q$ is a Schubert variety, then so is $X = \mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(Y) = G/P$. Moreover, $X$ may be determined as follows. Let $\sW^\mathfrak{q}} \def\bq{\mathbf{q}_\tmin$ be the set of (unique) minimal length representatives of the right cosets $\sW_\mathfrak{q}} \def\bq{\mathbf{q} \backslash \sW$, and let $\sW^\mathfrak{q}} \def\bq{\mathbf{q}_\mathrm{max} \subset \sW$ denote this set of (unique) maximal length representatives. If $w_0$ is the longest element of $\sW_\mathfrak{q}} \def\bq{\mathbf{q}$, then $\sW^\mathfrak{q}} \def\bq{\mathbf{q}_\mathrm{max}} \def\tmin{\mathrm{min} = \{ w_0 w \ | \ w \in \sW^\mathfrak{q}} \def\bq{\mathbf{q}_\tmin\}$. We have $\sW_\mathfrak{q}} \def\bq{\mathbf{q} \backslash \sW \simeq \sW^\mathfrak{q}} \def\bq{\mathbf{q}_\tmin \simeq \sW^\mathfrak{q}} \def\bq{\mathbf{q}_\mathrm{max}} \def\tmin{\mathrm{min}$. A proof of the following well--known lemma is given in \cite{MR3130568}.
\begin{lemma} \label{L:tits}
Let $w \in \sW^\mathfrak{q}} \def\bq{\mathbf{q}_\mathrm{max}} \def\tmin{\mathrm{min}$, and let $Y_w \subset G/Q$ denote the Schubert variety indexed by the coset $\sW_\mathfrak{q}} \def\bq{\mathbf{q} w$. Then the Tits transformation $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(Y_w)$ is the Schubert variety $X_w \subset G/P$ indexed by the coset $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w$.
\end{lemma}
\subsection{Lines through a point} \label{S:linespt}
The following, first observed in \cite{CR2}, is an immediate consequence of \eqref{SE:tits}.
\begin{lemma} \label{L:pt}
Suppose that $G/Q$ parameterizes a uniruling of $G/P$ by $\mathbb P} \def\cP{\mathcal P^1$s. Let $\Sigma = \tau(\eta^{-1}(o)) \subset G/Q$ denote the image of $o \in G/P$ under the Tits transform, and let $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(\Sigma) = \eta(\tau^{-1}(\Sigma)) \subset G/P$ denote the image of $\Sigma$ under the Tits transform back to $G/P$. Then $\Sigma$ is the set of lines in $G/P$ that are parameterized by $G/Q$ and pass through $o \in G/P$. Likewise,
$$
\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(\Sigma) \ = \ \bigcup_{o \in \mathbb P} \def\cP{\mathcal P^1 \in G/Q} \mathbb P} \def\cP{\mathcal P^1
$$
is the subset of $G/P$ swept out by these lines, and is naturally identified with a cone $\cC(\Sigma)$ over $\Sigma$ with vertex $o$.
\end{lemma}
\begin{remark} \label{R:pt}
By Lemma \ref{L:tits}, the variety $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(\Sigma)$ is a Schubert variety $X_w \subset G/P$.
\end{remark}
\begin{corollary} \label{C:pt}
Let $P$ be a maximal parabolic and let $Q$ be the parabolic of Section \ref{S:Pmax}(a). Then the uniruling $\cC$ of $G/P$ (Section \ref{S:uniruling}) is is precisely the uniruling $G/Q$ obtained through the Tits correspondence. In particular, the variety $\Sigma$ of Lemma \ref{L:pt} is the variety $\cC_o$ of \eqref{E:Co}, and the variety $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(\Sigma)$ of Lemma \ref{L:pt} is the variety $X$ of \eqref{E:Xo}.
\end{corollary}
\begin{proof}
By the maximality of $P$ we have $I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}) = \{\mathtt{i}} \def\sfi{\mathsf{i}\}$, cf.~ Section \ref{S:rhv}. Recall the Levi subgroup $G^0 \subset G$ whose Lie algebra ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$ is the zero--eigenspace of the grading element $\ttE = \ttS^\mathtt{i}} \def\sfi{\mathsf{i}$ associated with the parabolic $\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$, cf.~ Section \ref{S:rhv}. Then the simple roots of the semisimple Lie algebra ${\mathfrak{g}}} \def\tfg{\tilde\fg^0_\mathrm{ss} = [{\mathfrak{g}}} \def\tfg{\tilde\fg^0,{\mathfrak{g}}} \def\tfg{\tilde\fg^0]$ are $\mathscr{S} \backslash \{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}\}$, the simple roots of ${\mathfrak{g}}} \def\tfg{\tilde\fg$ minus the $\mathtt{i}} \def\sfi{\mathsf{i}$--th. As discussed in Section \ref{S:Pmax}(a), we have $\cC_o = G^0 /(G^0 \cap Q)$. It follows that $\cC_o = G^0_\mathrm{ss} /(G^0_\mathrm{ss} \cap Q)$, where $G^0_\mathrm{ss} \subset G^0$ is the semisimple subgroup with Lie algebra ${\mathfrak{g}}} \def\tfg{\tilde\fg^0_\mathrm{ss}$. Note the index set $I({\mathfrak{g}}} \def\tfg{\tilde\fg^0_\mathrm{ss} \cap \mathfrak{q}} \def\bq{\mathbf{q})$ associated with the parabolic $G^0_\mathrm{ss} \cap Q$ by \eqref{E:I} is $I(\mathfrak{q}} \def\bq{\mathbf{q})$.
As discussed in Section \ref{S:Pmax}(a), the set $I(\mathfrak{q}} \def\bq{\mathbf{q})$ indexes those nodes of the Dynkin diagram that are are adjacent to the $\mathtt{i}} \def\sfi{\mathsf{i}$--th node. By Remark \ref{R:P1}, $G/Q$ is the minimal rational $G$--homogeneous variety parameterizing a uniruling of $G/P$ by lines $\mathbb P} \def\cP{\mathcal P^1$s. The Tits transform $\Sigma = \mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(P/P) \subset G/Q$ is of the form $G'/P'$. From the descriptions of $G'$ and $P'$ in Section \ref{S:tits_pt} and the discussion of Remark \ref{R:P1} we see that $G' = G^0_\mathrm{ss}$ and $P' = G^0_\mathrm{ss} \cap Q$. Thus, $\cC_o = \Sigma$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{P:Xo}]
Part (a) of the lemma follows directly from Lemma \ref{L:pt} and Corollary \ref{C:pt}. The argument for part (b) is based on the Tits transform recipe given by Lemma \ref{L:tits} and the characterization of horizontal Schubert varieties \cite{KR2}. We will make use (without explicit mention) of observations made in the proof of Corollary \ref{C:pt}.
To begin, let $P \subset G$ be a maximal parabolic with index set $I(\mathfrak{p}} \def\olfp{\overline{\mathfrak p}) = \{\mathtt{i}} \def\sfi{\mathsf{i}\}$, cf.~ Section \ref{S:rhv}. Let $G/Q$ parameterize the uniruling of $G/P$ by $\mathbb P} \def\cP{\mathcal P^1$s. The right coset indexing the Schubert variety $o = P/P \in G/P22$ is $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w_0$, where $w_0$ is the longest word of the Weyl group $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ of ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$. By Lemma \ref{L:tits}, the Schubert variety $\Sigma = \mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(P/P)$ is indexed by the coset $\sW_\mathfrak{q}} \def\bq{\mathbf{q} w_0$. Let $w_1 \in \sW^\mathfrak{q}} \def\bq{\mathbf{q}_\mathrm{max}} \def\tmin{\mathrm{min}$ be the longest representative of $\sW_\mathfrak{q}} \def\bq{\mathbf{q} w_0 = \sW_\mathfrak{q}} \def\bq{\mathbf{q} w_1$. Again, Lemma \ref{L:tits} implies $\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(\Sigma)$ is the Schubert variety indexed by the right coset $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w_1$. By Corollary \ref{C:pt}, this is the Schubert variety $X$. Let $w \in \sW^\mathfrak{p}} \def\olfp{\overline{\mathfrak p}_\tmin$ be a shortest representative of $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w_1 = \sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w$. As discussed in Section \ref{S:HSV}, the Schubert variety $X$ is horizontal if and only if $\varrho_w(\ttE) = \tdim\,X$. So the substance of the proof is to compute the integer $\varrho_w(\ttE)$. Note that $\ttE = \ttS^\mathtt{i}} \def\sfi{\mathsf{i}$, cf.~\eqref{E:ttE}.
Let $v \in \sW_\tmin^\mathfrak{q}} \def\bq{\mathbf{q}$ be the shortest Weyl element indexing the Schubert variety $\Sigma$. Let $w_0' \in \sW_\mathfrak{q}} \def\bq{\mathbf{q}$ be the longest element of the Weyl subgroup, so that $w_0' v = w_1$ is the longest element indexing $\Sigma$. Then
$$
\sW_\mathfrak{q}} \def\bq{\mathbf{q} w_0 \ = \ \sW_\mathfrak{q}} \def\bq{\mathbf{q} v \ = \ \sW_\mathfrak{q}} \def\bq{\mathbf{q} w_0'v\,,
$$
and $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w_0' v$ indexes $X = \mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}(\Sigma)$. We claim that
\begin{equation} \label{E:c1}
\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} w_0' \ = \ \sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p} (\mathtt{i}} \def\sfi{\mathsf{i}) \,,
\end{equation}
where $(\mathtt{i}} \def\sfi{\mathsf{i}) \in \sW$ denotes the reflection associated with the simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$. It follows from \eqref{E:c1} that
$$
(\mathtt{i}} \def\sfi{\mathsf{i}) v \ = \ w \ \in \ \sW^\mathfrak{p}} \def\olfp{\overline{\mathfrak p}_\tmin
$$
is the shortest Weyl group element indexing $X$.
\begin{proof}[Proof of \eqref{E:c1}]
Observe that $\sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$ is generated by the simple reflections $\{ (j) \ | \ j \not=\mathtt{i}} \def\sfi{\mathsf{i}\}$. Likewise, $\sW_\mathfrak{q}} \def\bq{\mathbf{q}$ is generated by the simple reflections $\{ (j) \ | \ j \not= \mathtt{i}} \def\sfi{\mathsf{i}-1,\mathtt{i}} \def\sfi{\mathsf{i}+1\}$. In particular, any element $u$ of $\sW_\mathfrak{q}} \def\bq{\mathbf{q}$ may be written as $u' (\mathtt{i}} \def\sfi{\mathsf{i})$ with $u' \in \sW_\mathfrak{p}} \def\olfp{\overline{\mathfrak p}$. The claim follows.
\end{proof}
We return to the proof of Proposition \ref{P:Xo}(b). From the discussion of Section \ref{S:schub} (specifically \eqref{E:len=dim}) we see that
$$
|v| \ = \ |\Delta(v)| \ = \ \tdim\,\Sigma \quad\hbox{and}\quad
1 \,+\, |v| \ = \ |w| \ = \ |\Delta( w )| \ = \ \tdim\,X \,.
$$
The argument establishing \cite[Proposition 3.2.14(5)]{MR2532439} yields
$$
\Delta(w) \ = \ \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i} \} \ \cup \ (\mathtt{i}} \def\sfi{\mathsf{i}) \Delta(v) \,.
$$
Since $v$ indexes a \emph{homogeneous} Schubert variety $\Sigma = G'/P'$, we see from the discussion of Section \ref{S:schub} that
$$
\Delta(v) \ = \ \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \in \Delta({\mathfrak{g}}} \def\tfg{\tilde\fg^0) \ | \ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttF) > 0 \}\,,
$$
where
$$
\ttF \ = \ \sum_{i \in I(\mathfrak{q}} \def\bq{\mathbf{q})} \ttS^i
$$
is the grading element \eqref{E:ttE} associated with the parabolic $\mathfrak{q}} \def\bq{\mathbf{q}$. Since ${\mathfrak{g}}} \def\tfg{\tilde\fg^0$ is the zero--eigenspace of $\ttE = \ttS^\mathtt{i}} \def\sfi{\mathsf{i}$, this may be rewritten as
$$
\Delta(v) \ = \ \{ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \in \Delta \ | \ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttE) = 0 \,,\ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}(\ttF) > 0 \} \,.
$$
In particular, the roots of $\Delta(v)$ are precisely those positive roots $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} = m^i\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_i$ such that $m^\mathtt{i}} \def\sfi{\mathsf{i}=0$ and $m^j > 0$ for some $j \in I(\mathfrak{q}} \def\bq{\mathbf{q})$. Informally, the root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}$ does not involve the simple root $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$, but does involve some simple root adjacent to $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$. Whence the reflection
$$
(\mathtt{i}} \def\sfi{\mathsf{i})\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \ = \ \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \ - \ 2\frac{(\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a},\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i})}{(\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i},\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i})} \,\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}
$$
has the property that the integer $-2(\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a},\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i})/(\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i},\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}) \ge 1$ for every $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a} \in \Delta(v)$. It now follows from the second equality of \eqref{E:dfn_rho} that $\varrho_w(\ttE) > |w|$ if and only if $\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}_\mathtt{i}} \def\sfi{\mathsf{i}$ is short.
\end{proof}
\subsection*{\arabic{letcnt}: #1}
\stepcounter{letcnt}
}
\def\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}{\alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \alpha} \def\bara{{\bar\alpha}} \def\oa{{\overline \a}}}
\def\beta} \def\barb{{\bar\beta}} \def\ob{{\overline \b}{\beta} \def\barb{{\bar\beta}} \def\ob{{\overline \beta} \def\barb{{\bar\beta}} \def\ob{{\overline \b}}}
\def\delta} \def\od{{\overline \d}} \def\ud{\underline \d{\delta} \def\od{{\overline \delta} \def\od{{\overline \d}} \def\ud{\underline \d}} \def\ud{\underline \delta} \def\od{{\overline \d}} \def\ud{\underline \d}
\def\varepsilon} \def\bare{{\bar\varepsilon}} \def\oe{{\overline \e}{\varepsilon} \def\bare{{\bar\varepsilon}} \def\oe{{\overline \varepsilon} \def\bare{{\bar\varepsilon}} \def\oe{{\overline \e}}}
\def\rho{\rho}
\def\zeta{\zeta}
\def\mu{\mu}
\def\nu{\nu}
\def\sigma{\sigma}
\def\tau{\tau}
\def\Omega} \def\barO{{\bar\Omega}{\Omega} \def\barO{{\bar\Omega}}
\def\omega} \def\olw{\overline \omega{\omega} \def\olw{\overline \omega}
\def\theta{\theta}
\def\xi{\xi}
\def\mathcal{A}} \def\sfA{\mathsf{A}{\mathcal{A}} \def\sfA{\mathsf{A}}
\def\mathscr{A}} \def\bA{\mathbb{A}{\mathscr{A}} \def\bA{\mathbb{A}}
\def\mathfrak{a}} \def\sfa{\mathsf{a}{\mathfrak{a}} \def\sfa{\mathsf{a}}
\def\mathtt{a}} \def\ba{\mathbf{a}{\mathtt{a}} \def\ba{\mathbf{a}}
\def\mathrm{Ad}} \def\tad{\mathrm{ad}{\mathrm{Ad}} \def\tad{\mathrm{ad}}
\def\mathrm{Alt}} \def\tAnn{\mathrm{Ann}{\mathrm{Alt}} \def\tAnn{\mathrm{Ann}}
\def\mathrm{Aut}{\mathrm{Aut}}
\def\mathcal B}\def\sfB{\mathsf{B}{\mathcal B}\def\sfB{\mathsf{B}}
\def\mathbb{B}} \def\sB{\mathscr{B}{\mathbb{B}} \def\sB{\mathscr{B}}
\def\mathbf{b}} \def\sfb{\mathsf{b}{\mathbf{b}} \def\sfb{\mathsf{b}}
\def\mathfrak{b}} \def\tfb{{\tilde\fb}{\mathfrak{b}} \def\tfb{{\tilde\mathfrak{b}} \def\tfb{{\tilde\fb}}}
\def\mathtt{b}} \def\tbd{\mathrm{bd}{\mathtt{b}} \def\tbd{\mathrm{bd}}
\def\mathbb C} \def\cC{\mathcal C{\mathbb C} \def\cC{\mathcal C}
\def\mathscr{C}} \def\sfC{\mathsf{C}{\mathscr{C}} \def\sfC{\mathsf{C}}
\def\fc{\mathfrak{c}} \def\sfc{\mathsf{c}}
\def\mathbf{c}{\mathbf{c}}
\def\mathrm{Cl}} \def\tcl{\mathrm{cl}{\mathrm{Cl}} \def\tcl{\mathrm{cl}}
\def\mathrm{codim}} \def\tcpt{\mathrm{cpt}{\mathrm{codim}} \def\tcpt{\mathrm{cpt}}
\def\mathcal D} \def\sD{\mathscr{D}{\mathcal D} \def\sD{\mathscr{D}}
\def\mathfrak D{\mathfrak D}
\def\mathsf{D}} \def\ttD{\mathtt{D}{\mathsf{D}} \def\ttD{\mathtt{D}}
\def\mathfrak{d}} \def\bd{\mathbf{d}{\mathfrak{d}} \def\bd{\mathbf{d}}
\def\mathrm{d}} \def\sfd{\mathsf{d}{\mathrm{d}} \def\sfd{\mathsf{d}}
\def\mathrm{deg}} \def\tdet{\mathrm{det}{\mathrm{deg}} \def\tdet{\mathrm{det}}
\def\mathrm{Diff}{\mathrm{Diff}}
\def\mathrm{diag}} \def\tdim{\mathrm{dim}{\mathrm{diag}} \def\tdim{\mathrm{dim}}
\def\mathcal E} \def\sE{\mathscr{E}{\mathcal E} \def\sE{\mathscr{E}}
\def\mathsf{E}} \def\ttE{\mathtt{E}{\mathsf{E}} \def\ttE{\mathtt{E}}
\def\mathrm{End}} \def\tEuc{\mathrm{Euc}{\mathrm{End}} \def\tEuc{\mathrm{Euc}}
\def\mathrm{Ext}{\mathrm{Ext}}
\def\be{\mathbf{e}} \def\fe{\mathfrak{e}}
\def\mathsf{e}} \def\teven{\mathrm{even}{\mathsf{e}} \def\teven{\mathrm{even}}
\def\mathrm{exp}{\mathrm{exp}}
\def\mathcal F} \def\bbF{\mathbb{F}{\mathcal F} \def\bbF{\mathbb{F}}
\def\mathscr{F}{\mathscr{F}}
\def\ff{\mathfrak{f}} \def\olff{\overline \ff}
\def\mathsf{f}{\mathsf{f}}
\def\mathbf{F}} \def\sfF{\mathsf{F}} \def\ttF{\mathtt{F}{\mathbf{F}} \def\sfF{\mathsf{F}} \def\ttF{\mathtt{F}}
\def\mathrm{Flag}{\mathrm{Flag}}
\def\mathrm{filt}{\mathrm{filt}}
\def\mathbb{G}} \def\olG{{\overline G}{\mathbb{G}} \def\olG{{\overline G}}
\def\mathscr{G}{\mathscr{G}}
\def\mathcal G} \def\olcG{\overline{\mathcal G}{\mathcal G} \def\olcG{\overline{\mathcal G}}
\def\mathrm{GL}{\mathrm{GL}}
\def\mathrm{Gr}{\mathrm{Gr}}
\def{\mathfrak{g}}} \def\tfg{\tilde\fg{{\mathfrak{g}}} \def\tfg{\tilde{\mathfrak{g}}} \def\tfg{\tilde\fg}
\def\overline{\mathfrak g}{\overline{\mathfrak g}}
\def\mathfrak{gl}{\mathfrak{gl}}
\def\mathbb{H}{\mathbb{H}}
\def\mathcal H} \def\ttH{\mathtt{H}{\mathcal H} \def\ttH{\mathtt{H}}
\def\mathsf{H}} \def\sH{\mathscr{H}{\mathsf{H}} \def\sH{\mathscr{H}}
\def\mathrm{Hg}} \def\tHdg{\mathrm{Hdg}{\mathrm{Hg}} \def\tHdg{\mathrm{Hdg}}
\def\mathrm{Hom}{\mathrm{Hom}}
\def\mathfrak{h}} \def\olfh{\overline{\mathfrak h}{\mathfrak{h}} \def\olfh{\overline{\mathfrak h}}
\def\mathbf{h}{\mathbf{h}}
\def\mathbf{i}} \def\ti{\mathrm{i}{\mathbf{i}} \def\ti{\mathrm{i}}
\def\mathtt{i}} \def\sfi{\mathsf{i}{\mathtt{i}} \def\sfi{\mathsf{i}}
\def\mathfrak{i}{\mathfrak{i}}
\def\mathbf{I}} \def\cI{\mathcal I} \def\sI{\mathscr{I}{\mathbf{I}} \def\cI{\mathcal I} \def\sI{\mathscr{I}}
\def\mathsf{I}} \def\ttI{\mathtt{I}{\mathsf{I}} \def\ttI{\mathtt{I}}
\def\mathbf{II}{\mathbf{II}}
\def\mathrm{Id}} \def\tIm{\mathrm{Im}{\mathrm{Id}} \def\tIm{\mathrm{Im}}
\def\mathrm{id}} \def\tim{\mathrm{im}{\mathrm{id}} \def\tim{\mathrm{im}}
\def\mathrm{Int}} \def\tint{\mathrm{int}{\mathrm{Int}} \def\tint{\mathrm{int}}
\def\mathsf{j}} \def\ttj{\mathtt{j}} \def\fj{\mathfrak{j}{\mathsf{j}} \def\ttj{\mathtt{j}} \def\fj{\mathfrak{j}}
\def\mathcal J} \def\sJ{\mathscr{J}} \def\sfJ{\mathsf{J}{\mathcal J} \def\sJ{\mathscr{J}} \def\sfJ{\mathsf{J}}
\def\mathtt{J}{\mathtt{J}}
\def\mathcal K} \def\sfK{\mathsf{K}{\mathcal K} \def\sfK{\mathsf{K}}
\def\mathscr{K}{\mathscr{K}}
\def\mathfrak{k}} \def\bk{\mathbf{k}{\mathfrak{k}} \def\bk{\mathbf{k}}
\def\mathbbm{k}{\mathbbm{k}}
\def\mathsf{k}} \def\ttk{\mathtt{k}{\mathsf{k}} \def\ttk{\mathtt{k}}
\def\mathrm{Ker}} \def\tker{\mathrm{ker}{\mathrm{Ker}} \def\tker{\mathrm{ker}}
\def\mathcal L} \def\sL{\mathscr{L}{\mathcal L} \def\sL{\mathscr{L}}
\def\mathsf{L}} \def\ttL{\mathtt{L}{\mathsf{L}} \def\ttL{\mathtt{L}}
\def\mathfrak{l}} \def\ttl{\mathtt{l}{\mathfrak{l}} \def\ttl{\mathtt{l}}
\def\mathrm{lim}{\mathrm{lim}}
\def\mathrm{LG}{\mathrm{LG}}
\def\mathcal M} \def\tM{\mathrm{M}} \def\sM{\mathscr{M}{\mathcal M} \def\tM{\mathrm{M}} \def\sM{\mathscr{M}}
\def\mathfrak{m}} \def\bm{\mathbf{m}{\mathfrak{m}} \def\bm{\mathbf{m}}
\def\mathsf{m}} \def\ttm{\mathtt{m}{\mathsf{m}} \def\ttm{\mathtt{m}}
\def\mathrm{Mat}} \def\tMod{\mathrm{Mod}{\mathrm{Mat}} \def\tMod{\mathrm{Mod}}
\def\mathrm{max}} \def\tmin{\mathrm{min}{\mathrm{max}} \def\tmin{\mathrm{min}}
\def\mathrm{mod}} \def\tmult{\mathrm{mult}{\mathrm{mod}} \def\tmult{\mathrm{mult}}
\def\mathbb N} \def\cN{\mathcal N{\mathbb N} \def\cN{\mathcal N}
\def\mathsf{N}} \def\sN{\mathscr{N}{\mathsf{N}} \def\sN{\mathscr{N}}
\def\mathfrak{n}} \def\sfn{\mathsf{n}{\mathfrak{n}} \def\sfn{\mathsf{n}}
\def\mathbf{n}{\mathbf{n}}
\def\mathtt{n}} \def\tncpt{\mathrm{ncpt}{\mathtt{n}} \def\tncpt{\mathrm{ncpt}}
\def\mathrm{neg}} \def\tnilp{\mathrm{nilp}{\mathrm{neg}} \def\tnilp{\mathrm{nilp}}
\def\mathrm{Nilp}{\mathrm{Nilp}}
\def\mathrm{NL}{\mathrm{NL}}
\def\mathbb O} \def\cO{\mathcal O{\mathbb O} \def\cO{\mathcal O}
\def\mathrm{O}} \def\tOG{\mathrm{OG}{\mathrm{O}} \def\tOG{\mathrm{OG}}
\def\mathscr{O}{\mathscr{O}}
\def\mathfrak{o}} \def\todd{\mathrm{odd}{\mathfrak{o}} \def\todd{\mathrm{odd}}
\def\mathrm{op}{\mathrm{op}}
\def\mathbb P} \def\cP{\mathcal P{\mathbb P} \def\cP{\mathcal P}
\def\mathscr{P}} \def\ttP{\mathtt{P}{\mathscr{P}} \def\ttP{\mathtt{P}}
\def\mathsf{P}} \def\olP{{\overline P}{\mathsf{P}} \def\olP{{\overline P}}
\def\mathfrak{p}} \def\olfp{\overline{\mathfrak p}{\mathfrak{p}} \def\olfp{\overline{\mathfrak p}}
\def\mathsf{p}} \def\bp{\mathbf{p}{\mathsf{p}} \def\bp{\mathbf{p}}
\def\mathrm{prim}{\mathrm{prim}}
\def\mathrm{pt}{\mathrm{pt}}
\def\mathrm{Pic}{\mathrm{Pic}}
\def\tpos{\mathrm{pos}}
\def\mathrm{prin}} \def\tproj{\mathrm{proj}{\mathrm{prin}} \def\tproj{\mathrm{proj}}
\def\mathbb Q} \def\cQ{\mathcal Q} \def\sQ{\mathscr{Q}{\mathbb Q} \def\cQ{\mathcal Q} \def\sQ{\mathscr{Q}}
\def\mathsf{Q}} \def\ttQ{\mathtt{Q}{\mathsf{Q}} \def\ttQ{\mathtt{Q}}
\def\mathfrak{q}} \def\bq{\mathbf{q}{\mathfrak{q}} \def\bq{\mathbf{q}}
\def\mathsf{q}} \def\ttq{\mathtt{q}{\mathsf{q}} \def\ttq{\mathtt{q}}
\def\mathbb R} \def\cR{\mathcal R{\mathbb R} \def\cR{\mathcal R}
\def\mathsf{R}} \def\sR{\mathscr{R}{\mathsf{R}} \def\sR{\mathscr{R}}
\def\mathtt{R}{\mathtt{R}}
\def\mathfrak{r}{\mathfrak{r}}
\def\mathsf{r}} \def\ttr{\mathtt{r}{\mathsf{r}} \def\ttr{\mathtt{r}}
\def\mathrm{Re}} \def\tRic{\mathrm{Ric}{\mathrm{Re}} \def\tRic{\mathrm{Ric}}
\def\mathrm{rank}{\mathrm{rank}}
\def\mathbb S} \def\cS{\mathcal S{\mathbb S} \def\cS{\mathcal S}
\def\mathscr{S}{\mathscr{S}}
\def\fS{\mathfrak{S}} \def\sfS{\mathsf{S}}
\def\ttS{\mathtt{S}}
\def\mathfrak{s}{\mathfrak{s}}
\def\mathbf{s}} \def\sfs{\mathsf{s}{\mathbf{s}} \def\sfs{\mathsf{s}}
\def\mathrm{ss}{\mathrm{ss}}
\def\mathrm{SL}} \def\tSO{\mathrm{SO}{\mathrm{SL}} \def\tSO{\mathrm{SO}}
\def\mathrm{Sp}} \def\tSpin{\mathit{Spin}{\mathrm{Sp}} \def\tSpin{\mathit{Spin}}
\def\mathfrak{spin}{\mathfrak{spin}}
\def\mathrm{Seg}} \def\tSing{\mathrm{Sing}{\mathrm{Seg}} \def\tSing{\mathrm{Sing}}
\def\mathrm{Stab}} \def\tSym{\mathrm{Sym}{\mathrm{Stab}} \def\tSym{\mathrm{Sym}}
\def\mathrm{Spec}{\mathrm{Spec}}
\def\mathrm{SU}{\mathrm{SU}}
\def\mathrm{sing}} \def\tspan{\mathrm{span}{\mathrm{sing}} \def\tspan{\mathrm{span}}
\def\fsl{\mathfrak{sl}} \def\fso{\mathfrak{so}}
\def\mathfrak{sp}} \def\fsu{\mathfrak{su}{\mathfrak{sp}} \def\fsu{\mathfrak{su}}
\def\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}{\mathcal T} \def\sT{\mathscr{T}} \def\bT{\mathbb{T}}
\def\mathsf{T}} \def\ttT{\mathtt{T}{\mathsf{T}} \def\ttT{\mathtt{T}}
\def\mathfrak{t}} \def\tft{{\tilde\ft}{\mathfrak{t}} \def\tft{{\tilde\mathfrak{t}} \def\tft{{\tilde\ft}}}
\def\mathsf{t}} \def\bt{\mathbf{t}{\mathsf{t}} \def\bt{\mathbf{t}}
\def\mathbb{U}} \def\cU{\mathcal U} \def\tU{\mathrm{U}{\mathbb{U}} \def\cU{\mathcal U} \def\tU{\mathrm{U}}
\def\mathscr{U}{\mathscr{U}}
\def\mathbf{u}} \def\sfu{\mathsf{u}{\mathbf{u}} \def\sfu{\mathsf{u}}
\def\mathfrak{u}{\mathfrak{u}}
\def\mathbb{V}{\mathbb{V}}
\def\bV{\mathbf{V}} \def\cV{\mathcal V}
\def\mathsf{V}} \def\sV{\mathscr{V}{\mathsf{V}} \def\sV{\mathscr{V}}
\def\mathbf{v}} \def\sfv{\mathsf{v}{\mathbf{v}} \def\sfv{\mathsf{v}}
\def\mathfrak{v}{\mathfrak{v}}
\def\mathit{vol}{\mathit{vol}}
\def\mathcal W} \def\sW{\mathscr{W}{\mathcal W} \def\sW{\mathscr{W}}
\def\mathsf{W}} \def\olW{{\overline W}{\mathsf{W}} \def\olW{{\overline W}}
\def\mathsf{w}{\mathsf{w}}
\def\mathcal X} \def\sfX{\mathsf{X}} \def\fX{\mathfrak{X}{\mathcal X} \def\sfX{\mathsf{X}} \def\fX{\mathfrak{X}}
\def\mathscr{X}{\mathscr{X}}
\def\overline{X}{}{\overline{X}{}}
\def\mathbf{x}{\mathbf{x}}
\def\mathcal{Y}} \def\sfY{\mathsf{Y}{\mathcal{Y}} \def\sfY{\mathsf{Y}}
\def\mathtt{Y}} \def\sY{\mathscr{Y}{\mathtt{Y}} \def\sY{\mathscr{Y}}
\def\mathbf{y}{\mathbf{y}}
\def\mathcal Z} \def\sfZ{\mathsf{Z}} \def\bZ{\mathbb Z{\mathcal Z} \def\sfZ{\mathsf{Z}} \def\bZ{\mathbb Z}
\def\mathtt{Z}} \def\sZ{\mathscr{Z}{\mathtt{Z}} \def\sZ{\mathscr{Z}}
\def\mathfrak{z}} \def\olfz{\overline \fz{\mathfrak{z}} \def\olfz{\overline \mathfrak{z}} \def\olfz{\overline \fz}
\def\mathsf{z}} \def\bar z{\mathbf{z}{\mathsf{z}} \def\bar z{\mathbf{z}}
\def\tfrac{1}{2}{\tfrac{1}{2}}
\def\tfrac{\ti}{2}{\tfrac{\ti}{2}}
\def\tfrac{1}{3}{\tfrac{1}{3}}
\def\tfrac{2}{3}{\tfrac{2}{3}}
\def\tfrac{\sfi}{4}{\tfrac{\sfi}{4}}
\def\tfrac{1}{4}{\tfrac{1}{4}}
\def\tfrac{3}{4}{\tfrac{3}{4}}
\def\tfrac{3}{2}{\tfrac{3}{2}}
\def\mathbbm{1}{\mathbbm{1}}
\def\quad\hbox{and}\quad{\quad\hbox{and}\quad}
\def\partial{\partial}
\def\stackrel{\hbox{\tiny{dfn}}}{=}{\stackrel{\hbox{\tiny{dfn}}}{=}}
\def{\hbox{\tiny{$\bullet$}}}{{\hbox{\tiny{$\bullet$}}}}
\def{F^\sb}{{F^{\hbox{\tiny{$\bullet$}}}}}
\def\hookrightarrow{\hookrightarrow}
\def\twoheadrightarrow{\twoheadrightarrow}
\def\bar z{\bar z}
\def\oplus{\oplus}
\def\otimes{\otimes}
\def\underline{\underline}
\def\widetilde{\widetilde}
\def\widehat{\widehat}
\def\mathrm{d}x{\mathrm{d}x}
\def\mathrm{d}y{\mathrm{d}y}
\def\mathrm{d}z{\mathrm{d}z}
\def\mathrm{d} \bar{z}{\mathrm{d} \bar{z}}
\def\hbox{\small{$\lrcorner\, $}}{\hbox{\small{$\lrcorner\, $}}}
\def\hbox{\small $\bigwedge$}{\hbox{\small $\bigwedge$}}
\def\hbox{\tiny $\bigwedge$}{\hbox{\tiny $\bigwedge$}}
\def\textstyle{\bigwedge}{\textstyle{\bigwedge}}
\def\underline{\delta}{\underline{\delta}}
\def\underline{\varepsilon}{\underline{\varepsilon}}
\def{\Lambda_\mathrm{wt}}{{\Lambda_\mathrm{wt}}}
\def{\Lambda_\mathrm{wt}^+}{{\Lambda_\mathrm{wt}^+}}
\def{\Lambda_\mathrm{rt}}{{\Lambda_\mathrm{rt}}}
\def\bO\bP^2{\mathbb O} \def\cO{\mathcal O\mathbb P} \def\cP{\mathcal P^2}
\def\tGr(\bO^3,\bO^6){\mathrm{Gr}(\mathbb O} \def\cO{\mathcal O^3,\mathbb O} \def\cO{\mathcal O^6)}
\newcounter{numcnt}
\newenvironment{numlist}{
\begin{list}{{\small{(\arabic{numcnt})}}}
{\usecounter{numcnt}
\setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newcounter{cnt}
\newenvironment{reflist}{
\begin{list}{$[$\arabic{cnt}$]$}
{\usecounter{cnt}
\setlength{\leftmargin}{25pt}
\setlength{\itemsep}{0.05in}
\setlength{\listparindent}{20pt}
} }
{\end{list}}
\newenvironment{steplist}{
\begin{list}{\emph{Step \arabic{cnt}:}}
{\usecounter{cnt} \setlength{\itemsep}{2pt}
\setlength{\leftmargin}{20pt}
\setlength{\labelwidth}{15pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newcounter{acnt}
\newenvironment{a_list}{
\begin{list}{{(\alph{acnt})}}
{\usecounter{acnt} \setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{a_list_emph}{
\begin{list}{{\emph{(\alph{acnt})}}}
{\usecounter{acnt} \setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{a_list_bold}{
\begin{list}{{\bf{(\alph{acnt})}}}
{\usecounter{acnt} \setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newcounter{Acnt}
\newenvironment{A_list}{
\begin{list}{{(\Alph{Acnt})}}
{\usecounter{Acnt} \setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt}}
}
{\end{list}}
\newenvironment{A_list_sc}{
\begin{list}{{\textsc{(\alph{Acnt})}}}
{\usecounter{Acnt} \setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{Apar_list}{
\begin{list}{{{\bf(\Alph{Acnt})}}}
{\usecounter{Acnt} \setlength{\itemsep}{3pt}
\setlength{\leftmargin}{0pt}
\setlength{\rightmargin}{0pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newcounter{icnt}
\newenvironment{i_list_emph}{
\begin{list}{{\emph{(\roman{icnt})}}}
{\usecounter{icnt} \setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{i_list}{
\begin{list}{{(\roman{icnt})}}
{\usecounter{icnt}
\setlength{\itemsep}{3pt}
\setlength{\parsep}{0pt}
\setlength{\topsep}{0pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newcounter{Icnt}
\newenvironment{I_list}{
\begin{list}{{(\Roman{Icnt})}}
{\usecounter{Icnt}
\setlength{\itemsep}{3pt}
\setlength{\leftmargin}{25pt}
\setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt}}
}
{\end{list}}
\newcounter{exam_cnt}
\newenvironment{examlist}{
\begin{list}{{\bf (\arabic{exam_cnt}) }}
{\usecounter{exam_cnt} \setlength{\itemsep}{8pt}
\setlength{\leftmargin}{30pt}
\setlength{\labelwidth}{25pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{examlistA}{
\begin{list}{{\bf (\Alph{exam_cnt}) }}
{\usecounter{exam_cnt}
\setlength{\itemsep}{8pt}
\setlength{\leftmargin}{30pt}
\setlength{\labelwidth}{25pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newcounter{mccnt}\newenvironment{mclist}
{\begin{list}{ {\bf (\alph{mccnt})} }
{\usecounter{mccnt}
\setlength{\leftmargin}{5pt} \setlength{\itemsep}{5pt}
\setlength{\labelwidth}{0pt} \setlength{\itemindent}{0pt}
\setlength{\listparindent}{20pt} }}
{\end{list}}
\newenvironment{circlist}{
\begin{list}{$\circ$}
{\usecounter{cnt} \setlength{\itemsep}{2pt}
\setlength{\leftmargin}{15pt} \setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{bcirclist}{
\begin{list}{\boldmath$\circ$\unboldmath}
{\usecounter{cnt} \setlength{\itemsep}{2pt}
\setlength{\leftmargin}{15pt} \setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{blist}{
\begin{list}{$\bullet$}
{\usecounter{cnt} \setlength{\itemsep}{2pt}
\setlength{\leftmargin}{20pt} \setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{sblist}{
\begin{list}{${\hbox{\tiny{$\bullet$}}}$}
{\usecounter{cnt} \setlength{\itemsep}{2pt}
\setlength{\leftmargin}{20pt} \setlength{\labelwidth}{20pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
\newenvironment{blistnoindent}{
\begin{list}{$\bullet$}
{\usecounter{cnt} \setlength{\itemsep}{2pt}
\setlength{\leftmargin}{10pt} \setlength{\labelwidth}{10pt}
\setlength{\listparindent}{20pt} }
}
{\end{list}}
|
\section{Introduction}
\label{sec:intro}
It is now well established that the high redshift
intergalactic medium (IGM) is enriched with heavy metals
to metallicites of $10^{-3}$ to $10^{-2}$~${\rm Z}_{\odot}$
\citep[e.g.,][]{cowie95, schaye03, simcoe04, aguirre08}.
While metals only constitute a fraction of the total baryon
budget, they play an integral role in our understanding
of galaxy formation and evolution by providing a fossil record
of star formation, and by impacting upon cooling-rates which can
alter structure on many scales \citep[e.g.,][]{haas13a}.
Metals are synthesized in and released
from stars located in very overdense environments,
therefore they need to travel large distances to reach the
diffuse IGM, and this transport is likely driven by
feedback from star formation and active galactic nuclei (AGN).
Simulations have shown that metal pollution by galactic winds yields
reasonable enrichment statistics, without destroying the filamentary pattern
that gives rise to the \hone\ \lya\ forest \citep{theuns02}.
Although the need for inclusion of these processes
is clear, the mechanisms responsible
are not resolved, even in state-of-the-art cosmological simulations,
making their implementation uncertain.
By comparing observed and theoretical
metal-line absorption in the IGM, we may be able to constrain
enrichment mechanisms such as outflows.
Models and simulations of the IGM have been used
to make predictions about sources of metal pollution.
\citet{booth12} established that the observations of \citet{schaye03} of
\cfour\ associated with weak \hone\ at $z\approx3$
can only be explained if the low-density IGM was enriched primarily by low-mass
galaxies ($M_{\text{halo}} \leq 10^{10}$~\msol) that drive
outflows to distances of $\sim10^2$~proper~kiloparsecs (pkpc),
and calculated that
$>10$\% of the simulated volume and $>50$\% of the baryonic mass
in their successful model
was polluted by metals.
The simulations studied by \citet{wiersma10} indicate that
at least half of the metals found in the $z=2$ IGM were ejected from
galaxies at $z\geq3$, and that the haloes hosting these galaxies had masses less than
$M_{\text{halo}} = 10^{11}$~\msol.
This picture is consistent with that inferred from observations by
\citet{simcoe04}, who estimate that half of all baryons are enriched
to metallities $>10^{-3.5}$~Z$_{\odot}$ by $z\sim2.5$.
Studies of the IGM using the direct detection of individual metal lines
can typically probe only relatively
overdense gas, which constitutes a very small volume fraction of the
Universe. In this work, we employ an approach known as the pixel
optical depth method \citep{cowie98, ellison00, schaye00a, aguirre02, schaye03, turner14},
and provide a public version of the code at
\url{https://github.com/turnerm/podpy}.
This technique is a valuable tool for studying the IGM, as it enables
us to detect metals statistically even in low-density gas.
At the redshifts studied in this work, direct detection of metal-line absorption
in regions of the spectrum contaminated by \hone\ is nearly
impossible due to the density of the forest of absorption features. By using
the pixel optical depth method, we can correct for contamination and derive
statistical properties of the absorption by metals in this region.
In this work we take advantage of the fact that this technique is
fast and objective, and can be applied uniformly to both observations
and simulations.
Our observational sample consists of new spectra of eight $3.62\leq z \leq 3.92$ QSOs
with uniform coverage and high signal-to-noise (S/N). We compare the results to the
Evolution and Assembly of Galaxies and their Environments (EAGLE)
cosmological hydrodynamical simulations \citep{schaye15, crain15}.
The EAGLE simulations are ideal for studying metal-line absorption in the IGM, as they
have been run at relatively high resolution in a cosmologically
representative volume ($2\times1504^3$ particles in a 100~cMpc box).
The fiducial EAGLE model is able to reproduce the present-day galaxy stellar mass function, galaxy
sizes and the Tully Fisher relation \citep{schaye15}, and has been found to match observations
of galaxy colours \citep{trayford15} and the evolution of galaxy stellar masses \citep{furlong15}
and sizes \citep{furlong16}.
Furthermore, the simulations are in good agreement with a number of relevant observables,
including the properties of \hone\ absorption at $z\sim2$--$3$ \citep{rahmati15},
as well as the \osix\ and \cfour\ \citep{schaye15}
and the \hone\ \citep{crain16} column density distribution functions (CDDFs) at $z\sim0$ (although
for the latter we note that a higher than fiducial resolution is needed to
achieve agreement).
\citet{rahmati16} compared observed metal-line CDDFs
for various ions and redshifts with the predictions from EAGLE, finding generally
good agreement. Their study also included CIV, which was compared to the observations of
\citet[$z=1.5\text{--}5.5$]{songaila05},
\citet[$z=1.5\text{--}4.0$]{dodorico10}, and
\citet[$z=1.6\text{--}4.4$]{boksenberg15}.
Although the redshift ranges of these surveys overlap with ours, they are
much wider and hence likely affected by evolutionary trends. These observational
studies each used different, subjective methods to decompose the absorption
into Voigt profiles, which limits their utility for testing models. Indeed,
the different studies do not agree with each other, making it relatively easy
for a single model to agree with the data. Here we will instead employ a like-for-like
comparison of new data with virtual EAGLE observations tailored to mimic the
characteristics of our own quasar spectra and analysed using the same automated
method. We will consider relations between the pixel optical depths of
\hone, \cthree, \cfour, \sithree, \sifour\ and \osix.
This paper is structured as follows.
In \S~\ref{sec:method}, we describe the observations and simulations.
We also summarize the pixel optical depth method,
and how it is applied.
The results are presented in \S~\ref{sec:results},
and we give a discussion and conclusions in \S~\ref{sec:discussion}
and \ref{sec:conclusion}, respectively.
Throughout this work, we denote proper and comoving
distances as pMpc and cMpc, respectively.
Both simulations and observations use cosmological parameters
determined from the \textit{Planck} mission \citep{planck13}, i.e.
$H_{\rm 0}=67.8$~\kmps~Mpc$^{-1}$, $\Omega_{\rm m} = 0.307$,
and $\Omega_{\Lambda} = 0.693$.
\section{Method}
\label{sec:method}
\subsection{Observations}
We analyze a sample of eight QSOs with $3.62\leq z_{\text{QSO}} \leq 3.922$.
They were selected based on their redshift and the existence
of substantial, high S/N data taken with VLT/UVES. Initially,
there were already 76.0 hours of UVES data, excluding overheads,
of the QSOs.
Follow-up observations to fill in the gaps and improve S/N were
completed in 62.7 hours of on-source time in
programmes 091.A-0833, 092.A-0011 and 093.A-0575 (P.I. Schaye).
We note that for Q1422$+$23, the gaps in the UVES data were filled
using archival observations with Keck/HIRES of comparable S/N and resolution
(which is $\approx8.5$~\kmps).
The properties of the QSOs and the S/N of the spectra
are summarized in Table~\ref{tab:qso}.
The reduction of the UVES data was performed using the \texttt{UVES\_headsort} and
\texttt{UVES\_popler} software by
Michael T. Murphy, and binned to have a uniform velocity dispersion
of 1.3~\kmps.
The HIRES data was reduced using T. Barlow's \texttt{MAKEE} package, and binned on to 2.8~\kmps\ pixels.
The continuum fits for the spectra were performed by hand by M.\ Turner.
Any DLAs or Lyman break regions (i.e., due to
strong absorbers in \hone) were masked out, with the exception of DLAs in the \lya\ forest,
which were unmasked when recovering the \hone\ to be used for
subtraction of contaminating absorption by higher-order Lyman
series lines from \osix\ and \cthree\ optical depths.
To homogenize the continuum fitting errors, we implemented the automated
sigma-clipping procedure of \citet{schaye03} at wavelengths greater than
that of the QSO's \lya\ emission, which was applied to both the observed and simulated spectra.
The spectrum is divided into rest-frame bins of $\Delta \lambda = 20$~\AA,
which have central wavelength $\lambda_i$ and median flux $\bar{f_i}$.
A B-spline of order 3 is then interpolated through all $\bar{f_i}$ values, and any pixels with flux
$N_{\sigma}^{\rm cf}\times\sigma$ below the interpolated values are discarded, where $\sigma$ is
the normalized noise array.
We then recalculate $\bar{f_i}$ without the discarded pixels,
and repeat the procedure until convergence is reached.
We use $N_{\sigma}^{\rm cf}=2$, which has been shown to be optimal in the \cfour\ region
for spectra with a quality similar to ours,
as it induces errors that are smaller than the noise by at least an order of magnitude \citep{schaye03}.
\begin{table*}
\caption{Properties of the QSOs used in this work, and the median S/N in the \hone\ \lya\ and \cfour\
recovery regions. The columns list, from left to right,
name, right ascension, declination, redshift, magnitude and magnitude band (either $V$-band (V), $R$-band (R),
or photographic (P)) from \citet{veron10},
median signal-to-noise in the \lya\ forest
region and \cfour\ region, respectively (see \S~\ref{sec:redshift}
for the definition of these regions),
the median optical depth of \hone\ in the \lya\ forest, the factor $C_{\rm UVB}$ used to scale the UVB in the simulations,
and the redshift midpoint of eq.~\ref{eq:zlim}
that defines the \lya\ forest region.}
\begin{center}
\begin{tabular}{l l l l l l r r r r r}
\hline
Name & R.A. & Dec & $z_\text{QSO}$ & Mag & Band & S/N$_{\lyam}$ & S/N$_{\cfourm}$ & $\log_{10}\tau^{\rm med}_{\honem}$ & $C_{\rm UVB}$ & $z^{\rm mid}_{\lyam}$\\
\hline \hline
Q1422$+$23 & \tt{ 14:24:38 } & \tt{ +22:56:01 } & 3.620 & 15.84 & V & 87 & 82 & $-0.605$ & 0.896 & 3.24\\
Q0055$-$269 & \tt{ 00:57:58 } & \tt{ -26:43:14 } & 3.655 & 17.47 & P & 60 & 79 & $-0.583$ & 1.123 & 3.27\\
Q1317$-$0507 & \tt{ 13:20:30 } & \tt{ -18:36:25 } & 3.700 & 18.10 & P & 59 & 90 & $-0.549$ & 1.009 & 3.31\\
Q1621$-$0042 & \tt{ 16:21:17 } & \tt{ -23:17:10 } & 3.709 & 17.97 & V & 78 & 92 & $-0.546$ & 0.807 & 3.32\\
QB2000$-$330 & \tt{ 20:03:24 } & \tt{ -32:51:44 } & 3.773 & 17.30 & R & 105 & 83 & $-0.501$ & 0.809 & 3.38\\
PKS1937$-$101 & \tt{ 19:39:57 } & \tt{ -13:57:19 } & 3.787 & 17.00 & R & 96 & 64 & $-0.494$ & 1.293 & 3.39\\
J0124$+$0044 & \tt{ 01:24:03 } & \tt{ +00:44:32 } & 3.834 & 18.71 & V & 48 & 59 & $-0.445$ & 0.898 & 3.43\\
BRI1108$-$07 & \tt{ 11:11:13 } & \tt{ -15:55:58 } & 3.922 & 18.10 & R & 29 & 29 & $-0.390$ & 0.808 & 3.51\\
\hline
\end{tabular}
\end{center}
\label{tab:qso}
\end{table*}
\subsection{Simulations}
We compare the observations to predictions from the EAGLE cosmological
hydrodynamical simulations. EAGLE was run with a substantially modified version of the $N$-body TreePM
smoothed particle hydrodynamics (SPH) code \texttt{GADGET}~3
(last described in \citealt{springel05a}). EAGLE uses the package of hydrodynamics updates
``Anarchy'' (Dalla Vecchia, in prep.; see Appendix~A1 of \citealt{schaye15}) which invokes the pressure-entropy formulation
of SPH from \citet{hopkins13}, the time-step limiter from \citet{durier12},
the artificial viscosity switch from \citet{cullen10}, an artificial conduction
switch close to that of \citet{price08}, and the $C^2$ \citet{wendland95} kernel.
The influence of the updates is explored by \citet{schaller15}.
The fiducial EAGLE model
is run in a 100~cMpc periodic box with $1504^3$ of both dark matter and baryonic particles,
and is denoted Ref-L1001504. To test convergence with resolution and box size,
runs varying the number of particles and box size were also conducted, and are
listed in Table~\ref{tab:sims}.
The stellar feedback in EAGLE is implemented as described by \citet{dallavecchia12},
where thermal energy is injected
stochastically. While the temperature of heated particles is always increased
by $10^{7.5}$~K, the probability of heating varies with the local metallicity and density \citep{schaye15, crain15}.
The simulations include thermal AGN feedback \citep{booth09},
also implemented stochastically \citep{schaye15}.
Both stellar and AGN feedback have been calibrated such that the simulations
match the observed $z\sim0$ stellar mass function and galaxy--black hole mass relation,
and give sensible disc galaxy sizes. We note that of the two highest-resolution runs,
Ref-L025N072 has been realized with the same subgrid parameters used in the fiducial model,
while for the Recal-L025N072 the subgrid parameters were re-calibrated to better match
the observed galaxy stellar mass function.
EAGLE also includes a subgrid model for photo-heating
and radiative cooling via eleven elements: hydrogen, helium, carbon, nitrogen, oxygen,
neon, magnesium, silicon, sulphur, calcium and iron \citep{wiersma09a},
assuming a \citet{haardt01} UV and X-ray background.
Star formation is implemented with a gas metallicity-dependent density threshold \citep{schaye04}
as described by \citet{schaye08}, followed by stellar evolution
and enrichment from \citet{wiersma09b}.
Finally, details of the subgrid model for black hole seeding and growth
can be found in \citet{springel05b, rosas13} and \citet{schaye15}.
For each of our eight observed QSOs in Table~\ref{tab:qso},
we synthesize 100 corresponding mock spectra using the
\texttt{SPECWIZARD} package by Schaye, Booth, and Theuns (implemented
as in \citealt{schaye03}, see also Appendix~A4 of \citealt{theuns98}).
To create mock spectra that resemble the observed QSOs and whose absorption features
span a large redshift range, we follow \citet{schaye03} and stitch together
the physical state of the gas intersecting uncorrelated sightlines from
snapshots with different redshifts. The ionization balance of each gas particle
is estimated using interpolation tables generated from \texttt{Cloudy} \citep[][version 13.03]{ferland13}
assuming uniform illumination by an ultra-violet background (UVB).
We take the QSO+galaxy \citet{haardt01} UVB (denoted as ``HM01'')
to be our fiducial model,\footnote{ Our choice of the HM01 UVB is primarily
to maintain consistency with the EAGLE simulations, in which the \citet{haardt01} UVB is
also used to calculate radiative cooling rates.
While more recent
UVBs are available \citep[e.g.,][]{fauchergiguere09, haardt12}, they
do not necessarily provide a better match to observations \citep[e.g.,][]{kollmeier14}.
}
and have plotted the intensity
as a function of energy at $z=3.5$ in Fig.~\ref{fig:uvb}.
We also consider
the \citet{haardt01} background using quasars only (``Q-only''), which
is much harder than the fiducial model above $\approx4$~Ryd.
Furthermore, to explore the possible effects of a
delayed \hetwo\ reionization, we consider a UVB that is
significantly softer above 4~Ryd. To implement this,
we use the QSO+galaxy model and reduce the intensity above 4~Ryd
by a factor of 100, which we denote as ``4Ryd-100''.
Self-shielding for \hone\ was included by modifying the ionization fraction
using the fitting functions of \citet{rahmati13a}.
The normalization of the UVB is set such that the median recovered \hone\ \lya\ optical
depth of the simulated QSOs agrees with that of the observations at the same redshift.
A unique value of $C_{\rm UVB}$ is determined for each
observed QSO and corresponding set of 100 mock spectra,
and the values are presented in Table~\ref{tab:qso}.
In the EAGLE simulations, the dense particles
that represent the multiphase interstellar
medium (ISM, $n_{\rm H} > 0.1$~cm$^{-3}$) are not allowed to cool
below an effective equation of state (EoS),
and their temperature can be interpreted as a proxy for the pressure at which the warm
and cool ISM phases equilibriate.
We set their
temperatures to $10^4$~K when generating
the mock spectra, although we note that due to the small cross-section
of such dense absorbers the effect of including them is negligible.
\begin{figure}
\includegraphics[width=0.5\textwidth]{figures/uvb.pdf}
\caption{The intensity as a function of energy for the different UVB models.
The different models are: HM01 QSO+galaxy \citep{haardt01}, which is our
fiducial model; Q-only, which is also by \citet{haardt01} but
only considers an ionizing contribution from QSOs; and
4Ryd-100, the same as the fiducial model except that
the intensity is reduced by a factor of 100 above 4~Ryd.
The vertical light grey lines indicate the ionization energies
of ions of interest, where increasing line thickness denotes silicon,
carbon and oxygen ions, respectively.
All of the UVBs have been normalized
to have the same intensity as HM01 at 1~Ryd.
}
\label{fig:uvb}
\end{figure}
Each set of 100 mock spectra is synthesized to have redshifts
identical to that of their corresponding observed QSO, and we consider absorption
ranging from $1.5 < z < z_{\text{QSO}}$ in every case.
We include contributions from 31 \hone\ Lyman series transitions beginning with \lya,
and metal-line absorption from
\ctwo, \cthree, \cfour, \fetwo, \nfive, \osix, \sitwo, \sithree, and
\sifour\ (see Appendix~\ref{app:trans} for the rest wavelengths and oscillator strengths
of these transitions).
To match the spectral properties of the observations, the simulated spectra are convolved with a Gaussian
with a FWHM of 6.6~\kmps, and resampled on to pixels of 1.3~\kmps. For each observed
QSO, we have measured the RMS noise in bins of 150~\AA\ in wavelength and 0.2 in normalized flux.
We then use these measurements to add random Gaussian noise with the same variance to the simulations.
\begin{table*}
\caption{Characteristics of the EAGLE simulations. From left to right, the columns list the simulation
name, box size, number of particles,
initial baryonic particle mass, dark matter particle mass, comoving (Plummer-equivalent) gravitational softening, and
maximum physical softening.}
\begin{center}
\begin{tabular}{rrrccccl}
\hline
Simulation & $L$ & $N$ & $m_{\rm b}$ & $m_{\rm dm}$ & $\epsilon_{\rm com}$ & $\epsilon_{\rm prop}$ \\
& [cMpc] & & [\msol] & [\msol] & [ckpc] & [pkpc] \\
\hline
\hline
{Ref-L100N1504} & 100 & $2\times1504^3$ & $1.81 \times 10^6$ & $9.70 \times 10^6$ & 2.66 & 0.70 \\
{Ref-L050N0752} & 50 & $2\times752^3$ & $1.81 \times 10^6$ & $ 9.70 \times 10^6$ & 2.66 & 0.70 \\
{Ref-L025N0376} & 25 & $2\times376^3$ & $1.81 \times 10^6$ & $ 9.70 \times 10^6$ & 2.66 & 0.70 \\
{Ref-L025N0752} & 25 & $2\times752^3$ & $2.26 \times 10^5$ & $ 1.21 \times 10^6$ & 1.33 & 0.35 \\
{Recal-L025N0752} & 25 & $2\times752^3$ & $2.26 \times 10^5$ & $ 1.21 \times 10^6$ & 1.33 & 0.35 \\
\hline
\end{tabular}
\end{center}
\label{tab:sims}
\end{table*}
\subsection{Analyzed redshift range}
\label{sec:redshift}
The first step for the pixel optical depth
recovery involves choosing optimal
redshift limits. The fiducial redshift
range is selected to lie in the \lya\ forest, defined to be:
\begin{equation}
(1 + \zqsom) \dfrac{\lambda_{\lybm} }{ \lambda_{\lyam}}- 1 \leq z \leq \zqsom - (1+\zqsom) \dfrac{3000\,\kmpsm}{c}
\label{eq:zlim}
\end{equation}
where $\lambda_{\lyam}=1215.7$~\AA\ and $\lambda_{\lybm}=1025.7$~\AA\
are the \hone\ \lya\ and \lyb\ rest wavelengths, respectively.
The lower limit was chosen to avoid the \lyb\ forest and corresponds
to the \lyb\ transition at the
redshift of the QSO, while the upper limit is 3000~\kmps\ bluewards of the QSO redshift
to avoid any proximity effects.
For \hone, \cfour\ ($\lrestm = [1548.2, 1550.8]$~\AA) and
\cthree\ ($\lrestm = 977.0$~\AA) we use the above
redshift limits. For the remaining ions, we make slight modifications,
listed below,
in order to homogenize the contamination.
We use the notation $\lambda_{Z,k}$ to denote the rest wavelength of
multiplet component $k$ of the ion $Z$.
\begin{enumerate}
\item \osix\ ($\lrestm=[1031.9, 1037.6]$~\AA):
We limit the recovery to where \osix\ overlaps with the \lyb\ forest and
place a cut-off at the \lya\ forest region, which leads to
$z_{\rm max} = (1 + \zqsom) \lambda_{\honem,\lybm} / \lambda_{\osixm,2} - 1$
\item \sithree\ ($\lrestm=1206.6$~\AA):
We constrain the recovered optical depth region to not extend
outside of the \lya\ forest. For \sithree, which extends
slightly bluewards into the \lyb\ forest, we set
$z_{\rm min} =(1 + \zqsom) \lambda_{\lybm} / \lambda_{\sithreem}$.
\item \sifour\ ($\lrestm = [1393.8, 1402.8]$~\AA):
To avoid contamination by the \lya\ forest, we limit the blue end of the \sifour\ recovery
by setting
$z_{\rm min} = (1+\zqsom)\lambda_{\lyam}/\lambda_{\sifourm,1}-1$.
\end{enumerate}
\subsection{Pixel optical depth method}
\label{sec:pod}
We employ the pixel optical depth method, which we
use to study absorption on an individual pixel basis rather than
by fitting Voigt profiles to individual lines.
The goal is to obtain statistics on absorption by \hone\ and
various metal ions in the IGM, and on how their absorption relates to one another.
Our implementation is close to that of \citet{aguirre02}, but with
the improvements of \citet{turner14}, and the effects of varying the
chosen parameters can be found in both works.
The exact methodology
is described in full detail in Appendix~A of \citet{turner14},
and is summarized below.
A working version of the code can be found at
\url{https://github.com/turnerm/podpy}.
After restricting the redshift range, the next step is to convert the flux of every
pixel of ion $Z$ and multiplet component $k$ to an optical depth
$\tau_{Z,k}(z) = -\ln(F)$, where $F(\lambda)$ is the normalised flux at
$\lambda = \lambda_{k}(1+z)$. Then, depending on the ion, corrections are made
for saturation or contamination, as described below.
\begin{enumerate}
\item For \hone\ \lya, while there is very little contamination in the \lya\ forest, the absorption in
many of the pixels will be saturated, and we use the higher order Lyman series
transitions to correct for this. Specifically, if we consider a \lya\ pixel to be saturated,
we look to $N=16$ higher-order Lyman lines (beginning with \hone\ \lyb),
and take the minimum optical depth, scaled to that of \lya, of all unsaturated pixels at the same redshift (if any).
If we are unable to correct the pixel due to saturation of the higher-order transitions,
we set it to $10^4$ (so that these pixels can still be used for computing the median). Finally, we search for and discard
any contaminated pixels, by checking that higher-order transitions
do not have optical depth values significantly below what would be expected from the scaled \hone\ \lya\ optical depth.
\item For \osix\ and \cthree, we can use the corrected \hone\ \lya\ optical depths
to estimate and subtract contamination by \hone.
We do so beginning with \hone\ \lyb\ ($N=2$) and use higher-order Lyman series orders up to $N=5$.
For saturated \osix\ and \cthree\ pixels, the optical depth is not well defined
and therefore the above subtraction is not performed. Instead, we leave the pixel
uncorrected, unless the saturation can be attributed to \hone, in which case
the pixel is discarded.
\item \sifour\ and \osix\ are both closely-spaced doublets, and we can
use this fact to correct for contamination. To do so, we scale the optical depth
of the weak component to match that of the strong component, and take the minimum of the
two components modulo noise.
We only take the scaled optical depth of the weaker component if it is significantly
lower (when taking into account the noise array) than the stronger component.
\item For \cfour, which is a strong transition redward of the \lya\ forest,
the largest source of contamination is by its own doublet.
To correct for this, we perform an iterative self-contamination correction.
We first discard any pixels determined to be contaminated by other ions, by checking
whether the optical depth of a pixel is too high
to be explained by half of the associated stronger component combined with twice
the associated weaker component. We then subtract the estimated contribution of the
weaker component from each pixel, iterating until convergence is reached.
\end{enumerate}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/c4_vs_h1_q03_leg.png}
\end{center}
\caption{Example analysis of the \cfour(\hone) relation for Q1317$-$507.
The measured optical depths from the observations are shown by the points, and those recovered from the mock
spectra are indicated by the curves,
with the different colours and line styles representing variations in the model. $1\sigma$ errors are indicated by the bars (shaded region)
for the observations (fiducial simulation), and were determined by bootstrap resampling spectral chunks (mock spectra).
We indicate \taumin\ by the horizontal dashed and dotted lines for the observations and
simulations, respectively.
In Appendix~\ref{sec:single} we provide individual results for all eight QSOs,
and for all relations examined in this work.
}
\label{fig:example}
\end{figure}
\subsection{Analysis}
\label{sec:pod_as}
For the analysis, we would like to see how the absorption from one ion
correlates with that from another. The procedures used here are also described in
\S~3.4 and 4.2 of \citet{aguirre04}.
As an illustrative example, we will consider the ions
\cfour\ and \hone. For a single observed QSO, we use the recovered pixel
optical depths to construct a set of pixel pairs where each pair shares the same redshift.
We then divide the ions into bins of $\log\tau_{\honem}$, and take the median
$\tau_{\honem}$ and $\tau_{\cfourm}$ in each bin , to obtain
$\tau_{\cfourm}^{\text{med}}(\tau_{\honem})$,
which from this point forward we will denote as \cfour(\hone).
The result of this procedure applied to
one of our QSOs is shown in Fig.~\ref{fig:example},
and we briefly describe the characteristics here.
We note that the results from individual QSOs for all
relations examined here are given in Appendix~\ref{sec:single}.
We make note of two different regimes within the \cfour(\hone) relation.
The first is on the right-hand side of Fig.~\ref{fig:example},
where $\tau_{\honem}\gtrsim1$. Here, the median \cfour\ optical
depth increases with \hone, which indicates that the pixels are probing gas enriched by \cfour.
The value of $\tau_{\cfourm}^{\text{med}}$ constrains the
number density ratio of \cfour\ to \hone.
Next, we turn to the region with $\tau_{\honem}\lesssim1$, where
$\tau_{\cfourm}^{\text{med}}$ is approximately constant.
This behaviour arises because the median \cfour\ optical depths reach values
below the flat level \taumin, which is essentially a detection limit
set by noise, contamination, and/or continuum fitting errors.
An important caveat to keep in mind throughout this work is that
the median recovered metal-line optical depth is not necessarily representative
of \textit{typical} intrinsic pixel optical depths for a given \hone\ bin.
In particular, as the metal-line optical
depths approach the flat level, it is likely that many individual pixels in a given
\hone\ bin have intrinsic metal optical depths at or below the flat level itself. In this case, the median
recovered metal optical depth will be determined by the fraction of pixels
that have optical depths above the flat level.
To construct the \cfour(\hone) relation for the
observed spectra, \hone\ bins containing fewer than 25 pixels in total
are discarded.
Furthermore, we divide each spectrum into chunks of 5~\AA\ (chosen to be much greater than the line widths),
and discard any bins containing fewer than 5 unique chunks. This is done to ensure that the
median optical depths are obtained from more than just a few pixels in a very small spectral region.
To measure errors on $\tau_{\cfourm}^{\text{med}}$,
we create new spectra by bootstrap
resampling the chunks 1000 times with replacement. We then measure
\cfour(\hone)
for each bootstrap realization of the spectrum and take the error in each $\tau_{\honem}$ bin
to be the $1\sigma$ confidence interval of all
realizations.
For the simulated spectra, we measure
\cfour(\hone)
for each mock spectrum, and require that each $\tau_{\honem}$ bin
have at least 5 pixels in total.
Next, we combine the results for all 100 mock spectra
associated with a single observed QSO by measuring
the median \cfour\ optical depth in each $\tau_{\honem}$ bin
for all spectra, and we discard any bin containing contributions
from fewer than 5 spectra.
Errors are calculated by bootstrap resampling the spectra 1000 times.
Next, we compute the flat levels \taumin\ by taking
the median of all pixels that have
$\tau< \tau_c$, where we choose $\tau_c$ to be an optical depth below
which we do not find any correlations. As in \citet{aguirre08}, we take $\tau_c=0.1$ when binning in \hone,
and $0.01$ when binning in \cfour\ and \sifour. To estimate the error
on \taumin, for the observations we again divide the spectrum
into 5~\AA\ chunks, measure \taumin\ for 1000 bootstrap realizations,
and take the $1\sigma$ confidence interval. For the simulations, we calculate
\taumin\ for each spectrum, and take the final value to be the median
value from all 100 spectra.
Finally, below we outline the steps for combining the results from the different QSOs,
which is applied to both the observed relations as well as their respective counterparts from the mocks.
Because our sample is uniform in terms of S/N,
we simply combine the binned data points directly without subtracting \taumin.
However, because the implementation of the noise, continuum fitting errors and contamination
in simulations is not completely accurate, the flat levels differ from the observations.
To account for this offset, we linearly add the difference between flat levels
($\tauminm^{\text{obs}} - \tauminm^{\text{sims}}$) to the median
optical depths in the simulations. We have verified that performing this step
before the QSOs are combined does not modify the results.
Next, to measure the combined median values,
we perform $\chi^2$ fitting of a single value of $\tau_{\cfourm}^{\text{med}}$ to
all points in the bin, which is plotted against the central value of each \hone\ bin
(in contrast to the results from individual QSOs, which are plotted against
the median of all \hone\ pixel optical depths in each bin).
We discard any data points that have contributions from fewer than four QSOs,
and the $1\sigma$ errors are estimated by bootstrap resampling the QSOs.
The combined results for \cfour(\hone) can be seen in the left-hand panel of Fig.~\ref{fig:hone}.
\begin{figure*}
\includegraphics[width=0.33\textwidth]{{figures/c4_vs_h1_allqso_p50_force_tau_min_comb_leg}.pdf}
\includegraphics[width=0.33\textwidth]{{figures/si4_vs_h1_allqso_p50_force_tau_min_comb}.pdf}
\includegraphics[width=0.33\textwidth]{{figures/o6_vs_h1_allqso_p50_force_tau_min_comb}.pdf}
\caption{Median recovered pixel optical depths binned by \hone\ for \cfour\ (left),
\sifour\ (centre) and \osix\ (right).
Note that the dynamic range shown along the $y$-axis decreases strongly
from left to right.
The data from eight QSOs have been
combined, and the $1\sigma$ error bars are measured by bootstrap resampling the QSOs.
The black circles show the data, while the curves denote the results from
simulations, where different colours and line styles represent variations in the model and
we show the $1\sigma$ error region around the fiducial HM01 model. The flat level \taumin,
which is the same for the observations and simulations by construction,
is indicated by the dashed horizontal line. The data from the observations
is provided in Table~\ref{tab:data_0}.
The
simulation run with the fiducial UVB systematically underpredicts the median
\cfour\ and \sifour\ optical depths. The discrepancy is
lessened by invoking a softer UVB (4Ryd-100), higher metallicity (Zrel-10),
or unresolved turbulent broadening (bturb-100),
although the metal-line optical depths associated with the strongest \hone\ absorption
are still underestimated by $\approx0.5$~dex. In contrast, the predicted
\osix(\hone) relation (right-hand panel) is insensitive
to the UVB models, and in good agreement with the observations.
}
\label{fig:hone}
\end{figure*}
\begin{table}
\caption{Observational data shown in Fig.~\ref{fig:hone}. The left column
indicates the central $\log_{10}\tau_{\honem}$ value of each bin, and the next three columns list the $\log_{10}$ median
recovered optical depths for the \cfour(\hone), \sifour(\hone), and \osix(\hone) relations,
respectively, along with the $1\sigma$ errors. We note that for the bin where
$\log_{10}\tau_{\honem}=0.70$, the metal optical depths are
not defined because there were contributions from fewer that four of our eight quasars.}
\begin{center}
\input{tables/data_0.tab}
\end{center}
\label{tab:data_0}
\end{table}
\section{Results}
\label{sec:results}
\subsection{$\tau_{Z}$ as a function of $\tau_{\rm HI}$}
We begin by examining median metal-line pixel optical
depths as a function of \hone\ pixel optical depth in Fig.~\ref{fig:hone}, where we have plotted
\cfour(\hone), \sifour(\hone) and \osix(\hone) from left to right.
The grey points with error bars represent the observations,
while the curves show results from simulations, with different
colours indicating variations in the model. The data from
the observations is presented in Table~\ref{tab:data_0}.
The relations displayed in Fig.~\ref{fig:hone} depend on the following quantities
in the simulations, using \cfour\ as an illustrative example:
\begin{equation}
\log_{10} \dfrac{\tau_{\cfourm}}{\tau_{\honem}} =
\log_{10} \dfrac{(f\lambda)_{\cfourm}}{(f\lambda)_{\honem}}
\dfrac{n_{\cfourm}}{n_{\rm C}}
\dfrac{n_{\rm H}}{n_{\honem}}
+ [\text{C}/\text{H}]
+ (\text{C}/\text{H})_{\odot} ,
\label{eq:metallicity}
\end{equation}
where $f$ is the oscillator strength, $\lambda$ is the rest wavelength, and
$n$ is the number density.
While the oscillator strength and rest wavelength are fixed
empirical quantities, the element abundances are predicted, and
the ionization fractions (i.e., the ratio of ionized to total number density)
are determined using the particle temperatures and densities in the same manner
as was done to compute the cooling rates used in the simulation.
In the following analysis, we will consider the results in
two different regimes, separated by $\tau_{\honem}\approx10$. The reasons for this are:
(1) For $\tau_{\honem}\gtrsim10$ \hone\ pixel optical depths
will be highly saturated, and even though this is corrected for in our recovery procedure, the final values
still suffer from large uncertainties compared to their unsaturated counterparts.
(2) If the gas being probed is mainly in photoionization equilibrium,
which is a reasonable assumption for \cfour\ and \sifour\ \citep{schaye03, aguirre04}, then
\hone\ is considered a good tracer of the density \citep{schaye01}, even on an individual
pixel basis \citep{aguirre02}.
This means that higher \hone\ optical
depths probe dense regions closer to galaxies,
rather than the diffuse IGM.
To touch on this point more quantitatively, we have calculated
the optical depth-weighted overdensities for
our full sample of mock spectra. Fig.~\ref{fig:delta} shows the median optical depth-weighted
overdensity as a function of \hone\ for
\hone, \cfour, \sifour, and \osix, and we find a strong correlation
in every case. We have performed ordinary least-squares
fits to the curves using a power-law function:
\begin{equation}
\log_{10} \delta =A \times \log_{10} \tau_{\honem} + B,
\label{eq:delta_vs_hone}
\end{equation}
and give the resulting parameters in Table~\ref{tab:delta}. We note that our aforementioned division of
the \hone\ optical depths into two regimes at $\tau_{\honem}\approx10$ corresponds to an overdensity
of $\approx8$ for the gas responsible for the \hone\ absorption, but that the overdensity of the metal
ions at the same redshifts as the selected \hone\ pixels is typically a factor of a few lower.
We also caution that the quantities presented here are pixel-weighted,
and may therefore be biased to high temperatures since higher temperature lines will be more broadened
and therefore cover more pixels. This may also bias the densities to lower values, as intergalactic gas with $T\gtrsim10^5$~K
tends to correspond to lower densities than that found at lower temperatures.
\citep[e.g., Fig.~1 from][]{vandevoort11a}.
The fitted values for \hone\ can be compared to eq.~8 of \citet{rakic12}, who used
the relations of \citet{schaye01} to obtain an expression for overdensity as a
function of \hone\ \lya\ line centre optical depth $\tau_{0,\lyam}$, under the
assumption that the absorbers have
sizes of the order of the local Jeans length. Modifying their eq.~8
using the parameters from this work, we obtain:
\begin{equation}
\begin{split}
\delta \approx& \, 0.74\, \tau_{0,\lyam}^{2/3}\
\left(\dfrac{\Gamma_{12}}{0.85\times10^{12}\text{ s}^{-1}} \right)^{2/3}
\left( \dfrac{1+z}{4.36}\right)^{-3} \\
\times& \left(\dfrac{T}{2\times10^4\text{ K}} \right)^{0.17}
\left( \dfrac{f_{\rm g}}{0.15} \right)^{-1/3}
\left( \dfrac{b}{20 \text{ \kmps}} \right)^{2/3}.
\end{split}
\label{eq:rakic_eq8}
\end{equation}
To determine the redshift $z$, for each QSO we considered the
redshift at the centre of the \lya\ forest as defined by eq.~\ref{eq:zlim},
and took the mean of all of the values. To obtain the photoionization rate
$\Gamma_{12}$, we multiplied the photoionization rate from \citet{haardt01} at
each of the above redshifts
by the scale factor that was used to bring the median recovered \hone\ optical depth of
the mock spectra into agreement with that of the corresponding observed spectrum, and took the mean for
all set of mocks. As in \citet{rakic12}, we chose a temperature typical of
a moderately-dense region in IGM \citep[e.g.,][]{schaye00b, lidz10, becker11, rudie12b},
and assumed that the gas fraction $f_{\rm g}$ corresponds to the universal value of $\Omega_{\rm b}/\Omega_{\rm m}$.
Finally, we have taken $b$ to be $20$~\kmps, which is similar to
values measured by \citet{rudie12} for $z\approx2.5$.
Putting eq.~\ref{eq:rakic_eq8} into the form of eq.~\ref{eq:delta_vs_hone}, we obtain
$A=0.67$ and $B=-0.13$, compared with $A=0.66\pm0.01$ and $B=0.16\pm0.01$ measured from our simulations.
Hence, the slope $A$ is in excellent agreement with the theoretical scaling relation implied by the model
of \citet{schaye01}. The normalization agrees to within a factor of two, which we consider
good agreement given the uncertainties in the fiducial parameter values.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{{figures/delta_leg}.pdf}
\end{center}
\caption{Median optical depth-weighted overdensity as a function of \hone\ optical depth
for the simulated spectra. The different curves indicate the ion used for the
weighting. We have indicated the one $1\sigma$ error region around the \hone-weighted overdensity.
The parameters from an ordinary least-squares fits of the form
$\log_{10} \delta =A \times \log_{10} \tau_{\honem} +B$ to each of the relations
are given in Table~\ref{tab:delta}. The metal-line absorption typically arises in lower-density
gas than the \hone\ absorption with the same redshift.
}
\label{fig:delta}
\end{figure}
\begin{table}
\caption{Results from fitting a power-law $\log_{10} \delta =A \times \log_{10} \tau_{\honem} +B$ to
the relations in Fig.~\ref{fig:delta}, using ordinary least-squares.}
\begin{center}
\begin{tabular}{lrr}
\hline
Ion & $A$ & $B$ \\
\hline \hline
\hone & $ 0.66\pm 0.01$ & $ 0.16\pm 0.01$ \\
\cfour & $ 0.37\pm 0.02$ & $-0.02\pm 0.01$ \\
\sifour & $ 0.52\pm 0.01$ & $-0.24\pm 0.01$ \\
\osix & $ 0.36\pm 0.02$ & $-0.05\pm 0.01$ \\
\hline
\end{tabular}
\end{center}
\label{tab:delta}
\end{table}
With the above in mind, we can interpret the results of Fig.~\ref{fig:hone}.
Focusing first on the left-hand panel, we find that at fixed \hone, the observed
median \cfour\ optical depths are significantly higher than in the fiducial HM01 model.
The discrepancy increases from $\approx0.1$~dex at $\tau_{\honem}=1$ to $\approx0.5$~dex
at $\tau_{\honem}=10$ and $\approx1$~dex at $\tau_{\honem}=10^2$.
This suggests that at a given
gas overdensity, there is less \cfour\ in the simulations by $\approx0.5$~dex ($\approx1.0$~dex)
for $\tau_{\honem}\lesssim10$ ($\tau_{\honem}\gtrsim10$).
Turning to the different UVB models, while the harder Q-only background
provides a poorer match to the observations,
4Ryd-100 fares much better. Although this model still falls short of the
observed $\tau_{\cfourm}^{\text{med}}$ by about $\approx0.5$~dex
in the highest $\tau_{\honem}$ bin, the softest background is nearly
fully consistent with the observations for $\tau_{\honem}\lesssim10$.
From eq.~\ref{eq:metallicity},
it is apparent that an increase in [C/H] will lead to higher \cfour\ optical depths
at fixed \hone. Therefore, we have run \texttt{SPECWIZARD} with the
elemental abundances scaled linearly by a factor of ten, denoted as ``Zrel-10''.
Such a modification can be partially motivated by the fact that we expect some uncertainty
in the nucleosynthetic yields, of about a factor of two. We have chosen a factor much larger than this
since we find that increasing the metallicities does not scale the median recovered optical depth by the same factor
(contrary to what one might expect from eq.~\ref{eq:metallicity}).
This is because many pixels contributing to the median optical depth are noise dominated
(particularly at low \hone), and some fraction of these pixels have a true optical depth of zero,
so no matter how much the metallicity is increased, the optical depth will never change and exceed the noise.
Although Zrel-10 provides much better agreement between the model and data, even
with such an extreme choice of multiplicative factor it cannot fully account
for the discrepancy between the simulations and data.
This suggests that in the simulation too many \hone\ clouds have negligible metallicity.
To check this, we have calculated the mass
and volume filling factors of the metals using eq.~1 from
\citet{booth12}\footnote{We compute the mass filling factors
using SPH smoothed metallicities, but to avoid smoothing twice,
we compute the volume filling fractions using the particle
metallicities (see \citealt{wiersma09b} for a discussion on the use of
SPH-smoothed versus particle metallicities).}
The authors determined that $>10$\% of the volume and $>50$\% of
the mass need to be enriched to metallicities $Z> 10^{-3}$~Z$_\odot$ to
achieve agreement with observations of \cfour\ in the low-density IGM
at $z\approx3$. For the fiducial L100N1504 box at $z=3.5$ and $Z>0$ we
find volume and mass filling factors of 42\% and 68\%, but for
$Z>10^{-3}$~Z$_\odot$ these are reduced to 10\% and 19\%, respectively.
This suggests that while the fractions of the volume and mass with
non-zero metallicity may be sufficiently high, the metallicities
in the photo-ionised IGM are typically far too low.
Next, we describe another modification to our fiducial model denoted as
``bturb-100'', for which we have considered an unresolved
turbulent broadening term in addition to the
usual thermal broadening. Specifically, we add
$b_{\rm turb} = 100$~\kmps\ in quadrature
to the already included thermal broadening
$b_{\rm therm} (T, m)$ which is calculated for every ion
at each spectral pixel, and
depends on the local temperature $T$ and inversely on atomic mass $m$.
Because of this inverse dependence on atomic mass,
metal ions will be much more strongly affected by the
inclusion of turbulent broadening than \hone.\footnote{
We initially added the thermal broadening term to both \hone\ and metal ions. However,
this led to unphysical values for the flat level of metal ions recovered from
regions bluewards of the \lya\ forest, due to the extreme abundance of \hone\ lines.
Thus, for the model presented in this work we have added turbulent broadening only to
the metals and not to \hone, which
allows us to examine the effect of broadening on \osix, \cthree\ and \sithree\ more clearly.
This is likely also more physical, as the metal-bearing gas may well
be more turbulent than the gas that dominates the neutral hydrogen
absorption \citep[e.g.,][]{theuns02}.
We have confirmed that both methods produce equivalent results
for the \cfour(\hone) and \sifour(\hone) relations, likely because
the metal-line optical depth signal is mostly associated with the strongest and most clustered
hydrogen systems that are not significantly affected by an increased $b$-parameter.
}
Indeed, we find that bturb-100 provides a somewhat better
match to the observed \cfour(\hone) relation.
However, we note that our chosen broadening value $100$~\kmps\ should be considered
a very conservative upper limit, as individual
\cfour\ and \sifour\ components are not usually detected with such large $b$-parameters.
In the centre panel of Fig.~\ref{fig:hone}, we show \sifour(\hone), and find
results that are similar to those for \cfour(\hone).
For $\tau_{\honem}\gtrsim10$ the \sifour\ optical
depths are underestimated by the fiducial HM01 model by a factor
ranging from $\approx0.2$~dex at $\tau_{\honem}=10$
up to $\approx0.8$~dex at $\tau_{\honem}=10^2$.
Invoking 4Ryd-100 leads to near agreement
for all but the highest \hone\ optical depth, while Zrel-10 and bturb-100
also fare markedly better than the fiducial model.
We now consider the \osix(\hone) relation in the right-hand panel of Fig.~\ref{fig:hone}.
While \cfour\ and \sifour\ are expected to mainly probe
cool photoionized gas ($T\sim10^4$~K), \osix\ reaches its peak ionization fraction of 0.2 at
$T=3\times10^5$~K, which is close to the temperatures expected of shocks
associated with accretion events or winds.
Simulations predict that \osix\ around galaxies
is primarily collisionally ionized \citep[e.g.][]{tepper-garcia11, stinson12, ford13, shen13}.
Applying ionization modelling to observations also provides
evidence that \osix\ near moderate to strong \hone\ preferentially probes this
hot gas phase \citep[e.g,][]{simcoe04, aguirre08, danforth08, savage14, turner15}.
Indeed, the results from the right-hand panel of Fig.~\ref{fig:hone} differ considerably
from the previous two relations. Firstly, the simulation realized with the
fiducial model is almost fully consistent with the observations, with any discrepant
points offset by a maximum of 0.2~dex (note the smaller dynamic range
of the y-axis compared to the previous panels).
While
the alternate UVBs (Q-only and 4Ryd-100) have slightly lower $\tau_{\osixm}$ than the fiducial case,
overall we do not find significant differences between these models.
This suggests that in EAGLE the \osix(\hone) relation may be probing a
primarily collisionally ionized gas phase, for which variations in the ionization background
do not have a significant impact on the results.
Furthermore, the fact that the median \osix\ optical depth
appears to be largely insensitive to the addition of turbulent broadening
could indicate that \osix\ is already significantly thermally broadened.
We note that if the pixel optical
depths do not originate predominantly from photoionized gas, then $\tau_{\honem}$ can no
longer be used as a measure of the density.
\begin{figure*}
\includegraphics[width=0.45\textwidth]{{figures/c3_vs_c4_allqso_p50_force_tau_min_comb_leg}.pdf}
\includegraphics[width=0.45\textwidth]{{figures/si3_vs_si4_allqso_p50_force_tau_min_comb}.pdf}\\
\caption{The same as Fig.~\ref{fig:hone}, but showing \cthree(\cfour)
and \sithree(\sifour), and with the data from observations presented in
Table~\ref{tab:data_1}. Unlike for relations binned by \hone,
different ionization states of the same element are not sensitive to
the metallicity of the gas. We find that for \cthree(\cfour), the simulations and data
are in good agreement for the fiducial ionizing background, and the observations
particularly disfavour 4Ryd-100 and Zrel-10. The \sithree(\sifour) relation is somewhat
less constraining, and while the median \sithree\ optical depths from HM01 model
are slightly above the observed values, the discrepancy is no more than 0.1~dex and only
seen in the highest \sifour\ bins.
This indicates that
the temperature and density of the gas probed by pixels with detected \cfour\ and and \sifour\ is
well captured by the simulations, without needing to invoke modifications to the model.}
\label{fig:temp}
\end{figure*}
\begin{table*}
\caption{Observational data from Figs.~\ref{fig:temp} and \ref{fig:rel}.
The format is the same as Table~\ref{tab:data_0}, but here we present
relations binned by either \cfour\ or \sifour\ optical depths. The left column
indicates the central value of the \cfour\ or \sifour\ bin, and the
subsequent columns list the median recovered optical depths for the relation
denoted in the top row. }
\input{tables/data_1.tab}
\label{tab:data_1}
\end{table*}
\subsection{$\tau_Z$ as a function of $\tau_{\rm CIV}$ and $\tau_{\rm SiIV}$}
While metal ions as a function of $\tau_{\honem}$ can probe the metallicity-density relation,
examining different ionization states of a single element can provide insight into
the physical properties of the gas, because the ionization fractions that
set the relative optical depths will only depend on the temperature, the density,
and the UV radiation field (and not on the metallicity, but see below). These optical depth ratios
have previously been used to establish that the gas probed by \cfour\ and \sifour\ is
consistent with being in photoionization equilibrium \citep{schaye03, aguirre04}.
Fig.~\ref{fig:temp} examines \cthree(\cfour) and \sithree(\sifour), and the
observational data is provided in Table~\ref{tab:data_1}.
Looking first at \cthree(\cfour), we find that HM01 is consistent
with all of the \cfour\ bins. Both the bturb-100 and Q-only models
also agree with the data, which is notable in particular for Q-only as it is the most disfavoured by the
\cfour(\hone) relation.
Finally, we find that the 4Ryd-100 and Zrel-10 models fare
particularly poorly in this relation, and produce median \cthree\ optical depths lower than
the observations by up to 0.4~dex. In particular, for Zrel-10 such a discrepancy
may seem surprising, since changing the carbon abundance should not affect
the amount of one ionization state against another. However, the reason this occurs
is because \cfour\ increases more than \cthree, which is
due to the fact that many more \cthree\ pixels are contaminated and hence do not change.
Next, we find the \sithree(\sifour) relation
to be somewhat less constraining. While the fiducial HM01 model demonstrates one of the largest
discrepancies with the data,
the difference is not more than $\approx0.1$~dex when the errors are considered,
and is only seen in the highest \sifour\ bins.
Thus, we find good agreement between the data and the fiducial model for
both relations, which suggests that the temperature and density of the gas probed by
\cfour\ and \sifour\ pixels is consistent between the observations and simulations.
We note that the above result is not in tension with the results shown in Fig.~\ref{fig:hone}. Namely,
while Fig.~\ref{fig:hone} indicates that there is a lack of \cfour\ and \sifour\ in the simulations,
Fig.~\ref{fig:temp} demonstrates that the \sifour\ and \cfour\ that we do find in the mock spectra,
regardless of the amount,
likely resides in gas with similar temperature and density as in the observations.
\begin{figure*}
\includegraphics[width=0.33\textwidth]{{figures/si4_vs_c4_allqso_p50_force_tau_min_comb}.pdf}
\includegraphics[width=0.33\textwidth]{{figures/o6_vs_c4_allqso_p50_force_tau_min_comb_leg}.pdf}
\includegraphics[width=0.33\textwidth]{{figures/o6_vs_si4_allqso_p50_force_tau_min_comb}.pdf}
\caption{The same as Fig.~\ref{fig:hone}, except showing \sifour(\cfour),
\osix(\cfour) and \osix(\sifour) from left to right, which probe relative abundances.
The data from the observations is given in Table~\ref{tab:data_1}.
In the left-hand panel, we find that $\tau_{\sifourm}^{\text{med}}$
is underestimated by all models except bturb-100, and is insensitive to the
choice of UVB.
Next, for \osix(\cfour) and \osix(\sifour) we observe a stronger
sensitivity to different ionizing background models and turbulent broadening
compared to \osix(\hone). For these relations, we
observe a better match between the fiducial and hardest UVB models (HM01 and Q-only),
in tension with the results from \cfour(\hone) and \sifour(\hone) relations,
where we find a strong preference for the softer ionization backgrounds (see Fig.~\ref{fig:hone}).
}
\label{fig:rel}
\end{figure*}
In Fig.~\ref{fig:rel} we examine
relations between different metal ions, which trace relative abundances
and physical conditions. The data for this figure is provided in Table~\ref{tab:data_1}.
For example, Si/C, which can be estimated using the \sifour(\cfour) relation,
has been found to be greater than solar by a
factor of a few in the IGM \citep[e.g.,][]{songaila01, boksenberg03, aguirre04}.
In the left-hand panel of Fig.~\ref{fig:rel}, we plot the median
\sifour\ optical depth against \cfour. While the results are not very
sensitive to the choice of ionizing background, all models except bturb-100
present a paucity of \sifour\ with respect
to the observations. In particular, the median \sifour\ optical depth from the Zrel-10 model
shows an offset of $\leq-1$~dex from the observations at fixed
\cfour. Again, this is due to the fact that for \sifour,
many more pixels that contribute to the median optical depth are noise dominated compared to \cfour,
and therefore do not change when the metallicity is increased.
For the remaining models that use the metallicities directly from the simulations,
this may indicate that at $z\sim3.5$ the simulations have lower [Si/C] than observed.
Additionally, as evidenced by the bturb-100 model, turbulent
broadening, which has the strongest influence on the heavy silicon atoms,
could perhaps be invoked to alleviate this discrepancy.
We briefly draw attention to the bin centred at
$\log_{10}\tau_{\cfourm}=-0.3$, where the observed median \sifour\ optical
depth deviates starkly from the rest of the points. The same behaviour
is also seen in the central panel of Fig.~\ref{fig:rel}, in which
we examine \osix(\cfour). To find the origin of this inconsistency,
we turn to the relations of individual QSOs, in Figs.~\ref{fig:app_si4_vs_c4} and
\ref{fig:app_o6_vs_c4}. In the case of Q1317$-$507 (the upper right-hand panel
of both figures), the median \sifour\ and \osix\ optical depths are
unusually low in this \cfour\ bin, while having relatively
small error bars (the median optical depths of different QSOs are
combined in linear space). We conclude that these points from Q1317$-$507,
likely the result of small number statistics, are responsible for the
anomaly in the \sifour(\hone) and \cfour(\hone) relations.
The centre panel of Fig.~\ref{fig:rel}
shows $\tau_{\osixm}^{\text{med}}$ binned by $\tau_{\cfourm}$.
In contrast to the \osix(\honem) relation (Fig.~\ref{fig:hone}, right-hand panel),
it is apparent that the median \osix(\cfour) optical depth
depends strongly on the choice of UVB, and is sensitive to both an increase in the elemental
abundances and the addition
of turbulent broadening. This is consistent
with the picture that \cfour\ primarily traces photoionized gas,
which will depend on the choice ionizing background, and will not be
as thermally broadened as hot, collisionally ionized gas.
We find that the fiducial HM01 model is in broad agreement
with the data for this relation, even when including the bin centred
at $\log_{10}\tau_{\cfourm}=-0.3$,
while softer UVB models predict too weak \osix\ at high $\tau_{\cfourm}$.
If the \osix\ that is coincident with strong \cfour\ were photoionized,
then this would be a useful constraint. However, unlike \cfour, \osix\ may
well be collisionally ionized.
Finally, in the right-hand panel of Fig.~\ref{fig:rel} we show
\osix(\sifour). Except for $\log_{10}\tau_{\sifourm}\gtrsim-0.5$, we observe a much weaker
dependence on the models than for \osix(\cfour), but we still find that
HM01, in addition to Q-only and Zrel-10, provides the best match to the data.
\subsection{Physical conditions of the gas}
\begin{figure*}
\begin{center}
\includegraphics[width=0.45\textwidth]{{figures/odhist_temp_lin_h1min1.0max10.0_taucutnoise_leg}.pdf}
\includegraphics[width=0.45\textwidth]{{figures/odhist_delta_lin_h1min1.0max10.0_taucutnoise_leg}.pdf}
\end{center}
\caption{PDFs of the optical depth-weighted temperature (left) and overdensity (right), where the line
colour and style indicates the ion used for weighting. The narrower lines represent the full
sample of pixels that have $\tau_{\honem}>10$, and the thicker line PDFs were
calculated using this same sample but with the additional constraint that the metal-line optical
depth had to be $3\sigma$ above the flat level, \taumin. This second cut is used to isolate pixels that
have a significant optical depth detection. The PDF offset to the left of the main distribution denotes
pixels that have zero temperature and density. By definition these pixels have zero metallicity,
but a detected optical depth due to noise or contamination. We find that \cfour\ and \sifour\ probe
a bimodal range of temperatures, while pixels with higher \osix\ optical depths
primarily arise in hot gas.
}
\label{fig:pdf}
\end{figure*}
In this section, we examine the temperatures and densities of the
gas probed by our optical depth relations. In Fig.~\ref{fig:pdf}, we have plotted
the probability density functions (PDFs) of the optical depth-weighted temperatures (left) and
densities (right) from our simulations.
In particular, we are interested
in the physical properties of the regions from which we detect signal in Fig.~\ref{fig:hone}.
We therefore only consider pixels with
$\tau_{\honem}>10$, to focus on areas where
we find the largest discrepancy between
observations and mock spectra in this figure.\footnote{While we have used the
recovered optical depths for the \hone\ and
metal-line optical depth cuts made in Fig.~\ref{fig:pdf}, we note that the results
are unchanged when we use the true optical depths instead.}
The resulting PDFs are shown as the solid lines in Fig.~\ref{fig:pdf}, and demonstrate
high temperatures, which may be surprising for \sifour\ and \cfour, which we
expected to be at least somewhat photoionized.
However, these PDFs are biased to high temperatures because higher temperature gas is
more broadened. For a single absorption line at high temperature, the optical depths will hence
be spread over more pixels, and the individual pixel optical depth values will be lower
than for a lower temperature region.
This effect is further amplified by the fact that the pixel optical depth-weighted temperatures are averages
over the linear (rather than log) temperatures of different gas elements. This means
that gas elements with relatively low ion fractions but high temperatures can affect the
weighted means. Furthermore, at low density the ionization fraction peaks with temperature
become much less prominent if photoionization is included (see e.g. Fig.~1 of \citealt{rahmati16}).
In an effort to combat this bias,
we then make an additional cut, where we only take pixels that have a metal-line optical
depths $3\sigma$ above the corresponding flat level \taumin.
The result of making this additional cut is shown as the dashed line in Fig.~\ref{fig:pdf}.
First, we find that the temperature and density PDF for \osix\ is unchanged,
which indicates that most of the signal for \osix\ truly comes from gas with
higher temperatures (although the distribution is still quite broad, consistent
with \citealt{oppenheimer16}). On the other hand, for \cfour- and \sifour-weighted
quantities this cut reveals a bi-modal temperature distribution, where many of the pixels probe
cooler, $\approx10^{4.5}$~K gas.
Likewise, the \osix-weighted overdensities are not affected by the $3\sigma$ cut,
while for \cfour\ and \cfour\ the overdensity PDFs are still unimodal but have shifted to higher values.
Overall, this figure indicates that while most of the signal for \osix(\hone) comes from
hot ($T\gtrsim10^5$) gas with $\delta\sim1$--$10$, a substantial portion of the pixels that lead to the
\cfour(\hone) and \sifour(\hone) relations arise from cooler, likely photoionized
gas at $\sim10^{4.5}$~K with overdensities $\gtrsim10$.
\section{Discussion}
\label{sec:discussion}
In the previous section, we compared observations of pixel optical depth
relations to the EAGLE simulations. We considered a fiducial QSO+galaxy HM01 UVB \citep{haardt01},
as well a harder QSO-only model, and a softer UVB with
reduced intensity above 4~Ryd by a factor of 100.
For \osix(\hone), we found an insensitivity to the ionizing background model, and
saw good agreement between the simulations and the data. However,
the observed median optical depths from the \cfour(\hone) and \sifour(\hone) relations
were measured to be systematically higher than those derived from the simulations
using the fiducial UVB. The discrepancy is smaller than $\approx0.5$~dex below $\tau_{\honem}=10$
but can reach up to 1~dex for \hone\ bins above this threshold. For \sifour(\hone), invoking
4Ryd-100 fully alleviates the tension, while for \cfour(\hone)
we find this model still falls short of the data, but only for $\tau_{\honem}\gtrsim10$.
We also find that increasing the metallicity by a factor of ten (Zrel-10)
and manually broadening the absorption lines to take unresolved turbulence into consideration (bturb-100)
do not fully resolve the discrepancy.
In this section, we discuss in more detail possible reasons for the observed
mismatch.
Can the discrepancies between the observations and simulations
be attributed to differences in the UVB?
We have indeed found better agreement with the observed \cfour(\hone) and \sifour(\hone)
relations using our softest UVB intended to explore the effects of delayed
\hetwo\ reionization, 4Ryd-100. The reduced
intensity above 4~Ryd disfavours ionization to higher states,
increasing the abundances of \sifour\ and \cfour.
While \citet{haardt01} models take \hetwo\ reionization into account
and predict that the \hethree\ fraction already reaches 50\% at $z\approx6$,
recent studies suggest that the reionization process is patchy,
with \hetwo\ optical depths still high above $z\gtrsim3$
\citep[e.g.,][]{shull10, worseck11}. Thus, the work presented here probes
the epoch where the observed gas may be subject to a strongly fluctuating UVB
above 4~Ryd.
The much better match of the 4Ryd-100 UVB suggests that
\hetwo\ reionization could be complete too early in the simulations.
Turning to other optical depth relations, we find that \cthree(\cfour)
and \sithree(\sifour) do not strongly rule out the 4Ryd-100
model. While these soft UVBs are inconsistent with \osix(\cfour),
the problem occurs only for $\log_{10}\tau_{\cfourm}\gtrsim-0.7$, which
is higher than relevant for Fig.~\ref{fig:hone}.
An alternative effect could be the presence of ionization due
to stellar light from nearby galaxies, which is thought to be important for absorbers as rare as Lyman
limit systems \citep{schaye06, rahmati13b}. The strength of the ionizing radiation
emitted by galaxies drops sharply above 4~Ryd, but could strongly ionize \hone,
lowering the typical optical depths. If \hone\ optical depths are lower,
than at a fixed \hone\ the metal-line optical depths will be higher.
This could explain the larger discrepancy seen at
$\tau_{\honem}\gtrsim10$, where the pixel optical depths are probing denser gas at small
galactocentric distances compared to lower \hone\ optical
depths. However, since it is difficult to estimate the shape
and normalization of this ionizing radiation (and it likely should
not be applied uniformly), we leave testing of this explanation
to a future work.
We have also considered the effects of turbulent broadening.
It is certainly true that our fiducial model misses the unresolved
turbulence from the dense particles
on the imposed EoS (where by fixing the temperatures
to $10^4$~K we neglect a possibly
significant fraction of the energy), and possible
that turbulence in other regions is also underestimated.
Indeed, we find that by artificially broadening metal ion absorption lines, we are able to
bring the \cfour(\hone) and \sifour(\hone)
relations into slightly better agreement with observations. Furthermore, the inclusion of $b_{\rm turb}$
may help alleviate the tension between the observed and simulated \sifour(\cfour). However, we stress that our implementation
provides a very conservative upper limit on this effect, because (a) we use a very high
$b$-parameter (100~\kmps) and (b) we apply the turbulent term to all metal-line absorption pixels,
not just those in very high density regions or with contributions from particles on the EoS.
Another possibility for the observed discrepancy
is that the metallicity of the intergalactic gas in the simulations
is too low. Using our Zrel-10 model, we have examined the effect of increasing
the elemental abundances linearly by a factor of ten. While such a change in metallicity is
larger than the expected metal yield uncertainties,
it is still unable to increase the simulated median \cfour\ and \sifour\ at fixed
\hone\ enough to agree with observations. Furthermore, increasing the
metallicities of both carbon and silicon by equal amounts leads to disagreement
between the simulated and observed \sifour(\cfour) relation, since the optical depths
do not scale directly with metallicity in the same way due to differences in
contamination and noise.
A related issue may be that the that the metals are not transported far enough
into the IGM. An insufficient volume filling fraction of
enriched gas could arise if the simulations
do not resolve the low-mass galaxies thought to be important for metal
pollution \citep[e.g.,][]{wiersma10, booth12}. In Appendix~\ref{sec:restest},
we examined results from simulations with higher resolution than our fiducial model
(the Ref- and Recal-L025N0752 runs). These simulations can resolve galaxies (containing
at least 100 star particles) with $M_{\star}=2.3\times10^7$~\msol,
almost an order of magnitude below that of our fiducial model, where a
100 star particle galaxy would have stellar mass of $1.8\times10^8$~\msol.
Indeed, we find that relations involving \cfour\ are not fully converged
at our fiducial resolution, and invoking the highest-resolution model
for \cfour(\hone) results in an increase in $\tau_{\cfourm}^{\text{med}}$
of up to $\approx0.3$~dex in the highest \hone\ bins.
To investigate the reason for the resolution dependence, we have imposed the same metallicity-density
relation on both Ref-L025N0752 and Ref-L025N0376, finding no significant differences in the optical depth relations.
This implies that the better agreement with observations with increasing resolution is not caused by changes in the
density-temperature structure of the gas. We have also calculated the average metallicites for the various resolutions
in the 25~cMpc volume, and find that the differences are small (varying at most by a factor of 1.2).
Furthermore, we find that the mean gas-particle metallicity is lowest in Recal-L025N0752, and highest in Ref-L025N0188.
Therefore, the increase in \cfour(\hone) with resolution is not due to an increase in the total amount of metals,
but rather to an increase of the metallicity of the IGM. This suggests that winds ejected from galaxies with stellar masses below
$\approx1.8\times10^8$~\msol\ are likely important for IGM pollution.
While a simulation with higher resolution may bring the observations
and simulations closer to agreement, the effect does not appear to be strong enough to
fully explain the differences seen in the \cfour(\hone) relation, and furthermore,
the \sifour(\hone) relation shows almost no change when the resolution is
increased. Therefore, we believe that additional factors may be at play.
An important piece of information to consider is the much better
agreement between the observed and simulated \osix(\hone) relations.
The insensitivity of $\tau_{\osixm}^{\text{med}}$ (when binned by \hone) to the different UVB models
suggests that the gas is primarily collisionally ionized, and hence that \osix(\hone)
is probing a hotter ($T\gtrsim10^5$~K) gas phase than \cfour(\hone) and \sifour(\hone),
as also found for $\tau_{\honem}>10$ by \citet{aguirre08}.
From this, we can conclude that for \textit{hot} gas, the physical properties probed by the
pixel optical depth relations are consistent with observations of
the IGM at $z\sim3.5$. The lack of \cfour\ and \sifour, on the other hand,
may not be due to a too low metallicity or volume filling fraction,
but rather to an incorrect gas phase. If too much of the enriched gas
is excessively hot, then too much carbon and silicon will be ionized to states above
\cfour\ and \sifour, reducing the number of pixels with detectable
\cfour\ and \sifour\ absorption.
\citet{aguirre05} found an even more severe underestimation
of simulated median \cfour\ optical depths, with the tension also
being alleviated by invoking a softer UVB. In contrast to EAGLE,
the simulations in \citet{aguirre05} did not include metal-line cooling,
and that study found that most of the metals resided in an unrealistically-hot gas
phase ($10^5\lesssim T \lesssim 10^7$~K). The authors speculated that the simulations
could be brought into agreement with the observations by implementing
metal-line cooling, but here we have shown that this is not the case.
However, the inclusion of metal-line cooling may have aided in resolving
other issues. While we find good agreement between our observations and simulations
of the \cthree(\cfour) relation, \citet{aguirre05} measured a far too
low $\tau_{\cthreem}^{\text{med}}$, indicating a much stronger mismatch
in the temperature and/or density of the gas in their simulations.
It may be that the temperature of the metal-enriched gas in our
simulations is sensitive to the
details of the stellar feedback.
It is implemented thermally, using a
stochastic prescription in which the temperature of the directly heated
gas is guaranteed to initially exceed $10^{7.5}$~K \citep{dallavecchia12}. The probability
of heating events was calibrated to observations
of galaxy stellar masses and disc sizes at $z\approx0$,
but observations of the CGM were not considered.
In Fig.~\ref{fig:pdf} we established that much of the signal for the
\cfour(\hone) and \sifour(\hone)
relations comes from pixels with temperatures $T~\sim10^{4.5}$~K and
overdensities $\delta\gtrsim10$. This suggests that at $z\sim3.5$, the outflows driven
by stellar feedback may not entrain enough cool ($T\sim10^4$~K) gas.
While \citet{furlong15} found that galaxy star
formation rate densities and stellar masses are in good agreement with observations
for $z\sim3$--$4$, the work presented here suggests that other indicators may be needed to
test fully the feedback implementation at these redshifts.
\section{Conclusion}
\label{sec:conclusion}
In this work we used pixel optical depth relations to study the $z\sim3.5$ IGM,
using new, very high-quality data for a sample of eight $\langle z_{\text{QSO}} \rangle=3.75$ QSOs,
and compared our results with the EAGLE hydrodynamical simulations of galaxy formation.
The QSOs were observed with VLT/UVES, and their spectra all have similar
S/N and coverage. We employed the pixel optical depth technique to obtain
\hone\ and metal-line absorption partially corrected for the effects of
noise, contamination, and saturation. A public version of the code used
can be found at
\url{https://github.com/turnerm/podpy}.
The resulting pixel optical depth
relations were compared to those derived from mock spectra
generated from the EAGLE simulations. The mock spectra were synthesized to have a
resolution, pixel size, S/N and wavelength coverage closely matched to the observations.
We have considered a fiducial QSO+galaxy UVB \citep{haardt01},
as well as a harder QSO-only model and a model for which the intensity above 4~Ryd was reduced by
a factor of 100.
The fiducial EAGLE model was run in a cosmologically
representative box size (100~cMpc) at a relatively high resolution ($2\times1504^3$ particles), and
the feedback from star formation and AGN was calibrated to reproduce the
$z\approx0$ galaxy stellar mass function, galaxy-black hole mass relation,
and galaxy disc sizes.
Our conclusions are listed below.
\begin{itemize}
\item We detect strong correlations for the observed median
\cfour(\hone), \sifour(\hone), \osix(\hone) pixel optical depth relations (Fig.~\ref{fig:hone}), as well as for
\cthree(\cfour) and \sithree(\sifour) (Fig.~\ref{fig:temp}), and for
\sifour(\cfour), \osix(\cfour) and \osix(\sifour) (Fig.~\ref{fig:rel}).
\item We find that for the \cfour(\hone) and
\sifour(\hone) relations, the observed metal-line optical
depths are higher than in the simulations run with the fiducial HM01 UVB.
For \cfour(\hone), we find a discrepancy of up to $\approx0.1$~dex at $\tau_{\honem}=1$,
$\approx0.5$~dex at $\tau_{\honem}=10$, and $\approx1$~dex at $\tau_{\honem}=10^2$,
where we believe we are probing gas at high densities and small galactocentric
distances. For \sifour(\hone), while the agreement is slightly better,
the behaviour is qualitatively similar to that of \cfour(\hone),
and we find that the observed optical depths are higher than seen in the simulations
by up to $\approx0.2$~dex at $\tau_{\honem}=10$
and by up to $\approx0.8$~dex at $\tau_{\honem}=10^2$.
In contrast, \osix(\hone), which probes a hotter gas phase, exhibits
much better agreement (i.e. differences smaller than 0.2~dex) with the data for all \hone\ bins (Fig.~\ref{fig:hone}).
\item We consider UVBs that differ from the fiducial HM01 model, including a harder quasar-only background
(Q-only) and softer backgrounds with 100 times reduced intensity above 4~Ryd (4Ryd-100).
The softer models, which may be more realistic
than our fiducial background if \hetwo\ is still partially ionized at $z\sim3.5$,
are a better match to the \cfour(\hone) and \sifour(\hone) relations,
and can nearly reproduce the observations for $\tau_{\honem}\lesssim10$.
The results of the \osix(\hone) relation are however insensitive to the change
in UVB, which suggests that \osix\ is tracing predominantly collisionally ionized gas (Fig.~\ref{fig:hone}).
\item We also test a model where the elemental abundances are increased
linearly by a factor of 10 (Zrel-10), and a model where the absorption is broadened by
100~\kmps\ (bturb-100). These variations are meant to explore the effects of uncertainties
in the metal yields and of unresolved turbulence, respectively. In both
cases we find that the simulated \cfour(\hone) and \sifour(\hone) relations are in better (but not full)
agreement with the observations, and stress that in we have chosen very aggressive values
for our models such that we would expect the actual effects from uncertainties to be smaller (Fig.~\ref{fig:hone}).
\item Examining relations that investigate different ionization states of the same element, \cthree(\cfour) and
\sithree(\sifour), we find good agreement between the observations and simulations
for the fiducial HM01 model as well as Q-only and bturb-100. However, the observations disfavour
the 4Ryd-100 and Zrel-10 models (Fig.~\ref{fig:temp}).
\item Most models demonstrate a mild paucity of $\tau_{\sifourm}$ in the \sifour(\cfour) relation,
which suggests the simulations may have a slightly too low [Si/C]. The two exceptions are bturb-100, which is in good
agreement with the observations, and Zrel-100, which demonstrates substantially too little \sifour\ at
fixed \cfour\ (left-hand panel of Fig.~\ref{fig:rel}).
\item Unlike \osix(\hone), the \osix(\cfour) and \osix(\sifour) relations
exhibit sensitivity to the UVB for $\tau_{\cfourm}\gg10^{-1}$ and $\tau_{\sifourm}\gtrsim1$,
and we find that \osix(\cfour) is best described by the hardest models
(the fiducial HM01 and Q-only). The dependence on the ionizing background
suggests that strong
\cfour\ and \sifour\ typically probe a cooler ($T\sim10^4$~K), photoionized gas phase compared to the gas traced
by \osix(\hone) (centre and right-hand panels of Fig.~\ref{fig:rel}).
\item We use the simulations to examine the PDFs of the optical depth-weighted temperatures and densities of the pixels
responsible for the high optical depth values in Fig.~\ref{fig:hone}. We find that while
\cfour, \sifour\ and \osix\ all have a component probing hot $T\gtrsim10^5$ gas,
the \cfour\ and \sifour\ optical depths mainly arise from a phase of
cooler ($\sim10^{4.5}$~K) gas with $\delta\gtrsim10$ (Fig.~\ref{fig:pdf}).
\item We discuss possible reasons why \cfour\ and \sifour\ optical depths with
associated \hone\ are underestimated by the fiducial simulations, and we consider that perhaps
a combination of a number of explanations are responsible:
\begin{enumerate}
\item Ionization by local sources, which the simulations do not account for,
may play an important role.
Since the strength of the radiation emitted by stars typically falls sharply above 4~Ryd,
this would ionize \hone\ while having a much weaker effect on the metals, which
would increase the median metal-line absorption for a fixed \hone\ optical
depth (see e.g.\ \S~4.2 and Fig.~9 in \citealt{turner15}). This explanation
is particularly viable for $\tau_{\honem}\gtrsim10$, where we may
be probing small galactocentric distances.
\item The completion of \hetwo\ reionization in the HM01 simulations may
occur too early, or it may be too uniform, since the observations indicate that
it could be quite patchy around $z\sim3.5$ \citep[e.g.,][]{shull10, worseck11}.
This explanation is supported by the better agreement between the
4Ryd-100 model and the \cfour(\hone) and \sifour(\hone) observations.
However, even the 4Ryd-100 model cannot fully explain the \cfour(\hone) observations
for $\tau_{\honem}\gtrsim10$.
\item The magnitude of line-broadening, particularly in dense star-forming regions,
could be underestimated due to unresolved
turbulence in the simulations. While artificially adding a large turbulent broadening
term slightly increases the median \cfour\ and \sifour\ optical depths when binned by \hone,
the effect is not large enough to explain the observed discrepancy.
\item The metallicities may not be high enough in the simulations,
due to uncertainties in the yields. However, even scaling the metallicities
by a factor of 10 is not enough to achieve agreement in the case of \cfour(\hone). Furthermore,
this scaling creates tension in the \sifour(\cfour) relation, due to the fact that the median
recovered optical
\item The simulations may not resolve the low-mass galaxies
required to pollute the diffuse IGM. We find that the highest-resolution
simulations, Ref- and Recal-L025N0752, exhibit superior agreement with the observed \cfour(\hone) relation
by $\approx0.3$~dex at $\tau_{\honem}\approx10^2$. While resolution likely plays a role,
the magnitude of the effect does not appear large enough to explain fully the discrepancy,
particularly for the \sifour(\hone) relation, which we find to be almost insensitive to the
resolution increase.
\item The stellar feedback in the simulations may be driving outflows that
contain insufficient cool gas ($T\sim10^4$~K). The relatively good
agreement between the observed and simulated
\osix(\hone) relation, which probably traces collisionally ionized gas, indicates that
the simulations correctly capture
this hotter gas phase ($T\gtrsim10^5$~K), and that it contains enough metals.
However, if too much of the enriched gas is hot with respect to the observations, then more
\cfour\ and \sifour\ will be ionized to higher energy levels, leading to a paucity of pixels with
detected $\tau_{\cfourm}$ and $\tau_{\sifourm}$.
\end{enumerate}
\end{itemize}
Overall, while the EAGLE simulations qualitatively reproduce all of the pixel optical depth
correlations seen in our sample of QSOs, the mock spectra are found
to have less \cfour\ and \sifour\ at a given density than in the observations.
This suggests that the simulations are still missing one or more important components,
which we have tested in this work:
a more rigorous treatment of \hetwo\ reionization to create a softer UVB,
the resolution required to model turbulence that contribute to line broadening,
and/or higher metallicities and volume filling factors.
However, we do not find that any
of the above models are individually able to match the observations.
While it is possible that the addition of enhanced photoioniziation
of \hone\ by sources close to the absorbers may play an important role, the fact that the simulations agree with
the observed \osix(\hone) relation indicates that the fiducial model is at least able
to capture the hot gas phase correctly.
Therefore, we believe that it is likely that the outflows created by
energetic stellar feedback in the simulations entrain insufficient cool gas.
\section*{Acknowledgements}
This work used the DiRAC Data Centric system at Durham University,
operated by the Institute for Computational Cosmology on behalf of the STFC DiRAC HPC
Facility (www.dirac.ac.uk). This equipment was funded by BIS National E-infrastructure
capital grant ST/K00042X/1, STFC capital grants ST/H008519/1 and ST/K00087X/1, STFC DiRAC
Operations grant ST/K003267/1 and Durham University. DiRAC is part of the National
E-Infrastructure. We also gratefully acknowledge PRACE for
awarding us access to the resource Curie based in France at Tr\`{e}s
Grand Centre de Calcul. This work was sponsored by the Dutch
National Computing Facilities Foundation (NCF) for the use of supercomputer
facilities, with financial support from the Netherlands
Organization for Scientific Research (NWO). The research was supported in part by the
European Research Council under the European Union's Seventh
Framework Programme (FP7/2007-2013)/ERC grant agreement
278594-GasAroundGalaxies and the Interuniversity Attraction
Poles Programme of the Belgian Science Policy Office [AP
P7/08 CHARM]. RAC is a Royal Society URF.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{sec:Intro}
In particle physics models with local super symmetry~(SUSY) or supergravity,
one of the most important predictions
is the existing of gravitino, which is the superpartner of graviton and has a spin 3/2.
The gravitino mass $m_{3/2}$
is related to the energy scales of SUSY breaking,
and can vary from an order of eV up to of TeV.
In particular, scenarios with
light gravitino whose mass is $m_{3/2} \lesssim \mathcal{O}(10)$~eV
are very interesting
because they are free from the cosmological gravitino problem~\cite{Moroi:1993mb},
and can be
consistent with some baryogenesis scenarios
which require a high reheating temperature such as thermal leptogenesis~\cite{Fukugita:1986hr}.
Therefore, the scenarios with the light gravitino
are very attractive in cosmology
It is important to determine the gravitino mass
in order to understand the mechanism of SUSY breaking.
Although it can be probed by collider experiments~(e.g. LHC)~\cite{Hamaguchi:2007ge}
with direct and indirect signatures,
we can also obtain constraints on it
from cosmological observations~\cite{Pierpaoli:1997im}.
In the early Universe,
light gravitinos are produced from thermal plasma,
and they behave as warm dark matter~(WDM) at late epochs.
The light gravitinos influence the growth of density fluctuations
mainly through the following two effects.
First of all, they change the time of matter-radiation equality
because the light gravitinos behave as a radiation component at early epochs.
Secondly, the light gravitinos have large velocity dispersions
and propagate up
to the horizon scales until they become non-relativistic
Then, they erase their own density fluctuation
and suppress the growth of density fluctuations of matter
below the free-streaming scale.
However, the former effect is very small
because the energy density of light gravitinos
does not have a large fraction of the total energy of radiation.
Therefore,
we can probe signatures of the light gravitino mainly
through the latter effect.
From Lyman-$\alpha$ forest data in combination with WMAP \cite{Viel:2005qj},
a constraint on the light gravitino mass is obtained,
and its bound is $m_{3/2}<16$~eV (95\% C.L.).
Additionally, some authors have pointed out that
measurements of CMB lensing~\cite{Ichikawa:2009ir}
or weak lensing surveys of galaxies~\cite{Kamada:2013sya}
are quite effective in constraining the light gravitino mass.
By using the date of CMB lensing from Planck and cosmic shear from the CHFTLenS survey,
a stringent constraint is obtained,
and the upper bound is $m_{3/2}<4.7$~eV (95\% C.L.)~\cite{Osato:2016ixc}.
However,
it is difficult to obtain significant bounds of the light gravitino mass
if we treat the effective number of neutrino species
for light gravitino
$N_{3/2}\equiv \rho_{3/2}/\rho_{\nu}$ as a free parameter,
where $\rho_{3/2}$ and $\rho_{\nu}$
are the energy densities of light relativistic gravitinos and neutrinos, respectively.
Moreover,
discriminating signatures of light gravitino from
that of massive neutrino is also quite difficult.
Therefore, if one wants to discriminate light gravitinos from other possibilities,
it is mandatory to find a new powerful probe of cosmological signatures of the light gravitino.
In this paper, we particularly investigate
the issue of how accurately we can constrain the light gravitino mass
by using future observations of fluctuations of neutral hydrogen 21 cm line
which comes from the epoch of reionization (EoR), in addition to those of CMB.
By observing the power spectrum of cosmological 21 cm line fluctuations,
we will be able to obtain useful information on
a variety of cosmological parameters
\cite{McQuinn:2005hk,Loeb:2008hg,Pritchard:2008wy,
Pritchard:2009zz,Abazajian:2011dt,Oyama:2012tq,Kohri:2013mxa,Kohri:2014hea,Oyama:2015vva,Oyama:2015gma}.
Because observations of the 21 cm line can cover a wide redshift range,
they can be complementary to other observations such as CMB.
Additionally,
the effects of the light gravitino mainly appear on small scales,
which can be well measured by 21 cm observations.
In order to discuss expected constraints
from the future cosmological surveys
on the mass of light gravitino,
we make Fisher analysis by assuming the specifications
for planned observations of 21 cm line
such as Square Kilometre Array (SKA) \cite{SKA} and Omniscope~\cite{Tegmark:2009kv,Tegmark:2008au}.
In our analysis,
we also take into account
future CMB observations
such as the Simons Array~\cite{SimonsArray} and COrE+~\cite{COrE+}.
Besides, we consider
including information of a baryon acoustic oscillation (BAO) observation,
such as Dark Energy Spectroscopic Instrument (DESI) \cite{DESI:web}
and a direct measurement of the Hubble constant $H_0$.
This paper is organized as follows.
In Section~\ref{sec:Light-gravi},
we briefly review the effects of light gravitino on cosmology.
In Section~\ref{sec:methods},
we review analytical methods used in this paper,
paying particular attention to
21 cm line, CMB, BAO observations and the direct measurement of $H_0$.
We show our results in Section~\ref{sec:results},
and Section~\ref{sec:conclusion} is devoted to our conclusion.
\section{Light gravitino and its effects on large-scale structure in the Universe}
\label{sec:Light-gravi}
\subsection{Light gravitino}
\label{subsec:Light-gravi}
The existence of light gravitino is typically predicted in
gauge-mediated SUSY breaking (GMSB)
scenarios \cite{Dine:1981gu, Nappi:1982hm, AlvarezGaume:1981wy,
Dine:1993yw, Dine:1994vc, Dine:1995ag}.
SUSY breaking is the origin of
the masses of the gravitino and SUSY particles.
The SUSY breaking field $S$
has a vacuum expectation value as $\left< S\right> = M + F_S \theta^2$,
and $F_S$ gives SUSY breaking scale,
which is related to the gravitino mass $m_{3/2}$.
The gravitino mass is given by
\begin{align}
m_{3/2} = \frac{F_S}{\sqrt{3} M_{\textrm{pl}}},
\end{align}
where $M_{\textrm{pl}} \simeq 2.4\times 10^{18}$ GeV is the reduced Planck mass.
On the other hand, in GMSB scenarios,
sparticles in the standard model~(SM) sector acquire their masses
through messenger fields, whose mass scale is denoted as $M_{\textrm{mess}}$.
For example, in a GMSB model with
$N$ pairs of messenger particles,
gaugino masses $M_a$
($a=1,2,3$ is a gauge group index)
and sfermion masses squared $m_{\tilde{f}_i}^2$
($i$ is a flavor index) are typically given by
\begin{align}
M_{a} & =
N \left( \frac{\alpha_{a}}{4\pi}\right)
\frac{F_S}{M_{\textrm{mess}}}, \\
m_{\tilde{f}_i}^2 & =
2N \sum_a C_a^{(i)}
\left( \frac{\alpha_{a}}{4\pi}\right)^2
\left(\frac{F_S}{M_{\textrm{mess}}} \right)^2,
\end{align}
where
$C_a^{(i)}$ is Casimir operators for the sfermion $\tilde{f}_i$,
and $\alpha_a$ denotes the gauge coupling constants.
Although
$(F_S/M_{\textrm{mess}})\sim 100$~TeV is required
in order to obtain TeV scale masses,
still the SUSY breaking scale $F_S$ or the gravitino mass $m_{3/2}$
can take a wide range of values as
$\mathcal{O}(1)\textrm{eV}\lesssim m_{3/2}\lesssim \mathcal{O}(10)\textrm{GeV}$,
where the upper bound comes from the requirement that
the gravity-mediation does not dominate,
and existence of the lower bound
arises from avoiding destabilization of the messenger scalar
and not leading to unwanted vacua.
In the GMSB models,
stringent constraints on the gravitino mass
are obtained from the Higgs mass measured by LHC~\cite{Aad:2015zhl,Chatrchyan:2012xdj}.
In the minimal supersymmetric standard model (MSSM),
the stop mass is required to be as large as $\mathcal{O}(10-100)$~TeV
in order to achieve the measured large Higgs mass $m_h=125$~GeV,
and the bound can place a lower bound on the gravitino mass.
In a class of GMSB models with $N$ copies of messenger fields in the
$5+\bar{5}$ representation of SU(5),
we can obtain a bound
$300~\textrm{eV} < m_{3/2}$ with $N=1$, and
$60~\textrm{eV} < m_{3/2}$ with $N=5$~\cite{Ajaib:2012vc},
if the coupling between the messengers and the SUSY breaking field is perturbative.
Although a range of the light gravitino mass
which is allowed by present cosmological observations
are ruled out,
the bound by the LHC is model-dependent.
For example,
$\mathcal{O}(1-10)$~eV may be possible
if the coupling is non-perturbative or
a singlet Higgs is introduced (next to MSSM) \cite{Yanagida:2012ef}.
Additionally,
less stringent lower bounds are obtained
from direct SUSY searches in LHC~\cite{Aad:2014mra}.
Assuming a perturbative coupling,
we can obtain a lower bound
$3.7 \textrm{eV} < m_{3/2}$
in the same GMSB model as mentioned above
with $M_{\textrm{mess}}=250$~TeV and $N=3$ ($10+\bar{10}$ of SU(5)).
\subsection{Effects of light gravitino on the growth of density fluctuations}
\label{subsec:Light-gravi_effects}
If we take account of the presence of the light gravitino in the early Universe,
some difficulties arise in constructing a consistent cosmological scenarios,
which we are going to describe shortly below.
At the reheating era, gravitinos are efficiently produced
and the abundance of them can easily exceed
that of the present dark matter
unless the reheating temperature $T_R$ is very low~\cite{Moroi:1993mb}.
However, many known baryogenesis/leptogenesis scenarios
require high enough reheating temperature,
which may be inconsistent with
the upper bound coming from the observed abundance of dark matter.
For example, thermal leptogenesis scenario~\cite{Fukugita:1986hr}
requires $T_{\rm R}\gtrsim 10^{9}$~GeV
and the reheating temperature
seems to conflict with the gravitino problem
except for the very light gravitino mass range
$m_{3/2}\lesssim 100$~eV.
If gravitinos have such a small mass,
they are thermalized in the early Universe~\cite{Ichikawa:2009ir,Pierpaoli:1997im}.
In that case,
their abundance does not have dependency on
the reheating temperature
and is smaller than the dark matter density
if $m_{3/2}\lesssim 100$~eV.
This advantage is the reason why
we particularly focus on the light gravitino scenario.
Thermally produced gravitinos
decouple from the other particles at some point,
and their relic abundance is fixed.
The number density is determined by
the effective degrees
of freedom
$g_{*3/2}$ at the decoupling of the gravitinos.
In \cite{Ichikawa:2009ir,Pierpaoli:1997im}, the number density of the gravitinos
in GMSB models is evaluated by solving the Boltzmann equation,
and for a messenger mass scale $M_{\textrm{mess}}\sim 100$~TeV,
$g_{*3/2}$ becomes $g_{*3/2}\sim 90$~\cite{Ichikawa:2009ir}
with only mild dependence on $m_{3/2}$.
In consideration of the result,
we set $g_{*3/2} = 90$ as the fiducial value
in our analysis.
The thermally produced light gravitinos
behave as a warm dark matter component~\cite{Ichikawa:2009ir,Pierpaoli:1997im},
and they can be parametrized by their temperature and mass.
Since gravitinos interact with other particles
through their goldstino components in the GMSB scenario,
their phase-space distribution
is a Fermi-Dirac distribution with two degrees of freedom.
Because light gravitinos behave as relativistic particles
at early epoch,
we can parametrize its energy density $\rho_{3/2}$
by using the effective number of neutrino species,
and it is given by
\begin{align}
N_{3/2}
=\frac{\rho_{3/2}}{\rho_{\nu}}=\left(\frac{T_{3/2}}{T_{\nu}}\right)^{4}
=\left(\frac{g_{*\nu}}{g_{*3/2}}\right)^{4/3},
\label{eq:effective_num_gravi}
\end{align}
where $\rho_{\nu}$ is the energy density of one generation of neutrinos ,
and $g_{*3/2}$ and $g_{*\nu}$ are the effective degrees of freedom at decoupling of
light gravitinos and neutrinos, respectively.
In standard cosmology,
the degree of freedom of neutrinos at the neutrino decoupling
is $g_{*\nu}=10.75$.
$T_{3/2}$ and $T_{\nu}$ are temperatures of
light gravitinos and neutrinos, respectively.
From Eq.~\eqref{eq:effective_num_gravi},
the temperature of light gravitino at present is evaluated as
\begin{align}
T_{3/2}=\left( N_{3/2}\right)^{1/4}T_{\nu}=1.95\left( N_{3/2}\right)^{1/4} \textrm{K},
\label{eq:temp_gravi}
\end{align}
where we use the temperature of neutrinos in the standard cosmology.
At late epochs, light gravitinos lose their energy and
become non-relativistic particles due to the cosmic expansion.
Its present density parameter $\Omega_{3/2}h^2$ can be estimated as
\begin{align}
\Omega_{3/2}h^2=0.1269 \left(\frac{m_{3/2}}{100~\textrm{eV}}\right)\left(\frac{90}{g_{*3/2}}\right).
\label{eq:Omega_3_2}
\end{align}
In the following,
we assume that dark matter does not consists solely of the light gravitinos
because the gravitino mass $m_{3/2}$ needs to be about 90 eV in order to
be consistent with observed dark matter density $\Omega_{\textrm{DM}}\simeq 0.11$~\cite{Ade:2015xua},
which contradicts with
the constraint from Ly-$\alpha$ forest, $m_{3/2}<16$~eV~\cite{Viel:2005qj},
as well as a more recent one
$m_{3/2}<4.7$~eV from CMB and cosmic shear data~\cite{Osato:2016ixc}.
Therefore, we assume that dark matter consists of
the light gravitino and CDM components.
i.e. $\Omega_{\textrm{DM}}=\Omega_{\textrm{CDM}}+\Omega_{3/2}$,
and we define the fraction of gravitino
in the total dark matter density as
\begin{align}
f_{3/2}\equiv \frac{\Omega_{3/2}}{\Omega_{\textrm{DM}}}.
\end{align}
As the CDM component,
the QCD axion, a messenger baryon proposed in~\cite{Hamaguchi:2007rb}
and so on can be well-fitted
within the framework of the GMSB scenario.
From now on, we briefly explain effects of the light gravitinos on cosmological structure formation.
They behave as a warm dark matter component,
and their effects on structure formation can be understood by
considering the following two aspects:
(i) a contribution to the energy density of radiation
(ii) suppression of matter fluctuations on small scales through the free-streaming.
The first effect is due to the fact
that light gravitinos behave as relativistic particles at early epochs.
Therefore, the time of matter-radiation equality is slightly delayed
if light gravitinos exist in the Universe.
The delay alters the evolution of gravitational potential
and drives the integrated Sachs-Wolfe~(ISW) effect
in the CMB temperature anisotropy.
In addition, the matter fluctuations are suppressed at small scales
through stagspansion effect due to the delaying of matter-radiation equality.
However, its contribution is so small
(as we mentioned before,
the theoretical calculation predicts
that $g_{*3/2}$ is around 90, which corresponds to $N_{3/2}\simeq0.059$~\cite{Ichikawa:2009ir}),
and it is difficult to measure the impacts due to this effect
by observing
CMB anisotropies without lensing.
Therefore, constraints on the gravitino mass
mainly come from the second effect,
i.e. its free-streaming behavior.
Because light gravitinos have relatively large thermal velocity,
they
propagate up to their free-streaming scale
and erase own density fluctuation
in a similar manner to massive neutrinos.
Within the free-streaming scale,
light gravitinos do not contribute to the gravitational growth of the matter fluctuations.
Thus, matter fluctuations at small scales are suppressed in comparison to
the $\Lambda$CDM model.
Massive neutrinos also have similar effects on the growth of matter fluctuations,
but its temperature and energy density are different from those of light gravitinos.
Therefore, in principle, we can discriminate between
the effects of the light gravitino and the massive neutrino
through observing their free streaming scale.
\section{Forecasting methods}
\label{sec:methods}
\subsection{21 cm line}
Here, we briefly review a forecasting method
related to 21 cm line observations in our analysis.
For further details of the 21 cm line observations,
we refer to Refs.~\cite{Furlanetto:2006jb,Pritchard:2011xb}.
\subsubsection{Power spectrum of 21 cm radiation}
The 21 cm line of neutral hydrogen atom is emitted by transition
between the hyperfine splitting of the 1s ground state.
We can observe signals of 21 cm line which come from the epoch of reionization~(EoR)
or the cosmic dark ages
as the differential brightness temperature
relative to the temperature of CMB $T_{\rm CMB}$:
\begin{align}
\Delta T_{b} \left(\mbox{\boldmath $r$},z \right)
&=&
\frac{3c^{3}hA_{21}}{32\pi k_{B}\nu_{21}^{2}}
\frac
{x_{HI}(\mbox{\boldmath $r$},z)n_{H}(\mbox{\boldmath $r$},z) } {(1+z)H(z)}
\left(
1-\frac{T_{{\rm CMB}}(\mbox{\boldmath $r$},z) }{T_{S}(\mbox{\boldmath $r$},z)}
\right)
\left(
1-\frac{1+z}{H(z)}
\frac{dv_{p\|}(\mbox{\boldmath $r$},z)}{dr_{\|}}
\right),
\label{eq:obsbrightness2}
\end{align}
where
$\mbox{\boldmath $r$}$ is the comoving coordinates of the source of 21 cm line,
$z$ represents the redshift at emission/absorption,
$A_{21} \simeq 2.869 \times 10^{-15}{\rm s^{-1}}$
is the spontaneous decay rate of the hyperfine splitting,
$\nu_{21} \simeq 1.42$ GHz is the frequency of 21 cm line,
$n_{H}$ is the number density of hydrogen,
$x_{HI}$ is the fraction of neutral hydrogen,
and $dv_{p\parallel} / dr_{\parallel}$ is the gradient of peculiar velocity
along the line of sight.
$T_S$ is the spin temperature, which is
defined by $n_1/ n_0 = 3 \exp ( - T_{21} / T_S)$,
where $n_0$ and $n_1$ are the number densities of singlet
and triplet states of neutral hydrogen atom, respectively.
Here $T_{21}= hc / k_B
\lambda_{21}$ is the temperature corresponding to 21 cm line,
and $\lambda_{21}$ is its wavelength.
In this paper, we assume that $T_S \gg T_{\rm CMB}$
because we focus on the epoch of reionization
during which this condition is well satisfied.
In general, the brightness temperature is sensitive
to details of inter-galactic medium (IGM)
and astrophysical processes.
However, with a few reasonable assumptions,
we can eliminate the dependence from the 21 cm line
brightness temperature~\cite{Madau:1996cs,Furlanetto:2006tf,Pritchard:2008da}.
At the epoch of reionization long after star formation begins,
X-ray background produced by early stellar remnants heats the IGM.
Therefore, kinetic temperature of the IGM $T_{K}$ becomes much higher than
that of CMB $T_{\textrm{CMB}}$.
Furthermore, the star formation produces
a large amount of Ly$\alpha$ photons sufficient to couple
$T_{S}$ to $T_{K}$ through the Wouthuysen-Field effect~\cite{Wouthuysen:1952,Field:1958}.
In this scenario,
$T_{\textrm{CMB}} \ll T_{K} \sim T_{S} $ are justified at $z \lesssim 10$,
and $\Delta T_{b}$ does not depend on $T_{S}$.
Now, let us consider fluctuations of
the differential brightness temperature of 21 cm line $\Delta T_b ({\bm r})$.
By expanding the hydrogen number density $n_H$
and the ionization fraction $x_i$ ($x_i=1-x_{HI}$) as
$n_H(\bm{r}) = \bar{n}_H (1 + \delta (\bm{r}))$
and $x_i(\bm{r}) = \bar{x}_i (1 + \delta_x (\bm{r}) )$,
we can rewrite Eq.~\eqref{eq:obsbrightness2} as
\begin{align}
\Delta T_b (\bm{r},z)
=
\Delta \bar{T}_b(z) \left( 1 - \bar{x}_i ( 1+ \delta_x (\bm{r},z) ) \right)
(1+ \delta (\bm{r},z) )
\left(
1-\frac{1+z}{H(z)}
\frac{dv_{p\|}(\mbox{\boldmath $r$},z)}{dr_{\|}}
\right),
\end{align}
where we assume that $T_{\textrm{CMB}} \ll T_S $
and neglect the term including the spin temperature.
Here, $\Delta\bar{T}_b$ is the spatially averaged
differential brightness temperature at redshift $z$ and given by
\begin{align}
\Delta \bar{T}_b(z)
\simeq 26.8
\left(\frac{1-Y_p}{1-0.25} \right)
\left( \frac{\Omega_b h^2}{0.023} \right)
\left( \frac{0.15}{\Omega_m h^2} \frac{1+z}{10} \right)^{1/2}~\textrm{mK},
\end{align}
where $Y_p$ is the primordial $^4$He mass fraction.
By denoting the fluctuation of $\Delta T_b$ as
$\delta (\Delta T_b (\bm{r},z)) \equiv \Delta T_b(\bm{r},z) - \bar{x}_{H}(z)\Delta \bar{T}_b$(z),
the 21 cm line power
spectrum $P_{21} (\bm{k})$ in the $k$-space is defined by
\begin{align}
\left\langle
\delta (\Delta T^\ast_b ({\bm k})) \delta (\Delta T_b (\bm{k}')) \right\rangle
= (2\pi)^3 \delta^3 ( \bm{k-k'}) P_{21} (\bm{k}).
\label{eq:power}
\end{align}
Because the Fourier component of
the peculiar velocity term
$\delta_v \equiv (1+z)(dv_{p\|} / dr_{\|})/H(z)$
is given by
$ \delta_v ({\bm k})= -\mu^2 \delta ({\bm k})$
within the linear perturbation theory
($\mu = \hat{\bm k}\cdot \hat{\bm n}$
is the cosine of the angle between the wave vector and the line of sight),
the power spectrum can be written as
\begin{align}
P_{21} (\bm{k})
=
P_{\mu^0} (k) + \mu^2 P_{\mu^2} (k) + \mu^4 P_{\mu^4} (k),
\end{align}
where $k = |\bm{k}|$ and
\begin{align}
P_{\mu^0} & = \mathcal{P}_{\delta \delta} - 2 \mathcal{P}_{x\delta} + \mathcal{P}_{xx},
\\
P_{\mu^2} & = 2 \left( \mathcal{P}_{\delta \delta} - \mathcal{P}_{x\delta} \right),
\\
P_{\mu^4} & = \mathcal{P}_{\delta \delta}.
\end{align}
Here, $ \mathcal{P}_{\delta \delta} \equiv (\Delta \bar{T}_b)^2 \bar x_{HI}^2P_{\delta \delta},
\mathcal{P}_{x \delta} \equiv (\Delta \bar{T}_b)^2 \bar{x}_i \bar{x}_{HI}
P_{x \delta} $ and $ \mathcal{P}_{x x} \equiv (\Delta \bar{T}_b)^2 \bar{x}_i^2 P_{ x x} $,
where $P_{\delta\delta}, P_{x\delta}$ and $P_{xx}$
are the auto- and cross- power spectra
defined in the same manner as Eq.~\eqref{eq:power}
for the fluctuation of hydrogen number density $\delta$
and that of ionization fraction $\delta_x$.
Since $P_{\delta\delta}$ traces the fluctuation of matter,
the power spectrum of 21 cm line
has information on cosmological parameters.
$P_{x\delta}$ and $P_{xx}$ can be neglected
as long as we consider eras when the IGM is completely neutral.
However, after the reionization starts,
these two spectra significantly contribute
to the 21 cm line power spectrum.
In order to evaluate these spectra,
we adopt the treatment given in Ref.~\cite{Mao:2008ug},
where they assumed that
$\mathcal{P}_{x\delta}$ and $\mathcal{P}_{xx}$ have specific forms
which match simulations incorporating radiative transfer
in Refs.~\cite{McQuinn:2006et,McQuinn:2007dy}.
The explicit forms of the power spectra are parametrized to be
\begin{align}
\mathcal{P}_{x x} (k)
& =
b_{xx}^2 \left[ 1 + \alpha_{xx} (k R_{xx}) + (k R_{xx})^2 \right]^{-\gamma_{xx} / 2}
\mathcal{P}_{\delta\delta} (k), \label{eq:Pxx}
\\
\mathcal{P}_{x \delta} (k)
& =
b_{x\delta}^2 ~e^{ - \alpha_{x\delta} (k R_{x\delta}) - (k R_{x\delta})^2}
\mathcal{P}_{\delta\delta} (k),
\label{eq:Pxdelta}
\end{align}
where $b_{xx}$, $b_{x\delta}$, $\alpha_{xx}$, $\gamma_{xx}$ and $\alpha_{x\delta}$
are parameters which characterize the amplitudes and the shapes of these spectra,
and $R_{xx}$ and $R_{x\delta}$ represent the effective size of ionized bubbles.
In our analysis,
we adopt the values listed in Table~\ref{tab:Pxx_xdelta}
as the fiducial values of these parameters.
\begin{table}[t]
\centering
\begin{tabular}{ccccccccc}
\hline \hline
~~$z$~~ & ~~$\bar{x}_H$~~
& ~~$b_{xx}^2$~~ & ~~$R_{xx}$~~ & ~~$\alpha_{xx}$~~ & ~~$\gamma_{xx}$~~
& ~~$b_{x\delta}^2$~~ & ~~$R_{x\delta}$~~ & ~~$\alpha_{x\delta}$~~ \\
& & & $[{\rm Mpc}]$& & & &$[{\rm Mpc}]$ \\
\hline
$9.2$ & $0.9$ & $0.208$ & $1.24$ & $-1.63$ & $0.38$ & $0.45$ & $0.56$ & $-0.4$ \\
$8.0$ & $0.7$ & $2.12$ & $1.63$ & $-0.1$ & $1.35$ & $1.47$ & $0.62$ & $0.46$ \\
$7.5$ & $0.5$ & $9.9$ & $1.3$ & $1.6$ & $2.3$ & $3.1$ & $0.58$ & $2.0$ \\
$7.0$ & $0.3$ & $77.0$ & $3.0$ & $4.5$ & $2.05$ & $8.2$ & $0.143$ & $28.0$ \\
\hline \hline
\end{tabular}
\caption{Fiducial values of the parameters characterizing $\mathcal{P}_{xx}(k)$
and $\mathcal{P}_{x\delta}(k)$
(See Eqs.~\eqref{eq:Pxx} and \eqref{eq:Pxdelta}) \cite{Mao:2008ug}.
}\label{tab:Pxx_xdelta}
\end{table}
We note that
the power spectrum in the $k$-space $P_{21} (\bm{k})$
are not directly measured by 21 cm line observations.
Instead, the angular location on the sky and the frequency
are measured by an experiment,
and they can be specified by the following vector
\begin{align}
\bm{\Theta} = \theta_x \hat{e}_x + \theta_y \hat{e}_y + \Delta f \hat{e}_z
\equiv \bm{\Theta}_\perp + \Delta f \hat{e}_z,
\end{align}
where $\Delta f$ represents the frequency difference
from the central redshift $z$ of a given redshift bin.
Then, we can define the Fourier dual of $\bm{\Theta}$ as
\begin{align}
{\bm u} \equiv u_x \hat{e}_x + u_y \hat{e}_y + u_\parallel \hat{e}_z
\equiv {\bm u}_\perp + u_\parallel \hat{e}_z.
\end{align}
Notice that $u_\parallel$ has the unit of time
since it is the Fourier dual of $\Delta f$.
With the flat-sky approximation\footnote{
Even if we consider all-sky experiments, the flat-sky approximation
can be valid as long as we analyze
the data in a lot of small patches of the sky \cite{Mao:2008ug}.
},
we can linearize the relations between ${\bm r}$ and ${\bm \Theta}$,
and they are given by
\begin{align}
\bm{\Theta}_\perp = \bm{r}_\perp / d_A (z),
\qquad
\Delta f = \Delta r_\parallel / y(z),
\end{align}
where $\bm{r}_\perp$ is the vector perpendicular to the line of sight,
$\Delta r_\parallel$ is the comoving distance interval
corresponding to the frequency intervals $\Delta f$,
$d_A (z)$ is the comoving angular diameter distance,
and $y(z) \equiv \lambda_{21} (1+z)^2 / H(z)$.
Then, the relations between ${\bm k}$ and ${\bm u}$
can be written as
\begin{align}
\bm{u}_\perp = d_A \bm{k}_\perp,
\qquad
u_\parallel = y k_\parallel.
\end{align}
The power spectrum of $\Delta T_b$ in the $u$-space can be defined
in the same manner as the treatment in the $k$-space,
and the spectra are related each other by
\begin{align}
P_{21} (\bm{u}) = \frac{1}{d_A(z)^2 y (z)} P_{21} (\bm{k}).
\end{align}
We use the $u$-space power spectrum in the following analysis
because this quantity is directly measurable
without assuming cosmological parameters.
\subsubsection{Fisher matrix of 21 cm line observation}\label{subsubsec:Forecast}
In order to estimate errors of cosmological parameters,
we use the Fisher matrix analysis \cite{Tegmark:1996bz}.
The Fisher matrix of 21 cm line observations
is given by \cite{McQuinn:2005hk}
\begin{align}
F^{({\rm 21cm})}_{\alpha\beta} =
\sum_{\rm pixels}
\frac{1}{[ \delta P_{21}(\bm{u}) ]^2}
\left( \frac{\partial P_{21} (\bm{u})}{\partial \theta_{\alpha}} \right)
\left( \frac{\partial P_{21} (\bm{u})}{\partial \theta_{\beta}} \right),
\label{eq:Fisher_21}
\end{align}
where $\delta P_{21}(\bm{u})$ is the error
in the power spectrum measurements for a Fourier pixel $\bm{u}$,
and $\theta_{\alpha}$ represents a cosmological parameter with its index "$\alpha$".
The 1 $\sigma$ error of the parameter $\theta_{\alpha}$
is evaluated by the Fisher matrix,
and we can obtain the estimated error from
\begin{align}
\Delta \theta_{\alpha} \geq \sqrt{(F^{-1})_{\alpha\alpha}}.
\end{align}
When we differentiate $P_{21}({\bm u})$ with respect to
the cosmological parameters,
we fix ${\cal P}_{\delta \delta}(k)$
in Eqs.~(\ref{eq:Pxx})~and~(\ref{eq:Pxdelta})
so that we get conservative evaluations for
errors of cosmological parameters.
Then,
the information of the matter distribution only
comes from the ${\cal P}_{\delta \delta}(k)$
terms in $P_{\mu^0},P_{\mu^2}$ and $P_{\mu^4}$.
The error of the power spectrum
$\delta P_{21}({\bm u})$ consists of sample variances and experimental noise
\begin{align}
\delta P_{21}(\bm{u}) = \frac{ P_{21}(\bm{u}) + P_N (u_{\perp}) }{N_c^{1/2} },
\label{eq:variance_21cm}
\end{align}
where the first term on the right hand side
represents the sample variance,
and the second term gives contribution of experimental noises.
Here, $N_c = 2 \pi k_\perp \Delta k_\perp \Delta k_\parallel V(z) /(2\pi)^3$
is the number of independent cells in an annulus summing over the azimuthal angle,
$V(z) = d_A(z)^2 y(z) B \times {\rm FoV} $ is the survey volume,
where $B$ is the bandwidth,
and
FoV $\propto \lambda^{2}$ is the field of view of an interferometer.
\subsubsection{Specifications of experiments}\label{sec:21cm_spec}
\begin{table}
\centering
\begin{tabular}{c|ccccccc}
\hline \hline
& $N_{\rm ant}$ & $A_e~(z=8)$ & $L_{\rm min} $ & $ L_{\rm max}$ & FoV $(z=8)$ & Obs. time $t_0$ & z
\\
& & $[{\rm m}^2]$ & $[{\rm m}]$ & $[{\rm km}]$ & $[{\rm deg}^2]$ & [hour] &
\\
\hline
SKA phase 1 & 911/2 & 443 & 35 & 6 & 13.12 & 1000 & 6.8 -- 10
\\
SKA phase 2 & 911$\times$ 4 & 443 & 35 & 6 & 13.12 $\times$ 4 $\times$ 4 & 1000 & 6.8 -- 10
\\
Omniscope & $10^6$ & 1 & 1 & 1 & 2.063 $\times 10^4$ & 1000 & 6.8 -- 10
\\
\hline
\hline
\end{tabular}
\caption{Specifications of 21 cm observations adopted in the analysis. }
\label{tab:21obs}
\end{table}
Here, we show the specifications
of the 21 cm line observations which is considered in this paper.
\paragraph{Survey range:}
In our analyses, we assume that the redshift range
used for constraining cosmological parameters
is $ z = 6.75 - 10.05$,
which we divide into 4 bins: $z = 6.75 - 7.25, 7.25 - 7.75, 7.75-8.25$
and $8.25 - 10.05$.
For surveyed scales (wave number),
we set a minimum cut off $k_{{\rm min} \parallel} = 2 \pi / (y B) $
in order to avoid foreground contaminations \cite{McQuinn:2005hk},
and take a maximum value $ k_{\rm max} = 2~{\rm Mpc}^{-1}$
in order not to be affected by nonlinear effects
of matter fluctuations,
which becomes important on $k_{\rm max} \le k$.
For methods of foreground removals, see also recent discussions about
the independent component analysis (ICA) algorithm
(FastICA)~\cite{Chapman:2012yj}, which will be developed in terms of the
ongoing LOFAR observation~\cite{LOFAR}.
\paragraph{Noise power spectrum:}
The noise power spectrum, $P_N (u_{\perp})$ appeared
in Eq.~\eqref{eq:variance_21cm} is given by
\begin{align}
P_N (u_{\perp})
= \left( \frac{\lambda^{2} (z) T_{\rm sys} (z) }{A_e (z)} \right)^2
\frac{1}{t_0 n(u_\perp)},
\end{align}
where $\lambda$ is an observed wave length of redshifted 21 cm line,
$A_e$ is the effective collecting area per a antenna tile or a station,
$t_0$ is the observation time,
$n(u_{\perp})$ is the number density of baseline,
and
the system temperature $T_{\rm sys}$ is estimated to be
$T_{\rm sys} = T_{{\rm sky}} + T_{{\rm rcvr}}$
and is dominated by the sky temperature due to synchrotron radiation.
Here, $T_{{\rm sky}} = 60 (\lambda/[m])^{2.55} $ [K]
is the sky temperature,
and $T_{{\rm rcvr}} = 0.1T_{{\rm sky}} + 40 $[K]
is the receiver noise \cite{SKA}.
The effective collecting area is proportional to
the square of the observed wave length $A_e \propto \lambda^{2} $.
The number density of the baseline $n(u_\perp)$
depends on an antenna distribution.
As future observations of 21cm line fluctuations, in this paper
we consider SKA (phase~1 and phase~2) \cite{SKA,Mellema:2012ht}
and Omniscope \cite{Tegmark:2009kv,Tegmark:2008au},
whose specifications are shown in Table~\ref{tab:21obs}.
In order to estimate the number density of the baseline $n(u_\perp)$,
we assume a realization of antenna distributions for these arrays as follows.
The total collecting area of SKA phase~1 (SKA1)
is one-half as large as that of the originally planned SKA1.
Therefore, for SKA1,
we assume that the number of antenna station $N_{\textrm{ant}}$ is half as many as
that of the originally planned SKA1, which has 911 antenna stations,
and for SKA phase~2 (SKA2),
the number of antenna stations is 4 times as many as that of the originally planned SKA1.
The number density of baselines of the originally planned SKA1
is determined as follows.
We take the antenna stations
in a core region with a radius 3000 m,
which consists of 95\% of the total.
and the distribution has an antenna station density profile
of the originally planned SKA1 $\rho_{{\rm origSKA1}}(r)$
($r$: a radius from center of the array) as follows \cite{Kohri:2013mxa},
\begin{align}
\rho_{{\rm origSKA1}}(r) =
\left\{
\begin{array}{lll}
\rho_{0}r^{-1}, &\rho_{0} \equiv \frac{13}{16\pi\left(\sqrt{10}-1\right) } \ {\rm m}^{-2}
& \hspace{60pt} r \leq 400 \ {\rm m},\\
\rho_{1}r^{-3/2}, &\rho_{1} \equiv \rho_{0} \times 400^{1/2}, & \ \ \ 400 \ {\rm m} < r \leq 1000 \ {\rm m}, \\
\rho_{2}r^{-7/2}, &\rho_{2} \equiv \rho_{1} \times 1000^{2}, & \ \ 1000 \ {\rm m} < r \leq 1500 \ {\rm m}, \\
\rho_{3}r^{-9/2}, &\rho_{3} \equiv \rho_{2} \times 1500 , & \ \ 1500 \ {\rm m} < r \leq 2000 \ {\rm m}, \\
\rho_{4}r^{-17/2},&\rho_{4} \equiv \rho_{3} \times 2000^{4}, & \ \ 2000 \ {\rm m} < r \leq 3000 \ {\rm m}. \\
\end{array}
\right.
\end{align}
Here, we assume an azimuthally symmetric distribution of the antenna stations in SKA.
In this analysis, we ignore measurements from the sparse distribution of
the remaining 5\% of the total antenna stations which are outside the core region.
This distribution agrees with the specification of
the originally planned SKA1 baseline design.
Then we can evaluate the number density of baselines
of the originally planned SKA1 $n_{{\rm origSKA1}}(u_\perp)$
from this distribution.
Using the number density of baselines,
we can estimate that the number densities of baselines
of SKA1~(re-baseline design) and SKA2 are
\begin{align}
n_{{\rm SKA1}}(u_\perp)
&= n_{{\rm origSKA1}}(u_\perp) \times \left( \frac{1}{2} \right)^{2}, \\
n_{{\rm SKA2}}(u_\perp)
&= n_{{\rm origSKA1}}(u_\perp) \times 4^{2},
\end{align}
where $n_{{\rm SKA1}}(u_\perp)$
and $n_{{\rm SKA2}}(u_\perp)$
are the number densities of baseline of SKA1
or SKA2, respectively.
For Omniscope, which is a future square-kilometre collecting area array
optimized for 21 cm tomography, we take all of antenna tiles distributed
with a filled nucleus in the same manner as Ref. \cite{Mao:2008ug}.
\subsection{CMB}
In our analysis,
we focus on not only 21cm line observations
but also CMB observations,
especially gravitational lensing of CMB,
which has information on matter fluctuations at late times.
Although the 21 cm line observations are
powerful probes of the matter power spectrum,
particularly, on small scales,
the CMB observations greatly help to determine
other cosmological parameters
such as energy densities of the dark matter,
baryons and the dark energy.
Besides, CMB power spectra are sensitive to gravitino mass
through the CMB lensing.
Future precise CMB experiments are expected to
set stringent constraints on the light gravitino mass~\cite{Ichikawa:2009ir}.
Therefore, we take account of
combining the CMB experiments
with the 21 cm line observations.
\subsubsection{Fisher matrix of CMB}
We evaluate errors of cosmological parameters
by using the Fisher matrix of CMB,
which is given by~\cite{Tegmark:1996bz}
\begin{align}
F_{\alpha \beta}^{\rm (CMB)}
= \sum_{\ell}
\frac{\left( 2\ell+1\right)}{2}
\mathrm{Tr}
\left[
C_{\ell}^{-1}
\frac{\partial C_{\ell}}{\partial \theta_{\alpha}}
C_{\ell}^{-1}
\frac{\partial C_{\ell}}{\partial \theta_{\beta}}
\right],
\label{eq:Fisher_CMB}
\end{align}
\begin{align}
C_{\ell} = \left(
\begin{array}{ccc}
\hspace{5pt} C_{\ell}^{\mathrm{TT}}
+ N_{\ell}^{\mathrm{TT}} \hspace{5pt} &
\hspace{5pt} C_{\ell}^{\mathrm{TE}} \hspace{5pt} &
\hspace{5pt} C_{\ell}^{\mathrm{Td}} \hspace{5pt} \\
\hspace{5pt} C_{\ell}^{\mathrm{TE}} \hspace{5pt} &
\hspace{5pt} C_{\ell}^{\mathrm{EE}}
+N_{\ell}^{\mathrm{EE}} \hspace{5pt} &
0 \hspace{5pt} \\
\hspace{5pt} C_{\ell}^{\mathrm{Td}} \hspace{5pt} &
0 \hspace{5pt} &
\hspace{5pt} C_{\ell}^{\mathrm{dd}}
+ N_{\ell}^{\mathrm{dd}} \hspace{5pt}
\end{array}
\right),
\end{align}
where $\ell$ is the multipole of angular power spectra,
$C_{\ell}^{\mathrm{X}}
\left(\mathrm{X}=\mathrm{TT, EE, TE} \right)$
are the CMB power spectra,
$C_{\ell}^{\mathrm{dd}}$ is the deflection angle spectrum,
$C_{\ell}^{\mathrm{Td}}$ is the
cross correlation between the deflection angle and the temperature,
$N_{\ell}^{\mathrm{X'}}$
$\left(\mathrm{X'}=\mathrm{TT, EE} \right)$
and $N_{\ell}^{\mathrm{dd}}$
are the noise power spectra,
where $C_{\ell}^{\mathrm{dd}}$ is calculated by a lensing
potential~\cite{Okamoto:2003zw} and is related with
the lensed CMB power spectra.
The noise power spectra of CMB
$N_{\ell}^{\mathrm{X'}}$ are expressed with a beam size
$\sigma_{\mathrm{beam}}(\nu)=$ $\theta
_{\mathrm{FWHM}}(\nu)/\sqrt{8\ln 2}$,
where
$\sqrt{8\ln 2}\sigma_{{\rm beam}}$ means
the full width at half maximum of the Gaussian distribution,
and instrumental sensitivity
$\Delta _{\mathrm{X'}}(\nu)$ by
\begin{align}
N_{\ell}^{\mathrm{X'}}
= \left[
\sum_{i} \frac1{n_{\ell}^{\mathrm{X'}}(\nu_{i})}
\right]^{-1},
\label{eq:noise_CMB}
\end{align}
where $\nu_{i}$ is an observing frequency and
\begin{align}
n_{\ell}^{\mathrm{X'}}(\nu)=
\Delta ^2_{\mathrm{X'}}(\nu)\exp\left[\ell(\ell+1) \sigma ^2_{\mathrm{beam}}(\nu)\right].
\end{align}
The noise power spectrum of deflection angle
$N^{dd}_l$ is obtained assuming lensing reconstruction with the
quadratic estimator \cite{Okamoto:2003zw},
which is computed with FUTURCMB \cite{paper:FUTURCMB}.
In this algorithm,
$N_{\ell}^{dd}$ is estimated from
the noise $N_{\ell}^{X'}$,
and
lensed and unlensed power spectra of CMB temperature,
E-mode and B-mode polarizations.
Finally, the Fisher matrix in Eq.\eqref{eq:Fisher_CMB} is modified as follows
by taking into account the multipole range
[$\ell_{min}$, $\ell_{max}$]
and the fraction of the observed sky $f_{{\rm sky}}$,
\begin{equation}
F_{\alpha \beta}^{\rm (CMB)}
= \sum_{\ell = \ell_{min}}^{\ell_{max}}
\frac{\left( 2\ell+1\right)}{2}f_{\mathrm{sky}}
\mathrm{Tr}
\left[
C_{\ell}^{-1}
\frac{\partial C_{\ell}}{\partial \theta_{\alpha}}
C_{\ell}^{-1}
\frac{\partial C_{\ell}}{\partial \theta_{\beta}}
\right].
\end{equation}
\subsubsection{Specifications of the experiments}
Now, we show the specifications of
the CMB observations
which are considered in this paper.
In order to obtain the future constraints,
we consider Planck~\cite{Planck:2006aa},
the Simons Array\cite{SimonsArray}, which we occasionally abbreviate to SA in this paper,
and COrE+ \cite{COrE+}
whose experimental specifications
are summarized in Tables~\ref{tab:Planck_SA_spec} and \ref{tab:core_spec}.
The Simons Array is a near future
ground-based precise CMB polarization observation
and COrE+ is a planned satellite observing CMB.
When we combine observations of Planck and the Simons Array, we
evaluate the noise power spectra
$N_{\ell}^{\mathrm{X,Planck+SA}}$ of
the CMB polarization ($X = {\rm EE}$ or ${\rm BB}$)
with the following operation:
\begin{description}
\item[(1)] $2\leq \ell < 25$
\begin{align}
N_{\ell}^{\mathrm{X,Planck + SA}}
= N_{\ell}^{\mathrm{X,Planck}},
\end{align}
\item[(2)] $25\leq \ell \leq 3000$
\begin{align}
N_{\ell}^{\mathrm{X,Planck + SA}}
= [1/N_{\ell}^{\mathrm{X,Planck}}
+ 1/N_{\ell}^{\mathrm{X,SA}}]^{-1}.
\end{align}
\end{description}
Since we assume that
the CMB temperature fluctuation observed by the Simons Array
is not used for constraints on the cosmological parameters,
the temperature noise power spectrum
$N_{\ell}^{\mathrm{TT,\,Planck+SA}}$ is equal to
{\bf $N_{\ell}^{\mathrm{TT,\,Planck}}$}.
The reason for this
is that the CMB temperature fluctuation
observed by Planck reaches almost cosmic variance limit.
Therefore, the constraints are not significantly improved
if we include the CMB temperature fluctuation
observed by the Simons Array.
\begin{table}[tp]
\begin{center}
\begin{tabular}{c|ccccccc}
\hline
\hline
\shortstack{Experiment\\ \,}&
\shortstack{$\nu$ \\ $[\mathrm{GHz}]$}&
\shortstack{$\Delta _{\mathrm{TT}}$\\ $[\mathrm{\mu K\textrm{arcmin}}]$}&
\shortstack{$\Delta _{\mathrm{PP}}$\\ $[\mathrm{\mu K\textrm{arcmin}}]$}&
\shortstack{$\theta_{\mathrm{FWHM}}$\\ $[\mathrm{\textrm{arcmin}}]$}&
\shortstack{$f_{\mathrm{sky}}$\\ $ $} &
\shortstack{$\ell_{{\rm min}}$\\ $ $} &
\shortstack{$\ell_{{\rm max}}$\\ $ $}
\\
\hline
\hline
& 30 & 145 & 205 & 33 & & & \\
& 44 & 150 & 212 & 23 & & & \\
& 70 & 137 & 195 & 14 & & & \\
Planck & 100& 64.6& 104 & 9.5 & 0.65 & 2 & 3000 \\
& 143& 42.6& 80.9 & 7.1 & & & \\
& 217& 65.5& 134 & 5 & & & \\
& 353& 406 & 406 & 5 & & & \\
\hline
\hline
& 95 & - & 13.9 & 5.2 & & & \\
Simons Array & 150 & - & 11.4 & 3.5 & 0.65 & 25 & 3000 \\
& 220 & - & 38.8 & 2.7 & & & \\
\hline
\hline
\end{tabular}
\caption{
Experimental specifications of Planck and the Simons Array
assumed in our analysis~\cite{Oyama:2015gma}.
Here $\nu$ is the observation frequency,
$\Delta_{\rm TT}$ is the temperature sensitivity per $1'\times1'$
pixel, $\Delta_{\rm PP}=\Delta_{\rm EE}=\Delta_{\rm BB}$
is the polarization (E-mode and B-mode) sensitivity
per $1'\times1'$ pixel,
$\theta_{\rm FWHM}$ is the angular resolution defined to be the full width at
half-maximum, and $f_{\rm sky}$ is the observed fraction of the sky.
For the Planck experiment, we assume that
the three frequency bands ($70, 100, 143$ GHz)
are only used for the observation of CMB.
For the Simons Array, we assume that
the 95 and 150 GHz bands are used for observation of CMB,
and the 220 GHz band is not used for constraining cosmological parameters
because the band is found to be useful for the foreground removal~\cite{Oyama:2015vva,Oyama:2015gma}.
}
\label{tab:Planck_SA_spec}
\end{center}
\end{table}
\begin{table}
\begin{tabular}{c|ccccccc}
\hline
\hline
\shortstack{Experiment\\ \,}&
\shortstack{$\nu$ \\ $[\mathrm{GHz}]$}&
\shortstack{$\Delta _{\mathrm{TT}}$\\ $[\mathrm{\mu K\textrm{arcmin}}]$}&
\shortstack{$\Delta _{\mathrm{PP}}$\\ $[\mathrm{\mu K\textrm{arcmin}}]$}&
\shortstack{$\theta_{\mathrm{FWHM}}$\\ $[\mathrm{\textrm{arcmin}}]$}&
\shortstack{$f_{\mathrm{sky}}$\\ $ $} &
\shortstack{$\ell_{{\rm min}}$\\ $ $} &
\shortstack{$\ell_{{\rm max}}$\\ $ $}
\\
\hline
\hline
&45 & 5.2 & 9.0 & 23.3 & & & \\
&75 & 2.7 & 4.7 & 14 & & & \\
&105 & 2.7 & 4.6 & 10 & & & \\
&135 & 2.6 & 4.5 & 7.7 & & & \\
&165 & 2.6 & 4.6 & 6.4 & & & \\
&195 & 2.6 & 4.5 & 5.4 & & & \\
&225 & 2.6 & 4.5 & 4.7 & & & \\
COrE+&255 & 6.0 & 10.4& 4.1 & 0.65 & 2 & 3000 \\
&285 & 10.0& 17 & 3.7 & & & \\
&315 & 26.6& 46 & 3.3 & & & \\
&375 & 67.8& 117 & 2.8 & & & \\
&435 & 147.6&255 & 2.4 & & & \\
&555 & 218 & 589 & 1.9 & & & \\
&675 & 1268 & 3420 & 1.6 & & & \\
&795 & 7744 & 20881 &1.3 & & & \\
\hline \hline
\end{tabular}
\centering
\caption{Experimental specifications of COrE+ adopted in our analysis.
Because specifications of COrE+ are in the planning stage,
we use the values appeared in Ref.~\cite{COrE+},
which are original specifications of COrE.
In the same manner as Ref.~\cite{COrE+},
we assume that the CMB channels are 75, 105, 135, 165, 195 and 225~GHz.}
\label{tab:core_spec}
\end{table}
\subsection{BAO}
In our analysis, we take account of joint constraints from
CMB, 21cm line, baryon acoustic oscillation (BAO).
Therefore, before discussing future constraints,
we briefly summarize formalisms of our analysis methods related to
BAO, here.
\subsubsection[Fisher matrix of BAO]{Fisher matrix of BAO }
\label{subsubsec:BAOFisher}
For estimating future constrains,
we use the following Fisher matrix of BAO data.
The following method is based on
\cite{Albrecht:2006um} and \cite{Wu:2014hta}.
The observables of BAO are the comoving angular diameter
distance $d_{A}(z)$ and the Hubble parameter $H(z)$
(and more specifically, $\ln(d_{A}(z))$ and $\ln(H(z))$ are the observables).
Therefore, the Fisher matrix of BAO data is written as
\begin{align}
F^{(\textrm{BAO}) \ d,H}_{\alpha \beta} &=
\sum_{i} \frac{1}{\sigma_{d,H}^2(z_{i})}
\frac{\partial f_{i}^{d,H}}{\partial \theta_{\alpha}}
\frac{\partial f_{i}^{d,H}}{\partial \theta_{\beta}}, \\
f_{i}^{d} &= \ln(d_{A}(z_{i})), \\
f_{i}^{H} &= \ln(H(z_{i})),
\end{align}
where $i$ is the index of each redshift bin,
$\sigma_{d}(z_{i})$ and $\sigma_{H}(z_{i})$
are variances of $\ln(d_{A}(z_{i}))$
and $\ln(H(z_{i}))$, respectively.
We assume that an observed redshift range is divided into bins,
with the width and central redshift value of each bin respectively denoted as $\Delta z_i$ and $z_i$.
Note that cosmological parameters related to BAO data
are only $(\Omega_{m}h^2, \Omega_{\Lambda})$
or $(h, \Omega_{\Lambda})$
when we assume that the Universe is flat.
\subsubsection{Specifications of BAO data and the direct measurement of the Hubble constant}
\begin{table}
\begin{tabular}{c|c|c}
\hline \hline
Central redshift $z_{i}$ & $\sigma_{d}(z_{i}) \times 10^{2}$ & $\sigma_{H}(z_{i})\times 10^{2}$ \\
\hline
0.15 & 2.78 & 5.34 \\
0.25 & 1.87 & 3.51 \\
0.35 & 1.45 & 2.69 \\
0.45 & 1.19 & 2.20 \\
0.55 & 1.01 & 1.85 \\
0.65 & 0.87 & 1.60 \\
0.75 & 0.77 & 1.41 \\
0.85 & 0.76 & 1.35 \\
0.95 & 0.88 & 1.42 \\
1.05 & 0.91 & 1.41 \\
1.15 & 0.91 & 1.38 \\
1.25 & 0.91 & 1.36 \\
1.35 & 1.00 & 1.46 \\
1.45 & 1.17 & 1.66 \\
1.55 & 1.50 & 2.04 \\
1.65 & 2.36 & 3.15 \\
1.75 & 3.62 & 4.87 \\
1.85 & 4.79 & 6.55 \\
\hline \hline
\end{tabular}
\centering
\caption{Specification of DESI (14000 [${\rm deg}^2$]) adopted in the analysis.
These values are the same as those in Ref.\cite{Font-Ribera:2013rwa}. }
\label{tab:DESI_spec}
\end{table}
In this paper, we focus on the
Dark Energy Spectroscopic Instrument (DESI) \cite{DESI:web,Font-Ribera:2013rwa},
which is a future large volume galaxy survey.
The survey redshift range is $0.1<z<1.9$
(we do not include the Ly-$\alpha$ forest at $1.9<z$ for simplicity),
where we assume that the redshift range is divided into 18 bins,
in other words $\Delta z_{i}=0.1$,
and the observed solid angle is
$14000$~$[{\rm deg}^2]$.
We summarize the specifications of DESI in
Table.~\ref{tab:DESI_spec} \cite{Font-Ribera:2013rwa}.
Additionally, in the same manner as \cite{Wu:2014hta},
when we combine BAO with the other observations,
we add a 1\% $H_{0}$ prior,
which would be achievable by a direct measurements of the Hubble constant
in the next decade.
The Fisher matrix of the direct measurement of $H_0$ is expressed as
\begin{align}
F^{(H_{0})}_{\alpha \beta}
= \left\{
\begin{array}{c}
\hspace{-40pt} \frac{1}{(1\%\times H_{0,{\rm fid}})^2},
\hspace{30pt} \theta_{\alpha}=\theta_{\beta}=H_{0},\\
\hspace{20pt}0,
\hspace{60pt} {\rm the \ other \ components},
\end{array}
\right.
\end{align}
where $H_{0,{\rm fid}}$ is the fiducial value of $H_{0}$.
If we choose the Hubble parameter as
a dependent parameter,
it is necessary to translate the Fisher matrix
into that of a chosen parameter space.
Under the translation of
$(h, \Omega_{\Lambda})$ $\longrightarrow$
$(\Omega_{m}h^2, \Omega_{\Lambda})$,
the Fisher matrix in the new parameter space is written as
\begin{align}
\tilde{F}^{H_{0}} =
\left(
\begin{array}{cc}
\tilde{F}_{\Omega_{m}h^2 \Omega_{m}h^2} &
\tilde{F}_{\Omega_{m}h^2 \Omega_{\Lambda}} \\
\tilde{F}_{\Omega_{m}h^2 \Omega_{\Lambda}} &
\tilde{F}_{\Omega_{\Lambda} \Omega_{\Lambda}}
\end{array}
\right)
=
\frac{1}{(1\%\times H_{0,{\rm fid}})^2}
\left(\frac{1}{2\Omega_{m}h^2}\right)^2
\left(
\begin{array}{cc}
h^2 & h^4 \\
h^4 & h^6
\end{array}
\right).
\end{align}
\section{Results}
\label{sec:results}
In this section, we present our results for
projected constraints
by 21cm line~(SKA phase~1, phase~2, or Omniscope),
CMB~(Planck + Simons Array (SA) or COrE+),
BAO~(DESI) and a direct measurement of the Hubble constant
on cosmological parameters,
paying particular attention to
parameters related to the light gravitino,
i.e.
the fraction of light gravitinos
in the total dark matter density $f_{3/2}$,
and the effective number of neutrino species
for light gravitinos $N_{3/2}$.
When we calculate the Fisher matrices, we choose the following basic set
of cosmological parameters:
the energy density of matter $\Omega_{m}h^{2}$, baryons $\Omega_{b}h^{2}$
and the dark energy $\Omega_{\Lambda}$,
the scalar spectral index $n_{s}$, the scalar fluctuation amplitude $A_{s}$
(the pivot scale is taken to be $k_{{\rm pivot}}=$ $0.05 \ {\rm Mpc}^{-1}$),
the reionization optical depth $\tau$,
and the primordial value of the $^4$He mass fraction $Y_{\textrm{p}}$.
Fiducial values of these parameters
are taken to
$(\Omega_{m}h^{2},\Omega_{b}h^{2},\Omega_{\Lambda},n_{s},A_{s}\times10^{10},\tau,Y_{\textrm{p}})$
$=( 0.1417, 0.0223, 0.6911, 0.9667, 21.4
, 0.066, 0.25)$,
which are the best fit values of the Planck result \cite{Ade:2015xua}.
For the total neutrino mass $\Sigma m_{\nu} = m_{1}+m_{2}+m_{3}$,
we fix $\Sigma m_{\nu}$ to a fiducial value $\Sigma m_{\nu} = 0.06 \ \textrm{eV}$,
or vary it freely.
In the following analysis,
we fix the neutrino mass hierarchy to be the normal one
and the effective number of neutrino species $N_{\nu}$ to be $3.046$.
For the parameters related to the light gravitino,
we set the fiducial value of $N_{3/2}$ to be $0.059$
and that of $f_{3/2}$ to be
$0.01071$ or $0.05353$, which corresponds to $m_{3/2}=1$~eV
or $5$~eV, respectively when we fix $N_{3/2}$ to be $0.059$.
For $N_{3/2}$,
we fix $N_{3/2}=0.059$, or treat it as a free parameter.
To obtain Fisher matrices we use CAMB \cite{Lewis:1999bs,CAMB}\footnote{
In our analysis of 21 cm line,
we neglect non-linear effects for evolutions of the matter power spectrum
because
we adopt a 21 cm power spectrum only at linear regime.
For CMB lensing,
by performing a public code HALOFIT~\cite{Lewis:1999bs,CAMB},
we have checked that modifications
by including the non-linear effects
are much smaller than typical errors in our analyses
and have negligible impacts on our constraints.
}
for calculations of CMB anisotropies $C_{l}$ and matter power spectra $P_{\delta \delta}(k)$.
In order to combine the CMB experiments with the 21 cm line, BAO
and a Hubble $H_0$ measurement,
we calculate the combined Fisher matrix
\begin{equation}
F_{\alpha\beta}
= F^{\rm (21cm)}_{\alpha\beta} + F^{\rm (CMB)}_{\alpha\beta}
+ F^{\rm (BAO)}_{\alpha\beta} + \tilde{F}^{H_0}_{\alpha \beta}.
\end{equation}
In this paper, we do not use information for a possible correlation
between fluctuations of the 21 cm and the CMB.
\subsection{ Expected future constraints on light gravitino}
\label{subsec:constraints}
In Tables.~\ref{tab:gravi1eV_fixN32}-\ref{tab:gravi5eV_mnu_free},
we summarize constraints on cosmological parameters
from each combination of experiments.
In Tables~\ref{tab:gravi1eV_fixN32} and \ref{tab:gravi5eV_fixN32},
the constraints are
for the cases with fixed $N_{3/2}=0.059$ and $\Sigma m_{\nu}=0.06$~eV.
With Planck, the Simons Array, DESI~(BAO) and a direct measurement of $H_0$ combined,
we obtain a 1~$\sigma$ error on $f_{3/2}$
\begin{align}
\sigma(f_{3/2}) =& 0.00346,
\end{align}
for fiducial $f_{3/2}=0.01071$, which corresponds to $m_{3/2}=1$~eV.
Adding 21 cm line experiments to them,
we see that the constraint on $f_{3/2}$ can be significantly improved.
For example,
for fiducial $f_{3/2}=0.01071$, the combination of
SKA phase~1 and Planck + Simons Array + DESI + $H_0$ gives
\begin{align}
\sigma(f_{3/2}) =& 0.00263,
\end{align}
while the ones of SKA phase~2 and Planck + Simons Array + DESI + $H_0$
gives
\begin{align}
\sigma(f_{3/2}) =& 0.00165.
\end{align}
We can translate them into errors on the mass of light gravitino $m_{3/2}$.
For the case with Planck, the Simons Array, DESI and $H_0$ combined,
the error of $m_{3/2}$ is given as
\begin{align}
\sigma(m_{3/2}) =& 0.33~\textrm{eV}.
\end{align}
If we add SKA~phase~1 or phase~2 to them,
the error can be improved as
\begin{align}
\sigma(m_{3/2}) =& 0.25~\textrm{eV} \ \ \ \textrm{(SKA phase~1)},
\end{align}
\begin{align}
\sigma(m_{3/2}) =& 0.16~\textrm{eV} \ \ \ \textrm{(SKA phase~2)}.
\end{align}
If we
combine Omniscope with Planck + Simons Array + DESI + $H_0$,
the error can improved even further a
\begin{align}
\sigma(m_{3/2}) =& 0.067~\textrm{eV} \ \ \ \textrm{(Omniscope)}.
\end{align}
Thus, from these strong improvements,
we find that observations of 21 cm line are significantly
useful to constrain the mass of light gravitinos.
Next, in Figs.~\ref{fig:gravitino_1eV}-\ref{fig:gravitino_5eV_CMB},
we plot contours of 95\% confidence levels (C.L.) forecasts
of each combination of CMB, 21cm line, BAO and $H_0$
in $N_{3/2}$-$f_{3/2}$ plane.
In
the upper panels of Figs.~\ref{fig:gravitino_1eV}-\ref{fig:gravitino_5eV_CMB},
we fix the total neutrino mass $\Sigma m_{\nu}$,
and in those of the lower panels,
we treat $\Sigma m_{\nu}$ as a free parameter.
In Figs.~\ref{fig:gravitino_1eV} and \ref{fig:gravitino_1eV_CMB}
the fiducial value of $f_{3/2}$
is set to $f_{3/2} = 0.01071$,
which corresponds to $m_{3/2}=1$~eV when we fix $N_{3/2}$ to be $0.059$.
In Figs.~\ref{fig:gravitino_5eV} and \ref{fig:gravitino_5eV_CMB},
the fiducial value of $f_{3/2}$
is set to $f_{3/2} = 0.05353$,
which corresponds to $m_{3/2}=5$~eV when we fix $N_{3/2}$ to be $0.059$.
In the left two panels of Figs.~\ref{fig:gravitino_1eV}-\ref{fig:gravitino_5eV},
each contour represents constraints from Planck + Simons Array + BAO~(DESI) + $H_0$ measurement
or Planck + Simons Array + BAO~(DESI) + $H_0$ measurement + 21cm line (SKA phase~1, phase~2 or Omniscope).
In the right two panels of them, each contour represents constraints
from COrE+ + BAO~(DESI) + $H_0$ measurement
or COrE+ + BAO~(DESI) + $H_0$ measurement + 21cm line (SKA phase~1, phase~2 or Omniscope).
In the left two panels of Figs.~\ref{fig:gravitino_1eV_CMB}-\ref{fig:gravitino_5eV_CMB},
each contour represents constraints by CMB (Planck, Planck + Simons Array or COrE+) only.
In the right two panels of them, each contour represents constraints
by CMB (Planck, Planck + Simons Array or COrE+) + BAO~(DESI) + $H_0$ measurement.
From Figs.~\ref{fig:gravitino_1eV}-\ref{fig:gravitino_5eV_CMB},
we can see that constraints on $N_{3/2}$ and $f_{3/2}$
depend on the fiducial value of $f_{3/2}$.
As the fiducial value of $f_{3/2}$ becomes smaller,
the constraints on $f_{3/2}$ become better
while those
on $N_{3/2}$ become worse.
The dependences result from the following reasons.
From Eq.\eqref{eq:effective_num_gravi} and \eqref{eq:Omega_3_2},
the mass of light gravitinos behaves as
\begin{align}
m_{3/2} \propto
f_{3/2}N_{3/2}^{-\frac{3}{4}}.
\end{align}
From this equation,
we can find that
the variation of $m_{3/2}$ due to changing $N_{3/2}$
becomes smaller
as the fiducial value of $f_{3/2}$ becomes smaller.
Therefore,
the influence due to changing $N_{3/2}$
on the growth of perturbations
becomes less significant
if the fiducial value of $f_{3/2}$ is small.
On the other hand, we can see that
the constraints on $f_{3/2}$
depend on the fiducial value of $f_{3/2}$
mainly in CMB observation.
As the fiducial value of $f_{3/2}$ becomes larger,
the free streaming scale of light gravitinos becomes shorter,
which makes
it more difficult
to obtain the information of the gravitino mass from CMB observations
because we need to measure
higher multi-pole power spectra $C_{\ell}$
in order to obtain the information
of the free-streaming scales.
Next, from Figs.~\ref{fig:gravitino_1eV_CMB}-\ref{fig:gravitino_5eV_CMB},
adding the measurement of the Simons Array to the observation of Planck,
we see that
there are strong improvements on sensitivities to
$f_{3/2}$ and $N_{3/2}$
because the Simons Array can precisely observe the CMB polarizations,
which is quite useful for getting the information of CMB lensing.
COrE+ can further improve the measurement of CMB polarization and hence give tighter constraints.
Moreover, adding BAO data to CMB observations,
we find that
constraints on $N_{3/2}$ are improved somewhat
because several parameter degeneracies are broken by those combinations.
From these results,
we find that we can
detect the nonzero values of $f_{3/2}$ and $N_{3/2}$ at 2$\sigma$ level
by using combinations of next generation CMB observations with BAO data and $H_0$
if $f_{3/2}$, i.e. $m_{3/2}$ has a relatively large value
($f_{3/2}=$0.05353, i.e. $m_{3/2}=5$~eV).
However, it is difficult to obtain lower bounds of $N_{3/2}$
even by using COrE+
if $f_{3/2}$ has a relatively small value
($f_{3/2}=$0.01071, i.e. $m_{3/2}=1$~eV).
Additionally,
from the lower panels of Figs.\ref{fig:gravitino_1eV}-\ref{fig:gravitino_5eV_CMB},
these constraints becomes weaker
when we treat the total neutrino mass as a free parameter
because there is a degeneracy between
the effect of massive neutrino and that of light gravitino.
In that case,
we can obtain a lower bound of $f_{3/2}$
only by using the COrE+ experiment.
On the other hand,
from Figs.\ref{fig:gravitino_1eV}-\ref{fig:gravitino_5eV},
adding the 21 cm experiments to the CMB observations,
we see that there are substantial improvements.
In particular,
the combination of SKA phase~1 with Planck + Simons Array, DESI and $H_0$
has enough sensitivity to obtaining a lower bound of $f_{3/2}$ at 2~$\sigma$ level
even when the fiducial value of $f_{3/2}$ is 0.01071
and we treat the total neutrino mass as a free parameter.
Furthermore,
the combination of SKA phase~2 with Planck + Simons Array, DESI and $H_0$
can detect the nonzero value of $N_{3/2}$
except when we treat the total neutrino mass as a free parameter.
If we use the combination of SKA phase~2 with COrE+, DESI and $H_0$,
we can detect its nonzero value even in that case.
Of course, Omniscope has enough sensitivity
to detect the signature of light gravitino in any cases.
Moreover, in Figs.~\ref{fig:gravitino_mnu},
we plot contours of 95\% C.L. forecasts
of each combination of CMB, 21cm line, DESI and $H_0$
in $\Sigma m_{\nu}$-$f_{3/2}$ plane.
From the figure,
by using the combination of Planck + Simons Array with DESI and $H_0$,
it is difficult to discriminate between
effects of massive neutrino and light gravitino
if the fiducial value of $f_{3/2}$ is $0.01071$.
However, even in that case,
we can discriminate them and obtain a lower bound of $f_{3/2}$
if we use the combination of SKA phase~1 with Planck + Simons Array, DESI and $H_0$.
Additionally, if we use the combination
of SKA phase~2 with COrE+, DESI and $H_0$
or Omniscope,
we can also detect the nonzero neutrino mass, simultaneously.
From our results, we find that
21 cm line observations are
quite useful to constrain the mass of light gravitino,
and can significantly improve
constraints on $f_{3/2}$ and $N_{3/2}$
in combination
with CMB, BAO and $H_0$ observations.
Besides, by using 21 cm line observations,
we
will be able to
discriminate between
effects of massive neutrino and light gravitino
through measuring the difference of
their free streaming scales.
\begin{figure}[htbp]
\centering
\includegraphics[bb=47 142 566 654,width=1\linewidth]{gravi_1eV.pdf}
\caption{
Contours of 95\% C.L. forecasts in $N_{3/2}$-$f_{3/2}$ plane.
We assume that $f_{3/2}=0.01071$ and $N_{3/2}=0.059$, which correspond to $m_{3/2}=1$~eV.
In the upper panels, we fix the total neutrino mass $\Sigma m_{\nu}$,
and in the lower panels, we treat the total neutrino mass as a free parameter.
We show constraints from Planck + Simons Array~(SA) + DESI~(BAO) + $H_{0}$ (dotted purple line)
with SKA phase 1 (dashed yellow-green line), phase 2 (solid green line)
or Omniscope (thick blue line) in the left panels,
and COrE+ + DESI~(BAO) + $H_{0}$
(dotted black line) with SKA phase 1, phase 2 or Omniscope in the right panels.
}\label{fig:gravitino_1eV}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[bb=47 142 566 654,width=1\linewidth]{gravi_5eV.pdf}
\caption{
The same as Fig.\ref{fig:gravitino_1eV} but for $f_{3/2}=0.05353$
and $N_{3/2}=0.059$, which correspond to $m_{3/2}=5$~eV.
}\label{fig:gravitino_5eV}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[bb=47 142 566 654,width=1\linewidth]{gravi_1eV_CMB.pdf}
\caption{
Contours of 95\% C.L. forecasts in $N_{3/2}$-$f_{3/2}$ plane
by CMB combined with BAO~(DESI) and $H_{0}$.
We assume that $f_{3/2}=0.01071$ and $N_{3/2}=0.059$, which correspond to $m_{3/2}=1$~eV.
In the upper panels, we fix the total neutrino mass $\Sigma m_{\nu}$,
and in the lower panels, we treat the total neutrino mass as a free parameter.
We show constraints from CMB only (Planck, Planck + Simons Array(SA) and COrE+) in the left panels,
and combinations of CMB, BAO~(DESI) and $H_{0}$ in the right panels.
}\label{fig:gravitino_1eV_CMB}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[bb=47 142 566 654,width=1\linewidth]{gravi_5eV_CMB.pdf}
\caption{
The same as Fig.\ref{fig:gravitino_1eV_CMB} but for $f_{3/2}=0.05353$
and $N_{3/2}=0.059$, which correspond to $m_{3/2}=5$~eV.
}\label{fig:gravitino_5eV_CMB}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[bb=51 291 563 502,width=1\linewidth]{gravi_1eV_mnu.pdf}
\includegraphics[bb=51 291 563 502,width=1\linewidth]{gravi_5eV_mnu.pdf}
\caption{
Contours of 95\% C.L. forecasts in $\Sigma m_{\nu}$-$f_{3/2}$ plane.
We assume that $\Sigma m_{\nu} = 0.06$~eV and $N_{3/2}=0.059$.
In the upper panels,
we assume $f_{3/2}=0.01071$, which correspond to $m_{3/2}=1$~eV
and in the lower panels, we assume $f_{3/2}=0.05353$, which correspond to $m_{3/2}=5$~eV.
We show constraints from Planck + Simons Array~(SA) + DESI~(BAO) + $H_{0}$ (dotted purple line)
with SKA phase 1 (dashed yellow-green line), phase 2 (solid green line)
or Omniscope (thick blue line) in the left panels,
and COrE+ + DESI~(BAO) + $H_{0}$
(dotted black line) with SKA phase 1, phase 2 or Omniscope in the right panels.
}\label{fig:gravitino_mnu}
\end{figure}
\begin{table}[htbp]
\centering \scalebox{0.75}[0.75]{
\begin{tabular}{l|ccccc}
\hline \hline
& $\sigma( \Omega_{m}h^{2} )$ & $\sigma( \Omega_{b}h^{2} )$ & $\sigma( \Omega_{\Lambda} )$ & $ \sigma( n_s ) $ & $\sigma( A_{s} \times 10^{10} )$
\\
\hline
Planck
& $ 1.32 \times 10^{ -3} $ & $ 2.07 \times 10^{ -4} $ & $ 9.54 \times 10^{ -3} $ & $ 7.16 \times 10^{ -3} $ & $ 1.92 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 5.95 \times 10^{ -4} $ & $ 6.85 \times 10^{ -5} $ & $ 3.87 \times 10^{ -3} $ & $ 2.95 \times 10^{ -3} $ & $ 1.46 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$
& $ 4.63 \times 10^{ -4} $ & $ 6.28 \times 10^{ -5} $ & $ 2.79 \times 10^{ -3} $ & $ 2.65 \times 10^{ -3} $ & $ 1.42 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 4.17 \times 10^{ -4} $ & $ 6.09 \times 10^{ -5} $ & $ 2.50 \times 10^{ -3} $ & $ 2.57 \times 10^{ -3} $ & $ 1.35 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.31 \times 10^{ -4} $ & $ 5.82 \times 10^{ -5} $ & $ 1.88 \times 10^{ -3} $ & $ 2.32 \times 10^{ -3} $ & $ 1.24 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 6.26 \times 10^{ -5} $ & $ 1.40 \times 10^{ -5} $ & $ 4.68 \times 10^{ -4} $ & $ 1.43 \times 10^{ -3} $ & $ 1.04 \times 10^{ -1} $
\\
COrE+
& $ 4.88 \times 10^{ -4} $ & $ 5.20 \times 10^{ -5} $ & $ 3.08 \times 10^{ -3} $ & $ 2.47 \times 10^{ -3} $ & $ 8.33 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 4.08 \times 10^{ -4} $ & $ 4.97 \times 10^{ -5} $ & $ 2.46 \times 10^{ -3} $ & $ 2.35 \times 10^{ -3} $ & $ 8.27 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 3.70 \times 10^{ -4} $ & $ 4.87 \times 10^{ -5} $ & $ 2.23 \times 10^{ -3} $ & $ 2.28 \times 10^{ -3} $ & $ 8.11 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 2.91 \times 10^{ -4} $ & $ 4.73 \times 10^{ -5} $ & $ 1.66 \times 10^{ -3} $ & $ 2.05 \times 10^{ -3} $ & $ 7.67 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 6.14 \times 10^{ -5} $ & $ 1.37 \times 10^{ -5} $ & $ 4.57 \times 10^{ -4} $ & $ 1.36 \times 10^{ -3} $ & $ 6.12 \times 10^{ -2} $
\\
\hline
& $\sigma( \tau )$ & $\sigma( Y_{p} )$ & $\sigma( f_{3/2} )$
\\
\hline
Planck
& $ 4.25 \times 10^{ -3} $ & $ 1.13 \times 10^{ -2} $ & $ 1.86 \times 10^{ -2} $
\\
+ Simons Array (SA)
& $ 3.72 \times 10^{ -3} $ & $ 3.10 \times 10^{ -3} $ & $ 3.95 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$
& $ 3.57 \times 10^{ -3} $ & $ 2.96 \times 10^{ -3} $ & $ 3.46 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 3.41 \times 10^{ -3} $ & $ 2.89 \times 10^{ -3} $ & $ 2.63 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.10 \times 10^{ -3} $ & $ 2.75 \times 10^{ -3} $ & $ 1.65 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 2.47 \times 10^{ -3} $ & $ 1.31 \times 10^{ -3} $ & $ 7.13 \times 10^{ -4} $
\\
COrE+
& $ 2.12 \times 10^{ -3} $ & $ 2.48 \times 10^{ -3} $ & $ 2.69 \times 10^{ -3} $
\\
+ BAO + $H_{0}$
& $ 2.07 \times 10^{ -3} $ & $ 2.43 \times 10^{ -3} $ & $ 2.36 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 2.04 \times 10^{ -3} $ & $ 2.39 \times 10^{ -3} $ & $ 2.06 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 1.94 \times 10^{ -3} $ & $ 2.29 \times 10^{ -3} $ & $ 1.47 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 1.48 \times 10^{ -3} $ & $ 1.16 \times 10^{ -3} $ & $ 6.78 \times 10^{ -4} $
\\
\hline
\end{tabular} }
\caption{1~$\sigma$ errors on cosmological parameters for fiducial $f_{3/2}=0.01071$ ($m_{3/2}=1$~eV) for the cases with fixed $\Sigma m_{\nu}=0.06$~eV and $N_{3/2}=0.059$.}
\label{tab:gravi1eV_fixN32}
\vspace{10pt}
\centering \scalebox{0.75}[0.75]{
\begin{tabular}{l|ccccc}
\hline \hline
& $\sigma( \Omega_{m}h^{2} )$ & $\sigma( \Omega_{b}h^{2} )$ & $\sigma( \Omega_{\Lambda} )$ & $ \sigma( n_s ) $ & $\sigma( A_{s} \times 10^{10} )$
\\
\hline
Planck
& $ 1.19 \times 10^{ -3} $ & $ 2.03 \times 10^{ -4} $ & $ 8.08 \times 10^{ -3} $ & $ 6.81 \times 10^{ -3} $ & $ 1.92 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 4.57 \times 10^{ -4} $ & $ 6.84 \times 10^{ -5} $ & $ 2.85 \times 10^{ -3} $ & $ 2.86 \times 10^{ -3} $ & $ 1.43 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$
& $ 4.01 \times 10^{ -4} $ & $ 6.25 \times 10^{ -5} $ & $ 2.33 \times 10^{ -3} $ & $ 2.65 \times 10^{ -3} $ & $ 1.33 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 3.71 \times 10^{ -4} $ & $ 6.06 \times 10^{ -5} $ & $ 2.19 \times 10^{ -3} $ & $ 2.62 \times 10^{ -3} $ & $ 1.30 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.07 \times 10^{ -4} $ & $ 5.78 \times 10^{ -5} $ & $ 1.79 \times 10^{ -3} $ & $ 2.51 \times 10^{ -3} $ & $ 1.22 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 5.22 \times 10^{ -5} $ & $ 1.20 \times 10^{ -5} $ & $ 4.11 \times 10^{ -4} $ & $ 9.73 \times 10^{ -4} $ & $ 1.05 \times 10^{ -1} $
\\
COrE+
& $ 3.25 \times 10^{ -4} $ & $ 5.19 \times 10^{ -5} $ & $ 1.96 \times 10^{ -3} $ & $ 2.43 \times 10^{ -3} $ & $ 8.23 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 3.05 \times 10^{ -4} $ & $ 4.97 \times 10^{ -5} $ & $ 1.77 \times 10^{ -3} $ & $ 2.36 \times 10^{ -3} $ & $ 7.99 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 2.78 \times 10^{ -4} $ & $ 4.87 \times 10^{ -5} $ & $ 1.63 \times 10^{ -3} $ & $ 2.34 \times 10^{ -3} $ & $ 7.87 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 2.42 \times 10^{ -4} $ & $ 4.72 \times 10^{ -5} $ & $ 1.40 \times 10^{ -3} $ & $ 2.28 \times 10^{ -3} $ & $ 7.52 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 5.10 \times 10^{ -5} $ & $ 1.18 \times 10^{ -5} $ & $ 3.95 \times 10^{ -4} $ & $ 9.56 \times 10^{ -4} $ & $ 6.18 \times 10^{ -2} $
\\
\hline
& $\sigma( \tau )$ & $\sigma( Y_{p} )$ & $\sigma( f_{3/2} )$
\\
\hline
Planck
& $ 4.28 \times 10^{ -3} $ & $ 1.13 \times 10^{ -2} $ & $ 5.73 \times 10^{ -2} $
\\
+ Simons Array (SA)
& $ 3.71 \times 10^{ -3} $ & $ 3.12 \times 10^{ -3} $ & $ 8.97 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$
& $ 3.43 \times 10^{ -3} $ & $ 2.94 \times 10^{ -3} $ & $ 8.88 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 3.33 \times 10^{ -3} $ & $ 2.85 \times 10^{ -3} $ & $ 4.15 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.08 \times 10^{ -3} $ & $ 2.74 \times 10^{ -3} $ & $ 1.64 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 2.50 \times 10^{ -3} $ & $ 1.12 \times 10^{ -3} $ & $ 4.37 \times 10^{ -4} $
\\
COrE+
& $ 2.13 \times 10^{ -3} $ & $ 2.51 \times 10^{ -3} $ & $ 5.30 \times 10^{ -3} $
\\
+ BAO + $H_{0}$
& $ 2.06 \times 10^{ -3} $ & $ 2.43 \times 10^{ -3} $ & $ 5.22 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 2.03 \times 10^{ -3} $ & $ 2.37 \times 10^{ -3} $ & $ 3.52 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 1.93 \times 10^{ -3} $ & $ 2.31 \times 10^{ -3} $ & $ 1.57 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 1.47 \times 10^{ -3} $ & $ 9.93 \times 10^{ -4} $ & $ 4.32 \times 10^{ -4} $
\\
\hline
\end{tabular} }
\caption{Same as in Table \ref{tab:gravi1eV_fixN32} but for fiducial $f_{3/2}=0.05353$ ($m_{3/2}=5$~eV).}
\label{tab:gravi5eV_fixN32}
\end{table}
\begin{table}[htbp]
\centering \scalebox{0.75}[0.75]{
\begin{tabular}{l|ccccc}
\hline \hline
& $\sigma( \Omega_{m}h^{2} )$ & $\sigma( \Omega_{b}h^{2} )$ & $\sigma( \Omega_{\Lambda} )$ & $ \sigma( n_s ) $ & $\sigma( A_{s} \times 10^{10} )$
\\
\hline
Planck
& $ 5.44 \times 10^{ -3} $ & $ 2.12 \times 10^{ -4} $ & $ 2.04 \times 10^{ -2} $ & $ 7.37 \times 10^{ -3} $ & $ 2.06 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 2.06 \times 10^{ -3} $ & $ 6.86 \times 10^{ -5} $ & $ 7.14 \times 10^{ -3} $ & $ 2.97 \times 10^{ -3} $ & $ 1.71 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$
& $ 1.42 \times 10^{ -3} $ & $ 6.32 \times 10^{ -5} $ & $ 4.93 \times 10^{ -3} $ & $ 2.65 \times 10^{ -3} $ & $ 1.57 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 8.16 \times 10^{ -4} $ & $ 6.22 \times 10^{ -5} $ & $ 3.21 \times 10^{ -3} $ & $ 2.57 \times 10^{ -3} $ & $ 1.46 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 5.16 \times 10^{ -4} $ & $ 5.95 \times 10^{ -5} $ & $ 2.13 \times 10^{ -3} $ & $ 2.32 \times 10^{ -3} $ & $ 1.39 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 6.26 \times 10^{ -5} $ & $ 2.27 \times 10^{ -5} $ & $ 4.92 \times 10^{ -4} $ & $ 1.58 \times 10^{ -3} $ & $ 1.06 \times 10^{ -1} $
\\
COrE+
& $ 1.77 \times 10^{ -3} $ & $ 5.21 \times 10^{ -5} $ & $ 6.19 \times 10^{ -3} $ & $ 2.47 \times 10^{ -3} $ & $ 1.00 \times 10^{ -1} $
\\
+ BAO + $H_{0}$
& $ 1.31 \times 10^{ -3} $ & $ 4.98 \times 10^{ -5} $ & $ 4.58 \times 10^{ -3} $ & $ 2.35 \times 10^{ -3} $ & $ 9.32 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 7.93 \times 10^{ -4} $ & $ 4.94 \times 10^{ -5} $ & $ 3.06 \times 10^{ -3} $ & $ 2.29 \times 10^{ -3} $ & $ 8.70 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 5.05 \times 10^{ -4} $ & $ 4.81 \times 10^{ -5} $ & $ 2.06 \times 10^{ -3} $ & $ 2.08 \times 10^{ -3} $ & $ 8.44 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 6.14 \times 10^{ -5} $ & $ 2.11 \times 10^{ -5} $ & $ 4.81 \times 10^{ -4} $ & $ 1.50 \times 10^{ -3} $ & $ 6.42 \times 10^{ -2} $
\\
\hline
& $\sigma( \tau )$ & $\sigma( Y_{p} )$ & $\sigma( f_{3/2} )$ & $\sigma( N_{3/2} )$
\\
\hline
Planck
& $ 4.27 \times 10^{ -3} $ & $ 1.59 \times 10^{ -2} $ & $ 2.34 \times 10^{ -2} $ & $ 1.93 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 3.83 \times 10^{ -3} $ & $ 4.43 \times 10^{ -3} $ & $ 4.52 \times 10^{ -3} $ & $ 7.19 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$
& $ 3.65 \times 10^{ -3} $ & $ 3.83 \times 10^{ -3} $ & $ 3.75 \times 10^{ -3} $ & $ 5.07 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 3.55 \times 10^{ -3} $ & $ 3.42 \times 10^{ -3} $ & $ 2.79 \times 10^{ -3} $ & $ 2.98 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.39 \times 10^{ -3} $ & $ 3.17 \times 10^{ -3} $ & $ 1.92 \times 10^{ -3} $ & $ 2.07 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 2.57 \times 10^{ -3} $ & $ 2.06 \times 10^{ -3} $ & $ 7.23 \times 10^{ -4} $ & $ 6.43 \times 10^{ -3} $
\\
COrE+
& $ 2.13 \times 10^{ -3} $ & $ 3.63 \times 10^{ -3} $ & $ 3.26 \times 10^{ -3} $ & $ 5.89 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 2.08 \times 10^{ -3} $ & $ 3.22 \times 10^{ -3} $ & $ 2.66 \times 10^{ -3} $ & $ 4.48 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 2.07 \times 10^{ -3} $ & $ 2.86 \times 10^{ -3} $ & $ 2.16 \times 10^{ -3} $ & $ 2.73 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 2.03 \times 10^{ -3} $ & $ 2.65 \times 10^{ -3} $ & $ 1.61 \times 10^{ -3} $ & $ 1.77 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 1.58 \times 10^{ -3} $ & $ 1.80 \times 10^{ -3} $ & $ 6.88 \times 10^{ -4} $ & $ 5.80 \times 10^{ -3} $
\\
\hline
\end{tabular} }
\caption{1~$\sigma$ errors on cosmological parameters for fiducial $f_{3/2}=0.01071$ and $N_{3/2}=0.059$ ($m_{3/2}=1$~eV) for the cases with fixed $\Sigma m_{\nu}=0.06$~eV.}
\label{tab:gravi1eV}
\vspace{10pt}
\centering \scalebox{0.75}[0.75]{
\begin{tabular}{l|ccccc}
\hline \hlin
& $\sigma( \Omega_{m}h^{2} )$ & $\sigma( \Omega_{b}h^{2} )$ & $\sigma( \Omega_{\Lambda} )$ & $ \sigma( n_s ) $ & $\sigma( A_{s} \times 10^{10} )$
\\
\hline
Planck
& $ 2.98 \times 10^{ -3} $ & $ 2.09 \times 10^{ -4} $ & $ 1.40 \times 10^{ -2} $ & $ 7.19 \times 10^{ -3} $ & $ 1.97 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 1.02 \times 10^{ -3} $ & $ 6.90 \times 10^{ -5} $ & $ 5.26 \times 10^{ -3} $ & $ 3.07 \times 10^{ -3} $ & $ 1.53 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$
& $ 7.48 \times 10^{ -4} $ & $ 6.27 \times 10^{ -5} $ & $ 3.64 \times 10^{ -3} $ & $ 2.68 \times 10^{ -3} $ & $ 1.51 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 5.48 \times 10^{ -4} $ & $ 6.06 \times 10^{ -5} $ & $ 2.88 \times 10^{ -3} $ & $ 2.64 \times 10^{ -3} $ & $ 1.42 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.95 \times 10^{ -4} $ & $ 5.79 \times 10^{ -5} $ & $ 2.14 \times 10^{ -3} $ & $ 2.52 \times 10^{ -3} $ & $ 1.33 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 5.38 \times 10^{ -5} $ & $ 1.48 \times 10^{ -5} $ & $ 4.28 \times 10^{ -4} $ & $ 1.07 \times 10^{ -3} $ & $ 1.06 \times 10^{ -1} $
\\
COrE+
& $ 8.56 \times 10^{ -4} $ & $ 5.26 \times 10^{ -5} $ & $ 4.44 \times 10^{ -3} $ & $ 2.67 \times 10^{ -3} $ & $ 8.72 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 6.63 \times 10^{ -4} $ & $ 4.97 \times 10^{ -5} $ & $ 3.32 \times 10^{ -3} $ & $ 2.43 \times 10^{ -3} $ & $ 8.70 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 5.23 \times 10^{ -4} $ & $ 4.88 \times 10^{ -5} $ & $ 2.72 \times 10^{ -3} $ & $ 2.38 \times 10^{ -3} $ & $ 8.46 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 3.81 \times 10^{ -4} $ & $ 4.74 \times 10^{ -5} $ & $ 2.04 \times 10^{ -3} $ & $ 2.30 \times 10^{ -3} $ & $ 8.18 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 5.29 \times 10^{ -5} $ & $ 1.45 \times 10^{ -5} $ & $ 4.15 \times 10^{ -4} $ & $ 1.04 \times 10^{ -3} $ & $ 6.22 \times 10^{ -2} $
\\
\hline
& $\sigma( \tau )$ & $\sigma( Y_{p} )$ & $\sigma( f_{3/2} )$ & $\sigma( N_{3/2} )$
\\
\hline
Planck
& $ 4.29 \times 10^{ -3} $ & $ 1.14 \times 10^{ -2} $ & $ 6.47 \times 10^{ -2} $ & $ 6.33 \times 10^{ -2} $
\\
+ Simons Array (SA)
& $ 3.79 \times 10^{ -3} $ & $ 3.29 \times 10^{ -3} $ & $ 1.12 \times 10^{ -2} $ & $ 1.71 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$
& $ 3.70 \times 10^{ -3} $ & $ 2.97 \times 10^{ -3} $ & $ 1.08 \times 10^{ -2} $ & $ 1.42 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 3.54 \times 10^{ -3} $ & $ 2.87 \times 10^{ -3} $ & $ 4.29 \times 10^{ -3} $ & $ 9.33 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.29 \times 10^{ -3} $ & $ 2.76 \times 10^{ -3} $ & $ 1.72 \times 10^{ -3} $ & $ 6.65 \times 10^{ -3} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 2.52 \times 10^{ -3} $ & $ 1.12 \times 10^{ -3} $ & $ 4.83 \times 10^{ -4} $ & $ 9.10 \times 10^{ -4} $
\\
COrE+
& $ 2.14 \times 10^{ -3} $ & $ 2.62 \times 10^{ -3} $ & $ 6.57 \times 10^{ -3} $ & $ 1.26 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 2.10 \times 10^{ -3} $ & $ 2.45 \times 10^{ -3} $ & $ 6.40 \times 10^{ -3} $ & $ 1.04 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 2.08 \times 10^{ -3} $ & $ 2.39 \times 10^{ -3} $ & $ 3.79 \times 10^{ -3} $ & $ 7.62 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 2.01 \times 10^{ -3} $ & $ 2.32 \times 10^{ -3} $ & $ 1.60 \times 10^{ -3} $ & $ 5.29 \times 10^{ -3} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 1.50 \times 10^{ -3} $ & $ 9.93 \times 10^{ -4} $ & $ 4.73 \times 10^{ -4} $ & $ 8.93 \times 10^{ -4} $
\\
\hline
\end{tabular} }
\caption{Same as in Table \ref{tab:gravi1eV} but for fiducial $f_{3/2}=0.05353$ ($m_{3/2}=5$~eV).}
\label{tab:gravi5eV}
\end{table}
\begin{table}[htbp]
\centering \scalebox{0.75}[0.75]{
\begin{tabular}{l|ccccc}
\hline \hline
& $\sigma( \Omega_{m}h^{2} )$ & $\sigma( \Omega_{b}h^{2} )$ & $\sigma( \Omega_{\Lambda} )$ & $ \sigma( n_s ) $ & $\sigma( A_{s} \times 10^{10} )$
\\
\hline
Planck
& $ 5.51 \times 10^{ -3} $ & $ 2.35 \times 10^{ -4} $ & $ 2.47 \times 10^{ -2} $ & $ 7.61 \times 10^{ -3} $ & $ 2.07 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 2.09 \times 10^{ -3} $ & $ 7.27 \times 10^{ -5} $ & $ 9.53 \times 10^{ -3} $ & $ 3.39 \times 10^{ -3} $ & $ 1.77 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$
& $ 1.43 \times 10^{ -3} $ & $ 6.50 \times 10^{ -5} $ & $ 5.05 \times 10^{ -3} $ & $ 2.72 \times 10^{ -3} $ & $ 1.68 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 8.18 \times 10^{ -4} $ & $ 6.43 \times 10^{ -5} $ & $ 3.61 \times 10^{ -3} $ & $ 2.60 \times 10^{ -3} $ & $ 1.63 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 5.19 \times 10^{ -4} $ & $ 6.31 \times 10^{ -5} $ & $ 2.37 \times 10^{ -3} $ & $ 2.34 \times 10^{ -3} $ & $ 1.58 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 7.78 \times 10^{ -5} $ & $ 2.29 \times 10^{ -5} $ & $ 6.65 \times 10^{ -4} $ & $ 1.60 \times 10^{ -3} $ & $ 1.16 \times 10^{ -1} $
\\
COrE+
& $ 1.80 \times 10^{ -3} $ & $ 5.50 \times 10^{ -5} $ & $ 8.27 \times 10^{ -3} $ & $ 3.03 \times 10^{ -3} $ & $ 1.00 \times 10^{ -1} $
\\
+ BAO + $H_{0}$
& $ 1.32 \times 10^{ -3} $ & $ 5.04 \times 10^{ -5} $ & $ 4.72 \times 10^{ -3} $ & $ 2.35 \times 10^{ -3} $ & $ 9.50 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 7.94 \times 10^{ -4} $ & $ 5.00 \times 10^{ -5} $ & $ 3.51 \times 10^{ -3} $ & $ 2.29 \times 10^{ -3} $ & $ 9.08 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 5.10 \times 10^{ -4} $ & $ 4.94 \times 10^{ -5} $ & $ 2.34 \times 10^{ -3} $ & $ 2.08 \times 10^{ -3} $ & $ 8.87 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 7.46 \times 10^{ -5} $ & $ 2.15 \times 10^{ -5} $ & $ 6.50 \times 10^{ -4} $ & $ 1.51 \times 10^{ -3} $ & $ 7.30 \times 10^{ -2} $
\\
\hline%
& $\sigma( \tau )$ & $\sigma( Y_{p} )$ & $\sigma( f_{3/2} )$ & $\sigma( N_{3/2} )$ & $\sigma( \Sigma m_{\nu} )$
\\
\hline
Planck
& $ 4.29 \times 10^{ -3} $ & $ 1.66 \times 10^{ -2} $ & $ 2.82 \times 10^{ -2} $ & $ 2.38 \times 10^{ -1} $ & $ 2.02 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 4.07 \times 10^{ -3} $ & $ 4.85 \times 10^{ -3} $ & $ 6.63 \times 10^{ -3} $ & $ 9.55 \times 10^{ -2} $ & $ 8.44 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$
& $ 3.97 \times 10^{ -3} $ & $ 3.95 \times 10^{ -3} $ & $ 4.84 \times 10^{ -3} $ & $ 5.17 \times 10^{ -2} $ & $ 3.13 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 3.95 \times 10^{ -3} $ & $ 3.52 \times 10^{ -3} $ & $ 3.26 \times 10^{ -3} $ & $ 3.04 \times 10^{ -2} $ & $ 2.77 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.87 \times 10^{ -3} $ & $ 3.35 \times 10^{ -3} $ & $ 2.14 \times 10^{ -3} $ & $ 2.26 \times 10^{ -2} $ & $ 2.33 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 2.76 \times 10^{ -3} $ & $ 2.07 \times 10^{ -3} $ & $ 7.60 \times 10^{ -4} $ & $ 6.65 \times 10^{ -3} $ & $ 8.51 \times 10^{ -3} $
\\
COrE+
& $ 2.16 \times 10^{ -3} $ & $ 4.11 \times 10^{ -3} $ & $ 5.04 \times 10^{ -3} $ & $ 8.30 \times 10^{ -2} $ & $ 7.04 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 2.15 \times 10^{ -3} $ & $ 3.33 \times 10^{ -3} $ & $ 3.52 \times 10^{ -3} $ & $ 4.71 \times 10^{ -2} $ & $ 2.62 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 2.15 \times 10^{ -3} $ & $ 2.92 \times 10^{ -3} $ & $ 2.71 \times 10^{ -3} $ & $ 2.86 \times 10^{ -2} $ & $ 2.41 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 2.13 \times 10^{ -3} $ & $ 2.78 \times 10^{ -3} $ & $ 1.99 \times 10^{ -3} $ & $ 2.11 \times 10^{ -2} $ & $ 2.00 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 1.73 \times 10^{ -3} $ & $ 1.80 \times 10^{ -3} $ & $ 7.29 \times 10^{ -4} $ & $ 5.86 \times 10^{ -3} $ & $ 7.69 \times 10^{ -3} $
\\
\hline
\end{tabular} }
\caption{1~$\sigma$ errors on cosmological parameters for fiducial $f_{3/2}=0.01071$ and $N_{3/2}=0.059$ ($m_{3/2}=1$~eV) for the cases with freely varying $\Sigma m_{\nu}$.}
\label{tab:gravi1eV_mnu_free}
\vspace{10pt}
\centering \scalebox{0.75}[0.75]{
\begin{tabular}{l|ccccc}
\hline \hline
& $\sigma( \Omega_{m}h^{2} )$ & $\sigma( \Omega_{b}h^{2} )$ & $\sigma( \Omega_{\Lambda} )$ & $ \sigma( n_s ) $ & $\sigma( A_{s} \times 10^{10} )$
\\
\hline
Planck
& $ 3.45 \times 10^{ -3} $ & $ 2.43 \times 10^{ -4} $ & $ 2.48 \times 10^{ -2} $ & $ 7.58 \times 10^{ -3} $ & $ 1.97 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 1.23 \times 10^{ -3} $ & $ 7.19 \times 10^{ -5} $ & $ 9.65 \times 10^{ -3} $ & $ 3.17 \times 10^{ -3} $ & $ 1.72 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$
& $ 7.74 \times 10^{ -4} $ & $ 6.48 \times 10^{ -5} $ & $ 3.71 \times 10^{ -3} $ & $ 2.83 \times 10^{ -3} $ & $ 1.60 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 5.52 \times 10^{ -4} $ & $ 6.33 \times 10^{ -5} $ & $ 3.29 \times 10^{ -3} $ & $ 2.72 \times 10^{ -3} $ & $ 1.59 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.96 \times 10^{ -4} $ & $ 6.19 \times 10^{ -5} $ & $ 2.41 \times 10^{ -3} $ & $ 2.66 \times 10^{ -3} $ & $ 1.56 \times 10^{ -1} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 7.60 \times 10^{ -5} $ & $ 1.61 \times 10^{ -5} $ & $ 6.84 \times 10^{ -4} $ & $ 1.13 \times 10^{ -3} $ & $ 1.13 \times 10^{ -1} $
\\
COrE+
& $ 1.08 \times 10^{ -3} $ & $ 5.42 \times 10^{ -5} $ & $ 8.46 \times 10^{ -3} $ & $ 2.81 \times 10^{ -3} $ & $ 9.37 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 6.70 \times 10^{ -4} $ & $ 5.05 \times 10^{ -5} $ & $ 3.51 \times 10^{ -3} $ & $ 2.48 \times 10^{ -3} $ & $ 8.89 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 5.30 \times 10^{ -4} $ & $ 4.97 \times 10^{ -5} $ & $ 3.24 \times 10^{ -3} $ & $ 2.40 \times 10^{ -3} $ & $ 8.84 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 3.85 \times 10^{ -4} $ & $ 4.88 \times 10^{ -5} $ & $ 2.37 \times 10^{ -3} $ & $ 2.34 \times 10^{ -3} $ & $ 8.76 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 7.37 \times 10^{ -5} $ & $ 1.56 \times 10^{ -5} $ & $ 6.64 \times 10^{ -4} $ & $ 1.08 \times 10^{ -3} $ & $ 6.95 \times 10^{ -2} $
\\
\hline
& $\sigma( \tau )$ & $\sigma( Y_{p} )$ & $\sigma( f_{3/2} )$ & $\sigma( N_{3/2} )$ & $\sigma( \Sigma m_{\nu} )$
\\
\hline
Planck
& $ 4.29 \times 10^{ -3} $ & $ 1.16 \times 10^{ -2} $ & $ 6.48 \times 10^{ -2} $ & $ 6.72 \times 10^{ -2} $ & $ 1.73 \times 10^{ -1} $
\\
+ Simons Array (SA)
& $ 4.07 \times 10^{ -3} $ & $ 3.29 \times 10^{ -3} $ & $ 1.15 \times 10^{ -2} $ & $ 2.19 \times 10^{ -2} $ & $ 7.40 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$
& $ 3.96 \times 10^{ -3} $ & $ 3.24 \times 10^{ -3} $ & $ 1.12 \times 10^{ -2} $ & $ 1.92 \times 10^{ -2} $ & $ 3.36 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + SKA1
& $ 3.91 \times 10^{ -3} $ & $ 3.05 \times 10^{ -3} $ & $ 4.29 \times 10^{ -3} $ & $ 1.11 \times 10^{ -2} $ & $ 2.86 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + SKA2
& $ 3.85 \times 10^{ -3} $ & $ 2.97 \times 10^{ -3} $ & $ 1.73 \times 10^{ -3} $ & $ 7.75 \times 10^{ -3} $ & $ 2.42 \times 10^{ -2} $
\\
+ SA + BAO + $H_{0}$ + Omniscope
& $ 2.66 \times 10^{ -3} $ & $ 1.25 \times 10^{ -3} $ & $ 4.84 \times 10^{ -4} $ & $ 9.30 \times 10^{ -4} $ & $ 8.33 \times 10^{ -3} $
\\
COrE+
& $ 2.16 \times 10^{ -3} $ & $ 2.63 \times 10^{ -3} $ & $ 6.89 \times 10^{ -3} $ & $ 1.69 \times 10^{ -2} $ & $ 6.15 \times 10^{ -2} $
\\
+ BAO + $H_{0}$
& $ 2.16 \times 10^{ -3} $ & $ 2.61 \times 10^{ -3} $ & $ 6.72 \times 10^{ -3} $ & $ 1.50 \times 10^{ -2} $ & $ 2.96 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA1
& $ 2.15 \times 10^{ -3} $ & $ 2.49 \times 10^{ -3} $ & $ 3.80 \times 10^{ -3} $ & $ 9.80 \times 10^{ -3} $ & $ 2.54 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + SKA2
& $ 2.13 \times 10^{ -3} $ & $ 2.43 \times 10^{ -3} $ & $ 1.64 \times 10^{ -3} $ & $ 6.86 \times 10^{ -3} $ & $ 2.05 \times 10^{ -2} $
\\
+ BAO + $H_{0}$ + Omniscope
& $ 1.64 \times 10^{ -3} $ & $ 1.09 \times 10^{ -3} $ & $ 4.73 \times 10^{ -4} $ & $ 9.19 \times 10^{ -4} $ & $ 7.66 \times 10^{ -3} $
\\
\hline
\end{tabular} }
\caption{Same as in Table \ref{tab:gravi1eV_mnu_free} but for fiducial $f_{3/2}=0.05353$ ($m_{3/2}=5$~eV).}
\label{tab:gravi5eV_mnu_free}
\end{table}
\section{Conclusion
\label{sec:conclusion}
In this paper,
we have studied how well we can constrain the mass of light gravitino
$m_{3/2}<\mathcal{O}(10)$~eV,
or more specifically,
the fraction of light gravitinos in the total dark matter density $f_{3/2}$,
and the effective number of neutrino species
for
light gravitinos $N_{3/2}$,
which determine $m_{3/2}$,
by using observations of 21 cm line, CMB, BAO and direct measurements of $H_0$.
In the early Universe,
light gravitinos are produced from thermal plasma,
and they behave as warm dark matter~(WDM) at late epochs.
Thus, they imprint characteristic signatures
on density fluctuations,
and we can detect the features through cosmological observations,
such as CMB and 21 cm line.
Adding the measurement of the Simons Array,
which is a planned precise CMB polarization observation,
to the observation of Planck,
we see that there are strong improvements on sensitivities to constraints
on $f_{3/2}$ and $N_{3/2}$
because the Simons Array is quite useful for getting the information of CMB lensing.
If $f_{3/2}$, i.e. $m_{3/2}$ has a relatively large value
($f_{3/2}=$0.05353, which corresponds to $m_{3/2}=5$~eV),
by using Planck + Simons Array or COrE+,
we can detect the nonzero values of $f_{3/2}$ and $N_{3/2}$ at 2$\sigma$ level.
Besides, adding the 21 cm experiments to the CMB observations,
we see that there are substantial improvements.
For the cases with fixed $N_{3/2}=0.059$ and $\Sigma m_{\nu}=0.06$~eV,
by combining SKA phase~1 with Planck, the Simons Array, DESI and
a direct measurement of $H_0$ at 1\% accuracy,
we can obtain a 1~$\sigma$ error on the mass of light gravitinos,
$\sigma(m_{3/2})=0.25$~eV
for fiducial $f_{3/2}=0.01071$,
which corresponds to $m_{3/2}=1$~eV.
If we use SKA phase~2 or Omniscope,
the error can be improved as
$\sigma(m_{3/2})=0.16$~eV (SKA phase~2) or $\sigma(m_{3/2})=0.067$~eV (Omniscope), respectively.
In particular,
the combination of SKA phase~1 with Planck + Simons Array, DESI and $H_0$
has enough sensitivity to obtaining a lower bound of $f_{3/2}$ at 2~$\sigma$ level
even when the fiducial value of $f_{3/2}$ is %
as small as
0.01071
and we treat $N_{3/2}$ and the total neutrino mass as free parameters.
Furthermore,
the combination of SKA phase~2 with Planck + Simons Array, DESI and $H_0$
can detect the nonzero value of $N_{3/2}$
except when we treat the total neutrino mass as a free parameter.
Moreover,
if we use the combination of SKA phase~2 with COrE+, DESI and $H_0$,
we can detect the nonzero value of $N_{3/2}$ even in that case.
Although
it is difficult to discriminate between
the effects of massive neutrinos and light gravitinos
only by using Planck + Simons Array, BAO and a measurement of $H_0$,
it becomes feasible if a precise observation of 21 cm line is incorporated.
In particular,
the combination of SKA phase~2 with COrE+, DESI and $H_0$
has enough sensitivities to determine
the parameters of light gravitino and the total neutrino mass
at 2~$\sigma$ level, simultaneously.
If we use Omniscope, we can
detect features of light gravitinos and massive neutrinos
even with on-going CMB observations.
Our results indicate that
combining 21 cm line observations with CMB observations
has strong impacts on the determination of the mass of light gravitinos
and understanding the origin of matter in the Universe.
\section*{Acknowledgments}
We thank Tomo Takahashi for a useful correspondence
about cosmological effects of light gravitino.
Besides,
we appreciate the significant contribution made by
Toyokazu Sekiguchi.
This work is supported by MEXT KAKENHI Grant Number 15H05889 (M. K.),
JSPS KAKENHI Grant Number 25400248 (M. K.) and also by
the World Premier International Research Center Initiative (WPI), MEXT, Japan.
|
\section{Introduction}
Transition metal dichalcogenides (TMDs) are layered, weakly-coupled materials that can exist in few- and monolayer forms. Recently, this class of materials has attracted intense study due to the remarkable electronic and optical properties it exhibits, such as valley-selective circular dichroism, as well as coupling of spin and valley quantum numbers~\cite{cao12,xiao12,mak12} and the formation of strongly bound excitons and trions~\cite{ashwin12, qiu13, mak13, chernikov14, ugeda14, klots14, zhu14, ye14,hill15}. Molybdenum disulfide (\MoS2) is a prototypical TMD. In its most common semiconducting form (2H), monolayer \MoS2 consists of a layer of Mo atoms sandwiched between two layers of S atoms in a trigonal prismatic arrangement. In bulk and few-layer form, \MoS2 is an indirect gap semiconductor, but in monolayer form, it becomes a direct gap semiconductor, with a gap located at the $K$ and $K'$ points in the Brillouin zone~\cite{mak10,splendiani10}.
The optical spectrum of \MoS2 has been extensively studied experimentally. It has an optical gap of $1.9$~eV at room temperature~\cite{mak10, splendiani10}, which blue-shifts by as much as $0.1$~eV at low temperatures between 5 and 100K~\cite{tongay12, soklaski14}. The first peak in the optical spectrum is split by spin-orbit coupling by $0.15$~eV into two peaks commonly referred to as ``A'' and ``B''\cite{mak10}. The electronic quasiparticle bandgap is much harder to determine experimentally, but various experiments suggest that the bandgaps of monolayer \MoS2 and several other TMDs with the same structure lies between $0.2$ and $0.7$~eV above the optical gap~\cite{chernikov14, ye14, zhu14, ugeda14, klots14,zhang14,hill15}, indicating a large exciton binding energy.
There have also been numerous theoretical studies of the electronic and optical
properties of monolayer \MoS2 with widely differing results. The many-body
perturbation theory-based \textit{ab initio} GW approximation~\cite{hybertsen86} plus Bethe Salpeter equation (GW-BSE) approach~\cite{rohlfing98,albrecht98} is one the most common and accurate methods for computing quasiparticle (QP) bandstructures and optical response including electron-electron and electron-hole interactions. However, even within the general GW-BSE approach, there is a great deal of disagreement in the literature over everything from the magnitude and location of the QP bandgap to the exciton binding and excitation energies~\cite{ashwin12, lambrecht12, komsa12, wirtz13, qiu13, shi13, huser13, soklaski14,bernardi13,palummo15}. In this paper, we address the source of these inconsistencies and make note of computational issues in GW-BSE calculcations that arise for quasi-two-dimensional (quasi-2D) semiconductors and other reduced dimensional systems.
The main results of the paper are the following:
\begin{itemize}
\item The major computational challenges when dealing with mono- and
few-layers TMDs arise from the finite extent of atomic scale in one of the
spatial directions. This introduces rapid variations in the screening, which leads to complications in the computation of the quasiparticle and excitonic properties~\cite{qiu13,huser13}.
\item The convergence of quasiparticle gaps with respect to the k-point sampling, dielectric cutoff and number of bands included in the self-energy operator is much slower than what is reported in earlier work and is closely tied to the supercell size used and the treatment of the quasi-2D behaviour of the Coulomb interaction. The lack of convergence is sufficient to explain the varying results in the literature for GW-BSE calculations on monolayer \MoS2 and other TMDs.
\item We show that different numerical treatments of the divergence in the Coulomb interaction shifts the exciton continuum and can lead to false convergence of the binding energy with respect to k-point sampling. In particular, we find that it is possible to obtain an apparent agreement of the calculated optical gap with experiment, even though the exciton binding energy and the higher exctionic states are not computed correctly.
\end{itemize}
This paper is organized as follows. In section II, we discuss the dielectric screening in quasi-2D semiconductors and review the effect of the truncation of the Coulomb potential. In section III, we discuss the QP bandstructure, the convergence of the self-energy, including special considerations for quasi-2D systems, the effect of updating the Green's function $G$ in the $GW_0$ approach and the frequency dependence of the screening. In section IV, we discuss the effect of screening on the optical response of \MoS2, characterize the excitons and their wavefunctions, discuss how they converge in our calculations, and discuss the effects of quasiparticle lifetimes. We conclude in section V by summarizing our results.
\section{Electron-electron and electron-hole Interactions and Screening in 2D}
\subsection{\label{sect:truncation}Coulomb Truncation and Convergence}
First-principles calculations using planewave basis sets require periodic boundary conditions. This means that for 2D systems, such as monolayer \MoS2, it is necessary to increase the dimension $L_z$ of the unit cell in the aperiodic direction to avoid interactions between repeated monolayers~\cite{schluter75}. With conventional DFT functionals, such as LDA or GGA, there are no long range interactions for a neutral system, so a vacuum of $\sim\angs{5}$ ($L_z \sim \angs{10}$) is sufficient to converge the relative eigenvalues (other values, such as the work function and ionization energies, require a larger vacuum to prevent interactions between periodic images). However, when we compute the polarizability and related quantities in the GW approach, we end up calculating a response function that is long ranged, and it becomes computationally unfeasible to include enough vacuum to prevent periodic images from interacting.
One effective solution for this problem is to explitly truncate the Coulomb
interaction in real space along the aperiodic direction. This is implemented in the BerkeleyGW package~\cite{jdeslip11BGW} following Ismail-Beigi's scheme~\cite{beigi06}. The truncated Coulomb potential has a closed form in reciprocal space,
\begin{equation}
\label{eq:vT}
v^{\mathrm{trunc}}(\mathbf{q}) = \frac{4\pi}{q^2}\left[1-e^{-\frac{q_{xy} L_z}{2}}\cos\left(\frac{q_z L_z}{2}\right)\right],
\end{equation}
where $q_{xy} = (q_x^2+q_y^2)^{1/2}$. This allows us to directly compute the static RPA inverse dielectric matrix without spurious interactions between the repeated monolayers in our supercell geometry as
\begin{equation}
\epsilon^{-1}_{\mathbf{GG'}}(\mathbf{q}) = \delta_\mathbf{GG'} + v^{\mathrm{trunc}}(\mathbf{q+G})\chi_\mathbf{GG'}(\mathbf{q}),
\end{equation}
where $\chi_\mathbf{GG'}(\mathbf{q})$ is the static non-interacting RPA polarizability.
We now examine how the features of the dielectric matrix evolve with supercell size with and without Coulomb truncation. In isotropic bulk systems, the screening is dominated by the ``head'' element $\mathbf{G}{{=}}\mathbf{G'}{{=}}0$~\cite{hybertsen85,hybertsen86,hybertsen87,zhang89}. In quasi-2D systems, however, the $G_z$'s (the reciprocal lattice vectors along the aperiodic direction) are almost continuous, so it is no longer reasonable to look at the single element $G_z{=}0$. In Fig.~\ref{fig:epsinv}, we plot $\epsilon^{-1}_{\mathbf{GG}'}(\mathbf{q})$ for elements where $G_x{=}G_x'{=}G_y{=}G_y'{=}0$ and $G_z{=}G_z'$ for several different values of $G_z$. When the truncated Coulomb interaction is used, the behavior of $\epsilon^{-1}_{\mathbf{GG}'}(\mathbf{q})$ changes depending on whether $G_z$ is odd or even. $\epsilon^{-1}(\mathbf{q})$ goes smoothly to a value less than $1$ as $q$ goes to $0$, when $G_z$ is odd, and sharply returns to $1$ as $q$ goes to $0$, when $G_z$ is even. This behavior arises from the $\cos$ term in the truncated Coulomb interaction, and contrasts with the untruncated case, where $\epsilon^{-1}(\mathbf{q})$ goes to a number less than $1$ as $q$ goes to $0$ for all $G_z$'s, with most of the screening coming from $G_z{=}0$.
The screening behavior with and without Coulomb truncation also depends, unsurprisingly, on the amount of vacuum $L_z$. In both cases, consecutive $G_z$'s for $G_z>0$ become more similar as $L_z$ increases, since the separation between $G_z$'s is $\frac{2\pi}{L_z}$. Consequently, the number of $G_z$'s required to capture the screening behavior increases proportionally with $L_z$.
There is also a direct correlation between the $q$-dependence of the dielectric matrix with $L_z$ when we employ the truncated Coulomb interaction. As shown in the left panels in Fig.~\ref{fig:epsinv}, the ``dip'' feature for even $G_z$'s becomes sharper as $L_z$ increases, so the k-point sampling must be fine enough to resolve the features in $\epsilon^{-1}_{00}(\mathbf{q})$. An important consequence is that \emph{convergence of k-point sampling is tied to the size of the supercell}. Fig.~\ref{fig:kconv} shows the convergence of the QP gap with respect to k-point sampling and the size of the vacuum. When Coulomb truncation is used (Fig.~\ref{fig:kconv}~(b)), the QP gap converges more slowly for larger $L_z$'s, reflecting the need to resolve sharper features in $\epsilon^{-1}_{00}(\mathbf{q})$. However, the QP gap converges to the same value regardless of the supercell size when Coulomb truncation is used.
The picture is different and shows a significantly slower k-point convergence when we don't employ Coulomb truncation. As
shown in Fig.~\ref{fig:kconv}~(a), the QP gap still displays a very strong
dependence on k-point sampling at the densest grid size of $36\times36$. There are two important differences here with respect to
the case with truncated Coulomb potential: (1) these calculations converge to a
smaller incorrect QP gap, and (2) the convergence with respect to k-point sampling is not
monotonic but changes direction as the k-grid sampling becomes finer. Both
these facts are understood from the long wavelength behavior of the
screening. Whenever $q \lesssim 1/L_z$, the calculation without a truncated
Coulomb potential includes a spurious polarization due to the repeated
monolayers in the aperiodic direction. This spurious term screens out the
Coulomb interaction and decreases the QP gap.
Finaly, it's important to mention the dependence of the number of bands needed for the various quantities in the GW calculation on
$L_z$. The number of empty states included in our calculation is well
approximated by the number of planeawaves $\ket{\mathbf{G}}$ with kinetic energy
less than the dielectric cutoff $E=|\mathbf{G}|^2/2$, so it is proportional to
the supercell volume. If the number of bands is kept constant while $L_z$ is
increased, the screening will not be captured properly in the GW calculation,
and the QP gaps will be overestimated. We attritube the reason why some studies
found that the QP gaps increase much more when the vacuum is increased to this
false convergence~\cite{huser13,komsa12}.
\begin{figure}
\includegraphics[width=246.0pt]{Fig1_epsinv}
\caption{\label{fig:epsinv} (Color online) Evolution of the first few diagonal elements of the inverse dielectric matrix, $\epsilon^{-1}(\mathbf{q})$, for $G_x{=}G_y{=}G_x'{=}G_y'{=}0$ and $G_z{=}G_z'$ with (left) and without (right) Coulomb truncation for $L_z{=}\angs{15}$ (a,b), \angs{20} (c,d), and \angs{25} (e,f) supercell sizes. A cutoff of $35$~Ry and $6,000$ bands was used for calculating all $\epsilon^{-1}_{\mathbf{G,G'}}(\mathbf{q})$ in this figure. The value of $G_z$ is given in units of $\frac{2\pi}{L_z}$.
}
\end{figure}
\begin{figure}
\includegraphics[width=246.0pt]{Fig2_kconv}
\caption{\label{fig:kconv} (Color online) Convergence of the error in the QP gap with k-point sampling (a) without Coulomb truncation and (b) with Coulomb truncation, for supercell sizes $L_z=\angs{15}$ (black squares), $L_z=\angs{20}$ (blue circles), and $L_z=\angs{25}$ (red triangles). Zero is set to the QP gap with Coulomb truncation extrapolated to infinite k-point sampling for $L_z=\infty$.}
\end{figure}
\subsection{Effective 2D Dielectric Function}
For simplicity, we discuss here the static dielectric function. The same discussion caries over for the dynamic case. In general, the dielectric function of a material is defined as the following relation between the bare Coulomb potential v and the effective screened Coulomb interaction W:
\begin{equation}
\label{eq:epsinv}
W(\mathbf{r}_1, \mathbf{r}_2) \equiv \int \mathrm{d}^3 r_3\;
\epsilon^{-1}(\mathbf{r}_1, \mathbf{r}_3) v(|\mathbf{r}_3-\mathbf{r}_2|).
\end{equation}
Our goal now is to define an effective 2D dielectric function between two electrons in a monolayer material. Due to confinement, the modulus squared of the wavefunction (in a tight-binding framework) associated to either electron, $l=1,2$, can be written as $\rho_i(\mathbf{r}-\mathbf{s}_l)$, where $i=1,2$ labels different orbitals and $\mathbf{s}_l$ is a coordinate in the xy-plane around which the orbital is centered. In analogy to Eq.~\ref{eq:epsinv}, we define the effective 2D inverse dielectric function in terms of the strength of the electronic interaction integration between orbitals $i$ and $j$ as
\begin{align}
\label{eq:epsinv2D_1}
W_{i j}(\mathbf{s}_1, \mathbf{s}_2)
&\equiv \int \mathrm{d}^3 r_1 \; \mathrm{d}^3 r_2 \;
\rho_i(\mathbf{r}_1-\mathbf{s}_1) W(\mathbf{r}_1, \mathbf{r}_2) \rho_j(\mathbf{r}_2-\mathbf{s}_2) \notag \\
&\equiv \int \mathrm{d}^2 s_3 \;
\left(\epsilon^{-1}_\mathrm{2D}\right)_{i j}(\mathbf{s}_1, \mathbf{s}_3) v(|\mathbf{s}_3-\mathbf{s}_2|).
\end{align}
In order to gain further insight on the form of the response function, we assume that both $W$ and $\epsilon^{-1}_\mathrm{2D}$ are isotropic and depend only on $s\equiv|\mathbf{s}_2-\mathbf{s}_1|$. Such a simplification allows us to write the strength of the electronic interaction between the two orbitals in real space as
\begin{equation}
\label{eq:Ws}
W_{i j}(s) = \frac{1}{2\pi L_z} \mathcal{F}_0
\left[ \sum_{\mathbf{G}_z \mathbf{G}_z'}
\rho_i^*(\mathbf{q}+\mathbf{G}_z) W_{\mathbf{G}_z \mathbf{G}_z'}(q) \rho_j(\mathbf{q}+\mathbf{G}_z')
\right]\!(s),
\end{equation}
where $\rho(\mathbf{q}+\mathbf{G}_z)\equiv\int \mathrm{d}^3r e^{i(\mathbf{q}+\mathbf{G}_z)\cdot\mathbf{r}}\rho(\mathbf{r})$, and $\mathcal{F}_0 [f](s) \equiv 2\pi\int_0^\infty \mathrm{d}q\; q\,f(q)\,J_0(q s)$ is the Hankel transform of $f$.
In reciprocal space, the effective 2D inverse dielectric function is simply the
ratio between the 2D screened Coulomb interaction (2D Fourier transform of
Eq.~\ref{eq:Ws}) and the truly two-dimensional bare Coulomb potential, $v_\mathrm{2D}(q) = 2\pi e^2/q$.
The simplest choice of orbitals is a delta function at $\mathbf{r}=(\mathbf{s},z=0)$, which yields the effective 2D screening\begin{equation}
\label{eq:epsinv2D_2}
\epsilon^{-1}_\mathrm{2D}(q)
= \frac{q} {2\pi e^2 L_z} \sum_{\mathbf{G}_z \mathbf{G}_z'} W_{\mathbf{G}_z \mathbf{G}_z'}(q).
\end{equation}
Eq.~\ref{eq:epsinv2D_2} defines an effective 2D dielectric for a quasi-2D material,
where the complicated details of the screening in the out-of-plane direction $z$
have been integrated out. We note that our expression for $\epsilon^{-1}_\mathrm{2D}(q)$
differs from that defined in Refs.~\cite{huser13,latini15}, who define
it by the field in a region in the slab induced by a plane-wave external potential.
In contrast, Eq.~\ref{eq:epsinv2D_2} measures how
much the bare 2D Coulomb potential $v_\mathrm{2D}(q) = 2\pi e^2/q$ between two
point charges in the middle of the \MoS2
plane gets screened due to electronic screening. This is the relevant quantity to derive
low-energy Hamiltonians to model electron-electron and electron-hole interactions in
quasi-2D systems, including excitonic states and electron scattering.
In Fig.~\ref{fig:eps2D}~(c-f), we show the reciprocal-space effective 2D
dielectric function $\epsilon_\mathrm{2D}(q)$. The corresponding real-space
curves are obtained by taking the Hankel transform of Eq.~\ref{eq:epsinv2D_2}
and are shown in Fig.~\ref{fig:eps2D}~(a,b). There is a very sharp peak in
$\epsilon_\mathrm{2D}(s)$ at $s=\angs{1.5}$, which corresponds to roughly
half the thickness of the slab. This peak can be understood if we consider
the Coulomb interaction between two point charges embeded in a quasi-2D
semiconductor: as in 2D semiconductors, if two charges are very close together,
there is not enough space for the electronic cloud to polarize, so
$\epsilon_{\mathrm{2D}}(s{\rightarrow}0) = 1$. At the same time, if the two charges
are very far away, the field lines connecting the charges travel mainly through the vacuum,
so they are not much affected by the intrinsic dielectric environment of the quasi-2D semiconductor
and $\epsilon_{\mathrm{2D}}(s{\rightarrow}\infty) = 1$. Therefore, there is a finite
distance $s_\mathrm{max}$ where $\epsilon_{\mathrm{2D}}(s_\mathrm{max})$ must exhibit
its maximum. The value of the peak of $\epsilon_{\mathrm{2D}}(s_\mathrm{max})$ depends on
the polarizability and thickness of the material. We note that $L_z$ should have no effect on the effective 2D screening as long as it
is large enough to contain the charge density within the truncated Coulomb interaction approach. This is \textit{not} true for the untruncated case.
For very short distances ($s < \angs{1}$), the effective 2D dielectric screening
with and without truncation are similar, but at larger distances polarizability of the replica slab together with the long-range interaction results in drastic overscreening. Instead of
approaching $1$, $\epsilon_{\mathrm{2D}}(s)$ approaches a larger finite
constant, which is the macroscopic dielectric constant of a bulk system consisting of layers of \MoS2 separated by layers of vacuum. While this constant indeed approaches $1$ as $L_z \rightarrow \infty$, it does so very slowly. Thus, it is very important to truncate the Coulomb interaction to include correctly the effects of the dielectric response of quasi-2D systems.
Similar features are seen in the effective 2D dielectric function for the converged results in reciprocal space, as shown in Fig.~\ref{fig:eps2D}~(c,d). Specifically: (1) there is a peak in $\epsilon_{\mathrm{2D}}(q)$; (2) when the Coulomb interaction is truncated, $\epsilon_{\mathrm{2D}}(q)$ does not depend on $L_z$; and (3) while $\epsilon_{\mathrm{2D}}(q{\rightarrow}0)=1$ when we truncate the Coulomb potential, it incorrectly approaches a different and larger value when we don't truncate the potential.
We also show the effective screening for different energy cutoffs for
the dielectric matrix, in Fig.~\ref{fig:eps2D}~(e,f). The effect of changing the
dielectric cutoff is similar for both the truncated and untruncated Coulomb
interactions. For very small $q$'s, before the peak, screening does not depend
strongly on the cutoff. For larger $q$'s, decreasing the cutoff results in
overscreening. Therefore, depending on the property one is interested in
(quasiparticle or excitonic levels), different convergence parameters may have to be
used. In particular, the convergence of quasiparticle states, as computed within
the GW approximation, converges very slowly because the self energy depends on
$\epsilon$ at both short and long distances.
\begin{figure}
\includegraphics[width=246.0pt]{Fig3_epsilon2D}
\caption{\label{fig:eps2D} (Color online) Effective 2D screening between two point charges in the Mo plane with Coulomb truncation (left) and without Coulomb truncation (right). Panels (a) and (b) compare the effective screening in real space when $L_z=\angs{15}$ (solid blue) and $L_z=\angs{25}$ (dotted red) with a $35$~Ry cutoff. Panels (c) and (d) are the corresponding reciprocal space plots of (a) and (b). Panels (e) and (f) compare effective screening in reciprocal space when the cutoff is $35$~Ry (dashed red) and $8$~Ry (solid green) with $L_z=\angs{25}$.}
\end{figure}
Finally, we compare the effective 2D screening obtained from our \textit{ab initio} calculations with the screening model developed by Keldysh~\cite{keldysh}, which is frequently used to describe screening of excitons in quasi-2D materials~\cite{cudazzo11,pulci12,berkelbach13,chernikov14,wu15,latini15}. In the Keldysh model, which is based on a slab of constant dielectric value, the potential between two charges in a slab of thickness $d$ has the form
\begin{equation}
\label{eq:keldysh}
V_{\mathrm{2D}}(s)=\frac{\pi e^2}{2\rho_0}\left[H_0\left(\frac{s}{\rho_0}\right)-Y_0\left(\frac{s}{\rho_0}\right)\right],
\end{equation}
where $H_0$ and $Y_0$ are respectively the Struve and Bessel functions of the second kind and $\rho_0$ is a screening length, which is $\rho_0=\frac{d\epsilon}{2}$~\cite{keldysh}, where $\epsilon$ is the in-plane dielectric constant of the bulk material. If the slab is taken to be strictly 2D, it has been shown~\cite{cudazzo11} that the screening length is proportional to the 2D polarizability of the layer, and taking the 2D Fourier transform of Eq.~\ref{eq:keldysh} results in a dielectric function of the form
\begin{equation}
\epsilon_{\mathrm{2D}}(q) = 1 + \rho_0 q,
\end{equation}
where $\rho_0=2\pi\alpha_{\mathrm{2D}}$. Here, $\alpha_{\mathrm{2D}}$ is the 2D polarizability and can be related to the polarizability of the actual quasi-2D slab by fitting to the long wavelength limit of the \textit{ab initio} polarizability. We fit the Keldysh model to our \textit{ab intio} effective dielectric function at small $q$, as defined in Eq.~\ref{eq:epsinv2D_2}, and obtain an effective screening length of $\rho_0=35\mathrm{\AA}$ or an effective slab thickness of $d=6\mathrm{\AA}$, which is about twice the thickness of monolayer \MoS2 measured from the center of the sulfur atoms. A comparison of our \textit{ab intio} effective dielectric function with the best fit to the Keldysh model is shown in Fig.~\ref{fig:keldysh}. We see that the Keldysh model can be adjusted to give a good description of the form of the screening in the long wavelength limit and thus can describe the screening seen by excitons as long as the exciton radius is on the order of or larger than the screening length $\rho_0$, which is unknown without an \textit{ab initio} calculation. Moreover, for phenomena that depend on short-range or varying length scale screening, the Keldysh model would drastically overestimate the screening in quasi-2D systems.
\begin{figure}
\includegraphics[width=246.0pt]{Fig3_keldysh}
\caption{\label{fig:keldysh} (Color online) Comparison of the effective 2D screening as defined by Eq.~\ref{eq:epsinv2D_2} (red lines) with the Keldysh model (black lines) in real space (a) and reciprocal space (b). The Keldysh model uses an effective slab thickness of $d=6\mathrm{\AA}$ to obtain the best fit to the \textit{ab initio} results.}
\end{figure}
\section{Quasiparticle Bandstructure}
In this section, we discuss the computational details and results of our GW calculation of the QP bandstructure.
\subsection{Computational Details and Convergence}
We use density functional theory (DFT)~\cite{hohenberg64,kohn65}, as implemented in Quantum ESPRESSO~\cite{espresso},
in the local density approximation (LDA) to obtain a mean-field starting point for our
GW calculation~\cite{hybertsen86}. Different choices of the DFT functional and a relaxed versus
experimental crystal structure can result in about $0.1$~eV difference in the QP gap of \MoS2{}.
We find that relaxing the structure with an LDA functional increases the gap at K
by $0.04$~eV compared to the experimental structure. Given identical structures,
using a GGA functional decreases the gap by $0.03$~eV compared to LDA.
We use norm-conserving pseudopotentials and include the Mo $4s$ and $4p$ semicore states
and the $4d$ valence state. Including the semicore $4s$ and $4p$ states is necessary
to accurately capture the exchange contribution to the self energy. However, these deep
$4s$ and $4p$ states are not included in the charge density used in the Hybertsen-Louie
Generalized Plasmon Pole (HL-GPP) model~\cite{hybertsen86} to calculate the self energy,
since they are more than $35$~eV below the Fermi energy and, thus, do not
contribute to low-energy screening. We use a supercell with \angs{25} of vacuum in the aperiodic
direction, and we relax the supercell using a wavefunction cutoff of $350$~Ry and a
24x24x1 k-grid, resulting in an in-plane lattice constant of \angs{3.15}, which deviates less
than $1$\% from the experimental lattice constant of few-layer \MoS2~\cite{young68}.
Then, we generate wavefunctions used in the GW-BSE calculation using a wavefunction\cite{hybertsen86}
cutoff of $125$~Ry, which is sufficient to converge the bare exchange contribution~
to the QP gap to within $0.01$~eV.
Our GW calculation is performed with the BerkeleyGW package~\cite{jdeslip11BGW}.
We calculate the dielectric matrix using the truncated Coulomb interaction discussed
in section II and using a 24x24x1 k-point sampling to converge the QP gap to within
$0.05$~eV (see Fig.~\ref{fig:kconv}). We take into account dynamical screening effects
in the self energy through the HL-GPP model. We also use the
static remainder technique~\cite{jdeslip10SR} to reduce the number of necessary unoccupied states.
As discussed in our previous work~\cite{qiu13}, GW calculations on \MoS2 and TMDs in
general converge very slowly with respect to the energy cutoff ($E_S$) of the dielectric
matrix and the number of bands ($N_b$) included in the polarizability and Coulomb-hole
summations of the self energy. The slow convergence of $E_S$ arises from the presence of localized $d$ orbitals
near the Fermi energy and the different character of the valence and conduction bands.
The slow convergence of $N_b$ arises due to the large number of G-vectors in the dielectric
matrix and the supercell size, as discussed in section II.A. Our calculation required $N_b=6000$
bands and a dielectric cutoff of $E_S=35$~Ry to converge the QP gaps to better than $0.05$~eV,
for a total error bar of $\sim 0.1$~eV when combined with the error bar due to k-point sampling.
To test the convergence of the number of bands we calculated QP gaps with a dielectric cutoff
of up to $E_S=45$~Ry and up to $N_b=12000$ bands (Fig.~\ref{fig:QPconverge}).
As Shih \textit{et al.}~\cite{shih10} have noted, the dielectric cutoff and bands are interdependent parameters and
attempting to converge the number of bands using a dielectric cutoff that is too small or converge the
dielectric cutoff using too few bands will result in false convergence. The static remainder technique
speeds up convergence considerably when only a few bands are included, but for a precision
of greater than $0.1$~eV, the convergence with respect to bands for a fixed $E_S$ is about the
same with and without static remainder. The static remainder is still helpful, however,
because when using static remainder, convergence with respect to bands is in the opposite
direction as convergence with respect to $E_S$, resulting in some cancellation of error.
We also self-consistently update the eigenvalues of the Green's function, $G$, when building
the self-energy operator $\Sigma$. We find that going to $G_1W_0$ increases the QP gap at
$K$ by $0.08$~eV compared to $G_0W_0$. Further updating $G$ increases the QP gap at $K$ by
only $0.02$~eV, so we stop at the $G_1W_0$ level. The bandgap is 2.59 eV at the $G_0W_0$ level and
2.67 eV at the $G_1W_0$ level, with spin-orbit interactions included.
We also compare results obtained using the HL-GPP model with the full-frequency dielectric matrix calculated using the contour-deformation approach~\cite{contourdef,contourdef2}. At the $G_0W_0$ level, the full-frequency bandgap is 2.45 eV and increases to 2.54 eV after self-consistently updating the eigenvalues in $G$. Thus, inclusion of the explicit dynamical effects decreases the gap by 0.13 eV compared with the HL-GPP.
We include spin-orbit as a perturbation, and find that the valence band at $K$ is split by $0.15$~eV. The details of the implementation are discussed in section IV.A.3.
\begin{figure}
\includegraphics[width=246.0pt]{Fig4_QPconverge}
\caption{\label{fig:QPconverge} (Color online) Convergence of the QP gap at the M point with respect to the number of bands included in the partial sum for the Coulomb-hole contribution to the self energy, for dielectric cutoffs of $15$(blue), $25$(red), $35$(green), and $45$(magenta)~Ry. The static remainder correction is included. The dashed lines indicate the value of the QP gap extrapolated to infinite bands.}
\end{figure}
\subsection{Results}
The bandstructure of monolayer \MoS2 at the LDA and $G_1W_0$ levels are shown in Fig.~\ref{fig:bandstructure}. We find that monolayer \MoS2 is a direct bandgap material at all levels of theory. The direct gap at the $K$ point increases from $1.71$~eV at the LDA level to $2.59$~eV at the $G_0W_0$ level to $2.67$~eV at the $G_1W_0$ level. The spin-orbit splitting of the valence band at $K$ is $0.15$~eV.
The GW correction varies by k-point. The largest correction to the gap is $1.2$~eV at the $M$ point, and the smallest is $0.96$~eV at the $K$ point. The GW correction also changes the effective masses, making the electron mass smaller than the hole mass. At the LDA level, the electron and hole effective masses at the $K$ point are $0.5m_0$ and $0.6m_0$ respectively. At the $G_1W_0$ level, the electron and hole effective masses are $0.4m_0$ and $0.2m_0$ respectively.
\begin{figure}
\includegraphics[width=246.0pt]{Fig5_bs}
\caption{\label{fig:bandstructure} (Color online) LDA (dashed blue curve) and $G_1W_0$ (solid red curve) band structure of monolayer \MoS2. }
\end{figure}
\subsubsection{\label{sect:qp_comparison}Comparison with other Calculations}
There is significant disagreement on the electronic structure of monolayer \MoS2,
including whether it has a direct or indirect gap, at various levels of theory,
though it is well-known that the experimental gap is direct~\cite{mak10}.
We compare our results with previous GW calculations on monolayer MoS2 in
Table~\ref{tab:gw}. Several calculations~\cite{huser13, shi13} find an indirect gap
from $\Gamma$ to $K$ the $G_0W_0$ level, and Shi \textit{et al.}~\cite{shi13} argue that self-consistently
updating G makes the gap direct. We find a direct gap at the $K$ point at all levels
of theory regardless of k-point sampling and the truncation of the Coulomb interaction.
Different k-points converge with respect to $N_b$ and $E_S$ at different rates, and the
$\Gamma$ point converges much more quickly than the $K$ point, so the indirect gap seen
in some calculations is likely an artifact of a too small dielectric cutoff.
Because the self-energy correction is larger at $\Gamma$ than at $K$, self-consistently
updating $G$ may fortuitously restore the direct gap in those calculations.
Besides convergence, the largest source of differences across previous GW calculations
on monolayer MoS$_2$ is the use of a truncated Coulomb interaction. As discussed in
section~\ref{sect:truncation} and also seen in Refs.~\cite{komsa12,huser13}, not using Coulomb truncation in a calculation with periodic boundary
conditions results in over screening and decreases the QP gap by $100-300$~meV
depending on the supercell size used.
\input{tab_gw_ff}
\section{Optical Properties}
\subsection{Computational Details and Convergence}
\subsubsection{\label{sect:opt_conv}False Convergence and Shift of the Electron-hole Continuum}
As several works have noted, the optical properties of monolayer \MoS2,
as calculated using the Bethe-Salpeter Equation (BSE) formalism,
converge very slowly with respect to k-point sampling~\cite{komsa12,huser13,qiu13}.
In reduced-dimensional systems, the screening varies rapidly as $\mathbf{q}$ approaches the long
wavelength limit (See section II). Excitons at the $K$ point in \MoS2 are highly
localized in momentum space, which means they are extended in real space, so most of the screening comes from the rapidly varying
portion of $\epsilon_\mathrm{2D}(q)$. Hence, convergence with respect to k-point
sampling is slow because it is necessary to resolve the fast changes in spatial dependence in screening.
The extent of the exciton wavefunction in k-space is discussed in greater detail in
Section IV.B.2. We find that a 300x300x1 k-grid is required to converge the exciton
binding energy to within $0.1$~eV (Fig.~\ref{fig:q0}) for the lowest energy state. It is even more demanding for the excited exciton states.
The convergence of the excitation energies with k-point sampling varies depending
on the treatment of the divergent term $W(\mathbf{q}{=}0)$. For semiconductors,
the screened Coulomb interaction $W(\mathbf{q})$ diverges at $\mathbf{q}=0$, and it is common to avoid this
divergence by replacing the screened interaction, $W(\mathbf{q}{=}0)$, with an average
over a small region of the Brillouin zone~\cite{jdeslip11BGW,huser13} near $q=0$. We compare
two different methods of treating the $\mathbf{q}=0$ term. In the first, we average the
screened Coulomb interaction over a small volume in reciprocal space around
$\mathbf{q}=0$. That is, we replace the divergent term, $W_{00}(\mathbf{q}{\rightarrow}0)$, with
\begin{equation}
W_{00}^{\mathrm{avg}}(\mathbf{q})=\frac{N_\mathbf{q}V}{2\pi}\int_\mathrm{cell} \mathrm{d}^2q\; W_{00}(\mathbf{q}),
\end{equation}
where ``cell'' indicates an integral over the volume of the Voronoi cell around
$\mathbf{q}=0$, $N_\mathbf{q}$ is the total number of q-points, $V$ is the
volume of the unit cell in real-space, and $W_{00}$ refers to the divergent
``head'' element, $\mathbf{G}{=}\mathbf{G}'{=}0$.
This averaging treatment results in faster convergence of the excitation energies with k-point
sampling, but the convergence is non-variational -- i.e. the excitation
energy initially increases with k-point sampling (Fig.~\ref{fig:q0}~(a)).
The non-variational convergence occurs because replacing $W(\mathbf{q}{=}0)$
with its average means that a k-point-dependent value is being added to
the diagonal of the BSE matrix, which is equivalent to shifting the
exciton continuum by $W^{\mathrm{avg}}(\mathbf{q}{=}0)$.
We emphasize that, while the widely-used averaging scheme is useful for
improving the convergence of the excitation energies, it may lead
to misleading binding energies, defined as the difference between the
optical gap and the continuum of optical transitions. From Fig.~\ref{fig:q0}, the
excitation energy from a relatively coarse 24x24x1 k-grid appears to
agree better with experiment than finer k-grids, but if the shift
to the continuum energy is taken into account, the binding energy
is only $0.2$~eV. As k-grid sampling increases, the continuum energy
increases linearly with $1/\sqrt{N_\mathbf{k}}$. Even more surprisingly,
the excitation energy varies in a non-uniform way, and increases until we hit
a k-grid finer than about $90$x$90$. For k-grids finer than this, we start
to sample $\mathbf{q}$ vectors before the peak in the quasi-2D dielectric
screening. Because the excitons are fairly spread out in real space, it is
necessary to sample very small wave vectors to capture the small screenings
associated with these lengh scales.
In an alternative treatment of $\mathbf{q}=0$, we fix the exciton continuum at the QP gap
($E_c-E_v$) by setting $W_{00}(\mathbf{q}{=}0)=0$, which is the value of
$W^{\mathrm{avg}}$ in the limit of infinite k-points. In this scheme,
the excitation energies converge slower with respect to k-point sampling,
but the continuum does not move and the convergence is variational.
There is again a kink in the convergence of the excitation energy around
$90$x$90$, which comes from increased sampling in the small $\mathbf{q}$ region.
If we define the binding energy as the difference between the excitation energy
and the onset of the electron-hole or exciton continuum, the binding energy converges at roughly
the same rate regardless of the treatment of $W(\mathbf{q}{=}0)$.
Therefore, even though the comonly-used averaging scheme of the screened Coulomb
interaction typically converges the optical excitation faster, it does so by
moving the continuum of optical excitations and introduces errors in both the
excitonic wave functions and the energies of higher excited exciton states. This is
particularly important if one is interested in properties such as the radius of
the excitonic wave function OR the energies and characters of excited excitonic states.
As a final remark, we note that the fact that the exciton is tightly localized in k-space
reduces the dielectric cutoff, $E_S$, needed to capture the screening for exciton calculations as opposed to those for those for QP energies. As seen
in Fig.~\ref{fig:eps2D}, for $q < \angs{0.1}^{-1}$, the screening is the same for
$E_S=8$~Ry and $E_S=35$~Ry. Indeed, when we reduce the cutoff from $35$~Ry to
$8$~Ry the binding energies of the first $40$ excitonic states change by less
than $10$~meV.
\begin{figure}
\includegraphics[width=246.0pt]{Fig7_q0}
\caption{\label{fig:q0} (Color online) (a) Convergence with respect to k-point
sampling of the exciton continuum (dashed lines) and the 1st (1s) excitation
energy (solid lines) for the A series of excitons when setting $W(q\rightarrow
0)=0$ (red) or using $W^{\mathrm{avg}}(q\rightarrow 0)$ (black). (b)
Convergence of the binding energy, defined as the difference between the
continuum onset and the excitation energy of the 1st exciton in the A series
when setting $W(q\rightarrow 0)=0$ (red) or using
$W^{\mathrm{avg}}(q\rightarrow 0)$ (black). }
\end{figure}
\subsubsection{Computational Details}
In this section, we describe the techniques that allows us to solve the BSE with a very dense k-point sampling and include spin-orbit effects.
In Fig.~\ref{fig:q0}, we explicitly solve the BSE on k-grids with up to
600x600x1 k-points in the full Brillouin zone. However, to save computational
cost, we only included k-points within $\angs{0.2}^{-1}$ of the $K$ point. This
is reasonable for testing convergence, since more than 99\% of states contributing to the lowest energy exciton fall within $\angs{0.1}^{-1}$ of the K point. To obtain the entire optical spectrum, however, it is necessary to consider the entire Brillouin zone using a k-point sampling of at least 300x300x1, which is very computationally demanding.
Rohlfing and Louie~\cite{rohlfing98} originally proposed an interpolation scheme to eliminate this computational bottleneck using two distinct k-grids, a coarse one where the matrix elements for the BSE are calculated, and a fine one onto which the matrix elements are interpolated and on which the BSE Hamiltonian is diagonalized. However, this interpolation scheme is no longer accurate for quasi-2D materials since the dielectric matrix has a lot of structure for small $\mathbf{q}$'s, contrary to the case for bulk systems.
Here, we modify this interpolation scheme to fully capture these fast variations in $\epsilon^{-1}_{00}(\mathbf{q})$ for small $\mathbf{q}$'s. As in the original scheme, we use two k-grids: a fine 300x300x1 k-grid and a coarse 24x24x1 k-grid where we explicitly calculate the BSE matrix elements between all coarse k-points $\mathbf{k}_\mathrm{co}$. However, in addition to these matrix elements, we also calculate transitions from each coarse k-point to a number of fine k-points that form a cluster around each coarse k-point. We call this second set of matrix elements that capture small $\mathbf{q}$'s the \textit{cluster matrix elements}.
When we perform the interpolation of the matrix elements from the coarse to the
fine k-grid, we use the original scheme from Rohlfing and
Louie~\cite{rohlfing98} if a particular transition has a wave vector
$\mathbf{q}=\mathbf{k}_\mathrm{fi}-\mathbf{k}'_\mathrm{fi}$ larger than a given
threshold. Otherwise, we use the cluster matrix element. This interpolation
scheme explicitly captures the fast variation in screening at small
$\mathbf{q}$'s, and the resulting excitation energies of the first 20 exciton
states are within $20$~meV of excitation energies found by explicitly calculating the BSE matrix on a 300x300x1 k-grid.
In addition to k-point sampling, it is also important to consider spin-orbit interactions in the optical absorption spectrum. If one directly solves the BSE on a relativistic basis set that includes spin-orbit interactions, the time to diagonalize the BSE grows by a factor of $64$ compared to the non-relativistic case and would not allow one to use such fine k-point sampling. An alternative scheme to include spin-orbit interactions is therefore desirable.
Our solution is to take advantage of the facts that (1) spin-orbit
splitting is smaller than the exciton binding energy; and (2) spin along the z-axis is a good
quantum number at the $K$ and $K'$ points for monolayer TMDs~\cite{xiao12}. This allows us to efficiently
include spin-orbit effects as a perturbation. We perform both a spin-unpolarized
DFT calculation, which is used as the starting wave functions for our GW calculation, and a
non-collinear calculation with spin-orbit interactions included. We approximate the first-order
spin-orbit correction to the GW quasiparticle energies to be the difference between the two
Kohn-Sham eigenvalues. That is, we take
$\Delta \varepsilon_{\mathrm{GW}}^{\mathrm{SO}}(n\mathbf{k}\sigma) \approx \Delta \varepsilon_{\mathrm{LDA}}^{\mathrm{SO}}(n\mathbf{k}\sigma) \equiv \varepsilon^{\mathrm{non-col}}(n\mathbf{k}\sigma) - \varepsilon^{\mathrm{unpol}}(n\mathbf{k})$,
where $\sigma$ is the spinor index of the states in the non-collinear calculation.
This is a reasonable approximation since the overlaps between the spinor wave functions and
the scalar wave functions are exactly $1$ at $K$ and greater than $0.7$ in other regions
with spin-orbit splitting, in our LDA calculation.
To obtain the absorbance with spin-orbit interaction, we apply a first-order perturbation
theory to the solution of the Bethe-Salpeter equation, which is justifiable because the
quasiparticle gap ($\sim 2.7$~eV) is much larger than the spin-orbit splitting ($\sim 150$~meV).
Each excitonic state $\ket{S}$ can be expanded as a linear combination of pairs of
single-particle valence and conduction band states as
\begin{equation}
\ket{S} = \sum_{vc\mathbf{k}}A^{S}_{vc\mathbf{k}}\ket{vc\mathbf{k}}.
\end{equation}
We want to calculate the spin-orbit corrected exciton energies
$\Omega^{S}_{\sigma} = \Omega^{S} + \Delta\Omega^{S}_{\sigma}$, where $\Omega^{S}$ is the
energy of the $\ket{S}$-state, neglecting spin-orbit, and $\Delta\Omega^{S}_{\sigma}$ is the
first-order energy correction,
\begin{align}
\Delta\Omega^{S}_{\sigma} &\equiv \braket{ S | H^{\mathrm{SO}}_{\sigma} | S } \\
&= \sum_{vc\mathbf{k}}\sum_{v'c'\mathbf{k}'}
\left( A^{S}_{v'c'\mathbf{k}'} \right )^* A^{S}_{vc\mathbf{k}}
\braket{ v'c'\mathbf{k}' | H^{\mathrm{SO}}_{\sigma} | vc\mathbf{k} }, \notag
\end{align}
where the spin-orbit Hamiltonian, $H^{\mathrm{SO}}$, is block-diagonal in the spin-index
$\sigma$ and $H^{\mathrm{SO}}_{\sigma}$ is a block of the spin-orbit Hamiltonian for the spin $\sigma$.
We assume that $H^{\mathrm{SO}}_{\sigma}$ is diagonal in the $\ket{vc\mathbf{k}}$ basis,
which is valid due to the large overlap between the spinor and scalar wave functions.
Then, the spin-orbit correction to the excited-state energies becomes
\begin{equation}
\Delta\Omega^{S}_{\sigma} = \sum_{vc\mathbf{k}}\vert A^{S}_{vc\mathbf{k}}\vert^{2}\Delta \varepsilon^{\mathrm{SO}}_{vc\mathbf{k}\sigma}
\end{equation}
where $\Delta \varepsilon^{\mathrm{SO}}_{vc\mathbf{k}\sigma}$ are the spin-orbit corrected
differences in energy between the valence and conduction states
\begin{align}
\Delta \varepsilon^{\mathrm{SO}}_{vc\mathbf{k}\sigma} &=
(\varepsilon_{\mathrm{GW}}(c\mathbf{k}) + \Delta\varepsilon^{\mathrm{SO}}_{\mathrm{GW}}(c\mathbf{k}\sigma)) \notag \\
&- (\varepsilon_{\mathrm{GW}}(v\mathbf{k}) + \Delta\varepsilon^{\mathrm{SO}}_{\mathrm{GW}}(v\mathbf{k}\sigma)) .
\end{align}
Finally, the imaginary part of the dielectric function with spin-orbit interactions is
calculated using the spin-orbit corrected exciton energies,
\begin{equation}
\varepsilon_{2}(\omega)=\frac{16\pi^{2}e^{2}}{\omega^{2}}\sum_{S\sigma}\vert\mathbf{e}\cdot\bra{0}\mathbf{v}\ket{S\sigma}\vert^{2}\delta(\omega-\Omega^{S}_\sigma) ,
\end{equation}
where $\mathbf{e}$ is the polarization of the incoming light, $\mathbf{v}$ is the
velocity operator, and $\ket{S\sigma}=\ket{S}$.
\subsection{Optical Spectrum}
The absorption spectrum of monolayer \MoS2 with and without electron-hole interactions is shown in Fig.~\ref{fig:absp}. The lowest energy exciton, which forms peak A in the spectrum, has a binding energy of $0.63$~eV. Peaks A and B are spin-orbit split states that arise from excitons forming from transitions between the spin-orbit split valence band maximum and the conduction band minimum at the $K$ and $K'$ points in the Brillouin zone. Both A and B have bright excited states, which we label A', B', etc. The peak A'' overlaps with peak B'. We also see a large peak, which we label peak C, near the continuum onset at $2.7$~eV.
The lowest interband transition energies, i.e. the energies of direct transitions from the valence band to the conduction band throughout the Brillouin zone, are shown in Fig.~\ref{fig:absp}(d). The deepest valleys are parabolic valleys at $K$ and $K'$ points, which give rise to the A and B series of excitons. There is also a shallower Mexican-hat shaped valley around the $\Gamma$ point. Transitions from this Mexican hat valley give rise to peak C and its excited states.
The fine features due to excited states of peaks A and B, which appear in our calculated spectra, are broadened out in the experimental spectra. This is a signature of lifetime effects due to electron-phonon and other interactions. We account for the electron-phonon lifetime effects in our calculation following Marini~\cite{marini08}, and the result is plotted in Fig.~\ref{fig:absp}~(b). We consider both emission and absorption of phonons at $T=300$~K, and we extrapolate the scattering rate for quasiparticle energies larger than those computed by Li \textit{et al.}~\cite{li13}. This leaves the A and B peaks relatively sharp, while broadening out the intermediate peaks between B and C, resulting in excellent agreement with experiment for peak shape and position and the magnitude of the absorbance.
\begin{figure}
\includegraphics[width=246.0pt]{Fig8_absp}
\caption{\label{fig:absp} (Color online) (a) Absorption spectra of \MoS2 without (dashed red curve) and with (solid green curve) electron-hole interactions using a constant broadening of $25$~meV. (b) Same calculated data as in Fig.~\ref{fig:absp}~(a), but using an \textit{ab initio} broadening based on the electron-phonon interactions\cite{marini08,li13}. (c) Experimental absorbance\cite{mak10}. (d) Direct valence to conduction band transition energies in the 1st Brillouin zone.}
\end{figure}
\subsubsection{Comparison with other Calculations}
As with the QP bandgap, there is a wide range of disagreement in the literature about
the binding energy of the exciton giving rise to peak A at the GW-BSE level, with values
ranging an order of magnitude from $0.1-1.1$~eV. A comparison of values obtained in
different calculations is given in Table~\ref{tab:bse}. There is, however, a smaller spread in the
calculated values of the excitation energy of peak A. This is largely because errors
which result in over screening or under screening tend to affect the QP gap and binding
energy in opposite ways, resulting in a cancellation of error in the excitation energy.
The main sources of difference across various BSE calculations in the literature are:
(1) the k-grid sampling, as mentioned in Section~\ref{sect:opt_conv}; and
(2) the truncation of the Coulomb interaction. Coulomb truncation is especially
important to obtain the correct binding energy because, as
seen in Fig.~\ref{fig:eps2D}~(a,b), Coulomb truncation mainly affects screening
in the small-q region where the exciton wavefunction is sensitive because of its localization in k-space. For instance, Refs.~\cite{huser13}
and~\cite{wirtz13} have both noted that very fine k-point sampling is required to converge
the solution of the BSE, yet obtain drastically different results ($0.6$ and $0.15$~eV,
respectively) for the binding energy.
\input{tab_bse}
\subsection{Excitonic Spectrum of Series A and Comparison with Rydberg Series}
\begin{figure}
\includegraphics[width=246.0pt]{Fig8c_absp}
\caption{\label{fig:xct_spec}Comparison of the exciton state energy levels for the A series obtained from \textit{ab initio} GW-BSE calculation (left) with an effective 2D hydrogenic model (right). Bright [dark] exciton states are represented by opaque red [translucent blue] lines.}
\end{figure}
\begin{figure}
\includegraphics[width=246.0pt]{Fig9_xct_v2}
\caption{\label{fig:xct} (Color online) Electron-hole pair amplitudes of lowest energy exciton wavefunctions in reciprocal space for states (a) 1s, (b) 2p, (c) 2s, (d) 3d, (e) 3p, (f) 4f, (g) 3d, (h) 4d and (i) 4p. Each plot is centered around the $K$ point in the Brillouin zone.}
\end{figure}
We can obtain further insight of the structure of the excitonic states by comparing them
to a 2D hydrogenic model. In Fig.~\ref{fig:xct_spec}, we plot the energies of the excitons in the
series A obtained from our GW-BSE calculation with those from an effective 2D hydrogenic model
$H_\mathrm{hydrog} = -\frac{\nabla^2}{2m^*} + \frac{e^2}{\epsilon^* r}$.
This effective model is built by fitting the effective
dielectric constant $\epsilon^*$ to reproduce the binding energy of
peak A. Because there is very little coupling between the $K$ and $K'$ valleys~\cite{qiu15}, the A and B
series of excitons are both doubly degenerate, and so we focus here on the states in the
A series coming from a single valley.
As previously noted\cite{qiu13,ye14,zhu14,chernikov14}, the hydrogenic model deviates from the \textit{ab initio} results in two significant ways:
(1) first, the binding energies of excited states are much larger than expected from a 2D hydrogenic
model; and (2) states with higher angular momentum have a larger binding energy than states with
lower angular momentum. Additionally, there is also some splitting of states with the same
angular momentum, such as $2p$ and $3d$, due to the trigonal warping of the \MoS2 bandstructure
at the $K$ and $K'$ valleys. The $f$ states do not split because they have the same three-fold
symmetry as the bandstructure. Although the excitation energies of the solutions of the BSE
deviate from the hydrogenic model, for simplicity, we still label the states as $1s$, $2s$, $2p$, etc., using
the same notation as a 2D hydrogenic model, based on the number of radial and azimuthal
nodes in the envelope function of the exciton wavefunction.
To understand the physical reasons for these differences between the hydrogenic model and the \textit{ab initio} calculation, we will first analyze the character of the excitonic wave functions and see how the
actual \textit{ab initio} and \textbf{q}-dependent screening differs from the hydrogenic model.
Each excitonic state $\ket{S}$ can be expressed as a linear combination of the electron-hole
transitions $|vc\mathbf{k}\rangle$,
\begin{equation}
\ket{S} = \sum_{vc\mathbf{k}} A^{S}_{vc\mathbf{k}} \ket{vc\mathbf{k}} .
\end{equation}
The coefficients $A^{S}_{vc\mathbf{k}}$ describe the envelope function or electron-hole pair amplitude of the exciton wavefunction in
reciprocal-space. The envelopes of the wavefunctions of the first few states in the $A$ series
of excitons are plotted in Fig.~\ref{fig:xct}. The plots are centered around the $K$ point in the
Brillouin zone. The nodal structure of the envelope function of the states is apparent from this plot. The plots also show
that the excitonic wavefunctions are highly localized in k-space, with most transitions falling
within $\angs{0.1}^{-1}$ or about $5\%$ of the Brillouin zone. In fact, this is well
within the region of fast variation in screening seen in Fig.~\ref{fig:eps2D} and explains
for the most part why convergence with respect to k-point sampling is so slow, since the k-point
sampling must be fine enough to resolve \textit{both} the region before the peak in the dielectric
screening and the nodal structure of the exciton wavefunctions.
The deviations of the results of the \textit{ab initio}
calculation from those of the hydrogenic model may now be understood. If we compare the real-space screening $\epsilon(s)$
with the envelope of the exciton wavefunctions in real space, as shown in
Fig.~\ref{fig:rxct}, it is clear that the varying distribution of the
wavefunction in real-space results in different states experiencing different
screening and therefore different effective electron-hole interaction. In general, states with larger principal quantum number $n$ have a larger binding energy than expected from the hydrogenic model
because they have a larger radius and are thus less screened than states with smaller radii.
Similarly, states with larger angular momentum quantum number are more strongly bound than in the model because
there is a node in the wavefunction where screening is strongest.
Therefore, this effective state-dependent screening explains why
(1) excited excitonic states, such as $2s$ and $3s$, appear lower in energy than
what is predicted by a 2D hydrogenic model (which assumes a constant dielectric
constant), and (2) degenerate states with the same principal quantum number $n$
split, and the excitation energy for states with higher angular momentum is lower.
\begin{figure}
\includegraphics[width=246.0pt]{Fig10_rxct}
\caption{\label{fig:rxct} (Color online) (a) Modulus squared of the exciton wavefunction in real-space for the states 1s (solid blue line), 2s (red line with dash and dot), 2p (green dashed line), and 3d (cyan line with dash and two dots). (b) The effective 2D dielectric function over the same range in real space. }
\end{figure}
\section{Conclusion}
In summary, we find that many-body effects, namely, the electron-electron and electron-hole interactions for quasiparticle and optical excitations, in MoS$_2$ are well-described by the GW-BSE method, which gives results in good agreement with experimental optical spectra and conclusions about the bandgap. We find that, for \MoS2, / results do not differ qualitatively from /, as has been previously claimed. Instead, variations in GW-BSE results in the literature arise largely from different treatments of the long-range Coulomb interaction in periodic supercell calculations and convergence of k-grid sampling and cutoffs for the dielectric screening. We find that truncating the Coulomb interaction to prevent artificial over screening from periodic images is essential to obtain accurate results. The 2D nature of the system also gives rise to strong spatial variations in screening, which must be captured by very fine k-point sampling. The sharpest variation in screening is at small q-vectors (q$\lesssim\pi/d$, where $d$ is the layer thickness), where the screening rapidly vanishes as the wave vector $\mathbf{q}$ approaches zero. Even finer k-point sampling is required to converge the BSE, as the exciton electron-hole amplitude functions in \MoS2 are tightly localized in k-space. Finally, a large energy cutoff for the dielectric matrix is required to capture the spatial variation associated with the different characters of the VBM and CBM of \MoS2, and a correspondingly large number of empty states is required to avoid artificially truncating the dielectric matrix and capture the nearly continuous states arising from using a large vacuum. These are general conclusions that can be applied to GW-BSE calculations on any semiconductor in low dimensions.
We thank L. Yang and T. Cao for discussions. D. Y. Q. acknowledges support from the NSF Graduate Research Fellowship Grant No. DGE 1106400. Structural study and the work on calculating the electron-phonon interaction effects on the optical spectra were supported by NSF Grant No. DMR15-1508412. The GW-BSE calculations were supported by the Theory Program at the Lawrence Berkeley National Lab through the Office of Basic Energy Sciences, U.S. Department of Energy under Contract No. DE-AC02-05CH11231. This research used resources of the National Energy Research Scientific Computing Center, which is supported by the Office of Science of the U.S. Department of Energy, and the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1053575.
|
\section{Introduction}
Exact categories for additive categories were introduced by Quillen \cit
{Quillen}. Barr defined exact categories for non-additive categories \cite{Barr}. He introduced exact categories in order to define a good
notion of non-abelian cohomology.
We will mention regular categories closely related by exact categories. There exist in the literature many different definitions of regular categories, which are all equivalent under the assumptions that finite limits and coequalisers exist. We will recall the following definition from \cite{Tholen}, but a weaker version of it is given in \cite{Barr}.
A category $C$ is called regular if it satisfies the following three properties:
\begin{description}
\item[i)] $C$ is finitely complete,
\item[ii)] If $f:X\longrightarrow Y$ is a morphism in $C$, and
$$ \xymatrix@R=30pt@C=30pt{
Z \ar[d]_-{p_1} \ar[r]^-{p_0}
&X \ar[d]^-{f} \\
X \ar[r]_-{f}
& Y }
$$
\noindent is a pullback $(
then $Z\underset{p_{1}}{\overset{p_{0}}{\rightrightarrows }}X$ is called the kernel pair of $f)$, the coequaliser of
p_{0},p_{1}$ exists;
\item[iii)] If $f:X\longrightarrow Y$ is a morphism in $C$, and
$$ \xymatrix@R=30pt@C=30pt{
W \ar[d]_-{g} \ar[r]^-{}
&X \ar[d]^-{f} \\
Z \ar[r]_-{}
& Y }
$$
\noindent is a pullback, and if $f$ is a regular epimorphism, then $g$ is a regular epimorphism as well.
\end{description}
If a regular category additionally has the property that every equivalence relation is effective that is every equivalence
relation is a kernel pair, then it is called a Barr exact category.
\smallskip
Examples of exact categories are:
$\left( 1\right) $ The category $Set$ of sets.
$\left( 2\right) $ The category of non empty sets.
$\left( 3\right) $ Any abelian category.
$\left( 4\right) $ Every partially ordered set considered as a category.
$\left( 5\right) $ For any small category $C$, the functor category $\left(
\mathcal{C}^{op},Set\right) $.
Brown and Gilbert introduced in \cite{Brown1} the notion of braided regular
crossed module of groupoids and groups as an algebraic model for homotopy
3-types equivalent to Conduch\'{e}'s $2$-crossed module. The reduced case of
braided regular crossed module is called a braided crossed module of groups.
Braided crossed modules in the category of commutative algebras were defined by
Ulualan in \cite{Ulualan}.
The purpose of this paper is to answer the question whether the category of
braided crossed modules of commutative algebras is exact. We prove that the category of braided crossed modules of commutative algebras is an exact category.
\begin{quote}
\textbf{Conventions}
\end{quote}
Throughout this paper $k$ will be a fixed commutative ring with $0\neq 1$.
All $k$-algebras will be commutative and associative.
\section{Braided crossed modules}
In order to give the notion of braided crossed modules of commutative algebras, we will
recall the concept of crossed modules of commutative algebras. Crossed modules of groups
originate in algebraic topology and more particularly in homotopy theory. Mac Lane and Whitehead showed in \cite{Mac} that crossed modules of groups modelled homotopy 2-types (3-types in their notation). The commutative algebra case of crossed modules is contained in
the paper of Lichtenbaum and Schlessinger \cite{LS} and also in the work of
Gerstenhaber \cite{Gers} under different names. Some categorical results and
Koszul complex link are also given by Porter \cite{Porter}.
\begin{definition}\cite{Porter}
A crossed module of commutative algebras, $(C,R,\partial)$, is an $R$-algebra, $C$, together with an $R$-algebra morphism $\partial :C\longrightarrow
R $ such that for all $c,c^{\prime }\in C$
\begin{equation*}
\text{ }\partial \left(
c\right) \cdot c^{\prime }=cc^{\prime }
\end{equation*
where $R$ is a $k$-algebra.
A morphism of crossed modules from $\left( C,R,\partial \right) $ to $\left( C^{\prime },R^{\prime
},\partial ^{\prime }\right) $ is a pair of $k$-algebra morphisms, $\phi :C\longrightarrow C^{\prime }$
and $\psi :R\longrightarrow R^{\prime }$ such that
\begin{equation*}
\left( i\right) \text{ }\partial ^{\prime }\phi =\psi \partial \text{ \
\ \ and \ \ \ }\left( ii\right) \text{ }\phi \left( r\cdot c\right)
=\psi \left( r\right) \cdot \phi \left( c\right)
\end{equation*
for all $r\in R$ and $c\in C$. We thus get the category $\mathsf{XMod}$ of
crossed modules.
There is, for a fixed algebra $R$, a subcategory $\mathsf{XMod/R}$ of the category of crossed modules, which has as objects those crossed modules with $R$ as the
\textquotedblleft base\textquotedblright , i.e., all $\left( C,R,\partial
\right) $ for this fixed $R$, and having as morphisms from $\left(
C,R,\partial _{1}\right) $ to $\left( C^{\prime },R,\partial _{2}\right) $
just those $\left( f_{1},f_{0}\right) $ in $\mathsf{XMod}$ in which
f_{0}:R\longrightarrow R$ is the identity homomorphism on $R$.
\end{definition}
\begin{definition}\cite{Ulualan}
A braided crossed module of commutative algebras $\partial :C\rightarrow R$ is a
crossed module with the braiding function $\left\{ -,-\right\} :R\times
R\longrightarrow C$ satisfying the following axioms:
\begin{array}{ll}
BCM1) & \partial \left\{ r,r^{\prime }\right\} =rr^{\prime }
\end{array
$
\begin{array}{ll}
BCM2) & \left\{ \partial c,\partial c^{\prime }\right\} =cc^{\prime }
\end{array
$
\begin{array}{lllll}
BCM3) & \left\{ \partial c,r\right\} =r\cdot c & \text{and} & \left\{
r,\partial c\right\} =r\cdot c,&
\end{array
$
\begin{array}{ll}
BCM4) & \left\{ rr^{\prime },r^{\prime \prime }\right\} -\left\{ r,r^{\prime
}r^{\prime \prime }\right\} =0
\end{array
$
{\noindent}for all $r,r^{\prime },r^{\prime \prime }\in R$ and $c,c^{\prime
}\in C$.
We denote such a braided crossed module of commutative algebras by $\left\{
C,R,\partial \right\} $.
If $\left\{ C,R,\partial _{1}\right\} $ and $\left\{ C^{\prime },R^{\prime
},\partial _{2}\right\} $ are braided crossed modules, a morphism$\
\begin{equation*}
\left( f_{1},f_{0}\right) :\left\{ C,R,\partial _{1}\right\} \longrightarrow
\left\{ C^{\prime },R^{\prime },\partial _{2}\right\} ,
\end{equation*
of braided crossed modules is given by a morphism of crossed modules such
that
\begin{equation*}
\left\{ -,-\right\} \left( f_{0}\times f_{0}\right) =f_{1}\left\{
-,-\right\} .
\end{equation*}
We thus get the category $\mathsf{BXMod}$ of braided crossed modules of commutative algebras.
In the case of a morphism $\left( f_{1},f_{0}\right) $ between braided
crossed modules with the same base $R$, i.e. where $f_{0}$ is the identity on $R
$ with $f_{1}\left\{ -,-\right\} =\left\{ -,-\right\} $, then we say that
f_{1}$ is a morphism of braided crossed $R$-modules, $\partial _{1}:C\longrightarrow R$ is a braided crossed $R$-module and we use $\left\{C,\partial _{1} \right\} $ instead of $\left\{C,R,\partial _{1} \right\} $. This gives a
subcategory $\mathsf{BXMod/R}$ of $\mathsf{BXMod}$. Our results are obtained for this subcategory.
\end{definition}
Several well known examples of crossed modules give rise to braided crossed modules as follows.
\textbf{Examples of braided crossed modules}
\noindent $\left( 1\right) $ Any identity map of $k$-algebras $\partial
:X\longrightarrow X$ is a braided crossed module with $\left\{ x,y\right\}
=xy$.
\noindent $\left( 2\right) $ If $C$ is a $k$-algebra and $C^{2}$ is an ideal generated by $\left\{ c_{1}c_{2}\mid c_{1},c_{2}\in C\right\} $. Then $\partial :C^{2}\longrightarrow C$ is a braided
crossed module with $\left\{ c_{1},c_{2}\right\} =c_{1}c_{2}$, for
c_{1},c_{2}\in C$.
\noindent $\left( 3\right) $ Any $R$-module $M$ can be considered as an $R
-algebra with zero multiplication and hence the zero morphism
0:M\longrightarrow R$ is a braided crossed module with $\left\{ r,r^{\prime
}\right\} =0$.
\noindent $\left( 4\right) $ Let $\left\{ C,R,\partial _{1}\right\} $ and
\left\{ C^{\prime },R^{\prime },\partial _{2}\right\} $ be two braided
crossed modules, then $\left\{ C\times C^{\prime },R\times R^{\prime
},\partial \right\} $ is a braided crossed module.
\section{The Barr exactness property of braided crossed modules}
Our aim now is to obtain that $\mathsf{BXMod/R}$ is an exact category. For
this purpose, we have to prove some statements in this section.
\begin{proposition} \label{prop3}
In $\mathsf{BXMod/R}$ every pair of morphisms with common domain and
codomain has an equaliser.
\end{proposition}
\begin{proof}
Let $f,g:\left\{ C,\partial \right\} \longrightarrow \left\{ D,\delta
\right\} $ be two morphisms of braided crossed $R$-modules. Let $E$ denotes the
set $E=\left\{ c\in C\mid f\left( c\right) =g\left( c\right) \right\} .$ It can be easily checked that $\left\{ E,\varepsilon \right\} $ has the structure of
a braided crossed $R$-module, and the inclusion $u:\left\{ E,\varepsilon
\right\} \longrightarrow \left\{ C,\partial \right\} $ is a morphism of
braided crossed $R$-modules and clearly $fu=gu$.
Suppose that there exist a braided crossed $R$-module, $\left\{ E^{\prime
},\varepsilon ^{\prime }\right\} $ and a morphism $u^{\prime }:\left\{
E^{\prime },\varepsilon ^{\prime }\right\} \longrightarrow \left\{
C,\partial \right\} $ of braided crossed $R$-modules such that $fu^{\prime
}=gu^{\prime }.$ Then for all $x\in E^{\prime },$ $f\left( u^{\prime }\left(
x\right) \right) =g\left( u^{\prime }\left( x\right) \right) ,$ and hence
u^{\prime }\left( x\right) \in E.$ Thus, we have $\alpha :\left\{ E^{\prime
},\varepsilon ^{\prime }\right\} \longrightarrow \left\{ E,\varepsilon
\right\} $ by $\alpha \left( x\right) =u^{\prime }\left( x\right) $ from which we get
\begin{equation*}
\varepsilon \alpha \left( x\right) =\varepsilon u^{\prime }\left( x\right)
=\partial u^{\prime }\left( x\right) =\varepsilon ^{\prime }\left( x\right),
\end{equation*}
\begin{equation*}
\alpha \left( r\cdot x\right) =u^{\prime }\left( r\cdot x\right) =r\cdot
u^{\prime }\left( x\right)=r\cdot \alpha \left( x\right),
\end{equation*}
for all $x \in E^{\prime }, r \in R$. Since $u$, $u^{\prime }$ are braided crossed $R$-module morphisms, it is
clear that
$\alpha \left\{ r,r^{\prime }\right\} =\left\{ r,r^{\prime
}\right\} $ for all $r, r^{\prime } \in R$.
Let $\alpha ^{\prime }:E^{\prime }\longrightarrow E$ be a morphism of
braided crossed $R$-module such that $u\alpha ^{\prime }=u^{\prime }.$ Since
$\alpha \left( x\right) =u^{\prime }\left( x\right)
=u\alpha ^{\prime }\left( x\right) =\alpha ^{\prime }\left( x\right)$ for all $x\in E^{\prime }$,
we have that $\alpha$ is the unique morphism which makes the diagram
\begin{equation*}
\xymatrix@R=40pt@C=40pt{
\left\{ E,\varepsilon\right\} \ar@{^{(}->}[r]^-{u}& \left\{ C,\partial \right\} \ar@{->}@<2pt>[r]^-{f}
\ar@{->}@<-2pt> [r]_-{g} & \left\{ D,\delta\right\} \\ \left\{ E^{\prime},\varepsilon ^{\prime }\right\} \ar@{.>}[u]_{\alpha} \ar[ur]_{u^{\prime }} }
\end{equation*}
\noindent commutative. Hence $u$ is the equaliser of $\left( f,g\right),$ as required.
\end{proof}
\begin{proposition} \label{prop4}
$\mathsf{BXMod/R}$ has finite products.
\end{proposition}
\begin{proof}
Let $\left\{ C,\partial \right\} $ and $\left\{ D,\delta \right\}$ be braided crossed $R$-modules. The product $ \left\{ C\sqcap D,\tau \right\} $ is the pullback over the
terminal object $\left\{ R,i_{R}\right\} ,$
\begin{equation*}
\xymatrix@R=40pt@C=40pt{
\left\{ C\sqcap D,\tau \right\} \ar[d]_{\delta ^{\prime }} \ar[r]^{\partial ^{\prime }}
&\left\{ C,\partial \right\} \ar[d]^{\partial} \\
\left\{ D,\delta \right\} \ar[r]_{\delta}
& \left\{ R,i_{R}\right\} }
\end{equation*}
\noindent where $C\sqcap D=\left\{ \left( c,d\right) \mid \partial \left( c\right) =\delta
\left( d\right) \right\} $ and $\tau :C\sqcap D\longrightarrow R$ is defined
by $\tau=\partial \partial^{\prime }=\delta \delta ^{\prime }.$ Then by induction, $\mathsf{BXMod/R}$
has finite products.
\end{proof}
\begin{proposition}
\label{crll} $\mathsf{BXMod/R}$ is finitely complete.
\end{proposition}
\begin{proof}
Follows from Propositions \ref{prop3} and \ref{prop4}.
\end{proof}
\begin{proposition}
\label{prp2} In $\mathsf{BXMod/R}$ every morphism has a kernel pair, and the
kernel pair has a coequaliser.
\end{proposition}
\begin{proof}
Let $\left\{ A,\partial \right\} $ and $\left\{ B,\beta \right\} $ be two
braided crossed $R$-modules. Let $f:\left\{ A,\partial \right\}
\longrightarrow \left\{ B,\beta \right\} $ be a morphism of braided crossed
R$-modules. Then $\left( A,f\right) $ is a crossed $B$-module, where $B$
acts on $A$ via $\beta $ and the homomorphism $\alpha :A\times _{B}A\longrightarrow B$ defined by $\alpha \left( a,a^{\prime }\right) =f\left( a\right) =f\left(
a^{\prime }\right) $ is a crossed $B$-module, where $A\times _{B}A=\left\{
\left( a,a^{\prime }\right) \mid f\left( a\right) =f\left( a^{\prime
}\right) \right\} $ is a $B$-algebra:
\begin{array}{ll}
& \alpha \left( b\cdot \left( a,a^{\prime }\right) \right) =f\left( b\cdot
a\right) =b\cdot f\left( a\right) =b\cdot \alpha \left( a,a^{\prime }\right),
\end{array
$ \smallskip
\begin{array}{ll}
& \alpha \left( a,a^{\prime }\right) \cdot \left( a_{1},a_{1}^{\prime
}\right) =\left( f\left( a\right) \cdot a_{1},f\left( a^{\prime }\right)
\cdot a_{1}^{\prime }\right) =\left( a,a^{\prime }\right) \left(
a_{1},a_{1}^{\prime }\right),
\end{array
$
{\noindent }for all $\left( a,a^{\prime }\right) ,\left( a_{1},a_{1}^{\prime
}\right) \in A\times _{B}A, b \in B.$
We can define $\partial ^{\prime }:A\times _{B}A\longrightarrow R$ by
\partial ^{\prime }\left( a,a^{\prime }\right) =\partial \left( a\right)
=\partial \left( a^{\prime }\right) $, since $\beta \alpha =\beta f=\partial
$. It is easily checked that $\left( A\times _{B}A,R,\partial ^{\prime }\right)
$ is a crossed module:
\begin{array}{ll}
& \partial ^{\prime }\left( r\cdot \left( a,a^{\prime }\right) \right)
=\partial \left( r\cdot a\right) =r\cdot \partial \left( a\right) =r\cdot
\partial ^{\prime }\left( a,a^{\prime }\right),
\end{array
$
\begin{array}{llllll}
& \partial ^{\prime }\left( a,a^{\prime }\right) \cdot \left(
a_{1},a_{1}^{\prime }\right) & = & \partial \left( a\right) \cdot \left(
a_{1},a_{1}^{\prime }\right) & = & \left( \partial \left( a\right) \cdot
a_{1},\partial \left( a\right) \cdot a_{1}^{\prime }\right) \\
& & & & = & \left( \partial \left( a\right) \cdot a_{1},\partial \left(
a^{\prime }\right) \cdot a_{1}^{\prime }\right) \\
& & & & = & \left( aa_{1},a^{\prime }a_{1}^{\prime }\right) \\
& & & & = & \left( a,a^{\prime }\right) \left( a_{1},a_{1}^{\prime
}\right),
\end{array
$
{\noindent }for all $\left( a,a^{\prime }\right) ,\left( a_{1},a_{1}^{\prime
}\right) \in A\times _{B}A$, $r \in R$.
Below we will show that $\left\{ A\times _{B}A,\partial ^{\prime }\right\} $ is a braided
crossed $R$-module with the braiding map
\begin{equation*}
\left\{ -,-\right\} :R\times R\longrightarrow A\times _{B}A
\end{equation*
defined by $\left\{ r,r^{\prime }\right\} =\left( \left\{ r,r^{\prime }\right\}
,\left\{ r,r^{\prime }\right\} \right) $,
\begin{array}{ll}
\mathbf{BCM1)} & \partial ^{\prime }\left\{ r,r^{\prime }\right\}=\partial \left\{ r,r^{\prime }\right\}=rr^{\prime }
\end{array
$
\begin{array}{llll}
\mathbf{BCM2)} & \left\{ \partial ^{\prime }\left( a,a^{\prime }\right)
,\partial ^{\prime }\left( b,b^{\prime }\right) \right\} & = & \left(
\left\{ \partial a,\partial b\right\} ,\left\{ \partial a^{\prime },\partial
b^{\prime }\right\} \right) \\
& & = & \left( ab,a^{\prime }b^{\prime }\right) \\
& & = & \left( a,a^{\prime }\right) \left( b,b^{\prime }\right),
\end{array
$
\begin{array}{lrll}
\mathbf{BCM3)} & \left\{ \partial ^{\prime }\left( a,a^{\prime }\right)
,r\right\} & = & \left( \left\{ \partial a,r\right\} ,\left\{ \partial
a^{\prime },r\right\} \right) \\
& & = & \left( r\cdot a,r\cdot a^{\prime }\right) \\
& & = & r\cdot \left( a,a^{\prime }\right) \\
& \left\{ r,\partial ^{\prime }\left( a,a^{\prime }\right) \right\} & = &
\left( \left\{ r,\partial a\right\} ,\left\{ r,\partial a^{\prime }\right\}
\right) \\
& & = & \left( r\cdot a,r\cdot a^{\prime }\right) \\
& & = & r\cdot \left( a,a^{\prime }\right),
\end{array
$
\begin{array}{llll}
\mathbf{BCM4)} & \left\{ rr^{\prime },r^{\prime \prime }\right\} -\left\{
r,r^{\prime }r^{\prime \prime }\right\} & = & \left( \left\{ rr^{\prime
},r^{\prime \prime }\right\} ,\left\{ rr^{\prime },r^{\prime \prime
}\right\} \right) -\left( \left\{ r,r^{\prime }r^{\prime \prime }\right\}
,\left\{ r,r^{\prime }r^{\prime \prime }\right\} \right) \\
& & = & \left( \left\{ rr^{\prime },r^{\prime \prime }\right\} -\left\{
r,r^{\prime }r^{\prime \prime }\right\} ,\left\{ rr^{\prime },r^{\prime
\prime }\right\} -\left\{ r,r^{\prime }r^{\prime \prime }\right\} \right)
\\
& & = & \left( 0,0\right),
\end{array
$\bigskip
{\noindent }for all $\left( a,a^{\prime }\right) ,\left( b,b^{\prime }\right) \in
A\times _{B}A$, $r,r^{\prime },r^{\prime \prime }\in R$.
\\The following diagram
$$ \xymatrix@R=40pt@C=40pt{
A\times _{B}A\ar@{->}@<2pt>[r]^-{p_1}
\ar@{->}@<-2pt> [r]_-{p_2} \ar[d]_-{\partial ^{\prime }}& \ A \ar[r]^-{f} \ar[d]^-{\partial}
& B\ar[d]^-{\beta} \\R\ar@{=}[r]^{} & R\ar@{=}[r]^{} & R }
$$
{\noindent }commutes and the morphisms $p_{1}$ and $p_{2}$ above are
morphisms of braided crossed $R$-modules.
This construction satisfies
universal property: Let
\{E,\delta\} $ be a braided crossed $R$-module and $p_{1}^{\prime
},p_{2}^{\prime }:\{E,\delta\}\longrightarrow \{A,\partial
\} $ be any morphisms of braided crossed $R$-modules with
fp_{1}^{\prime }=fp_{2}^{\prime },$ then there exist a unique morphism
\begin{equation*}
h:\{E,\delta\} \longrightarrow \{
A\times _{B}A,\partial ^{\prime }\}
\end{equation*
given by $h\left( e\right) =\left( p_{1}^{\prime }\left( e\right)
,p_{2}^{\prime }\left( e\right) \right) ,$ for all $e\in E,$ which makes the diagram
$$ \xymatrix@R=35pt@C=35pt{
\{E,\delta\} \ar@/_/[ddr]_{p_{2}^{\prime }}
\ar@/^/[drr]^{p_{1}^{\prime }}
\ar@{.>}[dr]|-{h} \\
&\{A\times _{B}A,\partial ^{\prime }\} \ar[d]^-{p_{2}}
\ar[r]_-{p_{1}}
& \{A,\partial\} \ar[d]^-{f} \\
& \{A,\partial\}\ar[r]_-{f} &\{B,\beta\} } $$
{\noindent }commutative.
\\Then $\left( p_{1},p_{2}\right) $ is the kernel pair of
the morphism $f$.
Now we will show that the pair $\left( p_{1},p_{2}\right) $ has a
coequaliser. Let $I$ be an ideal of $A$ generated by
all the elements of the form $p_{1}\left( x\right) -p_{2}\left( x\right) ,$
for all $x=\left( a,a^{\prime }\right) \in A\times _{B}A$. We will define the braided crossed $R$-module $\delta
:A/I\longrightarrow R$ by $\delta \left( a+I\right) =\partial \left(
a\right) ,$ $\left\{ r,r^{\prime }\right\} =\left\{ r,r^{\prime }\right\} +I$
for $a\in A,$ $r,r^{\prime }\in R$. Since $I\subseteq
Ker\partial $, $\delta $ is well defined. By the definition of $\left\{
A/I,\delta \right\} $, the morphism $q:\left\{ A,\partial \right\}
\longrightarrow \left\{ A/I,\delta \right\} $ is the induced projection and
it is a morphism of braided crossed $R$-modules, i.e., the diagram
$$ \xymatrix@R=40pt@C=40pt{
A\times _{B}A \ar@{->}@<2pt>[r]^-{p_1}
\ar@{->}@<-2pt> [r]_-{p_2} \ar[dr]_-{\partial ^{\prime }}& \ A \ar[r]^-{q} \ar[d]^-{\partial}& A/I\ar[dl]^-{\delta} \\ & R & }
$$
{\noindent }commutes.
Suppose there exist a braided crossed $R$-module $\left\{ A^{\prime
},\alpha ^{\prime }\right\} $ and a morphism of braided crossed $R$-modules
q^{\prime } :\left\{ A,\partial \right\}
\longrightarrow \left\{ A^{\prime },\alpha ^{\prime }\right\} $ such that
q^{\prime }p_{1}=q^{\prime }p_{2},$ then there exists a unique morphism $
\varphi :\left\{ A/I,\alpha \right\} \longrightarrow \left\{
A^{\prime },\alpha ^{\prime }\right\} $ defined by $\varphi \left( a+I\right)
=q^{\prime }\left( a\right) $, satisfying $\varphi q=q^{\prime },$ i.e. $q$
is the universal among all the morphisms $q^{\prime } :\left\{ A,\partial \right\} \longrightarrow \left\{ A^{\prime },\alpha
^{\prime }\right\} $ for any braided crossed $R$-module $\left\{ A^{\prime
},\alpha ^{\prime }\right\} .$ So we get the following commutative diagram:
$$ \xymatrix@R=40pt@C=40pt{
A\times _{B}A \ar@{->}@<2pt>[r]^-{p_1}
\ar@{->}@<-2pt> [r]_-{p_2} \ar[dr]_-{\partial ^{\prime }}& \ A \ar[r]^-{q} \ar[d]^-{\partial} \ar[dr]|(0.7){q^{\prime }}& A/ I\ar[dl]|(0.7){ \ \ \ {\delta} \ \ \ } \ar@{.>}[d]^{\varphi}\\ & R & A^{\prime }\ar[l]^{\delta ^{\prime }} }
$$
Then $q$ is the coequaliser of the pair $\left( p_{1},p_{2}\right) .$
Therefore in $\mathsf{BXMod/R}$, the kernel pair of every morphism exists
and has a coequaliser.
\end{proof}
\begin{proposition}
\label{prp1}In $\mathsf{BXMod/R}$ every regular epimorphisms are stable under pullback.
\end{proposition}
\begin{proof}
Let $\Phi :\left\{ C,\mu \right\} \longrightarrow \left\{ D,\sigma \right\} $
be a regular epimorphism in $\mathsf{BXMod/R,}$ which means that
C\longrightarrow D$ is a regular epimorphism of $k$-algebras. Let $\eta
:\left\{ G,\theta \right\} \longrightarrow \left\{ D,\sigma \right\} $ be
any morphism in $\mathsf{BXMod/R.}$ The pullback of $\Phi $ along $\eta $ is
$\Phi ^{\ast }:\left\{ C\times _{D}G,\rho \right\} \longrightarrow \left\{
G,\theta \right\} $ given by $\Phi ^{\ast }\left( c,g\right) =g$ where
C\times _{D}G=\left\{ \left( c,g\right) \mid \Phi \left( c\right) =\eta
\left( g\right) \right\} $ and $\rho :C\times _{D}G\longrightarrow R$ is a
braided crossed $R$-module defined by $\rho \left( c,g\right) =\mu \left(
c\right) =\theta \left( g\right) $ with $\left\{ r,r^{\prime }\right\}
=\left( \left\{ r,r^{\prime }\right\} ,\left\{ r,r^{\prime }\right\} \right)
$, for all $r,r^{\prime }\in R,$ $c\in C,$ $g\in G.$ Since in the category
of $k$-algebras the regular epimorphisms are characterised as the surjective
homomorphisms, these are closed in this way under pullback. Thus $\Phi
^{\ast }$ is a surjective homomorphism.
\begin{equation*}
\xymatrix@R=40pt@C=40pt{
\left\{ C\times
_{D}G,\rho \right\} \ar[d]_{\Phi ^{\ast }} \ar[r]^{}
&\left\{ C,\mu \right\} \ar[d]^{\Phi} \\
\left\{ G,\theta \right\} \ar[r]_{\eta}
& \left\{ D,\sigma \right\} }
\end{equation*}
\noindent We claim that the surjective morphism $\Phi
^{\ast }$ is a regular epimorphism, that is $\Phi ^{\ast }$ is the coequaliser of a pair of morphisms. Define
\begin{equation*}
E=\left\{ \left( x,y\right) \in \left( C\times _{D}G\right) \times \left(
C\times _{D}G\right) \mid \Phi^{\ast } \left( x\right) =\Phi^{\ast } \left( y\right)
\right\}
\end{equation*
and let $\left\{ E,\alpha \right\} \underset{q}{\overset{p}{\rightrightarrows }
\left\{ C\times _{D}G,\rho \right\} $ be the first and second projections. Since $G$ is isomorphic to the quotient of $C\times
_{D}G$, $\Phi ^{\ast }$ is the coequaliser of $p
$ and $q$. Thus we get that if $\Phi $ is regular epimorphism, so is $\Phi ^{\ast }$, as required.
\end{proof}
\begin{theorem}
\label{theo} $\mathsf{BXMod/R}$ is regular.
\end{theorem}
\begin{proof}
The proof is a direct consequence of Propositions \ref{crll}, \ref{prp2} and \ref{prp1}.
\end{proof}
Now we recall the definition of equivalence relation from \cite{Wells}.
\begin{definition}
Let $A$ be a category with finite limits. If $A$ is an object, a subobject
\left( d^{0},d^{1}\right) :E\longrightarrow A\times A$ is called an
equivalence relation if it is
\noindent $\mathbf{ER1.}$ Reflexive: There is an arrow $r:A\longrightarrow E$ such that
d^{0}r=d^{1}r=id_{A};$
\noindent $\mathbf{ER2.}$ Symmetric: There is an arrow $s:E\longrightarrow E$ such that
d^{0}s=d^{1}$ and $d^{1}s=d^{0};$
\noindent $\mathbf{ER3.}$ Transitive: If
\begin{equation*}
\xymatrix@R=30pt@C=30pt{
T \ar[d]_{q_{2}} \ar[r]^{q_{1}}
&E \ar[d]^{d^{0}} \\
E \ar[r]_{d^{1}}
& A }
\end{equation*}
\noindent is a pullback, there is an arrow $t:T\longrightarrow E$ such that
d^{1}t=d^{1}q_{1}$ and $d^{0}t=d^{0}q_{2.}$
\end{definition}
\begin{proposition}
\label{prp3} Every equivalence relation
$$ \xymatrix@R=40pt@C=40pt{
\{ E,\partial\}\ar@{->}@<2pt>[r]^-{u}
\ar@{->}@<-2pt> [r]_-{v}& \ \{A,\alpha\} }
$$
{\noindent }in the category of braided crossed $R$-modules is effective
\end{proposition}
\begin{proof}
Let $A/E$ be the set of all equivalence classes $\left[ a\right] $ with
respect to $E,$ i.e., $\left[ a\right] =\left\{ b\in A\mid \left( a,b\right)
\in E\right\} .$ $A/E$ has the structure of an $R$-algebra. $\overline
\alpha }:A/E\longrightarrow R$ induced by $\alpha $ is well defined since
\alpha u=\alpha v.$ We can form the braided crossed $R$-module $\overline
\alpha }$ with braiding map $\left\{ r,r^{\prime }\right\} =\left[ \left\{
r,r^{\prime }\right\} \right] $ and get the following diagram,
$$ \xymatrix@R=40pt@C=40pt{
E \ar@{->}@<2pt>[r]^-{u}
\ar@{->}@<-2pt> [r]_-{v} \ar[dr]_-{\partial}& \ A \ar[r]^-{q} \ar[d]^-{\alpha}
& A/ E\ar[dl]^-{\overline{\alpha }} \\ & R & }
$$
{\noindent }of morphisms of braided crossed $R$-modules, where $q$ is the
projection onto $A/E$. By the definition of an equivalence relation on $A$, we have $\left( u\left(
x\right) ,v\left( x\right) \right) \in E,$ for $x\in E$. Since
\begin{equation*}
E\subseteq A\times _{R}A=\left\{ \left( a,a^{\prime }\right) \mid \alpha
\left( a\right) =\alpha \left( a^{\prime }\right) \right\} ,
\end{equation*}
{\noindent we get }$\alpha u\left( x\right) =\alpha v\left( x\right) $, thus
$\left( 0,u\left( x\right) -v\left( x\right) \right) \in E,$ therefore $qu=qv
$. Suppose that there exist a braided crossed $R$-module $\left\{ D,\omega
\right\} $ with $u^{\prime },v^{\prime }:\left\{ D,\omega \right\}
\longrightarrow \left\{ A,\alpha \right\} $ such that $qu^{\prime
}=qv^{\prime },$ so $\left[ u^{\prime }\left( d\right) \right] =\left[
v^{\prime }\left( d\right) \right] ,$ i.e. $\left( u^{\prime }\left(
d\right) ,v^{\prime }\left( d\right) \right) \in E,$ and therefore there
exists a unique morphism $\theta :\left\{ D,\omega \right\} \longrightarrow
\left\{ E,\partial \right\} ,$ such that the diagram
$$ \xymatrix@R=35pt@C=35pt{
\{D,\omega\}\ar@/_/[ddr]_{v^{\prime }} \ar@/^/[drr]^{u^{\prime }}
\ar@{.>}[dr]|-{ \ \ {\theta} \ \ } \\
& \{E,\partial\}\ar[d]^-{v} \ar[r]_-{u}
& \{A,\alpha\} \ar[d]^-{q} \\
& \{A,\alpha\}\ar[r]_-{q} & \{A/E,\overline{\alpha }\}
} $$
{\noindent }commutes.
Thus $\left( u,v\right) $ is the kernel pair of a morphism $q:\left\{
A,\alpha \right\} \longrightarrow \left\{ A/E,\overline{\alpha }\right\} $
in the category of braided crossed $R$-modules.
\end{proof}
\begin{theorem}
The category $\mathsf{BXMod/R}$ of braided crossed $R$-modules is a Barr exact category.\label{th}
\end{theorem}
\begin{proof}
The conditions of exact category for $\mathsf{BXMod/R}$ are satisfied by Theorem \ref{theo} and Proposition \ref{prp3}
, which completes the proof.
\end{proof}\newline
\textbf {Acknowledgments}
The first author was partially supported by T\"{U}B\.{I}TAK (The Scientific and
Technological Research Council Of Turkey).
|
\section{Introduction}
The Rydberg blockade mechanism introduced in \cite{Jaksch2000} has been demonstrated to be capable of creating bipartite entanglement with fidelity of $\sim 0.7 - 0.8$\cite{Maller_PRA_92_022336,Jau2016}. There is good reason to believe that the fidelity achieved to date is not a fundamental limit, but is due to experimental perturbations and the high sensitivity of Rydberg states to external fields \cite{Saffman2016}. With the expectation that experimental techniques will continue to improve it is important to address the question of the intrinsic fidelity limit of the Rydberg blockade gate. Detailed analysis with constant amplitude Rydberg excitation pulses revealed a Bell state fidelity limit of $F_B\sim 0.999$ in Rb or Cs atoms in a 300 K environment\cite{Zhang_PRA_85_042310}. Other work has sought to improve on this with optimal control pulse shapes\cite{Muller2011,Goerz_PRA_90_032329}, adiabatic excitation\cite{Muller_PRA_89_032334,Rao2014}, or simplified protocols that use a single Rydberg pulse\cite{Han2016,Su2016}. However, none of the analyses to date that consistently account for Rydberg decay and excitation leakage to neighboring Rydberg states have provided a fidelity better than 0.999. This leaves open the question of whether or not the Rydberg gate will be capable of reaching the $0.9999$ level or better that appears necessary for scalable quantum computation with a realisitc overhead in terms of qubit numbers for logical encoding\cite{Devitt2013}.
In this work we show that Rydberg gates with $F_B>0.9999$ are possible with Cs atoms in a 300 K environment and $F_B>0.99999$ in a 4K environment. This advance is made possible using simple and smooth analytic shaped pulses that are designed to suppress leakage at a discrete set of frequencies\cite{Motzoi_PRL_103_110501} corresponding to neighboring Rydberg states. By suppressing the leakage orders of magnitude more effectively than is possible with square, or simple Gaussian pulses, we are able to run the gate at least an order of magnitude faster than previous protocols, which is fast enough to keep the spontaneous emission error low and achieve high fidelity. Drastically reducing the gate time also has advantages in the short term, by avoiding the onset of other experimental errors that increase with time, such as technical noise. We find a gate time close to 50 ns, which is fast enough to be competitive with superconducting qubits while retaining much longer coherence times \cite{Egger_SUST_2013,Ghosh_PRA_87_022309}.
\begin{figure}
\includegraphics[width=.98\linewidth]{pulsesequence_new4.pdf}
\caption{\label{fig:sequence}(color online) a) DRAG pulse sequence (blue) and initial Gaussian waveform, (thin black) to implement a two-qubit entangling gate. Pulse durations are $\tau_c, \tau_t$ for the control and target atoms. The control amplitudes are shown on the same scale. b) Level diagram for one-photon Rydberg excitation with laser frequency $\omega_d$. c) Detail of the Rydberg level structure and detunings. d) Spectrum of pulse on control atom for Gaussian (black) and DRAG (blue) waveforms.}
\end{figure}
\section{Rydberg excitation} The free evolution and gate Hamiltonians $\op{H}_d$ and $\op{H}_g$, respectively, of a single Rydberg atom in its lab frame, are given by ($\hbar=1$ everywhere)
\begin{subequations}\label{eq:ham_lab}\begin{align}
\op{H}_{d} & = \omega_g\ketbra{g}{g}+\omega_q\ketbra{1}{1} + \sum\limits_{r'}\omega_{r'}\ketbra{r'}{r'}\\
\op{H}_{g} & = \Omega(t)\sum\limits_{r'}\left(\frac{n}{n'}\right)^{3/2}\left(\ketbra{r'}{0}+\ketbra{r'}{1}\right) + \text{h.c.}
\end{align}\end{subequations}
whereby $\ket{g}$ denotes some auxiliary level we will use to model decay. Here $r'$ is shorthand for the set of quantum numbers specifying the Rydberg states and $n,n'$ are the principal quantum numbers. The matrix elements and the Rabi coupling for single photon excitation to high lying Rydberg states scale as $1/n^{3/2}$. For Cs, the ground hyperfine splitting is $\omega_q/2\pi=9.1926\;\rm{GHz}$. The set of states $\{\ket{r'}\}$ describes all relevant Rydberg states. We assume there is negligible coupling of any of the states to $\ket{g}$ due to the control $\Omega(t)$, hence without loss of generality we set its energy to zero, i.e. $\omega_g=0$. The control field has in-phase control only,
\begin{align}\label{eq:control_field}
\Omega(t) & = \varepsilon_x(t)\cos{\omega_d t}.
\end{align}
Usually, atoms are driven on resonance with the $\ket{1}\leftrightarrow\ket{r}$ transition, so that $\omega_d=\omega_r-\omega_1$ with $\omega_r$ being the frequency of the target Rydberg state $\ket{r}$. In order to remove any oscillation on the order of $\omega_d$ from the dynamics, we choose to work in a frame rotating with $\omega_d$ in the remainder of this work. The pulse sequence which we will use to implement a two-qubit entangling gate is illustrated in \reffig{sequence}a). When control and target atoms are initially prepared in their $\ket{1}$ state, the desired Rydberg state $\ket{r}$ of the target atom will be Rydberg-blockaded by ${\sf B}_0$ during the $2\pi$-pulse due to the control atom's $\ket{r}$ state being populated. Hence, the $2\pi$-pulse will ideally produce a phase shift of $\pi$ on the state $\ket{1}$ of the target atom. This scheme implements an entangling C$_{\rm Z}$ gate\cite{Jaksch2000}, $\op{U}_{\rm C_Z} = \mathrm{diag}(1,-1,1,1)$ in the computational basis $\{\ket{00},\ket{01},\ket{10},\ket{11}\}$. This differs from the phase gate matrix of \cite{Jaksch2000} due to our use of $-\pi$ instead of $\pi$ for the last pulse which results in slightly better gate fidelity.
The Hamiltonian of the compound system, control and target atom, can be written as
\begin{align}\label{eq:ham_tot}
\op{H} & = \op{H}_{\rm control} \otimes \mathds{\hat{1}} + \mathds{\hat{1}} \otimes \op{H}_{\rm target} + \sum\limits_{i,j}{\sf B}_{r_i,r_j}\ketbra{r_i,r_j}{r_i,r_j}.
\end{align}
Here, the ${\sf B}_{r_i,r_j}$ quantify the Rydberg interaction strength between all relevant Rydberg states $\ket{r_i}$ of the control atom and $\ket{r_j}$ of the target atom, including all possible leakage levels depicted in \reffig{sequence}c). The desired excitation is resonant between $\ket{1}$ and $\ket{r}$, with leakage channels to both $(n\pm 1)p_{3/2}$ states (detunings $\mathrm{\Delta}_{\pm}$).
To remove leakage to $np_{1/2}$ states we assume a specific implementation in Cs atoms where qubit state $\ket{1}$ is mapped to $\ket{1'}=\ket{f=4,m=4}$ before and after the Rydberg gate. With $\sigma_+$ polarized excitation light $\ket{1'}$ only couples to states $\ket{np_{3/2},f=5,m_f=5}$ so there is no leakage to $np_{1/2}$ states, and errors due to coupling to multiple hyperfine states within the $np_{3/2}$ levels are also suppressed. For compactness of notation we refer to $\ket{1'}$ as $\ket{1}$ in the following.
In addition to leakage to the blockaded target Rydberg state during the $2\pi$ pulse, significant leakage channels exist for the $\pi$ pulse when the control qubit is initially in the $\ket{0}$ state (see \reffig{sequence}c). The $\ket{0}$ state is coupled to Rydberg states $\ket{n'p_{1/2,3/2}}$ and $\ket{n''p_{1/2,3/2}}$. The Cs $6s_{1/2}-np_{1/2}$ oscillator strength is anomalously small, as was first explained by Fermi\cite{Fermi1930}, and for the states S1, S2 of primary interest in \reftab{params} we estimate the ratio of Rabi coupling strengths to $np_{1/2}$ states as compared to $np_{3/2}$ states as $<1/300$\cite{Lorenzen1978}. The leakage to $np_{1/2}$ states in Cs with Gaussian pulses is therefore negligible. Nevertheless we have still included possible leakage to $np_{1/2}$ states in order to substantiate the generality of our approach. The detunings for these transitions to $np_{1/2,3/2}$ states are $\mathrm{\Delta}',\mathrm{\Delta}''$. In what follows we will refer to the interaction between two target Rydberg states $\ket{np_{3/2}}$ as ${\sf B}_0$.
\section{Design of DRAG pulses} An analytic tool to minimize leakage errors is the \emph{Derivative Removal by Adiabatic Gate} (DRAG) method \cite{Motzoi_PRL_103_110501} which is based on shaping both in- and out-of-phase control of the system. The method has been further developed \cite{Motzoi_PRA_88_062318} and provides a general toolbox to design frequency-selective pulses, a form of counter-diabatic driving \cite{Demirplak_Chem_107_9937,Torrontegui_shortcuts}. Under the assumption -- which can be derived from a Magnus expansion \cite{Warren_JChemPhys_81_5437} in the interaction picture -- that the finite Fourier transform
\begin{align}\label{eq:fourier}
S(f,\delta)=\int\limits_0^{T}\mathrm{d} t\;f(t)e^{i\delta t}
\end{align}
gives a good first-order estimate of the evolution, DRAG pulses can be alternately and more simply be derived so that they have no spectral power at certain frequencies $\{\delta_j\}$ with $j=1,\ldots,m$ . In contrast to previous work on DRAG controls, we will utilize only a shaped in-phase control $\varepsilon_x(t)$ and no additional out-of-phase quadrature, which has been suggested in earlier work \cite{Gambetta_PRA_83_012308}. This simplifies the experimental implementation. A key requirement for the method to work is that the first $N=2m$ derivatives of a pulse $f(t)$ vanish at its beginning (0) and end (T), so that we can use integration by parts to show that
\begin{align}\label{eq:partint}
S(f,\delta) = \left(\frac{i}{\delta}\right)^N\int\limits_0^{T}\mathrm{d} t\;\frac{\mathrm{d}^N f(t)}{\mathrm{d} t^N}e^{i\delta t}.
\end{align}
To obtain a control shape $\varepsilon_x(t)$ which satisfies $S(\varepsilon_x,\delta_j)=0$ for all $j=1,\ldots,m$ we make the expansion
\begin{align}\label{eq:ansatzDrag}
\varepsilon_x(t) = \varepsilon_x^{(0)}(t) + \sum\limits_{k=1}^{N/2}\alpha_{2k}\frac{\mathrm{d}^{2k} \varepsilon_x^{(0)}(t)}{\mathrm{d} t^{2k}},
\end{align}
whereby $\varepsilon_x^{(0)}(t)$ is some smooth initial shape, e.g. a Gaussian pulse, which satisfies Eq.\refeq{partint}. This particular ansatz as an expansion in terms of derivatives is motivated by a sequence of adiabatic transformations that yield instantaneous-time control and aid analytical solutions to the dynamics. Demanding $S(\varepsilon_x,\delta_j)=0$ for all $j=1,\ldots m$ and utilizing Eq.\refeq{partint} as well as Eq.\refeq{ansatzDrag} leads to a system of $m$ equations for the coefficients $\alpha_k$,
\begin{align}
1+\sum\limits_{k=1}^{N/2}\alpha_{2k}\left(-i\delta_j\right)^{2k} = 0,\quad j=1,\ldots,m.
\end{align}
For instance, if two leakage transitions at $\delta_1$ and $\delta_2$ need to be suppressed, the corresponding real-valued solutions for the coefficients $\alpha_k$ in Eq.\refeq{ansatzDrag} read
$\alpha_2 = -\left(\frac{1}{(\delta_1)^2}+\frac{1}{(\delta_2)^2}\right)$, $\alpha_4 = \frac{1}{(\delta_1)^2(\delta_2)^2}$. It is important to note that the solutions presented here minimize the error at every instant of time as can be seen from a rigorous iterative application of adiabatic transformations which essentially arrives at the same result \cite{Motzoi_PRA_88_062318}. This substantiates that the ansatz in Eq.\refeq{ansatzDrag} is preferable over other possible waveforms $f(t)$ that solely satisfy $S(f,\delta_j)=0$.
\section{Gate analysis}
\subsection{Population error}
We proceed to demonstrate how Gaussian pulses with DRAG components help to improve over previous methods by several orders of magnitude. Since the main advantage of Gaussian and DRAG shapes is an exponential suppression of leakage, we first focus on population error arising from leakage channels to other Rydberg states as shown in \reffig{sequence}c). For our simulations we use the system parameters that are listed in \reftab{params}. The two different settings, S1 and S2, respectively, belong to two possible one-photon-excitation schemes starting from the Cs $6s_{1/2}$ state. Leakage errors are expected to be worse in S2 due to smaller energy splittings at higher Rydberg states, whilst lifetimes in S2 are better by roughly a factor of two. As initial pulses for DRAG and Gaussian control, we utilize generalized Gaussians of duration $T$
\begin{align}
\varepsilon_G(t) & = A_{\theta}\left[\exp{-\frac{(t-T/2)^2}{2\sigma^2}}-\exp{-\frac{(T/2)^2}{2\sigma^2}}\right]^{N+1}
\end{align}
with a standard deviation $\sigma=2T/3$ and a pulse area $\theta$ determined by the value of $A_{\theta}$\cite{alphanote}. The exponent $N+1$ ensures that the first $N=2m$ derivatives of the Gaussian vanish at times $t=0$ and $t=T$. Note that $N$ here is the same as e.g. in \refeq{ansatzDrag}, so that we meet the conditions for Eq.\refeq{partint} to hold. Unless stated otherwise, we fix the pulse length $\tau_c$ for the $\pm\pi$-pulses on the control atom to $\tau_c=\tau_t/2$.
\begin{table}[]
\centering
\label{my-label}
\begin{tabular}{|c|c|c||c|cc|}
\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{Parameter}} & \multicolumn{2}{c||}{Value} & \multirow{2}{*}{Parameter} & \multicolumn{2}{c|}{Value} \\
\multicolumn{1}{|c|}{} & S1 & S2 & & \multicolumn{1}{c|}{S1} & S2 \\ \hline\hline
$n$ & 107 & 141 & $\tau_n\;(\mu s)$ & \multicolumn{1}{c|}{538} & 969 \\
$n'$ & 106 & 138 & $\mathrm{\Delta}_+/2\pi\;({\rm GHz})$ & \multicolumn{1}{c|}{-5.534} & -2.507 \\
$n''$ & 105 & 137 & $\mathrm{\Delta}_-/2\pi\;({\rm GHz})$ & \multicolumn{1}{c|}{5.694} & 2.562 \\
$\mathrm{\Delta}'_{1/2}/2\pi\;({\rm GHz})$ & -2.961 & -1.245 & $\mathrm{\Delta}'_{3/2}/2\pi\;({\rm GHz})$& \multicolumn{1}{c|}{-3.161} & -1.333 \\
$\mathrm{\Delta}''_{1/2}/2\pi\;({\rm GHz})$ & 3.256 & 1.495 & $\mathrm{\Delta}''_{3/2}/2\pi\;({\rm GHz})$& \multicolumn{1}{c|}{3.051} & 1.405 \\
${\sf B}_0/2\pi\;({\rm GHz})$ & 1.54 & 0.68 & $b_{n,n}$ & \multicolumn{2}{c|}{1}\\
$b_{n,n'}$ & \multicolumn{2}{c||}{0.85} & $b_{n,n''}$ & \multicolumn{2}{c|}{0.80}\\
$b_{n,n+1}$ & \multicolumn{2}{c||}{1.02} & $b_{n,n-1}$ & \multicolumn{2}{c|}{0.97}
\\\hline
\end{tabular}
\caption{\label{tab:params}System parameters for simulation of $np_{3/2}$ states in Cs for two different single-photon excitations, S1 and S2, at temperature $T=300\;{\rm K}$. Lifetimes are calculated using expressions in \cite{Beterov_PRA_79_052504}. The relative blockades $b_{n,m}$ between Rydberg states $\ket{r}$ and $\ket{r_i}$ with principal quantum numbers $n,m$ are given in units of ${\sf B}_0$.}
\end{table}
In \reffig{error_compare} we show the overall population error for a Rydberg blockade entangling gate according to the pulse sequence given in \reffig{sequence}a). Conventional square pulses perform very poorly due to a high degree of leakage. Gaussian pulses (we always compare Gaussian and DRAG pulses with equal values of $N$) show an improvement by $2$ to $2.5$ orders of magnitude over the square pulse sequence. This is attributed to Gaussians exponentially suppressing excitations to off-resonant transitions in the Fourier space, whilst square pulses only achieve a polynomial suppression.
Leakage can further be reduced by minimizing the main leakage channel into the $\ket{n'}$ subset of the control atom while also avoiding blockade leakage into the target Rydberg state $\ket{r}$ of the target atom with the aid of analytical DRAG pulses, whereby control and target pulses can be shaped independently of each other. Hence, for the area $\pi$ pulses we use $N=4$ in Eq.\refeq{ansatzDrag} to simultaneously suppress both $\mathrm{\Delta}'$ transitions, whereas $N=2$ is sufficient for the $2\pi$ pulse since only leakage to the blockade-shifted target Rydberg state is significant. Note, however, that the error from the $2\pi$ pulse is more significant than that from the $\pi$ pulse since ${\sf B}_0$ is about half the value of $\mathrm{\Delta}'$. Note also that we do not suppress the $\ket{n''}$ subset since this would require us to use $N=8$, which in turn increases the amplitude of the control pulse $\varepsilon_x$. The spectral argument that leads to our pulses only holds for $\varepsilon_x/\delta\ll 1$ ($\delta$ being the smallest detuning) so that increasing amplitudes deteriorate the gate. Owing to the frequencies $\mathrm{\Delta}'$ and $\mathrm{\Delta}''$ being very similar, the spectral power at both $\mathrm{\Delta}''$ transitions is sufficiently low even though they are not explicitly nulled out, as can be seen in the spectrum shown in \reffig{sequence}d). Using these frequency-selective shapes additionally yields $1.5$ orders of magnitude improvement over Gaussians, hence improving over square pulses by up to four orders of magnitude. Best population errors are achieved for excitations in S1, owing to larger separations of atomic levels. Under these conditions, DRAG pulses allow speeding up gates by a factor of three compared to Gaussians, while achieving the same error. Compared to square pulses, the speed up lies in the range of several orders of magnitude.
\begin{figure}[!t]
\centering
\includegraphics[width=.95\linewidth]{error_Square_Gauss_DRAG_compare.pdf}
\caption{\label{fig:error_compare}(color online) Population error for a two-qubit Rydberg blockade entangling gate as a function of gate time $t_g=\tau_t+2\tau_c$. Gaussian pulses reduce leakage errors by up to $2.5$ orders of magnitude compared to conventional square controls, while additional supplementation with DRAG further improves by another $1.5$ orders of magnitude for reasonable gate times. The DRAG pulses are designed to minimize primarily leakage into the $\ket{n'}$ subset of the control atom as well as blockade leakage in the target atom. The Rydberg blockades ${\sf B}_0/2\pi$ are $1.54\;{\rm GHz}$ and $0.68\;{\rm GHz}$ for S1 and S2, respectively.}
\end{figure}
\subsection{Optimal Rydberg blockade} The performance of the Rydberg entangling gate strongly depends on the value of the blockade shifts. Scanning over the value of ${\sf B}_0$ for a fixed gate time ($\tau_t=30\;{\rm ns}$) reveals that the optimal value for ${\sf B}_0/2\pi$ is around $1.5(0.7)\;{\rm GHz}$, for settings S1(S2) as illustrated in \reffig{scan_B}. This is explained qualitatively by analyzing the energies of all involved Rydberg states. For this purpose, we assume for simplicity that all blockades ${\sf B}_{r_i,r_j}\sim {\sf B}_0$. Starting in the initial state $\rho_{in}=\ketbra{10}{10}$ we see that due to the first $\pi$-pulse populating $\ket{np_{3/2}}$, the Rydberg levels of the target atom are blockade-shifted by ${\sf B}_0$. As a consequence, for instance the leakage transitions into the $\ket{n''}$ subset are almost resonantly driven by the $2\pi$-pulse if ${\sf B}_0\sim (\mathrm{\Delta}''_{1/2}+\mathrm{\Delta}''_{3/2})/2$, leading to even more undesired excitation. On the other hand, too small a blockade will produce large population errors since the $2\pi$-pulse will leave population inside the almost resonant blockade-shifted $\ket{np_{3/2}}$ state. This motivates a careful analysis of the Rydberg energies since an unsophisticated choice of ${\sf B}_0$ might introduce severe frequency crowding issues and tremendously lower gate fidelities.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{error_DRAG_scan_blockade.pdf}
\caption{\label{fig:scan_B}(color online) Population error for a fixed gate time $\tau_t=30\;{\rm ns}$ as a function of the blockade shift ${\sf B}_0$ in settings S1 and S2. The error is minimized for a value of ${\sf B}_0/2\pi \sim 0.7\;{\rm GHz}$ in S2. In S1, the population error is optimal for blockade shifts in the range of $0.7-2.7\;{\rm GHz}$.}
\end{figure}
Analytically estimating the optimal value for the blockade shift is possible by minimizing excitation to harmful levels\cite{Zhang_PRA_85_042310}. The matrix element for a transition to a Rydberg state $\ket{n}$ scales $\propto n^{-3/2}$. The probability of exciting states at detunings $\mathrm{\Delta}_k$ scales $\propto \mathrm{\Delta}_k^{-2}$ so that we can write the sum of all probabilities to excite harmful leakage states as
\begin{align}\label{eq:leak_prob}\begin{split}
P_{\rm leak} & \propto \frac{1}{(n+1)^3(\mathrm{\Delta}_1+{\sf B}_0)^2} + \frac{1}{(n-1)^3(\mathrm{\Delta}_1-{\sf B}_0)^2}\\
& + \frac{1}{(n')^3(\mathrm{\Delta}_2-{\sf B}_0)^2} + \frac{1}{(n'')^3(\mathrm{\Delta}_2+{\sf B}_0)^2}
+ \frac{1}{n^3{\sf B}_0^2} .
\end{split}\end{align}
Here, we have set $\mathrm{\Delta}_1 = (\mathrm{\Delta}_++\mathrm{\Delta}_-)/2$ and $\mathrm{\Delta}_2 = (\mathrm{\Delta}'_{1/2}+\mathrm{\Delta}'_{3/2}+\mathrm{\Delta}''_{1/2}+\mathrm{\Delta}''_{3/2})/4$. Finding the roots of $\mathrm{d}P_{\rm leak}/\mathrm{d}{\sf B}_0$ in order to minimize Eq.\refeq{leak_prob} for e.g. setting S2, yields an optimal value for the blockade, ${\sf B}_0/2\pi \sim 0.68\;{\rm GHz}$ which is in very good agreement to the optimal value found numerically in \reffig{scan_B}. Note that the shape of $P_{\rm leak}$ may be very flat around its exact minimum. As a consequence, it may be possible for certain setups to achieve similar performance with blockades that are clearly below the analytical estimate, as we see from the blue line in \reffig{scan_B}. There, the analytical estimate is $1.54\;{\rm GHz}$ which is twice as much as the lowest optimal value found numerically.
\subsection{Entanglement fidelity} The ideal unitary after the sequence in \reffig{sequence} is
\begin{align}\label{eq:unit_cz_like}
\op{U}_{\mathrm{C_Z},\vec{\phi}} & = \mathrm{diag}(e^{i\phi_{00}},e^{i\phi_{01}},e^{i\phi_{10}},e^{i\phi_{11}}),
\end{align}
with $\phi_{ij}\equiv\phi_{ij,ij}$ being a shorthand notation for phases on the diagonal elements. To turn the $C_Z$-like gate in Eq.\refeq{unit_cz_like} into an entangling CNOT-like gate we slightly modify the procedure that turns a $C_Z$ into a CNOT gate: Applying a Hadamard on the target qubit before and after the $C_Z$ results in a CNOT gate. Similarly, we find that a general $\pi/2$ rotation
\begin{align}\label{eq:hadamardgen}
\op{R}(\vec{h}) & = \frac{1}{\sqrt{2}}\begin{pmatrix}e^{ih_{00}} & e^{ih_{01}}\\varepsilon^{ih_{10}} & e^{ih_{11}}\end{pmatrix}
\end{align}
with phases $\vec{h}=(h_{00},h_{01},h_{10},h_{11})$ can be used to turn, up to relative phases, Eq.\refeq{unit_cz_like} into a CNOT. If the entangling phase $ \phi_{\rm ent} = \phi_{00}-\phi_{01}-\phi_{10}+\phi_{11}$ of Eq.\refeq{unit_cz_like} is exactly $\pi$, the transformation
\begin{align}\label{eq:cnotlike}
\left(\mathds{\hat{1}}\otimes\op{R}(\pi,\tilde{\phi},-\tilde{\phi},0)\right)\op{U}_{\mathrm{C_Z},\phi}\left(\mathds{\hat{1}}\otimes\op{R}(0,0,0,\pi)\right)
\end{align}
with $\tilde{\phi}=\phi_{10}-\phi_{11}$ produces a maximally entangling CNOT-like gate. In order to quantify the degree of entanglement of our pulse sequence, we pick $(\ket{00}+\ket{10})/\sqrt{2}$ as an initial state. Ideally, under Eq.\refeq{cnotlike} this yields, up to local phases, the maximally entangled Bell state $\ket{\Phi_+}=(\ket{00}+\ket{11})/\sqrt{2}$. To quantify the performance we evaluate the overlap fidelity between two density matrices $\rho,\rho_{\rm id}$ \cite{Raginsky_PLA_290_11}
\begin{align}\label{eq:fid_nu}
F = \left(\PTrace{\mathbb{Q}}{\sqrt{\sqrt{\rho}\rho_{\rm id}\sqrt{\rho}}}\right)^2.
\end{align}
Here, we take the partial trace over the computational subspace $\mathbb{Q}=\mathrm{span}\{\ket{00},\ket{01},\ket{10},\ket{11}\}$ to disregard irrelevant information about non-computational states. For $\rho_{\rm id}=\ketbra{\Phi_+}{\Phi_+}$ we denote the fidelity as Bell state fidelity $F_B$. The results are shown in the upper plot of \reffig{bellfid} whereby we assume that the $\pi/2$ gates on the qubit subspace are perfect gates. We observe that Gaussian controls seem to achieve better results than a naive DRAG control. However, the main reason for DRAG pulses to perform poorly at a first glance is wrong phases. Originally, it was proposed to change the drive frequency $\omega_d$ as a function of time to account for this effect \cite{Motzoi_PRL_103_110501}. However, it is also possible to employ a constant detuning $\Lambda$ from resonance, i.e. $\omega_d=\omega_r-\omega_q+\Lambda$ \cite{Theis_PRA_93_012324}, with the benefit of less experimental effort being required. We find, that detuning the target $2\pi$ pulse is sufficient to achieve low enough errors. As a consequence of off-resonant drive, rotation errors will be induced which can be corrected by rescaling the amplitudes of the pulses (by up to $3\%$ only for the fastest gates). The difference between the solid black line and the dotted red one in \reffig{bellfid} illustrates that a constant detuning and a rescaled amplitude indeed account for this induced error, yielding at least two orders of magnitude improvement over Gaussian waveforms. As one would expect from previous results \cite{Motzoi_PRL_103_110501}, the detuning scales proportionally to the Rabi frequency squared, yielding approximately a $1/\tau_t^2$ power law whereby the optimal detuning for a $2\pi$ pulse of $25\;{\rm ns}$ is $124.07\;{\rm MHz}$. We find that we are able to produce Bell states with a fidelity of $0.9999$ for $t_g\sim 50\;{\rm ns}$ using detuned DRAG pulses with amplitude correction.
An alternate approach to account for phase issues is by waiting an appropriate time $t_{\rm gap}$ between the pulses \cite{Maller_PRA_92_022336} or to track phases in software and correct for them afterwards. The former approach will noticeably prolong the gate times compared to our approach. Overall, detuned DRAG pulses yield an improvement of more than two orders of magnitude compared to conventional shapes. Furthermore, the necessary gate times are less than $10^{-7}$ of the few second coherence times that have been demonstrated with neutral atom qubits\cite{YWang2015}, substantiating that Rydberg gates are a promising approach for scalable quantum computing.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{unitary_BellError_Square_Gauss_DRAG_compare_set1_set3.pdf}
\caption{\label{fig:bellfid}(color online) Unitary Bell state infidelity as a measure for entanglement generated by the pulse sequence in \reffig{sequence} using square pulses, Gaussians, DRAG, detuned (d) DRAG controls and detuned DRAG controls with amplitude correction (d/r) for the setting S1 as well as optimized DRAG controls in S2. Detuning DRAG pulses on the target atom accounts for wrong phases and combines less leakage with high degrees of entanglement. The necessary detuning $\Lambda$ decreases proportionally to $1/\tau_t^2$ with a value of $\Lambda/2\pi=124.07\;{\rm MHz}$ at $\tau_t=25\;{\rm ns}$.}
\end{figure}
\subsection{Including spontaneous emission} All results in the previous section are based on unitary evolution of the atoms. A more realistic model incorporates decay due to finite lifetimes of the energy levels. We employ a Lindbladian model to simulate the effects of decoherence, whereby we assume that population of Rydberg levels decays by a fraction of $7/8$ into some auxiliary level $\ket{g}$ that has zero effect on the rest of the dynamics. The residual part decays with equal probabilities into the states $\ket{0}$ and $\ket{1}$ of the atoms. Hence, the full dynamics of our system are goverened by the Lindblad master equation for the density operator $\op{\rho}$
\begin{align}
\dot{\op{\rho}} & = -i\comut{\op{H}}{\op{\rho}} -\frac{1}{2}\sum\limits_r \left(\op{C}_r^{\dagger}\op{C}_r\rho + \rho\op{C}_r^{\dagger}\op{C}\right) + \sum\limits_r\op{C}_r\op{\rho}\op{C}_r^{\dagger}.
\end{align}
Here, the operators $\op{C}_r=\op{c}_r \otimes \mathds{\hat{1}} + \mathds{\hat{1}} \otimes \op{c}_r$ describe decay of all relevant Rydberg states $\ket{r}$ in both atoms into $\ket{g}$, $\ket{0}$ and $\ket{1}$, i.e.
\begin{align}
\op{c}_r & = \sqrt{\Gamma_r}\LR{\frac{7}{8}\ketbra{g}{r}+\frac{1}{16}\ketbra{0}{r}+\frac{1}{16}\ketbra{1}{r}}.
\end{align}
The decay rate $\Gamma_r$ is the inverse of the lifetime $\tau_r$ of a Rydberg state $\ket{r}$. Values for the target Rydberg states in both settings are given in \reftab{params}. For an experiment at room temperature ($\sim 300\;{\rm K}$) in setting S1, we find that Bell states are generated with a fidelity of better than $0.9999$ at a gate time of $\lesssim 60\;{\rm ns}$. The results for optimized DRAG pulses are plotted in \reffig{bellfid_nu}. As expected because of shorter lifetimes, non-unitary errors become visible earlier in S1 than in S2. However, unitary errors are dominant, so that gates in S1 apear to be more promising than those in S2, despite the shorter lifetimes. Since the $\pi$ pulses on the control atom do not require blockade effects, we may run them faster without losing performance. The dotted red curve in \reffig{bellfid_nu} confirms this observation. For $\tau_c=\tau_t/3$ we achieve slightly better results, yielding errors less than $10^{-4}$ at only $50\;{\rm ns}$ gate time. In a $4\;{\rm K}$ environment lifetimes will be on the order of a few ms, allowing for performance very similar to that for the unitary analysis.
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{nonunitary_BellError_DRAG_compare_set1_set3_amplitude_detuned.pdf}
\caption{\label{fig:bellfid_nu}(color online) Bell state infidelity including decay from all Rydberg levels for optimized detuned DRAG controls in both settings S1 and S2. Bell states are generated with a fidelity of $0.9999$ at a total gate time of only $50\;{\rm ns}$.}
\end{figure}
We have characterized the gate performance in terms of the Bell state fidelity. While the fidelity is the most widely used measure of gate performance, others have been proposed\cite{Gilchrist2005}. In particular the trace distance
has been shown to be linearly sensitive to Rydberg gate phase errors that affect the fidelity only quadratically\cite{Zhang_PRA_85_042310}.
Using the rescaled and detuned DRAG gates that optimize the fidelity we find that the trace distance error is an order of magnitude larger.
As it is an open question as to which performance measure is most relevant for specific quantum computational tasks we have not studied the trace distance in more detail, although we anticipate that the trace distance error could also be reduced with appropriate pulse design,
\section{Summary} In conclusion we have presented DRAG pulses with $x$ quadrature control for Rydberg blockade gates that lead to Bell state fidelity $F_B>0.9999$ with gate times of 50 ns. The pulses are generated with an analytical method that could readily be extended to the level structure of other atoms. The results fully account for all the dominant leakage channels as well as Rydberg decay in a room temperature environment. The 50 ns gate time is orders of magnitude faster than high fidelity trapped ion gates, about the same speed as state of the art superconducting qubit gates, while the ratio of coherence time to gate time is orders of magnitude better.
Together with recent progress in high fidelity single qubit gates\cite{Xia2015,YWang2015} DRAG pulses establish neutral atom qubits with Rydberg gates as a promising candidate for scalable quantum computation. Our result specifically applies to the case of one-photon Rydberg excitation. We leave extension to the more common case of two-photon excitation for future work. We also emphasize that the predicted gate fidelity assumes no technical errors and ground state laser cooling. Demonstrating real performance close to the theoretical level established here remains an outstanding challenge.
\section*{Acknowledgments}
MS acknowledges funding from the IARPA MQCO program through ARO contract W911NF-10-1-0347, the ARL-CDQI through cooperative agreement W911NF-15-2-0061 and NSF award 1521374. LST, FM and FKW were supported by the European Union through SCALEQIT.
|
\section{Introduction}
The trend in wireless communications towards moving infrastructure access points off of tall towers and bringing them into the user environment is well established. Small pico-cell and femto-cell heterogeneous cellular deployment in outdoor urban environments is one of the most popular methods for increasing capacity in next generation cellular networks \cite{ghosh-a-2012}. In several vehicular ad-hoc network proposals, communication between vehicles and roadside wireless access points is a requirement \cite{hartenstein-h-2008}. Finally, most sensor network applications involve deploying sensors close to the ground \cite{romer_k1}.
When high density wireless networks are deployed, providing a backhaul network to carry access point traffic is a challenge in most urban centers. When placing access points on utility poles, street lights or traffic lights, there typically is no wired communications infrastructure to carry backhaul traffic. In these cases, a backhaul consisting of fixed peer-to-peer wireless links between access points is clearly attractive.
From a propagation perspective, the transmit and receive antennas for these backhaul links will be mounted well below traditional micro-cell antenna height and the transmitter and receiver will both be surrounded by similar scattering environments. The antennas will be close to vehicular and pedestrian traffic which will influence small scale fading \cite{ahumada-l-2006}. Most heterogeneous cellular and vehicular network applications will require an access point approximately every city block, which means link distances of approximately 150~m to 200~m. If considering a North American context, the regular grid layout of most urban centers will allow line of sight (LOS) or near-LOS backhaul channels between access points.
The outdoor fixed wireless link has received considerable research attention and there have been a number of fixed wireless link propagation measurement campaigns for carrier frequencies close to 2~GHz \cite{greenstein-lj-2009, liou-a-2009, michelson-dg-2009}, 2.5~GHz \cite{erceg-v-2004, gans-mj-2002}, 3.5~GHz \cite{ahumada-l-2005, hong-cl-2003a, hong-cl-2003b} and 5~GHz \cite{durgin-gd-1998, domazetovic-a-2003, skentos-nd-2006}. These studies all characterize the fixed wireless access channel where a fixed user terminal is communicating with a base station tower positioned above the scattering environment experienced by the user. The studies that do adjust antenna height \cite{greenstein-lj-2009, durgin-gd-1998, ahumada-l-2005, gans-mj-2002} adjust the receive antenna to observe the effect of raising the receiver out of the clutter, rather than lowering the transmit antenna into it. These studies also consider typical macro-cellular link distances ranging from several hundreds of meters to several kilometers.
Many of the existing studies that consider transmit and receive antennas that are both down in the same scattering environment focus on large scale channel effects only. The below rooftop to below rooftop (BRT to BRT) LOS transmission in \cite{802.16-channel-2007} provides path loss values that would be representative of the fixed wireless backhaul scenario. Other studies that measure large scale fading values when both transmit and receive antennas are below rooftop include \cite{ahumada-l-2013, xia-h-1994, xia-h-1993, feuerstein-mj-1994} but these campaigns did not set transmit and receive antennas at the same height, as would be typical for a backhaul scenario. The work in \cite{durgin-gd-2003} is the only study that utilized the same antenna height for the transmitter and receiver. However, the antenna heights used in \cite{durgin-gd-2003} are only 1.5~m. The WINNER channel model fixed street level feeder scenario, case B5B in \cite{winner-2005}, is the most relevant to the measurements in this paper and will be used as a point of comparison for both small and large scale fading results.
The contribution of this paper is to present the results of a propagation measurement campaign that characterizes both large and small scale channel effects for fixed wireless links meant to carry backhaul traffic for heterogeneous cellular, vehicular and sensor network infrastructure in a dense urban environment. Both the transmit and receive antennas are mounted at the same height as would be typical for a backhaul scenario. The results from this campaign will complement existing fixed wireless propagation work that has focused on access scenarios with cellular style towers and very long link distances. Dual antenna multiple input/single output (MISO) measurements are collected for both spatially separated omni antennas and dual-polarized directional antennas. The results will demonstrate how antenna pattern affects channel statistics and the correlation between antennas for the urban wireless backhaul scenario.
An additional contribution of this paper is to focus specifically on the properties of the propagation environment that will affect the ability of a wireless backhaul unit to simultaneously support several spatial data streams. In a grid-style peer-to-peer backhaul network where radio nodes located at each intersection are serving as relays for other nodes, it will be common for one node to be communicating with devices in multiple directions. One way to handle these streams simultaneously is to use multiple antenna techniques: either spatial multiplexing or directional antenna space division multiple access (SDMA). To help evaluate the feasibility of SDMA, this study will measure the spatial rejection achieved in the off-broadside direction using directional antennas. The feasibility of spatial multiplexing will be determined by evaluating the correlation between spatially separated omni antennas both in the broadside and off-broadside directions.
Section~\ref{sec:meas} discusses the radio channel measurement equipment, measurement collection procedure and the environment where the measurements were captured. The method used to analyze the propagation measurements is discussed in Section~\ref{sec:analysis} and the results of this analysis are presented in Section~\ref{sec:results}. Concluding remarks are made in Section~\ref{sec:concl}.
\section{Propagation Measurements}
\label{sec:meas}
In this section, Subsection~\ref{ssec:equip} describes the radio channel measurement equipment and Subsection~\ref{ssec:proc} discusses how the equipment is used in the propagation environment to capture wireless channel response data.
\subsection{Equipment}
\label{ssec:equip}
The measurement apparatus used for this study collects $2 \times 1$ MISO channel impulse response measurements. The transmit unit consists of two field programmable gate array (FPGA) software radio boards installed inside a commercial pico-cell base station enclosure. Each radio is connected to a different transmit antenna port and both are locked to a common time and carrier frequency reference. Each radio generates the same 25~Mchip/sec, length $2^{19}-1$ maximal length pseudo-random noise (PN) sequence \cite{proakis_jg1} and utilizes an offset of $2^{18}$ chips to ensure orthogonality. The offset sequences are denoted $p_1[n]$ and $p_2[n]$. After digital to analog conversion, the signals have an approximate 3~dB double sided bandwidth of 25~MHz. They are amplified to a maximum transmit power of 36~dBm equivalent isotropically radiated power (EIRP) and transmitted at a carrier frequency of 2.4724~GHz.
The receive unit contains a single FPGA software radio board that implements two real time correlators, one correlating the received signal with $p_1[n]$ and the other with $p_2[n]$. The correlation is performed over 100,000 chips so that a $2 \times 1$ MISO channel impulse response is captured every 4~ms for a measurement capture rate of 250~Hz. The channel response from antenna $i$ is represented by matrix $\Mx{h}_i \in {\cal C}^{N \times L}$ where $N$ is the number of channel impulse responses captured and $L$ is the number of discrete channel taps. The element at row $n$ and column $l$ of $\Mx{h}_i$ is denoted $h_i(n,l)$.
Time and frequency synchronization for the transmit and receive units is provided by SIM940 rubidium clock references \cite{sim940}. The references provide both 1 pulse per second and 10~MHz reference signals. They are synchronized to each other after a warm up period and are battery powered to ensure synchronization is maintained during the measurements.
Measurements are collected using either omni-directional antennas or directional antennas. For the omni antenna measurements, broadside is defined as perpendicular to the line through the two transmit antenna elements. Broadside for the directional antenna measurements is perpendicular to the panel antenna face in the direction of maximum antenna gain. Off broadside for both cases is 90$^{\circ}$ off of the broadside direction in azimuth.
The omni antennas are the W1030 quarter wavelength dipole antennas manufactured by Pulse Electronics \cite{w1030-antenna-2014}. The antenna has a gain of 2.0~dBi and an omni-directional horizontal radiation pattern. The separation between the two transmit omni dipole antennas is 23~cm. This separation is dictated by the location of the antenna ports on the pico-base station enclosure used to house the transmitter software radios.
The directional antennas are the MT-344048/ND cross polarized panel antennas manufactured by MTI Wireless \cite{mt344048-antenna-2014}. The panel antennas mount on the front of the base station enclosure. They have a 30$^{\circ}$ beamwidth and have antenna gain patterns that satisfy the ETSI EN 302 326-3 cross polarized directional antenna standard. This is an antenna standard created by the European Telecommunications Standards Institute governing antenna requirements for fixed wireless applications. The details on this antenna standard, including quantitative values for the antenna gain masks can be found in \cite{etsi-ants}.
It is important to verify that the system has sufficient signal to noise ratio to ensure measurement noise will not be mistaken as fading when measuring very stable wireless links \cite{ahumada-l-2005}. To this end, the measurement transmitter and receiver were connected with a cable that included 80~dB of attenuation. The resulting signal to noise ratio of the system, as defined in \cite{ahumada-l-2005}, was 44.3~dB. The methods in \cite{ahumada-l-2005} can be used to show this is high enough for measurement noise to cause only negligible error for the channels measured in this study.
The parameters for the measurement equipment section are summarized in Table~\ref{tb.equip}.
\begin{table}[htbp]
\centering
\begin{tabular}{|c|c|}
\hline
Parameter & Value \\
\hline\hline
Carrier Frequency & 2.4724~GHz \\
Signal Bandwidth & 25~MHz \\
Maximum Transmit Power & 36~dBm~EIRP \\
MISO Configuration & 2$\times$1 \\
Impulse Response Capture Rate & 250~Hz \\
Omni Antenna Gain & 2~dBi \\
Directional Antenna Gain & See \cite{etsi-ants}\\
\hline
\end{tabular}
\caption{Measurement equipment parameters.}
\label{tb.equip}
\end{table}
\subsection{Measurement Procedure}
\label{ssec:proc}
At the start of each measurement session, the transmit unit is attached to a traffic light pole, as shown in Fig.~\ref{fg.setup}a. The target mounting height for the transmit unit is 3~m but this would vary $\pm$0.25m from site to site depending on the configuration of the pedestrian walk lights on the traffic light pole. The receive equipment, shown in Fig.~\ref{fg.setup}b, is placed in a wheeled cart equipped with an antenna mast. In order for the measurements to be representative of a peer-to-peer wireless backhaul scenario, the mast is adjusted so that the receive antenna is the same height as the antennas on the transmit unit.
\begin{figure}[htbp]
\centering
\bMiniPage{1.5in}
\centering
a) Transmitter\\[2mm]
\includegraphics[width=1.4in]{TrafficLightMount}
\end{minipage}%
\bMiniPage{1.5in}
\centering
b) Receiver\\[2mm]
\includegraphics[width=1.4in]{Cart}
\end{minipage}
\caption{Measurement equipment.}
\label{fg.setup}
\end{figure}
A map of the measurement environment is shown in Fig.~\ref{fg.map}. The area is dominated by office buildings that range in height from 20 to 50 storeys. There is either single or double lane vehicular traffic on each street with typical vehicular speeds not exceeding 50~km/hour. The streets also include broad pedestrian sidewalks. All measurements were collected during the business day so that vehicular and pedestrian traffic levels ranged from moderate to heavy.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{map}}
\caption{Measurement locations.}
\label{fg.map}
\end{figure}
In Fig.~\ref{fg.map}, the stars indicate the different locations where the transmit unit was mounted to a traffic light pole. For each of these locations, the receiver cart is used to collect data while moving down the street in both the broadside and off-broadside directions, as indicated in Fig.~\ref{fg.map}. A single trip down the street with the receiver cart is referred to as a measurement {\em route}. All routes start with the cart no more than 5~m away from the transmit unit and end with the cart one block away from the transmit unit, as indicated by the solid and dotted lines in Fig.~\ref{fg.map}.
For each transmitter location, the broadside and off-broadside directions are measured once for the omni antenna case and once for the directional antenna case using the following procedures.
\subsubsection{Omni Antenna Measurement Procedure}
\label{ssec:omni-proc}
When moving the receive cart along a particular route, a pair of $2\times 1$ measurement channel impulse response matrices, $\Mx{h}_1$ and $\Mx{h}_2$, are captured at measurement locations spaced by approximately 5~m. This spacing would vary somewhat due to obstructions on the sidewalk. Approximately 20 of these measurements are captured for the off-broadside routes and 35 measurements for the broadside routes.
When the cart arrives at measurement location, the receive correlator is activated and the apparatus begins to capture MISO channel responses at a rate of 250~Hz, as described in Section~\ref{ssec:equip}. In order to capture small scale fading variation due to spatial offsets, the receiver cart is moved within a square of at least 0.5~m $\times$ 0.5~m while the MISO channel impulse responses are being captured. The size of this 0.5~m $\times$ 0.5~m measurement grid falls within the range of grid sizes reported for capturing small scale fading in a local neighbourhood \cite{chiu-s-2010,chee-kl-2012}. The measurements captured using this approach are referred to as {\em spatial measurements}.
To characterize small scale fading variation due to moving objects in the environment, measurements were also collected while the transmit and receive antennas were completely stationary. These are referred to as {\em temporal measurements}. One temporal measurement is captured at the end of each measurement route where the separation between transmitter and receiver was at its maximum. The duration of each of these measurements was 5-6~minutes and the measurement capture rate was 250~Hz. Any channel variation longer than this 5-6~minute duration will not be reflected in the analysis presented in this paper.
\subsubsection{Directional Antenna Measurement Procedure}
\label{ssec:dir-proc}
In order to reflect a typical backhaul scenario, it is important that the transmit and receive directional antennas are pointed directly at each other for all propagation measurements. This alignment is difficult to maintain while moving the cart from side to side in a square grid. The risk is that the temporary misalignment of the beams due to cart motion may cause artificial fading.
To address this concern, the receive correlator is enabled at the start of a measurement route and MISO channel impulse responses are captured continuously as the cart is moved in a straight line to the end of the route. This allows the antenna orientation to be properly maintained. Spatial small scale fading variations are characterized by dividing the measurement track into 2~m segments and processing all channel impulse responses captured within a segment to determine spatial small scale fading statistics in that local area.
Unlike the omni antenna, the directional panel antenna used on the receive cart has two spatial polarizations. As a result, a separate set of measurements are collected along each route for each of the polarizations.
\section{Measurement Analysis}
\label{sec:analysis}
In this section, the procedure for extracting large scale channel information from the measurements is described in Section~\ref{ssec:lscale} and the small scale fading analysis is described in Section~\ref{ssec:sscale}.
\subsection{Large Scale Channel Effects}
\label{ssec:lscale}
The large scale channel effect analysis consists of estimating both the path loss exponent and shadowing distribution for each measurement route. The first step is to determine the average received power from transmit antenna $i$ at a particular measurement location according to
\begin{equation}
P_i = \frac{1}{N}\sum_n\sum_l |h_i(n,l)|^2
\label{eq.AvgRxPwr}
\end{equation}
\noindent
For the omni antenna case, values of $P_i$ are calculated by performing the averaging in (\ref{eq.AvgRxPwr}) on each $0.5 \times 0.5$~m spatial measurement described in Section~\ref{ssec:omni-proc}. For the directional antennas, a value of $P_i$ is calculated for each of the 1~m measurement track segments described in Section~\ref{ssec:dir-proc} by applying (\ref{eq.AvgRxPwr}) to all the channel impulse responses captured in each segment.
As described in \cite{rappaport_ts1}, the path loss exponent is calculated as the slope of the best fit line on a log-log plot of average received power values versus distance between the transmit and receive radios. The shadowing values are the deviation of each average received power point to that best fit line. Separate path loss exponents and shadowing average power deviation values are calculated for each transmit antenna and for the two different polarizations of the directional receive antenna.
For this study, a separate line of best fit is calculated for each measurement route. This produces several path loss exponent values and also results in a tighter shadowing distribution than if a single line of best fit was calculated for all received power values from all streets. This is desirable since this tighter distribution more accurately reflects shadowing conditions on a particular street. Note that a single shadowing distribution is still calculated by aggregating average power deviation values from all streets into a single histogram. However, these deviation values are determined using the best fit lines calculated on a per street basis.
\subsection{Small Scale Channel Effects}
\label{ssec:sscale}
To analyze spatial small scale fading variation, all channel impulse responses captured in a $0.5 \times 0.5$~m grid, for the omni case, or a 1~m linear segment, for the directional case, are normalized to have unit average power. To analyze temporal small scale variation, the temporal measurement is divided into 60~s segments and each segment is normalized and analyzed separately. The responses are then transformed to the frequency domain by taking the discrete Fourier transform of each impulse response, $h_i(n,l)$, along index $l$. Let $H_i(n,w)$ denote the transformed impulse response where $w$ is the discrete frequency index and the definitions of $n$ and $i$ are unchanged. The channel frequency response matrix is defined as $\Mx{H}_i \in {\cal C}^{N \times L}$, where the $n$th row is the Fourier transform of the channel impulse response captured at time $n$.
If $W_{\rm coh}$ is the discrete coherence bandwidth of the channel and $W_{\rm sig}$ is the discrete bandwidth of the measurement signal, then $R = \lfloor W_{\rm sig}/W_{\rm coh} \rfloor$ is the number of channel frequency response values with independent small scale fading. Since $N$ fading values are captured for each of these frequency points, this allows the creation of a $1 \times NR$ {\em fading vector} for the $i$th transmit antenna defined as $\Mx{e}_i = [\ |H_i(0,0)|\ \ldots\ |H_i(N-1,0)|\ |H_i(0,W_{\rm coh})| \ldots\ |H_i(N-1,RW_{\rm coh})|\ ]$.
Antenna correlation is calculated by determining the correlation coefficient between $\Mx{e}_1$ and $\Mx{e}_2$. While this is only a $2 \times 1$ system, it is important to point out that the correlation between the two antenna elements at the transmitter would be representative of the correlation between elements of a larger array with the same element spacing.
To quantify the severity of fading in the channel, the amount of fading (AF) metric is used \cite{alouini-2005}. AF is defined as $AF = \mbox{variance}\left(\Mx{e}_i^2\right)/\Exp{\Mx{e}_i^2}^2$. As will be discussed in Section~\ref{sec:results}, there are several instances where the fading of $\Mx{e}_i$ cannot be modeled as Ricean. AF is therefore a more appropriate fading metric since it is independent of a particular fading distribution.
Two different types of AF values are determined. {\em Temporal AF} values are calculated from measurements where both the transmit and receive antennas are stationary and channel variation is due to movement of scattering objects in the environment. {\em Spatial AF} values are calculated from measurements where the receive antenna is moved about a small local area of a few wavelengths during the measurement. In the spatial case, fading is also caused by the motion of the antenna and scatterers.
It is important for fixed wireless network designers to understand both the spatial and temporal fading. The temporal variations are clearly the fluctuations that the wireless link must contend with after deployment. However, spatial fading characterizes the possible variation in received power that may occur due to small offsets in the mounting position for the antenna. While spatial variations are constant once the antenna is mounted, the network designer must still allow a fading margin in the link budget calculation to account for them. Basing a link budget fading margin on temporal fading alone would result in transmit powers that are too low.
\section{Measurement Results}
\label{sec:results}
In this section, Subsections~\ref{ssec:lsanalysis} and \ref{ssec:ssanalysis} present the large scale and small scale analysis results, respectively. Subsection~\ref{ssec:rejanalysis} discusses the measurements used to evaluate the feasibility of using multiple antennas to support simultaneous communication in the broadside and off-broadside directions.
\subsection{Large Scale Path Loss Analysis}
\label{ssec:lsanalysis}
As discussed in Section~\ref{ssec:lscale}, separate path loss exponents are determined for all intersections. Cumulative distribution function (CDF) plots of the path loss exponent values for both the omni and directional antennas in the broadside direction are shown in Fig.~\ref{fg.pathloss}.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{PathLossExp}}
\caption{Path loss exponent values.}
\label{fg.pathloss}
\end{figure}
We can note from Fig.~\ref{fg.pathloss} that the path loss exponent values are quite low with the means of both antenna types close to but below 2. This is consistent with path loss values in \cite{xia-h-1994, xia-h-1993, feuerstein-mj-1994, winner-2005} and is due to the urban canyon acting as a waveguide that serves to reduce path loss exponent. A slightly higher pathloss exponent was observed in \cite{ahumada-l-2013} but in that study, the authors note that even their line of sight measurements were typically obscured by trees. We also note that the directional antennas exhibit a slightly higher path loss exponent since omni signals are more diffuse in the environment and are less affected by obstructions in the direct line of sight between transmitter and receiver.
Fig.~\ref{fg.shadow} shows the shadowing distribution for the two antenna types. Both antenna types experience shadowing that is approximately log-normal with the directional antenna experiencing a larger standard deviation. The smaller standard deviation for the omni antenna is again due to the signal being more diffuse in the propagation environment. The shadowing standard deviation for both cases is comparable to the B5B scenario in \cite{winner-2005}.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{Shadowing}}
\caption{Shadowing distributions.}
\label{fg.shadow}
\end{figure}
As described in Section~\ref{ssec:lscale}, the path loss and shadowing values are calculated on a per-measurement route basis. Since some streets with more obstructions may be expected to have both a higher path loss exponent and a higher shadowing variation, it is possible that some correlation may exist between the two quantities. This can be determined from the scatterplot in Fig.~\ref{fg.nAndShdwCorr} which shows the path loss exponent for each measurement route plotted versus the shadowing standard deviation for that route.
Fig.~\ref{fg.nAndShdwCorr} shows little relationship between path loss and shadowing standard deviation for the omni antenna but a stronger correlation for the directional antenna. To support this, a calculation of correlation coefficient between the two values shows that path loss exponent and shadowing standard deviation have a correlation coefficient of 0.048 for the omni case but 0.39 for the directional case. The higher correlation for the directional case is most likely due to the sensitivity of narrow beams to obstructions in the line of sight path between transmitter and receiver.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{PathLossVsShadowScatter}}
\caption{Path loss exponent versus shadowing standard deviation.}
\label{fg.nAndShdwCorr}
\end{figure}
\subsection{Broadside Small Scale Fading Analysis}
\label{ssec:ssanalysis}
As described in Section~\ref{ssec:sscale}, the first step in assembling the fading vectors for analysis is to calculate the coherence bandwidth and determine the number of frequencies in the channel response with independent fading. The mean coherence bandwidths for the omni and directional antennas, calculated as described in \cite{rappaport_ts1}, are 2.33~MHz and 5.50~MHz, respectively. The higher coherence bandwidth for the directional antenna is expected since the beam pattern will reject some scatterers and reduce the delay spread of the channel.
Next, each small scale fading vector, $\Mx{e}_i$, is subjected to a Chi-square goodness-of-fit test \cite{navidi_wc1} to determine if it fits a Ricean distribution. A significance level of 5\% is used, meaning there is a 5\% chance that measurements that are actually Ricean will be discarded. The percentage of measurements that fit a Ricean distribution is 100\% and 88\% for the omni antenna spatial and temporal scenarios, respectively. For the directional antenna, 38\% and 68\% of the spatial and temporal measurements, respectively, satisfy the Ricean fit. The directional spatial measurements, in particular, were a poor fit to the Ricean distribution primarily due to the straight line measurement procedure described in Section~\ref{ssec:dir-proc}. The reduced number of spatial samples in each measurement results in distributions concentrated around a small number of discrete fading values.
Fig.~\ref{fg.AF} shows spatial and temporal linear AF CDFs for the omni and directional antennas. The mean linear AF values for the omni spatial, omni temporal, directional spatial and directional temporal scenarios are 0.778, 0.592, 0.313 and 0.730, respectively. As expected, the omni spatial channel exhibits more severe fading than the directional spatial channel since the omni antenna gain pattern admits more scattering rays than the directional pattern. While the directional temporal fading is slightly more severe than the omni temporal fading, fading for the two scenarios is still quite close. Temporal fading is expected to be similar for the two antenna types since most of the scattering objects are close to the line of sight path due to the large transmitter/receiver separations used for the temporal measurements. At this distance, the directional antenna pattern is not narrow enough to reject the signals from these scatterers. Finally, when the B5B K-factor in \cite{winner-2005} is translated to AF, the result is 0.174. This indicates that the fading measured in this campaign for all scenarios is considerably more severe than that assumed by the WINNER model.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{AF}}
\caption{Omni and directional AF distributions.}
\label{fg.AF}
\end{figure}
Fig.~\ref{fg.brdCorr} shows correlation coefficients for all measurement and antenna types in the broadside direction with the directional antenna curves marked with circles. The results in Fig.~\ref{fg.brdCorr} indicate very similar correlation performance in all cases in the broadside direction which suggests that both spatial and polarization diversity are good options in this environment. It should be noted that no real dependence with distance was observed in the correlation coefficient values. This is in contrast with \cite{erceg-v-2004} which did observe some decrease in antenna correlation with distance. The primary reason for the difference here is that the transmitter/receiver separation distances in this study are much smaller than \cite{erceg-v-2004}. The environment is also very uniform along each measurement route due to the low antenna heights and urban canyon environment.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{BroadFadingCorr}}
\caption{Omni and directional correlation coefficient in the broadside direction.}
\label{fg.brdCorr}
\end{figure}
\subsection{Broadside and Off-broadside Propagation Conditions}
\label{ssec:rejanalysis}
In this subsection, propagation conditions are compared in the broadside and off-broadside directions. These results will shed some light on the feasibility of using multiple antenna techniques to allow a node to simultaneously communicate in these two directions.
First, the ability of the directional antenna to reject signals from the off-broadside direction is investigated. A high degree of rejection would indicate that SDMA using fixed beams is a viable option. The analysis starts by comparing the difference in average received power levels for the broadside and off-broadside directions.
Let $P_{xy,B}$ and $P_{xy,OB}$ denote the average received powers in the broadside and off-broadside directions, respectively, measured for the same transmitter/receiver separation, $d$. The subscripts $x$ and $y$ indicate transmit and receive antenna polarizations, respectively, so that $x,y \in \{ V, H \}$. Cross-polar rejection in the broadside direction is quantified by the distribution of all $P_{VV,B} - P_{VH,B}$ and $P_{HH,B} - P_{HV,B}$ values calculated for all $d$. Cross-polar rejection in the off-broadside direction is the distribution of all $P_{VV,B} - P_{VH,OB}$ and $P_{HH,B} - P_{HV,OB}$ values for all $d$. Finally, co-polar rejection in the off-broadside direction is the distribution of all $P_{VV,B} - P_{VV,OB}$ and $P_{HH,B} - P_{HH,OB}$ values for all $d$.
A plot of the CDF of the antenna rejection values is shown in Fig.~\ref{fg.reject}. First, the average cross-polar rejection of 11.0~dB in the broadside direction is close to the 9~dB cross-polar rejection value for the B5B scenario in \cite{winner-2005}. This verifies the accuracy of the measurements. The figure also indicates very similar rejection in the off-broadside direction for both the co-polar and cross-polar measurements. Since almost all measurement locations in the off-broadside direction were obscured, the off-broadside signals experienced heavy scattering that served to spread the signal across both polarizations. As a result, the primary mode of signal rejection for both cases was the directional antenna pattern.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{DirRejection}}
\caption{Directional antenna rejection.}
\label{fg.reject}
\end{figure}
Spatial multiplexing is the second option for simultaneously accommodating broadside and off-broadside communication. Fig.~\ref{fg.oBrdCorr} shows the spatial and temporal correlation coefficients in the off-broadside direction for the omni antenna measurements. Omni antenna results are shown based on the assumption that the spatial multiplexing would be implemented using a series of spaced antenna array elements.
The figure shows that both spatial and temporal fading spatial correlation is low enough to support a spatial multiplexing scheme. It should be emphasized that these low correlations are achieved using linearly spaced antenna elements that are approximately in line with the direction of transmission. The figure shows that the temporal correlation coefficient is noticeably lower than the spatial correlation. This is due primarily to the fact that the temporal measurements are collected only at the far end of each off-broadside measurement route where the direct path between the transmitter and receiver is typically obscured.
\begin{figure}[htbp]
\centerline{\includegraphics[width=3.5in]{OffBroadFadingCorr}}
\caption{Off-broadside omni fading correlation coefficient.}
\label{fg.oBrdCorr}
\end{figure}
\section{Conclusion}
\label{sec:concl}
This paper has presented a propagation study for fixed wireless links in a dense urban environment where the antennas are mounted on utility poles or street lights. The antenna heights are much lower than traditional micro-cellular base stations and the transmit and receive units experience similar physical scattering environments.
Large scale fading analysis indicates that the wireless link is best characterized with a very low path loss exponent and shadowing standard deviation relative to traditional dense urban cellular propagation models. This reflects both the near-LOS conditions of the links and the per-street method used to calculate the shadowing values. Antenna type does have an effect on large scale fading with directional antennas experiencing larger shadowing variation.
Small scale fading analysis indicates moderate fading with the majority of measurements experiencing less than Rayleigh fading severity. Directional antennas successfully reduce fading severity but only in the spatial measurement case. In the broadside direction, both polarization and spatial diversity would offer equivalent performance.
Finally, multi-antenna techniques show some promise for supporting simultaneous communication in the broadside and off-broadside directions. A spatial multiplexing approach would be more attractive due to the very low fading correlation observed in the off-broadside direction. SDMA is also attractive since directional antennas demonstrate good rejection in the off-broadside direction with approximately equivalent performance for co-polarized and cross-polarized transmission.
\section*{Acknowledgment}
This work was supported by TEKTELIC Wireless Communications Inc., SRD Innovations Inc. and the Canadian Natural Sciences and Engineering Research Council (NSERC).
\printbibliography
\begin{IEEEbiography}
[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{MikePic}}]{Michael Wasson}
Michael W. Wasson received a B.Sc. degree in electrical engineering in 2010, and is expecting to receive his M.Sc. in electrical engineering in 2016, both from the University of Calgary in Alberta, Canada.
\end{IEEEbiography}
\begin{IEEEbiography}
[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{GeoffPic}}]{Geoffrey Messier}
Geoffrey Messier (S'91 - M'98) received his B.S. in Electrical
Engineering and B.S. in Computer Science degrees from the University of
Saskatchewan, Canada with great distinction in 1996. He received his
M.Sc. in Electrical Engineering from the University of Calgary, Canada
in 1998 and his Ph.D. degree in Electrical and Computer Engineering from
the University of Alberta, Canada in 2004.
From 1998 to 2004, he was employed in the Nortel Networks CDMA Base
Station Hardware Systems Design group in Calgary, Canada. At Nortel
Networks, he was responsible for radio channel propagation measurements
and simulating the physical layer performance of high speed CDMA and
multiple antenna wireless systems. Currently, Dr. Messier is a
Professor in the University of Calgary Department of
Electrical and Computer Engineering. His research interests include
data networks, physical layer communications and communications
channel propagation measurements.
\end{IEEEbiography}
\begin{IEEEbiography}
[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{DevinPic}}]{Devin Smith}
Devin Peter Smith was born in Calgary, Alberta, Canada, in 1988. He received his BSc. in Computer Engineering from the University of Calgary in 2012, and is expecting to complete his MSc. in Electrical Engineering in 2015. Currently, his main research interests have been focused on wireless propagation and sensor networks, specifically within forest environments.
\end{IEEEbiography}
\end{document}
|
\section{Introduction}
A two-receiver discrete memoryless broadcast channel $p(y, z| x)$ is a channel with one sender and two receivers. The goal of the sender is to send private messages to the receivers over multiple uses of the channel. The capacity region of the channel is the set of all rate pairs $(R_1, R_2)$ such that private information with asymptotically vanishing error can be sent to the receivers at rate $R_1$ and $R_2$ respectively.
It is easy to verify that the capacity region of the broadcast channel $p(y, z|x)$ depends only on the \emph{marginal} channels $p(y|x)$ and $p(z|x)$ and not on the whole $p(y, z|x)$. So we may assume with no loss of generality that $p(y, z|x)=p(y|x) p(z|x)$. That is, we may think of a broadcast channel as two point-to-point channels with the same input sets.
There are some known inner and outer bounds for the capacity region of the broadcast channel~\cite{ElGamalKim}, yet deriving a single letter formula for the capacity region is a long standing open problem.
The best inner bound for the capacity region of the broadcast channel is due to Marton~\cite{Marton} and is described as follows. Let $p_{UVWX}$ be an arbitrary distribution that induces the distribution $p_{UVWXYZ}$. Then any pair of non-negative numbers $(R_1, R_2)$ satisfying
\begin{align}\label{eq:Marton}
R_1&\leq I(U, W; Y),\nonumber\\
R_2 & \leq I(V, W; Z),\nonumber\\
R_1+R_2&\leq \min\big\{I(W; Y),\, I(W;Z) \big \} + I(U; Y|W) +I(V; Z|W) - I(U; V|W),
\end{align}
is an achievable rate pair.
The best outer bound on the capacity region of the broadcast channel is called the $UV$ outer bound~\cite{UV}. According to this outer bound for any achievable rate pair $(R_1, R_2)$ there is a distribution $p_{UVX}$ with the induced distribution $p_{UVXYZ}$ such that
\begin{align}\label{eq:UV}
R_1&\leq I(U; Y),\nonumber\\
R_2 & \leq I(V; Z),\nonumber\\
R_1+R_2&\leq \min\big\{I(U; Y) +I(V; Z|U), \, I(V; Z) +I(U; Y|V)\big\}.
\end{align}
As mentioned above, the Marton inner bound~\eqref{eq:Marton} and the UV outer bound~\eqref{eq:UV} do not match in general, and the capacity region of an arbitrary broadcast channel is not known.
A simple achievable rate region is derived by \emph{time division}.
Let $C_1=\max_{p_X} I(X; Y)$ be the capacity of the first channel $p(y|x)$ and $C_2=\max_{p_X} I(X; Z)$ be the capacity of the second channel $p(z|x)$. By ignoring the second receiver, the sender can transmit information to the first receiver at the highest possible rate, namely $C_1$. Thus $(R_1, R_2)=(C_1, 0)$ is achievable. Similarly $(R_1, R_2)= (0, C_2)$ is in the capacity region. Moreover, the sender can use time sharing; she can send information to the first receiver in $\alpha\in [0,1]$ fraction of uses of the channel, and then send information to the second receiver in the remaining uses of the channel. Then the rate pair $(R_1, R_2)=(\alpha C_1, (1-\alpha)C_2)$, for any $\alpha\in [0,1]$, is achievable. More precisely, the whole set
$$\mathcal R_{\text{\rm TD}}:=\Big\{ (R_1, R_2):\, \frac{R_1}{C_1} + \frac{R_2}{C_2}\leq 1,\, R_1, R_2\geq 0 \Big\},$$
which we call the \emph{time division} rate region, is in the capacity region.
The main result of this paper is a characterization of broadcast channels for which the time division rate region $\mathcal R_{\text{\rm TD}}$ is equal to the capacity region. Here is an informal statement of our main result.
\begin{theorem}\label{thm:informal} (Informal) Let $p(y|x)$ and $p(z|x)$ be two point-to-point channels with capacities $C_1$ and $C_2$ respectively. Suppose that $C_1\geq C_2$ and that the channels $p(y|x), p(z|x)$ satisfy some technical assumptions. Then $\mathcal R_{\text{\rm TD}}$ is equal to the capacity region of the broadcast channel $p(y, z|x)=p(y|x) p(z|x)$ if and only if either $C_1< C_2$ and
\begin{align}\label{eq:main-ineq}
\frac{I(X; Y)}{C_1} \leq \frac{I(X;Z)}{C_2}, \qquad \forall p_X,
\end{align}
or $C_1=C_2$ and the two channels are more capable comparable.
\end{theorem}
Recall that a channel $p(y|x)$ is called \emph{more capable} than $p(z|x)$ if for all input distributions $p_X$ we have
$$I(X; Y) \geq I(X;Z).$$
We say that two channels $p(y|x)$ and $p(z|x)$ are more capable comparable if either $p(y|x)$ is more capable than $p(z|x)$ or vice versa.
A partial characterization of \emph{degraded} broadcast channels for which time division is optimal is provided in~\cite[Theorem 3]{GGA14} that is similar to our characterization.
To prove this theorem we use some known facts about the set of \emph{capacity achieving distributions} of a point-to-point channel~\cite[Theorem 13.1.1]{CoverThomas}. For the convenience of the reader we also present the proofs of these facts.
\section{Capacity achieving distributions}
We use quite standard notations in this paper (see, e.g.,~\cite{ElGamalKim}). Sets are denoted by calligraphic letters such as $\mathcal X$, and a distribution on such a set is specified by a subscript as in $p_X$. Discrete memoryless point-to-point channels are determined by a set of conditional distributions $\{p_{Y|x}, x\in \mathcal X\}$, on a set $\mathcal Y$. For ease of notation we denoted such a channel by $p(y|x)$. For $\lambda\in [0,1]$ we use the notation $\bar \lambda:=1-\lambda$. To avoid confusions, when a mutual information $I(X; Y)$ is computed with respect to a distribution $p_{XY}$ we denote it by $I(X; Y)_p$.
Let $p(y|x)$ be a discrete memoryless point-to-point channel with capacity $C=\max_{p_X} I(X; Y)$.
For arbitrary distributions $p_X$ and $r_Y$ define
$$\psi(p_X, r_Y) = \sum_x p(x)D(p_{Y|x} \| r_Y ) = D(p_{XY}\| p_X r_Y) = I(X; Y)_p + D(p_Y\| r_Y),$$
where $D(\cdot \| \cdot)$ is the KL divergence, and $p_{XY}$ is the induced distribution on the input and output of the channel with input distribution $p_X$.
Then by the joint convexity of KL divergence and Sion's minimax theorem we have
$$\max_{p_X} \min_{r_Y} \psi(p_X, r_Y) = \min_{r_Y} \max_{p_X} \psi(p_X, r_Y).$$
Let us compute each side of the above equation. By the non-negativity of KL divergence we have
$$\max_{p_X} \min_{r_Y} \psi(p_X, r_Y) =\max_{p_X} \min_{r_Y} I(X; Y)_p + D(p_Y\| r_Y) = \max_{p_X} I(X; Y) = C.$$
On the other hand, by the linearity of $\psi$ in $p_X$ have
$$\min_{r_Y} \max_{p_X} \psi(p_X, r_Y) = \min_{r_Y} \max_{x_0} D(p_{Y|x_0}\| r_Y).$$
As a result,
\begin{align}\label{eq:min-max-q-Y}
\min_{r_Y} \max_{x_0} D(p_{Y|x_0}\| r_Y) = C.
\end{align}
Observe that the minimum in~\eqref{eq:min-max-q-Y} is achieved. So
let $r^*_Y$ be some optimal distribution there, i.e., $r^*_Y$ is such that
$$\max_{x_0} D(p_{Y|x_0}\| r^*_Y)=\max_{p_X} I(X; Y)_p + D(p_Y\| r^*_Y)=C.$$
Then for every $p_X$ we have
\begin{align}\label{eq:CA-ineq}
\psi(p_X, r^*_Y) =I(X; Y)_p + D(p_Y\| r^*_Y)\leq C.
\end{align}
Let $\Pi$ be the set of capacity achieving distributions:
$$\Pi=\arg\max_{p_X} I(X; Y).$$
Then by~\eqref{eq:CA-ineq} for every $p_X\in\Pi$ we have $D(p_Y\| r^*_Y) = 0$, i.e., $p_Y=r^*_Y$. This means that, for any capacity achieving distribution $p_X\in \Pi$ its induced distribution on the output of the channel is fixed, i.e., $r^*_Y=p_Y$. Indeed, the \emph{optimal output distribution} of a channel is unique, and is the unique distribution $r_Y^*$ that achieves the minimum in~\eqref{eq:min-max-q-Y}.
Let us define
\begin{align}\label{eq:def-K}
\mathcal K:=\{x:\, D(p_{Y|x}\| r^*_Y) = C \}.
\end{align}
Note that by the above discussion $\mathcal K$ is non-empty. Let $p_X$ be some distribution with $\text{supp}(p_X)\subseteq \mathcal K$, where
$\text{supp}(p_X):=\{x:\, p(x)>0 \}$. Then we have
\begin{align*}
I(X; Y)_p + D(p_Y\| r^*_Y) = \sum_x p(x) D(p_{Y|x}\| r^*_Y) = \sum_{x\in \mathcal K} p(x) D(p_{Y|x}\| r^*_Y) = C.
\end{align*}
This means that the inequality in~\eqref{eq:CA-ineq} becomes an equality for all $p_X$ with $\text{supp}(p_X)\subseteq \mathcal K$.
On the other hand, let $p_X\in \Pi$ be some capacity achieving distribution. Then by the above discussion, $p_Y = r^*_Y$. Moreover, we have
\begin{align*}
C & = I(X; Y)_p = \sum_x p(x) D(p_{Y|x}\| r^*_Y) \\
&= \sum_{x\in \mathcal K} p(x) D(p_{Y|x}\| r^*_Y) + \sum_{x\notin \mathcal K} p(x) D(p_{Y|x}\| r^*_Y) \\
& = p(\mathcal K) C + p(\mathcal X\setminus \mathcal K) \max_{x\notin \mathcal K} D(p_{Y|x}\| r^*_Y).
\end{align*}
As a result, we must have $p(\mathcal X\setminus \mathcal K) = 0$, i.e., $\text{supp}(p_X) \subseteq \mathcal K$.
Finally, suppose that $p_X$ is some distribution with $\text{supp}(p_X)\subseteq \mathcal K$ and $p_Y=r^*_Y$. Then~\eqref{eq:CA-ineq} becomes equality for $p_X$ and since $D(p_Y\| r^*_Y)=0$, $p_X$ is capacity achieving.
We summarize the above findings in the following proposition.
\begin{proposition}\label{prop:def-Pi}
For any point-to-point channel $p(y|x)$ there is a unique distribution $r^*_Y$ such that for all capacity achieving distributions $p_X\in \Pi$ we have $p_Y=r^*_Y$. Moreover, for any $p_X\in \Pi$ we have $\text{supp}(p_X)\subseteq \mathcal K$ where $\mathcal K$ is defined in~\eqref{eq:def-K}. Indeed, a given distribution $p_X$ is capacity achieving if and only if $\text{supp}(p_X)\subseteq \mathcal K$ and $p_Y= r^*_Y$. In particular $\Pi$ is convex.
\end{proposition}
The above proposition motivates the following definition. Define $\mathcal K_0$ to be the union of the supports of capacity achieving distributions, i.e.,
\begin{align}\label{eq:def-K-0}
\mathcal K':=\bigcap_{p_X\in \Pi} \text{supp}(p_X).
\end{align}
By the above proposition $\mathcal K'\subseteq \mathcal K$.
For a channel that has a capacity achieving distribution with full support (e.g., a channel for which the uniform distribution is capacity achieving) we have $\mathcal K'=\mathcal K=\mathcal X$. For example, this equality holds for binary symmetric and binary erasure channels.
Later we will see an example of a channel for which the inclusion $\mathcal K'\subseteq \mathcal K$ is strict.
\begin{proposition}\label{prop:K-p}
There exists $r_X\in \Pi$ such that $\text{supp}(r_X)=\mathcal K'$. Moreover, for any $p_X$ with $\text{supp}(p_X)\subseteq \mathcal K'$ we have
$$I(X; Y)_p + D(p_Y\| r^*_Y)=\sum_{x\in \mathcal K'} p(x)D(p_{Y|x} \| r^*_Y ) = C.$$
\end{proposition}
\begin{proof}
The existence of $r_X\in \Pi$ with $\text{supp}(r_X)=\mathcal K'$ follows from the definition of $\mathcal K'$ and the convexity of $\Pi$ established in Proposition~\ref{prop:def-Pi}. The second claim follows from $\mathcal K'\subseteq \mathcal K$.
\end{proof}
\section{Proof of the main result}
Let $p(y|x)$ and $p(z|x)$ be two channels with capacities $C_1$ and $C_2$ respectively. Let $r^*_Y$ and $s^*_Z$ be the optimal output distributions of the channels (as defined in the previous section) respectively. Also let $\Pi_1$ and $\Pi_2$ be their associated sets of capacity achieving distributions respectively. Finally let $\mathcal K_1, \mathcal K'_1$ and $\mathcal K_2, \mathcal K'_2$ be their associated subsets of $\mathcal X$ defined by~\eqref{eq:def-K} and~\eqref{eq:def-K-0}.
Here is the formal statement of our main result.
\begin{theorem}\label{thm:main-thm}
Suppose that $\mathcal K'_1=\mathcal K'_2=\mathcal X$ and $C_1\geq C_2$. Then the time division rate region $\mathcal R_\text{\rm TD}$ is the capacity region of the broadcast channel $p(y, z|x)=p(y|x)p(z|x)$ if and only if either $C_1< C_2$ and
\begin{align}\label{eq:main-ineq}
\frac{I(X; Y)}{C_1} \leq \frac{I(X;Z)}{C_2}, \qquad \forall p_X,
\end{align}
or $C_1=C_2$ and the two channels are more capable comparable.
\end{theorem}
\vspace{.09in}
The rest of this section is devoted to the proof of this theorem.\\
\noindent
$(\Rightarrow)$ First suppose that the time division region is the capacity region. That is, for any achievable rate pair $(R_1, R_2)$ we have
\begin{align}\label{eq:TD}
\frac{R_1}{C_1} + \frac{R_2}{C_2}\leq 1.
\end{align}
Let $r_{XU}$ and $s_{XV}$ be arbitrary distributions.
Define $p_{Q W \widetilde U \widetilde V X}$ by
\begin{align}\label{eq:margin-Q}
p(Q=0)=\lambda, \qquad p(Q=1)=\bar \lambda= 1-\lambda,
\end{align}
and according to the following table:
\begin{center}
\begin{tabular}{c|c|c|c}
& $W$ & $\widetilde U$ & $\widetilde V$ \\
\hline
$Q=0$ & $U$ & $X$ & \text{Const.}\\
$Q=1$ & $V$ & \text{Const.} & $X$
\end{tabular}
\end{center}
\label{default}
This table should be understood as follows. Firstو we have $\mathcal Q=\{0,1\}$ and the marginal distribution $p_Q$ is given by~\eqref{eq:margin-Q}. Second, we have $\mathcal W = \mathcal U\, \dot{\cup}\,\mathcal V$, $\widetilde{\mathcal U} = \mathcal X\,\dot{\cup}\,\{u^*\}$ and $\widetilde{\mathcal V} = \mathcal X\,\dot{\cup}\, \{v^*\}$ for two distinguished elements $u^*, v^*$. Third, the conditional distribution $p(w, \tilde u, \tilde v, x|q)$ is given by
\begin{align*}
p(w, \tilde u, \tilde v, x| Q=0) = \begin{cases}
r(X=x, U=w), & \tilde u=x, \tilde v=v^*, w\in \mathcal U,
\\
0, & \text{ otherwise,}
\end{cases}
\end{align*}
and
\begin{align*}
p(w, \tilde u, \tilde v, x| Q=1) = \begin{cases}
s(X=x, V=w), & \tilde v=x, \tilde u=u^*, w\in \mathcal V,
\\
0, & \text{ otherwise.}
\end{cases}
\end{align*}
Now let $\widetilde W = (Q, W)$ and consider the distribution $p_{\widetilde W\widetilde U\widetilde V XYZ}$ induced by the channel. Observe that $I(\widetilde U; \widetilde V| \widetilde W) = I(\widetilde U; \widetilde V| Q, W) =0$. Then by Marton's coding theorem~\eqref{eq:Marton} the rate pair $(R_1, R_2)$ given by
\begin{align*}
\begin{cases}
R_2= I(\widetilde V \widetilde W; Z) = I(\widetilde W; Z) + I(\widetilde V; Z|\widetilde W),
\\ R_1+ R_2 = \min\big\{ I(\widetilde W; Y), I(\widetilde W; Z) \big\} + I(\widetilde U; Y| \widetilde W) + I(\widetilde V; Z| \widetilde W),
\end{cases}
\end{align*}
is achievable.
Therefore, by our assumption we must have
\begin{align*}
\frac{R_1}{C_1} + \frac{R_2}{C_2} = \frac{R_1+R_2}{C_1} + \big(\frac{1}{C_2} - \frac{1}{C_1} \big) R_2 \leq 1,
\end{align*}
and then
\begin{align}\label{eq:bound-21}
\frac{1}{C_1} \min\big\{ I(\widetilde W; Y), I(\widetilde W; Z) \big\} + \big(\frac{1}{C_2} - \frac{1}{C_1} \big) I(\widetilde W; Z) + \frac{1}{C_1}I(\widetilde U; Y| \widetilde W) + \frac{1}{C_2}I(\widetilde V; Z|\widetilde W)\leq 1.
\end{align}
Let us compute individual terms in the above equation. We have
\begin{align}\label{eq:I-w-y}
I(\widetilde W; Y) = I(Q, W; Y) = I(Q; Y) + I(W; Y|Q) = I(Q; Y) + \lambda I(U; Y)_r + \bar \lambda I(V; Y)_s.
\end{align}
We similarly have
\begin{align}\label{eq:I-w-z}
I(\widetilde W; Z) = I(Q; Z) + \lambda I(U; Z)_r + \bar\lambda I(V; Z)_s.
\end{align}
Moreover, observe that
\begin{align*}
I(\widetilde U; Y| \widetilde W) & = \lambda I(X; Y| U)_r,
\end{align*}
and
\begin{align*}
I(\widetilde V; Z| \widetilde W) & = \bar \lambda I(X; Z| V)_s.
\end{align*}
Putting the above two equations in~\eqref{eq:bound-21} we find that
\begin{align}\label{eq:gen-ineq}
\frac{1}{C_1} \min\big\{ I(\widetilde W; Y), I(\widetilde W; Z) \big\} + \big(\frac{1}{C_2} - \frac{1}{C_1} \big) I(\widetilde W; Z) + \frac{\lambda}{C_1}I(X; Y| U)_r + \frac{\bar \lambda}{C_2}I(X; Z|V)_s\leq 1.
\end{align}
We can now consider two cases: either $I(\widetilde W; Y) \geq I(\widetilde W; Z)$ or $I(\widetilde W; Y) < I(\widetilde W; Z)$.
Then using $C_1\geq C_2$, equations~\eqref{eq:I-w-y} and~\eqref{eq:I-w-z}, and ignoring some non-negative terms in~\eqref{eq:gen-ineq} we find that
\begin{align*}
\begin{cases}
\frac{1}{C_1} \Big(\lambda I(U; Y)_r +\bar \lambda I(V; Y)_s \Big) +\big(\frac{1}{C_2} - \frac{1}{C_1} \big)\bar\lambda I(V; Z)_s+ \frac{\lambda}{C_1}I(X; Y| U)_r + \frac{\bar \lambda}{C_2}I(X; Z|V)_s\leq 1,\\
~\text{or}\\
\frac{1}{C_2} \Big(\lambda I(U; Z)_r +\bar \lambda I(V; Z)_s \Big) + \frac{\lambda}{C_1}I(X; Y| U)_r + \frac{\bar \lambda}{C_2}I(X; Z|V)_s\leq 1.
\end{cases}
\end{align*}
Now suppose that we chose $r_{XU}$ and $s_{XV}$ such that $r_X\in \Pi_1$ and $s_X\in \Pi_2$. Then using $C_1=I(X; Y)_r = I(UX; Y)_r = I(U; Y)_r + I(X; Y|U)_r$ and $C_2=I(X; Z)_s = I(VX; Z)_s = I(U; Z)_s + I(X; Y|V)_s$, by a simple algebra we arrive at
\begin{align*}
\begin{cases}
I(V; Y)_s \leq I(V; Z)_s,\\
~\text{or}\\
\frac{1}{C_2} I(U; Z)_r \leq \frac{1}{C_1}I(U; Y)_r.
\end{cases}
\end{align*}
Observe that the first inequality here depends only on $s_{XV}$ and the second one is solely in terms of $r_{XU}$. Then either the first one holds for every valid choice of $s_{XV}$ or the second one holds for every valid choice of $r_{XU}$. This means that either
\begin{align}\label{eq:O1}
I(V; Y) \leq I(V; Z), \qquad \forall s_{VX} \,\text{ s.t. }\, s_X\in \Pi_2,
\end{align}
or
\begin{align}\label{eq:O2}
\frac{1}{C_2}I(U; Z) \leq \frac{1}{C_1}I(U; Y), \qquad \forall r_{UX} \,\text{ s.t. }\, r_X\in \Pi_1.
\end{align}
Let us suppose that~\eqref{eq:O1} holds. Fix $s_X\in \Pi_2$ to be a capacity achieving distribution for $p(z|x)$ with $\text{supp}(s_X)=\mathcal K'_2=\mathcal X$ whose existence is guaranteed by Proposition~\ref{prop:K-p}. Let $p_X$ be an arbitrary distribution. Define $s_{VX}$ as follows. Let $\mathcal V=\{0,1\}$ and define $s(V=0)=\epsilon$ and $s(V=1)=1-\epsilon$. Also let
$$s(x|V=0)=p(x), \qquad s(x|V=1)=\frac{1}{1-\epsilon} s(x) - \frac{\epsilon}{1-\epsilon}p(x).$$
Observe that $\text{supp}(p_X)\subseteq \text{supp}(s_X)=\mathcal X$, so for sufficiently small $\epsilon>0$, both $s(x|V=0)$ and $s(x|V=1)$ are valid distributions. Then we obtain a distribution $s_{VX}$ on $\{0,1\}\times \mathcal X$ whose marginal on $\mathcal X$ is the distribution $s_X\in \Pi_2$ we started with. Since we assumed that~\eqref{eq:O1} holds, for any sufficiently small $\epsilon>0$ we have $I(V; Y)\leq I(V;Z)$. Now a simple computation verifies that $I(V;Y) = \epsilon D(p_Y\| s_Y)+\Theta(\epsilon^2)$ and $I(V; Z) = \epsilon D(p_Z\| s_Z) + \Theta(\epsilon^2)$. Therefore, we have
\begin{align}\label{eq:op-01}
D(p_Y\| s_Y) \leq D(p_Z\| s_Z^*), \qquad \forall p_X.
\end{align}
Starting from~\eqref{eq:O2} and following similar arguments we find that
\begin{align}\label{eq:op-02}
\frac{1}{C_2} D(p_Z \| r_Z) \leq \frac{1}{C_1}D(p_Y\| r^*_Y), \qquad \forall p_X,
\end{align}
where $r_X\in \Pi_2$ is a capacity achieving distribution for $p(y|x)$ with $\text{supp}(r_X)=\mathcal K'_1=\mathcal X$. Then either~\eqref{eq:op-01} or~\eqref{eq:op-02} is satisfied.
Let us in~\eqref{eq:op-01} and~\eqref{eq:op-02} restrict ourself to $p_X$ of the form $p(x) = \delta_{x, x_0}$, where $\delta_{x, x_0}$ denotes the Kronecker delta function and $x_0\in \mathcal X$ is arbitrary. Then either
\begin{align}\label{eq:op-11}
D(p_{Y|x_0}\| s_Y) \leq D(p_{Z|x_0}\| s_Z^*)=C_2, \qquad \forall x_0 ,
\end{align}
or
\begin{align}\label{eq:op-12}
\frac{1}{C_2} D(p_{Z|x_0} \| r_Z) \leq \frac{1}{C_1}D(p_{Y|x_0}\| r^*_Y) = 1, \qquad \forall x_0,
\end{align}
holds.
If~\eqref{eq:op-01} and then~\eqref{eq:op-11} hold, we have $\max_{x_0} D(p_{Z|x_0} \| s_Y) \leq C_2$, which using~\eqref{eq:min-max-q-Y} gives $C_1\leq C_2$. Then by our assumption $C_1\geq C_2$ we arrive at $C_1=C_2$. Therefore,~\eqref{eq:op-01} does not hold if $C_1> C_2$. Moreover, if $C_1=C_2$,~\eqref{eq:op-01} and~\eqref{eq:op-02} are symmetric. So in both cases, with no loss of generality we may assume that~\eqref{eq:op-02} and then~\eqref{eq:op-12} are satisfied.
We note that~\eqref{eq:op-12} implies that $\max_{x_0} D(p_{Z|x_0} \| r_Z)\leq C_2$. Thus $r_Z$
is an optimal output distribution for $p(z|x)$, and by its uniqueness $r_Z=s^*_Z$. Then~\eqref{eq:op-02} reduces to
$$\frac{1}{C_2} D(p_Z \| s^*_Z) \leq \frac{1}{C_1}D(p_Y\| r^*_Y),\qquad \forall p_X.$$
On the other hand, by Proposition~\ref{prop:K-p} we have $D(p_Y\| r^*_Y) = C_1- I(X; Y)_p$ and $D(p_Z\| s^*_Z)= C_2-I(X; Z)_p$.
Using these in the above inequality gives~\eqref{eq:main-ineq}.
\vspace{.2in}
\noindent $(\Leftarrow)$ We now prove the converse. We assume that either $C_1> C_2$ and~\eqref{eq:main-ineq} holds, or $C_1=C_2$ and the two channels are more-capable comparable. In the latter case, by symmetry with no loss of generality we assume that $p(y|x)$ is more capable than $p(z|x)$. Then in both cases~\eqref{eq:main-ineq} holds. Assuming this
we show that time division is optimal for the broadcast channel $p(y|x)p(z|x)$.
Let $(R_1, R_2)$ be an achievable rate pair. By the $UV$ outer bound~\eqref{eq:UV}, there exists $p_{UVX}$ such that
\begin{align*}
R_2&\leq I(V; Z),\\
R_1+R_2&\leq I(V; Z) + I(U; Y|V)\leq I(V; Z) + I(X; Y|V).
\end{align*}
Then using the fact that $C_1\geq C_2$ we find that
\begin{align*}
\frac{R_1}{C_1} + \frac{R_2}{C_2} &\leq \frac{I(X; Y|V)}{C_1} + \frac{I(V; Z)}{C_2}\\
& \leq \frac{I(X; Z|V)}{C_2} + \frac{I(V; Z)}{C_2}\\
& \leq \frac{I(X; Z)}{C_2}\\
& \leq 1,
\end{align*}
where in the second line we use~\eqref{eq:main-ineq}. We are done.
\section{Example}
In the statement of Theorem~\ref{thm:main-thm} we assume that $\mathcal K'_1=\mathcal K'_2 = \mathcal X$. This may seem an unnecessary technical assumption that is forced by our proof method.
Here we give an example to illustrate that Theorem~\ref{thm:main-thm} does not hold without it.
Let $\mathcal A, \mathcal B$ be two finite disjoint sets with $|\mathcal A| \geq |\mathcal B|\geq 2$. Let $\mathcal X=\mathcal A\cup \mathcal B$, $\mathcal Y=\mathcal A$ and $\mathcal Z=\mathcal B$. Define channels $p(y|x)$ and $p(z|x)$ by
\begin{align*}
p(y| x) =
\begin{cases}
\delta_{x,y} &\quad x\in \mathcal A,\\
\frac{1}{|\mathcal B|} &\quad x\in \mathcal B,
\end{cases}
\qquad\qquad
p(z| x) =
\begin{cases}
\frac{1}{|\mathcal B|} &\quad x\in \mathcal A,\\
\delta_{z, x} &\quad x\in \mathcal B,
\end{cases}
\end{align*}
where $\delta$ denotes the Kronecker delta function.
The uniform distribution on the subset $\mathcal A\subset \mathcal X $ is the unique capacity achieving distribution of $p(y|x)$, and its optimal output distribution is the uniform distribution on $\mathcal Y$. Then $\mathcal K'_1 = \mathcal A$. Likewise the unique capacity achieving distribution of $p(z|x)$ is the uniform distribution on $\mathcal B\subset \mathcal X$ and we have $\mathcal K'_2 = \mathcal B$.
Moreover, $C_1=\log |\mathcal A| \geq \log |\mathcal B|= C_2$.
Also note that~\eqref{eq:main-ineq} does not hold, nor the two channels are more-capable comparable. Indeed, if $p_X$ is a non-trivial distribution supported only on $\mathcal A$, then $I(X; Y)_p>0$ while $I(X; Z)_p=0$. Similarly if $p_X$ is non-trivial and supported only on $\mathcal B$, then $I(X; Y)_p=0$ while $I(X; Z)_p>0$.
Nevertheless, we show in the following that time division is optimal for the broadcast channel $p(y|x)p(z|x)$.
Let $(R_1, R_2)$ be an achievable rate pair. Again by the $UV$ outer bound~\eqref{eq:UV}, there exists $p_{UVX}$ such that
\begin{align*}
R_2&\leq I(V; Z),\\
R_1+R_2&\leq I(V; Z) + I(U; Y|V)\leq I(V; Z) + I(X; Y|V).
\end{align*}
Then using the fact that $C_1\geq C_2$ we find that
\begin{align}\label{eq:A123}
\frac{R_1}{C_1} + \frac{R_2}{C_2} &\leq \frac{I(X; Y|V)}{C_1} + \frac{I(V; Z)}{C_2}\nonumber\\
& \leq \frac{I(X; Y)}{C_1} + \frac{I(V; Z)}{C_2},
\end{align}
where in the second line we use the fact that $V-X-Y$ forms a Markov chain.
Let $Q$ be a binary random variable that equals $0$ if $X\in \mathcal A$ and equals $1$ if $X\in \mathcal B$.
Then we have
\begin{align*}
I(X; Y) & = I(XQ; Y)\\
& = H(Y) - p(Q=0) H(Y| X, Q=0) -p(Q=1) H(Y| X, Q=1)\\
& = H(Y) - p(Q=1) \log |\mathcal A |\\
& \leq p(Q=0) C_1.
\end{align*}
We similarly have $I(X; Z) \leq p(Q=1) C_2$. Putting these in~\eqref{eq:A123} we find that $R_1/C_1 + R_2/C_2\leq 1$. Therefore, time division is optimal for $p(y|x)p(z|x)$.
\vspace{.23in}
\noindent\textbf{Acknowledgements.} The author is thankful to Chandra Nair for several fruitful discussions about broadcast channels. The author was supported in part by Institute of Network Coding of CUHK and by GRF grants 2150829 and 2150785.
|
\section{Introduction}
Bike-sharing transportation systems have been presented as an ecological and user-friendly transportation
mode, which appears to be well complementary to classic public transportation
systems (\cite{midgley2009role}). The quick propagation of many implementations
of such systems across the world confirms the
interesting potentialities that bike-sharing can offer \cite{demaio2009bike}. \noun{O'Brien} \& \textit{al}. propose in \cite{o2013mining} a review on the current state of bike-sharing across the world. Inspired by the relatively good success of such systems in Europe,
possible key factors for their quality have been questioned and
transposed to different potential countries such as
China (\cite{liu2012solving,geng2009bike}) or the United States (\cite{gifford2004will}).
The understanding of system mechanisms is essential for its optimal exploitation. That can be done
through statistical analysis with predictive statistical models (\cite{borgnat2009modelisation,borgnat2009spatial,borgnat2009studying,borgnat2011shared})
or data-mining techniques (\cite{o2013mining,kaltenbrunner2010urban}),
and can give broader results such as structure
of urban mobility patterns. Concerning the implementation, a crucial point in the design of the
system is an optimal location of stations. That problem have been
extensively studied from an Operational Research point of view (\cite{lin2011hub,lin2011strategic}
for example). The next step is a good exploitation of the system.
By nature, strong asymmetries appear in the distribution of bikes: docking stations in residential areas are emptied during the day contrary to working areas. That causes in most
cases a strong decrease in the level
of service (no parking places or no available bikes for example).
To counter such phenomena, operators have redistribution strategies that have also been well studied and for which optimal plans have been proposed (\cite{kek2006relocation,nair2011fleet,nair2013large}).
However, all these studies always approach the problem from a top-down
point of view, in the sense of a centralized and global approach of the issues, whereas bottom-up strategies (i. e. local actions that would allow the emergence of desired patterns) have been to our knowledge
not much considered in the literature. User-based methods have been
considered in \cite{barth1999simulation,barth2004trb} in the case
of a car-sharing system, but the problem stays quite far from a behavioral
model of the agents using the system, since it explores the possibility
of implication of users in the redistribution process, or of shared
travels what is not relevant in the case of bikes. Indeed the question
of a precise determination of the influence of users behaviors and parameters on the
level of service of a bike-sharing systems remains open. We propose an agent-based model of simulation in order to
represent and simulate the system from a bottom-up approach, considering
bikers and parking as stations as agents and representing their interactions
and evolutions in time. That allows to explore user-targeted strategies for an increase of the level of service, as the incitation to use online information media or to be more flexible on the destination point. Note that our work aims to explore effects of user-based policies, but does not pretend to give recommendations to system managers, since our approach stays technical and eludes crucial political and human aspects that one should take into account in a broader system design or management context.
The rest of the paper is organized as follows. The model and indicator used to quantify its behavior are described in Section 2. Next, Section 3 presents the implementation and results, including internal and external validations of the model by sensitivity analysis and simplified calibration on real data, and also exploration of possible bottom-up strategies for system management. We conclude by a discussion on the applicability of results and on possible developments.
\section{Presentation of the model}
\paragraph{Introduction}
The granularity of the model is the scale of the individual
biker and of the stations where bikes are parked. A more integrated
view such as flows would not be useful to our purpose since we want
to study the impact of the behavior of individuals on the overall
performance of the system. The global working scheme consists in agents
embedded in the street infrastructure, interacting with particular
elements, what is inspired from the core structure of the Miro model
(\cite{banos2011simuler}). Spatial scale is roughly the scale of the district; we don't consider the whole system for calculation
power purposes (around 1300 stations on all the system of Paris, whereas
an interesting district have around 100 stations), what should not
be a problem as soon as in- and outflows allow to reconstruct travels entering and getting out of the area. Tests on larger spatial zones showed that generated travel were quite the same, justifying this choice of scale. Focusing on some particular districts is important since issues with level of service occur only in narrow areas. Time scale of a run is logically one full day because
of the cyclic nature of the process (\cite{vogel2011understanding}).
\begin{wrapfigure}[15]{o}{0.4\columnwidth}%
\centering
\includegraphics[trim=1cm 3cm 1cm 3.5cm ,scale=0.28]{flowchart.pdf}
\caption{\footnotesize Flowchart of the decision process of bikers, from the start of their travel to the drop of the bike.}
\label{fig:1}
\end{wrapfigure}%
\paragraph{Formalisation}
The street network of the area is an euclidian network $(V\subset\mathbb{R}^{2},E\subset V\times V)$
in a closed bounded part of $\mathbb{R}^{2}$. The time is discretized
on a day, so all temporal evolution are defined on $T=[0,24]\cap\tau\mathbb{N}$
with $\tau$ time step (in hours). Docking stations $S$ are particular
vertices of the network for which constant capacities $c(s\in S)$
are defined, and that can contain a variable number of bikes $p_{b}(s)\in\{0,\ldots,c\}^{T}$.
We suppose that temporal fields $O(x,y,t)$ and $D(x,y,t)$ are defined,
corresponding respectively to probabilities that a given point at
a given time becomes the expected departure (resp. the expected arrival)
of a new bike trip, knowing that a trip starting (resp. arriving) at that time exists. Boundaries conditions
are represented as a set of random variables $(N_{I}(i,t))$.
For each possible entry point $i\in I$ ($I\subset V$ is a given
set of boundaries points) and each time, $N_{I}(i,t)$ gives the number
of bikes trips entering the zone at point $i$ and time $t$. For
departures, a random time-serie $N_{D}(t)$ represents the number
of departures in the zone at time $t$. Note that these random variables
and probabilities fields are sufficient to built the complete process
of travel initiation at each time step. Parametrization of the model
will consist in proposing a consistent way to construct them from
real data.
Docking stations are fixed agents, only their functions $p_{b}$ will
vary through time. The other core agents are the bikers, for which
the set $B(t)$ is variable. A biker $b\in B(t)$ is represented
by its mean speed $\bar{v}(b)$, a distance $r(b)$ corresponding to its ``propensity to walk'' and a boolean $i(b)$ expressing the capacity of having access to information on the whole system at any time (through a mobile device and the dedicated application for example). The initial set of bikers $B(0)$ is taken empty, as $t=0$ corresponds to 3a.m. when there is approximately no travels on standard days.
We define then the workflow of the model for one time
step. The following scheme is sequentially executed for each $t\in T$,
representing the evolution of the system on a day.
For each time step the evolution of the system follows this process :
\begin{itemize}
\item Starting new travels. For a travel within the area, if biker has information,
he will adapt his destination to the closest station of its destination
with free parking places, if not his destination is not changed.
\begin{itemize}
\item For each entry point, draw number of new traveler, associate to each
a destination according to $D$ and characteristics (information drawn
uniformly from proportion of information, speed according to fixed
mean speed, radius also).
\item Draw new departures within the area according to $O$, associate either
destination within (in proportion to a fixed parameter $p_{it}$,
proportion of internal travels) the area, or a boundary point (travel
out of the area). If the departure is empty, biker walks to
an other station (with bikes if has information, a random one if not)
and will start his travel after a time determined by mean walking
speed and distance of the station.
\item Make bikers waiting for start for which it is time begin their journey
(correspond to walkers for which a departure station was empty at
a given time step before)
\end{itemize}
\item Make bikers advance of the distance
corresponding to their speed. Travel path is taken as the shortest path between origin and destination, as effective paths are expected to have small deviation from the shortest one in urban bike travels~\cite{borgnat2009spatial}.
\item Finish travels or redirect bikers
\begin{itemize}
\item if the biker was doing an out travel and is on a boundary point, travel
is finished (gets out of the area)
\item if has no information, has reached destination and is not on a station, go to a random station
within $r(b)$
\item if is on a station with free places, drop the bike
\item if is on a station with no places, choose as new destination either
the closest station with free places if he has information, or a random
one within $r(b)$ (excluding already visited ones, implying the memory
of agents).
\end{itemize}
\end{itemize}
Fig. \ref{fig:1} shows the decision process for starting and arriving bikers.
Note that walking radius $r(b)$ and information $i(b)$ have implicitly
great influence on the output of the model, since dropping station
is totally determined (through a random process) by these two parameters
when the destination is given.
\paragraph{Evaluation criteria}
In order to quantify the performance of the system, to compare different
realizations for different points in the parameter space or to evaluate
the fitness of a realization towards real data, we need to define
some functions of evaluation, proxies of what are considered as ``qualities''
of the system.
\paragraph{Temporal evaluation functions}
These are criteria evaluated at each time step and for which the output
on the all shape of the time-series will be compared.
\begin{itemize}
\item Mean load factor
$\bar{l}(t)=\frac{1}{\left|S\right|}\sum_{s\in S}\frac{p_{b}(s)}{c(s)}$
\item Heterogeneity of bike distribution: we aggregate spatial heterogeneity
of load factors on each station through a standard normalized heterogeneity
indicator, defined by
$h(t)=\frac{2}{\sum_{s\neq s'\in S}\frac{1}{d(s,s')}}\cdot\sum_{
s\neq s'\in S}\frac{\left|\frac{p_{b}(s,t)}{c(s)}-\frac{p_{b}(s',t)}{c(s')}\right|}{d(s,s')}
$
\end{itemize}
\paragraph{Aggregated evaluation functions}
These are criteria aggregated on a all day quantifying the level of
service integrated on all travels. We note $\mathcal{T}$ the set
of travels for a realization of the system and $\mathcal{A}$ the
set of travel for which an ``adverse event'' occured, i. e. for
which a potential dropping station was full or a starting station
was empty. For any travel $v\in\mathcal{T}$, we denote by $d_{th}(v)$
the theoretical distance (defined by the network distance between
origin and initial destination) and $d_{r}(v)$ the effective realized
distance.
\begin{itemize}
\item Proportion of adverse events: proportion of users for which the quality
of service was doubtful.
$A=\frac{\left|\mathcal{A}\right|}{\left|\mathcal{T}\right|}$
\item Total quantity of detours: quantification of the deviation regarding
an ideal service
$D_{tot}=\frac{1}{\left|\mathcal{T}\right|}\cdot\sum_{v\in\mathcal{T}}\frac{d_{r}(v)}{d_{th}(v)}$
\end{itemize}
We also define a fitness function used for calibration of the model
on real data. If we note $(lf(s,t))_{s\in S,t\in T}$ the real time-series
extracted for a standard day by a statistical analysis on real data,
we calibrate on the mean-square error on all time-series, defined
for a realization of the model by
\[
MSE=\frac{1}{\left|S\right|\left|T\right|}\sum_{t\in T}\sum_{s\in S}(\frac{p_{b}(s,t)}{c(s)}-lf(s,t))^{2}
\]
\section{Results}
\subsection{Implementation and parametrization}
\paragraph{Implementation}
The model was implemented in NetLogo (\cite{NetLogo}) including GIS
data through the GIS extension. Preliminary treatment of GIS data
was done with QGIS (\cite{QGIS_software}). Statistical pre-treatment
of real temporal data was done in R (\cite{R}), using the NL-R extension
(\cite{thiele2012agent}) to import directly the data. For complete reproducibility, source code (including data collection scripts, statistical R code and NetLogo agent-based modeling code) and data (raw and processed) are available on the open git repository of the project at \url{http://github.com/JusteRaimbault/CityBikes}.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{running}
\caption{\footnotesize Example of the graphical output of the model for
a particular district (Chatelet). The map shows docking stations, for each the color gradient from green to red gives the current loading factor (green : empty, red : full).}
\label{fig:2}
\end{figure}
Concerning the choice of the level of representation in the graphical
interface, we followed \noun{Banos} in \cite{banos2013HDR} when he
argues that such exploratory models can really be exploited only if
a feedback through the interface is possible. It is necessary to find
a good compromise for the quantity of information displayed in the
graphical interface. In our case, we represent a map of the district,
on which link width is proportional to current flows, stations display
their load-factor by a color code (color gradient from green, $lf(s)=0$,
to red, $lf(s)=1$). Bikes are also represented in real time, what
is interesting thanks to an option that allow to follow some individuals
and visualize their decision process through arrows representing original
destination, provenance and new destination (should be implemented
in further work). This feature could be seen as superficial at this
state of the work but it appears as essential regarding possible further
developments of the project (see discussion section). Fig. 2 shows
an example of the graphical interface of the implementation of the
model of simulation.
\paragraph{Data collection}
All used data are open data, in order to have good reproducibility
of the work. Road network vector layer was extracted from OpenStreetMap
(\cite{bennett2010openstreetmap}). Time-series of real stations statuts
for Paris were collected automatically%
\footnote{from the dedicated website api.jcdecaux.com%
} all 5 minutes during 6 month and were imported into R for treatment with \cite{couture2013rjson}
and the point dataset of stations was created from the geographical
coordinates with \cite{keitt2011rgdal}.
\paragraph{Parametrization}
The model was designed in order to have real proxies for most of parameters.
Mean travel speed is taken as $\bar{v}=$14km.h$^{-1}$ from \cite{jensen2010characterizing},
where data of trips where studied for the bike system of the city
of Lyon, France. To simplify, we take same speed for all bikers : $v(b)=\bar{v}$. A
possible extension with tiny gaussian distribution around mean speed
showed in experiments to bring nothing more. It has been shown in
\cite{o2013mining} that profiles of use of bike systems stays approximatively
the same for european cities (but can be significantly different for
cities as Rio or Taipei), what justify the use of these inferred data
in our case. We also use the determined mean length of travel from
\cite{nair2013large} (here that parameter should be more sensible
to the topology so we prefer extract it from this second paper although
it seems to have subsequent methodological bias compared to the first
rigorous work on the system of Lyon), which is 2.3km, in order to
determine the diameter of the area on which our approach stays consistent.
Indeed the model is built in order to have emphasis on travels coming
from the outside and on travels going out, internal travels have to
stay a small proportion of all travels. In our case, a district of
diameter 2km gives a proportion of internal travels $p_{it} \approx 20\%$.
We will take districts of this size with this fixed proportion in
the following.
\begin{wrapfigure}[28]{o}{0.5\columnwidth}%
\centering
\includegraphics[trim=0cm 0cm 0cm 6cm,width=0.5\textwidth]{histogram}
\caption{\footnotesize
Statistical analysis of outputs.\\
For some aggregated outputs (here the overall quantity of detours
and the proportion of adverse events), we plotted histograms of the
statistical distribution of the functions on many realizations of
the model for a point in the parameter space. Two points of the parameter space, corresponding to $(r=300,p_{info}=50,\sigma=80)$ (green histogram) and $(r=700,p_{info}=50,\sigma=80)$ (red) are plotted
here as examples. Gaussian fits are also drawn. The relative good fit shows
the internal consistence of the model and we are able to quantify
the typical number of repetitions needed when applying the model : supposing normal distributions for the indicator and its mean, a 95\% confidence interval of size $\sigma/2$ is obtained with $n=(2\cdot2\sigma\!\cdot\!1.96/\sigma)^2\approx60$
}
\label{fig:3}
\end{wrapfigure}%
The crucial part of the parametrization is the construction of $O,D$
fields and random variables $N_{I},N_{D}$ from real data. Daily data were reduced through sampling of time-series
of load-factors of all stations and dimension of the representation
of a day was significantly reduced through a $k$-means clustering
procedures (classically used in time-series clustering as it is described
in \cite{warren2005clustering}). These reduced points were then clustered
again in order to isolate typical weekdays from week-ends, where the
use profiles are typically different and from special days such as
ones with very bad climate or public transportation strikes. That
allowed to create the profile of a ``standard day'' that was used
to infer $O,D$ fields through a spatial Gaussian multi-kernel estimation
(see \cite{tsybakov2004introduction}). The characteristic size of
kernels $1/\sigma$ is an essential parameter for which we have no
direct proxy, and that will have to be fixed through a calibration procedure.
The laws for $N_{I},N_{D}$ were taken as binomial: for an actual
arrival, we consider each possible travel and increase the number
of drawing of each binomial law of entries by 1 at the time corresponding
to mean travel time (depending on the travel distance) before arrival
time. Probabilities of binomial laws are $\nicefrac{1}{Card(I)}$
since we assume independence of travels. For departure, we just increase
by one drawings of the binomial law at current time for an actual departure.
\begin{wrapfigure}[30]{o}{0.5\columnwidth}%
\centering
\includegraphics[trim=1cm 2cm 1cm 1cm,width=0.5\textwidth]{calib3d}
\caption{\footnotesize Simplified calibration procedure.\\
We plot the surface of the mean-square error on time-series of load-factors
as a function of the two parameters on which we want to calibrate.
For visibility purpose, only one surface was represented out of the
different obtained for different values of walking radius. The absolute
minimum obtained for very large kernel has no sense since such value
give quasi-uniform probabilities because of total recovering of Gaussian
kernels. We take as best realization the second minimum, which is
located around a kernel size of 50 and a quantity of information of
30\%, which seem to be reasonable values afterwards.}
\label{fig:3}
\end{wrapfigure}
What we call parameter space in the following consists in the 3 dimensional
space of parameters that have not been fixed by this parametrization,
i. e. the walking radius $r$ (taken as constant on all bikers, as
for the speed), the information proportion $p_{info}$ what is the probability for a new biker to have information and the "size"
of the Gaussian kernels $\sigma$ (note that the spread of distributions is decreasing with $\sigma$).
\subsection{Robustness assessment, exploration and calibration}
\paragraph{Internal consistence of the model}
Before using simulations of the model to explore possible strategies, it is necessary to assess that the results produced are internally consistent, i. e. that the randomness introduced in the parametrization and in the internal rules do not lead to divergences in results. Simulations were launched on a large number of repetitions for different points in the parameter space and statistical distribution of aggregated outputs were plotted. Fig. 3 shows example of these results. The relative good gaussian fits and the small deviation of distributions confirm the internal consistence of the model. We obtain the typical number of repetitions needed to have a 95\% confidence interval of length half of the standard deviation, what is around 60, and we take that number in all following experiments and applications. These experiments allowed a grid exploration of the parameter space, confirming expected behavior of indicators. In particular, the shape of $MSE$ suggested to use the simplified calibration procedure presented in the following.
\paragraph{Robustness regarding the study area}
The sensitivity of the model regarding geometry of the area was also tested. Experiments described afterwards were run on comparable districts (Ch{\^a}telet, Saint-Lazare and Montparnasse), leading to the same conclusions, what confirms the external robustness of the model.
\paragraph{Reduced calibration procedure}
Using experiments launched during the grid exploration of the parameter
space, we are able to assess or the regularity of some aggregated
criteria, especially of the mean-square error on loads factors of
stations. We calibrate on kernel size and quantity of information.
For different values of the walking radius, the obtained area for
the location of the minimum of the mean-square error stays quite the
same for reasonable values of the radius (300-600m). Fig. 4 shows
an example of the surface used for the simplified calibration. We
extract from that the values of around 50 for kernel size and 30 for
information proportion. The most important is kernel size since we
cannot have real proxy for that parameter. We use these values for
the explorations of strategies in the following.
\subsection{Investigation of user-based strategies}
\paragraph{Influence of walking radius}
Taking for kernel-size and quantity of information the values
given by the calibration, we can test the influence of walking radius
on the performance of the system. Note that we make a strong assumption,
that is that the calibration stay valid for different values of the
radius. As we stand previously, this stays true as soon as we stay
in a reasonable range of values (we obtained 300m to 600m) for the
radius. The influence of variations of walking radius on indicators were tested. Most interesting results are shown in figure \ref{fig:5}. Concerning the indicators evaluated on time-series ($h$ and $\bar{l}(t)$), it is hard to
have a significant conclusion since the small difference that one
can observe between curves lies inside errors bars of all curves.
For $A$, we see a decreasing of the indicator
after a certain value (300m), what is significant if we consider that
radius under that value are not realistic, since a random place in
the city should be at least in mean over 300m from a bike station. However, the results concerning the radius are not so concluding, what could be due to the presence of periodic negative feedbacks: when the mean distance between two stations is reached, repartitions concerns neighbor stations as expected, but the relation is not necessarily positive, depending on the current status of the other station. A deeper understanding and exploration of the behavior of the model regarding radius should be the object of further work.
\begin{figure}
\centering
\subfloat[{\footnotesize Time series of heterogeneity
indicator $h(t)$ for different values of walking radius. Small
differences between means could mislead to
a positive effect of radius on heterogeneity, but the error bars
of all curves recover themselves, what makes any conclusion non-significant.}]
{\includegraphics[trim=1cm 0.5cm 1cm 0cm,width=0.49\textwidth]{hetero}}
\hfill{}
\subfloat[{\footnotesize Influence of walking radius
on the quantity of adverse events $A$. After 400m, we observe a relative decrease
of the proportion. However, values under 300-400m should be ignored since these
are smaller than the mean distance of a random point to a station.}]
{\includegraphics[trim=0cm 0.5cm 1cm 0cm,width=0.49\textwidth]{adverseRadius}}
\caption{Results on the influence of walking
radius.}
\label{fig:5}
\end{figure}
\paragraph{Influence of information}
For the quantity of information, we are on the contrary able to draw
significant conclusions. Again, behavior of indicators were studied according to variations of $p_{info}$. Most significant are shown on figure \ref{fig6}. Results from time-series are also not concluding, but concerning aggregated indicators, we have a constant and regular
decrease for each and for different values of the radius. We are able
to quantify a critical value of the information for which most of
the progress concerning indicator $A$ (adverse events) is done, that is around 35\%.
We observe for this value an amelioration of 4\% in the quantity of
adverse events, that is interesting when compared to the total number
of bikers. Regarding the management strategy for an increase in the level of service, that implies an increase of the penetration rate of online information tools (mobile application e. g.) if that rate is below 50\%. If it is over that value, we have shown that efforts for an increase of penetration rate would be pointless.
\section{Discussion}
\subsection{Applicability of the results}
We have shown that increases of both walking radius and information quantity could have positive consequences on the level of service of the system, by reducing the overall number of adverse events and the quantity of detours especially in the case of the information. However, we can question the possible applicability of the results. Concerning walking radius, first a deeper investigation would be needed for confirmation of the weak tendency we observed, and secondly it appears that in reality, it should be infeasible to play on that parameter. The only way to approach that would be to inform users of the potential increase in the level of service if they are ready to make a little effort, but that is quite optimistic to think that they will apply systematically the changes, either because they are egoistic, because they won't think about it, or because they will have no time.
Concerning the information proportion, we cannot also force users to have information device (although a majority of population owns such a device, they won't necessarily install the needed software, especially if that one is not user-friendly). We should proceed indirectly, for example by increasing the ergonomics of the application. An other possibility would to improve information displayed at docking stations that is currently difficult to use.
\begin{figure}
\centering
\subfloat[{\footnotesize Influence of proportion
of information on adverse events $A$ for two different values of walking
radius. We can conclude significantly that the information has a positive
influence. Quantitatively, we extract the threshold of 35\% that corresponds
to the majority of decrease, that means that more is not necessarily
needed.}]
{\includegraphics[trim=0cm 0cm 0cm 1cm,width=0.48\textwidth]{adverse}
}\hfill{}
\subfloat[{\footnotesize Influence of information
on quantity of detours $D_{tot}$. Curves for $r=300m$ and $r=700m$ are shown (scale color). Here also, the influence is positive. The
effect is more significant for high values of walking radius. The inflection is around 50\% of users informed, what is more than for adverse events.}]
{\includegraphics[trim=0cm 0cm 0cm 1cm,width=0.48\textwidth]{detours}
}
\caption{Results on the influence of proportion
of information.}
\label{fig6}
\end{figure}
\subsection{Possible developments}
\paragraph{Other possible management strategies}
Concerning user parameters, other choices could have been made, as including waiting time at a fixed place, either for a parking or a bike. The parameters chosen are both possible to influence and quite adapted to the behavioral heuristic used in the model, and therefore were considered. Including other parameters, or changing the behavioral model such as using discrete choice models may be possible developments of our work. Furthermore, only the role of user was so far explored. The object of further investigation could be the role of the ``behavior'' of docking stations. For example, one could fix rules to them, as close all parkings over a certain threshold of load-factor, or allow only departures or parkings in given configurations, etc. Such intelligent agents would surely bring new ways to influence the overall system, but will also increase the level of complexity (in the sense of model complexity, see \cite{varenne2013modeliser}), and therefore that extension should be considered very carefully (that is the reason why we did not integrate it in this first work).
\paragraph{Towards an online bottom-up pilotage of the bike-sharing system}
Making the stations intelligent can imply making them communicate
and behave as a self-adapting system. If they give information to
the user, the heterogeneity of the nature and quantity of information
provided could have strong impact on the overall system. That raises
of course ethical issues since we are lead to ask if it is fair to
give different quantities of information to different users. However, the perspective of a bottom-up piloted system could be of great interest from a theoretical and practical point of view.
One could think of online adaptive algorithms
for ruling the local behavior of the self-adapting system, such as
ant algorithms (\cite{monmarche2004algorithmes}), in which bikers
would depose virtual pheromones when they visit a docking station (corresponding
to their information on travel that is easy to obtain), that would allow the system to take some local decisions of redirecting bikers or closing stations for a short time in order to obtain an overall better level of service. Such methods have already been studied to improve level of service in other public transportation systems like buses~\cite{10.1371/journal.pone.0021469}.
\section*{Conclusion}
This work is a first step of a new bottom-up approach of bike-sharing
systems. We have implemented, parametrized and calibrated a basic
behavioral model and obtained interesting results for user-based strategies
for an increase of the level of service. Further work will consist in a deeper validation of the model, its application on other data. We suggest also to explore developments such as extension to other types of agents (docking stations), or the study of possible bottom-up online algorithm
for an optimal pilotage of the system.
\tiny
\bibliographystyle{unsrt}
|
\section{Introduction}
The currently progressing serious reductions in diversity of ecosystems force us to consider in depth the relationships between ecosystem stability and species abundance distributions (SADs) quantifying community diversity~\cite{McCann:00,May:99}. Theoretical knowledge about SADs for systems at a given trophic level has greatly advanced in the last decade, based on Hubbell's neutral theory~\cite{Hubbell:01,Volkov:03,Alonso:06,Etienne:07}.
Meanwhile, two pioneering works~\cite{Gardner:70,May:72} opened the possibility of theoretical treatment of more complicated systems with multiple trophic levels, using linear models which continue to provide useful suggestions~\cite{Allesina:12}.
Beyond the linear model, a nonlinear model called replicator dynamics (RD) has been employed to study ecosystems, and has offered qualitative knowledge about the global behavior of population dynamics~\cite{Rieger:89,Diederich:89,Oliveira:00,Oliveira:02,Oliveira:03,Tokita:04,Galla:06}. In the RD, the population of species evolves by its own fitness function which consists of two contributions: interactions with other species and self productivity of the species itself. Depending on the complicacy of the interactions, the RD yields various SADs~\cite{Oliveira:00,Oliveira:02,Oliveira:03,Tokita:04}. Although it is a highly nontrivial task to identify the fitness functions of species in a given real community, the RD provides a useful description of real ecosystems in a qualitative level in SADs. However, analytical treatment of it has so far been limited to the case where each species interacts with all other species. These “dense” interactions are not only seemingly unrealistic but also involve an undesirable simplification in SADs, due to the extensive sum of contributions in the fitness function. Thus, it is expected that novel and various SADs can be observed when “sparse” interactions, or specialist-specialist interactions, are employed.
\section{Problem setting}
Consider a community consisting of $N$ species, denote the fitness function of $i$th species by $F_i$, and assume that $F_i$ consists of two terms of pairwise interactions $J_{ij}$ and of self productivity $u$
\begin{eqnarray}
F_{i}(\V{x}|\V{J},u)=\frac{1}{2}\sum_{j(\neq i)}J_{ij}x_j-\frac{1}{2}u x_i,
\end{eqnarray}
where $x_i(\geq 0)$ denotes the $i$th species' population. We assume the productivity $u$ is common among all the species, to purely see the effect of interactions. In the RD, each species is governed by the following equation
\begin{eqnarray}
\frac{d x_{i}}{dt}=x_i \left( F_{i} -\bar{F}\right),
\Leq{RE}
\end{eqnarray}
where the averaged fitness is introduced as
\begin{eqnarray}
\bar{F}(\V{x}|\V{J},u)=\frac{\sum_{i}x_iF_{i}(\V{x}|\V{J},u)}{\sum_i x_i}.
\end{eqnarray}
The total population $\sum_{i=1}^{N}x_i$ is preserved in the RD, and without loss of generality we normalize this to the number of species as $\sum_{i=1}^{N}x_i=N$.
Properties of interactions crucially influence the dynamics. Here we treat “sparse” interactions: the number of interacting species of the $i$th species, $c_i$, is bounded by a fixed constant $c_{\rm max}$ independent of $N$. In addition, we investigate the case of symmetric relations, where $J_{ij} = J_{ji}$. This case is relatively simple, since the dynamics necessarily converges to a fixed point depending on the initial condition. Despite this simplicity, symmetric RDs can describe several communities such as a certain class of selection equation in population genetics and a classical model of community competing for resources~\cite{MacArthur:67}.
For symmetric interactions, the averaged fitness $\bar{F}$ plays the role of a Lyapunov function whose local maxima correspond to fixed points of the dynamics. Among those maxima, we focus on the global maximum stated as
\begin{eqnarray}
\V{x}^*=
\mathop{\rm arg~max}\limits_{ \{ x_i \geq 0\}_{i=1}^{N} }
\left\{
\bar{F} \left( \V{x}| \V{J},u \right)
\right\},
\Leq{extremization}
\end{eqnarray}
and the SAD is defined as
\begin{eqnarray}
P(x)=\frac{1}{N}\sum_{i=1}^{N}\delta\left( x-x_i^*\right),
\end{eqnarray}
where $\delta(x)$ is the Dirac delta function. A typical shape of the SAD resembles a skewed lognormal distribution~\cite{Preston:62a,May:75}. We investigate how these quantities change if the interactions become sparse.
\section{Result}
Our analysis is based on direct evaluation of \Req{extremization}, neglecting the non-negativity of $x_i$. This treatment provides exact results as long as no extinct species exist, which is the case for moderate values of $u$, but becomes less precise as the proportion of extinct species grows. Despite this limitation, our analysis is sufficient to observe novel interesting SADs as seen below.
Taking a direct variation of the averaged fitness with respect to $\V{x}$ on the above assumption and imposing the constraint $\sum_{i}x_i=N$, we get
\begin{eqnarray}
x_{i}^*=N\frac{\sum_{j}K^{-1}_{ij}}{\sum_{i,j}K^{-1}_{ij}},
\Leq{x-general}
\end{eqnarray}
where $K=uI-J$ and $I$ is the unit matrix. To obtain clear information from \Req{x-general}, the perturbative expansion with respect to $u^{-1}$ is performed under the assumption that $u$ is sufficiently large. The inverse of the matrix $K$ is expanded as
\begin{eqnarray}
K^{-1}=\left( u I - J \right)^{-1}=\frac{1}{u}\sum_{p=0}^{\infty} u^{-p}J^{p}.
\end{eqnarray}
Inserting this into \Req{x-general} gives
\begin{widetext}
\begin{eqnarray}
x^{*}_{i}=\frac{
1+u^{-1}\sum_{j}J_{ij}+u^{-2}\sum_{j,k}J_{ij}J_{jk}+u^{-3}\sum_{j,k,l}J_{ij}J_{jk}J_{kl}+\cdots
}{
1+u^{-1}\frac{1}{N}\sum_{i,j}J_{ij}+u^{-2}\frac{1}{N}\sum_{i,j,k}J_{ij}J_{jk}+
u^{-3}\frac{1}{N}\sum_{i,j,k,l}J_{ij}J_{jk}J_{kl}+\cdots
}.
\Leq{x-series}
\end{eqnarray}
\end{widetext}
Retaining higher order terms gives more precise results, but the computation becomes more complicated. Up to the second order of $u^{-1}$, the topology of the network does not affect the result and a clear discussion is possible. It is instructive to see an explicit formula of the SAD, the distribution of $x^{*}_i$, in the first-order expansion
\begin{widetext}
\begin{eqnarray}
P(x)\approx \int d\V{J} P(\V{J}) \delta\left( x- \left\{ 1+u^{-1}\left( \sum_{j}J_{ij}-(1/N)\sum_{i,j}J_{ij} \right) \right\} \right),
\Leq{P(x)-first}
\end{eqnarray}
\end{widetext}
where $P(\V{J})$ is the distribution of the interactions $\V{J}$. Hence, if the distribution of interactions exhibits a certain discrete nature, i.e. if the marginal distribution $P_{ij}(J_{ij})\equiv \int \prod_{ \Ave{k,l}( \neq \Ave{i,j}) } dJ_{kl} P(\V{J})$ consists of discriminable multiple peaks, the SAD correspondingly takes multiple peaks. The mechanism of multiple peaks here is based only on two assumptions: that $u$ is sufficiently large and that the distribution of interactions is discrete, providing the possibility of theoretically explaining multiple peaks observed in several field studies~\cite{Dornelas:08,Gray:05,Magurran:03}. Note that similar multiple peaks were observed in the RD with a specific distribution of dense interactions~\cite{Oliveira:02,Oliveira:03}, but we stress that our mechanism producing multiple peaks is completely different from the one in~\cite{Oliveira:02,Oliveira:03}. Their model's interactions are given by Hebb's rule of $p$ binary traits, meaning that all species are strictly categorized into $p+1$ groups by the symmetry in the $N\to \infty$ limit, and accordingly the SAD consists of $p+1$ (or less if multiple groups are extinct) delta peaks. In contrast to those delta peaks, our theory can naturally provide rounded discrete peaks and as well standard (lognormal like) functional forms by changing a single parameter $u$. Thus far such a flexibility has been absent in the densely interacting networks~\cite{Rieger:89,Diederich:89,Oliveira:00,Oliveira:02,Oliveira:03,Tokita:04,Galla:06}
One interesting outcome of treating sparse interactions is all species coexistence at moderate values of $u$. This coexistence starts to collapse at a critical value of $u_c$. A general upper bound of the critical value, $u_{c}^{\rm (UB)}$, can be derived by examining the condition where the numerator of \Req{x-series} vanishes. Each term in the numerator is bounded as
\begin{eqnarray}
\left| \sum_{j_1,j_2,\cdots,j_{p}}J_{j_1j_2}J_{j_2 j_3}\cdots J_{j_{p-1}j_p} \right| \leq c_{\rm max}^{p}J^{p}_{\rm \max},
\end{eqnarray}
where $J_{\rm max}$ is the maximum absolute value of the pairwise interaction $J_{ij}$. Thus, the numerator can be bounded from below
\begin{eqnarray}
&&
1+u^{-1}\sum_{j}J_{ij}+u^{-2}\sum_{j,k}J_{ij}J_{jk}+\cdots
\nonumber \\
&&
\geq 1-\sum_{p=1}^{\infty}\left( \frac{c_{\rm max} J_{\rm max}}{u} \right)^p=\frac{u-2c_{\rm max} J_{\rm max}}{u-c_{\rm max}}.
\Leq{u_c-bound}
\end{eqnarray}
This gives the general upper bound as
\begin{eqnarray}
u_c^{\rm (UB)}=2c_{\rm max}J_{\rm max}.
\Leq{upper bound}
\end{eqnarray}
As long as $u>u_c^{\rm (UB)}$, all species coexistence is guaranteed. This is actually observed in \Rfig{PD} for a specific choice of $P(\V{J})$ (see below for the detail). A similar general upper bound can be derived even if the self interactions can vary depending on species as long as its mean is large enough. In that case, the expansion with respect to $J/u$ is replaced to the one with respect to $J/U$ where $U$ is a diagonal matrix whose entries are species-dependent self interactions.
Beyond the perturbation and the general bounding method, we can provide a non-perturbative theory. Terms of higher orders than $2$ can be categorized into those with and without loops. For example, in the case of third-order terms, $\sum_{j,k,l}J_{ij}J_{jk}J_{kl}$, terms of $J_{ij}J_{jk}J_{ki}~(j\neq i,j\neq k,k\neq i)$ have a closed loop, and other terms do not have loops. It is known that summing terms without loops is possible for all orders in a systematic manner, which is known to be equivalent to the so-called Bethe approximation, or cavity method~\cite{ADVA,INFO,Kabashima:12}. For a single instance of network, this method evaluates influences to a newcomer species from an existing species, which are termed cavity biases, in a self-consistent manner.
For clarity of the following discussion based on the cavity method, we treat a single degree network $c_i=c$, and fix the distribution of interactions as
\begin{eqnarray}
P(J_{ij})=\frac{1+\Delta}{2}\delta(J_{ij}-1)+\frac{1-\Delta}{2}\delta(J_{ij}+1).
\Leq{P(J_{ij})}
\end{eqnarray}
This simple function maintains a discrete nature and also has a parameter, $\Delta$, controlling the ratio of mutualistic and competitive relations. This setup enables us to see the effects of productivity $u$, of the ratio of mutualistic relations $\Delta$, and of the degree of network $c$, in an unified manner. The distribution of the cavity biases, $\hat{q}(\hat{H})$, satisfies the following self-consistent equation:
\begin{eqnarray}
\hspace{-5mm}
\hat{q}(\hat{H})=\int\prod_{l=1}^{c-1}d\hat{H}_{l}\hat{q}(\hat{H}_{l})
\overline{
\delta \left( \hat{H}-J\hat{a} \left( r+\sum_{l=1}^{c-1}\hat{H}_{l} \right) \right)
},
\Leq{B2B}
\end{eqnarray}
where $\overline{\cdots}$ denotes the average over the interaction $J$, and $\hat{a}$ and $r$ are given by the external parameters as
\begin{eqnarray}
&&
\hat{a}=\frac{u - \sqrt{u^2-4(c-1)}}{2(c-1)},
\\
&&
r=( u-c\hat{a} )\left( 1- \frac{ c\hat{a}\Delta }{ 1+\hat{a}\Delta } \right).
\end{eqnarray}
Using the solution of \Req{B2B}, the SAD $P(x)$ is assessed as
\begin{eqnarray}
P(x)=
\int \prod_{l=1}^{c} d\hat{H}_l \hat{q}(\hat{H}_l) \, \delta\left( x- \frac{r+\sum_{l=1}^{c}\hat{H}_l}{u-c\hat{a}}\right).
\Leq{P(x)}
\end{eqnarray}
Solving \Req{B2B} and inserting the solution into \Req{P(x)} give the results shown in \Rfigss{PD}{diversity-u}. This gives an exact treatment for the interaction network without loops, as well as for the randomly generated network which has some global loops but their influence can be neglected in the large size limit.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.7\columnwidth]{Figure1.eps}
\caption{Critical values of productivity $u_c$ plotted against $\Delta$ for $c=3$. The points are the non-perturbative results, and the dashed line corresponds to the second-order perturbation. Both of the values locate well below the general upper bound $u_{c}^{\rm (UB)}=2c=6$. }
\Lfig{PD}
\end{center}
\end{figure}
\Rfig{PD} displays $u_c$ plotted against $\Delta$ for $c=3$. The points are based on the non-perturbative theory, and the dashed line is derived by keeping only up to the second-order terms of $u^{-1}$ in \Req{x-series}. The qualitative shape of the $u_c$ curve is captured by the second-order approximation, but the quantitative deviation is not small. An interesting, and perhaps somewhat counter-intuitive, observation is the dependence on $\Delta$. Larger values of $\Delta$ provide more mutualistic relations, and hence \Rfig{PD} shows more mutualistic communities tend to be more extinct-prone, in the sense that extinct species start to appear even at larger $u$. The critical value $u_c$ drastically drops off around $\Delta=\pm1$, which exhibits singularities at those points. This is natural since $\Delta=\pm 1$ corresponds to the case where all the species are equal and thus $x_i=1(\forall i)$ holds irrespective of the value of $u$. We also calculate the $c$-dependence of $u_c$ for fixed $\Delta$ and find that $u_c$ monotonically increases as $c$ grows.
\begin{figure*}
\begin{center}
\includegraphics[width=0.6\columnwidth]{Figure2a.eps}
\includegraphics[width=0.6\columnwidth]{Figure2b.eps}
\includegraphics[width=0.6\columnwidth]{Figure2c.eps}
\caption{ The SADs for $\Delta=-0.8,0$ and $0.8$ at $c=3$.}
\Lfig{SADs}
\end{center}
\end{figure*}
\Rfig{SADs} shows the SAD for several values of $u$ and $\Delta$ at $c=3$. For $u=5.5(>u_c)$, the distribution is symmetric about $x=1$ for $\Delta=0$, though it is biased to $x>1$ or to $x<1$ for $\Delta \neq 0$. For the competitive case $\Delta=-0.8$, the largest peak appears with $x<1$, and the long tail persists in the $x>1$ region, while for the mutualistic case $\Delta=0.8$ the opposite is true. As $u$ decreases, the discreteness becomes weaker and the functional forms become closer to a standard skewed lognormal distribution, irrespective of $\Delta$, as seen at $u=3(<u_c)$. Hence, our theory smoothly connects the standard functional form and that with multiple peaks, by a single parameter $u$. The $c$-dependence of the SAD appears as the number of peaks for large $u$ which is equal to $c+1$, as understood by \Req{P(x)-first}. For small $u$, tails of the SADs in $x<0$ appear because we neglect the non-negativity constraint. These tails mean there occurs a partial extinction in the corresponding parameters. In this situation, our analysis is not exact, but it is possible to interpret the cumulative distribution in $x<0$ as an approximation of the ratio of the extinct species.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\columnwidth]{Figure3a.eps}
\includegraphics[width=0.48\columnwidth]{Figure3b.eps}
\caption{The diversity $\alpha(0)$ and $\alpha(1)$ are plotted against $\Delta$ for $u=6$ (left) and $3$ (right) at $c=3$. The dependence on $\Delta$ is not monotonic. }
\Lfig{diversity-u}
\end{center}
\end{figure}
An interesting finding appears in the survival function $\alpha(y)=1-\int_{0_{-}}^{y}P(x)dx$, an index quantifying community diversity~\cite{Tokita:04}. \Rfig{diversity-u} shows $\alpha(1)$ and $\alpha(0)$ plotted against $\Delta$ for $u=6(>u_c)$ and $u=3(<u_c)$ at $c=3$. We observe non-monotonic behavior of diversity,
and particularly $\alpha(1)$ shows an oscillating behavior as $\Delta$ changes. This is related to the multiple peaks of the SAD: the height and location of the highest peak and the tail sensitively change as $\Delta$ deviates, leading to nontrivial behavior of the diversity.
To see the robustness of SAD discreteness in large $u$ regions discussed so far, we perform a numerical simulation. \Rfig{hist-univ} shows the result for the case where the interaction network is a random graph of the degree $c=3$ and the values of $J_{ij}$ are drawn from a sum of two Gaussian distributions
\begin{eqnarray}
P(J_{ij})=\frac{1}{\sqrt{8V\pi}}\left\{
e^{-\frac{(J_{ij}-1)^2}{2V} }+e^{ -\frac{(J_{ij}+1)^2}{2V} }
\right\},
\Leq{J-dist-Gauss}
\end{eqnarray}
with $V=0.1$. Two peaks are symmetrical and thus correspond to $\Delta=0$. The result clearly demonstrates the presence of multiple peaks. This is not trivial, since the value of $J_{ij}$ drawn from \Req{J-dist-Gauss} is not bounded.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\columnwidth]{Figure4.eps}
\caption{The SAD with the interactions generated from \Req{J-dist-Gauss}, the network structure of which is a random graph of single degree $c=3$. The parameters are $u=6.03$, $N=16000$. }
\Lfig{hist-univ}
\end{center}
\end{figure}
We also perform other numerical simulations in several different situations, and the result is given in \Rfig{SAD-several}. The left panel is for the square lattice; $N=16384$ and $u=8$. The middle panel is for a heterogeneous random network with degrees $c=3,4,5,6$, each ratio of which is $p=0.45,0.35,0.1,0.1$, respectively; $N=1000$ and $u=10$. The right panel is for random self interactions, whose values are uniformly taken from $( 9.4,10.6 )$, and the network is single degree $c=3$ and the simulated size is $N=1000$. In all these cases the interactions are unbiased binary.
\begin{figure*}
\begin{center}
\vspace{0mm}
\includegraphics[width=0.6\columnwidth]{SAD-square.eps}
\includegraphics[width=0.6\columnwidth]{SAD-hetero.eps}
\includegraphics[width=0.6\columnwidth]{SAD-randu.eps}
\vspace{0mm}
\caption{The SADs for the square lattice with unbiased binary interactions (left), a heterogeneous random network with degrees $c=3,4,5,6$ and unbiased binary interactions (middle), and random self interactions on a single degree network $c=3$ with unbiased binary interactions. }
\Lfig{SAD-several}
\end{center}
\end{figure*}
All these numerical simulations universally exhibit multiple peaks in the SADs, as long as $P(J_{ij})$ has a discrete nature and the self interactions are sufficiently large, and thus the presented mechanism is fairly robust. Therefore, we again stress the importance of this mechanism in explaining actually observed multiple peaks in several field studies~\cite{Dornelas:08,Gray:05,Magurran:03}.
\section{Conclusion}
Our analysis of the RD with sparse interactions has provided several nontrivial behavior in the SADs: coexistence of all the species, multiple peaks. These consequences were transparently understood from the perturbative expansion assuming the self interaction $u$ is large enough and the interactions have a certain discrete nature. The general upper bound of critical value of $u$, below which extinct species emerge, has also been evaluated.
We also provided a non-perturbative theory which is exact on tree-like networks without loops, to obtain more quantitative information. As a result, exact critical values of $u$ and the SADs' dependence on model parameters have been calculated for a specific network. We have found a nontrivial dependence of diversity on the ratio of mutualistic relations and a drastic change of the abundance distribution's shape from the one with multiple peaks to a standard skewed lognormal distribution. We stress that this drastic change is controlled by a few parameters, the self interaction $u$ and the ratio of mutualistic relations $\Delta$. Thus our theory provides a possibility of unifying different shapes of the abundance distributions. Comparison with experimental observations is highly desired.
Our results so far are derived by investigating the RD with a few assumptions, and thus they can be applied to other contexts in which the RD appears, such as population genetics, game theory, and chemical networks in living cells~\cite{Hofbauer:98,Nowak:06}.
\section*{Acknowledgements} This work was supported by a Grant-in-Aid for JSPS Fellows (No. 2011) (TO), KAKENHI Nos. 26870185 (TO), 25120013 (YK), and 24570099 (KT). KT also acknowledges support in part by the project ``Creation and Sustainable Governance of New Commons through Formation of Integrated Local Environmental Knowledge'', at the Research Institute for Humanity and Nature (RIHN), and the project ``General Communication Studies'', at the International Institute for Advanced Studies (IIAS).
|
\section{Introduction}
Two key signatures of the Quark Gluon Plasma- collective flow and jet quenching- have been thoroughly measured experimentally over the years at the Relativistic Heavy Ion Collider RHIC at BNL and the Large Hadron Collider LHC at CERN. Initially, it was thought that the soft physics regime (low transverse momentum $p_T$ where collective flow is measured) and the hard physics regime (high $p_T$ where jet physics is relevant) could be tackled separately. The early measurements of high $p_T$ elliptic flow \cite{Adare:2010sp} demonstrated flow was possible even within the hard sector though theoretical calculations \cite{Wang:2000fq,Gyulassy:2000gk,Shuryak:2001me} taking into account only energy loss physics and simplified assumptions for the medium evolution were not able to reproduce the large measured $v_2$ and, in fact, the difficulties in simultaneously describing $R_{AA}\otimes v_2$ remained for years to come (see the discussion and references in \cite{Betz:2014cza,Xu:2014tda}).
\begin{figure}[h]\centering
\includegraphics[width=26pc]{dis2030pic.pdf}\hspace{2pc}%
\caption{\label{flowdis}Distributions of $\varepsilon_2$ at LHC for MCGlauber and MCKLN (with two different smoothing scales) produce initial conditions that are at one end very circular and at the other end elliptical.}
\end{figure}
Meanwhile, comparisons of relativistic hydrodynamical models to collective flow measurements had a breakthrough circa 2010 \cite{Alver:2010gr} with the inclusion of event-by-event calculations \cite{Paiva:1996nv,Gyulassy:1996br,Aguiar:2001ac,Socolowski:2004hw,Andrade:2006yh} to make comparisons to nonzero triangular flow measurements \cite{Alver:2010gr}. Event-by-event fluctuating initial conditions imply that within the same centrality class (i.e. holding the density constant) there is a wide range of initial shapes (eccentricities $\varepsilon_{n}$) that relativistic hydrodynamics map into the final flow harmonics \cite{Gardim:2011xv,Gardim:2014tya}. In Fig. \ref{flowdis} the distribution of $\varepsilon_2$ are shown for MCGlauber and MCKLN \cite{Drescher:2006ca} (smoothing out initial energy density fluctuations to $\lambda=0.3$ fm and $\lambda=1$ fm- see \cite{Noronha-Hostler:2015coa} for more details). Regardless of the initial conditions, a wide range of initial elliptical shapes are produced (and more complex shapes such as ``triangles", ``squares" etc are also produced), which experimentally are proven via $v_n$ distributions \cite{Aad:2013xma}.
\begin{figure}[h]\centering
\includegraphics[width=20pc]{mcklnvn2.pdf}\hspace{2pc}%
\caption{\label{softv2} $v_2\{2\}$ and $v_3\{2\}$ for MCGlauber and MCKLN (with two different smoothing scales) at LHC.}
\end{figure}
Event-by-event viscous hydrodynamical models have been enormously successful at describing soft physics observables such as the cumulants of flow harmonics (as shown in Fig.\ \ref{softv2}) as well as even making predictions for the highest LHC energies \cite{Noronha-Hostler:2015uye}. While there has been some progress made in combining the knowledge of hydrodynamical backgrounds \cite{Betz:2015mlf,Crkovska:2016flo} and fluctuating initial conditions \cite{Zhang:2012ha} with energy loss models, it was not until \cite{Noronha-Hostler:2016eow} that full event-by-event viscous hydrodynamical backgrounds (v-USPhydro \cite{Noronha-Hostler:2013gga,Noronha-Hostler:2014dqa}) were combined with an energy loss model (BBMG \cite{Betz:2014cza,Betz:2011tu,Betz:2012qq}). The effects of event-by-event fluctuations as well as the correlation between soft and hard flow harmonics in \cite{Noronha-Hostler:2016eow} provided the key step needed in order to solve the longstanding $R_{AA}\otimes v_2$ puzzle.
\section{Event-by-event viscous hydrodynamics+energy loss}
In \cite{Noronha-Hostler:2016eow} there were two crucial differences between all previous calculations of $v_2(p_T>10 GeV)$:
\begin{enumerate}
\item Full event-by-event viscous hydrodynamical backgrounds were fed into the energy loss model to obtain $v_2^{hard}$.
\item $v_2^{exp}(p_T)$ from Eq. (\ref{eqn:e2exp}) was calculated correlating the soft integrated $v_2^{soft}$ and hard $v_2^{hard}(p_T)$ to simulate the two particle (one soft, one hard) correlations used to measure elliptical flow experimentally.
\end{enumerate}
The soft $v_2^{soft}$ was taken using the best fitting parameters in the soft sector, see \cite{Noronha-Hostler:2016eow} for more details. The $v_2^{hard}$ (the 2nd Fourier coefficient of $R_{AA}(p_T,\phi)$) is computed using the BBMG code \cite{Betz:2014cza,Betz:2011tu,Betz:2012qq} in the pQCD-like scenario. In order to make comparisons to experimental data then the soft and hard flow harmonics are correlated via
\begin{equation}\label{eqn:e2exp}
v^{exp}_n(p_T)=\frac{\langle v^{soft}_n\,v_n^{hard}(p_T)\cos\left[n\left(\psi^{soft}_n-\psi^{hard}_n(p_T)\right]\right) \rangle}{\sqrt{\left\langle \left(v^{soft}_n\right)^{2}\right\rangle}},
\label{softhard}
\end{equation}
where $\langle\ldots\rangle$ denote event averages and $\psi^{soft}_n$ and $\psi^{hard}_n(p_T)$ are the soft event plane angle and hard event plane angle, respectively, as was discussed in \cite{Luzum:2013yya}. Note that if one assumes a smoothed, averaged hydrodynamical background then there is only one ``event" and Eq.\ (\ref{eqn:e2exp}) reduces down to $v^{exp}_n(p_T)|_{(ic=avg)}=v_n^{hard}(p_T)$.
The enhancement of $v_2$ due to the soft-hard $v_n$ correlation in Eq.\ (\ref{eqn:e2exp}) becomes clear if one considers linear response (a good approximation as long as one is not considering peripheral collisions \cite{Noronha-Hostler:2015dbi}). In the soft sector we can assume $v_2^{soft}\sim c\, \varepsilon_2$ and for the hard sector since we have a $p_T$ dependent quantity one takes $v_2^{hard}\sim \chi_2(p_T) \varepsilon_2$. Substituting them into Eq.\ (\ref{eqn:e2exp}), one finds
\begin{equation}\label{eqn:rms}
v^{exp}_2(p_T)\sim\chi_2(p_T) \sqrt{\langle \varepsilon_2^2\rangle}.
\end{equation}
Since $\langle\varepsilon_2\rangle<\sqrt{\langle \varepsilon_2^2\rangle}$ one can see that event-by-event fluctuations will always increase the flow harmonics (soft $v_2$ is also proportional to the root mean squared not the mean \cite{Ollitrault:2009ie}). Note that the use of Eq. (\ref{eqn:e2exp}) is not the only source of enhancement for $v_2$, rather it is the combination of full event-by-event hydrodynamically expanding backgrounds with the correct calculation of Eq.\ (\ref{eqn:e2exp}) that leads to allows for a simultaneous description and solution of the $R_{AA}\otimes v_2$ puzzle.
\section{Results}
\begin{figure}[h]\centering
\includegraphics[width=26pc]{RAA_v2.pdf}\hspace{2pc}%
\caption{\label{RAAv2}Comparison of event-by-event calculations in
mid-central $\sqrt{s}=2.76$ TeV Pb+Pb collisions of $R_{AA}$ and $v_2$ from v-USPhydro+BBMG \cite{Noronha-Hostler:2016eow} to experimental data at LHC from ALICE \cite{ALICE_RAA,ALICE_v2_v3}, CMS \cite{CMS_RAA,CMS_v2}, and ATLAS \cite{ATLAS_v2}.}
\end{figure}
In Fig. \ref{RAAv2} the result of the v-USPhydro+BBMG calculations are shown for $R_{AA}$ and $v_2$ compared to ALICE \cite{ALICE_RAA,ALICE_v2_v3}, CMS \cite{CMS_RAA,CMS_v2}, and ATLAS \cite{ATLAS_v2} data at the LHC. It is clear that the inclusion of event-by-event fluctuations with soft-hard $v_2$ correlation was a needed ingredient in order to bring up the elliptical flow to the experimental data. In fact, one can see in Fig. \ref{RAAv2} that while $R_{AA}$ is relatively insensitive to the effect of initial conditions and event-by-event fluctuations, $v_2$ has a clear splitting between MCGlauber and MCKLN initial conditions exactly as in the soft sector (see that in Fig. \ref{softv2} the $v_2$ from MCKLN is always larger than from MCGlauber due to the larger eccentricities of MCKLN).
\begin{figure}[h]\centering
\includegraphics[width=13pc]{v3.pdf}\hspace{2pc}%
\caption{\label{v3}Comparison of event-by-event calculations in
mid-central $\sqrt{s}=2.76$ TeV Pb+Pb collisions of $v_3$ from v-USPhydro+BBMG \cite{Noronha-Hostler:2016eow} to experimental data at LHC from ALICE \cite{ALICE_v2_v3}.}
\end{figure}
In fact, the similarities in the soft sector also hold true for $v_3$ at high $p_T$ as seen in Fig. \ref{v3}. MCKLN produces smaller $\varepsilon_3$ than MCGlauber, which correlates to a smaller $v_3$ both in the soft sector Fig. \ref{softv2} and in the hard sector Fig. \ref{v3}. Furthermore, the simple existence of a non-zero $v_3$ at high $p_T$ demonstrates that event-by-event fluctuations are needed. Finally, an interesting avenue to explore for the future is the decorrelation between the soft and hard the event-planes for triangular flow (and above) \cite{Jia:2012ez}.
\section{Conclusions and Outlook}
In this talk, it was shown that that event-by-event fluctuations combined with the proper calculation of $v_2^{exp}$ at high $p_T$ provided a natural solution to the $R_{AA}\otimes v_2$ puzzle. Event-by-event fluctuations have been clearly measured already at both RHIC and LHC and they have been shown to be necessary to describe $v_3$ \cite{Alver:2010gr}, $v_n$ distributions \cite{Aad:2013xma}, event shape engineering \cite{Adam:2015eta}, $SC(n,m)$ correlations \cite{ALICE:2016kpq,Giacalone:2016afq}, and higher order $v_n\{m\}$ cumulants \cite{Chatrchyan:2012ta,Abelev:2014mda,Aad:2014vba}. Thus, a proper comparison to high $p_T$ $v_2$ experimental data necessarily requires the inclusion of event-by-event hydrodynamic fluctuations as well as the calculation of the soft-hard correlation in Eq.\ (\ref{softhard}).
\begin{figure}[h]\centering
\includegraphics[width=26pc]{ese.pdf}\hspace{2pc}%
\caption{\label{ese}Within the same centrality class the path length can vary strongly on an event-by-event basis producing different $R_{AA}$'s.}
\end{figure}
Once one accepts that event-by-event fluctuations also play a key role in the hard sector, it opens up an entire new field of possible calculations. In \cite{Noronha-Hostler:2016eow} it was suggested that event-shape engineering could be preformed also in the hard sector. Fig.\ \ref{ese} illustrates the underlying concept. For the same centrality class, e.g. density, one has a wide range of eccentricities $\varepsilon_n$'s. It is already known that the shape of the event determines the path length seen by the jet, which is why centrality class selection is so important. However, using centrality class alone as an estimate for the path length is inherently flawed because within a single centrality class there are multiple possible path lengths depending on the eccentricities of the event. Thus, one possibility is to select on events with the same soft $p_T$ flow harmonic in order to hold the path length constant within the centrality class. Additionally, comparisons can be made while holding the density fixed and changing only the path length by selecting on a variety of soft $p_T$ flow harmonics to study the effect of energy loss for a very specific path length. Qualitatively, one can see from Fig.\ \ref{ese} that highly eccentric events have a shorter path length (since statistically its much more likely for a jet to cross the thin part of the event vs. the long part) whereas circular events have the longest average path length.
Experimental studies at ATLAS \cite{Aad:2015lwa} and more recently at CMS \cite{CMS:2016uwf} have already begun comparing soft-hard flow harmonics. Both demonstrate a very clear linear correlation between the soft and hard $v_2$ (as shown in \cite{Noronha-Hostler:2016eow}) further solidifying the connection between $v_2$ and $\sqrt{\langle \varepsilon_2^2\rangle}$ in Eq.\ (\ref{eqn:rms}). Additionally, CMS \cite{CMS:2016uwf} found a convergence of the cumulants of flow harmonics that is a good indication that there is collective flow even at high $p_T$.
The success of v-USPhydro+BBMG in fitting $v_2(p_T>10 GeV)$ makes one think whether the hard flow harmonics are also just as sensitive to the usual hydrodynamic parameters such as the initial time $\tau_0$, the switching temperature $T_{sw}$, the transport coefficients etc. While the heavy flavor shows a strong dependence on the heavy quark drag coefficient \cite{Das:2015ana}, it also shows a very strong dependence on a temperature dependent shear viscosity to entropy density ratio $\eta/s(T)$ \cite{Esha:2016svw}. It is natural to wonder how a temperature dependent $\eta/s(T)$ would affect high $p_T$ flow given its connection with the jet transport coefficient $\hat{q}$ \cite{Majumder:2007zh}. Furthermore, using jet physics one can even extract a minimum at $T_c$ \cite{Xu:2014tda} analogous to what is found in the hadronic phase \cite{NoronhaHostler:2008ju,NoronhaHostler:2012ug}. Since there is still a wide variation in $\eta/s(T)$ calculations \cite{Noronha-Hostler:2015qmd} any additional constraint on the values of transport coefficients is sorely needed.
The most obvious step towards the future of combining soft and hard physics is to include hard scattering effects within relativistic hydrodynamics in order to study the intermediate $p_T$ range of flow harmonics that are not possible with only relativistic hydrodynamics or energy loss models on their own. Initial work has been done in this direction \cite{Crkovska:2016flo,Andrade:2014swa,Pang:2012he} with the hope that this will be explored in more depth in the near future.
\section*{References}
|
\section{Introduction}
The domain of representation learning has undergone tremendous advances due to improved supervised learning techniques. However, unsupervised learning has the potential to leverage large pools of unlabeled data, and extend these advances to modalities that are otherwise impractical or impossible.
One principled approach to unsupervised learning is generative probabilistic modeling. Not only do generative probabilistic models have the ability to create novel content, they also have a wide range of reconstruction related applications including inpainting \citep{theis2015generative, oord2016pixel, DBLP:conf/icml/Sohl-DicksteinW15}, denoising \citep{balle2015density}, colorization \citep{zhang2016colorful}, and super-resolution \citep{bruna2015super}.
As data of interest are generally high-dimensional and highly structured, the challenge in this domain is building models that are powerful enough to capture its complexity yet still trainable.
We address this challenge by introducing \emph{real-valued non-volume preserving (real NVP) transformations}, a tractable yet expressive approach to modeling high-dimensional data.
This model can perform {\em efficient and exact} inference, sampling and log-density estimation of data points. Moreover, the architecture presented in this paper enables {\em exact and efficient} reconstruction of input images from the hierarchical features extracted by this model.
\section{Related work}
\begin{figure}
\begin{center}
{\renewcommand{\arraystretch}{2}
\begin{tabular}{m{1.25in}ccc}
& Data space ${\mathcal{X}}$ & & Latent space ${\mathcal{Z}}$
\\
\parbox{1.25in}{
\textbf{Inference}
$\begin{aligned}
\qquad x &\sim \hat{p}_{X} &\\
\qquad z &= f\left(x\right)
\end{aligned}$
}
&
\raisebox{-.43\height}{
\includegraphics[width=1.5in]{example_x.pdf}
}
&
\parbox{.25in}{
$\begin{aligned}
\Rightarrow
\end{aligned}$
}
&
\raisebox{-.43\height}{
\includegraphics[width=1.5in]{example_z.pdf}
}
\\[0.5in]
\parbox{1.25in}{
\textbf{Generation}
$\begin{aligned}
\qquad z &\sim p_{Z} &
\\
\qquad x &= f^{-1}\left(z\right)
\end{aligned}$
}
&
\raisebox{-.43\height}{
\includegraphics[width=1.5in]{sample_x.pdf}
}
&
\parbox{.25in}{
$\begin{aligned}
\Leftarrow
\end{aligned}$
}
&
\raisebox{-.43\height}{
\includegraphics[width=1.5in]{sample_z.pdf}
}
\end{tabular}}
\end{center}
\caption{Real NVP learns an invertible, stable, mapping between a data distribution $\hat{p}_{X}$ and a latent distribution $p_{Z}$ (typically a Gaussian).
Here we show a mapping that has been learned on a toy 2-d dataset.
The function $f\left(x\right)$ maps samples $x$ from the data distribution in the upper left into approximate samples $z$ from the latent distribution, in the upper right. This corresponds to exact inference of the latent state given the data.
The inverse function, $f^{-1}\left(z\right)$, maps samples $z$ from the latent distribution in the lower right into approximate samples $x$ from the data distribution in the lower left. This corresponds to exact generation of samples from the model.
The transformation of grid lines in ${\mathcal{X}}$ and ${\mathcal{Z}}$ space is additionally illustrated for both $f\left(x\right)$ and $f^{-1}\left(z\right)$.
}
\label{fig:spaghetti}
\end{figure}
Substantial work on probabilistic generative models has focused on training models using maximum likelihood.
One class of maximum likelihood models are those described by \emph{probabilistic undirected graphs}, such as \emph{Restricted Boltzmann Machines} \citep{smolensky1986information} and \emph{Deep Boltzmann Machines} \citep{salakhutdinov2009deep}. These models are trained by taking advantage of the conditional independence property of their bipartite structure to allow efficient exact or approximate posterior inference on latent variables. However, because of the intractability of the associated marginal distribution over latent variables, their training, evaluation, and sampling procedures necessitate the use of approximations like \emph{Mean Field inference} and \emph{Markov Chain Monte Carlo}, whose convergence time for such complex models remains undetermined, often resulting in generation of highly correlated samples. Furthermore, these approximations can often hinder their performance \citep{berglund2013stochastic}.
\emph{Directed graphical models} are instead defined in terms of an \emph{ancestral sampling procedure}, which is appealing both for its conceptual and computational simplicity. They lack, however, the conditional independence structure of undirected models, making exact and approximate posterior inference on latent variables cumbersome \citep{saul1996mean}. Recent advances in \emph{stochastic variational inference} \citep{hoffman2013stochastic} and \emph{amortized inference} \citep{dayan1995helmholtz, mnih2014neural, kingma2013auto, rezende2014stochastic}, allowed efficient approximate inference and learning of deep directed graphical models by maximizing a variational lower bound on the log-likelihood \citep{neal1998view}. In particular, the \emph{variational autoencoder algorithm} \citep{kingma2013auto, rezende2014stochastic} simultaneously learns a \emph{generative network}, that maps gaussian latent variables $z$ to samples $x$, and a matched \emph{approximate inference network} that maps samples $x$ to a semantically meaningful latent representation $z$, by exploiting the \emph{reparametrization trick} \citep{williams1992simple}.
Its success in leveraging recent advances in backpropagation \citep{rumelhart1988learning, lecun2012efficient} in deep neural networks resulted in its adoption for several applications ranging from speech synthesis \citep{chung2015recurrent} to language modeling \citep{bowman2015generating}. Still, the approximation in the inference process limits its ability to learn high dimensional deep representations, motivating recent work in improving approximate inference \citep{maaloe2016auxiliary, rezende2015variational, salimans2014markov, tran2015variational, burda2015importance, DBLP:conf/icml/Sohl-DicksteinW15, kingma2016improving}.
Such approximations can be avoided altogether by abstaining from using latent variables. \emph{Auto-regressive models} \citep{frey1998graphical, bengio1999modeling, larochelle2011neural, DBLP:journals/corr/GermainGML15} can implement this strategy while typically retaining a great deal of flexibility. This class of algorithms tractably models the joint distribution by decomposing it into a product of conditionals using the \emph{probability chain rule} according to a fixed ordering over dimensions, simplifying log-likelihood evaluation and sampling. Recent work in this line of research has taken advantage of recent advances in \emph{recurrent networks} \citep{rumelhart1988learning}, in particular \emph{long-short term memory} \citep{DBLP:journals/neco/HochreiterS97}, and \emph{residual networks} \citep{DBLP:journals/corr/HeZR016, DBLP:journals/corr/HeZRS15} in order to learn state-of-the-art generative image models \citep{theis2015generative, oord2016pixel} and language models \citep{DBLP:journals/corr/JozefowiczVSSW16}.
The ordering of the dimensions, although often arbitrary, can be critical to the training of the model \citep{vinyals2015order}.
The sequential nature of this model limits its computational efficiency. For example, its sampling procedure is sequential and non-parallelizable, which can become cumbersome in applications like speech and music synthesis, or real-time rendering..
Additionally, there is no natural latent representation associated with autoregressive models, and they have not yet been shown to be useful for semi-supervised learning.
\emph{Generative Adversarial Networks} (GANs) \citep{DBLP:conf/nips/GoodfellowPMXWOCB14} on the other hand can train any differentiable generative network by avoiding the maximum likelihood principle altogether. Instead, the generative network is associated with a \emph{discriminator network} whose task is to distinguish between samples and real data. Rather than using an intractable log-likelihood, this discriminator network provides the training signal in an adversarial fashion. Successfully trained GAN models \citep{DBLP:conf/nips/GoodfellowPMXWOCB14, DBLP:conf/nips/DentonCSF15, DBLP:journals/corr/RadfordMC15} can consistently generate sharp and realistically looking samples \citep{DBLP:journals/corr/LarsenSW15}. However, metrics that measure the diversity in the generated samples are currently intractable \citep{DBLP:journals/corr/TheisOB15, gregor2016towards, im2016generating}. Additionally, instability in their training process \citep{DBLP:journals/corr/RadfordMC15} requires careful hyperparameter tuning to avoid diverging behavior.
Training such a generative network $g$ that maps latent variable $z \sim p_{Z}$ to a sample $x \sim p_{X}$ does not in theory require a discriminator network as in GANs, or approximate inference as in variational autoencoders. Indeed, if $g$ is bijective, it can be trained through maximum likelihood using the \emph{change of variable formula}:
\begin{align}
p_{X}(x) = p_{Z}(z) \left\vert \det\left(\frac{\partial g(z)}{\partial z^T}\right)\right\vert^{-1}
.
\end{align}
This formula has been discussed in several papers including the maximum likelihood formulation of \emph{independent components analysis} (ICA) \citep{bell1995information, hyvarinen2004independent}, gaussianization \citep{NIPS1994_901, chen2000gaussianization} and deep density models \citep{bengio1991artificial, rippel2013high, dinh2014nice, balle2015density}.
As the existence proof of nonlinear ICA solutions \citep{hyvarinen1999nonlinear} suggests, auto-regressive models can be seen as tractable instance of maximum likelihood nonlinear ICA, where the residual corresponds to the independent components.
However, naive application of the change of variable formula produces models which are computationally expensive and poorly conditioned, and so large scale models of this type have not entered general use.
\section{Model definition}
In this paper, we will tackle the problem of learning highly nonlinear models in high-dimensional continuous spaces through maximum likelihood. In order to optimize the log-likelihood, we introduce a more flexible class of architectures that enables the computation of log-likelihood on continuous data using the change of variable formula. Building on our previous work in \citep{dinh2014nice}, we define a powerful class of bijective functions which enable exact and tractable density evaluation and exact and tractable inference.
Moreover, the resulting cost function does not to rely on a fixed form reconstruction cost such as square error \citep{DBLP:journals/corr/LarsenSW15, DBLP:journals/corr/RadfordMC15}, and generates sharper samples as a result. Also, this flexibility helps us leverage recent advances in batch normalization \citep{ioffe2015batch} and residual networks \citep{DBLP:journals/corr/HeZRS15, DBLP:journals/corr/HeZR016} to define a very deep multi-scale architecture with multiple levels of abstraction.
\subsection{Change of variable formula}
Given an observed data variable $x \in X$,
a simple prior probability distribution $p_{Z}$ on a latent variable $z \in Z$,
and a bijection $f: X \rightarrow Z$ (with $g = f^{-1}$),
the change of variable formula defines a model distribution on $X$ by
\begin{align}
p_{X}(x) &= p_{Z}\big(f(x)\big) \left|\det\left(\cfrac{\partial f(x)}{\partial x^T} \right)\right|
\label{eq:change-variables}\\
\log\left(p_{X}(x)\right) &= \log\Big(p_{Z}\big(f(x)\big)\Big) + \log\left(\left|\det\left(\frac{\partial f(x)}{\partial x^T}\right)\right|\right)
,
\end{align}
where $\frac{\partial f(x)}{\partial x^T}$ is the Jacobian of $f$ at $x$.
Exact samples from the resulting distribution can be generated by using the inverse transform sampling rule \citep{devroye1986sample}. A sample $z \sim p_{Z}$ is drawn in the latent space, and its inverse image $x = f^{-1}(z) = g(z)$ generates a sample in the original space. Computing the density on a point $x$ is accomplished by computing the density of its image $f(x)$ and multiplying by the associated Jacobian determinant $\det\left(\frac{\partial f(x)}{\partial x^T}\right)$. See also Figure \ref{fig:spaghetti}. Exact and efficient inference enables the accurate and fast evaluation of the model.
\subsection{Coupling layers}
\begin{figure}
\centering
\subfigure[Forward propagation]{
\includegraphics[width=.2\textwidth]{coupling.pdf}
\qquad
\label{fig:coupling}} ~~~~~~
\subfigure[Inverse propagatio
]{
\qquad
\includegraphics[width=.2\textwidth]{inverse_coupling.pdf}
\label{fig:inverse_coupling}}
\caption{Computational graphs for forward and inverse propagation. A coupling layer applies a simple invertible transformation consisting of scaling followed by addition of a constant offset to one part $\mathbf x_{2}$ of the input vector conditioned on the remaining part of the input vector $\mathbf x_{1}$. Because of its simple nature, this transformation is both easily invertible and possesses a tractable determinant.
However, the conditional nature of this transformation, captured by the functions $s$ and $t$, significantly increase the flexibility of this otherwise weak function.
The forward and inverse propagation operations have identical computational cost. }
\end{figure}
Computing the Jacobian of functions with high-dimensional domain and codomain and computing the determinants of large matrices are in general computationally very expensive. This combined with the restriction to bijective functions makes Equation \ref{eq:change-variables} appear impractical for modeling arbitrary distributions.
As shown however in \citep{dinh2014nice}, by careful design of the function $f$, a bijective model can be learned which is both tractable and extremely flexible.
As computing the Jacobian determinant of the transformation is crucial to effectively train using this principle, this work exploits the simple observation that \emph{the determinant of a triangular matrix can be efficiently computed} as the product of its diagonal terms.
We will build a flexible and tractable bijective function by stacking a sequence of simple bijections.
In each simple bijection,
part of the input vector is updated using a function which is simple to invert,
but which depends on the remainder of the input vector in a complex way.
We refer to each of these simple bijections as an \emph{affine coupling layer}.
Given a $D$ dimensional input $x$ and $d < D$, the output $y$ of an affine coupling layer follows the equations
\begin{align}
y_{1:d} &= x_{1:d}\\
y_{d+1:D} &= x_{d+1:D} \odot \exp\big(s(x_{1:d})\big) + t(x_{1:d})
,
\end{align}
where $s$ and $t$ stand for scale and translation, and are functions from $R^{d} \mapsto R^{D - d}$, and $\odot$ is the Hadamard product or element-wise product (see Figure \ref{fig:coupling}).
\subsection{Properties}
The Jacobian of this transformation is
\begin{align}
\frac{\partial y}{\partial x^T} = \left[\begin{array}{cc}
{\mathbb{I}}_{d} & 0 \\
\frac{\partial y_{d+1:D}}{\partial x_{1:d}^T} & \diag\big(\exp\left[s\left(x_{1:d}\right)\right]\big)
\end{array} \right]
,
\end{align}
where $\diag\big(\exp\left[s\left(x_{1:d}\right)\right]\big)$ is the diagonal matrix whose diagonal elements correspond to the vector $\exp\left[s\left(x_{1:d}\right)\right]$. Given the observation that this Jacobian is triangular, we can efficiently compute its determinant as $\exp\left[\sum_{j}{s\left(x_{1:d}\right)_j}\right]$. Since computing the Jacobian determinant of the coupling layer operation does not involve computing the Jacobian of $s$ or $t$, those functions can be arbitrarily complex. We will make them deep convolutional neural networks. Note that the hidden layers of $s$ and $t$ can have more features than their input and output layers.
Another interesting property of these coupling layers in the context of defining probabilistic models is their invertibility. Indeed, computing the inverse is no more complex than the forward propagation (see Figure \ref{fig:inverse_coupling}),
\begin{align}
\begin{cases}
y_{1:d} &= x_{1:d} \\
y_{d+1:D} &= x_{d+1:D} \odot \exp\big(s(x_{1:d})\big) + t(x_{1:d})
\end{cases}\\
\Leftrightarrow
\begin{cases}
x_{1:d} &= y_{1:d} \\
x_{d+1:D} &= \big(y_{d+1:D} - t(y_{1:d})\big) \odot \exp\big(-s(y_{1:d})\big),
\end{cases}
\end{align}
meaning that sampling is as efficient as inference for this model.
Note again that computing the inverse of the coupling layer does not require computing the inverse of $s$ or $t$, so these functions can be arbitrarily complex and difficult to invert.
\subsection{Masked convolution}
Partitioning can be implemented using a binary mask $b$, and using the functional form for $y$,
\begin{align}
y = b \odot x + (1 - b) \odot \Big(x \odot \exp\big(s(b \odot x)\big) + t(b \odot x)\Big)
.
\end{align}
We use two partitionings that exploit the local correlation structure of images: spatial checkerboard patterns, and channel-wise masking (see Figure \ref{fig:squeezing}). The spatial checkerboard pattern mask has value $1$ where the sum of spatial coordinates is odd, and $0$ otherwise. The channel-wise mask $b$ is $1$ for the first half of the channel dimensions and $0$ for the second half.
For the models presented here, both $s(\cdot)$ and $t(\cdot)$ are rectified convolutional networks.
\begin{figure}
\centering \includegraphics[width=.6\textwidth]{masks.pdf}
\vspace{-10pt}
\caption{Masking schemes for affine coupling layers. On the left, a spatial checkerboard pattern mask. On the right, a channel-wise masking. The squeezing operation reduces the $4 \times 4 \times 1$ tensor (on the left) into a $2 \times 2 \times 4$ tensor (on the right). Before the squeezing operation, a checkerboard pattern is used for coupling layers while a channel-wise masking pattern is used afterward.
}
\label{fig:squeezing}
\end{figure}
\subsection{Combining coupling layers}
Although coupling layers can be powerful, their forward transformation leaves some components unchanged.
This difficulty can be overcome by composing coupling layers in an alternating pattern, such that the components that are left unchanged in one coupling layer are updated in the next (see Figure \ref{fig:composition}).
The Jacobian determinant of the resulting function remains tractable, relying on the fact that
\begin{align}
\cfrac{\partial (f_{b} \circ f_{a})}{\partial x_{a}^{T}}(x_{a}) &= \cfrac{\partial f_{a}}{\partial x_{a}^{T}}(x_{a}) \cdot \cfrac{\partial f_{b}}{\partial x_{b}^{T}}\big(x_{b}=f_{a}(x_{a})\big) \\
\det(A \cdot B) &= \det(A) \det(B).
\end{align}
Similarly, its inverse can be computed easily as
\begin{align}
(f_{b} \circ f_{a})^{-1} = f_{a}^{-1} \circ f_{b}^{-1}.
\end{align}
\begin{figure}[h]
\centering \subfigure[In this alternating pattern, units which remain identical in one transformation are modified in the next.]{
\includegraphics[width=.6\textwidth]{composition.pdf}
\label{fig:composition}} ~~~~~~
\subfigure[Factoring out variables. At each step, half the variables are directly modeled as Gaussians, while the other half undergo further transformation.]{
\includegraphics[width=.25\textwidth]{factor_out.pdf}
\label{fig:factor_out}}
\caption{Composition schemes for affine coupling layers.}
\end{figure}
\subsection{Multi-scale architecture}
\label{section:multiscale}
We implement a multi-scale architecture using a squeezing operation: for each channel, it divides the image into subsquares of shape $2 \times 2 \times c$, then reshapes them into subsquares of shape $1 \times 1 \times 4c$. The squeezing operation transforms an $s \times s \times c$ tensor into an $\frac{s}{2} \times \frac{s}{2} \times 4c$ tensor (see Figure \ref{fig:squeezing}), effectively trading spatial size for number of channels.
At each scale, we combine several operations into a sequence: we first apply three coupling layers with alternating checkerboard masks, then perform a squeezing operation, and finally apply three more coupling layers with alternating channel-wise masking. The channel-wise masking is chosen so that the resulting partitioning is not redundant with the previous checkerboard masking (see Figure \ref{fig:squeezing}). For the final scale, we only apply four coupling layers with alternating checkerboard masks.
Propagating a $D$ dimensional vector through all the coupling layers would be cumbersome, in terms of computational and memory cost, and in terms of the number of parameters that would need to be trained.
For this reason we follow the design choice of \citep{simonyan2014very} and factor out half of the dimensions at regular intervals (see Equation \ref{eq:factor_out}).
We can define this operation recursively (see Figure \ref{fig:factor_out}),
\begin{align}
h^{(0)} &= x \\
(z^{(i + 1)}, h^{(i + 1)}) &= f^{(i + 1)}(h^{(i)})
\label{eq:factor_out} \\
z^{(L)} &= f^{(L)}(h^{(L - 1)})
\label{eq:final_factor}\\
z &= (z^{(1)}, \dots, z^{(L)})
\label{eq:concat_factors}
.
\end{align}
In our experiments, we use this operation for $i < L$.
The sequence of coupling-squeezing-coupling operations described above is performed per layer when computing $f^{(i)}$ (Equation \ref{eq:factor_out}).
At each layer, as the spatial resolution is reduced, the number of hidden layer features in $s$ and $t$ is doubled.
All variables which have been factored out at different scales are concatenated to obtain the final transformed output (Equation \ref{eq:concat_factors}).
As a consequence, the model must Gaussianize units which are factored out at a finer scale (in an earlier layer) before those which are factored out at a coarser scale (in a later layer). This results in the definition of intermediary levels of representation \citep{salakhutdinov2009deep, rezende2014stochastic} corresponding to more local, fine-grained features as shown in Appendix \ref{app:latent}.
Moreover, Gaussianizing and factoring out units in earlier layers has the practical benefit of distributing the loss function throughout the network, following the philosophy similar to guiding intermediate layers using intermediate classifiers \citep{lee2014deeply}. It also reduces significantly the amount of computation and memory used by the model, allowing us to train larger models.
\subsection{Batch normalization}
To further improve the propagation of training signal, we use deep residual networks \citep{DBLP:journals/corr/HeZRS15, DBLP:journals/corr/HeZR016} with batch normalization \citep{ioffe2015batch} and weight normalization \citep{badrinarayanan2015understanding, salimans2016weight} in $s$ and $t$.
As described in Appendix \ref{sec batch norm} we introduce and use a novel variant of batch normalization which is based on a running average over recent minibatches, and is thus more robust when training with very small minibatches.
We also use apply batch normalization to the whole coupling layer output.
The effects of batch normalization are easily included in the Jacobian computation, since it acts as a linear rescaling on each dimension. That is, given the estimated batch statistics $\tilde{\mu}$ and $\tilde{\sigma}^{2}$, the rescaling function
\begin{align}
x \mapsto \frac{x - \tilde{\mu}}{\sqrt{\tilde{\sigma}^{2} + \epsilon}}
\end{align}
has a Jacobian determinant
\begin{align}
\left(\prod_{i}{(\tilde{\sigma}_{i}^{2} + \epsilon)}\right)^{-\frac{1}{2}}.
\end{align}
This form of batch normalization can be seen as similar to reward normalization in deep reinforcement learning \citep{mnih2015human, van2016learning}.
We found that the use of this technique not only allowed training with a deeper stack of coupling layers, but also alleviated the instability problem that practitioners often encounter when training conditional distributions with a scale parameter through a gradient-based approach.
\section{Experiments}
\subsection{Procedure}
The algorithm described in Equation \ref{eq:change-variables} shows how to learn distributions on unbounded space. In general, the data of interest have bounded magnitude. For examples, the pixel values of an image typically lie in $[0, 256]^{D}$ after application of the recommended jittering procedure \citep{uria2013rnade, DBLP:journals/corr/TheisOB15}. In order to reduce the impact of boundary effects, we instead model the density of $\mbox{logit}(\alpha + (1 - \alpha) \odot \frac{x}{256})$, where $\alpha$ is picked here as $.05$.
We take into account this transformation when computing log-likelihood and bits per dimension. We also augment the CIFAR-10, CelebA and LSUN datasets during training to also include horizontal flips of the training examples.
We train our model on four natural image datasets: \emph{CIFAR-10} \citep{krizhevsky2009learning}, \emph{Imagenet} \citep{russakovsky2015imagenet}, \emph{Large-scale Scene Understanding (LSUN)} \citep{yu2015construction}, \emph{CelebFaces Attributes (CelebA)} \citep{liu2015faceattributes}. More specifically, we train on the downsampled to $32 \times 32$ and $64 \times64$ versions of Imagenet \citep{oord2016pixel}. For the LSUN dataset, we train on the \emph{bedroom}, \emph{tower} and \emph{church} outdoor categories. The procedure for LSUN is the same as in \citep{DBLP:journals/corr/RadfordMC15}: we downsample the image so that the smallest side is $96$ pixels and take random crops of $64 \times64$. For CelebA, we use the same procedure as in \citep{DBLP:journals/corr/LarsenSW15}: we take an approximately central crop of $148 \times 148$ then resize it to $64 \times 64$.
We use the multi-scale architecture described in Section \ref{section:multiscale} and use deep convolutional residual networks in the coupling layers with rectifier nonlinearity and skip-connections as suggested by \citep{oord2016pixel}. To compute the scaling functions $s$, we use a {\em hyperbolic tangent} function multiplied by a learned scale, whereas the translation function $t$ has an affine output. Our multi-scale architecture is repeated recursively until the input of the last recursion is a $4 \times 4 \times c$ tensor. For datasets of images of size $32 \times 32$, we use $4$ residual blocks with $32$ hidden feature maps for the first coupling layers with checkerboard masking. Only $2$ residual blocks are used for images of size $64 \times 64$. We use a batch size of $64$. For CIFAR-10, we use $8$ residual blocks, $64$ feature maps, and downscale only once. We optimize with ADAM \citep{kingma2014adam} with default hyperparameters and use an $L_{2}$ regularization on the weight scale parameters with coefficient $5 \cdot 10^{-5}$.
We set the prior $p_{Z}$ to be an isotropic unit norm Gaussian. However, any distribution could be used for $p_{Z}$, including distributions that are also learned during training, such as from an auto-regressive model, or (with slight modifications to the training objective) a variational autoencoder.
\subsection{Results}
\begin{table}
\begin{center}
\begin{tabular}{| c | c | c | c | c |}
\hline
\textbf{Dataset} & \textbf{PixelRNN} \citep{oord2016pixel} & \textbf{Real NVP} & \textbf{Conv DRAW \citep{gregor2016towards}} & \textbf{IAF-VAE \citep{kingma2016improving}} \\ \hline
\textbf{CIFAR-10} & 3.00 & 3.49 & < 3.59 & < 3.28 \\ \hline
\textbf{Imagenet ($32 \times 32$)} & 3.86 (3.83) & 4.28 (4.26) & < 4.40 (4.35) & \cellcolor{black!25} \\ \hline
\textbf{Imagenet ($64 \times 64$)} & 3.63 (3.57) & 3.98 (3.75) & < 4.10 (4.04) & \cellcolor{black!25} \\ \hline
\textbf{LSUN (bedroom)} & \cellcolor{black!25} & 2.72 (2.70) & \cellcolor{black!25} & \cellcolor{black!25} \\ \hline
\textbf{LSUN (tower)} & \cellcolor{black!25} & 2.81 (2.78) & \cellcolor{black!25} & \cellcolor{black!25} \\ \hline
\textbf{LSUN (church outdoor)} & \cellcolor{black!25} & 3.08 (2.94) & \cellcolor{black!25} & \cellcolor{black!25} \\ \hline
\textbf{CelebA} & \cellcolor{black!25} & 3.02 (2.97) & \cellcolor{black!25} & \cellcolor{black!25} \\ \hline
\end{tabular}
\caption{Bits/dim results for CIFAR-10, Imagenet, LSUN datasets and CelebA. Test results for CIFAR-10 and validation results for Imagenet, LSUN and CelebA (with training results in parenthesis for reference).
\label{fig:quant}}
\end{center}
\end{table}
We show in Table \ref{fig:quant} that the number of bits per dimension, while not improving over the Pixel RNN \citep{oord2016pixel} baseline, is competitive with other generative methods. As we notice that our performance increases with the number of parameters, larger models are likely to further improve performance.
For CelebA and LSUN, the bits per dimension for the validation set was decreasing throughout training, so little overfitting is expected.
\begin{figure}
\begin{center}
\includegraphics[width=.47\textwidth]{fig_cifar10_examples.jpg} \hfill
\includegraphics[width=.47\textwidth]{fig_cifar10_samples.jpg} \\\vspace{1mm}
\includegraphics[width=.47\textwidth]{fig_imnet_32_examples.jpg} \hfill
\includegraphics[width=.47\textwidth]{fig_imnet_32_samples.jpg} \\\vspace{1mm}
\includegraphics[width=.47\textwidth]{fig_imnet_64_examples.jpg} \hfill
\includegraphics[width=.47\textwidth]{fig_imnet_64_samples.jpg} \\\vspace{1mm}
\includegraphics[width=.47\textwidth]{fig_celeba_examples.jpg} \hfill
\includegraphics[width=.47\textwidth]{fig_celeba_samples.jpg} \\\vspace{1mm}
\includegraphics[width=.47\textwidth]{fig_bedroom_examples.jpg} \hfill
\includegraphics[width=.47\textwidth]{fig_bedroom_samples.jpg}
\caption{On the left column, examples from the dataset. On the right column, samples from the model trained on the dataset. The datasets shown in this figure are in order: CIFAR-10, Imagenet ($32 \times 32$), Imagenet ($64 \times 64$), CelebA, LSUN (bedroom)
}
\label{fig:samples}
\end{center}
\end{figure}
We show in Figure \ref{fig:samples} samples generated from the model with training examples from the dataset for comparison. As mentioned in \citep{DBLP:journals/corr/TheisOB15, gregor2016towards}, maximum likelihood is a principle that values diversity over sample quality in a limited capacity setting. As a result, our model outputs sometimes highly improbable samples as we can notice especially on CelebA. As opposed to variational autoencoders, the samples generated from our model look not only globally coherent but also sharp. Our hypothesis is that as opposed to these models, real NVP does not rely on fixed form reconstruction cost like an $L_{2}$ norm which tends to reward capturing low frequency components more heavily than high frequency components. Unlike autoregressive models, sampling from our model is done very efficiently as it is parallelized over input dimensions.
On Imagenet and LSUN, our model seems to have captured well the notion of background/foreground and lighting interactions such as luminosity and consistent light source direction for reflectance and shadows.
\begin{figure}
\begin{center}
\includegraphics[width=.47\textwidth]{fig_celeba_manifold.jpg} \hfill
\includegraphics[width=.47\textwidth]{fig_imnet_64_manifold.jpg} \\\vspace{1mm}
\includegraphics[width=.47\textwidth]{fig_bedroom_manifold.jpg} \hfill
\includegraphics[width=.47\textwidth]{fig_tower_manifold.jpg}
\end{center}
\caption{Manifold generated from four examples in the dataset. Clockwise from top left: CelebA, Imagenet ($64 \times 64$), LSUN (tower), LSUN (bedroom).}
\label{fig:manifold}
\end{figure}
We also illustrate the smooth semantically consistent meaning of our latent variables.
In the latent space, we define a manifold based on four validation examples $z_{(1)}$, $z_{(2)}$, $z_{(3)}$, $z_{(4)}$, and parametrized by two parameters $\phi$ and $\phi'$ by,
\begin{align}
z = \cos(\phi) \left(\cos(\phi') z_{(1)} + \sin(\phi') z_{(2)}\right) + \sin(\phi) \left(\cos(\phi') z_{(3)} + \sin(\phi') z_{(4)}\right)
.
\label{eq:manifold}
\end{align}
We project the resulting manifold back into the data space by computing $g(z)$. Results are shown Figure \ref{fig:manifold}. We observe that the model seems to have organized the latent space with a notion of meaning that goes well beyond pixel space interpolation. More visualization are shown in the Appendix. To further test whether the latent space has a consistent semantic interpretation, we trained a class-conditional model on CelebA, and found that the learned representation had a consistent semantic meaning across class labels (see Appendix \ref{app attr change}).
\section{Discussion and conclusion}
In this paper, we have defined a class of invertible functions with tractable Jacobian determinant, enabling exact and tractable log-likelihood evaluation, inference, and sampling.
We have shown that this class of generative model achieves competitive performances, both in terms of sample quality and log-likelihood.
Many avenues exist to further improve the functional form of the transformations, for instance by exploiting the latest advances in dilated convolutions \citep{yu2015multi} and residual networks architectures \citep{DBLP:journals/corr/TargAL16}.
This paper presented a technique bridging the gap between auto-regressive models, variational autoencoders, and generative adversarial networks. Like auto-regressive models, it allows tractable and exact log-likelihood evaluation for training. It allows however a much more flexible functional form, similar to that in the generative model of variational autoencoders. This allows for fast and exact sampling from the model distribution.
Like GANs, and unlike variational autoencoders, our technique does not require the use of a fixed form reconstruction cost, and instead defines a cost in terms of higher level features, generating sharper images.
Finally, unlike both variational autoencoders and GANs, our technique is able to learn a semantically meaningful latent space which is as high dimensional as the input space. This may make the algorithm particularly well suited to semi-supervised learning tasks, as we hope to explore in future work.
Real NVP generative models can additionally be conditioned on additional variables (for instance class labels) to create a structured output algorithm. More so, as the resulting class of invertible transformations can be treated as a probability distribution in a modular way, it can also be used to improve upon other probabilistic models like auto-regressive models and variational autoencoders. For variational autoencoders, these transformations could be used both to enable a more flexible reconstruction cost \citep{DBLP:journals/corr/LarsenSW15} and a more flexible stochastic inference distribution \citep{ rezende2015variational}. Probabilistic models in general can also benefit from batch normalization techniques as applied in this paper.
The definition of powerful and trainable invertible functions can also benefit domains other than generative unsupervised learning. For example, in reinforcement learning, these invertible functions can help extend the set of functions for which an $\argmax$ operation is tractable for {\em continuous Q-learning} \citep{gu2016continuous} or find representation where local linear Gaussian approximations are more appropriate \citep{watter2015embed}.
\section{Acknowledgments}
The authors thank the developers of Tensorflow \citep{abadi2016tensorflow}.
We thank Sherry Moore, David Andersen and Jon Shlens for their help in implementing the model.
We thank Aäron van den Oord, Yann Dauphin, Kyle Kastner, Chelsea Finn, Maithra Raghu, David Warde-Farley, Daniel Jiwoong Im and Oriol Vinyals for fruitful discussions.
Finally, we thank Ben Poole, Rafal Jozefowicz and George Dahl for their input on a draft of the paper.
\small
\setlength{\bibsep}{0pt plus 0.12ex}
\bibliographystyle{plain}
|
\section{Introduction}
The study of quantum integrable systems in the context of Quantum Inverse Scattering Method (QISM) has been one of the great successes of the Leningrad school \cite{FadST79,FadT79,FadLH96}. As one of the early members of this school, Petr Kulish has made several important contributions to the field. Among them, it is worth noticing the study of quantum integrable models solvable by the nested algebraic Bethe ansatz \cite{Kul81,KulRes81,KulRes83}, the search for solutions to the Yang-Baxter equation \cite{KulS80,KulRes82}, or the development of QISM for models with open boundaries \cite{KulSas93,Damk03,AvaKR11}. He also contributed to the generalisations to the supersymmetric versions of these models \cite{KulS80,Kul85}.
In this article, we would like to make a contribution to the field he was found of, that is supersymmetric quantum integrable models solvable by the nested algebraic Bethe ansatz. This paper is the continuation of a series of articles devoted to quantum integrable models
possessing a $\mathfrak{gl}(2|1)$ symmetry.
In \cite{PakRS16a} we have built explicit representations for the Bethe vectors in these models.
In \cite{HutLPRS16a} we found multiple actions of the monodromy matrix entries onto these vectors. Now we consider the problem of scalar products
of Bethe vectors.
Scalar products of Bethe vectors play a very important role in the Algebraic Bethe ansatz. They are a necessary tool for
calculating form factors and correlation functions within this framework. The first results in this field concern
$\mathfrak{gl}(2)$-based models and their $q$-deformations. They were obtained in \cite{Kor82,IzeK84,Ize87}, where in particular the Izergin--Korepin formula (for scalar products) was given. Concerning scalar
products in models with $\mathfrak{gl}(3)$-invariant
$R$-matrix, the first result was obtained in \cite{Res86}. There, an analog of the Izergin--Korepin
formula for scalar product of generic Bethe vectors was produced (the Reshetikhin formula) and a determinant representation
for the norm of the eigenvectors of the transfer matrix was found. Similar results for the models based on a $q$-deformed $\mathfrak{gl}(3)$ algebra
were obtained in \cite{PakRS14a,Sla15a}.
In this paper we consider scalar products in the models with $\mathfrak{gl}(2|1)$ symmetry. At this stage of study our goal is to obtain an analog
of Reshetikhin formula for these models. This means that we are going to find an explicit representation for the scalar product
of generic Bethe vectors, in which the result is given as a sum over partitions of the Bethe parameters (sum formula). Certainly, this type
of formulas is not convenient for direct applications. However, they give a key for studying particular cases of scalar product, in which the sum over partitions can be reduced to a single determinant. Such determinant representations
for particular cases of scalar products were already found for the models with $\mathfrak{gl}(2)$ and $\mathfrak{gl}(3)$ symmetries
\cite{Sla89,KitMT99,BelPRS12b,Sla15a}. In all these cases the starting point was a sum formula.
The article is organized as follows. In section~\ref{S-N} we introduce the model under consideration and
specify our conventions and notation. Section~\ref{S-NT} contains explicit formulas for the Bethe vectors and multiple
actions of the monodromy matrix entries onto them.
In section~\ref{S-GFSP} we present the main result of the paper. There we define scalar products
and give a sum formula for them. The sum formula is proved in the remaining part of the paper.
In section~\ref{S-SA} we study successive actions of the monodromy matrix entries onto Bethe vectors.
In section~\ref{S-HC} we obtain an explicit determinant representation for the highest coefficient of the scalar product.
Section~\ref{S-SP11} deals with a particular case of the scalar products in the $\mathfrak{gl}(1|1)$-based models.
Section~\ref{S-DRHC} contains several alternative representations for the highest coefficient.
We have collected auxiliary formulas in four appendices. In particular, appendix~\ref{A-IHC} is devoted to the properties of
the partition function of six-vertex model with domain wall boundary conditions. In appendix~\ref{A-SumInt} we describe a relationship of sums over
partitions and multiple contour integrals. Appendix~\ref{A-SF} contains several identities allowing us to calculate certain multiple sums of rational functions.
Finally, in appendix~\ref{A-RPJ} we present reduction properties of a determinant introduced in section~\ref{S-HC}.
\section{Description of the model\label{S-N}}
\subsection{$\mathfrak{gl}(2|1)$-based models}
The $R$-matrix of $\mathfrak{gl}(2|1)$-based models acts in the tensor product $\mathbb{C}^{2|1}\otimes \mathbb{C}^{2|1}$,
where $\mathbb{C}^{2|1}$ is the $\mathbb{Z}_2$-graded vector space with the grading $[1]=[2]=0$, $[3]=1$.
The $R$-matrix has the form
\be{R-mat}
R(u,v)=\mathbb{I}+g(u,v)P, \qquad g(u,v)=\frac{c}{u-v},
\end{equation}
where $\mathbb{I}$ is the identity matrix, $P$ is the graded permutation operator \cite{KulS80} and $c$ is a constant. The
monodromy matrix $T(u)$ is also graded according to the rule $[T_{ij}(u)]=[i]+[j]$. It satisfies the $RTT$-relation
\be{RTT}
R(u,v)\bigl(T(u)\otimes \mathbb{I}\bigr) \bigl(\mathbb{I}\otimes T(v)\bigr)= \bigl(\mathbb{I}\otimes T(v)\bigr)
\bigl( T(u)\otimes \mathbb{I}\bigr)R(u,v).
\end{equation}
Equation \eqref{RTT} holds in the tensor product $\mathbb{C}^{2|1}\otimes \mathbb{C}^{2|1}\otimes\mathcal{H}$,
where $\mathcal{H}$ is a Hilbert space of a Hamiltonian under consideration. Tensor products
of $\mathbb{C}^{2|1}$ spaces are graded as follows:
\be{tens-prod}
(\mathbb{I}\otimes e_{ij})\,\cdot\,(e_{kl}\otimes \mathbb{I}) = (-1)^{([i]+[j])([k]+[l])}\,e_{kl}\otimes e_{ij},
\end{equation}
where $e_{ij}$ are the elementary units: $(e_{ij})_{ab}=\delta_{ia}\delta_{jb}$.
Algebra \eqref{RTT} possesses an antimorphism
\cite{PakRS16a}
\be{psi}
\psi\bigl(T_{ij}(u)\bigr)=(-1)^{[i][j]+[i]}T_{ji}(u),\qquad
\psi\bigl(AB\bigr)=(-1)^{[A][B]}\psi\bigl(B\bigr)\psi\bigl(A\bigr),
\end{equation}
where $A$ and $B$ are arbitrary operators of fixed gradings. It follows from \eqref{psi} that
\be{psiAn}
\psi\bigl(A_1\dots A_n\bigr)=(-1)^{\vartheta_n}\psi\bigl(A_n\bigr)\dots\psi\bigl(A_1\bigr), \qquad
\vartheta_n=\sum_{1\le i<j\le n} [A_i]\cdot [A_j].
\end{equation}
The $RTT$-relation \eqref{RTT} implies a set of scalar commutation relations for the monodromy matrix elements.
For our purpose, we will need only the following ones:
\be{odd-comm}
\begin{aligned}
&[T_{ik}(v_1),T_{ik}(v_2)]=0,\qquad \text{if}\quad [T_{ik}]=0,\\
& h(v_1,v_2)T_{j3}(v_1)T_{j3}(v_2)=h(v_2,v_1)T_{j3}(v_2)T_{j3}(v_1),\qquad j=1,2.\\
&h(v_2,v_1)T_{3j}(v_1)T_{3j}(v_2)=h(v_1,v_2)T_{3j}(v_2)T_{3j}(v_1),\qquad j=1,2.
\end{aligned}
\end{equation}
Finally, the graded transfer matrix is defined as the supertrace of the monodromy matrix
\be{transfer.mat}
\mathcal{T}(u)=\mathop{\rm str} T(u)= \sum_{j=1}^{3} (-1)^{[j]}\, T_{jj}(u).
\end{equation}
It is a generating function of the integrals of motion, due to the relation $[\mathcal{T}(u)\,,\,\mathcal{T}(v)]=0$.
\subsection{Notation and known results}
In this paper we use the same notation and conventions as in \cite{HutLPRS16a}. Besides the function $g(u,v)$
we shall use also two functions
\be{desand}
f(u,v)=1+g(u,v)=\frac{u-v+c}{u-v} \mb{and}
h(u,v)=\frac{f(u,v)}{g(u,v)}=\frac{u-v+c}{c}.
\end{equation}
These functions have the following obvious properties:
\be{propert}
g(u,v)=-g(v,u),\quad h(u,v+c)=\frac1{g(u,v)},\quad f(u,v+c)=\frac1{f(v,u)}.
\end{equation}
We denote sets of variables by bar: $\bar x$, $\bar u$, $\bar v$ etc.
Individual elements of the sets are denoted by Latin subscripts: $v_j$, $u_k$ etc. As a rule, the number of elements in the
sets is not shown explicitly in the equations, however we give these cardinalities in
special comments to the formulas. The notation $\bar u\pm c$ means that all the elements of the set $\bar u$ are
shifted by $\pm c$: $\bar u\pm c=\{u_1\pm c,\dots,u_n\pm c\}$. A union of sets is denoted by braces: $\{\bar u,\bar v\}\equiv
\bar u\cup \bar v$.
As a rule, subsets of variables are labeled by Roman subscripts: $\bar u_{\scriptscriptstyle \rm I}$, $\bar v_{\rm ii}$, $\bar x_{\scriptscriptstyle \rm I\hspace{-1pt}I}$ etc. The only exception
will be the notation $\bar u_j$ for a subset that is complementary to the element $u_j$, that is $\bar u_j=\bar u\setminus \{u_j\}$.
A notation $\bar u\Rightarrow\{\bar u_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ means that the
set $\bar u$ is divided into two subsets $\bar u_{\scriptscriptstyle \rm I}$ and $\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}$ such that $\{\bar u_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}=\bar u$ and $\bar u_{\scriptscriptstyle \rm I}\cap\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}
=\emptyset$. We assume that the elements in every subset of variables are ordered in such a way that the sequence of
their subscripts is strictly increasing. We call this ordering natural order.
In order to avoid too cumbersome formulas we use a shorthand notation for products of functions depending on one or two variables.
Namely, if the functions $g$, $f$, $h$ depend
on a set of variables, this means that one should take the product over the corresponding set.
For example,
\be{SH-prodllll}
h(\bar u,v)= \prod_{u_j\in\bar u} h(u_j,v);\quad
f(u_k,\bar v_j)=\prod_{\substack{v_l\in\bar v\\ l\ne j}} f(u_k,v_l);\quad
g(\bar v_{\scriptscriptstyle \rm I},\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I})=\prod_{v_j\in\bar v_{\scriptscriptstyle \rm I}}\prod_{v_k\in\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I}} g(v_j,v_k).
\end{equation}
We use the same prescription for the products of even commuting operators $T_{ij}$ and for vacuum eigenvalues
$\lambda_k$ and $r_k$ (see \eqref{Tjj}, \eqref{ratios}), for instance,
\be{SH-prodlllo}
\lambda_2(\bar v)= \prod_{v_j\in\bar v} \lambda_2(v_j);\quad r_3(\bar u)= \prod_{u_j\in\bar u} r_3(u_j);\quad T_{21}(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I})= \prod_{v_j\in \bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I}} T_{21}(v_j).
\end{equation}
For the odd operators $T_{j3}$ and $T_{3j}$, that exchange with a multiplication factor, we introduce for an arbitrary set
$\bar v=\{v_1,\dots,v_n\}$
\be{bTc-def}
\mathbb{T}_{j3}(\bar v)= \frac{T_{j3}(v_1)\dots T_{j3}(v_n)}{\prod_{n\ge \ell>m\ge 1} h(v_\ell,v_m)}, \qquad
\mathbb{T}_{3j}(\bar v)= \frac{T_{3j}(v_1)\dots T_{3j}(v_n)}{\prod_{n\ge \ell>m\ge 1} h(v_m,v_\ell)}, \qquad j=1,2.
\end{equation}
Due to commutation relations \eqref{odd-comm} the operator products \eqref{bTc-def} are symmetric over
the parameters $\bar v$ and play the same role as the bosonic products $T_{12}(\bar u)$.
We would like to stress that the convention on the shorthand notation for the products concerns the functions (operators)
depending on one or two variables only. It should not be applied to the functions, which by definition might depend
on many variables.
An example of a function depending on two sets of variables
is the partition function of the six-vertex model with domain wall boundary conditions (DWPF) \cite{Kor82,Ize87}. We denote it by
$K_n(\bar u|\bar v)$. It depends on two sets of variables $\bar u$ and $\bar v$, and the subscript shows that
$\#\bar u=\#\bar v=n$. The function $K_n$ has the following determinant representation
\begin{equation}\label{K-def}
K_n(\bar u|\bar v)
=\Delta'_n(\bar u)\Delta_n(\bar v)h(\bar u,\bar v)
\det_n \left(\frac{g(u_j,v_k)}{h(u_j,v_k)}\right),
\end{equation}
where $\Delta'_n(\bar u)$ and $\Delta_n(\bar v)$ are
\be{def-Del}
\Delta'_n(\bar u)
=\prod_{j<k}^n g(u_j,u_k),\qquad {\Delta}_n(\bar v)=\prod_{j>k}^n g(v_j,v_k).
\end{equation}
It is easy to see that $K_n$ is symmetric over $\bar u$ and symmetric over $\bar v$, however $K_n(\bar u|\bar v)\ne
K_n(\bar v|\bar u)$. Some other properties of DWPF are given in appendix~\ref{A-IHC}.
\section{Necessary tools\label{S-NT}}
In this section we recall some explicit formulas for the Bethe vectors and describe the multiple actions
of the operators $T_{ij}$ onto them.
\subsection{Bethe vectors\label{S-BV}}
Generic Bethe vectors and their dual vectors are denoted respectively by $\mathbb{B}_{a,b}(\bar u;\bar v)$
and $\mathbb{C}_{a,b}(\bar u;\bar v)$. They are parameterized by two sets of
complex parameters (Bethe parameters) $\bar u=u_1,\dots,u_a$ and $\bar v=v_1,\dots,v_b$ with $a,b=0,1,\dots$. The reader
can find several explicit representations for the Bethe vectors in terms of the monodromy matrix entries in \cite{PakRS16a}.
Here we use some of them (see below). Vectors $|0\rangle=\mathbb{B}_{0,0}(\emptyset;\emptyset)$ and $\langle0|=\mathbb{C}_{0,0}(\emptyset;\emptyset)$
respectively are called a pseudovacuum vector and a dual pseudovacuum vector. They
are eigenvectors of the diagonal entries of the monodromy matrix
\be{Tjj}
T_{ii}(u)|0\rangle=\lambda_i(u)|0\rangle, \qquad \langle0|T_{ii}(u)=\lambda_i(u)\langle0|,
\qquad i=1,2,3,
\end{equation}
where $\lambda_i(u)$ are some scalar functions. In the framework of the generalized model \cite{Kor82} considered in this paper, they remain free functional parameters.
Below it will be convenient to deal with ratios of these functions
\be{ratios}
r_1(u)=\frac{\lambda_1(u)}{\lambda_2(u)}, \qquad r_3(u)=\frac{\lambda_3(u)}{\lambda_2(u)}.
\end{equation}
Now let us give one of explicit representations for Bethe vectors obtained in \cite{PakRS16a,Sla16}
\be{Phi-expl1}
\mathbb{B}_{a,b}(\bar u;\bar v)=\sum K_n(\bar v_{\scriptscriptstyle \rm I}|\bar u_{\scriptscriptstyle \rm I})\frac{ f(\bar u_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar v_{\scriptscriptstyle \rm I})}
{\lambda_2(\bar v)\lambda_2(\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar v,\bar u)} \mathbb{T}_{13}(\bar v_{\scriptscriptstyle \rm I})\,\mathbb{T}_{23}(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I})\,T_{12}(\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I})|0\rangle.
\end{equation}
Here $K_n(\bar v_{\scriptscriptstyle \rm I}|\bar u_{\scriptscriptstyle \rm I})$ is the DWPF \eqref{K-def}. The sum is taken over partitions $\bar u\Rightarrow\{\bar u_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ and $\bar v\Rightarrow\{\bar v_{\scriptscriptstyle \rm I},\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$,
where $\#\bar v_{\scriptscriptstyle \rm I}=\#\bar u_{\scriptscriptstyle \rm I}=n$, and $n=0,1,\dots,\min(a,b)$. Recall that the notation $T_{12}(\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I})$, $ g(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar v_{\scriptscriptstyle \rm I})$,
and so on means the products of the operators (functions) over the corresponding subset (see \eqref{SH-prodllll}--\eqref{bTc-def}).
Explicit representations for dual Bethe vectors will play more important role in our calculations. We use two of them \cite{PakRS16a,Sla16}:
\be{dPhi-expl1}
\mathbb{C}_{a,b}(\bar u;\bar v)=(-1)^{\frac{b^2-b}2}\sum K_n(\bar v_{\scriptscriptstyle \rm I}|\bar u_{\scriptscriptstyle \rm I})\frac{ f(\bar u_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar v_{\scriptscriptstyle \rm I})}
{\lambda_2(\bar v)\lambda_2(\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar v,\bar u)} \langle0|T_{21}(\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I})\,\mathbb{T}_{32}(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I})\mathbb{T}_{31}(\bar v_{\scriptscriptstyle \rm I}),
\end{equation}
and
\be{dPhi-expl2}
\mathbb{C}_{a,b}(\bar u;\bar v)=(-1)^{\frac{b^2-b}2}\sum g(\bar v_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I}) \frac{ f(\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar u_{\scriptscriptstyle \rm I})f(\bar v_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar v_{\scriptscriptstyle \rm I})h(\bar u_{\scriptscriptstyle \rm I},\bar u_{\scriptscriptstyle \rm I})}
{\lambda_2(\bar u)\lambda_2(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar v,\bar u)}\;
\langle0|\mathbb{T}_{32}(\bar v_{\scriptscriptstyle \rm I\hspace{-1pt}I})\,\mathbb{T}_{31}(\bar u_{\scriptscriptstyle \rm I})\, T_{21}(\bar u_{\scriptscriptstyle \rm I\hspace{-1pt}I}).
\end{equation}
Here the sum is taken over the same partitions of the sets $\bar u$ and $\bar v$ as in \eqref{Phi-expl1}.
\subsection{Multiple actions of $T_{ij}$.\label{S-act}}
Explicit formulas for the multiple actions of the operators $T_{ij}$ onto Bethe vectors were obtained
in \cite{HutLPRS16a}. For our goal we need the actions of $T_{ij}$ with $i>j$. Below we give the corresponding formulas.
Everywhere in this section $\bar\eta=\{\bar u,\bar z\}$, $\bar\xi=\{\bar v,\bar z\}$, $\#\bar u=a$, $\#\bar v=b$, and $\#\bar z=n$.
\subsubsection{Multiple action of $T_{21}$.\label{S-act21}}
The multiple action of the operators $T_{21}$ onto Bethe vector reads
\begin{multline}\label{A-T21}
T_{21}(\bar z)\mathbb{B}_{a,b}(\bar u;\bar v)=\lambda_2(\bar z)h(\bar\xi,\bar z)\sum r_1(\bar\eta_{\scriptscriptstyle \rm I})
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}
{h(\bar\xi_{\scriptscriptstyle \rm I},\bar z)f(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})}\\
\times K_n(\bar z|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c)K_n(\bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{\scriptscriptstyle \rm I}+c)
\mathbb{B}_{a-n,b}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I};\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}).
\end{multline}
Here the functions $K_n$ are the DWPF \eqref{K-def}. The sum is taken over partitions $\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ and
$\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ with $\#\bar\xi_{\scriptscriptstyle \rm I}=\#\bar\eta_{\scriptscriptstyle \rm I}=\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n$.
If $n>a$, then the product $T_{21}(\bar z)$ annihilates $\mathbb{B}_{a,b}(\bar u;\bar v)$.
\subsubsection{Multiple action of $T_{31}$.\label{S-act31}}
The multiple action of $T_{31}(z)$ has the following form:
\begin{multline}\label{A-T310}
\mathbb{T}_{31}(\bar z)\mathbb{B}_{a,b}(\bar u;\bar v)=(-1)^{\frac{n(n+1)}2}\lambda_2(\bar z)h(\bar\xi,\bar z)\sum r_3(\bar\xi_{\scriptscriptstyle \rm I})r_1(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
\frac{g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}
{h(\bar\eta_{\scriptscriptstyle \rm I},\bar z)h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar z)}\\
\times \frac{f(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) f(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) h(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I})}{f(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I})f(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}
\;K_n(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c)\;\mathbb{B}_{a-n,b-n}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I};\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}).
\end{multline}
Here the sum is taken over partitions $\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ and
$\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ with $\#\bar\xi_{\scriptscriptstyle \rm I}=\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\#\bar\eta_{\scriptscriptstyle \rm I}=\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n$.
If $n>\min(a,b)$, then the product $\mathbb{T}_{31}(\bar z)$ annihilates $\mathbb{B}_{a,b}(\bar u;\bar v)$.
Actually, for this action we will need only the particular case $n=a$. Then
\begin{multline}\label{A-T31}
\mathbb{T}_{31}(\bar z)\mathbb{B}_{a,b}(\bar u;\bar v)=(-1)^{\frac{n(n+1)}2}\lambda_2(\bar z)h(\bar\xi,\bar z)\sum r_3(\bar\xi_{\scriptscriptstyle \rm I})r_1(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
\frac{g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}
{h(\bar\eta_{\scriptscriptstyle \rm I},\bar z)h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar z)}\\
\times \frac{f(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
h(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I})}{f(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}
\;K_n(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c)\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}).
\end{multline}
Here the sum is taken over partitions $\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ and
$\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ with $\#\bar\xi_{\scriptscriptstyle \rm I}=\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\#\bar\eta_{\scriptscriptstyle \rm I}=\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n$.
\subsubsection{Multiple action of $T_{32}$.\label{S-act32}}
The multiple action of $T_{32}(z)$ reads
\begin{multline}\label{A-T32}
\mathbb{T}_{32}(\bar z)\mathbb{B}_{a,b}(\bar u;\bar v)=(-1)^{\frac{n(n-1)}2}\lambda_2(\bar z)h(\bar\xi,\bar z)\sum r_3(\bar\xi_{\scriptscriptstyle \rm I})
\frac{f(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}
{h(\bar\eta_{\scriptscriptstyle \rm I},\bar z)h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar z)f(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}\\
\times h(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I})\;\mathbb{B}_{a,b-n}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I};\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}).
\end{multline}
Here the sum is taken over partitions $\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ and
$\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ with $\#\bar\xi_{\scriptscriptstyle \rm I}=\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\#\bar\eta_{\scriptscriptstyle \rm I}=n$. If $n>b$,
then the result of this action vanishes.
If $a=0$ and $n=b$, then
\begin{equation}\label{A-T32b2}
\mathbb{T}_{32}(\bar z)\mathbb{B}_{0,b}(\emptyset;\bar v)=(-1)^{\frac{b(b-1)}2}\lambda_2(\bar z)\sum r_3(\bar\xi_{\scriptscriptstyle \rm I})
g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})|0\rangle.
\end{equation}
The sum is taken over partitions $\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ with $\#\bar\xi_{\scriptscriptstyle \rm I}=\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n$.
\section{General form of the scalar product\label{S-GFSP}}
The scalar product of Bethe vectors is defined as
\be{Def-SP}
S_{a,b}\equiv S_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})=\mathbb{C}_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C})\mathbb{B}_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})\,,
\end{equation}
where all the Bethe parameters are generic complex numbers. We have added the superscripts $C$ and $B$
to the sets $\bar u$, $\bar v$ in order to stress that the vectors
$\mathbb{C}_{a,b}$ and $\mathbb{B}_{a,b}$ may depend on different sets of parameters.
Being a scalar function, the scalar product is invariant under the action of the antimorphism $\psi$ \eqref{psi}
\be{psiSP1}
\psi\Bigl(S_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})\Bigr)=S_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B}).
\end{equation}
On the other hand, acting with $\psi$ on the r.h.s. of \eqref{Def-SP} and using the explicit representations
\eqref{Phi-expl1} and \eqref{dPhi-expl1} for the Bethe vectors we find
\be{psiSP2}
\psi\Bigl(\mathbb{C}_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C})\mathbb{B}_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})\Bigr)=\mathbb{C}_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})\mathbb{B}_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C})=S_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B}|\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}).
\end{equation}
Here we have used $\psi(T_{j3})=T_{3j}$ and $\psi(T_{3j})=-T_{j3}$ for $j=1,2$ (see \eqref{psi}). Then using \eqref{psiAn}, and the fact that the total
number of odd operators $T_{3j}$ and $T_{j3}$ with $j=1,2$ in the scalar product is equal $2b$, we arrive at \eqref{psiSP2}.
Thus, we conclude that the scalar product is invariant under the permutation of the sets
$\{\bar{u}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle C}\}\leftrightarrow\{\bar{u}^{\scriptscriptstyle B},\bar{v}^{\scriptscriptstyle B}\}$:
\be{SP-CB}
S_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})=S_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B}|\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}).
\end{equation}
In order to calculate the scalar product one can take an explicit formula for the dual Bethe vector (\eqref{dPhi-expl1} or \eqref{dPhi-expl2})
and then use the formulas of the multiple actions \eqref{A-T21}--\eqref{A-T32b2}.
Basing on these formulas we can present
the scalar product of Bethe vectors in the following schematic form:
\be{Sab-genform}
S_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})=\sum r_1(\bar w_{\rm i})r_3(\bar w_{\rm ii})W_{\text{part}}(\bar w_{\rm i};\bar w_{\rm ii};\bar w_{\rm iii}).
\end{equation}
Here a set $\bar w$ is the union of all the Bethe parameters: $\bar w=\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B},\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$. The sum is taken over partitions
of this set into three subsets $\bar w\Rightarrow\{\bar w_{\rm i},\bar w_{\rm ii},\bar w_{\rm iii}\}$. The functions $W_{\text{part}}$ are some
rational coefficient. Their explicit forms are not important for now. We stress in \eqref{Sab-genform} that a part of the Bethe parameters $\bar w_{\rm i}$ becomes the arguments of the functions $r_1$, while the parameters $\bar w_{\rm ii}$ become the arguments of the functions $r_3$. The remaining
parameters $\bar w_{\rm iii}$ enter the rational functions $W_{\text{part}}$ only.
Let us call the set $\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}$ the parameters of $u$-type. Correspondingly, we call the set $\{\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$
the parameters of $v$-type.
\begin{prop}\label{uv-type}
The set $\bar w_{\rm i}$ in \eqref{Sab-genform} consists of parameters of the $u$-type only, while the set $\bar w_{\rm ii}$
consists of the parameters of $v$-type, that is, $\bar w_{\rm i}\subset \{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}$ and $\bar w_{\rm ii}\subset \{\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$.
Moreover, $\#\bar w_{\rm i}=a$ and $\#\bar w_{\rm ii}=b$.
\end{prop}
{\sl Proof.} Let us prove that $\bar w_{\rm i}\subset \{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}$. For this we take the dual Bethe vector in the form \eqref{dPhi-expl2}.
Let us fix a partition $\bar{u}^{\scriptscriptstyle C}\Rightarrow\{\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ in \eqref{dPhi-expl2}, such that $\#\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=n$, $n=0,1,\dots,\min(a,b)$.
Calculating the scalar product we first act with the operators $T_{21}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})$ onto the Bethe vector. Then due to
\eqref{A-T21} we obtain a sum over partitions of the set $\{\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}\}$. The terms of this sum are proportional to the
products of the functions $r_1(\bar\eta_{\scriptscriptstyle \rm I})$, where $\bar\eta_{\scriptscriptstyle \rm I}\subset \{\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}\}$ and $\#\bar\eta_{\scriptscriptstyle \rm I}=a-n$. Hence, the
parameters $\bar\eta_{\scriptscriptstyle \rm I}$ are of the $u$-type.
Next, we act with the operators $\mathbb{T}_{31}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})$ onto obtained Bethe vectors via \eqref{A-T31}. We get new partitions of the set
$\bigl\{\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}\setminus\bar\eta_{\scriptscriptstyle \rm I}\bigr\}$ and new products of functions $r_1$, say, $r_1(\bar\eta_{\scriptscriptstyle \rm I'})$. Obviously,
$\bar\eta_{\scriptscriptstyle \rm I'}\subset \bigl\{\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}\setminus\bar\eta_{\scriptscriptstyle \rm I}\bigr\}$ and $\#\bar\eta_{\scriptscriptstyle \rm I'}=n$. Thus, the total number of the
functions $r_1$ is equal to $a$, and all their arguments are of the $u$-type.
Finally, we should act with the product of the operators $\mathbb{T}_{32}(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})$. But due to \eqref{A-T32} this action does not
produce new functions $r_1$. Thus, we have proved that $\bar w_{\rm i}\subset \{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}$ and $\#\bar w_{\rm i}=a$.
Similarly, one can prove that $\bar w_{\rm ii}\subset \{\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$ and $\#\bar w_{\rm ii}=b$. However, for this one should take representation
\eqref{dPhi-expl1} for the dual Bethe vector. Then all the functions $r_3$ will be produced under the successive actions of the operators
$\mathbb{T}_{32}(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})$ and $\mathbb{T}_{31}(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})$. Repeating the considerations above we prove that all the parameters
$\bar w_{\rm ii}$ are of the $v$-type and their total number is equal to $b$.\hfill\nobreak\hbox{$\square$}\par\medbreak
Due to proposition~\ref{uv-type} one can recast \eqref{Sab-genform} in the form
\be{Sab-genform1}
S_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})=\sum r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})W_{\text{part}}(\bar\eta_{\scriptscriptstyle \rm I};\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\xi_{\scriptscriptstyle \rm I};\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}).
\end{equation}
Here $\bar\eta=\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}$ and $\bar\xi=\{\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$. The sum is taken over partitions
$\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ and $\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$, such that
$\#\bar\eta_{\scriptscriptstyle \rm I}=a$ and $\#\bar\xi_{\scriptscriptstyle \rm I}=b$. Setting in \eqref{Sab-genform1}
\be{newsets}
\begin{array}{lll}
\bar\eta_{\scriptscriptstyle \rm I}=\{\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\;\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}, &\qquad \bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\{\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\;\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}\},&\qquad \#\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}=\#\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=k,\quad k=1,\dots,a;\\
%
\bar\xi_{\scriptscriptstyle \rm I}=\{\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\;\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\} , &\qquad \bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\{\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\;\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}\},&\qquad \#\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}=\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=n,\quad n=1,\dots,b ,
\end{array}
\end{equation}
we arrive at a representation
\be{Sab-genform2}
S_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle B})=\sum \frac{r_1(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})r_1(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})r_3(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})r_3(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})}
{f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})}
W_{\text{part}}\begin{pmatrix}\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},&\bar{u}^{\scriptscriptstyle C}_{{\scriptscriptstyle \rm I}},\bar{u}^{\scriptscriptstyle B}_{{\scriptscriptstyle \rm I}}\\
\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},&\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\end{pmatrix}.
\end{equation}
Here the sum runs over all the partitions $\bar{u}^{\scriptscriptstyle C}\Rightarrow\{\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$,
$\bar{u}^{\scriptscriptstyle B}\Rightarrow\{\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$, $\bar{v}^{\scriptscriptstyle C}\Rightarrow\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ and $\bar{v}^{\scriptscriptstyle B}\Rightarrow\{\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$
with $\#\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}$ and $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}$.
The functions $W_{\text{part}}$ are rational coefficients, which depend on the partitions but do not
depend on the functions $r_1$ and $r_3$. We also have extracted explicitly the product $f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})^{-1}f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})^{-1}$ that plays
the role of a normalization factor.
\begin{Def}
We call the highest coefficient $Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})$ the function $W_{\text{\rm part}}$ that corresponds to the extreme partitions
$\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}=\emptyset$ and $\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\emptyset$:
\be{def:Zl}
W_{\text{\rm part}}\begin{pmatrix}\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B},&\emptyset,\emptyset\\
\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B},&\emptyset,\emptyset\end{pmatrix}=Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B}).
\end{equation}
In other words, $Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})$ is the coefficient of the product $r_1(\bar{u}^{\scriptscriptstyle C})r_3(\bar{v}^{\scriptscriptstyle B})$.
\end{Def}
One also can define a conjugated highest coefficient corresponding to the extreme partition $\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\emptyset$ and $\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}=\emptyset$, that is the coefficient of the product $r_1(\bar{u}^{\scriptscriptstyle B})r_3(\bar{v}^{\scriptscriptstyle C})$. However, due to \eqref{SP-CB} it is clear that
\be{def:Zr}
W_{\text{part}}\begin{pmatrix}\emptyset,\emptyset,&\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B},\\
\emptyset,\emptyset,&\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\end{pmatrix}=Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C}).
\end{equation}
We will show that all other coefficients $W_{\text{part}}$ are equal to bilinear combinations of the highest coefficient and its conjugated.
\begin{prop}\label{W-2lin}
For a fixed partition with $\#\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}=k$ and $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}=n$, (where $k=0,\dots,a$ and
$n=0,\dots,b$), the coefficient $W_{\text{\rm part}}$ has the form
\begin{multline}\label{W-Reshet}
W_{\text{\rm part}}\begin{pmatrix}\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},&\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}\\
\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},&\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\end{pmatrix}=f(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}) f(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})
f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}) f(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I})\\\rule{0pt}{20pt}
\times
Z_{a-k,n}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I};\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I};\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}) \;Z_{k,b-n}(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I};\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}|\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})\;.
\end{multline}
\end{prop}
The main goal of this paper is to find an explicit formula for the highest coefficient $Z_{a,b}$ and to prove the representation \eqref{W-Reshet} for the coefficient $W_{\text{\rm part}}$.
Comparing \eqref{W-Reshet} with the analogous formula for the $\mathfrak{gl}(3)$-based models \cite{Res86} one can see that they are
very similar. It is enough to replace the product $g(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})$ in \eqref{W-Reshet} with the product
$f(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})$ and we obtain the formula of the paper \cite{Res86}. One should remember, however, that
the highest coefficients also have different representations in the models described by $\mathfrak{gl}(2|1)$ and $\mathfrak{gl}(3)$ algebras.
In particular, we will see that in the case under consideration the highest coefficient $Z_{a,b}$ admits a single determinant representation,
while in the $\mathfrak{gl}(3)$ case such a determinant formula is not known.
\section{Successive actions\label{S-SA}}
In the previous section we have described how the scalar product depends on the functions $r_k$. Our goal now is to find
explicitly the rational coefficients $W_{\text{part}}$. For this we calculate successive action of the operators
$T_{ij}$ with $i>j$ onto a generic Bethe vector. This calculation is quite similar to the one of the work \cite{HutLPRS16a},
where we computed the multiple actions of the monodromy matrix entries.
\subsection{Successive action of $\mathbb{T}_{31}(\bar x)T_{21}(\bar y)$}
We start with the successive action of the products $\mathbb{T}_{31}(\bar x)T_{21}(\bar y)$.
Let $\#\bar x=n$ and $\#\bar y=a-n$ where $n=0,1,\dots,\min(a,b)$. Define
\be{def-G}
G_{n,a}(\bar x,\bar y)=\frac{\mathbb{T}_{31}(\bar x)T_{21}(\bar y)}{\lambda_{2}(\bar x)\lambda_{2}(\bar y)}\mathbb{B}_{a,b}(\bar u;\bar v).
\end{equation}
Using successively \eqref{A-T21} and \eqref{A-T31} we obtain
\begin{multline}\label{A-T3121}
G_{n,a}(\bar x,\bar y)=(-1)^{\frac{n(n+1)}2}h(\bar v,\bar y)h(\bar y,\bar y)\sum r_1(\bar\eta_{\scriptscriptstyle \rm I})
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}
{h(\bar\xi_{\scriptscriptstyle \rm I},\bar y)f(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})}\\
\times K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c)K_{a-n}(\bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{\scriptscriptstyle \rm I}+c)
h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)h(\bar x,\bar x)r_3(\bar\xi_{\rm i})r_1(\bar\eta_{\rm ii})
\frac{g(\bar\xi_{\rm ii},\bar\xi_{\rm i})g(\bar\xi_{\rm iii},\bar\xi_{\rm ii})g(\bar\xi_{\rm iii},\bar\xi_{\rm i})}
{h(\bar\eta_{\rm i},\bar x)h(\bar\xi_{\rm i},\bar\eta_{\rm i})h(\bar\xi_{\rm ii},\bar x)}\\
\times \frac{f(\bar\eta_{\rm i},\bar\eta_{\rm ii})
h(\bar\eta_{\rm i},\bar\eta_{\rm i})}{f(\bar\xi_{\rm i},\bar\eta_{\rm ii})f(\bar\xi_{\rm iii},\bar\eta_{\rm ii})}
\;K_n(\bar\eta_{\rm ii}|\bar\xi_{\rm ii}+c)\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\rm iii}).
\end{multline}
The sum is organized as follows. First the sets $\{\bar y,\bar u\}$ and $\{\bar y,\bar v\}$ are divided respectively into subsets
$\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ and $\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$
with the restriction $\#\bar\xi_{\scriptscriptstyle \rm I}=\#\bar\eta_{\scriptscriptstyle \rm I}=\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=a-n$. Then the sets $\{\bar x,\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ and $\{\bar x,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$
are divided respectively into subsets
$\{\bar\eta_{\rm i},\bar\eta_{\rm ii}\}$ and $\{\bar\xi_{\rm i},\bar\xi_{\rm ii},\bar\xi_{\rm iii}\}$ with the restriction
$\#\bar\xi_{\rm i}=\#\bar\xi_{\rm ii}=\#\bar\eta_{\rm i}=\#\bar\eta_{\rm ii}=n$.
It is convenient to introduce the sets $\bar\eta=\{\bar y,\bar x,\bar u\}$ and $\bar\xi=\{\bar y,\bar x,\bar v\}$. Then we can understand the sum in
\eqref{A-T3121} as the sum over partitions $\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i},\bar\eta_{\rm ii}\}$ and
$\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\rm i},\bar\xi_{\rm ii},\bar\xi_{\rm iii}\}$ with the restrictions mentioned above and an additional
constrain $\bar x\cap\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I}\} =\emptyset$. Hereby $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}= \{\bar\eta_{\rm i},\bar\eta_{\rm ii}\}\setminus \bar x$
and $\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I} = \{\bar\xi_{\rm i},\bar\xi_{\rm ii},\bar\xi_{\rm iii}\}\setminus \bar x$. Then we have
\be{project-1}
f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})=\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm ii})
f(\bar\eta_{\rm i},\bar\eta_{\scriptscriptstyle \rm I})f(\bar\eta_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})}{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)f(\bar x,\bar\eta_{\scriptscriptstyle \rm I})},
\end{equation}
and
\be{project-2}
\frac{g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}{f(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})}=
\frac{g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm iii},\bar\xi_{\scriptscriptstyle \rm I})f(\bar x,\bar\eta_{\scriptscriptstyle \rm I})}
{g(\bar x,\bar\xi_{\scriptscriptstyle \rm I})f(\bar\xi_{\rm i},\bar\eta_{\scriptscriptstyle \rm I})f(\bar\xi_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})f(\bar\xi_{\rm iii},\bar\eta_{\scriptscriptstyle \rm I})}.
\end{equation}
Observe that the restrictions $\bar x\cap\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I} =\emptyset$ and $\bar x\cap\bar\xi_{\scriptscriptstyle \rm I} =\emptyset$ hold automatically
due to the presence of the product $f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)$ in the denominator of \eqref{project-1} and the product $g(\bar x,\bar\xi_{\scriptscriptstyle \rm I})$ in the
denominator of \eqref{project-2}. Indeed, $1/f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)=0$ as soon as $\bar x\cap\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I} \ne \emptyset$ and $1/g(\bar x,\bar\xi_{\scriptscriptstyle \rm I})=0$ as soon
as $\bar x\cap\bar\xi_{\scriptscriptstyle \rm I} \ne\emptyset$. Actually, one can easily see that the condition $\bar x\cap\bar\eta_{\scriptscriptstyle \rm I} =\emptyset$ also holds, although
the product $f(\bar x,\bar\eta_{\scriptscriptstyle \rm I})$ in the denominator of \eqref{project-1} is compensated by the same product in the numerator of
\eqref{project-2}. Indeed, we have seen that $\bar x\cap\bar\xi_{\scriptscriptstyle \rm I} =\emptyset$, that is to say, $\bar x\subset\{\bar\xi_{\rm i},\bar\xi_{\rm ii},\bar\xi_{\rm iii}\}$.
But in this case $\bar x\cap\bar\eta_{\scriptscriptstyle \rm I} =\emptyset$ due to the products of the $f$-functions in the denominator of \eqref{project-2}.
Thus, we can recast \eqref{A-T3121} as follows:
\begin{multline}\label{A-T3121-1}
G_{n,a}(\bar x,\bar y)=(-1)^{\frac{n(n+1)}2}\frac{h(\bar\xi,\bar y)h(\bar\xi,\bar x)}{h(\bar x,\bar y)}\sum r_1(\bar\eta_{\scriptscriptstyle \rm I})r_1(\bar\eta_{\rm ii})r_3(\bar\xi_{\rm i})\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm ii})
f(\bar\eta_{\rm i},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\rm i},\bar\eta_{\rm ii})h(\bar\eta_{\rm i},\bar\eta_{\rm i})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)f(\bar\xi_{\rm i},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\xi_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})f(\bar\xi_{\rm iii},\bar\eta_{\scriptscriptstyle \rm I})
f(\bar\xi_{\rm i},\bar\eta_{\rm ii}) f(\bar\xi_{\rm iii},\bar\eta_{\rm ii}) }\\\rule{0pt}{20pt}
\times
\frac{g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm iii},\bar\xi_{\scriptscriptstyle \rm I})
g(\bar\xi_{\rm ii},\bar\xi_{\rm i})g(\bar\xi_{\rm iii},\bar\xi_{\rm ii})g(\bar\xi_{\rm iii},\bar\xi_{\rm i})}
{h(\bar\xi_{\scriptscriptstyle \rm I},\bar y)h(\bar\xi_{\scriptscriptstyle \rm I},\bar x)h(\bar\eta_{\rm i},\bar x)h(\bar\xi_{\rm i},\bar\eta_{\rm i})h(\bar\xi_{\rm ii},\bar x)
g(\bar x,\bar\xi_{\scriptscriptstyle \rm I})}\\\rule{0pt}{20pt}
\times
K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c) K_{a-n}(\bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{\scriptscriptstyle \rm I}+c)\;K_n(\bar\eta_{\rm ii}|\bar\xi_{\rm ii}+c)\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\rm iii}).
\end{multline}
Here we have also used
\be{h-h}
h(\bar x,\bar x)h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)=\frac{h(\bar\xi,\bar x)}{h(\bar\xi_{\scriptscriptstyle \rm I},\bar x)}.
\end{equation}
In \eqref{A-T3121-1} the sum is taken over partitions $\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i},\bar\eta_{\rm ii}\}$ and
$\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\rm i},\bar\xi_{\rm ii},\bar\xi_{\rm iii}\}$. The restriction are imposed on the cardinalities of
the subsets only.
One can reduce the number of subsets in \eqref{A-T3121-1}. Let $\bar\eta_{0}=\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\rm ii}\}$. Then \eqref{A-T3121-1} takes the form
\begin{multline}\label{A-T3121-1a}
G_{n,a}(\bar x,\bar y)=(-1)^{\frac{n(n+1)}2}\frac{h(\bar\xi,\bar y)h(\bar\xi,\bar x)}{h(\bar x,\bar y)}\sum r_1(\bar\eta_{0})r_3(\bar\xi_{\rm i})
\frac{ f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{0})f(\bar\eta_{\rm i},\bar\eta_{0})}{f(\bar\xi_{\rm iii},\bar\eta_{0})
f(\bar\xi_{\rm i},\bar\eta_{0})}\;K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c)\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i})h(\bar\eta_{\rm i},\bar\eta_{\rm i}) g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm iii},\bar\xi_{\scriptscriptstyle \rm I})
g(\bar\xi_{\rm ii},\bar\xi_{\rm i})g(\bar\xi_{\rm iii},\bar\xi_{\rm ii})g(\bar\xi_{\rm iii},\bar\xi_{\rm i})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) h(\bar\xi_{\scriptscriptstyle \rm I},\bar y)h(\bar\xi_{\scriptscriptstyle \rm I},\bar x)h(\bar\eta_{\rm i},\bar x)h(\bar\xi_{\rm i},\bar\eta_{\rm i})h(\bar\xi_{\rm ii},\bar x)
g(\bar x,\bar\xi_{\scriptscriptstyle \rm I})}\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\rm iii})\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})}{f(\bar\xi_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})}
K_{a-n}(\bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{\scriptscriptstyle \rm I}+c)\;K_n(\bar\eta_{\rm ii}|\bar\xi_{\rm ii}+c).
\end{multline}
We see that the sum over partitions $\bar\eta_{0}\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\rm ii}\}$ involves the terms in the last line
only. This sum can be computed via lemma~\ref{main-ident}. Using \eqref{Sym-Part-old1} we find
\begin{multline}\label{Appl-ML}
\sum_{\bar\eta_{0}\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\rm ii}\}}
K_{a-n}(\bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{\scriptscriptstyle \rm I}+c)\;K_n(\bar\eta_{\rm ii}|\bar\xi_{\rm ii}+c)\frac{f(\bar\eta_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})}{f(\bar\xi_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})}\\
=\frac{(-1)^n}{f(\bar\xi_{\rm ii},\bar\eta_{0})}
\sum_{\bar\eta_{0}\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\rm ii}\}}
K_{a-n}(\bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{\scriptscriptstyle \rm I}+c)\;K_n(\bar\xi_{\rm ii}|\bar\eta_{\rm ii})f(\bar\eta_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I})
=\frac{(-1)^a K_a(\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\rm ii}\}|\bar\eta_{0})}{f(\bar\xi_{\rm ii},\bar\eta_{0})f(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{0})}.
\end{multline}
Thus, \eqref{A-T3121-1a} takes the form
\begin{multline}\label{A-T3121-2}
G_{n,a}(\bar x,\bar y)=(-1)^{a+\frac{n(n+1)}2}\frac{h(\bar\xi,\bar y)h(\bar\xi,\bar x)}{h(\bar x,\bar y)}\sum r_1(\bar\eta_{0})r_3(\bar\xi_{\rm i})
K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c)\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{0}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i})
f(\bar\eta_{\rm i},\bar\eta_{0})h(\bar\eta_{\rm i},\bar\eta_{\rm i})
g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm iii},\bar\xi_{\scriptscriptstyle \rm I})
g(\bar\xi_{\rm ii},\bar\xi_{\rm i})g(\bar\xi_{\rm iii},\bar\xi_{\rm ii})g(\bar\xi_{\rm iii},\bar\xi_{\rm i})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) f(\bar\xi_{\rm ii},\bar\eta_{0})f(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{0}) f(\bar\xi_{\rm iii},\bar\eta_{0})
f(\bar\xi_{\rm i},\bar\eta_{0})
h(\bar\xi_{\scriptscriptstyle \rm I},\bar y)h(\bar\xi_{\scriptscriptstyle \rm I},\bar x)h(\bar\eta_{\rm i},\bar x)h(\bar\xi_{\rm i},\bar\eta_{\rm i}) h(\bar\xi_{\rm ii},\bar x)
g(\bar x,\bar\xi_{\scriptscriptstyle \rm I}) }\\\rule{0pt}{20pt}
\times K_a(\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\rm ii}\}|\bar\eta_{0})\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\rm iii}).
\end{multline}
Now we define $\bar\xi_{0}=\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\rm ii}\}$. Then \eqref{A-T3121-2} can be written as
\begin{multline}\label{A-T3121-2a}
G_{n,a}(\bar x,\bar y)=(-1)^{a+\frac{n(n-1)}2}\frac{h(\bar\xi,\bar y)h(\bar\xi,\bar x)}{h(\bar x,\bar y)}\sum r_1(\bar\eta_{0})r_3(\bar\xi_{\rm i})
K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c)K_a(\bar\xi_{0}|\bar\eta_{0})\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{0}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i})
f(\bar\eta_{\rm i},\bar\eta_{0})h(\bar\eta_{\rm i},\bar\eta_{\rm i})
g(\bar\xi_{\rm i},\bar\xi_{0})g(\bar\xi_{\rm iii},\bar\xi_{0})g(\bar\xi_{\rm iii},\bar\xi_{\rm i})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) f(\bar\xi_{0},\bar\eta_{0}) f(\bar\xi_{\rm iii},\bar\eta_{0})
f(\bar\xi_{\rm i},\bar\eta_{0})h(\bar\eta_{\rm i},\bar x)h(\bar\xi_{\rm i},\bar\eta_{\rm i})h(\bar\xi_{0},\bar x)}
\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\rm iii}) \\\rule{0pt}{20pt}
\times\frac{g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})}{ h(\bar\xi_{\scriptscriptstyle \rm I},\bar y)g(\bar x,\bar\xi_{\scriptscriptstyle \rm I})}.
\end{multline}
The sum over partitions $\bar\xi_{0}\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\rm ii}\}$ involves
the terms in the last line only. It can be computed via \eqref{Sym-Part-new1}:
\begin{multline}\label{anal-ML}
\sum \frac{g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})}{g(\bar x,\bar\xi_{\scriptscriptstyle \rm I})h(\bar\xi_{\scriptscriptstyle \rm I},\bar y)}=\frac{(-1)^{n(a-n)}}{g(\bar\xi_{0},\bar x)}
\sum \frac{g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar x)}{h(\bar\xi_{\scriptscriptstyle \rm I},\bar y)}\\
=\frac{(-1)^{n(a-n)}}{g(\bar\xi_{0},\bar x)}
\sum g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar x)g(\bar\xi_{\scriptscriptstyle \rm I},\bar y-c)=\frac{h(\bar x,\bar y)}{h(\bar\xi_{0},\bar y)}.
\end{multline}
Thus, we arrive at
\begin{multline}\label{A-T3121-3}
G_{n,a}(\bar x,\bar y)=(-1)^{a+\frac{n(n-1)}2}h(\bar\xi,\bar y)h(\bar\xi,\bar x)\sum \frac{r_1(\bar\eta_{0})r_3(\bar\xi_{\rm i})}
{h(\bar\xi_{0},\bar x)h(\bar\xi_{0},\bar y)}
K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c) K_a(\bar\xi_{0}|\bar\eta_{0})\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{0}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\rm i})
f(\bar\eta_{\rm i},\bar\eta_{0})h(\bar\eta_{\rm i},\bar\eta_{\rm i}) g(\bar\xi_{\rm i},\bar\xi_{0}) g(\bar\xi_{\rm iii},\bar\xi_{0})
g(\bar\xi_{\rm iii},\bar\xi_{\rm i})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) f(\bar\xi_{0},\bar\eta_{0}) f(\bar\xi_{\rm iii},\bar\eta_{0})
f(\bar\xi_{\rm i},\bar\eta_{0}) h(\bar\eta_{\rm i},\bar x)h(\bar\xi_{\rm i},\bar\eta_{\rm i}) }
\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\rm iii}).
\end{multline}
Finally, after a relabeling of the subsets
$\bar\eta_{0}\to \bar\eta_{\scriptscriptstyle \rm I}$, $\bar\eta_{\rm i}\to \bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}$, $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}\to \bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}$,
$\bar\xi_{\rm i}\to\bar\xi_{\scriptscriptstyle \rm I}$, $\bar\xi_{0}\to\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}$, $\bar\xi_{\rm iii}\to\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}$,
we recast \eqref{A-T3121-3} as follows:
\begin{multline}\label{A-T3121-4}
G_{n,a}(\bar x,\bar y)=(-1)^{a+\frac{n(n-1)}2}h(\bar\xi,\bar y)h(\bar\xi,\bar x)\sum \frac{r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})}
{h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar y)}
K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c) K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I})\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})
g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar x) f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I}) h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }
\;\mathbb{B}_{0,b-n}(\emptyset;\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}).
\end{multline}
We recall that the cardinalities of the subsets are
\be{card-new0}
\begin{aligned}
\#\bar\eta_{\scriptscriptstyle \rm I}=a,\qquad & \#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n,\qquad & \#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=a-n,\\
\#\bar\xi_{\scriptscriptstyle \rm I}=n,\qquad & \#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=a,\qquad & \#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=b-n.
\end{aligned}
\end{equation}
\begin{rmk}\label{REM-lim}
Strictly speaking, the sets $\bar\eta$ and $\bar\xi$ in equation \eqref{A-T3121-4} should be understood as
\be{eps}
\begin{array}{l}
\bar\eta=\{\bar x+\epsilon_1,\bar y+\epsilon_1, \bar u+\epsilon_1\},\\
\bar\xi=\{\bar x+\epsilon_2,\bar y+\epsilon_2, \bar v+\epsilon_2\},
\end{array}
\qquad\text{at}\quad \epsilon_{k}\to 0,\qquad k=1,2.
\end{equation}
The point is that individual
factors in \eqref{A-T3121-4} may have singularities, if we set $\epsilon_k=0$. For instance, if $\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\cap\bar\eta_{\scriptscriptstyle \rm I}\ne\emptyset$,
then the DWPF $K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I})$ is singular. However, these poles are compensated by the product $f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I})^{-1}$.
Therefore, for appropriate evaluating the limit we should have $\epsilon_k\ne 0$. In order to lighten the formulas we do
not write these auxiliary parameters $\epsilon_k$ explicitly, but one has to keep them in mind when doing the calculations.
\end{rmk}
\subsection{Successive action of $\mathbb{T}_{32}(\bar z)\mathbb{T}_{31}(\bar x)T_{21}(\bar y)$}
Let now $\#\bar z=b-n$. Then we define
\be{ewa-act}
\frac{\mathbb{T}_{32}(\bar z)\, \mathbb{T}_{31}(\bar x)\, T_{21}(\bar y)}{\lambda_2(\bar z)\, \lambda_2(\bar x)\,\lambda_2(\bar y)}
\mathbb{B}_{a,b}(\bar u,\bar v)=\frac{\mathbb{T}_{32}(\bar z)}{\lambda_2(\bar z)}G_{n,a}(\bar x,\bar y)=H_{n,a,b}(\bar x,\bar y,\bar z)|0\rangle.
\end{equation}
In order to act with $\mathbb{T}_{32}(\bar z)$ onto $G_{n,a}(\bar x,\bar y)$ we should use \eqref{A-T32b2}. Let us denote the union
$\{\bar x,\bar y\}$ as $\bar{u}^{\scriptscriptstyle C}$ (as it will be in the case of the scalar product). Then we obtain
\begin{multline}\label{A-T32T3121-1}
H_{n,a,b}(\bar x,\bar y,\bar z)=(-1)^{a+\frac{n(n-1)}2+\frac{(b-n)(b-n-1)}2}h(\bar v,\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})
\sum \frac{r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\rm i})}{h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})}
\\\rule{0pt}{20pt}
\times K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c) K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I})
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar x) f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I}) }\\
\times \frac{ g(\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})
g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}
{h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x)h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }g(\bar\xi_{\rm ii},\bar\xi_{\rm i}).
\end{multline}
Here the partitions of the set $\bar\eta$ remain the same as in \eqref{A-T3121-4}. The partitions of the remaining variables
are organized as follows. We first have the partitions of the set $\{\bar{u}^{\scriptscriptstyle C},\bar v\}=\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$.
Then we combine $\{\bar z,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ and obtain additional partitions $\{\bar z,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}\Rightarrow\{\bar\xi_{\rm i},\bar\xi_{\rm ii}\}$ with
the restriction $\#\bar\xi_{\rm i}=\#\bar\xi_{\rm ii}=b-n$.
We should substitute $\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=\{\bar\xi_{\rm i},\bar\xi_{\rm ii}\}\setminus \bar z$ into \eqref{A-T32T3121-1}. Then, using
\be{subs-use}
\frac{g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})}{f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I})}=
\frac{g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})}
{f(\bar v,\bar\eta_{\scriptscriptstyle \rm I})f(\bar{u}^{\scriptscriptstyle C},\bar\eta_{\scriptscriptstyle \rm I})g(\bar z,\bar\xi_{\scriptscriptstyle \rm I})g(\bar z,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})},
\end{equation}
we arrive at
\begin{multline}\label{A-T32T3121-2}
H_{n,a,b}(\bar x,\bar y,\bar z)=(-1)^{a+\frac{n(n-1)}2+\frac{(b-n)(b-n-1)}2}h(\bar v,\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})
\sum \frac{r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\rm i}) }{h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})}
\\\rule{0pt}{20pt}
\times K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c) K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I})
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar z,\bar\eta_{\scriptscriptstyle \rm I})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar x) f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I}) }\\
\times \frac{g(\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})
g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I}) g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm ii},\bar\xi_{\rm i})}
{h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar z,\bar\xi_{\scriptscriptstyle \rm I})g(\bar z,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }.
\end{multline}
Here we have denoted by $\bar\xi$ the union $\{\bar z,\bar{u}^{\scriptscriptstyle C},\bar v\}$. This set is divided into four subsets
$\bar\xi\Rightarrow\{\bar\xi_{\rm i},\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ with the cardinalities
$\#\bar\xi_{\rm i}=\#\bar\xi_{\rm ii}=b-n$, $\#\bar\xi_{\scriptscriptstyle \rm I}=n$, and $\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=a$.
Let $\bar\xi_{0}=\{\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I}\}$. Then
\begin{multline}\label{A-T32T3121-2a}
H_{n,a,b}(\bar x,\bar y,\bar z)=(-1)^{a+n(b+1)+\frac{b(b-1)}2}h(\bar v,\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})
\sum \frac{r_1(\bar\eta_{\scriptscriptstyle \rm I}) r_3(\bar\xi_{0}) }{h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})}K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c)
\\\rule{0pt}{20pt}
\times K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I})
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar z,\bar\eta_{\scriptscriptstyle \rm I})
g(\bar\xi_{0},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})
g(\bar\xi_{\rm ii},\bar\xi_{0}) }
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar x) f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I}) h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) g(\bar z,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }\\
\times
\frac{g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I}) }{ h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar z,\bar\xi_{\scriptscriptstyle \rm I})}.
\end{multline}
The sum over partitions $\bar\xi_{0}\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\rm i}\}$ involves
the terms in the last line only. It can be computed via \eqref{Sym-Part-new1}:
\be{sum-OM}
\sum\frac{g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I}) } {h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar z,\bar\xi_{\scriptscriptstyle \rm I})}=
\frac{(-1)^{b-n}}{g(\bar z,\bar\xi_{0})}\sum\frac{g(\bar\xi_{\rm i},\bar\xi_{\scriptscriptstyle \rm I})g(\bar\xi_{\rm i},\bar z) } {h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}
=\frac{h(\bar z,\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}{h(\bar\xi_{0},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}.
\end{equation}
Substituting this into \eqref{A-T32T3121-2a} we find
\begin{multline}\label{A-T32T3121-3}
H_{n,a,b}(\bar x,\bar y,\bar z)=(-1)^{a+n(b+1)+\frac{b(b-1)}2}h(\bar v,\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})
\sum \frac{r_1(\bar\eta_{\scriptscriptstyle \rm I}) r_3(\bar\xi_{0}) }{h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})}
\\\rule{0pt}{20pt}
\times
\frac{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar z,\bar\eta_{\scriptscriptstyle \rm I})g(\bar\xi_{0},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})
g(\bar\xi_{\rm ii},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\xi_{\rm ii},\bar\xi_{0})h(\bar z,\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}
{f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar x) f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) h(\bar\xi_{0},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar z,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }\\\rule{0pt}{20pt}
\times
K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c) K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I}).
\end{multline}
Finally, relabeling $\bar\xi_{0}\to \bar\xi_{\scriptscriptstyle \rm I}$ and $\bar\xi_{\rm ii}\to \bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}$ we arrive at
\begin{multline}\label{H-1}
H_{n,a,b}(\bar x,\bar y,\bar z)=(-1)^{a+n(b+1)+\frac{b^2-b}2}\frac{h(\bar\xi,\bar{u}^{\scriptscriptstyle C})}{h(\bar z,\bar{u}^{\scriptscriptstyle C})}\sum r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})
\,K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I}) K_{a-n}(\bar y|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c)\\
\times\frac{f(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar z,\bar\eta_{\scriptscriptstyle \rm I})h(\bar z,\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
g(\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }
{ h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I}) h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar x)g(\bar z ,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) }.
\end{multline}
Recall that in this formula $\bar\eta=\{\bar x,\bar y,\bar u\}$, $\bar\xi=\{\bar z, \bar x,\bar y,\bar v\}$, and we denote $\bar{u}^{\scriptscriptstyle C}=\{\bar x,\bar y\}$.
The sum is taken over partitions
$\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ and $\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$.
Hereby $\bar\eta_{0}=\{\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$. The cardinalities of the subsets are
\be{card}
\begin{aligned}
&\#\bar\eta_{\scriptscriptstyle \rm I}=a,\qquad &\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n,\qquad &\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=a-n,\\
&\#\bar\xi_{\scriptscriptstyle \rm I}=b,\qquad &\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=a,\qquad &\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=b-n.
\end{aligned}
\end{equation}
\section{Highest coefficients\label{S-HC}}
Equation \eqref{H-1} allows us to obtain an explicit representation for the scalar product of Bethe vectors. Using
\eqref{dPhi-expl2} we find
\be{SP-first}
S_{a,b}=\frac{ (-1)^{\frac{b^2-b}2} }{f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})}
\sum g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}) f(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})h(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})
\;H_{n,a,b}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}).
\end{equation}
The sum is taken over partitions $\bar{u}^{\scriptscriptstyle C}\Rightarrow\{\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ and $\bar{v}^{\scriptscriptstyle C}\Rightarrow\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$,
where $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=n$, and $n=0,1,\dots,\min(a,b)$. The function $H_{n,a,b}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})$ itself
is given as a sum over partitions described in \eqref{H-1}. Namely, the union $\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}$ is divided into three subsets and
the union $\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$ also is divided into three subsets. Although the resulting formula is explicit, it is inconvenient for later use.
Therefore, we will try to simplify it. To do this, we introduce a new function.
\begin{Def}
Let $\bar x$, $\bar y$, $\bar t$, $\bar s$, and $\bar\beta$ be five sets of generic complex numbers with cardinalities
$\#\bar x=n$, $\#\bar y=m$, and $\#\bar \beta=n+m$. The cardinalities of the sets $\bar t$ and $\bar s$ are not fixed.
Define a function
\be{def-J}
J_{n,m}(\bar x;\bar y|\bar t;\bar s|\bar\beta)=
\Delta_{n+m}(\bar\beta)\Delta'_n(\bar x)\Delta'_m(\bar y)\;
\det_{n+m}\mathcal{J}_{jk},
\end{equation}
where
\be{matelJ}
\begin{aligned}
&\mathcal{J}_{jk}=\frac{g(\beta_j,x_k)}{h(\beta_j,x_k)},&\qquad k=1,\dots,n;\\
&\mathcal{J}_{j,k+n}=g(\beta_j,y_k)\frac{h(\beta_j,\bar t)}{h(\beta_j,\bar s)},&\qquad k=1,\dots,m;
\end{aligned}
\qquad j=1,\dots,n+m.
\end{equation}
\end{Def}
Developing the determinant in \eqref{def-J} with respect to the first $n$ columns (see Appendix~\ref{A-SF} for more details) we obtain a presentation of $J_{n,m}$ as a sum
over partitions of the set $\bar\beta$:
\be{J-sum}
J_{n,m}(\bar x;\bar y|\bar t;\bar s|\bar\beta)=\sum K_n(\bar\beta_{\scriptscriptstyle \rm I}|\bar x)\frac{g(\bar\beta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\beta_{\scriptscriptstyle \rm I})g(\bar\beta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar y)h(\bar\beta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar t)}
{h(\bar\beta_{\scriptscriptstyle \rm I},\bar x)h(\bar\beta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar s)}.
\end{equation}
Here the sum is taken over partitions $\bar\beta\Rightarrow\{\bar\beta_{\scriptscriptstyle \rm I},\bar\beta_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$, such that $\#\bar\beta_{\scriptscriptstyle \rm I}=n$ and $\#\bar\beta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=m$.
\subsection{First representation for the highest coefficient}
Let us find the highest coefficient $Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})$. We recall that up to the normalization factor
$\bigl(f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})\bigr)^{-1}$ it is the rational coefficient of the product $r_1(\bar{u}^{\scriptscriptstyle B})r_3(\bar{v}^{\scriptscriptstyle C})$
(see \eqref{def:Zr}, \eqref{W-Reshet}).
Obviously, for this we should set $\bar\eta_{\scriptscriptstyle \rm I}=\bar{u}^{\scriptscriptstyle B}$ and $\bar\xi_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle C}$ in \eqref{H-1}. However,
$\bar\xi_{\scriptscriptstyle \rm I}\subset\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$. Hence, one can have $\bar\xi_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle C}$ if and only if $\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\bar{v}^{\scriptscriptstyle C}$, and thus,
$\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\emptyset$. But $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=n$ in \eqref{SP-first}, therefore $\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\emptyset$ and
$n=0$. Thus, \eqref{SP-first} takes the form
\be{HC-rep1}
\frac{r_1(\bar{u}^{\scriptscriptstyle B})r_3(\bar{v}^{\scriptscriptstyle C})\;Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})}{f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})}=\frac{ (-1)^{\frac{b^2-b}2} }{f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})}
\;H_{0,a,b}(\emptyset,\bar{u}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle C})\Bigr|_{\bar\eta_{\scriptscriptstyle \rm I}=\bar{u}^{\scriptscriptstyle B};~\bar\xi_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle C}}.
\end{equation}
Substituting the conditions $\bar\eta_{\scriptscriptstyle \rm I}=\bar{u}^{\scriptscriptstyle B}$ and $\bar\xi_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle C}$ into \eqref{H-1} we should take into account that
$\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n=0$ (see \eqref{card}). Hence, $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\emptyset$, which implies $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=\bar\eta_{0}=\bar{u}^{\scriptscriptstyle C}$. Thus,
substituting these subsets into \eqref{H-1} we find
\begin{multline}\label{Z-pc1}
r_1(\bar{u}^{\scriptscriptstyle B})r_3(\bar{v}^{\scriptscriptstyle C})\;Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})=(-1)^{a}h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C}) r_1(\bar{u}^{\scriptscriptstyle B})r_3(\bar{v}^{\scriptscriptstyle C})
\\
\times K_{a}(\bar{u}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle C}+c)\sum K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar{u}^{\scriptscriptstyle B})
\frac{ g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }
{ h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) },
\end{multline}
where the sum is taken over partitions $\{\bar{u}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}=\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ with $\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=a$ and
$\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=b$. Due to \eqref{K-K} we conclude that $K_{a}(\bar{u}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle C}+c)=(-1)^a$, and we arrive at
\begin{equation}\label{Z-pc2}
Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})=h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})\sum K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar{u}^{\scriptscriptstyle B})
\frac{ g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }
{ h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) }.
\end{equation}
Finally, using $\{\bar{u}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}=\bar\xi$ we recast \eqref{Z-pc2} as follows:
\begin{multline}\label{Z-pc3}
Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})=\sum K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar{u}^{\scriptscriptstyle B})
g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})\\
=h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B})\sum K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar{u}^{\scriptscriptstyle B})
\frac{ g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) }
{ h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}) }.
\end{multline}
Comparing \eqref{Z-pc3} and \eqref{J-sum} we conclude that
\be{Z-J}
Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})=h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B})\;J_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C}|\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\{\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C}\}).
\end{equation}
Thus, we have obtained an explicit representation for the highest coefficient $Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})$
in terms of the determinant of the $(a+b)\times(a+b)$ matrix $\mathcal{J}_{jk}$ \eqref{matelJ}.
\subsection{Second highest coefficient}
In order to obtain the second highest coefficient $Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})$ it is enough to make the
replacements $\bar{u}^{\scriptscriptstyle C}\leftrightarrow\bar{u}^{\scriptscriptstyle B}$ and $\bar{v}^{\scriptscriptstyle C}\leftrightarrow\bar{v}^{\scriptscriptstyle B}$ in \eqref{Z-J}. On the other hand, this coefficient should
arise if we set $\bar\eta_{\scriptscriptstyle \rm I}=\bar{u}^{\scriptscriptstyle C}$ and $\bar\xi_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle B}$ in \eqref{H-1}. However, if we do so, then we do not obtain
\eqref{Z-J} with the replacements mentioned above. Instead, we obtain much more sophisticated formula involving many sums over partitions. This `break of symmetry'
occurs because we use a specific representation \eqref{dPhi-expl2} for the dual Bethe vector. If we would use equation \eqref{dPhi-expl1}
for $\mathbb{C}_{a,b}(\bar u;\bar v)$, then we would have an analog of \eqref{Z-J} for $Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})$, however, we would
have a more complex formula for $Z_{a,b}(\bar{u}^{\scriptscriptstyle B};\bar{u}^{\scriptscriptstyle C}|\bar{v}^{\scriptscriptstyle B};\bar{v}^{\scriptscriptstyle C})$.
A `complex' formula for the highest coefficient provides
us with a very non-trivial identity for $Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})$, that will be used later. In order to obtain this
identity we first make several additional summations in \eqref{SP-first}. Let us rewrite this equation explicitly
\begin{multline}\label{SP-expl1}
S_{a,b}=\frac{ h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})} {f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})}
\sum (-1)^{a+n(b+1)} g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}) g(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})\\
\times r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})
\,K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I}) K_{a-n}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c)\\
\times\frac{f(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
g(\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }
{ h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) f(\bar\xi,\bar\eta_{\scriptscriptstyle \rm I}) h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}) f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I} ,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) }.
\end{multline}
The sum over partitions into subsets $\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}$ and $\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}$, as well as the sum over partitions $\bar{u}^{\scriptscriptstyle C}\Rightarrow\{\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$
can be computed in terms of
the function $J$ \eqref{def-J}. Let $\bar\xi_{0}=\{\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$. Then
\begin{multline}\label{sum-xi}
\sum_{\bar\xi_{0}\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}}
\frac{ K_a(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I})g(\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I}) g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})}
{g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I} ,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})}\\
=\frac{(-1)^{ab+n+b}g(\bar\xi_{0},\bar\xi_{\scriptscriptstyle \rm I})h(\bar\xi_{0},\bar\eta_{\scriptscriptstyle \rm I})}
{g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{0}) h(\bar\xi_{0},\bar{u}^{\scriptscriptstyle C})}J_{a,b-n}(\bar\eta_{\scriptscriptstyle \rm I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}| \bar{u}^{\scriptscriptstyle C}; \bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{0}).
\end{multline}
Similarly, one can verify that
\begin{multline}\label{sum-u}
\sum_{\bar{u}^{\scriptscriptstyle C}\Rightarrow\{\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}}\frac{ K_{a-n}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}+c)
g(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}) h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})}
{h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})}\\
=(-1)^{a+n+an}\frac{h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C})}{g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})}J_{a-n,n}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}| \bar{u}^{\scriptscriptstyle C}+c; \bar\eta_{0}+c|\bar{u}^{\scriptscriptstyle C}-c).
\end{multline}
Substituting these results into \eqref{SP-expl1} we find
\begin{multline}\label{SP-expl2}
S_{a,b}=\frac{ h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C})h(\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})} {f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})}
\sum (-1)^{b+ab+n(a+b+1)} r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})J_{a,b-n}(\bar\eta_{\scriptscriptstyle \rm I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}| \bar{u}^{\scriptscriptstyle C}; \bar\eta_{\scriptscriptstyle \rm I}|\bar\xi_{0})\\
\times J_{a-n,n}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}| \bar{u}^{\scriptscriptstyle C}+c; \bar\eta_{0}+c|\bar{u}^{\scriptscriptstyle C}-c)f(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
\\
\times\frac{ h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}) g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})
g(\bar\xi_{0},\bar\xi_{\scriptscriptstyle \rm I})h(\bar\xi_{0},\bar\eta_{\scriptscriptstyle \rm I}) }
{g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) f(\bar{u}^{\scriptscriptstyle C},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar{v}^{\scriptscriptstyle B},\bar\eta_{\scriptscriptstyle \rm I}) h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{0})h(\bar\xi_{0},\bar{u}^{\scriptscriptstyle C}) }.
\end{multline}
The sum is taken over partitions:
\\
\begin{tabular}{l}
(1) $\bar{v}^{\scriptscriptstyle C}\Rightarrow\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ with $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=n$ and $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=b-n$;\\
(2) $\{\bar{u}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}=\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{0}\}$ with $\#\bar\xi_{\scriptscriptstyle \rm I}=b$ and $\#\bar\xi_{0}=b+a-n$;\\
(3) $\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}=\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{0}\}$ and $\bar\eta_{0}\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}$ with
$\#\bar\eta_{\scriptscriptstyle \rm I}=a$, $\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n$, and $\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=a-n$.
\end{tabular}\\
In all these partitions $n=0,1,\dots,\min(a,b)$.
Due to proposition~\ref{uv-type} the function $r_3$ depends on the variables of the $v$-type only. Hence, $\bar\xi_{\scriptscriptstyle \rm I}\cap
\bar{u}^{\scriptscriptstyle C}=\emptyset$, that is $\bar{u}^{\scriptscriptstyle C}\subset\bar\xi_{0}$. Therefore, we can set $\bar\xi_{0}=\{\bar{u}^{\scriptscriptstyle C},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$. Substituting this
into \eqref{SP-expl2} we obtain
\begin{multline}\label{SP-expl3}
S_{a,b}=\frac{ h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C})} {f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})}
\sum (-1)^{b+n(a+b+1)} r_1(\bar\eta_{\scriptscriptstyle \rm I})r_3(\bar\xi_{\scriptscriptstyle \rm I})J_{a,b-n}(\bar\eta_{\scriptscriptstyle \rm I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}| \bar{u}^{\scriptscriptstyle C}; \bar\eta_{\scriptscriptstyle \rm I}|\{\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}\})\\
\times J_{a-n,n}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}| \bar{u}^{\scriptscriptstyle C}+c; \bar\eta_{0}+c|\bar{u}^{\scriptscriptstyle C}-c)f(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I})h(\bar\eta_{0},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
\\
\times\frac{ h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}) g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})
g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I})g(\bar\xi_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}) }
{g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) g(\bar{u}^{\scriptscriptstyle C},\bar\eta_{\scriptscriptstyle \rm I}) f(\bar{v}^{\scriptscriptstyle B},\bar\eta_{\scriptscriptstyle \rm I}) h(\bar\xi_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})h(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) }.
\end{multline}
In this formula $\bar\xi=\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}\}$, $\#\bar\xi_{\scriptscriptstyle \rm I}=b$ and $\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=b-n$. All other subsets are the same as in \eqref{SP-expl2}.
Now everything is ready to formulate the second representation for the highest coefficient $ Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})$.
For this we set $\bar\eta_{\scriptscriptstyle \rm I}=\bar{u}^{\scriptscriptstyle C}$ and $\bar\xi_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle B}$. Then
automatically
$\bar\eta_0=\bar{u}^{\scriptscriptstyle B}$, $\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}$ and we also can set $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}$, $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}$.
Substituting this into \eqref{SP-expl3} and keeping in mind remark \ref{REM-lim} we obtain
\begin{multline}\label{SP-3pc1}
Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})=f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})f(\bar{u}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C})\sum (-1)^{n(a+b+1)+b}
g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}) \\
%
\times J_{a-n,n}(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}|\bar{u}^{\scriptscriptstyle C}+c;\bar{u}^{\scriptscriptstyle B}+c|\bar{u}^{\scriptscriptstyle C}-c)\frac{h(\bar{u}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})
g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}) h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) }
{g(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I} ,\bar{u}^{\scriptscriptstyle C})h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C}) }
\, \\\rule{0pt}{20pt}
\times \lim_{\substack{\bar\eta_{\scriptscriptstyle \rm I}\to\bar{u}^{\scriptscriptstyle C}\\ \bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\to\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}}}
\frac{J_{a,b-n}(\bar\eta_{\scriptscriptstyle \rm I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}| \bar{u}^{\scriptscriptstyle C}; \bar\eta_{\scriptscriptstyle \rm I}|\{\bar{u}^{\scriptscriptstyle C},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\})}{g(\bar{u}^{\scriptscriptstyle C},\bar\eta_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I} ,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})}.
\end{multline}
Using \eqref{limW-1} and \eqref{limW-2} we find
\be{lim}
\lim_{\substack{\bar\eta_{\scriptscriptstyle \rm I}\to\bar{u}^{\scriptscriptstyle C}\\ \bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\to\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}}}
\frac{J_{a,b-n}(\bar\eta_{\scriptscriptstyle \rm I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}| \bar{u}^{\scriptscriptstyle C}; \bar\eta_{\scriptscriptstyle \rm I}|\{\bar{u}^{\scriptscriptstyle C},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\})}{g(\bar{u}^{\scriptscriptstyle C},\bar\eta_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I} ,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I})}
=(-1)^{b+n}g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I} ,\bar{u}^{\scriptscriptstyle C}),
\end{equation}
and we thus arrive at
\begin{multline}\label{Ident-1-main}
Z_{a,b}(\bar{u}^{\scriptscriptstyle C};\bar{u}^{\scriptscriptstyle B}|\bar{v}^{\scriptscriptstyle C};\bar{v}^{\scriptscriptstyle B})= f(\bar{u}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})\sum (-1)^{n(a+b)}
J_{a-n,n}(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}|\bar{u}^{\scriptscriptstyle C}+c;\bar{u}^{\scriptscriptstyle B}+c|\bar{u}^{\scriptscriptstyle C}-c)\\
\times h(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})f(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})\;\frac{f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}) h(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})
g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B})}
{h(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})}.
\end{multline}
Here the sum is taken over partitions $\bar{u}^{\scriptscriptstyle B}\Rightarrow\{\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}\}$ and $\bar{v}^{\scriptscriptstyle C}\Rightarrow\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$
with $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=n$ and $\#\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=a-n$.
This is the second representation for the highest coefficient discussed above. It will play the key role below, therefore we formulate
it as a proposition.
\begin{prop}\label{non-tr-prop}
For arbitrary sets of complex numbers $\bar t$, $\bar x$, $\bar s$, and $\bar y$ with cardinalities
$\#\bar t=\#\bar x=a$ and $\#\bar s=\#\bar y=b$ the following identity holds:
\begin{multline}\label{Ident-0-main}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)= f(\bar x,\bar t)f(\bar y,\bar x)\sum (-1)^{n_1(a+b)}
J_{\ell_2,n_1}(\bar x_{\scriptscriptstyle \rm I\hspace{-1pt}I};\bar s_{\scriptscriptstyle \rm I}|\bar t+c;\bar x+c|\bar t-c)\\
\times h(\bar x_{\scriptscriptstyle \rm I},\bar x_{\scriptscriptstyle \rm I})f(\bar x_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x_{\scriptscriptstyle \rm I})\;\frac{f(\bar s_{\scriptscriptstyle \rm I},\bar t) h(\bar s_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x_{\scriptscriptstyle \rm I})
g(\bar s_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar s_{\scriptscriptstyle \rm I})g(\bar s_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar y)}
{h(\bar y,\bar x_{\scriptscriptstyle \rm I})g(\bar x_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar t)}.
\end{multline}
Here $\ell_2=\#\bar x_{\scriptscriptstyle \rm I\hspace{-1pt}I}$, $n_1=\#\bar s_{\scriptscriptstyle \rm I}$. The sum is taken over partitions
\be{part-0}
\bar x\Rightarrow\{\bar x_{\scriptscriptstyle \rm I},\bar x_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}, \qquad \bar s\Rightarrow\{\bar s_{\scriptscriptstyle \rm I},\bar s_{\scriptscriptstyle \rm I\hspace{-1pt}I}\},
\end{equation}
with a restriction $\#\bar s_{\scriptscriptstyle \rm I}=\#\bar x_{\scriptscriptstyle \rm I}$ (which is equivalent to $\ell_2+n_1=a$).
\end{prop}
{\sl Proof}. Setting in \eqref{Ident-1-main} $\bar{u}^{\scriptscriptstyle C}=\bar t$, $\bar{u}^{\scriptscriptstyle B}=\bar x$, $\bar{v}^{\scriptscriptstyle C}=\bar s$, and $\bar{v}^{\scriptscriptstyle B}=\bar y$ we obtain
\eqref{Ident-0-main}.
\subsection{General formula for the scalar product}
Now we turn back to equation \eqref{SP-expl3}. To proceed further we should specify all the subsets. Let
$\bar{v}^{\scriptscriptstyle C}=\{\bar{v}^{\scriptscriptstyle C}_{\rm i},\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle C}_{\rm iii}\}$ and $\bar{v}^{\scriptscriptstyle B}=\{\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle C}_{\rm ii}\}$. We set
\be{part-v}
\begin{aligned}
& \bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle C}_{\rm iii},\qquad & \bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\{\bar{v}^{\scriptscriptstyle C}_{\rm i},\bar{v}^{\scriptscriptstyle C}_{\rm ii}\}, \\
& \bar{v}^{\scriptscriptstyle B}=\{\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle B}_{\rm ii}\},\qquad &{} \\
&\bar\xi_{\scriptscriptstyle \rm I}=\{\bar{v}^{\scriptscriptstyle C}_{\rm i},\bar{v}^{\scriptscriptstyle B}_{\rm ii}\},\qquad &\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=\{\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle B}_{\rm i}\},
\end{aligned}
\mb{ with }
\begin{cases}
\#\bar{v}^{\scriptscriptstyle C}_s=n_s,\\
\#\bar{v}^{\scriptscriptstyle B}_s=m_s,
\end{cases}
\qquad s=\rm i, \rm ii,\rm iii.
\end{equation}
It is easy to see that the
following conditions for the cardinalities hold:
\be{restr-v}
n_{\rm iii}=n,\quad n_{\rm i}+n_{\rm ii}=b-n, \quad m_{\rm i}+m_{\rm ii}=b, \quad m_{\rm i}=n_{\rm i}.
\end{equation}
Let also
\be{part-u}
\begin{aligned}
& \bar{u}^{\scriptscriptstyle C}=\{\bar{u}^{\scriptscriptstyle C}_{\rm i},\bar{u}^{\scriptscriptstyle C}_0\},\qquad & \bar{u}^{\scriptscriptstyle C}_{0}=\{\bar{u}^{\scriptscriptstyle C}_{\rm ii},\bar{u}^{\scriptscriptstyle C}_{\rm iii}\},\\
& \bar{u}^{\scriptscriptstyle B}=\{\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle B}_0\},\qquad & \bar{u}^{\scriptscriptstyle B}_{\rm i}=\{\bar{u}^{\scriptscriptstyle B}_{\rm ii},\bar{u}^{\scriptscriptstyle B}_{\rm iii}\}, \\
& \bar\eta_{\scriptscriptstyle \rm I}=\{\bar{u}^{\scriptscriptstyle C}_{\rm i},\bar{u}^{\scriptscriptstyle B}_0\},\qquad & \bar\eta_{0}=\{\bar{u}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle B}_{\rm i}\},%
\end{aligned}
\mb{ with }
\begin{cases} \#\bar{u}^{\scriptscriptstyle C}_p=k_p, \\ \#\bar{u}^{\scriptscriptstyle B}_p=\ell_p, \end{cases}
\qquad p=0,\rm i, \rm ii,\rm iii.
\end{equation}
We have the following conditions for the cardinalities:
\be{restr-u}
k_{\rm i}+k_0=\ell_{\rm i}+\ell_0=a,\quad k_{\rm i}=\ell_{\rm i}, \quad k_0=\ell_0
\end{equation}
Observe that we do not fix a distribution of the parameters $\bar{u}^{\scriptscriptstyle C}$ and $\bar{u}^{\scriptscriptstyle B}$ among the subsets $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}$
and $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}$. It is important, however, that $\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n=n_{\rm iii}$
and $\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}=a-n$.
Using \eqref{limW-1} and \eqref{limW-2} we obtain
\begin{multline}\label{red-W}
\frac{J_{a,b-n}(\bar\eta_{\scriptscriptstyle \rm I};\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}| \bar{u}^{\scriptscriptstyle C}; \bar\eta_{\scriptscriptstyle \rm I}|\{\bar{u}^{\scriptscriptstyle C},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\})}
{g(\bar{u}^{\scriptscriptstyle C},\bar\eta_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I} ,\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }=(-1)^{b-n}
\frac{g(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{u}^{\scriptscriptstyle C})g(\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle C}_{\rm i}) h(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{u}^{\scriptscriptstyle C}_0)}
{g(\bar{u}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle B}_0)g(\bar{v}^{\scriptscriptstyle B}_{\rm i} ,\bar{v}^{\scriptscriptstyle C}_{\rm i}) h(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{u}^{\scriptscriptstyle B}_0) }\\
\times J_{\ell_0,n_{\rm i}}(\bar{u}^{\scriptscriptstyle B}_0;\bar{v}^{\scriptscriptstyle C}_{\rm i}| \bar{u}^{\scriptscriptstyle C}_0; \bar{u}^{\scriptscriptstyle B}_0|\{\bar{u}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm i}\}).
\end{multline}
Due to \eqref{Z-J} this function reduces to the highest coefficient
\be{W-Z1}
J_{\ell_0,n_{\rm i}}(\bar{u}^{\scriptscriptstyle B}_0;\bar{v}^{\scriptscriptstyle C}_{\rm i}| \bar{u}^{\scriptscriptstyle C}_0; \bar{u}^{\scriptscriptstyle B}_0|\{\bar{u}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm i}\})=
\frac{Z_{\ell_0,n_{\rm i}}(\bar{u}^{\scriptscriptstyle B}_0,\bar{u}^{\scriptscriptstyle C}_0|\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle C}_{\rm i})}
{h(\bar{u}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle B}_0)h(\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle B}_0)}.
\end{equation}
Now we substitute \eqref{red-W}, \eqref{W-Z1} into \eqref{SP-expl3}. We also write explicitly the products $g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})$,
$g(\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\xi_{\scriptscriptstyle \rm I})$, $f(\bar\eta_0,\bar\eta_{\scriptscriptstyle \rm I})$, and combine $\{\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle C}_{\rm iii}\}=\bar{v}^{\scriptscriptstyle C}_0$. Then we have
\begin{multline}\label{SP-4}
S_{a,b}=\frac{1}{ f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})}\sum
r_1(\bar{u}^{\scriptscriptstyle C}_{\rm i})r_1(\bar{u}^{\scriptscriptstyle B}_0)r_3(\bar{v}^{\scriptscriptstyle C}_{\rm i})r_3(\bar{v}^{\scriptscriptstyle B}_{\rm ii}) \;Z_{\ell_0,n_{\rm i}}(\bar{u}^{\scriptscriptstyle B}_0,\bar{u}^{\scriptscriptstyle C}_0|\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle C}_{\rm i})
\frac{f(\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle C}_{\rm i})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar{u}^{\scriptscriptstyle C})}
{f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B}_{0})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar{u}^{\scriptscriptstyle C}_{\rm i})}\\
%
\times f(\bar{u}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle C}_{\rm i})f(\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle B}_0)g(\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle C}_{\rm i})g(\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle B}_{\rm ii})\quad
\Bigl\{ J_{a-n,n}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\rm iii}|\bar{u}^{\scriptscriptstyle C}+c;\bar\eta_{0}+c|\bar{u}^{\scriptscriptstyle C}-c) \\\rule{0pt}{20pt}
\times (-1)^{n(a+b-n_{\rm i})} h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
\frac{f(\bar{v}^{\scriptscriptstyle C}_{\rm iii},\bar{u}^{\scriptscriptstyle C})g(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle C}_{\rm iii})g(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle B}_{\rm ii})h(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }
{ g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})h(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }\Bigr\}.
\end{multline}
Here the sum is organized as follows. First, we have partitions
\be{Afin-part}
\begin{aligned}
&\bar{u}^{\scriptscriptstyle C}\Rightarrow\{\bar{u}^{\scriptscriptstyle C}_{\rm i},\bar{u}^{\scriptscriptstyle C}_0\},\qquad &\bar{u}^{\scriptscriptstyle B}\Rightarrow\{\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle B}_0\},\qquad &\#\bar{u}^{\scriptscriptstyle C}_0=\#\bar{u}^{\scriptscriptstyle B}_0=\ell_0,\\
&\bar{v}^{\scriptscriptstyle C}\Rightarrow\{\bar{v}^{\scriptscriptstyle C}_{\rm i},\bar{v}^{\scriptscriptstyle C}_0\},\qquad &\bar{v}^{\scriptscriptstyle B}\Rightarrow\{\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle B}_{\rm ii}\}, \qquad &\#\bar{v}^{\scriptscriptstyle C}_{\rm i}=\#\bar{v}^{\scriptscriptstyle B}_{\rm i}=n_{\rm i}.
\end{aligned}
\end{equation}
After this we have two additional partitions: the set $\bar{v}^{\scriptscriptstyle C}_0$ is divided into subsets $\bar{v}^{\scriptscriptstyle C}_{\rm ii}$ and $\bar{v}^{\scriptscriptstyle C}_{\rm iii}$;
the union of the subsets $\bar\eta_0=\{\bar{u}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle B}_{\rm i}\}$ is divided into subsets $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}$ and $\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}$ (see the terms in braces in \eqref{SP-4}).
Hereby we have one restriction for the cardinalities $\#\bar{v}^{\scriptscriptstyle C}_{\rm iii}=\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=n=n_{\rm iii}$. Let us write separately this additional sum over partitions
in braces of \eqref{SP-4}
\begin{multline}\label{AddSum}
\mathcal{F}(\bar{u}^{\scriptscriptstyle C};\bar\eta_0;\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii})= \sum_{\substack{\bar\eta_0 \Rightarrow \{\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I}\}\\
\bar{v}^{\scriptscriptstyle C}_0 \Rightarrow \{\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle C}_{\rm iii}\} } }
J_{a-n,n}(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I};\bar{v}^{\scriptscriptstyle C}_{\rm iii}|\bar{u}^{\scriptscriptstyle C}+c;\bar\eta_{0}+c|\bar{u}^{\scriptscriptstyle C}-c) \\\rule{0pt}{20pt}
\times (-1)^{n(a+b-n_{\rm i})} h(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})
\frac{f(\bar{v}^{\scriptscriptstyle C}_{\rm iii},\bar{u}^{\scriptscriptstyle C})g(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle C}_{\rm iii})g(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar{v}^{\scriptscriptstyle B}_{\rm ii})h(\bar{v}^{\scriptscriptstyle C}_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }
{ g(\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle C})h(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }.
\end{multline}
Comparing \eqref{AddSum} with \eqref{Ident-0-main} one can see that they coincide after appropriate identification of
the subsets and their cardinalities. Namely, \eqref{AddSum} turns into \eqref{Ident-0-main} under the
replacements $b-n_{\rm i} \to b$, $\bar{u}^{\scriptscriptstyle C}\to\bar t$, $\bar\eta_0\to\bar x$, $\bar{v}^{\scriptscriptstyle C}_0\to\bar s$, $\bar{v}^{\scriptscriptstyle B}_{\rm ii}\to\bar y$. Thus, due to Proposition~\ref{non-tr-prop}
we obtain
\begin{equation}\label{sum-nontr}
\mathcal{F}(\bar{u}^{\scriptscriptstyle C};\bar\eta_0;\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii})=\frac{Z_{a,b-n_{\rm i}}(\bar{u}^{\scriptscriptstyle C};\bar\eta_0|\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii})}
{f(\bar\eta_0,\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar\eta_0)}.
\end{equation}
Thus, substituting this into \eqref{SP-4} we arrive at
\begin{multline}\label{SP-5}
S_{a,b}=\frac{1}{ f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})}\sum
r_1(\bar{u}^{\scriptscriptstyle C}_{\rm i})r_1(\bar{u}^{\scriptscriptstyle B}_0)r_3(\bar{v}^{\scriptscriptstyle C}_{\rm i})r_3(\bar{v}^{\scriptscriptstyle B}_{\rm ii}) \;Z_{\ell_0,n_{\rm i}}(\bar{u}^{\scriptscriptstyle B}_0,\bar{u}^{\scriptscriptstyle C}_0|\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle C}_{\rm i})
\frac{f(\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle C}_{\rm i})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar{u}^{\scriptscriptstyle C})}
{f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B}_{0})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar{u}^{\scriptscriptstyle C}_{\rm i})}\\\rule{0pt}{20pt}
%
\times f(\bar{u}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle C}_{\rm i})f(\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle B}_0)g(\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle C}_{\rm i})g(\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle B}_{\rm ii})\,
\frac{Z_{a,b-n_{\rm i}}(\bar{u}^{\scriptscriptstyle C};\bar\eta_0|\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii})}
{f(\bar\eta_0,\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar\eta_0)}.
\end{multline}
It remains to simplify the ratio $Z_{a,b-n_{\rm i}}(\bar{u}^{\scriptscriptstyle C};\bar\eta_0|\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii})/f(\bar\eta_0,\bar{u}^{\scriptscriptstyle C})$. It can be done via
\eqref{Z-J}, \eqref{limW-1}:
\be{red-Z2}
\frac{Z_{a,b-n_{\rm i}}(\bar{u}^{\scriptscriptstyle C};\bar\eta_0|\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii})}
{f(\bar\eta_0,\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar\eta_0)}=\frac{f(\bar{v}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle C}_0)Z_{a-\ell_0,b-n_{\rm i}}(\bar{u}^{\scriptscriptstyle C}_{\rm i};\bar{u}^{\scriptscriptstyle B}_{\rm i}|\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii})}
{f(\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle C}_{\rm i})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar{u}^{\scriptscriptstyle B}_{\rm i})f(\bar{v}^{\scriptscriptstyle B}_{\rm ii},\bar{u}^{\scriptscriptstyle C}_0)}.
\end{equation}
Substituting this into \eqref{SP-5} we obtain
\begin{multline}\label{Sab-exp}
S_{a,b}=
\sum r_1(\bar{u}^{\scriptscriptstyle C}_{\rm i})r_1(\bar{u}^{\scriptscriptstyle B}_0)r_3(\bar{v}^{\scriptscriptstyle C}_{\rm i})r_3(\bar{v}^{\scriptscriptstyle B}_{\rm ii})f(\bar{u}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle C}_{\rm i})f(\bar{u}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle B}_0)g(\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle C}_{\rm i})g(\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle B}_{\rm ii})\\
\times \frac{f(\bar{v}^{\scriptscriptstyle C}_0,\bar{u}^{\scriptscriptstyle C}_0)f(\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{u}^{\scriptscriptstyle B}_{\rm i})}{ f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})}\;
Z_{\ell_0,n_{\rm i}}(\bar{u}^{\scriptscriptstyle B}_0,\bar{u}^{\scriptscriptstyle C}_0|\bar{v}^{\scriptscriptstyle B}_{\rm i},\bar{v}^{\scriptscriptstyle C}_{\rm i})\; Z_{a-\ell_0,b-n_{\rm i}}(\bar{u}^{\scriptscriptstyle C}_{\rm i},\bar{u}^{\scriptscriptstyle B}_{\rm i}|\bar{v}^{\scriptscriptstyle C}_0,\bar{v}^{\scriptscriptstyle B}_{\rm ii}).
\end{multline}
It is easy to see that after appropriate relabeling the subsets we arrive at
\begin{multline}\label{Sab-fin}
S_{a,b}=
\sum r_1(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})r_1(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})r_3(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})r_3(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})f(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})f(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})\\
\times \frac{f(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I})f(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I})}{ f(\bar{v}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle C})f(\bar{v}^{\scriptscriptstyle B},\bar{u}^{\scriptscriptstyle B})}\;
Z_{a-k,n}(\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}|\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})\;Z_{k,b-n}(\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}|\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}),
\end{multline}
where $k=\#\bar{u}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{u}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}$ and $n=\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\#\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}$.
Comparing this result with \eqref{Sab-genform2} and \eqref{W-Reshet} we see that proposition~\ref{W-2lin} is proved.
\section{Scalar product in the $\mathfrak{gl}(1|1)$ sector\label{S-SP11}}
Consider a particular case of the subalgebra $\mathfrak{gl}(1|1)$, generated by the operators $T_{23}(u)$, $T_{22}(u)$,
$T_{33}(u)$ and $T_{32}(u)$. In this case one should set $\bar{u}^{\scriptscriptstyle C}=\bar{u}^{\scriptscriptstyle B}=\emptyset$ in
the formulas for the scalar product. Then the highest coefficient simplifies as
\be{Z-det11}
Z_{0,b}(\emptyset;\emptyset|\bar s;\bar y)=\Delta_{b}(\bar s)\Delta'_b(\bar y)\;
\det_{b}g(s_j,y_k)=g(\bar s,\bar y),
\end{equation}
where we used an explicit representation for Cauchy determinant
\be{Cauchy-0}
\det_m g(u_j,v_k)=\frac{g(\bar u,\bar v)}{\Delta'_m(\bar u)\Delta_m(\bar v)}.
\end{equation}
The scalar product \eqref{Sab-fin} takes the form
\begin{equation}\label{Sa0-fin}
S_{0,b}=
\sum r_3(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})r_3(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I})g(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})
g(\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I})g(\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}),
\end{equation}
where the sum is taken over partitions $\bar{v}^{\scriptscriptstyle C}\Rightarrow\{\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ and $\bar{v}^{\scriptscriptstyle B}\Rightarrow\{\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I},\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$
such that $\#\bar{v}^{\scriptscriptstyle C}_{\scriptscriptstyle \rm I}=\bar{v}^{\scriptscriptstyle B}_{\scriptscriptstyle \rm I}$. It is easy to see that this sum reduces to a single determinant
\begin{equation}\label{Sa0-det}
S_{0,b}=\Delta'_b(\bar{v}^{\scriptscriptstyle C})\Delta_b(\bar{v}^{\scriptscriptstyle B})\det_b\Bigl[g(v^{\scriptscriptstyle C}_j,v^{\scriptscriptstyle B}_k)\bigl(r_3(v^{\scriptscriptstyle B}_k)-r_3(v^{\scriptscriptstyle C}_j)\bigr)\Bigr].
\end{equation}
Indeed, developing the determinant in \eqref{Sa0-det} via Laplace formula and using \eqref{Cauchy-0}, \eqref{Del-sig} we
obtain the sum \eqref{Sa0-fin}.
Thus, the scalar product of Bethe vectors in $\mathfrak{gl}(1|1)$ integrable models admits a determinant representation
without any restriction on the Bethe parameters. This is not surprising, as these models are equivalent to free fermions
\cite{FoeK93,GohM98}.
\section{Different representations for the highest coefficient\label{S-DRHC}}
If $\bar{v}^{\scriptscriptstyle C}=\bar{v}^{\scriptscriptstyle B}=\emptyset$, then formula \eqref{Sab-fin} describes the scalar product in the $\mathfrak{gl}(2)$-based models.
In this case the scalar product admits a determinant representation, if one of the Bethe vectors is an eigenvector of the transfer matrix.
One expects that in the general $\mathfrak{gl}(2|1)$ case the sum over partitions in \eqref{Sab-fin} also can be reduced to a single
determinant for some particular cases of Bethe vectors. To make this reduction
one should have different representations for the highest coefficient
$ Z_{a,b}(\bar t;\bar x|\bar s;\bar y)$. In this section we give several formulas for $Z_{a,b}$ in terms of sums over partitions and multiple
contour integrals.
We have already obtained an expression for $Z_{a,b}$ as the determinant of an $(a+b)\times(a+b)$ matrix
\be{Z-detJ}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=h(\bar w,\bar t)\Delta_{a+b}(\bar w)\Delta'_a(\bar t)\Delta'_b(\bar y)\;
\det_{a+b}\mathcal{J}_{jk},
\end{equation}
where $\bar w=\{\bar x,\bar s\}$ and the matrix $\mathcal{J}_{jk}$ is defined in \eqref{matelJ}:
\be{matelJ1}
\begin{aligned}
&\mathcal{J}_{jk}=\frac{g(w_j,t_k)}{h(w_j,t_k)},&\qquad k=1,\dots,a;\\
&\mathcal{J}_{j,k+a}=g(w_j,y_k)\frac{h(w_j,\bar x)}{h(w_j,\bar t)},&\qquad k=1,\dots,b;
\end{aligned}
\qquad j=1,\dots,a+b.
\end{equation}
Developing the determinant with respect to the $a$ first columns we
obtain
\be{RHC-IHC}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=\sum K_{a}(\bar w_{\scriptscriptstyle \rm I}|\bar t) h(\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) g(\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar y) g(\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar w_{\scriptscriptstyle \rm I}).
\end{equation}
%
The sum is taken over partitions $\{\bar x,\bar s\}=\bar w\Rightarrow\{\bar w_{\scriptscriptstyle \rm I},\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ with $\#\bar w_{\scriptscriptstyle \rm I}=a$ and $\#\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I}=b$.
Let us give several alternative representations for the highest coefficient.
\begin{itemize}
\item As a sum over partitions of $\bar t$ and $\bar y$:
\be{Al-RHC-IHC}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=f(\bar s,\bar t)f(\bar y,\bar x)
\sum g(\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}) \frac{h(\bar t,\bar \eta_{\scriptscriptstyle \rm I\hspace{-1pt}I})}{h(\bar s,\bar \eta_{\scriptscriptstyle \rm I\hspace{-1pt}I} )} K_a(\bar x|\bar\eta_{\scriptscriptstyle \rm I}).
\end{equation}
Here the sum is taken over partitions $\{\bar t,\bar y+c\}=\bar\eta\Rightarrow\{\bar\eta_{\scriptscriptstyle \rm I},\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ such that $\#\bar\eta_{\scriptscriptstyle \rm I}=a$ and $\#\bar\eta_{\scriptscriptstyle \rm I\hspace{-1pt}I}=b$.
\item As a sum over partitions of $\bar t$ and $\bar x$:
%
\begin{multline}\label{GF1}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=
(-1)^{a} h(\bar x, \bar x) h(\bar s, \bar x) g(\bar x,\bar y) g(\bar s,\bar y) \\
%
\sum K_{a}(\bar t - c |\bar \xi_{\scriptscriptstyle \rm I}) \frac{h(\bar \xi_{\scriptscriptstyle \rm I}, \bar t)
g(\bar x, \bar \xi_{\scriptscriptstyle \rm I}) g(\bar s, \bar \xi_{\scriptscriptstyle \rm I})}{g(\bar \xi_{\scriptscriptstyle \rm I}, \bar y)} g(\bar \xi_{\scriptscriptstyle \rm I},\bar \xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}).
\end{multline}
Here the sum is taken over partitions $\{\bar t,\bar x-c\}=\bar\xi\Rightarrow\{\bar\xi_{\scriptscriptstyle \rm I},\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ such that
$\#\bar\xi_{\scriptscriptstyle \rm I}=\#\bar\xi_{\scriptscriptstyle \rm I\hspace{-1pt}I}=a$.
\item As a sum over partitions of $\bar s$ and $\bar y$:
\be{S-GF1}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=(-1)^{a+b} f(\bar x,\bar t) f(\bar s,\bar t) \sum g(\bar \nu_{\scriptscriptstyle \rm I},\bar \nu_{\scriptscriptstyle \rm I\hspace{-1pt}I})
K_{a+b}(\{\bar \nu_{\scriptscriptstyle \rm I}, \bar t - c\}|\{\bar x,\bar s\})
\end{equation}
Here the sum is taken over partitions $\{\bar s-c,\bar y\}=\bar\nu\Rightarrow\{\bar\nu_{\scriptscriptstyle \rm I},\bar\nu_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ such that
$\#\bar\nu_{\scriptscriptstyle \rm I}=\#\bar\nu_{\scriptscriptstyle \rm I\hspace{-1pt}I}=b$.
\end{itemize}
All the sum formulas listed above follow from \eqref{RHC-IHC} and can be proved via reduction of the sums over partitions to multiple contour
integrals of Cauchy type. Let us show how this method works.
Consider a $b$-fold integral
\be{Int-Or-for}
\mathcal{I}=\frac{(-1)^{b}}{(2\pi ic)^b b!}\oint\limits_{\bar w} K_{a+b}(\bar w|\{\bar t, \bar z + c\})
h(\bar z, \bar x) \frac{ g(\bar z, \bar y) g(\bar z, \bar w)}{\Delta_b(\bar z) \Delta'_b(\bar z)}
\,d\bar z,
\end{equation}
%
where $\bar w=\{\bar s,\bar x\}$ and $d\bar z=dz_1,\dots,dz_b$. We have used a subscript $\bar w$ on the integral symbol in order
to stress that the integration contour for every $z_j$ surrounds the set $\bar w$ in
the anticlockwise direction. We also assume that the integration contours do not contain any
other singularities of the integrand. Similar prescription will be kept for all other integral representations
considered below.
The only poles of
the integrand within the integration contours are the points $z_j=w_k$. Evaluating the integral
by the residues in these poles we obtain (see appendix~\ref{A-SumInt} for details)
\be{Int-Or-for-twin-res}
\mathcal{I}=(-1)^b\sum K_{a+b}(\bar w|\{\bar t,\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c\}) h(\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar x) g(\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar y) g(\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I},\bar w_{\scriptscriptstyle \rm I}),
\end{equation}
%
where the sum is taken over partitions of $\bar w$ into subsets $\bar w_{\scriptscriptstyle \rm I}$ and $\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I}$ with
$\#\bar w_{\scriptscriptstyle \rm I}=a$ and $\#\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I}=b$. Due to
\eqref{K-K} we have
\be{red-Kab}
K_{a+b}(\bar w|\{\bar t,\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c\})=(-1)^b K_{a}(\bar w_{\scriptscriptstyle \rm I}|\bar t),
\end{equation}
and comparing the obtained sum with \eqref{RHC-IHC} we immediately obtain $\mathcal{I}=Z_{a,b}(\bar t;\bar x|\bar s;\bar y)$.
Similarly, one can check that the sum over partitions in \eqref{RHC-IHC} can be presented as an
$a$-fold contour integral
\be{Int-Or-for-twin}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=(-1)^b\frac{h(\bar w,\bar x)g(\bar y,\bar w)}{(2\pi ic)^a a!}\oint\limits_{\bar w}
\frac{K_a(\bar z|\bar t) g(\bar z,\bar w) }{h(\bar z,\bar x)g(\bar z, \bar y)\Delta_a(\bar z) \Delta'_a(\bar z)}
\,d\bar z,
\end{equation}
%
where now $d\bar z=dz_1,\dots,dz_a$. Indeed, taking the residues in the points $\bar z=\bar w_{\scriptscriptstyle \rm I}$ we obtain
\be{Int-AOr-for-twin}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=(-1)^b h(\bar w,\bar x)g(\bar y,\bar w)\sum
\frac{K_a(\bar w_{\scriptscriptstyle \rm I}|\bar t) g(\bar w_{\scriptscriptstyle \rm I},\bar w_{\scriptscriptstyle \rm I\hspace{-1pt}I}) }{h(\bar w_{\scriptscriptstyle \rm I},\bar x)g(\bar w_{\scriptscriptstyle \rm I}, \bar y)}.
\end{equation}
%
Multiplying the terms of the sum with the prefactor $h(\bar w,\bar x)g(\bar y,\bar w)$ we arrive at \eqref{RHC-IHC}.
Let us turn back to the integral \eqref{Int-Or-for}. Obviously, it can be calculated taking the residues in
the poles outside the original integration contour. It
is easy to see that for arbitrary $z_j$ the integrand behaves as $1/z_j^3$ at $z_j\to\infty$. Hence, the
residue at infinity vanishes. The poles outside the original integration contours are in $z_j=y_k$ and $z_j=s_k-c$
(the poles at $z_j=x_k-c$ are compensated by the zeros of the product $h(\bar z, \bar x)$). Thus, we can move the original
contour surrounding $\bar w$ to the points $\bar \nu=\{\bar y,\bar s-c\}$
\be{Int-Or-for-twin1}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=\frac{1}{(2\pi ic)^b b!}\oint\limits_{\bar \nu} K_{a+b}(\bar w|\{\bar t, \bar z + c\})
h(\bar z, \bar x)\frac{ g(\bar z, \bar y) g(\bar z, \bar w)}{\Delta_b(\bar z) \Delta'_b(\bar z)}
\,d\bar z.
\end{equation}
%
It is convenient to transform the integrand, applying \eqref{Red-K} to
$K_{a+b}(\bar w|\{\bar t, \bar z + c\})$. Then substituting $\bar w = \{\bar x, \bar s\}$ and using elementary properties of $f(z,w)$ we obtain
\be{Int-Or-for-twin2}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=\frac{(-1)^{a+b}}{(2\pi ic)^b b!}\oint\limits_{\bar \nu}
\frac{K_{a+b}(\{\bar t - c, \bar z\}| \bar w)f(\bar w,\bar t) h(\bar z, \bar x)g(\bar z, \bar y) g(\bar z, \bar x) g(\bar z, \bar s)}
{f(\bar z, \bar x) f(\bar z,\bar s)\Delta_b(\bar z) \Delta'_b(\bar z)}
\,d\bar z,
\end{equation}
%
and after simplification we arrive at
\be{Int-Or-for-twin3}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=\frac{(-1)^{a+b} f(\bar w,\bar t)}{(2\pi ic)^b b!}\oint\limits_{\bar \nu}K_{a+b}(\{\bar t - c, \bar z\}| \bar w)
\frac{g(\bar z, \bar \nu)}{\Delta_b(\bar z) \Delta'_b(\bar z)}
\,d\bar z.
\end{equation}
Now all the poles are explicitly combined in the product $g(\bar z,\bar\nu)$. Hence, the
result of the integration gives the sum over partitions of $\bar\nu\Rightarrow\{\bar\nu_{\scriptscriptstyle \rm I},\bar\nu_{\scriptscriptstyle \rm I\hspace{-1pt}I}\}$ with
$\#\bar\nu_{\scriptscriptstyle \rm I}= \# \bar\nu_{\scriptscriptstyle \rm I\hspace{-1pt}I}=b$, which coincides with \eqref{S-GF1}.
Applying \eqref{K-K} to the DWPF $ K_{a+b}(\{\bar \nu_{\scriptscriptstyle \rm I}, \bar t - c\}|\{\bar x,\bar s\})$ in \eqref{S-GF1}, we have
\begin{multline}
K_{a+b}(\{\bar \nu_{\scriptscriptstyle \rm I}, \bar t - c\}|\{\bar x,\bar s\})= (-1)^bK_{a+2b}(\{\bar \nu, \bar t - c\}|\{\bar x,\bar s,\bar \nu_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c\})\\
=(-1)^bK_{a+2b}(\{\bar y, \bar s-c,\bar t - c\}|\{\bar x,\bar s,\bar \nu_{\scriptscriptstyle \rm I\hspace{-1pt}I}+c\})
= K_{a+b}(\{\bar y, \bar t - c\}|\{\bar x,\bar \nu_{\scriptscriptstyle \rm I\hspace{-1pt}I} + c\}).
\end{multline}
Then the sum over partitions in \eqref{S-GF1} is equivalent to a multiple contour integral
\be{Int-Or-for-twin4}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=\frac{(-1)^{a} f(\bar x,\bar t)f(\bar s,\bar t)}{(2\pi ic)^b b!}\oint\limits_{\bar \nu}
K_{a+b}(\{\bar y,\bar t - c\}| \{\bar x, \bar z+c\})
\frac{g(\bar z, \bar \nu)}{\Delta_b(\bar z) \Delta'_b(\bar z)}
\,d\bar z.
\end{equation}
Using \eqref{Red-K} we recast \eqref{Int-Or-for-twin4} as
\be{Int-Or-for-twin5}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=\frac{(-1)^{b}f(\bar y,\bar x) f(\bar s,\bar t)}{(2\pi ic)^b b!}\oint\limits_{\bar \nu}
\frac{K_{a+b}(\{\bar x-c, \bar z\}|\{\bar y,\bar t - c\})}{f(\bar z,\bar y)f(\bar z,\bar t-c)}
\frac{g(\bar z, \bar \nu)}{\Delta_b(\bar z) \Delta'_b(\bar z)}
\,d\bar z.
\end{equation}
Setting now $\bar\eta=\{\bar t,\bar y+c\}$ we obtain
\be{Int-Or-for-twin6}
Z_{a,b}(\bar t;\bar x|\bar s;\bar y)=\frac{(-1)^{b}f(\bar y,\bar x) f(\bar s,\bar t)}{(2\pi ic)^b b!}\oint\limits_{\bar \nu}
K_{a+b}(\{\bar x-c, \bar z\}|\bar\eta-c)\frac{h(\bar z,\bar t)g(\bar z, \bar\eta-2c)}{h(\bar z,\bar s)\Delta_b(\bar z) \Delta'_b(\bar z)}
\,d\bar z.
\end{equation}
Now we can evaluate the integral by the residues outside the integration contours. All of them are collected in the
product $g(\bar z, \bar\eta-2c)$. Taking the residues in the points $\bar\eta-2c$ and using $h(x-2c,y)=-h(y,x)$ we immediately
arrive at \eqref{Al-RHC-IHC}.
Similarly, starting with the integral representation \eqref{Int-Or-for-twin} one can obtain the sum formula \eqref{GF1}.
\section*{Conclusion}
In this paper we have derived a sum formula for the scalar product of Bethe vectors in the models with $\mathfrak{gl}(2|1)$
symmetry. We considered the case of generic Bethe vectors. This means that the Bethe parameters are generic complex numbers.
However, one can also use this sum formula, if the Bethe parameters obey some constraints. In some of these particular cases the
sums over partitions can be taken explicitly, leading eventually to determinant representations for scalar products. In the second
part of this paper we will consider these particular cases in more details. We will show that if a part of Bethe parameters satisfy
Bethe equations, then the sum formula in the $\mathfrak{gl}(2|1)$-based models reduces to a single determinant.
To conclude this paper we would like to mention that our results can be applied to the models with $\mathfrak{gl}(1|2)$
symmetry as well. Indeed, due to an isomorphism between Yangians $Y\bigl(\mathfrak{gl}(1|2)\bigr)$ and
$Y\bigl(\mathfrak{gl}(2|1)\bigr)$ (see \cite{PakRS16a}), it is enough to make the replacements $\{\bar{u}^{\scriptscriptstyle C},\bar{u}^{\scriptscriptstyle B}\}
\leftrightarrow \{\bar{v}^{\scriptscriptstyle C},\bar{v}^{\scriptscriptstyle B}\}$ and $a\leftrightarrow b$ in the sum formula. The obtained expression describes the
scalar product of Bethe vectors in $\mathfrak{gl}(1|2)$-based models.
\section*{Acknowledgements}
The work of A.L. has been funded by the Russian Academic Excellence Project 5-100 and by joint NASU-CNRS project F14-2016.
The work of S.P. was supported in part by the RFBR grant 14-01-00474 and the grant
of the Scientific Foundation of NRU HSE.
N.A.S. was supported by the grants RFBR-15-31-20484-mol-a-ved and RFBR-14-01-00860-a.
|
\section{Introduction}
An overpartition \cite{lcc} is a partition in which the first occurrence of each distinct number may be overlined. For example, the $14$ overpartitions of $4$ are
\begin{equation*}
\begin{gathered}
4, \overline{4}, 3+1, \overline{3} + 1, 3 + \overline{1},
\overline{3} + \overline{1}, 2+2,\\ \overline{2}
+ 2, 2+1+1, \overline{2} + 1 + 1, 2+ \overline{1} + 1, \\
\overline{2} + \overline{1} + 1, 1+1+1+1, \overline{1} + 1 + 1 +1.
\end{gathered}
\end{equation*}
Let $\overline{N}(n,m)$ denote the number of overpartitions of $n$ whose rank is $m$ and $\overline{M}(n,m)$ the number of overpartitions of $n$ whose first residual crank is $m$. Andrews, Chan, Kim and Osburn \cite{acko} defined the positive rank and crank moments for overpatitions:
\begin{equation*}
\overline{N}_{k}^{+}(n) := \sum_{m=1}^{\infty} m^{k} \overline{N}(m,n)
\end{equation*}
\noindent and
\begin{equation*}
\overline{M}_{k}^{+}(n) := \sum_{m=1}^{\infty} m^{k} \overline{M}(m,n).
\end{equation*}
\noindent They proved the inequality
\begin{equation} \label{oddrcover}
\overline{M}_{1}^{+}(n) > \overline{N}_{1}^{+}(n)
\end{equation}
holds for all $n \geq 1$. They also gave the difference
$ \overline{M}_{1}^{+}(n) - \overline{N}_{1}^{+}(n)$ a combinatorial interpretation. In order to prove (\ref{oddrcover}), Andrews, Chan, Kim and Osburn \cite{acko} introduced the function
\begin{equation*} \label{hqdefn}
h(q) := \sum_{n=1}^{\infty} \frac{(-1)^{n+1} q^{n(n+1)/2}}{1-q^n},
\end{equation*}
and conjectured that
$$\label{con}
\frac{1}{(q)_{\infty}} ( h(q) - m h(q^m) )
$$
has positive coefficients for all $m\geq 3$, where we use the standard $q$-series notation, $(a)_{\infty} = (a;q)_{\infty} = \prod_{n=1}^{\infty} (1-aq^{n-1} ).$ In \cite{kks}, Byungchan Kim, Eunmi Kim and Jeehyeon Seo provided a combinatorial interpretation for the coefficients of $ \frac{1}{(q)_{\infty}} h( q^m )$. According to their definition, a $m$-string in an ordinary partition is the parts consisting of $m(1+k)$, $m(3+k)$, $\ldots$, $m(2j-1+k)$ with a positive integer $j$ and a nonnegative integer $k \le j$, and a weight of $m$-string is 1 if $k=0$ or $j$, and 2, otherwise. They defined $C_m (n)$ as the weighted sum of the number of $m$-strings along the partitions of $n$, i.e.
$$
C_m (n) = \sum_{\lambda \vdash n} \sum_{\substack{\text{$\pi$ is a} \\ \text{$m$-string of $\lambda$}}} \operatorname{wt} (\pi).
$$
It is clear that
$$
\frac{1}{(q)_{\infty}} h(q^m) = \sum_{n \ge 1} C_m (n) q^n.
$$
So the conjecture of Andrews, Chan, Kim and Osburn \cite{acko} can be intepretated as there are more (weighted count of) $1$-strings than $m$ times of (weighted count of) $m$-strings along the partitions of $n$. In the paper \cite{kks}, Byungchan Kim, Eunmi Kim and Jeehyeon Seo, by using the circle method of Wright \cite{w} and some results
from \cite{bm2}, proved the conjecture of Andrews-Chan-Kim-Osburn is asymptotically true. In this note, we will prove this conjecture is true if $m$ is any power of $2$. Moreover, we show that in order to prove the conjecture, it is only to prove it is true for all primes $m$. Here are our main results.
\begin{theorem}\label{main}
For all integer $m\geq 2$,
$$
\frac{1}{(q)_{\infty}}(h(q)-2^{m}h(q^{2^{m}})
$$
has positive power series coefficients for all positive powers of $q$.
\end{theorem}
\begin{theorem}\label{second main}
Suppose for a prime $p$,
$$
\frac{1}{(q)_{\infty}}(h(q)-ph(q^{p})
$$
has positive power series coefficients for all positive powers of $q$.
Then for all integer $m\geq 2$,
$$
\frac{1}{(q)_{\infty}}(h(q)-p^{m}h(q^{p^{m}})
$$
has positive power series coefficients for all positive powers of $q$.
\end{theorem}
We don't know why primes appeared here, in Section 3, we give a stronger conjecture related primes.
\section{ Proof of Theorem 1.1. and Theorem 1.2}
\noindent For a prime $p\geq 1$, we define the function
$
M_{p}(q)=\frac{1}{(q)_{\infty}}(h(q)-ph(q^p)),
$
\noindent then we have
$M_{2}(q)$ has positive power series coefficients for all positive powers of $q^{n}$ for all $n$ except that the coefficients of $q^{2}$ and $q^{4}$ are zero. This lemma is the Corollary 2.4 of \cite{acko}.
{\it Proof of Theorem 1.1.} $m\geq 2$,
\begin{multline*} \label{1}
\begin{aligned}
&M_{2^{m}}(q)
= \frac{1}{(q)_{\infty}}(h(q)-2^{m}h(q^{2^{m}})) \\
& = \frac{1}{(q)_{\infty}}(h(q)-2h(q^2)+2h(q^{2})-4h(q^4)+\cdot\cdot\cdot
+2^{m-1}h(q^{2^{m-1}})-2^{m}h(q^{2^{m}}))\\
& =\frac{h(q)-2h(q^2)}{(q)_{\infty}}+2\frac{h(q^{2})-2h(q^4)}{(q)_{\infty}}+\cdot\cdot\cdot
+2^{m-1}\frac{h(q^{2^{m-1}})-2h(q^{2^{m}})}{(q)_{\infty}}\\
&=\frac{h(q)-2h(q^2)}{(q)_{\infty}}+
\frac{2}{(1-q)(1-q^{3})(1-q^{5})\cdot\cdot\cdot}\cdot\frac{h(q^{2})-2h(q^{4})}{\prod^{\infty}_{n=1}(1-q^{2n})}+\cdot\cdot\cdot \\
&+\frac{2^{m-1}}{\prod_{n\neq 2^{m-1}k, k\geq 1}(1-q^{n})}\cdot \frac{h(q^{2^{m-1}})-2h(q^{2^{m}})}{\prod_{n=1}(1-q^{2^{m-1}n})}\\
&=M_{2}(q)+\frac{2}{(1-q)(1-q^{3})(1-q^{5})\cdot\cdot\cdot}\cdot M_{2}(q^{2}) + \cdot\cdot\cdot
+ \frac{2^{m-1}}{\prod_{n\neq 2^{m-1}k, k\geq 1}(1-q^{n})}\cdot M_{2}(q^{2^{m-1}}).
\end{aligned}
\end{multline*}
\noindent By the lemma above, $M_{2}(q)$ has positive coefficients of $q^{n}$ for all $n$ with except 2 and 4, therefore for all $m\geq 2$, $M_{2}(q^{2^{m-1}})$ has positive coefficients except that the coefficients of $q^{2^{m+1}}$ and $q^{2^{m+2}}$, but the sum of the left hand will have positive coefficients of $q^{n}$ for all $\geq 1$.
\qed
{\it Proof of Theorem 1.2.}
For $p\geq 3$ a prime and $m\geq 2$,
\begin{equation*}
\begin{aligned}
&M_{p^{m}}(q)
= \frac{1}{(q)_{\infty}}(h(q)-p^{m}h(q^{p^{m}})) \\
& = \frac{1}{(q)_{\infty}}(h(q)-ph(q^p)+ph(q^{p})-p^{2}h(q^{p^{2}})+\cdot\cdot\cdot
+p^{m-1}h(q^{p^{m-1}})-p^{m}h(q^{p^{m}}))\\
& =\frac{h(q)-ph(q^p)}{(q)_{\infty}}+p\frac{h(q^{p})-ph(q^{p^{2}})}{(q)_{\infty}}+\cdot\cdot\cdot
+p^{m-1}\frac{h(q^{p^{m-1}})-ph(q^{p^{m}})}{(q)_{\infty}}\\
&=\frac{h(q)-ph(q^p)}{(q)_{\infty}}+
\frac{p}{\prod_{n\neq pk, k\geq 1}(1-q^{n})}\cdot\frac{h(q^{p})-2h(q^{p^{2}})}{\prod^{\infty}_{n=1}(1-q^{pn})}+\cdot\cdot\cdot \\
&+\frac{p^{m-1}}{\prod_{n\neq p^{m-1}k, k\geq 1}(1-q^{n})}\cdot \frac{h(q^{p^{m-1}})-ph(q^{p^{m}})}{\prod_{n=1}(1-q^{p^{m-1}n})}\\
&=M_{p}(q)+\frac{p}{\prod_{n\neq pk, k\geq 1}(1-q^{n})}\cdot M_{p}(q^{p}) + \cdot\cdot\cdot
+ \frac{p^{m-1}}{\prod_{n\neq p^{m-1}k, k\geq 1}(1-q^{n})}\cdot M_{p}(q^{p^{m-1}}).
\end{aligned}
\end{equation*}
Since each summand of the right hand side of the above has positive coefficients of $q^{n}$ for all positive integers $n$, $M_{p^{m}}$ will have positive coefficients
of $q^{n}$ for all integers $n$.
\qed
{\it Remark.}
By our method, it can be easily seen that if the conjecture is true for all primes $m=p$, then it is true for any other natural numbers $m$. For example, consider the case $m=6$,
\begin{equation*} \label{1}
\begin{aligned}
&\frac{h(q)-6h(q^{6})}{(q)_{\infty}}\\
& = \frac{h(q)-2h(q^{2})}{(q)_{\infty}}+\frac{2h(q)-6h(q^{6}))}{(q)_{\infty}}\\
&=M_{2}(q)+\frac{2}{(1-q)(1-q^{2})(1-q^{3}\cdot\cdot\cdot)}\cdot \\
&\frac{h(q^{2})-3h(q^{6})}{(1-q^{2})(1-q^{4})(1-q^{6})\cdot\cdot\cdot}\\
&=M_{2}(q) + \frac{2}{(1-q)(1-q^{3})(1-q^{5})\cdot\cdot\cdot}\cdot M_{3}(q^{2}).
\end{aligned}
\end{equation*}
\noindent We see that the positivity of coefficients of the power series $M_{2}(q)$ and $M_{3}(q)$ will imply the positivity of the coefficients of the power series of $M_{6}(q)$.
\section{A Stronger Conjecture }
Kim et al. proved that the conjecture of Andrews et al is asymptotically true by using circle method, But Andrews et al. originally expected to find $q$-theoretic or combinatorial proofs for this conjecture. Here basing on the numerical results, we make the following stronger conjecture. We also expected find $q$-theoretic or combinatorial proofs for this conjecture. We can easily see that this conjecture implies the conjecture of Andrews et al.
\begin{conjecture}
Let $p_{1}>p_{2} \geq2$ be two primes. Then the function
$$
\frac{1}{(q)_{\infty}} ( p_{1}h(q^{p_{1}}) - p_{2}h(q^{p_{2}}) )$$
has positive power series coefficients of $q^{n}$ for all $n\geq p_{2}$ and has nonnegative power series coefficients of $q^{n}$ for all $n\geq 1$.
\end{conjecture}
We verified this conjecture for the first $100,000$ coefficients of the power series for each prime pair cases which are less than $50$ by using Mathematica. We provide some coefficients of power series of $\frac{1}{(q)_{\infty}} ( h(q) - m h(q^m) )$ for small prime $m$, which are also obtained by using Mathematica.
\begin{equation*}
\begin{aligned}
&\frac{1}{(q)_{\infty}} ( h(q) - 3 h(q^3) )\\
&=q + 2 q^2 + 3 q^4 + 3 q^5 + 4 q^6 + 5 q^7 + 9 q^8 + 10 q^9 \\
&+16 q^{10} + 19 q^{11} + 26 q^{12} + 33 q^{13} + 46 q^{14} + 56 q^{15} +\cdot\cdot\cdot \\
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned} & \frac{1}{(q)_{\infty}} ( h(q) - 5 h(q^5) \\
& =q + 2 q^2 + 3 q^3 + 6 q^4 + 4 q^5 + 11 q^6 + 13 q^7 + 21 q^8 \\
&+27 q^9 + 36 q^{10} + 46 q^{11} + 67 q^{12} + 82 q^{13} + 111 q^{14} \\
&+ 141 q^{15} +\cdot\cdot\cdot
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned} & \frac{1}{(q)_{\infty}} ( h(q) - 7 h(q^7) )=q + 2 q^2 + 3 q^3 + 6 q^4 + 9 q^5 \\
&+ 16 q^6 + 16 q^7 + 29 q^8 +
38 q^9 + 55 q^{10} \\
&+ 71 q^{11} + 103 q^{12 }+ 130 q^{13} + 174 q^{14 }+
225 q^{15} +\cdot\cdot\cdot\\
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}& \frac{1}{(q)_{\infty}} ( h(q) - 11 h(q^{11}) ) =q + 2 q^2 + 3 q^3 + 6 q^4 + 9 q^5\\
& +16q^6 + 23q^7 + 36 q^8
+52q^9 + 76q^{10 }+ 95q^{11}\\
& + 141 q^{12} + 185 q^{13} +253 q^{14} + 331 q^{15}+\cdot\cdot\cdot \\
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned}
& \frac{1}{(q)_{\infty}} ( h(q) - 13 h(q^{13}) )=q + 2 q^2 + 3 q^3 + 6 q^4 + 9 q^5\\
& + 16 q^6 + 23 q^7 + 36 q^8 +
52 q^9 + 76 q^{10}\\
& + 106 q^{11} + 152 q^{12} + 192 q^{13} + 273 q^{14} +360q^{15}+ \cdot\cdot\cdot\\
\end{aligned}
\end{equation*}
\begin{equation*}
\begin{aligned} & \frac{1}{(q)_{\infty}} ( h(q) - 17 h(q^{17}) )=q + 2 q^2 + 3 q^3 + 6 q^4 + 9 q^5\\
&+16q^6 + 23 q^7 + 36 q^8 +
52 q^9 + 76q^{10}\\
& + 106 q^{11} + 152 q^{12} + 207 q^{13} + 286 q^{14 }+
386 q^{15} + \cdot\cdot\cdot\\
\end{aligned}
\end{equation*}
\section*{Acknowledgments}
The author would like to thank Professor Peter Paule for his valuable comments on an earlier version of this paper and encouragements.
The author also would like to thank Professor Gorge Andrews for his suggestion of the rewriting some parts of this note.
|
\section*{Model}
Following Desai and Fisher (\citeyear{desai_2007}), the population is divided into discrete fitness classes differing by multiples of a constant fitness increment $s$ (Fig.\ \ref{fig:travellingwave}). Population size $N$ is assumed to be large enough that the abundances of most fitness classes behave deterministically. Desai and Fisher (\citeyear{desai_2007}) is formulated in terms of relative fitness, and $N$ is constant. We instead use a simple logistic model of absolute fitness,
\begin{equation}
\frac{1}{n_i}\frac{d n_i}{d t}=b\left(1-\frac{N}{\kappa}\right) - (d+is). \label{eq:dni_bd}
\end{equation}
Here $n_i$ is the abundance of fitness class $i$ (so that $N=\sum_i n_i$), $b$ is the intrinsic birth rate, $d$ is the mortality rate of perfectly adapted ($i=0$) individuals, and $i s$ is the additional mortality associated with fitness class $i$. The indices $i$ count the number of fitness classes from perfection, and can therefore be interpreted as a measure of lag load \citep{smith_76}. $\kappa$ is the maximum possible population size without deaths, representing territorial or resource limitations. Henceforth, the index $i$ will be used to refer to any possible fitness class, which could be empty, while $j$ will be used to refer specifically to the most fit class that has abundance large enough that it behaves approximately deterministically according to Eq.\ \eqref{eq:dni_bd} (Fig.\ \ref{fig:travellingwave}).
$\kappa$ is distinct from the maximum achievable abundance (the carrying capacity), where births balance deaths ($d N/d t=0$). If all individuals are in the same fitness class $j$, the carrying capacity is $K_j=(b-d-j s) \kappa/b$, which decreases with decreasing fitness. The population is not viable if deaths exceed births at low population density i.e. $d+i s>b$ for all occupied fitness classes. This defines an extinction threshold $i_e\approx (b-d)/s$, given by the $i$ where $d+i s$ first exceeds $b$.
We only consider beneficial mutations, and all mutations have the same fitness effect $s$ irrespective of genetic background. We assume that the beneficial mutation rate in fitness class $i$ is $Ui$ per birth, where $U$ is a constant. This represents a ``running out of mutations'' effect, where there are more ways for genetic novelty to improve fitness in poorly-adapted genotypes (and no ways to improve a perfect genotype). We return to our running out of mutations assumption, specifically how it differs from diminishing returns epistasis, in the Discussion.
\setcounter{figure}{0}
\renewcommand{\figurename}{Fig.}
\begin{figure}[h]
\includegraphics[scale=0.6]{fig1.pdf}
\caption{\label{fig:travellingwave} Absolute fitness classes $i$ representing fitness increments of size $s$, with abundances $n_i$. Environmental deterioration intermittently reduces population fitness by $s$. Fitness classes grow or decline relative to each other depending on whether their fitness is respectively greater or smaller than the mean fitness $\overline{i}$ (small vertical arrows). Population size $N$ changes dynamically with fitness (double-headed vertical arrow). At the nose of the distribution, mutant establishment is stochastic (hatched bar). The fittest established class is $j$, and mutants are $q_j=\overline{i}-(j-1)$ fitness classes away from the mean (their fitness advantage is $q_js$). The extinction threshold $i_e$ is shown with a vertical dashed line.}
\end{figure}
There is no sex, so mutations only matter in the leading deterministic class $j$, producing mutants appear in the stochastic ``nose'' $j-1$. Mutations on poorer genetic backgrounds --- away from the nose --- are doomed to be outcompeted by nose mutants (multiple-mutations interference; \citealt{desai_2007}). Thus, the only relevant mutation rate in our model is that feeding the nose $Uj$.
Mutant lineages initially have low abundance (starting from a solitary mutant), and are therefore strongly affected by demographic stochasticity. Only some mutant lineages avoid going extinct in the initial stochastic phase and attain a large enough abundance that they grow deterministically according to Eq.\ \eqref{eq:dni_bd} (a process called ``establishment''). The probability of establishment at the nose, denoted $p_{j-1}$, is approximately $q_j s/(d+js)$ for most mutations (Appendix A), where $q_j=\overline{i}-(j-1)$ is the number of fitness classes that the nose is ahead of the mean (Fig.\ \ref{fig:travellingwave}). However, $p_{j-1}$ can be substantially smaller if environmental change occurs during the establishment process. The calculation of $p_{j-1}$ in this case is discussed in Appendix A.
Once a mutant lineage established, it becomes the new most fit established class with dynamics governed by Eq.\ \eqref{eq:dni_bd}. The initial abundance for deterministic growth $\nu$, which is applied at the time that the mutation occurs, is a random variable that represents the stochasticity in the time that the mutant lineage takes to establish. The cumulative density function for $\nu$ is given by \citep[Eq.\ 40]{uecker_2011}
\begin{equation}
P(\nu\leq\nu_0)=1-e^{-\nu_0 p_{j-1}},\label{eq:nu_dist}
\end{equation}
so that the mean of $\nu$ is $1/p_{j-1}$ \cite[also see][Eq. 16]{desai_2007}.
Note that there is a clean separation between the deterministic bulk obeying Eq.\ \eqref{eq:dni_bd}, with fittest class $j$, and the stochastic nose in fitness class $j-1$, as shown in Fig. \ref{fig:travellingwave}. This clean separation holds when establishing mutants make up a small fraction of the population ($Ns\gg 1$), and $Uj$ is small enough that mutant lineages rarely produce double-mutants before establishment (${\rm birth\,rate}\times Uj\ll s$; \citealt{desai_2007}), as is the case here.
Environmental deterioration occurs in discrete events where the entire fitness distribution is shifted backwards by one fitness class ($i\rightarrow i+1$ for all of the population's fitness classes). These events are assumed to follow a Poisson process with mean time $T$ between successive changes. Thus, in addition to being a measure of lag load, $i$ can also be interpreted as the number of environmental challenges facing the individuals in fitness class $i$. In this interpretation, the linear dependence of the mutation rate $Uj$ on $j$ can be interpreted as saying that each environmental deterioration event opens up one new possible beneficial mutation that addresses the new environmental challenge.
Environmental challenges may be biotic or abiotic in our model; fitness differences simply represent differences in mortality without specifying causes. We could easily attribute fitness differences in Eq. \eqref{eq:dni_bd} to births or a mixture of births and deaths instead, but this would not substantially alter our model's behavior.
\section*{Results}
The model described above is simulated numerically (implementation is summarized in Supplement A). In addition to these simulations, we show that the population's long-term evolution can be approximated with a much simpler discrete-time Markov chain (MC).
\subsection*{Markov chain approximation}
Our model has two distinct adaptive regimes: the ``successional'' regime, and the ``multiple mutations'' regime. We first describe our MC approximation in the ``successional'' regime, where fixation (the growth of a newly established mutant to a frequency of $1$) is much faster than the typical time between mutant establishments. The population spends most of the time in equilibrium with all individuals in one fitness class $j$ ($N\approx K_j$), waiting for adaptive mutant establishment or environmental change. Adaptive advances occur at a mean rate $v_j$ equal to the equilibrium birth rate $K_j b(1-K_j/\kappa)$ multiplied by the mutation rate $Uj$ and establishment probability $p_{j-1}$,
\begin{equation}
v_j=K_j b(1-K_j/\kappa) Uj p_{j-1}. \label{eq:vsuccessional}
\end{equation}
Our MC approximation amounts to taking regular ``snapshots'' at intervals given by the characteristic fixation time (Appendix B). In the vast majority of snapshots, the population will be in equilibrium, and will jump between fitness classes with per-snapshot probabilities proportional to $1/T$ for environmental change $j\rightarrow j+1$, and $v_j$ for adaptive mutant fixation $j\rightarrow j-1$. These are the MC transition probabilities. The current MC state is $j$, with possible values $i=0,\ldots,i_e$ (Fig.\ \ref{fig:chainfigure}a), and each MC iteration represents the fixation time. The extinction threshold $i_{e}$ is an absorbing state because the corresponding $N=0$ equilibrium is attained within a single iteration (the scenario where the extinction threshold has been crossed but the population manages to recover by producing higher-fitness individuals with $i<i_e$, called ``evolutionary rescue'', is not possible in our MC approximation).
In the ``multiple mutations'' regime, fixation is slower than the rate at which mutants destined for establishment are produced, and there is standing fitness variation ($q_j>0$). By invoking beneficial mutation-selection balance, the steady-state adaptation rate can be approximated analytically for a given mutation rate, population size and establishment probability \citep{desai_2007}. The latter quantities depend on fitness, particularly the position of the most fit established class $j$. Thus, the corresponding steady state also depends on $j$, giving (Appendix C)
\begin{equation}
q_{j}\approx \frac{2\ln (K_j s)}{\ln{(s/Uj)}},\quad v_j\approx \frac{2\ln (K_j s)-\ln(s/Uj)}{\ln^2(s/Uj)}s. \label{eq:vmm}
\end{equation}
This is a straightforward generalization of Eqs. (40) and (41) in \cite{desai_2007}, which assumed constant mutation rate, population size and establishment probability.
Unlike the successional regime MC approximation, adaptation cannot be treated as memoryless (independent of the population's history) in the multiple mutations MC approximation. The mutant-generating class $j$ grows over time, and thus so does the overall rate of mutant production at the nose. Consequently, mutant establishment is much less likely shortly after the previous establishment, while class $n_j$ is still small. Moreover, previous growth at the nose is not ``forgotten'' when environmental changes occur. Thus, mutant establishments occur more regularly than memoryless events with mean rate $v_j$. Accordingly, we use two MC approximations to bound the actual behavior of the multiple mutations regime. In the first, we ignore memory so that $v_j$ from Eq.\ \eqref{eq:vmm} is the $j\rightarrow j-1$ transition probability, analogous to the successional case (Fig.\ 2a). In the second, we assume that mutant establishment occurs periodically at given intervals (Fig.\ 2b). Each iteration of the periodic-adaptation MC represents the time required for mutant establishment $1/v_j$, and exactly one establishment happens every iteration. Note that the memoryless and periodic-adaptation MC chains differ only in whether or not \textit{adaptation} is memoryless: in both cases, the transition probabilities are memoryless, as they must be in a Markov chain. The mathematical details of our MC approximation are given in Appendix B.
\begin{figure}
\includegraphics[scale=.6]{fig2.png}
\caption{\label{fig:chainfigure} In addition to simulations of the traveling wave model illustrated in Fig.\ \ref{fig:travellingwave}, two Markov chain approximations are used to model long-term evolution. (a) Memoryless adaptation with $j\rightarrow j-1$ transition probability proportional to $v_j$ where $j$ is the fittest established class, and $v_j$ is given by Eq.\ \eqref{eq:vsuccessional} (successional regime) and Eq.\ \eqref{eq:vmm} (multiple-mutations regime). (b) Periodic adaptation in the multiple mutations regime, where each iteration represents the establishment timescale $1/v_j$. Exactly one adaptation event occurs each iteration, but a variable number of environmental change events can occur.}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.5]{fig3.pdf}
\caption{\label{fig:meantimesmm} Time to extinction $t_e$ increases abruptly with increasing $T$. (a) As $T$ increases, the population transitions from ``always losing'' to ``sometimes winning''. Arrows show mean direction of fitness change near points where $1/T=v_j$; $j/i_e\approx 0.4$ is an ``attractor''. (b) Comparison of simulated $t_e$ and mean $t_e$ predicted from MC approximation (Appendix B). Parameters: $b=2,d=1,U=10^{-6},s=0.02,\kappa=4\times 10^6, j_{\rm initial}/i_e=0.4$ (multiple mutations regime).}
\end{figure}
\subsection*{Extinction times}
Long-term evolution is primarily controlled by the difference between the opposing rates of environmental change $1/T$ and adaptation $v_j$, where $v_j$ tends to zero at perfection $j=0$ (no beneficial mutations) and extinction $j=i_e$ ($N=0$), and exhibits a peak between these extremes (Fig.\ \ref{fig:meantimesmm}a). Figure \ref{fig:meantimesmm}b shows the predicted pattern of time to extinction $t_e$ versus $T$ (time is measured in generations, implemented by setting $d=1$). When environmental change is relatively rapid ($T<150$), the population cannot keep up with environmental deterioration ($1/T>v_j$ for all $j$), and extinction occurs rapidly (thousands of generations). Modestly slowing environmental change ($T=180$) allows the population to beat the environment ($1/T<v_j$) over part of the fitness domain, dramatically increasing the mean and variance in $t_e$ \cite[compare][Fig. 1b]{burger_95}. This sudden transition to long-term persistence occurs at $T\approx 160$. Fig.\ \ref{fig:meantimesmm} is in the multiple mutations regime; the successional regime gives essentially identical results, except with lower persistence times for given $T$ (since, by definition, there are far fewer adaptive mutants establishing).
MC mean extinction times $\langle t_e\rangle$ (Appendix B) closely follow the full simulation results (Supplement A) in Fig.\ \ref{fig:meantimesmm}b. The periodic-adaptation MC performs better than the memoryless MC, confirming the importance of mutant establishment ``memory'' in the multiple mutations regime.
Fig.\ \ref{fig:extimehist}a shows the distribution of extinction times for a population with low initial fitness $j_{\rm initial}/i_e=0.8$ \cite[compare][Fig. 4]{burger_95}. This could represent a newly establishing population at the start of peripatric speciation, for example. As expected, the distribution is sharply peaked near zero, reflecting a high risk of early extinction. The distribution also has a long tail (Fig.\ \ref{fig:extimehist}a inset), reflecting cases where the population manages to reach the stable ``attractor'' at $j/i_e\approx 0.4$ (Fig.\ \ref{fig:meantimesmm}a). Once the attractor is reached, long-term persistence is possible.
\begin{figure}
\centering
\includegraphics[scale=0.6]{fig4.pdf}
\caption{\label{fig:extimehist} The distribution of extinction times for fixed $T$ is not exponential, but has an exponential tail. Main figure: $t_e$ from memoryless MC simulations (histogram; same parameters as Fig.\ \ref{fig:meantimesmm} except $j_{\rm initial}/i_e=0.8$) compared with the corresponding (same mean) exponential distribution (curve). Inset: same simulations omitting the initial spike of rapid extinction $t_e<4\times10^5$ (histogram). Curve shows exponential distribution with mean given by MC mean $t_e$ (using Eq.\ \eqref{eq:meante_linsys}) assuming $j_{\rm initial}/j_e=0.4$ (the attractor in Fig.\ \ref{fig:meantimesmm}).}
\end{figure}
The tail in the distribution of $t_e$ is exponential (Fig.\ \ref{fig:extimehist}a inset), indicating that extinction risk is effectively constant over time. The reason for this constant risk is that the population remains near the attractor for most of its existence. When extinction does occur, it is due to an abnormally rapid sequence of environmental changes and/or slow sequence of mutant establishments rather than a gradual erosion of fitness. The resulting decline in fitness is rapid compared to the mean persistence time. Fig.\ \ref{fig:beforeext} shows the average decline in fitness immediately prior to extinction obtained by averaging over simulated trajectories in the full model as well as a backward-time variant of our MC approximation (Supplement B); moving from the attractor to the extinction threshold only takes around $1\%$ of mean persistence time.
This explains why such a substantial discrepancy exists between the memoryless MC and simulations after the transition to persistence ($T>160$), but not before it ($T\approx 100$): after the transition, large fluctuations in the number of adaptive establishments are much more likely when establishments are memoryless. Before the transition, there is no attractor, mean time to extinction is determined by the average decline in fitness due to the fact that $1/T<v_j$ (specifically, $\langle t_e \rangle = j_{\rm initial}/(1/T-v_j)$), and fluctuations are only of secondary importance.
\begin{figure}[h]
\centering
\includegraphics[scale=0.6]{fig5.pdf}
\caption{\label{fig:beforeext} For persistent populations, fitness declines rapidly at the time of extinction compared to mean $t_e$. Shown is the expectation of $j$ for the backward-time trajectory immediately preceding extinction, using both backward-time memoryless MC (Supplement B) and averaging over simulated trajectories in the full model. Mean persistence time is $\approx 2\times 10^6$ generations. Same parameters as Fig.\ \ref{fig:meantimesmm} with $T=180$.}
\end{figure}
\section*{Discussion}
It is difficult to directly compare our predictions with fossil data because we have only considered a single population adapting to local environmental changes. Fossil extinction times also reflect larger scale processes in which environmental heterogeneity, range shifts and migration are potentially important. Nevertheless, population-level processes should have a strong influence at larger spatial and temporal scales, particularly for species with relatively small ranges.
For concreteness, consider the example of mollusc species, which feature prominently in the fossil record, and can have tiny geographic ranges \citep{stanley_1986}. A representative generation time for fossil mollusc species is on the order of $10$ years \citep{powell_2011}. With this generation time, our extinction time predictions (Fig.\ \ref{fig:meantimesmm}) are easily large enough to be consistent with typical fossil species persistence times of several million years \citep{raup_1994}. The shortest interval between environmental changes consistent with long-term persistence is $T\approx160$ generations, or approximately a $2\times 10^3$ years between $2\%$ increases in mortality rate (since $d=1$ and $s=0.02$). By comparison, the major glacial cycles in the last $1$ million years occurred at intervals of roughly $10^5$ years \citep{augustin_2004}. Thus, in the absence of adaptation, the mortality rate would roughly double over the course of each major glaciation cycle. This rate of environmental change is certainly gradual compared to catastrophes such as the aftermath of bolide impacts, but is still rapid enough to present a significant threat to long-term persistence. Accordingly, the long-term persistence of our population is not a trivial consequence of a negligible environmental threat. Our population size, which is order $10^6$ individuals (except when close to extinction), is considerably smaller than is typical for extant molluscs \citep{stanley_1986}, and can be viewed as a conservative lower bound (in any case, $v_j$ only increases logarithmically with $N$ in Eq. \eqref{eq:vmm}). These results provide a rare bridge between micro-evolutionary population genetic models and macro-evolutionary phenomena
The fossil record contains many instances of abundant, widely-distributed species that have suddenly disappeared. This is commonly cited in support of a catastrophic view of extinction \citep{raup_1994}. In contrast, Darwin seems to have regarded sudden disappearance as a fossilization artifact, holding that species typically disappear gradually ``first from one spot, then from another, and finally from the world'' \cite[pp. 317]{darwin_1859}, driven by inter-specific competition \citep{raup_1994}. Thus, there is no need to ``invoke cataclysms to desolate the world'' \cite[pp. 73]{darwin_1859}. These viewpoints share the questionable assumption that sudden disappearance --- assuming it is not an artifact --- indicates a severe, sudden driver of extinction. This clearly need not be true in light of the suddenness of extinction in our gradualist model (Fig.\ \ref{fig:beforeext}), which would appear as a long period of relatively stable abundance followed by sudden disappearance. Sudden disappearance is driven entirely by gradual evolutionary processes, not the one or few extreme environmental changes that characterize catastrophes. In a sense it is still a ``catastrophe'' --- an abnormally large fitness fluctuation --- but this fluctuation reflects poor adaptive performance just as much as environmental pressure. Thus, sudden disappearance alone does little to distinguish between catastrophic or gradual extinction scenarios. The case for a catastrophic interpretation is much stronger if many thriving taxa disappear synchronously (mass extinction), but this excludes much of the fossil record of extinction \citep{raup_1994}, which could therefore plausibly be driven by gradual processes instead.
Long-term evolution in the vicinity of a fitness attractor is a form of Red Queen evolution in which fitness gains are continually thwarted by environmental deterioration, resulting in effectively stagnant mean absolute fitness \citep{van_1973}. As a consequence, persistent populations will have exponentially distributed times to extinction $t_e$, because extinction risk will be essentially independent of population age (ignoring short-term fitness fluctuations). However, if fitness is initially low, say because young populations tend to be colonizers in unfamiliar environments, then the risk of early extinction will be elevated (Fig.\ \ref{fig:extimehist}), and older populations will be less extinction-prone than younger ones. Intriguingly, fossil genera do exhibit reduced extinction risk with age, even after controlling for geographic range and species richness \citep{finnegan_2008}. Our results raise the possibility that population-level evolutionary processes contribute to this pattern (even without major differences in mutation rate or population size), provided that the predicted initial elevation of extinction risk lasts long enough to leave a fossil signature.
It is interesting to consider the role of genetic load in our model, since different interpretations of load have featured prominently in previous discussions about adaptation rates and extinction risk. In particular, substitutional load arguments directly contributed to the formulation and popularity of neutral theory \citep{kimura_1968}, but their interpretation was controversial. Kimura argued that most substitutions must be neutral because a many-locus version of Haldane's single-locus substitution implies extremely large substitutional loads. However, calculating a cost of selection in this way presumes that the perfect genotype (i.e. with the fittest allele at all loci considered) is present in the population \cite[pp. 78]{ewens_2004}. This effectively conflates the relative substitutional load (proportional to the fitness of the fittest genotype present minus mean fitness; \citealt{crow_1968}) with absolute lag load (proportional to the fitness of the perfect genotype minus mean fitness; \citealt{smith_76}). In our model, the substitutional load is $q_js$, a crucial determinant of the rate of adaptation $v_j$, while the lag load is $\overline{i}s$, a measure of extinction risk. The two are interdependent. In steady state this follows immediately from Eq.\ \eqref{eq:vmm}: for given population parameters, the steady state values of $q_j$ depend on $j$ (and therefore mean fitness $\overline{i}$). Or, looking at it another way, the fitness advantage of new mutants $q_js$ determines $v_j$, which in turn determines the location of the fitness attractor (and if one exists). Interdependence between these relative (substitutional) and absolute (lag) loads is a natural consequence of evolution on an absolute fitness axis driven by relative fitness differences. Substitutional and lag load can therefore be seen as complementary --- but not independent --- aspects of a population's long-term evolutionary status.
Our model is superficially similar to models of mutation load accumulation \citep{lynch_93,kondrashov_1995}, where, instead of environmental change, deleterious mutations gradually erode fitness. However, the effects of accumulating deleterious mutations throughout the population is potentially considerably more complicated, and much weaker, than the population-wide fitness deterioration induced by environmental shifts. For the large asexual populations considered here, provided that the deleterious mutation rate $U_d$ is not very large ($U_d/s\ll 1$), deleterious mutations have little effect on the overall rate of fitness gain regardless of their fitness effect \citep{desai_2007}. If $U_d$ is large enough, a reversible Muller's ``ratchet'' will begin to turn, shifting the entire population one fitness class at time, much like environmental change. However, either $U_d$ would need to be very large or $N$ very small for this mutation-induced deterioration to overpower beneficial mutant establishment and pose an extinction risk \citep{xiaoqian_2011,goyal_2012}.
Some of the results presented here were anticipated by \cite{burger_95}, particularly the pattern shown in Fig. \ref{fig:meantimesmm}b. A major focus of their analysis is what determines the critical rate of environmental change consistent with persistence ($1/T=\max_j v_j$ in our model). Their modeling assumptions are quite different from ours: sex is obligate with free recombination, and population size is small (at most $512$ individuals) and effectively constant (except for after the population crosses the extinction threshold). Consequently, their population has very little linkage disequilibrium, and $N$ is so small that genetic drift and demographic stochasticity are important factors. Our population has high linkage disequilibrium, and $N$ is large enough that stochasticity only plays a role in the establishment of new beneficial mutations. Our approaches can therefore be viewed as complementary, but given the drastic difference in population sizes, it is hard to compare any of their specific predictions to ours.
Probably the biggest limitation of our model is that there is no genetic recombination. Sexual recombination is nearly universal among fossil species and the macroorganisms of interest to conservation biologists. This makes no difference for small populations, which are in the successional regime regardless of recombination. But for large populations, recombination substantially increases the rate of adaptation $v_j$ \citep{neher_2010,weissman_2012}. For relatively simple models of recombination, this increase is fairly well understood \citep{neher_2010,weissman_2012,neher_2013}. Changing $v_j$ does not alter the basic qualitative features of our model, particularly the central role of the fitness attractor, but would affect the quantitative predictions of persistence for given population parameters.
We have assumed that evolution slows down as the population approaches perfection because the availability of beneficial mutations is limited, represented by the fitness-dependent mutation rate $Uj$ (running out of mutations (RM)). An alternative mechanism for slowing evolution is that beneficial mutations are less effective on fitter genetic backgrounds (diminishing returns epistatis (DR)). Since the relative importance of these alternatives is unresolved \citep{wiser_2013,good_2014}, we checked whether our model is sensitive to the choice of RM versus DR (Supplement C). Even for relatively strong DR, the main effect of using DR instead of RM is causing $v_j$ to have a greater peak value which occurs at lower fitness. This does not alter our conclusions.
\subsection*{Acknowledgements}
This work was financially supported by Wissenschaftskolleg zu Berlin and the National Science Foundation (DEB-1348262). KG is funded by NIH Grant GM084905. We thank Taylor Kessinger for early contributions to the project and Taylor Kessinger, two anonymous reviewers and the associate editor for constructive comments on the manuscript, and Mike Sanderson and Karl Flessa for helpful discussions.
\pagebreak
\nocite{fumagalli_2015}
\bibliographystyle{plainnat}
|
\section{Introduction}
In human societies social life consists of the flow and exchange of norms, values, ideas, goods as well as other social and cultural resources, which are channeled through a network of interconnections. In all the social relations between people {\em trust} is a fundamental component~\cite{gamb}, such that the quality of the dyadic relationships reflects the level of trust between them. From the personal perspective social networks can be considered structured in a series of layers whose sizes are determined by person's cognitive constraints and frequency and quality of interactions~\cite{dunbar}, which in turn correlate closely with the level of trust that the dyad of individuals share. As one moves from the inner to the outer layers of an individual's social network, emotional closeness diminishes, as does trust. Despite its key role in economics, sociology, and social psychology, the detailed psychological and social mechanisms that underpin trust remain open. In order to provide a systematic framework to understand the role of trust, one needs to create metrics or quantifiable measures as well as models for describing plausible mechanisms producing complex emergent effects due to social interactions of the people in an interconnected societal structure.
One example of such social interaction phenomena, in which trust plays an important role, is trading between buyers and sellers. Such an economic process is influenced by many apparently disconnected factors, which make it challenging to devise a model that takes them into account. Therefore, models that have been proposed, necessarily select a subset of factors considered important for the phenomena to be described. For instance, there are studies of income and wealth distribution~\cite{sinha,wd}, using gas like models~\cite{bikas2007}, life-cycle models~\cite{wang}, game models~\cite{bikas2015}, and so on. For a review of various agent based models we refer to~\cite{sornette}. In addition, we note that detailed studies of empirical data and analysis of the distribution functions~\cite{yuqing2007} seem to lend strong support in favour of gas-like models for describing economic trading exchanges.
In order to consider the role of trust in trading relations we focus on the simplest possible situation in which trust clearly plays a definite role. This is the case of trading goods or services for money through dyadic interactions or exchange, which takes place either as a directional flow of resources from one individual to another individual or vice versa. When an agent is buying, trust plays a role, as people prefer to buy from a reliable and reputable selling agent, i.e. agent they trust. It should be noted that the dyadic relationship does not have to be symmetric, i.e. a seller does not need to trust the buyer. A key ingredient in the trading interactions is profit that an agent makes when providing goods or services, and it can realistically be assumed that a seller wants to put the highest possible price to its goods, while the buyer tends to perform operations with agents offering a low price.
In this study we propose an agent based "gas-like" model to take into account the above mentioned important features of trading. The model describes dyadic transactions between agents in a random network. The amount of goods and money are considered conserved in time, but the price of goods and trust, we measure as reputation, vary according to the specific situation in which trade is made. In section~\ref{model} we describe the model and set up the dynamic equations of the system. In section~\ref{results} we present the results of extensive numerical calculations and explore their dependence on the parameters of the model. Here we also compare our numerical results with available real data and discuss the predictions of the model as well as possible extensions to it. Finally, in section~\ref{conclusion} we conclude by making remarks concerning the role of trust in trade and social relations.
\section{The model}
\label{model}
\subsection{The basic model}
First we introduce the basic model, which describes the dynamic development of a random network of $N$ agents such that the state of agent $i$ is defined by two time-dependent state variables, $(x_i,y_i)$, where $x_i$ stands for the amount of money and $y_i$ for the amount of goods or services. The pairwise connectivities between agents in the network are described by $N\times N$ adjacency matrix $C$.
It is necessary to distinguish the direction of the flow of goods and money in the network, since agent $i$ could buy from agent $j$, or vice versa. At time $t=0$ we define two symmetric matrices, $A(t=0)$ and $B(t=0)$, with an average of $Z/2$ random entries per row, for the flow of money or goods, respectively. Then the adjacency matrix is simply $C=A+B$, and $Z$ stands for the mean degree.
The elements of $A(t)$ and $B(t)$ are defined as the normalised probabilities of transactions per unit time $\alpha$ and $\beta$, respectively and they could become asymmetric. These matrices represent the buying or selling transactions, according to the individual agent instantaneous situation.
The dynamic equations for the state variables $x$ (money) and $y$ (goods) initialised randomly $\in [ 0,1]$ are,
\begin{subequations}\label{nm}
\begin{align} \frac{\mathrm{d}x_{i}}{\mathrm{d}t} & = \sum_{j} \left[- x_{i}\beta_{ij}-s_jy_j \alpha_{ji}+ x_{j} \beta_{ji} + s_i y_i\alpha_{ij},\right] \label{second} \\
\frac{\mathrm{d}y_{i}}{\mathrm{d}t} & = \sum_{j} \left[ \frac{x_i \beta_{ij}}{s_j} +y_j\alpha_{ji}- \frac{x_{j}\beta_{ji}}{s_i}-y_i\alpha_{ij}\right].\label{third}
\end{align}
\end{subequations}
where $s_i$ is the price of the goods as decided by seller $i$, and its value depends on time. In both Eqs.~(\ref{nm}) the first and second terms on the right represent the transactions in which agent $i$ is buying goods from agent $j$. Note that there is an outflow of money (negative $\beta_{ij}$) and an inflow of goods (negative $\alpha_{ji}$). The third and last terms represent selling goods to $j$. Observe that we could simply use the first and third terms in Eq.~(\ref{second}), to represent the exchange of money, or use the second and fourth terms, if the exchanged money is represented as its transaction value $sy$ ($s$ has units of money per unit good). The same reasoning is applied to Eq.~(\ref{third}). We preferred to keep the equations in a symmetric form as regards to goods and money.
The elements of the matrices $\alpha$ and $\beta$ in the equations represent the proportion of money and goods, respectively, that an agent is distributing amongst its links in a given transaction. Therefore, the simplest expressions for the corresponding transaction coefficients are linear functions of the respective variables, thus we propose,
\begin{equation}
\begin{aligned}
\alpha_{ij}(t) &= \frac{1}{\Delta t}\frac{x_{j} \Theta(s_{i}y_i - x_{j} )}{\sum_j x_j\Theta(s_{i}y_i - x_{j} )},\\
\beta_{ij}(t)&=\frac{1}{\Delta t}\frac{y_{j} \Theta(x_i-s_{j}y_j)}{\sum_j y_j \Theta(x_i-s_{j}y_j)},
\end{aligned}
\label{ab}
\end{equation}
provided $\sum_i A_{ij}\ne0$ for $\alpha$ and $\sum_i B_{ij}\ne0$ for $\beta$, respectively. The unit time $\Delta t$ could conveniently be taken as one. The reason to include the Heaviside functions $\Theta$ is that agent $i$ buys goods from agent $j$ only if it has enough money to pay for the price agent $j$ is asking for its goods (in $\beta_{ij}$) or if agent $j$ has enough goods to satisfy agent $i$ (in $\alpha_{ji}$). By the same token, agent $i$ cannot sell to agent $j$ if it has not enough goods (as in $\alpha_{ij}$) or if the buyer has not enough money (as in $\beta_{ji}$).
The diagonal elements of these matrices represent the quantity of money or goods that are not used in the transaction. If they are set to zero, all the money that an agent has is used in all the transactions, meaning that there are no savings. The elements of these matrices cannot be negative, hence, whenever an element is about to become negative, it should be set to zero, which is equivalent to say that the link between agents is irrelevant for that particular transaction.This amounts to reshaping the trading network.
These dynamical equations constitute the basic model for trading transactions. It should be noted that the model conserves the amount of goods $(\sum_i x_i)$ and money $(\sum_i y_i)$. This restriction could be easily relaxed, if needed, by adding up sources and sinks to the equations. We decided to keep the conserved version for simplicity.
\subsection{Trust is built from reputation}
As emphasised above trust plays a key role in all social and trading transactions. In order to include trust into the transactions, we need first to assume that trust is something that you either gain or lose with time.
Therefore, it has to be related to a quantity that measures the performance of agents as traders, which we here assume to be {\it reputation} $R_i(t)$. Accordingly, we write
\begin{equation}
\begin{aligned}
R_i(t) & = \frac{ \mathcal{R}_i(t) }{\max_i \{ \mathcal{R} (t) \}}, \qquad \text{ where } \\
\mathcal{R}_i (t) & = \int_{0}^t \sum_{j=1}^N \left[ \alpha_{ij}(t') + \alpha_{ji}(t') + \beta_{ij}(t') + \beta_{ji}(t') \right] \mathrm{d} t'.
\end{aligned}
\label{R}
\end{equation}
Observe that the numerator of Eq.~(\ref{R}) at a given time increases with the number of successful transactions, since the elements of the trading matrices $\alpha$ and $\beta$ are positive definite. In order to maintain the scale of variation of all the variables between zero and one, we normalise the variable $R$ by dividing it with the maximum value of $R$, encountered in the network. Note that the reputation of a given agent could decrease in time because of the time dependent normalisation.
Let us now introduce a non-symmetrical matrix $F$, whose elements $F_{ij}$ regulate the amount of goods and money in each transaction between a pair of agents. The entries of the matrix $F$ range from zero (no transaction possible) to unity (larger amount of goods and money exchanged). We assume that agent $i$ prefers to buy if the value of the goods in its possession is smaller than the amount of money it has, and favours selling otherwise, in order to maintain a balanced stock of money and goods, and the same for agent $j$. To reflect this, one could define the matrix elements of $F$ as follows,
\begin{equation}
F_{ij}(t) = R_j \Theta(\hat{s}(t) y_i(t)-x_i(t)).
\label{F1}
\end{equation}
where $\hat{s}(t)$ is the average price over the whole network of the goods at time $t$, which could be very different from the price the individual $i$ is pricing its goods. Here it is assumed that agent $i$ is buying from agent $j$, and therefore the reputation of agent $j$ is the factor that matters. Trust on sellers when buying is important because one buys from a reliable agents only. On the contrary, when selling one is happy to do it to whomever is able to pay, independently of the reputation of the buyer.
Note that the transpose $F_{ji}$, should be used on the transactions in which agent $i$ sells goods, in which case the reputation of agent $i$ matters.
Taking this new feature into account, the dynamical equations of our model read as follows,
\begin{equation}
\begin{aligned}
\frac{\mathrm{d}x_{i}}{\mathrm{d}t} = \sum_{j} & \left[ - x_{i}\beta_{ij}F_{ij} - s_jy_j \alpha_{ji}F_{ij} \right. \\
& \qquad \left. + x_{j} \beta_{ji}F_{ji} + s_i y_i\alpha_{ij}F_{ji},\right], \\
& \\
\frac{\mathrm{d}y_{i}}{\mathrm{d}t} = \sum_{j} & \left[ \frac{x_i \beta_{ij}F_{ij}}{s_j} +y_j\alpha_{ji}F_{ij} \right. \\
& \qquad \left. - \frac{x_{j}\beta_{ji}F_{ji}}{s_i}-y_i\alpha_{ij}F_{ji}\right].
\end{aligned}
\label{nmF}
\end{equation}
Notice that these model equations reduce to those of the basic model (Eqs.~(\ref{nm})) if one eliminates the effect of trust by setting $F_{ij}=F_{ji},=1, \; \forall i,j$. In this sense, the effect of trust is to regulate the amount of goods and money exchanged in each transaction, but also in deciding if one sells or buys.
\subsection{Evolution of prices $s(t)$}
The main quantity that regulates the dynamics of goods and money in our trading model is the time dependence of the price of the goods. Therefore, one needs to suggest a mechanism by which an agent modifies the price of its goods. We assume that in general people would like to sell at the highest possible price, but we need to consider to what extend it is limited by the agent's reputation.
If the agent is reputed to be a successful trader, it could increase its prices by a large amount, provided that the price the agent $i$ is using is smaller than the average price the agents connected to it are offering, otherwise lowering the value of $s_i$ is a way of attracting more transactions. In a mathematical language, one can represent this situation by the following coupled first order equations,
\begin{equation}
\label{sd1}
\begin{aligned}
\frac{\mathrm{d} \tau_i}{\mathrm{d}t} & = - \frac{1}{t_0} R_i(t) \left[ s_i(t) - \frac{1}{k_i} \sum_{j=1}^{k_i}s_j(t) \right], \\
& = -\frac{1}{t_0} R_i(t) \left[ s_i(t) - s^k_i(t) \right], \\
\frac{\mathrm{d} s_i}{\mathrm{d}t} & = \frac{1}{t_0} \left[ R_i \tau_i(t) - g s_i(t) \right],
\end{aligned}
\end{equation}
where $\tau_i$ is a variable that enables agent $i$ to decide how to modify its prices according to trust and the information available to it. The parameter $t_0$ is the time scale for price changes and $g$ is a constant that represents reluctance for the agent to lower the prices, which for simplicity is assumed to be very small and the same for all the agents.
We can eliminate the variable $\tau$ from this equations and obtain a second order equation for $s$, which we identify as a set of damped harmonic oscillators, which admits an
analytical solution, namely,
\begin{equation}
\label{dho}
s_i(t)=s_i(0)e^{-\gamma t}\left[\cos (\omega_i t)-\frac{\gamma}{\omega_i}\sin (\omega_i t) \right] + s_i^k(t),
\end{equation}
where $\gamma=g/2$ and $\omega_i=\sqrt{(4R_i^2-g^2)}/2$. Notice that this solution is approximate and valid only when the prices do not change appreciably within a time step $\mathrm{d}t$, which means that $t_0$ is large and $g$ is small. In any case one could find the solution of this non-linear system of coupled oscillators numerically. The solutions of Eqs.~(\ref{sd1}) produce an apparently chaotic behaviour for $s(t)$ showing qualitatively similar to the price variations of real markets (there are even some models of the stock market that use damped oscillators to analyse the price variation~\cite{sor}).
\section{Results.}
\label{results}
The numerical calculations were made in a random network of 500 agents with the average degree of 12, using a simple Euler method with random initial conditions for $x$, and $y \, \in [0, 1]$. The appropriate step size for convergence was found to be $\mathrm{d}t = 0.05$, reached within 200000 iterations. At time $t = 0$ we set $s(0)$ randomly using a flat distribution with the mean $\bar{s}=1.2$ and the width $\mathrm{d}s = 0.4$, and set a time scale $t_0 = 400$ and $g = \mathrm{d}t / t_0$. For this $t_0$ the system reaches dynamical equilibrium within the running time. The diagonal elements of the trading matrices $A(t)$ and $B(t)$ were set to zero.
In Fig.~\ref{fig1} we depict the structure of the final network by showing only the active links, where the height of the 3D plot is stratified according to the wealth of the agent $i$, i.e. $w_i = x_i + y_i$. Here it is seen that there are poor agents and only a few very rich ones. In addition we observed in numerous calculations that if the width of the distribution of prices is increased, the rich become richer and the number of poor agents increases.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=6cm]{fig1}
\caption{Final configuration of an active trading network of buyers and sellers when trust is included in trading transactions. The colour code for the links between agents is as follows: Black $\alpha\ne 0$ and $\beta \ne0$, Red $\alpha\ne 0$ and $\beta =0$, and Blue $\alpha=0$ and $\beta \ne0$. The colour of the nodes is chosen according to the wealth ($w_i=x_i+y_i)$ such that light green, red, dark green, blue, grey, black are in decreasing order of $w$.}
\label{fig1}
\end{center}
\end{figure}
In Fig.~\ref{fig2} we show the time histories of the variables for the typical calculation of the previous figure. Note that most agents have converged to certain amounts of money and goods, but the ones who posses more wealth are prevented from taking it all, which is an effect of trust. Also observe oscillations of prices and apparently chaotic behaviour of trust. With the chosen value of $g$ (reluctance for the agent to lower the prices) the average price is maintained around a constant value, but with increasing $g$ the average price diminishes, because one is less reluctant to lower its prices. If all prices are set to the same value, the dynamical behaviour of $R$ turns out to be less chaotic, and shows less variations.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig2}
\caption{The time histories of the variables for amounts of traders' or agents' money and goods, and prices of goods and sellers' reputations.}
\label{fig2}
\end{center}
\end{figure}
In Fig.~\ref{fig3} the behaviour of the average reputation of sellers and price of goods are depicted. Here we see that the mean value of reputation of agents decreases with time, which reflects the fact that there are less successful transactions, while the reputation of some selected individuals increases with time.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig3}
\caption{The time histories of the average price of goods and reputation of sellers in the network.}
\label{fig3}
\end{center}
\end{figure}
In order to assess the effect of including trust in the transactions, we made a calculation in which no trust variables were included, that is $F_{ij}(t)=1$ and $R_i=1$, $\forall (i,t)$, and another calculation including the dynamics of the trust variables. In Fig.~\ref{fig4} we show a representation of the network as in Fig.~\ref{fig1} and a 2D graph of the two calculations.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig4}
\caption{Results of a typical network in which no trust variables are included (left hand side of the vertical black line), and including trust dynamics (right hand side). The yellow lines represent inactive links. }
\label{fig4}
\end{center}
\end{figure}
Here we observe that trust has two main effects: 1) it reinforces the trade network, the number of active links is increased noticeably, while in the calculation without trust (indiscriminate trading) the network is dismembered in isolated subgraphs with very few active links. 2) The wealth distribution between agents turns out to be more fair, in the sense that there are less poor people and quite a large proportion in the middle class. The backbone network, in which active trading took place, is noticeably larger and with more "black" or active interactions ($\alpha$ and $\beta \ne0$), the dramatic effect of trust is: A disconnected trading network without trust becomes a robust and connected network when trust is affecting agents in deciding transactions.
We have also observed that in the calculations with trust the state variables remain positive or zero, while in the calculations without trust a small number of agents have negative values. This means that these agents are not only poor but in debt but the number of such agents is very small, such that on average there are 160 agents with negative values in a population of 5000 agents after 200000 time steps.
As one of our main research foci we compare the distribution of wealth calculated from our model with that of the actual statistical data. For comparison we depict in Fig.~\ref{fig5a} the histograms of the actual wealth distribution in the USA (in blue), together with our results from 10 realisations of the network with 500 agents, without and with trust included. It should be noted that our model predicts a concentration of wealth in the hands of few agents, regardless of the role of trust, although the concentration of such agents is much less pronounced than what we see in reality. In the calculation including trust the distribution amongst middle and bottom percentages agree quantitatively with the data, but we are not able to find a good match with data for the upper 40$\%$ of the wealth. This could be due to the small size and randomness of our model network, nevertheless the trend is already noticeable.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig5}
\caption{Wealth distribution in the USA (in blue, taken from~\cite{dom}) and numerically calculated wealth distribution from our model (in yellow) including trust (top) and without including it (bottom). In both calculations the average price was $\bar{s} = 1.2$ and the spread $\mathrm{d}s = 0.4$.}
\label{fig5a}
\end{center}
\end{figure}
In Ref.~\cite{usa} one finds an interesting exercise in which people are asked to construct wealth distribution that they consider ideal, and also an estimated distribution, based on their information, and these data are compared with actual data for wealth distribution in the USA. In Fig.~\ref{fig5b} we compare the published results with our calculations.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=6.5cm]{fig6}
\caption{Ideal, estimated, and actual wealth distribution in the USA (in blue, taken from~\cite{usa}) compared with numerically calculated wealth distribution from our model (in yellow) using the parameters indicated in each histogram.}
\label{fig5b}
\end{center}
\end{figure}
Here we find that our calculation results seem to compare very well with the ideal distribution when trust is included and no spread of prices is allowed. This situation probably represents a country that has strict control of unique prices for goods, set by the government or some monopoly. The estimated distribution is well reproduced with a rather small spread of prices and having trust included. However, the actual situation is somewhat disappointing, since the best fit is with a large spread of prices. In the bottom right panel we show the situation in which the network agents lack trust in their transactions thus rendering the outcome unrealistic.
The situation is quite different for such egalitarian societies as Denmark, in which case we have found that for very small spread of prices and including trust our model agrees extremely well with the actual data (taken from ~\cite{din}) of the income distribution for the 1992 statistics. Using data from the money variable only, we show the comparison in Fig.~\ref{fig6}, which turns out to be very favourable.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig7}
\caption{ Actual income distribution in Denmark (1992) (in blue, taken from~\cite{din}) compared with numerically calculated wealth distribution from our model (in yellow) with trust included and having a price spread of $\mathrm{d}s = 0.02$.}
\label{fig6}
\end{center}
\end{figure}
Another maybe better way to compare our results with real statistical data is to investigate the distribution of money in different systems, since it depends on the actual mechanisms of acquiring money. For instance, there are data on the annual income of people in Europe, which ranges from 0 to millions of euros. In Fig.~\ref{fig7}(a) we compare the actual data (in red) with a histogram from our numerical calculation with 1000 agents but without including trust. In Fig.~\ref{fig7}(b) we compare the same data with the results of a calculation with 250 agents and a dispersion of prices twice as large as in (a) but once again without including trust. Both these calculations do not seem to fit with the real data, neither do averages over many realisations. \\
However, a rather different situation is encountered in a closed system as we have found out when we investigated data from the annual salaries that all the players in the NFL earned in two different years. In Fig.~\ref{fig7}(c) we show in red the distribution of salaries in 1998 and compare it with a numerical result without including trust. In Fig.~\ref{fig7}(d) we show the numerical results for the same system including trust and compare them with the NLF data from 2011. Here we can observe that both distributions fit the data fairly well, which allows us to think that trust does not play much of a role in the mechanisms of deciding salaries in a system like NFL, in which a few of the star players began to receive exaggeratedly good salaries.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig8}
\caption{(a) Histogram of the annual salary of european people (in red) compared with a numerical calculation without including trust in a system with 1000 agents (blue). (b) Same as (a), but the calculation was performed in a system with 250 agents. (c) Histogram of the salary of players in the NFL in 1998 (in red), compared with results from the model without including trust effects. (d) Histogram of the salary of players in the NFL in 2011 (in red), compared with results from the model including trust effects. }
\label{fig7}
\end{center}
\end{figure}
In order to investigate such a situation in more detail we have adopted the approach presented by Jun-ichi Inoue et al. ~\cite{bikas}, in which they define indices to measure social inequality in various fields, including income and trading. In Fig.~\ref{fig8} we present the inequality for our model calculations. We see that these Lorentz inequality curves vary quite sensitively with the spread of prices of the goods. We also see that the real situation presents itself as rather unequal such that for the same spread of prices, the lack of trust generates more inequity. For the curve tagged with an arrow we found $g=0.5611$ and $k=0.701$, which is very similar to the indexes calculated for USA: $g=0.54-0.6$ and $k=0.69-0.71$~\cite{bikas}. We have also detected that the results do not vary much when the size of the network is increased to 1000.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig9}
\caption{Lorentz curves from our calculations for several values of the spread of prices. The straight black line corresponds to perfect equality. The curve indicated by the arrow matches with real data. This corroborates our claim that trading without trust makes wealth distribution skewed or unequal.}
\label{fig8}
\end{center}
\end{figure}
\subsection{Dynamical behaviour}
As for the dynamical behaviour of trading there are data available for the wealth share of various countries~\cite{WTDB}. In Fig.~\ref{fig9} we show the top 1, 5, and 10\% countries' income share, and compare it with the numerical calculations for a network size of 500 agents. All the calculations were set to run up to 200000 iterations, of which only half of them were selected for the comparisons. The parameters of the model $\mathrm{d}s$ and $t_0$ were varied to find the best fit.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=\columnwidth]{fig10}
\caption{Comparison between numerical results and actual data of the top 10 \% income share with data taken from ref.\cite{WTDB} for six selected countries.}
\label{fig9}
\end{center}
\end{figure}
For all the cases a value of $t_0 = 400$ in the model calculation is found to fit well with the data, except in the case of Italy, where we chose $t_0 = 600$ for the best fit. It is interesting to notice that the case of China is the only one that fits better with the 100000 initial iterations of the calculation, and the dispersion of prices is smaller ($\mathrm{d}s = 0.15$). All the other countries are best fitted once the variables attain a final distribution, which occurs during the last 100000 iterations. This could reflect the fact that China is a newly emergent economic power and that their rules of trading are tighter. For Japan and Australia the dispersion of prices is larger than for China ($\mathrm{d}s = 0.2$). This could be interpreted such that these countries have more free trading rules, in which case one could vary prices more widely without loosing competitiveness.
For developed countries with long history in economic traditions, $\mathrm{d}s = 0.4$ is quite large, probably reflecting the influence of many strategies to allow prices to vary without the loss of competitiveness. For instance the dispersion of gas prices nationally in the USA ranges from 1.31 to 2.37 dollars per gallon, and the prices are unevenly distributed geographically ~\cite{gas}. Observe that USA and Canada are practicable indistinguishable, which is to be expected as they are similar and tightly linked. The case of Italy, the only European country selected, is interesting, since it is the only one in which a good fit is obtained by increasing the time scale for price changes. this seems to suggest that in Italy and similar European countries prices tend to change more slowly than in the very dynamic American economies.
Interestingly, the fit for the 1\% top population is not as good as the others, meaning that our model predicts less extremely rich people than in reality there is. The extreme social and economic inequality of the present world is most likely due to the fact that the economic alliances are not random (as in our model) and impose some trading preferences, other than the ones considered in our model. Also, our approximation of a conserved system
result in a constraint on the amount of goods or money an individual could gather, and this constraint is not present in the actual economic picture, in which money could be printed and goods could be produced and destroyed. In the real world countries cannot be considered as closed systems, although the global economy could be considered as such.
\section{Conclusions}
\label{conclusion}
In conclusion our agent-based trading model gives rise to results that overall seem to compare very favourably with the findings from the real data, even though the model takes into account only a subset of the known factors affecting trading. One striking result is the effect of trust in trading relations. First of all, it was found that trust reinforces trading transactions in such a way that the network with active links remains fully connected when trust is included, and becomes a set of disconnected graphs if it is not. Secondly, trust helps to make society more even and it is seen that the distribution of wealth is fairer when trust plays a role in trading. If trust is not included then the society seems to have a number of poor people in debt, unlike in a society with trust.
One important conclusion of this work is that even in the simplified case of having a conserved system, agreement with real data on the distribution of wealth is not only possible but also quite good. This could be interpreted indicating that including the production and deterioration of goods and money, which is essential to the idea of creation of wealth~\cite{galle}, does not seem to be a fundamental property of the economy in general. Furthermore, as far as the distribution of wealth is concerned, the fundamental issue seems to be the spread of prices, rather than the production of wealth. It also turns out that our model predicts inequality in a closed economy without production from simple rules of buying and selling between agents, which illustrates the fact that inequality can arise naturally in a rudimentary economy.
It should be noted that contrary to the general trend in classical economic models to consider various representative classes of agents (consumers or producers)~\cite{st} we have here considered a single class of heterogeneous trading agents. The behaviour emerging from such economy is an aggregate of individual decisions. Hence it seems that an unbalanced economy emerges from particular decisions of individuals~\cite{art}.
It is also important to recognise that the structure of real trading networks is more similar to a scale free networks than to random networks, which stays fixed during the dynamic trading process. In order to test the effect of topologically changing network structure we have introduced a rewiring scheme in which the agents whose connections are not working are deleted from the network, and new agents are added following a scale free method. As a result we have found that the final network structure evolved after a long run of the dynamics resembling roughly a scale free network. It is found that the structure of the final network still depends very much on the trust variable $R$, namely if $R=1$ for all agents, then agents that become "hubs" at a certain time are very likely to be deleted, and eventually the network is fragmented into small isolated trading groups. On the other hand, if one includes $R$ in the dynamics, the hubs that are formed remain trading and the network, although changing, remains cohesive. The network that one obtains after many rewiring processes has always all the links working perfectly (black lines). However, the wealth distribution in the network is very similar to the ones reported here without rewiring, and thus we decided to leave the detailed study of the rewiring problem for the future, since the above-described rewiring scheme breaks the conservation of wealth, which is an essential assumption in our current model. Our justification to do so is that in the calculations shown here the time span for transactions is not long enough to rewire the network. \\
As the main message of this work we would like to suggest that in trade the prices seem to be the main cause of impoverishment. Wide spread of prices tends to augment the differences between poor and rich people, and the results seem much more sensitive to a change in the allowed spread of prices than to the average price. Also we conclude that trust seems to be important in regulating trading, since it is a way in which agents decide to trade preferentially amongst the agents they are linked with. In indiscriminate transactions without including trust, the network turns out to be disrupted, while with trust few links are always reinforced. In this way trust could be considered to favour the appearance of monopolies, since the agents that have a good history seem better off regardless of the high prices and quality of goods, and are thus in position of engulfing small traders.
As a final remark we could conclude that our simple model conserving the amount of goods and money is able to reproduce some salient features found in trading networks, with the additional advantage that we could fairly easily add new features to the model and analyse them in depth.
\section*{Acknowledgments}
RB, ERG, and TG would like to acknowledge financial support from Conacyt (Mexico) through project 179616 and KK from Academy of Finland through project 276439. We are grateful to Prof. Larissa Adler, whose original ideas were the primary source of inspiration for this work. We are also grateful to Prof. Juan M. Hern\'andez for careful reading of the manuscript and valuable comments and criticisms.
|
Subsets and Splits